paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
1906.01957 | 1 | 1906 | 2019-06-05T11:52:50 | Maximizing Energy Battery Efficiency in Swarm Robotics | [
"cs.MA",
"cs.RO"
] | Miniaturization and cost, two of the main attractive factors of swarm robotics, have motivated its use as a solution in object collecting tasks, search & rescue missions, and other applications. However, in the current literature only a few papers consider energy allocation efficiency within a swarm. Generally, robots recharge to their maximum level every time unconditionally, and do not incorporate estimates of the energy needed for their next task. In this paper we present an energy efficiency maximization method that minimizes the overall energy cost within a swarm while simultaneously maximizing swarm performance on an object gathering task. The method utilizes dynamic thresholds for upper and lower battery limits. This method has also shown to improve the efficiency of existing energy management methods. | cs.MA | cs |
Maximizing Energy Battery Efficiency in Swarm
Robotics
Anthony Chen, John Harwell, and Maria Gini
Department of Computer Science and Engineering, University of Minnesota
{chen4714,harwe006,gini}@umn.edu
Abstract. Miniaturization and cost, two of the main attractive factors of swarm
robotics, have motivated its use as a solution in object collecting tasks, search &
rescue missions, and other applications. However, in the current literature only a
fewpapersconsiderenergyallocationefficiencywithinaswarm.Generally,robots
recharge to their maximum level every time unconditionally, and do not incorpo-
rate estimates of the energy needed for their next task. In this paper we present
an energy efficiency maximization method that minimizes the overall energy cost
within a swarm while simultaneously maximizing swarm performance on an ob-
ject gathering task. The method utilizes dynamic thresholds for upper and lower
battery limits. This method has also shown to improve the efficiency of existing
energy management methods.
Keywords: Swarm Robotics · Energy Efficiency · Foraging.
Introduction
1
Swarm intelligence has been developed through the observation of cooperation in natu-
ral swarms, such as fishes, birds, and bacteria [2]. In the context of a foraging task, ant
colonies have been studied as they can execute simple, efficient, localized foraging algo-
rithms that result in an intelligent division of labor based on assigning roles according to
pastperformance[4].Antscanalsocommunicatewithoneanotherthroughpheromones,
through which they achieve an advantageous collective foraging behavior. Similarly, in
swarm robotics, virtual methods to imitate pheromones have been simulated to commu-
nicate between robots [8]. In general, localized communication is necessary to ensure
system scalability to thousands of robots [4].
A common application in swarm robotics is foraging, in which robots search widely
for resources to bring back to a central location (the "nest"). When comparing foraging
methods to determine which one is most suited to a particular application, some of the
desired features are simplicity, scalability, decentralization, sensing, and parallelism [6].
Systems with these properties are much better equipped to handle applications with
multiple objectives in dynamic environments, such as post-disaster relief and geological
surveying [12].
In most prior work, the metric for efficiency is the time spent foraging [11]. Time
spent foraging does not fully encapsulate the cost it takes to perform the task, because it
ignores actual energy used or stored, as well as the time taken to re-charge a given robot.
Anthony Chen, John Harwell, and Maria Gini
2
Although it is important for swarms to maintain an optimized speed and efficiency [12],
time is not the only metric for energy usage. Another measure is the remaining battery
residing in each robot. It is important to maintain an efficient use of the energy in the
robots as well as the energy being depleted while foraging. If robots were charged to
100% without consideration to how much energy is actually needed, the swarm would
be wasteful. It would be wasteful to charge robots to their maximum capacity if it was
alreadyknownthattherewereresourcesnearthenest.Thatrobotcoulduseitsremaining
energy to perform those simple tasks. Other robots that are struggling to find resources
in other areas would be the ones that have to pay attention to getting charged and using
more of their maximum capacity.
Very few papers consider the energy in the battery when measuring energy effi-
ciency, so when they try to move the algorithms to the real world, their algorithms might
be impractical. Previous work [7,10,5] only focused on measuring energy during forag-
ing, and does not account for the unused energy in each robot's battery. That energy is
wastedforthetask andthereforetheswarmis lessefficientthananenergy-awareswarm.
Themostcommonlyusedmetricforcostissearchandretrievaltime;anadditionaluseful
metric would include the measurement of the energy remaining in each robot's battery.
Regardless of performance on a foraging task, systems need to be energy aware, as
the "income" from foraging (i.e., how many objects are gathered in a given amount of
time) needs to outweigh the cost of obtaining it [1]. In this work, cost is measured in
terms of swarm energy consumption.
As a motivating example, consider a swarm of robots in which each robot has a high
accuracy image recognition camera, making the swarm as a whole capable of highly
preciselocalizationandmapping.Realtimeimagerecognitionishighlycomputationally
expensive while the camera hardware attached to each robot is physically expensive.
Such swarm would necessarily consume a high amount of energy per-capita, potentially
making its usage intractable [3]. Furthermore, the production of hundreds of thousands
of such robots would likely be sub-optimal in terms of the increased "income" received
during foraging vs the additional energy expenditure.
Some of the most popular applications of swarm-robotic foraging are in rescue mis-
sions and agricultural foraging [6]. Such tasks require an efficient use of energy as un-
necessary foraging may deplete energy that could instead be used in other tasks. For
example, energy savings achieved during foraging could be used later by the swarms to
further process resources if there was a finite amount of energy available for swarms to
use for recharging. In addition, in cases of non-uniform object distributions, robots in
areas with a lot of easy to access resources would not need to be charged as often. Non-
uniform robot charging would enable robots to be charged with the necessary energy for
their next foraging task. Collectively the swarm would then be more flexible and robust
as it can invest energy where it is most needed.
In this paper, we introduce a method in which each robot stores energy thresholds
and capacity variables to indicate how much energy should be allocated and used during
foraging.In ourproposedapproach, thesevariables adaptbasedon eachrobot'senviron-
mental encounters (i.e., obstacle collisions, resource pickup, etc) every time the robot
returns to the nest after foraging. Using this method, the swarm obtains a collective for-
aging strategy that maximizes both overall energy efficiency and successful income of
Maximizing Energy Battery Efficiency in Swarm Robotics
3
resources simultaneously. To validate the effectiveness of our method, we conducted
experiments that investigated energy and performance patterns over time as the amount
of energy consumed by each robot changes depending on the environment. Our results
show we can reach our goals and be better than prior work because of the flexibility of
our method, which can be used to improve the efficiency of existing methods, including
other rivaling energy efficiency methods.
2 Related Work
Liu focused on energy efficiency by changing the foraging time for each robot [7]. The
time spent foraging for resources depended on various cues such as personal successful
food retrievals, collisions with teammates for food, and success among other robots in
foodretrieval.Sothemethodwasabletofindanadaptableandoptimaltimeforforaging,
but did not optimize the amount of energy allocated for it.
Stirling took a novel approach at finding energy efficient searching algorithms for
flying swarms in indoor environments [10]. The strategy involves having robots create a
network of beacons that communicate with each other to direct where other robots in the
swarmshouldgo.Whenanexploringrobotarrivesatanunexploredlocation,itbecomes
a beacon to help sense for the other exploring robots. They found that launching the
robots incrementally rather than all at once decreased total energy consumption as well
as collision rate but increased search time. Their method provides a trade-off between
energy consumption and search time. Once again the metric is energy consumption and
not energy allocation so it ignores the unused charge in each robot after the task was
done.
Labellatookinspirationfromants.Hecreatedanadaptationmethodthatcontrolsthe
number of robots foraging in the environment [5]. Each robot has a probability variable
that increases and decreases based on the number of successes and failures the robot had
when foraging. The probability variable dictates the probability that the specific robot
would leave the nest and start foraging. This allows for robots who consistently find and
retrieve prey to keep foraging, while also eventually reaching an equilibrium or state in
which the necessary number of robots are active. Thus, it efficiently uses energy when
needed, similarly to [7,10] in that they measure the time it takes to finish foraging but
not how much energy was already allocated.
3 Proposed Method
3.1 Problem Statement
We assume K foraging robots initialize at a nest. The environment outside the nest has
a finite R number of resources. Each of the K robots needs to forage for one of these
R resources and bring it back to the nest without using more energy than needed. The
robot can charge at the nest as the nest has an infinite source of energy. If the robot's
battery level reaches 0, it will be permanently lost. The total time, , it takes for a robot
to complete a single round of foraging from leaving the nest to returning to the nest
can be broken down into time while searching, , and time while retreating, . The
Anthony Chen, John Harwell, and Maria Gini
4
energylevel decreasesatdifferentratesdependingonwhethertherobotissearchingor
retreating. So the overall change in energy level for a single robot k during is modeled
as:
(1)
Ɗ = −
Energy spent Searching
−
()
Energy spent collecting if successful
−
Energy spent Retreating
where istheenergylostrateforsearchingand istheenergylostrateforretreating.
is the energy cost for finding and collecting a resource, , while () is a binary function
that returns if 1 or 0 upon successful or unsuccessful resource retrieval.
Collectively, the swarm solves the following multi-objective optimization problem:
(2)
We want to maximize the amount of successful resource retrievals while also mini-
ࢠ
mizing the energy allocation of every robot.
Ɗ()
ࢣ
max
()
min
3.2 Adaptive Energy Parameters
With Ɗ wedenotethechangeintheenergylevelofarobot afteraroundofforaging.
We now introduce how the battery energy is partitioned for use in the next round.
() = + ()
Energy allocated = Energy to Forage + Energy to Retreat
(3)
(4)
First, energy is allocated for foraging resources (searching + collecting). Second,
energy is allocated for travelling back to the nest to ensure a safe return.
The energy allocated for foraging resources will be represented by a capacity . The
size of the capacity is described by the difference between an upper threshold, (),
and lower threshold, () value. If the energy level reaches below (), then the
energy allocated for retreating back to the nest is used. Therefore, the amount of energy
allocated to return to nest will be determined by what level () is at.
The amount of energy for foraging and the amount of energy for retreating needs to
be adaptable to changes in the environment through the changes of values and (),
respectively. To ensure energy is allocated properly for the task, both values will be
changed based on successful foraging (), number of robots encountered, and speed
of task completion.
Ɗ = − () (0, − Ɗ)1
If robot finishes too early
+ ¬()2
If no resource found
+
3
If encountered other bots
Each of these three components is accompanied by weight variables 1,2, and
3.Whentheenvironmentsizeisheldfixed, shouldincreasewhen(1)therearemany
(5)
Maximizing Energy Battery Efficiency in Swarm Robotics
5
Fig. 1. Diagram of battery and battery variables
robots in the environment and (2) there are few resources in the environment. It should
then decrease when the opposite is true. A high robot density in an environment means
there is more competition among robots for resources so more energy is needed to look
for those resources. The competition for resources is measured by collision encounters
with other robots. Similarly, few resources also means longer time to forage and so more
energy is needed. An increase in foraging also results in farther distance from the nest
which requires () to change similarly:
Ɗ() = − (0, − Ɗ)1
If robot finishes too early
+
¬()2
If no resource found
+
3
If encountered other bots
(6)
Each of these three components is accompanied by weight variables 1,2, and 3.
The new upper energy threshold () can then be determined by and ().
() = () +
(7)
A clear emergent property of our proposed algorithm is per-robot energy allocation
such that upon return to the nest a robot has very little energy left. This is desirable
because it indicates that very little wasted energy is used and we efficiently allocate only
what is needed for the task. So even if the energy source gets cut off or taken away, each
robot should be able to finish their foraging task with little wasted energy.
3.3 Finite State Machine Implementation
Our proposed mathematical model is implemented with the help of a per-robot state
machine, described below. A robot can be in one of 4 states:: Charging, Searching,
Collecting, and Retreating.
6
Anthony Chen, John Harwell, and Maria Gini
=
searching,
collecting,
retreating,
charging,
if = () ࢱ () = 0
if robot encounters resource
if <= () ࢱ () = 1
if k is in the nest
(8)
Eq. 7. State Transitions based on Energy Levels/Thresholds
The robot starts in Searching from the nest until it finds a resource. Once a resource
is found, the robot tries Collecting it. If it succeeds, it begins Retreating to nest with
it. If not, it returns to Searching for another resource. If while Searching, the robot's
energylevelisbelowthelowerenergycapacitythreshold (),thenitstarts Retreating
without a resource. Once the robot has retreated back to the nest, it starts Charging until
the energy level is at the desired upper energy capacity threshold (). Once fully
charged, the robot starts Searching and a new round of foraging begins. A nest delay
timer is added to the robot while it is charging to simulate the time delay it takes to get
a robot charged and ready.
Fig. 2. Diagram of finite state machine of behavior for a battery efficient swarm
How each robot runs each round of foraging and cycle of states for a robot is defined
in the Adapting Battery Algorithm (see Algorithm 1). The algorithm also handles the
adaptive behavior of the energy capacity and thresholds while the robot is in the nest.
3.4 Adaptive Battery Critical Point
At some point, the adaptive thresholds of the battery may increase to cover the entirety
of the battery. This is because as time passes, most of the resources will have been
foraged while the number of robots remains the same. To find the remaining resources,
the robots will have to use their maximum battery capacity to forage. At this point, the
robot cannot increase its thresholds any more which means it is near the end of foraging
most of the resources. The only threshold that can be modified is the energy threshold
Maximizing Energy Battery Efficiency in Swarm Robotics
7
if == () then
end if
ࢎ −
ࢎ −
ࢎ
() ࢎ () − (0, − Ɗ)1 + ¬()2 + 3
ࢎ − () (0, − Ɗ)1 + ¬()2 + 3
() ࢎ () +
ࢎ ()
() ࢎ 0
ࢎ 0
else if == Ò then
ࢎ Ò
if > 0 then
ࢎ − 1
else
Algorithm 1 Adapting Battery Algorithm
1: robot k is initialized with thresholds()and().
2: ࢎ ()
3: () ࢎ 0
4: ࢎ 0
5: nest delay ࢎ 20
6: while > 0 ࢱ all K robots inactive or dead do
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
end if
36: end while
else if == Ò then
else if < () and = Ò then
else if Encounter resource r then
else if == then
ࢎ
() ࢎ 1
= −
else if Encounter robot then
else if Encounter Nest then
ࢎ + 1
= Ò
ࢎ 20
for when to retreat. This threshold will update similarly to Equation 5 and Equation 6.
We call this period Energy Efficiency Endgame (EEE). We investigate three different
options for how each robot will handle being in this state.
Well-informed. Inthismethod,weassumethateachrobotiswell-informed.Thismeans
that each robot no longer needs to change the thresholds of the battery for when it needs
to retreat. Instead, each robot will simply continue but with an increasing nest delay
timer each time the robot re-enters the nest. The nest delay timer increases by seconds
after each round of foraging.
Anthony Chen, John Harwell, and Maria Gini
8
Ill-informed. In the ill-informed method, the robot believes that it has autonomously
decided on a wrong threshold for when to retreat to the nest. So along with increasing
the nest delay timer, the robot also continues to adapt the lower energy threshold using
equation 6.
Null-informed. Null-informed is an option that makes each robot stop foraging and
remain still in the nest. Once the robot has reached its battery full capacity, it stops and
relies on the still adapting robots to find the remaining resources.
Table 1. Table of Energy Efficiency Endgame (EEE) options
EEE Option Increase nest delay Adaptive
Stops When
Well-informed
Ill-informed
Null-informed
No robot is foraging
No robot is foraging
Each robot reaches EEE
Yes
Yes
No
No
Yes
No
3.5 Experimental Framework
The experiments in this paper were implemented and executed using the ARGoS sim-
ulator, which allows real-time simulation of large swarms of robots [9]. The simulation
is tested in a three-dimensional space with s-bot models.
The following assumptions were made for all experiments:
-- The robots are homogeneous.
-- The robots cannot communicate.
-- The number of resources does not affect a robot's performance.
Fig. 3. Visualization of the simulation with 16 robots.
Maximizing Energy Battery Efficiency in Swarm Robotics
9
-- The environment is flat and open with no unknown objects/obstacles.
-- The nest has unlimited storage capacity.
-- Robots can be charged to a specific amount.
-- The resources in the environment respawn until at least 100 are collected. At that
point, 25 resources will remain.
-- The energy expenditure rates for searching and retreating are constant.
-- The energy expenditure Ɗ of a robot can be accurately measured and recorded
-- All resources are identical.
The goal is to find a better method for energy efficiency. Let be the number of
collected resources and be the total foraging time. The efficiency of a swarm is often
defined as:
=
(9)
However, in this work, we also consider the remaining energy in each battery upon task
completion as part of our efficiency definition (i.e., robots try to ensure that the foraging
ends with no leftover energy). Thus, our efficiency is represented as:
ࢣ
is energy remaining in battery of robot .
(10)
is energy depleted
where denotes all robots, denotes a single robot instance,
and
Table 2 shows the values for the adapting equation that we used in our simulations.
These values were empirically chosen to balance exploration and exploitation. For ex-
ample, the values 1 and 1 were chosen especially small as they represent the coeffi-
cient for number of robot collisions. Because a robot could experience many collisions
injustoneforaginground,wewantedthisweighttobesmalltoreducenoiseandprevent
random collisions to overshadow the other parts of the model.
߰ =
+
Table 2. Values used in the experiments
Init () Init 1 2 3 1 2 3
0.5 0.3 0.1 0.005 0.2 0.1 0.005 10
0.3
4 Results
For each method, we experimented with swarm sizes of ࢠ {2 ࢼ = 1 … 8} in or-
der to observe efficiency across scales. Each experimental run was terminated when
the swarm successfully collected all the resources or the swarm fulfilled the Energy
Efficiency Endgame stopping criteria. Performance is measured by as described in
Eq. (10). For each experiment, 20 simulations were averaged. In Fig.4, we can see the
10
Anthony Chen, John Harwell, and Maria Gini
Fig. 4. Graph of the efficiency of the different battery allocation methods
battery energy allocation method plotted with three different lines each representing a
different EEE method. A fourth line-plot was added that forages without any energy ef-
ficiency strategy to represent the baseline to show how our method improves efficiency.
The results indicate that each energy battery allocation method improved overall swarm
efficiencywiththeNull-informedEEEmethodraisingoverallefficiencyconsistentlythe
most with a variety of swarm sizes. The Null-informed version most likely performed
the best because it prevented robots from re-entering the environment when only a few
robots were needed to finish collecting the resources. In some cases we can see that hav-
ing battery conscious robots in the swarm doubles the energy efficiency use. However,
it appears that it is not always consistently the best. Specially as we increase the swarm
size, the difference in efficiency between a conscious and non-conscious battery energy
swarm decreases. For higher swarm densities, this method is not that advantageous,
most likely due to competition among robots to retrieve resources. When competition
is high, the entire battery capacity is needed so trying to partition the battery usage is
futile. But it is important to note that although it appears that difference between energy
efficiency is low, it is still between twice to three times more efficient than without it.
So the small difference in energy efficiency improvement is more due to the arena ra-
tio than the efficiency method itself. In a second set of experiments, we explored how
the efficiency improves when applying our conscious battery energy allocation method
compared to other existing energy efficiency methods. Previously discussed in Section
2, Labella's group used an adapting probability value to control when a robot should
leave the nest [5]. We implemented the algorithm with the same experimental values
= 0.0015, = 0.05, = 0.033, and = 0.005.
Similarly, Liu's group had an adaptable time variable that denoted how long the
robot should spend searching [7]. However, since they used pheromones as social cues
for their time variable, we will have to modify it since we are not using pheromones.
Maximizing Energy Battery Efficiency in Swarm Robotics
11
Fig. 5. Efficiency of different methods with or without battery allocation methods. Dotted lines
(red and blue) are related energy efficiency works that do not use our battery conscious allocation
strategy. The solid lines (green and yellow) are those related works methods in conjunction with
our method.
Instead we simply have the robot keep track its own successes and failures rather than
those of other robots. The constant values used from their experiment equation remain
the same.
In Fig 5, the dotted lines denote related approaches without our conscious battery
energy method and the solid lines denote related approaches with the conscious battery
method. Our method almost consistently increases the energy efficiency in every exper-
iment compared to the other energy efficient methods. This is indicated in the graph by
the fact that the solid lines (methods with Null EEE) have higher energy efficiency than
the dotted lines (methods without Null EEE). It is also interesting to note that if you
compare 5 with 4, our battery conscious strategy by itself is about as efficient or almost
as efficient as the related works methods. And then combining the two improves effi-
ciency significantly. This shows the advantage of using our conscious battery allocation
in future swarm foraging methods.
5 Conclusions and Future Work
Wepresentedanadaptive,energyefficient,andenergy-awareapproachtoforagingswarm
robotics. Through conscious allocation of energy our method has shown that energy ef-
ficiency can be greatly increased when taking into account the energy allocated in the
battery. When using battery conscious efficiency methods such as the Null-informed
EEE, swarm strategies increase energy efficiency significantly.
One direction for future work is to further explore how combinations of energy ef-
ficient methods perform. It would be interesting to see if combining our method with
Anthony Chen, John Harwell, and Maria Gini
12
otherexistingefficiencymethodwouldincreaseefficiencyoriftoomuchcomplexitywill
produce a diminishing return. Another aim could be to explore local communication or
signaling among robots to see if there is a way to communicate energy information.
With this, robots could adjust energy allocation and usage with other robots are taken
into account instead of just the robot itself.
Another direction is to explore multiple nest locations each with limited charging
capacity. In real life, each nest for swarm robots is not going to have unlimited energy
supply. So a future research topic is to explore energy efficiency when supply is limited
and perhaps explore robot communication to direct which nest has more energy supply.
As listed in the Experimental Framework (Section 3.5) we assume that the energy
expenditure rates for searching and retreating are constant. We also had our sim-
ulation be able to accurately calculate and record the total expenditure Ɗ. The issue
with these two assumptions is that it is difficult to implement this model on robot hard-
ware due to how accurate the real-time energy measurements must be. One direction we
would like to explore is trying to run real life swarm robot experiments that can handle
this model. One possibility is having the energy levels be measured in time, so the en-
ergy remaining is measure in how many seconds left before the robot dies. This would
of course require lots of battery testing to see how the energy expenditure rates change
in real time.
In order to facilitate future research and collaboration, the code for this work is open
source. It can be found at https://github.com/swarm-robotics.
Acknowledgments
We gratefully acknowledge Amazon Robotics, the MnDRIVE RSAM initiative at the
University of Minnesota, and the Minnesota Supercomputing Institute (MSI) for their
support of this work.
References
1. Campo, A., Dorigo, M.: Efficient multi-foraging in swarm robotics. In: Almeida e Costa,
F., Rocha, L.M., Costa, E., Harvey, I., Coutinho, A. (eds.) Advances in Artificial Life. pp.
696 -- 705. Springer, Berlin, Heidelberg (2007)
2. Dorigo, M.: Swarm robotics: The coordination of robots via swarm intelligence principles.
In: Hinchey, M., Pagnoni, A., Rammig, F.J., Schmeck, H. (eds.) Biologically-Inspired Col-
laborative Computing. pp. 1 -- 1. Springer US, Boston, MA (2008)
3. Haverinen, J., Parpala, M., Roning, J.: A miniature mobile robot with a color stereo camera
system for swarm robotics research. In: Proceedings of the 2005 IEEE International Confer-
ence on Robotics and Automation. pp. 2483 -- 2486 (April 2005)
4. Khaldi, B., Foudil, C.: An overview of swarm robotics: Swarm intelligence applied to multi-
robotics. International Journal of Computer Applications 126, 31 -- 37 (09 2015)
5. Labella, T.H., Dorigo, M., Deneubourg, J.L.: Efficiency and task allocation in prey retrieval.
In: Ijspeert A.J., Murata M., W.N. (ed.) Biologically Inspired Approaches to Advanced Infor-
mation Technology, pp. 274 -- 289. LNCS volume 3141, Springer (2004)
6. Liekna, A., Grundspenkis, J.: Towards practical application of swarm robotics: Overview of
swarm tasks. Engineering for Rural Development 13, 271 -- 277 (01 2014)
Maximizing Energy Battery Efficiency in Swarm Robotics
13
7. Liu, W., Winfield, A., Sa, J., Chen, J., Dou, L.: Strategies for energy optimisation in a swarm
of foraging robots. In: Sahin, E., Spears, W.M., Winfield, A.F.T. (eds.) Swarm Robotics. p.
14âf26. Springer Berlin Heidelberg, Berlin, Heidelberg (2007)
8. Payton, D., Daily, M., Estowski, R., Howard, M., Lee, C.: Pheromone robotics. Autonomous
Robots 11(3), 319 -- 324 (Nov 2001)
9. Pinciroli, C., Trianni, V., O'Grady, R., Pini, G., Brutschy, A., Brambilla, M., Mathews, N.,
Ferrante, E., Di Caro, G., Ducatelle, F., Stirling, T., GutiÃ
rrez, A., M Gambardella, L.,
Dorigo, M.: Argos: a pluggable, multi-physics engine simulator for heterogeneous swarm
robotics (02 2011)
10. Stirling, T., Wischmann, S., Floreano, D.: Energy-efficient indoor search by swarms of sim-
ulated flying robots without global information. Swarm Intelligence 4 (Jun 2010)
11. Stirling,T.S.,Floreano,D.:Energy-timeefficiencyinaerialswarmdeployment.In:Proc.Int'l
Symp. on Distributed Autonomous Robotic Systems (DARS) (2010)
12. Tan, Y., Zheng, Z.Y.: Research advance in swarm robotics. Defence Technology 9(1), 18 -- 39
(2013)
|
1604.04728 | 1 | 1604 | 2016-04-16T11:42:38 | Reaching Unanimous Agreements Within Agent-Based Negotiation Teams With Linear and Monotonic Utility Functions | [
"cs.MA",
"cs.AI",
"cs.GT"
] | In this article, an agent-based negotiation model for negotiation teams that negotiate a deal with an opponent is presented. Agent-based negotiation teams are groups of agents that join together as a single negotiation party because they share an interest that is related to the negotiation process. The model relies on a trusted mediator that coordinates and helps team members in the decisions that they have to take during the negotiation process: which offer is sent to the opponent, and whether the offers received from the opponent are accepted. The main strength of the proposed negotiation model is the fact that it guarantees unanimity within team decisions since decisions report a utility to team members that is greater than or equal to their aspiration levels at each negotiation round. This work analyzes how unanimous decisions are taken within the team and the robustness of the model against different types of manipulations. An empirical evaluation is also performed to study the impact of the different parameters of the model. | cs.MA | cs | Reaching Unanimous Agreements within
Agent-Based Negotiation Teams with Linear and
Monotonic Utility Functions
Victor Sanchez-Anguix, Vicente Julian, Vicente Botti, and Ana Garc´ıa-Fornes
6
1
0
2
r
p
A
6
1
]
A
M
.
s
c
[
1
v
8
2
7
4
0
.
4
0
6
1
:
v
i
X
r
a
Abstract -- In this article, an agent-based negotiation model
for negotiation teams that negotiate a deal with an opponent
is presented. Agent-based negotiation teams are groups of agents
that join together as a single negotiation party because they
share an interest that is related to the negotiation process. The
model relies on a trusted mediator that coordinates and helps
team members in the decisions that they have to take during
the negotiation process: which offer is sent to the opponent,
and whether or not the offers received from the opponent are
accepted. The main strength of the proposed negotiation model
is the fact that it guarantees unanimity within team decisions
since decisions report a utility to team members that is greater
than or equal to their aspiration levels at each negotiation round.
This work analyzes how unanimous decisions are taken within
the team and the robustness of the model against different types
of manipulations. An empirical evaluation is also performed to
study the impact of the different parameters of the model.
Index Terms -- Negotiation teams, automated negotiation,
agreement technologies, multiagent systems.
I. INTRODUCTION
A negotiation team is a group of two or more interdepen-
dent individuals who join together as a single negotiation
party because of their similar interests and objectives related
to the negotiation and who are all present at the bargaining
table [1]. Therefore, this group of individuals unites because
their members share goals that are related to a negotiation
with an opponent. For instance, negotiation teams formed by
different stakeholders are usually sent to the negotiation table
when a company decides to sell a product line to another
company. Nevertheless, as it has been stated in social sciences,
negotiation teams are not necessarily unitary players since
team members may have different preferences regarding the
possible outcomes of the negotiation process [2]. Thus, given
the divergence in preferences between teammates, the team
has to agree upon, not only a negotiation strategy to carry
out with the opponent, but also upon those agreements that
are acceptable to the team. Despite being studied in social
sciences to some extent [1], [2], as far as we know, negotiation
teams have been overlooked by artificial intelligence research.
We argue that a negotiation team is also an element that
may be appropriate for some scenarios involving software
V. Sanchez-Anguix, V. Julian, V. Botti, and A. Garcia-Fornes are with
the Departamento de Sistemas Inform´aticos y Computaci´on, Universidad
Polit´ecnica de Valencia, Valencia, Spain, Cam´ı de Vera s/n, 46022 e-mails:
{sanguix,vinglada,vbotti,agarcia}@dsic.upv.es
This work is supported by TIN2008-04446, PROMETEO/2008/051,
TIN2009-13839-C03-01, CSD2007-00022 of the Spanish government, and
FPU grant AP2008-00600 awarded to V´ıctor S´anchez-Anguix.
agents. For instance, let us imagine an example based on
an electronic market for travel and tourism. In this system,
a group of friends (each friend is represented by a software
agent) has decided to go on a trip together. This goal requires
a negotiation with a travel agency agent. The fact that the
group of agents (group of friends) has a common and shared
goal (which is going on a trip together),
is clear, and it
requires an agreement with an opponent (the travel agency
agent). However,
it also seems reasonable to assume that
friends may have different preferences regarding the negotiable
trip conditions: hotel quality, price, number of days to stay,
etc. For example, while some friends may care more about
comfort, others may be more interested in money. In this
type of scenario, and specially in open multi-agent systems,
mediated preference aggregation is complicated since i) agents
may be inclined to exaggerate their preferences in order to
ensure a certain level of utility; ii) preferences are delicate
information which may not be revealed to anyone; iii) utility
functions may be different and require an extensive and costly
aggregation. Hence, mechanisms that allow an agent-based
negotiation team to handle intra-team conflict (divergences in
preferences) while trying to get a deal from travel agencies'
agents are needed. Thus, it is necessary to provide agent-based
models for negotiation teams.
This article describes a mediated negotiation model for
agent-based negotiation teams that negotiate with an opponent.
The negotiation model defines the communications protocol
with the opponent, what decisions are taken by the negotiation
team, and how and when these decisions are taken (i.e.,
team dynamics) [3]. More specifically, our preliminary study
presented about this model (Full Unanimity Mediated or FUM)
in [4] is extended further. FUM is able to guarantee unanimity
in decisions taken within the negotiation team as long as team
members share the same type of monotonicity for valuation
functions of the attributes. This assumption is relatively natural
in buyer-seller settings found in electronic commerce (e.g., a
team of buyers may value with the same type of valuation
function attributes like the price, the quality of the product, and
the time of dispatch). The proposed model relies on unanimity
rules regarding the opponent's offer acceptance and an iterated
offer construction process to determine which offer is sent to
the opponent. This article is organized as follows. First, we
describe our negotiation model. Then, we analyze how una-
nimity is assured within the team as well as the robustness of
this model against different types of attacks. Then, we evaluate
the empirical response of the proposed model depending on
the impact of the different model parameters and we analyze
possible incentives that team members may have to deviate
from the proposed behavior. We then relate our work to other
works found in artificial intelligence. Finally, we summarize
the conclusions of this work and discuss our future research.
II. NEGOTIATION MODEL
Traditionally, a negotiation model is composed of a negotia-
tion protocol, which defines the set of actions that are available
for agents at each instant, and the negotiation strategy, which
defines the decision making mechanisms employed by agents
during the negotiation. In this article, we present a negotiation
model where a negotiation team negotiates with an opponent.
Despite resembling a bilateral negotiation scenario, negoti-
ations that have teams as participants are slightly different
since team dynamics also play a key role. Thus, three dif-
ferent elements have to be specified when a team negotiation
model is proposed: negotiation protocol with the opponent, the
negotiation strategy used by the opponent, and the intra-team
negotiation strategy followed by team members in order to
decide the actions to perform during the negotiation process.
An intra-team strategy defines which decisions are taken by
the team, and how and when these decisions are taken. More
specifically, it is defined by the negotiation protocol followed
among team members within the team and by the strategies
followed by agents within the team. In this section, the general
assumptions of our negotiation model and our negotiation
model itself are described. Special attention is focused on
the interactions among team members, which are carried out
before and during the negotiation process and the negotiation
strategy followed by team members within the team.
A. General Assumptions
• In our model, a group of agents has formed a team A =
{a1, a2, ..., aM} whose goal is to negotiate a successful
deal with an opponent op. However, each team member
ai may have different preferences about the negotiation
issues.
• Communications between the team and the opponent are
carried out by means of a mediator that is trusted by
the team. This mediator sends team decisions to the
opponent and receives, and later broadcasts, decisions
from the opponent to team members. Thus, the fact that
the opponent is communicating with a team is not known
by the opponent, which only interacts with the trusted
mediator. The trusted mediator also performs other tasks
that allow team members to reach unanimous decisions
regarding the offer that
to the opponent and
whether or not the opponent's offer is accepted.
is sent
attributes whose domain is [0, 1]. Thus,
number of offers is [0, 1]n.
• The negotiation domain is comprised of n real-valued
the possible
• A complete offer is represented as X = {x1, x2, ..., xn},
where xi is the value assigned to the i-th attribute. The
notation X t
i→j is employed to indicate that X is the offer
sent by agent/team i to agent j at round t.
2
• Team composition will remain static during the nego-
tiation process. It is acknowledged that team members
may leave or join the group in certain specific situations.
However, membership dynamics is not considered in this
article, and it is designated as future work.
• All of the agents use linear utility functions to represent
their private preferences. Negotiation attributes are sup-
posed to be independent. Thus, the value of a specific
attribute does not affect the valuation of other attributes'
values. These functions can be formalized as follows:
n(cid:88)
Ui(X) =
wi,j Vi,j(xj)
(1)
j=1
Weights are normalized so that(cid:80)n
where Vi,j(.) is a monotonic valuation function that trans-
forms the attribute value to [0, 1], and wi,j is the weight or
importance that is given by the agent i to the j-th attribute.
j=1 wi,j = 1 holds for
every utility function. It is assumed that teammates share
the same type of monotonic valuation function (either
increasing or decreasing) for each negotiation attribute. In
contrast to team members, the opponent valuation func-
tion is always the opposite type of monotonic function.
Thus, if team members employ monotonically increasing
functions, the opponent will be modeled using monotoni-
cally decreasing functions. It is reasonable to assume this
model for valuation functions in e-commerce scenarios.
Buyers usually share the same type of valuation function
for attributes such as the price (monotonically decreas-
ing), product quality (monotonically increasing), and the
dispatch time (monotonically decreasing), whereas sellers
usually use the opposite type of monotonic functions
(monotonically increasing for price, monotonically de-
creasing for product quality, and monotonically increas-
ing for dispatch time). As for attributes' weights, it is
considered that each team member may assign different
weight/importance to each negotiation issue. Therefore,
differences among teammates are introduced by assigning
different weights to negotiation attributes. Nevertheless,
it should be highlighted that since team members share
the same type of monotonic function, if one of the team
members increases its welfare by increasing/decreasing
one of the attribute values, the other team members will
stay at the same welfare level or they will also increase
their welfare. Thus, there is potential for cooperation
among team members. Weights given by the opponent
to attributes may also be different to those given by
teammates. Agents do not know the form of other agents'
utility functions, even if they are teammates.
• The opponent has a private deadline Top, which defines
his maximum number of negotiation rounds. Once Top
has been reached in the negotiation process, the oppo-
nent will exit the process and the negotiation will end
in failure. The team has a private joint deadline TA,
which is common information for team members. Once
this deadline has been reached, the team will exit the
negotiation process and the negotiation will end in failure.
We consider that TA has been agreed upon by team
members before the negotiation process starts.
• The opponent has a reservation utility RUop. Any offer
whose utility is lower than RUop is rejected. Each team
member has a private reservation utility RUai, where ai
is a team member. This individual reservation utility is
not shared among teammates. Therefore, a team member
ai rejects any offer whose value is under RUai. In this
setting, reservation utilities represent the individual utility
of each agent if the negotiation process fails.
B. Negotiation Protocol with the Opponent
In this section, the negotiation process between a negotiation
team and an opponent is studied. The fact that one of the
parties is a team is transparent
to the other party. Thus,
superficially, the scenario resembles a bilateral negotiation
scenario. Because of this, we decided to model the interaction
between the team and the opponent as an alternating bilateral
negotiation process [5]. In this protocol, one of the two agents
is the initiator and sends the first offer to the other party
or responding agent. The responding agent receives the offer
and decides whether he/she accepts the offer or he/she sends
a counter-offer as response. If the responding agent sends a
counter-offer, the initiator agent has to decide whether he/she
accepts the counter-offer or not. If the counter-offer is rejected,
a new round starts and the process is repeated again until a
deal is accepted (successful negotiation) or one of the parties
decides to quit the negotiation since its deadline has been
reached (failed negotiation). In our negotiation model, we
consider that a trusted mediator is responsible for sending team
decisions to the opponent and broadcasting opponent decisions
to team members. Nevertheless, this mediator not only acts
as a coordinator but also helps the team to reach unanimous
agreements by means of an iterated process for offer genera-
tion and unanimity rules for opponent offer acceptance.
C. Opponent Negotiation Strategy
A negotiation strategy defines the decision-making of an
agent in a negotiation process. In this case, the negotiation
strategy is constituted by the concession strategy, which marks
the aspiration level of the opponent in terms of utility at
each negotiation round, and the acceptance criterion, which
determines whether the team offer is accepted or not.
• It is assumed that the opponent uses a concession strategy
to carry out during the negotiation process. A concession
strategy typically (although not necessarily) starts by de-
manding the maximum aspiration and, as the negotiation
process advances, the aspiration demanded tends to be
lowered. The amount of concession/reduction applied at
each step may depend on the specific tactic selected
by the opponent. In this article, our main focus of
interest is the general behavior of the proposed intra-team
strategy. Thus, a set of well-known negotiation tactics
was selected as the opponent negotiation strategy: time-
dependent tactics [6], [7]. We formalized time-dependent
tactics as suggested by [7]:
sop(t) = 1 − (1 − RUop)(
t
Top
1
βop
)
(2)
3
Fig. 1.
followed by the team in our model.
This figure represents an activity diagram of the general process
where t is the current negotiation round and βop is a
negotiation strategy parameter (concession speed) which
determines how concessions are made towards RUop. On
the one hand, when βop = 1 the concession is linear and
each negotiation round the same amount of concession is
performed, and when βop < 1 the concession is boulware
and very little is conceded at the start of the negotiation
process but the agent concedes faster as the negotiation
deadline approaches. On the other hand, when βop > 1
the concession is conceder and the agents concede fast
towards the reservation utility in the first rounds.
• The opponent uses an offer acceptance criterion acop(.)
during the negotiation process. It is formalized as follows:
if sop(t + 1) ≤ Uop(X)
otherwise
(3)
(cid:26) accept
reject
acop(X) =
where t is the current round, X is the offer received from
the team, Uop(.) is the utility function of the opponent,
and sop(.) is the opponent concession strategy. Thus, an
offer is accepted by op if it reports a utility that is equal
to or greater than the utility of the offer that op would
propose in the next round.
D. Intra-Team Strategy: Negotiation Protocol within the Team
The negotiation protocol followed by team members for
team communications can be divided into two different phases:
the protocol followed during the pre-negotiation, and the
protocol used during the negotiation process. Both of them
are thoroughly described below. A general overview of the
process followed by team members can be observed in Fig. 1.
1) Pre-negotiation: In the pre-negotiation, team members
confidentially share certain information about their preferences
with the trusted mediator. Each team member specifies which
attributes' decision rights it is willing to hand over when
(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:3)(cid:6)(cid:7)(cid:8)(cid:9)A(cid:8)(cid:9)(cid:7)(cid:5)B(cid:3)(cid:6)(cid:7)(cid:8)(cid:9)A(cid:8)(cid:9)(cid:7)(cid:5)C(cid:3)D(cid:9)A(cid:8)(cid:7)(cid:2)AEFE(cid:16)(cid:8)(cid:7)(cid:17)(cid:3)A(cid:18)C(cid:3)(cid:18)(cid:19)(cid:3)(cid:2)E(cid:20)A(cid:8)(cid:8)(cid:2)(cid:9)(cid:19)(cid:21)(cid:8)(cid:3)E(cid:5)(cid:7)(cid:8)(cid:9)(cid:5)(cid:8)(cid:3)(cid:2)(cid:3)E(cid:8)(cid:9)(cid:5)(cid:6)(cid:22)(cid:23)(cid:24)(cid:24)(cid:3)(cid:2)(cid:16)(cid:25)(cid:7)(cid:5)E(cid:8)(cid:2)(cid:21)(cid:26)(cid:8)(cid:9)(cid:7)(cid:5)(cid:23)(cid:24)(cid:24)(cid:3)(cid:2)(cid:16)(cid:27)(cid:26)(cid:26)(cid:3)(cid:28)(cid:8)A(cid:5)(cid:26)(cid:3)C(cid:3)D(cid:9)A(cid:8)(cid:7)(cid:2)(cid:16)AEFE(cid:16)(cid:8)(cid:7)A(cid:26)(cid:8)(cid:9)(cid:29)(cid:3)(cid:16)(cid:8)(cid:3)A(cid:18)(cid:18)(cid:3)(cid:18)(cid:19)(cid:3)(cid:2)E(cid:16)(cid:9)(cid:5)(cid:8)(cid:3)(cid:2)(cid:3)E(cid:8)(cid:3)D(cid:9)(cid:5)(cid:16)(cid:30)(cid:20)(cid:16)(cid:1)A(cid:2)(cid:8)(cid:9)A(cid:31)(cid:16)(cid:23)(cid:24)(cid:24)(cid:3)(cid:2)(cid:16)(cid:9)E !"(cid:16)#A(cid:31)(cid:21)(cid:3)(cid:16)(cid:5)(cid:3)(cid:3)D(cid:3)D(cid:16)(cid:24)(cid:7)(cid:2)A(cid:8)(cid:8)(cid:2)(cid:9)(cid:19)(cid:21)(cid:8)(cid:3)(cid:16)(cid:30)(cid:22)C(cid:3)D(cid:9)A(cid:8)(cid:7)(cid:2)A(cid:6)(cid:6)(cid:2)(cid:3)(cid:6)A(cid:8)(cid:3)E(cid:16)(cid:29)A(cid:31)(cid:21)(cid:3)E"C(cid:3)D(cid:9)A(cid:8)(cid:7)(cid:2)(cid:16)AEFE$%(cid:3)(cid:8)%(cid:3)(cid:2)(cid:16)(cid:5)(cid:3)$(cid:28)A(cid:2)(cid:8)(cid:9)A(cid:31)(cid:16)(cid:7)(cid:24)(cid:24)(cid:3)(cid:2)(cid:16) !(cid:16)(cid:9)EA(cid:26)(cid:26)(cid:3)(cid:28)(cid:8)A(cid:19)(cid:31)(cid:3)(cid:16)(cid:8)(cid:7)(cid:3)(cid:29)(cid:3)(cid:2)&(cid:16)A(cid:26)(cid:8)(cid:9)(cid:29)(cid:3)(cid:16)(cid:8)(cid:3)A(cid:18)(cid:18)(cid:3)(cid:18)(cid:19)(cid:3)(cid:2)'(cid:3)(cid:5)D(cid:23)(cid:24)(cid:24)(cid:3)(cid:2)C(cid:3)D(cid:9)A(cid:8)(cid:7)(cid:2)(cid:16)AEFE$%(cid:3)(cid:8)%(cid:3)(cid:2)(cid:7)(cid:28)(cid:28)(cid:7)(cid:5)(cid:3)(cid:5)(cid:8)!E(cid:7)(cid:24)(cid:24)(cid:3)(cid:2)(cid:16)(cid:9)EA(cid:26)(cid:26)(cid:3)(cid:28)(cid:8)A(cid:19)(cid:31)(cid:3)(cid:1)(cid:2)(cid:3)(cid:3)(cid:4)(cid:5)(cid:5)(cid:6)(cid:7)(cid:8)(cid:9)(cid:2)A(cid:4)(cid:17)(cid:3)A(cid:18)(cid:16)(cid:2)(cid:3)(cid:30)(cid:3)(cid:26)(cid:8)E(cid:17)(cid:3)A(cid:18)(cid:16)A(cid:26)(cid:26)(cid:3)(cid:28)(cid:8)E(cid:23)(cid:28)(cid:28)(cid:7)(cid:5)(cid:3)(cid:5)(cid:8)(cid:16)(cid:26)(cid:7)(cid:21)(cid:5)(cid:8)(cid:3)(cid:2)(cid:7)(cid:24)(cid:24)(cid:3)(cid:2)E(cid:23)(cid:28)(cid:28)(cid:7)(cid:5)(cid:3)(cid:5)(cid:8)(cid:16)A(cid:26)(cid:26)(cid:3)(cid:28)(cid:8)E(cid:23)(cid:28)(cid:28)(cid:7)(cid:5)(cid:3)(cid:5)(cid:8)(cid:16)((cid:9)(cid:8)%D(cid:2)A$ !(cid:16)(cid:9)E(cid:16)(cid:26)(cid:7)(cid:18)(cid:28)(cid:31)(cid:3)(cid:8)(cid:3) !(cid:16)(cid:9)E(cid:16)(cid:5)(cid:7)(cid:8)(cid:16)A(cid:16)(cid:26)(cid:7)(cid:18)(cid:28)(cid:31)(cid:3)(cid:8)(cid:3)(cid:16)(cid:7)(cid:24)(cid:24)(cid:3)(cid:2)(cid:17)(cid:3)A(cid:18)(cid:16)D(cid:3)AD(cid:31)(cid:9)(cid:5)(cid:3)(cid:16)(cid:2)(cid:3)A(cid:26)%(cid:3)D"(cid:16)(cid:17)(cid:3)A(cid:18)(cid:16)((cid:9)(cid:8)%D(cid:2)A$E {Start pre-negotiation phase}
∀i, j, I(i, j) = true
Ask for N Iai to each ai
Receive responses N Iai from each ai
for ai ∈ A do
for j ∈ N Iai do
I(i, j) = f alse
end for
end for
{End pre-negotiation phase}
Fig. 2. Pre-negotiation protocol followed by team A and the trusted mediator.
In this schema, we show the protocol from the point of view of the mediator.
The protocol followed by team members is analogous and straight-forward.
the team proposes an offer. It is reasonable that each team
member may be willing to sacrifice decision rights pertaining
to negotiation attributes that have little importance or no
importance at all for one's own interests. This decision may
help to find a more satisfactory agreement for the opponent
while maintaining good quality for one's own utility. The fact
that some attributes may yield little or no importance at all for
some team members is also feasible in a team setting, since
some of these attributes may have been introduced to satisfy
the interests of a subgroup of team members.
Therefore, for each team member ai, the trusted mediator
asks the set of attributes N Iai whose decision rights conform
the set of rights that ai is willing to hand over when building
an offer for the opponent. This information is annotated by the
mediator in an interest matrix I. Each matrix position I(i, j)
indicates whether the team member ai holds decision rights
for attribute j (I(i, j) = true) or not (I(i, j) = f alse). How
each team member ai specifies this set of attributes N Iai is
described later in Subsection II-E. The communication pro-
tocol carried out during the pre-negotiation phase, described
from the point of view of the mediator, is shown in Fig. 2.
2) Negotiation: Three possible actions can be carried out
by the negotiation team at each negotiation round t: (a)
op→A) (Offer acceptance
accept/reject opponent offer acA(X t
in Fig. 1); (b) send an offer/counter-offer X t
A→op (Offer
construction in Fig. 1); (c) abandon the negotiation process.
This last action is performed when the team deadline TA
is reached. Thus, when t > TA, the mediator informs the
opponent about
the team's withdrawal. The mediator also
has a very important role in the coordination mechanisms
employed by the team to decide upon action (a) and (b). The
coordination processes (a) and (b) are described in a detailed
way. Furthermore, a finite state machine formalization of the
negotiation from the point of view of the mediator can be
observed in Fig. 3. It shows the offer construction and the
opponent offer acceptance processes.
a) Offer acceptance: When team A has to decide
op→A)),
whether or not to accept the opponent's offer (acA(X t
first, the mediator receives the opponent's offer X t
op→A. This
offer is publicly announced to all of the team members by the
mediator. Then, the mediator opens a private voting process
where each team member ai should specify whether or not
op→A).
it supports acceptance of the opponent's offer acai(X t
4
Fig. 3. Finite state machine formalization of the negotiation from the point
of view of the mediator. The transitions follow the format Events / Actions.
Messages sent by the mediator follow the format <Recipient,Message>.
op→A in A (start voting process)
op→A) from each team member ai
op→A
Receive X t
Broadcast X t
Receive response acai (X t
in A
votes = 0
for ai ∈ A do
if acai (X t
votes = votes + 1
end if
end for
if votes = A then
op→A) = true then
acA(X t
op→A) = true
else
acA(X t
op→A) = f alse
end if
Fig. 4. Negotiation protocol followed by team A and the trusted mediator
to decide the acceptance of the opponent offer. In this schema, we show the
protocol from the point of view of the mediator. The protocol followed by
team members is analogous and straight-forward.
How each team member decides whether or not the opponent
offer is supported will be described in Subsection II-E. Once
every vote has been received, the mediator counts the number
of positive votes (votes that support the opponent's offer). The
offer is accepted if the number of positive votes is equal to the
number of team members. Otherwise, the offer is rejected. A
complete view of this communication protocol can be observed
in Fig. 4.
b) Offer construction: The mediator coordinates an it-
erated offer-building process in order to ensure unanimity in
A→op sent to the opponent. For that purpose, each
the offer X t
(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:9)(cid:7)A(cid:8)(cid:2)(cid:3)BC(cid:7)D(cid:8)(cid:2)(cid:1)(cid:2)EFD(cid:16)(cid:17)F(cid:18)(cid:19)(cid:20)(cid:17)D(cid:18)(cid:20)(cid:21)(cid:18)(cid:22)(cid:7)FD(cid:16)(cid:17)(cid:7)(cid:23)(cid:24)(cid:25)(cid:26)(cid:18)(cid:17)F(cid:18)(cid:19)(cid:20)(cid:17)D(cid:18)(cid:20)(cid:21)(cid:18)(cid:22)(cid:2)(cid:27)(cid:16)(cid:17)(cid:2)(cid:28)(cid:28)(cid:29)(cid:30)(cid:18)(cid:30)(cid:24) attribute value is adjusted one by one. Before the iterated
process starts, the mediator considers every team member ai as
an active member in the offer construction process. Once the
iterated offer construction process starts, the trusted mediator
selects an attribute j from the set of attributes that have not
(cid:48)t
yet been set. Given the partial offer X
A→op built until the
moment, the mediator asks each active team member ai who
is interested in j (I(i, j) = true) about the value xai,j needed
to get as close as possible to its current aspiration level sai(t).
When private responses have been gathered from every team
member, the mediator decides a value xj for the attribute
j. Here, the morphology of the proposed utility functions
comes into play. Due to the fact that team members share
the same type of monotonicity for valuation functions, the
trusted mediator can aggregate agents' opinions by means
of the max function (monotonically increasing) or the min
function (monotonically decreasing). As will be proved in
Section III, the value decided for xj ensures unanimity among
team members under certain assumptions. The value xj is set
(cid:48)t
A→op which is publicly announced to
in a new partial offer X
team members. Then, the mediator asks every active agent in
the offer construction process whether or not the new partial
offer is satisfactory at round t, ac(cid:48)
(cid:48)t
A→op). Those agents
(X
ai
(cid:48)t
A→op are eliminated from
that agree with the current state of X
the active list. Those attributes which are not interesting for
any team member are maximized according to the opponent's
preferences. The process steps back to the selection of a new
attribute j until all of them have been set. A more detailed
description of this process can be observed in Fig. 5. As will
be reviewed in Section III, when agents comply with certain
assumptions, the proposed iterated process is able to reach an
offer that is supported by all of the team members at each
negotiation round (unanimity). Obviously, the resultant offer
depends on the agenda of attributes employed by the mediator.
Since unanimity is guaranteed independently of the agenda
employed by the mediator, the offer constructed should be
as satisfactory as possible for the opponent. Ideally, the team
should try to fulfill its own interests with those attributes that
are less important for the opponent.
E. Intra-Team Strategy: Team Members' Strategy within the
Team
This subsection specifies how team members answer the me-
diator's petitions. On the one hand, during the pre-negotiation,
team members decide upon which attribute decision rights are
handed over. On the other hand, during the negotiation, agents
have to decide whether they accept the opponent's offer and
which values should be set for the offer to be sent to the
opponent. The behavior of team members in these decision
making processes is described below.
1) Pre-negotiation: In the pre-negotiation phase, the me-
diator asks each team member the set of attributes' decision
rights that it is willing to hand over. Its size may range from
0 attributes to the whole set of attributes. How many decision
rights ai is willing to hand over depends on an individual and
private value ai ∈ [0, 1]. When ai = 0, the agent is only
willing to hand over rights that yield no interest at all (i.e.,
5
A→op = ∅
(cid:48)t
X
A(cid:48) = A
{For each attribute}
for j ∈ N do
V = ∅
{Check opinion of team members who are active in the
building phase and are interested in the attribute}
for ai ∈ A(cid:48) ∧ I(i, j) = true do
Ask for xai,j
Receive xai,j
V = V (cid:83) xai,j
end for
{Aggregate agents' opinions}
if V = 0 then
xj = best value for opponent(j)
else
(cid:83) xj
xj = max(V ) or min(V )
(cid:48)t
A→op
(cid:48)t
A→op among team members
end if
(cid:48)t
X
A→op = X
Make public new X
{Update list of agents who are active in the building
phase}
for ai ∈ A(cid:48) do
Ask for ac(cid:48)
Receive ac(cid:48)
ai
if ac(cid:48)
(cid:48)t
A→op) = true then
(X
(cid:48)t
A→op)
(cid:48)t
A→op)
ai
(X
(X
A(cid:48) = A(cid:48) − ai
ai
end if
end for
end for
X t
A→op = X
(cid:48)t
A→op
Fig. 5. Negotiation protocol followed by team A and the trusted mediator
to build an offer to be sent to the opponent. In this schema, we show the
protocol from the point of view of the mediator. The protocol followed by
team members is analogous and straight-forward.
attributes j whose wai,j = 0), whereas when ai = 1, the
agent is willing to hand over all of the attributes' decision
rights. The set of attributes N Iai whose decision rights are
handed over by agent ai follow this equation:
wai,j ≤ ai
(4)
(cid:88)
j∈N Iai
Thus, ai acts as an upper limit that determines the total
importance given by ai to the attributes whose decision rights
are handed over. Given a value for ai, there are multiple sets
N Iai that fulfill Eq. 4. A reasonable heuristic is to assume
that the agent is willing to concede as many decision rights
as possible since this will enhance the possibility of finding
an agreement with the opponent. Hence, each team member
ai chooses the largest possible set N Iai that fulfills Eq. 4.
2) Negotiation: In the negotiation process, two different
decisions are taken by the team: whether or not they accept
the opponent's offer, and which offer is sent to the opponent.
First, the decision making mechanism acai(.), used by each
agent ai to decide whether or not it supports the opponent's
offer, is described. It seems appropriate to assume that the
agent will accept the opponent's offer if it reports a utility
that is greater than or equal to the aspiration level marked by
the concession strategy in the next round:
(cid:26) true
f alse
acai (X) =
if sai (t + 1) ≤ Uai(X)
otherwise
(5)
where true means that the agent supports the opponent's offer,
f alse has the opposite meaning, and sai(.) is the concession
strategy employed by agent ai
to calculate the aspiration
level at each negotiation round t. Regarding the concession
strategy employed by team members, it is considered that
team members have agreed upon a time-based concession
strategy with a common βA. Thus, the concession strategy
sai(.) followed by each team member ai can be formalized
as depicted below. It is a modification of the well-known
concession strategy used in Equation 2.
sai (t) = (1 − ai) − (1 − ai − RUai)(
1
βA
)
t
TA
(6)
For the expression above, it can be observed that each agent's
aspiration level, despite being governed by the same βA,
depends on the private reservation utility of each agent RUai.
ai acts as a limit for the maximum utility demanded by the
concession strategy. Since the agent has handed over decision
rights for a set of attributes whose weights sum up to ai, the
maximum utility that the agent is able to demand by itself is
(1 − ai). This is reflected in the equation above.
Second, in the case of the iterated construction process, team
members take two decisions: which value xai,j is requested
(cid:48)t
A→op, and whether or
for attribute j given the partial offer X
not the new partial offer is acceptable ac(cid:48)
(cid:48)t
A→op). When
requesting a value for j, each team member communicates
anonymously the value xai,j which gets as close as possible to
its desired aspiration level sai(t). This value can be calculated
by obtaining the attribute value xai,j whose weighted utility
(wai,jVai,j(xai,j)) is the closest to the utility needed by the
partial offer in order to reach the desired utility level (sai(t)−
Uai(X
(X
ai
(sai(t) − Uai(X
A→op) − wai,jVai,j(x)) (7)
(cid:48)t
(cid:48)t
A→op)) :
xai,j = argmin
x∈[0,1]
where xai,j is set so that the new offer's utility does not exceed
the aspiration level marked by the concession strategy:
A→op) − wai,jVai,j(xai,j) ≥ 0
(cid:48)t
sai(t) − Uai(X
(8)
(cid:48)t
Once the partial offer X
A→op has been updated by the
mediator, those agents that are still active in the construction
process are asked whether or not the new offer is acceptable
for the current negotiation round. Again, we consider that a
partial offer is acceptable for an agent ai if it reports a utility
that is greater than or equal to the aspiration level marked by
its concession strategy:
(cid:26) true
f alse
ac(cid:48)
ai
(X) =
if Uai (X) ≥ sai(t)
otherwise
(9)
where true indicates that the partial offer is acceptable at its
current state for agent ai, and f alse indicates the opposite. A
simplistic trace of a negotiation round, and how team members
would behave, can be found in Fig. 6.
6
III. THEORETICAL ANALYSIS
In this section, we analyze some of the important character-
istics of our negotiation model in depth. The two main aspects
that we analyze are how unanimous decisions are guaranteed
regarding team decisions and how robust the proposed model
is against manipulations. In the first case, the model assures
that each team member gets a utility that is greater than or
equal to its current aspiration level. In the second case, we
analyze how the proposed model is robust against agents from
the opponent, but it is easily attacked by agents from the
competition that try to sabotage a deal with the opponent.
A. Unanimity within the Team
As mentioned in this article, the proposed negotiation model
allows team members to reach unanimity in team's decisions.
These decisions include the offer that is sent to the opponent
and the acceptance/rejection of the opponent's offers. In the
latter, it is clear that the proposed acceptance mechanism en-
sures unanimity since an opponent offer is only accepted when
it is considered acceptable by all of the team members. In the
former, the definition of unanimity is not straightforward.
We define that an offer sent to the opponent X t
A→op is a
strict unanimous decision for the team when, for any team
member ai, the offer reports a utility that is greater than or
equal to its current aspiration level sai(t):
∀ai∈AUai(X t
A→op) ≥ sai(t)
(10)
Achieving this definition of unanimity within the team en-
sures that if a final agreement is found, it reports a utility that is
greater than or equal to each agent's private reservation utility.
In order to achieve the proposed definition of unanimous
decision, some assumptions have to be made regarding the
behavior of team members. These assumptions have already
been presented in this article. Basically, team members have
to be truthful in their responses to the mediator, following the
behavior specified in Eq. 7,8 and 9. Next, we prove that, if
team members follow these behaviors, unanimity is achieved
in team's decisions according to Equation 10.
Proof: ∀ai∈AUai(X t
A→op) ≥ sai(t)
subject to: Eq. 7, Eq. 8, Eq. 9, and the same type of monotonic-
ity for valuation functions Vai,j in team members' utility func-
tions. For the sake of simplicity, we assume that team mem-
bers' valuation functions are monotonically increasing for any
negotiation attribute. It should be pointed out that, in that case,
the aggregation operation carried out by the trusted mediator is
the max operator. In any case, for any attribute j, its value will
be determined as xj = max(xa1,j, xa2,j, ..., xaM ,j) and then
it holds true that ∀ai ∈ A, wai,jVai,j(xj) ≥ wai,jVai,j(xai,j).
The proof is quite straightforward. When the mediator declares
that an attribute j must be set, three different situations may
arise for an agent ai:
• ai has already reached its aspiration level with the partial
A→op) ≥ sai(t). Therefore, the value deter-
(cid:48)t
offer Uai(X
mined for xj will add utility to the partial offer and the
utility reported to ai will further exceed its aspirations
Uai(X
A→op) + wai,jVai,j(xj) ≥ sai(t).
(cid:48)t
7
The figure shows a simplistic trace of the proposed model during one negotiation round t. The team is composed of three buyers (a1, a2, a3),
Fig. 6.
whereas the opponent is a seller. The upper left table shows the utility functions of team members and their aspiration levels at round t and t + 1. The upper
right table shows team members' valuation functions for negotiation attributes and the agenda learnt by the mediator. The negotiation attribute are p (price,
monotonically decreasing), q (quality, monotonically increasing), and d (payment date time, monotonically increasing). The lower table is the trace of the
negotiation round. In (1), the first attribute of the agenda xd is set, but it should be highlighted that a3 does not participate in its construction since a3 = 0
and, thus, a3 has handed the decision rights over xd during the pre-negotiation. After (4), a2 does not participate in the offer construction since it has reached
its desired aspiration with the value set for attributes xd and xq. Finally, in (7) the opponent's counteroffer is rejected since the offer is not acceptable for a3.
• ai can reach its current aspiration level sai(t) if it asks
(cid:48)t
for a value xai,j. Thus, Uai(X
A→op)+wai,jVi,j(xai,j) =
sai(t). Since the aggregation operation is xj =
max(xa1,j, xa2,j, ..., xaM ,j), the new partial offer will
have a utility that is equal to or greater than its aspirations,
Uai (X
A→op) + wai,jVai,j(xj) ≥ sai(t).
(cid:48)t
• ai cannot reach its aspirations by just setting xj. In
this case, ai will demand the maximum possible value
for j and then xj = xai,j. ai will have to reach its
aspiration level by adjusting the next attributes in the
agenda. In the worst case scenario, the next attribute to
be set xN is the last one in the agenda. This means that
ai has demanded the maximum value for the previous
attributes and succeeded in getting its desired value for
them. Thus, before the last attribute is set, the utility
N(cid:80)
wai,j. Since
reported by the partial offer to ai is
wai,j = 1 and 0 ≤ sai(t) ≤ 1, the agent will reach
j
its aspiration level by demanding a value for xN that
N−1(cid:80)
j
fulfills Vai,N (xN ) ≥
, which is ensured
thanks to the morphology of the valuation functions
(0 ≤ Vai,j(x) ≤ 1).
wai,N
j
sai(t)−N−1(cid:80)
wai,j
One might wonder whether or not it is reasonable to think
that agents are truthful in this process. However, members
are not tempted to demand lesser value for attributes since
the process would not ensure that the final agreement would
achieve its current aspiration level. On the other hand, it is
true that agents may be inclined to demand a greater value for
attributes since the process ensures that the offer will be more
profitable for them. Nevertheless, it should be pointed out that,
generally, if more value is demanded for attributes the offer
may be less profitable for the opponent and the probabilities
of reaching an agreement may be greatly reduced. This issue
is studied in Subsection IV-D, where we analyze whether or
not team members have strong incentives to deviate from the
proposed behavior.
B. Manipulation within the Team
1) Opponent: Here, we refer to agents that infiltrate the
team in order to increase the quality of the final agreement
from the point of view of the opponent party. In a negotiation
team setting formed by buyers, we are concerned about the
fact that some seller parties may attempt to introduce agents
among team members. This way, opponents may be able to
maximize their own preferences by manipulating the decisions
taken by the team. However, our proposed negotiation model
is robust to this kind of manipulation.
Let us imagine a situation where a negotiation team wants
to buy a product and a seller has been able to infiltrate agents
in the team. Due to the mechanism that is employed to build
the offer sent to the opponent and the mechanism employed to
decide upon whether or not to accept the opponent's offer, it is
not possible for opponent agents to manipulate the decisions
taken within the team. Regarding the iterated offer construction
process, an opponent agent would try to demand values that
are close to the preferences of the opponent. In a generic
electronic commerce application, an opponent agent might
demand high values for the price and the dispatch date and low
values for the product quality. However, the aggregation rules
employed by the trusted mediator (max or min depending on
the type of monotonic function that represents the preferences
of the team members) will ensure that team preferences prevail
independently of the number of infiltrated opponent agents.
As for the unanimous voting process, opponent agents might
try to engage team members in accepting the opponent's offer.
However, this is not possible due to the fact that as long as one
team member does not support the opponent's offer, it will not
be accepted. Thus, it does not require further demonstration.
For instance, let us imagine a negotiation team, formed by
5 buyers, that negotiates with a seller. Two agents (a1, a2) of
such team are real buyers, whereas the other three (a3, a4, a5)
are agents infiltrated by the seller. The negotiation problem is
based on two negotiation attributes, price and quality, whose
domains have been scaled to [0, 1]. The valuation function
used for the price in the case of the buyers is assumed to be
monotonically decreasing (buyers prefer low prices to high
prices), and the type of monotonic function used for the
quality is assumed to be monotonically increasing (buyers
prefer high quality to low quality). Thus, the mediator uses
the min function to aggregate the price attribute, and the max
function to aggregate the quality attribute. Assuming that the
opponent's valuation functions are of the opposite monotonic
type to those of the buyers, team members first demand the
following values for the price: xa1,price = 0.1, xa2,price = 0.2,
xa3,price = 0.9, xa4,price = 0.85, xa5,price = 1. The mediator
aggregates such values and the final value for the attribute
price is xprice = min(0.1, 0.2, 0.9, 0.85, 1) = 0.1, which is
actually preferred by the two real buyers. Thus, even if the
number of opponent agents is larger than the number of real
buyers, infiltrated agents from the opponent are not able to
manipulate the team. The example for the quality attribute
is analogous and does not require further explanation. In the
end, the preferences of real buyers will prevail over opponent
agents' demands and the team is not manipulated.
2) Competitors: Another kind of possible manipulation
is the one carried out by competitor agents. Competitors
are buyer agents (in the case that the team is made up of
buyer agents) that are interested in the same product as the
team. Some competitors may be interested in sabotaging team
deals if that assures that competitors get better deals from
the opponent. This is especially true in environments where
goods or services are limited (e.g., personal sellers on Ebay).
Thus, competitor agents may attempt to prevent the team from
reaching an agreement with the opponent.
This is the case even in situations where the group of opponent
agents is larger than the number of real team members.
Even though the proposed model is robust against opponent
agents, robustness is not maintained when dealing with infil-
trated competitor agents. In that case, the strengths shown by
the model become its weaknesses. In the voting process carried
out to decide upon whether or not to accept the opponent's
offer, only a single agent is needed to manipulate the process
and prevent the team from accepting the opponent's offers. On
the other hand, competitor agents may manipulate the offer
construction phase by being highly demanding. In a generic
electronic commerce application, the competitor agent would
demand very low values for the price, short dispatch dates and
very high product quality. This way, competitor agents make
offers extremely undesirable for opponent agents, preventing
the team from reaching a final agreement with the opponent.
Due to the aggregation operators employed by the trusted
mediator, only one competitor agent is needed to manipulate
the offer construction process. Thus, this model should be
employed only when team members are extremely sure that no
competitor agent has infiltrated the team. It would be possible
to employ sophisticated mechanisms to detect these agents;
8
however this is a topic for future research.
IV. EMPIRICAL EVALUATION
In this section, we explore the impact of the different
parameters of our proposed model. More specifically, we
study the importance of the agenda of issues imposed by
the mediator on the negotiation process, the impact of the
number of decision rights that are handed over during the
pre-negotiation, the empirical robustness of the model against
attacks (i.e., agents from the competition that try to sabotage
the negotiation), and whether or not
team members have
incentives to deviate from the proposed strategy.
A. Studying the Impact of Intra-Team Agenda
The mediator uses an agenda to determine which attributes
are set first in the iterated building process. A reasonable
heuristic is to try to satisfy team members with those attributes
that are less important for the opponent. Otherwise, the resul-
tant offer may be too demanding and the negotiation process
may end in failure. Thus, ideally, the agenda should order the
attributes in ascendant order of importance for the opponent.
However, it is acknowledged that the situations where the
opponent may reveal its full ranking of preferences are very
limited or almost non-existent. Thus, it is necessary to provide
the mediator with mechanisms that approximately learn the
opponent's preferences. In this article, we propose a simple
learning mechanism that is based on the idea that one agent
may concede less in its important attributes during the first
negotiation rounds. The mechanism takes into account the
offers received in the first k negotiation rounds and sums up
the accumulated amount of concession for each negotiation
attribute. Then, the mediator orders the attributes in descendant
order according to the amount of concession and it becomes
the agenda of attributes for the negotiation. If the number of
current rounds is lower than k, the agenda is built based on
the available information. Thus, in the first few rounds, the
learning mechanism is not expected to accurately match the
opponent's preferences; however, as the negotiation process
advances, more information is available and the learning
mechanism should match the opponent's preferences better.
In the first experiment, we decided to study the importance
of the agenda on the negotiation process. While every team
member gets a utility that is greater than or equal to its desired
aspiration level, the offer may be more or less demanding for
the opponent. If the offer is less demanding for the opponent,
it is more probable that it will be accepted by him. Therefore,
we decided to study the utility reported by the teams' offer
to the opponent at each negotiation round. We simulated a
negotiation process where offers are not accepted (i.e.,
it
always reaches the negotiation deadline) just to observe the
utility of the offers proposed by the team from the opponent's
perspective. Two different environments were tested: one with
a short deadline Top = TA = 10, and one with a long deadline
Top = TA = 50. Other parameters were set to the standard
values of our negotiation model: βop = βA = 1, ai = 0,
and RUop = RUai = 0. Three different types of agendas
for the FUM model were compared: a perfect agenda where
the mediator knows perfectly the order of importance given
by the opponent (FUM-perfect); the simple learning method
described above (FUM-simple); and a random agenda that
is built at each negotiation round (FUM-random). For FUM-
simple, the number of initial negotiation rounds to be taken
4 (cid:99). Additionally, the proposed
into account was set to k = (cid:98) TA
negotiation model is compared with a representative model
(RE) and a similarity simple voting model (SSV) which were
proposed in our previous work [4]. On the one hand, the
representative model does not assure any kind of consensus
among team members. One of the team members is chosen
as representative and decides on behalf of the team according
to its own private utility function. On the other hand, SSV
uses majority/plurality to take team decisions. Each round,
each team member is allowed to propose an offer to be sent
to the opponent. This offer is proposed based on a similarity
heuristic that considers the last opponent's offer and the last
offer proposed by team members in the previous round. These
two models are expected to be less demanding in terms of
utility due to the fact that less conflict is introduced with
the opponent (i.e., a fewer number of team members may
reach their aspiration level). A total of 100 random teams with
size M = 4 and random utility functions (4 attributes) were
confronted with 11 randomly generated opponents. In order
to capture stochastic variations in the different models, each
possible negotiation was repeated 4 times. Thus, a total of
4400 negotiations were carried out per model and environment
(i.e., short/long deadline). The results for this first experiment
can be observed in Fig. 7.
As can be observed in the short deadline scenario (Fig. 7),
the offers proposed by the representative model are more at-
tractive for the opponent. This is reasonable since, in this case,
the representative only negotiates attending to its own utility
function. Therefore, it results in less conflict with the opponent
and more trade-off possibilities. The behavior observed for the
perfect agenda model and the similarity simple voting model
are more surprising. Even though, in the first rounds, SSV
proposes offers that report more utility for the opponent than
those built by the perfect agenda model, as the negotiation
advances, the perfect agenda model outperforms SSV. This
happens at negotiation round 6. This may be explained by
the fact that, at that point, more trade-off possibilities arise
between all of the team members and the opponent, and the
perfect agenda model is capable of exploiting them while
assuring the desired aspiration level for each teammate. As
for the simple agenda model, it performs slightly better than
the random agenda model, but worse than the other methods in
the experiment. This is explainable by the fact that, since the
negotiation deadline is short, limited information can be used
to learn the opponent's preferences. Consequently, the agenda
built is closer to a random agenda than to the perfect agenda.
In the case of the long deadline scenario, a similar tendency
can be observed. Nevertheless,
there are some differences
that are worth highlighting. First, the representative model
is still the one that is the most attractive for the opponent's
interests. However, in this scenario, both the perfect agenda
model and the simple agenda model are able to outperform
SSV at some points of the negotiation process. Obviously, this
9
Fig. 7. The upper graphic shows the average utility reported to the opponent
by the team's proposal at each negotiation round for the short deadline
scenario. The lower graphic shows the results for the long deadline scenario.
happens earlier for the perfect agenda model since it represents
perfect knowledge about the opponent's preferences. Hence, it
is able to take advantage of possible trade-offs earlier in the
negotiation. It happens approximately at round 22. Regarding
the simple agenda model, it is able to outperform SSV around
round 33. Differently to the first scenario, since the amount of
information to learn from is greater, the simple agenda model
is able to get closer to the perfect agenda and offer more
attractive offers to the opponent.
In conclusion, methods proposed in the literature like RE
and SSV (in the short deadline scenario) are less demanding
for the opponent; however it should be pointed out that they
do not ensure unanimity as FUM. Thus, the preferences of
all the team members are not represented in the deals found
by RE and SSV. In fact, we ran additional tests to ascertain
this conclusion. The experimental conditions were set to the
same parameters found in this first batch of experiments, but
in this experiment the two parties were able to accept offers,
thus ending the negotiation before the deadline. We tested
the performance of RE, SSV, and FUM-simple according to
different quality measures such as the minimum utility of the
team members and the average utility of the team members.
The results of this experiment can be observed in Table I.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 2 3 4 5 6 7 8 9 10UtilityNegotiation RoundUtility reported to the opponent from the team's offer at each roundRepresentativeSimilarity Simple VotingFull Unanimity Mediated (Random Agenda)Full Unanimity Mediated (Simple Agenda)Full Unanimity Mediated (Perfect Agenda) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 10 20 30 40 50UtilityNegotiation RoundUtility reported to the opponent from the team's offer at each roundRepresentativeSimilarity Simple VotingFull Unanimity Mediated (Random Agenda)Full Unanimity Mediated (Simple Agenda)Full Unanimity Mediated (Perfect Agenda) Method
RE
SSV
FUM-simple
Min.
[0.10-0.11]
[0.33-0.34]
[0.41-0.42]
Long deadline
Min.
[0.11-0.12]
[0.30-0.31]
[0.50-0.51]
Ave.
[0.45-0.46]
[0.56-0.57]
[0.74-0.75]
Short deadline
Ave.
[0.42-0.43]
[0.52-0.53]
[0.68-0.69]
TABLE I
THIS TABLE SHOWS THE QUALITY OF THE FINAL AGREEMENT. VALUES
SHOWN ARE CONFIDENCE INTERVALS (95%) FOR THE MEAN. MIN:
MINIMUM UTILITY OF TEAM MEMBERS, AVE: AVERAGE UTILITY OF
TEAM MEMBERS
As shown, FUM-simple is able to obtain better agreements in
terms of utility (both measures) for the team members. Thus,
the results suggest the aforementioned claim: even though RE
and SSV are less demanding for the opponent, they do not
represent the preferences of the team members as FUM does.
B. Studying the Impact of ai
In this second experiment, we decided to study the impact
of ai on the team's performance. It seems reasonable to think
that low values of this parameter should help to construct
offers that are more interesting for the opponent, but high
values should impact negatively on the utility obtained by
ai. We devised an experiment where the value of ai was
set in a uniform way for all of the team members. More
specifically, we used the values 0, 0.02, 0.05, 0.07, 0.1, 0.12,
0.15, 0.17, 0.2 for ai. For the quality measures, we observed
the minimum and the average utility of the team members.
Two different environments were tested: short/long deadline,
whose lengths are drawn from the uniform distributions Top =
TA = U [5, 10], Top = TA = U [30, 60], respectively. The
concession speed for both parties was set to be drawn from
βop = βA = U [0.4, 0.99] since initial experiments have
suggested that boulware strategies may provide more utility
for both parties in absence of other outside options [4]. The
reservation utility for the agents was drawn from a uniform
distribution RUop = RUai = U [0, 0.25]. In this case, the
learning method for the agenda was set to FUM-simple and
the number of initial rounds to be taken into account was
set to k = (cid:98) TA
4 (cid:99). A total of 100 randomly generated teams
with size M = 4 and random utility functions (4 attributes)
were confronted with 12 randomly generated opponents. Each
possible negotiation was repeated 4 times. Thus, a total of
4800 negotiation were carried out per model and environment.
The results for this experiment are shown in Table II.
The results show a slight decrease in the utility (minimum
utility and average utility) as ai gets larger. This behavior
is found in almost every scenario tested. Those configurations
that do not show this pattern usually obtain very similar results
for all of the configurations. Thus, the agents should choose
ai = 0 independently of the type of scenario where they
negotiate. In the best case, the agent will get a slightly better
utility than other values of the parameter. In the worst case
scenario, the agent will get a very similar utility to other values
of the parameter ai. The value ai = 0 corresponds to the
agents only handing over those decision rights associated to
attributes that yield no interest at all for the agent.
10
Long deadline
Ave.
Min.
0.72
0.49
0.50
0.71
0.68
0.49
0.67
0.48
0.66
0.49
0.65
0.48
0.48
0.64
0.63
0.47
0.46
0.61
Short deadline
Ave.
Min.
0.60
0.35
0.37
0.61
0.58
0.37
0.57
0.37
0.56
0.37
0.56
0.37
0.37
0.55
0.54
0.38
0.38
0.55
ai
0.00
0.02
0.05
0.07
0.10
0.12
0.15
0.17
0.20
TABLE II
THIS TABLE SHOWS THE AVERAGE IMPACT OF ai ON TEAM
PERFORMANCE. MIN: MINIMUM UTILITY OF TEAM MEMBERS, AVE:
AVERAGE UTILITY OF TEAM MEMBERS
It can also be observed that the average utility is impacted
more negatively by increment of the ai parameter in the
long deadline scenario than in the short deadline scenario.
A thorough analysis of our results gave an answer to this
phenomenon. The results suggest that higher values of ai
reduce the average utility for the team members. However,
the number of negotiations that ended with no agreement in
the long deadline scenario when ai = 0 was 151 (3.1% of
the negotiation cases ended with an average utility equal to 0),
whereas the number of failed negotiations was 404 (8.41%)
when ai = 0 and the deadline was short. As ai was increased
to 0.2, the number of failed negotiations decreased to 35
(0.7%) in the long deadline scenario and 91 (1.8%) in the
short deadline scenario. Thus, higher values for ai contribute
to reaching an agreement in cases where no deal was found.
This effect is more notorious in the short deadline scenario.
Since the number of failed negotiations is greatly reduced in
the short deadline scenario, the negative effect of higher ai is
moderated since the new negotiations contribute with values
for the average utility that are greater than or equal to 0.
Despite this, the reduction in the number of failed negotiations
is not enough to counter the negative impact of ai.
In general, ai can be considered as some sort of moderator
for the initial demand. According to our results, in general,
agents should not give up any decision right over an attribute
that yields interest for him. Only those decision rights associ-
ated to attributes that yield no interest at all should be handed
over. Hence, team members should always start demanding
their highest aspiration level. This situation resembles results
obtained in bilateral negotiation [6], where it was found that
if the deadline is reasonably long,
the agent should start
demanding values close to their maximum utility.
C. Studying the Impact of Infiltrated Competitors
In Section III, we described the robustness of the proposed
negotiation model against agents that
try to sabotage the
negotiation team. It was shown that agents from the oppo-
nent may not be able to manipulate the negotiation process.
Nevertheless, when agents from the competition infiltrate the
team, they may be able to stop the team from reaching an
agreement. Thus, one of our concerns is how different levels
of risk may affect the performance of teams acting according
to our model. We decided to test the performance of FUM in
different adverse negotiation scenarios.
These scenarios differ in the probability P that at least
one of the team members comes from the competition. The
infiltrated agent
tries to stop the team from reaching an
agreement with the opponent. To do so, the agent always
rejects the offer received from the opponent and is extremely
demanding when asking for value in the iterated offer build-
ing process. Since an agent that always asks for the most
demanding value can be easily spotted (i.e., the agent does
not concede at all), we decided to model this competitor agent
as an agent that tries to mimic a team member with a high
reservation utility RUai. This way, the agent concedes during
the negotiation process, but its requests are always high since
its reservation utility is high. Hence, we decided that the
infiltrated agent would follow the same concession strategy
as the rest of the teammates, but the infiltrated agent would
have an unexpectedly high reservation utility drawn from the
uniform distribution RUai = U [0.8, 1.0]. According to this
strategy, an infiltrated agent would be more difficult to identify
than an agent that always asks for the most demanding value.
However, this does not assure that the team will be sabotaged.
This will also depend on other factors such as deadline lengths
and concession strategies carried out by both parties.
the FUM model
In this experiment, we propose to analyze the per-
formance of
in scenarios where P =
{0, 25, 50, 75, 100%}. Additionally, we include two modifi-
cations of FUM with two different unanimity levels: 50%
(FUM50), and 75% (FUM75). These levels of unanimity
are applied when accepting the opponent's offer. The infil-
trated agent will act according to the βA = U [0.4, 0.99]
imposed by the team, but he will act as an agent with a
high reservation utility RU = U [0.8, 1.0]. The reservation
utility for the rest of the agents was drawn from a uniform
distribution RUop = RUai = U [0, 0.25]. As in previous
experiments, the concession speed of the opponent was set to
βop = U [0.4, 0.99] and we simulated two different scenarios:
one with short deadlines (TA = Top = U [5, 10]) and one
with long deadlines (TA = Top = U [30, 60]). For ai, it was
set to ai = 0 since the previous experiment showed that it
may be more beneficial in terms of utility to team members.
The learning method for the agenda was set to FUM-simple
and the number of initial negotiation rounds to be taken into
4 (cid:99). A total of 100 randomly
account was set to k = (cid:98) TA
generated teams with size M = 4 and random utility functions
(4 attributes) were confronted with 12 randomly generated
opponents. Each negotiation was repeated 4 times. Thus, a
total of 4800 negotiations were carried out per model and
environment. The results are shown in Table III.
The results showed the expected tendency: as the probability
P increased,
the utilitarian values for the team members
decreased. This effect is observable due to the fact that the
environment was more distrustful and the agents were able to
successfully sabotage the team by acting as highly demanding
team members. The average utility of the team members was
reduced by 67%-73% in the highest risk scenario for FUM.
As a solution for this problem, other unanimity rules could
be useful. In fact, it can be observed that when P is high,
11
P
0.0%
25%
50%
75%
100%
Long
0.71
0.60
0.47
0.33
0.24
Short
0.61
0.50
0.38
0.28
0.17
FUM
FUM75
FUM50
Long
0.67
0.65
0.61
0.59
0.56
Short
0.59
0.56
0.53
0.48
0.46
Long
0.58
0.58
0.56
0.55
0.53
Short
0.53
0.51
0.50
0.50
0.48
TABLE III
THIS TABLE SHOWS HOW THE PROBABILITY P THAT AT LEAST ONE OF
THE TEAM MEMBERS IS A COMPETITOR IMPACTS ON THE TEAM'S
PERFORMANCE. VALUES SHOW THE MEAN FOR THE AVERAGE UTILITY OF
TEAM MEMBERS
FUM50 and FUM75 perform better than FUM and are not
so affected by the infiltrated agent (16%-33% performance
reduction for FUM75 in the highest risk scenario, and 9%-
10% performance reduction for FUM50). Thus, it is acknowl-
edged that without any additional mechanism (e.g., trust and
reputation models [8]) the proposed negotiation model is not
convenient for scenarios where it is very likely that competitor
agents may enter a team. In cases where there is a high risk of
encountering manipulators, models based on majority/plurality
voting paradigms such as SSV [4] or modifications of FUM
like FUM50 and FUM75 may prove to be more fit since a
large number of competitor agents may be needed to sabotage
the negotiation. However, the unanimity would not be assured
anymore, which we consider highly desirable for teams.
D. Strategy Deviation
The proposed model assumes that team members state the
truth when asked about which attribute values they need to
reach their desired utility level during the offer construction
phase. When dealing with selfish agents, one risk faced is
the fact that selfish agents may not tell the truth in order to
maximize their own utility. In this case, it seems clear that
team members have no incentives to ask for less attribute
value than they need since it may end up in an agreement
with a utility inferior to the desired level of utility. However,
team members may have incentives to demand more value
if that maximizes their utilities (be more demanding). For
a team member to play strategically, it would need to have
some knowledge about team members' and opponent's utility
functions, deadlines, reservation utilities, and other agents'
strategies. We aim to propose negotiation models for open
environments, where information is private. Therefore, agents
usually have limited and uncertain information regarding the
negotiation conditions. This leads to the question of whether or
not team members would achieve higher utilities by deviating
from the proposed strategy.
In this subsection we analyze whether or not team members
have incentives to deviate from the proposed strategy in the
offer construction phase. For this matter, we designed two
types of deviated team members. The first type of deviated
agent, which we will name slightly deviated, behaves exactly
as the standard behavior proposed for team members in this
article. However, during the iterated offer construction phase,
the agent does not ask for the value it needs from attribute j,
but a value that reports higher utility than it needs. The amount
(di × (sai(t)− Uai(X
of extra utility that it attempts to achieve is controlled by a
parameter di. When di > 1, the team member demands more
value than it needs, as it can be appreciated in the formula:
A→op))− wai,jVai,j(x))
(cid:48)t
xai,j = argmin
x∈[0,1]
(11)
When the utility of the partial offer exceeds or equals the
desired utility level sai(t),
the agent abandons the offer
construction phase at that round. The effect of this behavior
is that, when the agent is asked to set an attribute which
can report the desired utility, it demands more value for that
attribute and then leaves the iterated building process. For
instance, if a seller agent needs 0.50 for the price attribute
in order to reach its desired utility level and di = 1.25, it
will ask for 0.50 × 1.25 = 0.625 instead. The second type
of deviated team member, named highly deviated, behaves as
the slightly deviated team member but when it has reached
its desired utility level, it stays an additional attribute in the
iterated building process. When asked about the value of that
extra attribute, the highly deviated agent asks for a random
value that reports between 10% and 50% of the attribute's
utility. For instance, assuming that the price is scaled between
0 and 1, a highly deviated seller that has reached its desired
utility level would ask for a price value between 0.1 and 0.5.
After setting the extra negotiation attribute, the highly deviated
team member leaves the offer construction phase.
We set the parameters of our model to the same values
used in the previous experiment: TA = Top = U [30, 60]
for long deadline scenarios, TA = Top = U [5, 10] for
short deadline scenarios, RUai = RUop = U [0, 0.25],and
βA = βop = U [0.4, 0.99]. A total of 100 randomly generated
teams with size M = 4 and random utility functions (4
attributes) were confronted with 12 randomly generated oppo-
nents. Each possible negotiation was repeated 4 times. Thus,
a total of 4800 negotiations were carried out per model and
environment. We studied the effect of the number of slightly
deviated agents Asd = {1, 2, 3, 4} (the rest of team members
having the standard behavior), the effect of the number of
highly deviated agents Ahd = {1, 2, 3, 4} (the rest of team
members having the standard behavior), and different values
for di = {1.25, 1.50, 1.75} (all of the deviated agents were
set to have the same di). The quality measure studied was
the average utility since an increment in the utility of one of
the team members will always have a positive effect on the
average utility (same type of valuation functions). The results
of the experiment are depicted in Table IV. We only show the
results for the long deadline scenario, but it should be noted
that the same pattern was found for short deadline scenarios. It
can be observed that all the combinations obtain similar results
in terms of average utility. There is only a slight decrement in
the average utility as we move to more demanding attitudes
(e.g., Ahd = 4, di = 1.75). Even though, the differences
between the most demanding behaviors and other behaviors
are not large enough to be considered significant. Thus, the
experimental results suggest that team members may not have
incentives to deviate much from the proposed strategy. A
closer look at the negotiation traces explained the previous
results. While being more demanding may obtain higher
A = 4
Asd = 1
Asd = 2
Asd = 3
Asd = 4
Ahd = 1
Ahd = 2
Ahd = 3
Ahd = 4
A = 4
Asd = 1
Asd = 2
Asd = 3
Asd = 4
Ahd = 1
Ahd = 2
Ahd = 3
Ahd = 4
di = 1
[0.71-0.72]
-
-
-
-
-
-
-
-
di = 1
206
-
-
-
-
-
-
-
-
di = 1.25
di = 1.5
di = 1.75
12
-
[0.70-0.71]
[0.70-0.72]
[0.69-0.70]
[0.69-0.71]
[0.70-0.71]
[0.69-0.71]
[0.69-0.70]
[0.68-0.69]
di = 1.75
-
199
236
243
287
236
274
302
324
-
[0.70-0.72]
[0.71-0.72]
[0.71-0.72]
[0.70-0.72]
[0.70-0.72]
[0.70-0.71]
[0.69-0.71]
[0.69-0.70]
di = 1.25
-
[0.71-0.72]
[0.71-0.72]
[0.70-0.72]
[0.70-0.72]
[0.71-0.72]
[0.70-0.72]
[0.69-0.71]
[0.69-0.70]
di = 1.5
-
208
202
199
248
202
241
256
299
-
205
230
240
267
189
246
301
292
TABLE IV
THE UPPER TABLE SHOWS CONFIDENCE INTERVALS FOR THE AVERAGE
UTILITY DEPENDING ON THE NUMBER OF deviated AGENTS AND di. THE
LOWER TABLE SHOWS THE NUMBER OF FAILED NEGOTIATIONS FOR EACH
CASE. SOME OF THE COMBINATIONS ARE EMPTY SINCE THEY DO NOT
MAKE SENSE IN PRACTICE (E.G., 0 deviated AGENTS AND di > 1).
utilities in successful negotiations, it may also lead to a higher
number of failed negotiations, thus leading to lower or equal
average utilities. These results can be observed also at Table
IV, where there is a clear tendency for the number of failed
negotiations to increase as team members deviate further from
the standard behavior.
V. RELATED WORK
In the last few years, there has been growing interest in
multiagent systems as a support for complex and distributed
systems. Among these complex systems,
there is special
interest
in scenarios where multiple agents, with possibly
conflicting goals, cooperate with each other to reach their
own goals. The benefits of cooperation and coordination are
well known, and, as stated by Klein [9], computer systems
may help us to identify and apply the appropriate coordina-
tion mechanism. Due to the inherent conflict among agents,
techniques that allow agents to solve their own conflicts and
cooperate are needed. This need is what has given birth to a
group of technologies which have recently been referred to as
agreement technologies [10]. Trust and reputation [8], norms
[11], agent organizations [12], [13], argumentation [14], [15]
and automated negotiation [16], [17] are part of the core that
makes up this new family of technologies.
Despite being part of agreement technologies, automated
negotiation has been studied by scholars for a few years.
Automated negotiation consists of an automated search pro-
cess for an agreement between two or more parties where
participants exchange proposals. Two different research trends
can be distinguished in automated negotiation models. The first
type of model aims to calculate the optimum strategy given
certain information about the opponent and the negotiation
environment [18], [19]. The second type of model encloses
heuristics that do not calculate the optimum strategy but obtain
results that aim to be as close to the optimum as possible [6],
Work
Teams Mediated
Faratin et al. [6], [21]
Jonker et al. [20]
Lai et al. [7]
Sanchez-Anguix et al. [17]
Ehtamo et al. [22]
Klein et al. [23]
Ito et al. [24]
Sanchez-Anguix et al. (2011) [4]
No
No
No
No
No
No
No
Yes
No
No
No
No
Yes
Yes
Yes
Yes
Parties
2
2
2
2
n
n
n
2
A BRIEF OVERVIEW OF STATE-OF-THE-ART NEGOTIATION MODELS AND
TABLE V
THEIR FEATURES
[20], [21], [7]. These models assume imperfect knowledge
about
the opponent and the environment, and aim to be
computationally tractable while obtaining good results. This
present work can be classified into the latter type of models.
Most of the research has concentrated on bilateral models
where each party is a single individual. Our article studies
bilateral negotiations where at least one of the parties is a
negotiation team, made up of more than a single individual.
It should be noted that the problem of finding an agreement
for a negotiation team is inherently complex since it not only
requires finding an agreement with the other party but it also
entails reaching some type of consensus within the team.
Even though communications with the opponent party may
be similar to classic bilateral models, negotiation teams may
require an additional level of negotiation that involves team
members. Thus, classical bilateral models cannot be applied
directly if a certain level of consensus is necessary regarding
team decisions. As far as we know, our previous work is [4]
is the only work that focuses on negotiation teams. Despite
that, bilateral negotiation is perhaps the most similar topic to
our current research. Hence, we describe some of the most
important negotiation models that use imperfect knowledge.
A brief overview of these models can be observed in Table V.
Faratin et al. [6] propose a non-mediated bilateral ne-
gotiation model for service negotiation where agents apply
and mix different concession tactics (i.e.,
time-dependent,
imitative and resource-dependent). In their work, they analyze
the impact of the model's parameters and determine which
configurations work better in different scenarios by means
of experiments. Our proposed work also assumes the use of
time-dependent concession strategies for the calculation of
agents' aspirations at each negotiation round. Additionally, we
also take an experimental approach to validate the impact of
our model's parameters. Later, the authors proposed a non-
mediated bilateral negotiation model [21] whose main novelty
was the use of trade-offs to improve agreements between two
parties. A trade-off consists of reducing the utility obtained
from some negotiation issues with the goal of obtaining the
same exact utility from other negotiation issues. The rationale
behind trade-offs is to make the offer more likable for the
opponent while maintaining the same level of satisfaction for
the proposing agent. For that purpose, the authors propose a
fuzzy similarity heuristic that proposes the most similar offer
to the last offer received from the opponent. Our model does
not leave room for trade-offs since the offer that is calculated
13
at each negotiation round is deterministic with respect to the
agenda of issues and the current aspiration level of team
members. However, its main strength lies in the fact that it
is capable of guaranteeing the desired level of utility for each
team member at each round.
Jonker and Treur propose the Agent-Based Market Place
(ABMP) non-mediated model [20] where agents, engage in
bilateral negotiations. ABMP is a negotiation model where
proposed bids are concessions to previous bids. The amount
of concession is regulated by the concession factor (i.e.,
reservation utility), the negotiation speed, the acceptable utility
gap (maximal difference between the target utility and the
utility of an offer that
is acceptable), and the impatience
factor (which governs the probability of the agent leaving the
negotiation process).
Lai et al. [7] propose a non-mediated bilateral negotiation
model where agents are allowed to propose up to k different
offers at each negotiation round. Offers are proposed from the
current iso-utility curve according to a similarity mechanism
that selects the most similar offer to the last offer received from
the opponent. The selected similarity heuristic is the Euclidean
distance since it is general and does not require domain-
specific knowledge and information regarding the opponent's
utility function. Results showed that the strategy is capable of
reaching agreements that are very close to the Pareto Frontier.
Sanchez-Anguix et al. [17] proposed an enhancement for this
non-mediated strategy in environments where computational
resources are very limited and utility functions are complex.
It relies on genetic algorithms to sample offers that are
interesting for the agent itself and creates new offers during the
negotiation process that are interesting for both parties. Results
showed that the model is capable of obtaining statistically
equivalent results to similar models that had the full iso-utility
curve sampled, while being computationally more tractable.
Another topic that resembles team negotiations are multi-
party negotiations. Several works have been proposed in
the literature along this line [22], [23], [24]. For instance,
Ehtamo et al. [22] propose a mediated multi-party negotiation
protocol which looks for joint gains in an iterated way. The
algorithm starts from a tentative agreement and moves in
a direction according to what
the agents prefer regarding
some offers' comparison. Results showed that the algorithm
converges quickly to Pareto optimal points. Klein et al. [23]
propose a mediated negotiation model which can be extended
to multiple parties. Their main goal is to provide solutions
for negotiation processes that use complex utility functions
to model agents' preferences. The negotiation attributes are
no longer independent, and, thus, preference spaces cannot
be explored as easily as in the linear case. Later, Ito et
al. [24] proposed different types of utility functions (cube
and cone constraints) and multiparty mediated negotiation
models that obtain good quality results for the proposed utility
functions. The main difference between our work and multi-
party negotiations lies in the nature of the conflict and how
protocols are devised. Even though each team member could
be viewed as a participant in a multi-party negotiation with the
opponent, it is natural to think that team members' preferences
are more similar (e.g., a team of buyers, a group of friends,
etc.) and they trust other teammates more than the opponent
(i.e., they may share more information). Furthermore, multi-
party negotiation models may be unfair for agents that are
alien to the team if the number of team members exceeds the
number of other participants. In that case, multi-party models
may be inclined to move the negotiation towards agreements
that maximize the preferences of team members.
Multi-agent teamwork is also a close research topic. Agent
teams have been proposed for a variety of tasks such as
Robocup [25], rescue tasks [26], and transportation tasks [27].
However, as far as we know, there is no published work that
considers teams of agents negotiating with an opponent. Most
works in agent teamwork consider fully cooperative agents that
work to maximize shared goals. The team negotiation setting
is different since, even though team members share a common
interest related to the negotiation, there may be competition
among team members to maximize one's own preferences.
Finally, given the results obtained by our proposed model,
consensus building should be mentioned as a close research
topic outside agent research. Our proposed model is capable of
attaining consensus/unanimity regarding team decisions under
the assumption of private information. Consensus building
works like [28], [29] usually take the assumption that all
the information regarding parties is available to a trusted
mediator. Cook et al. [28], generalize the use of distance-based
measures to obtain consensus over multiple decision makers
with ordinal preferences by assigning utility weights to ordinal
positions. They show that this representation is equivalent to
the commonly used model of using ordinal positions as utility
weights for options. Herrera-Viedma et al. [29] propose a
computational model that is able to help humans/experts to
reach soft consensus over a set of alternatives. The model is
based on an iterative process where two measures are used to
achieve this result: a soft consensus measure, and a proximity
measure. Both measures are used to evaluate how close the
individual expert opinion with respect to the collective opinion
is, and help the computational system to provide feedback to
experts that are far from the group's opinion. Even though the
goal is similar to our work, the assumptions are different. We
advocate for open systems like e-commerce systems, where
agents act semi-automatically on behalf of their users. Since
any type of agent can be found in open environments, privacy
is a big concern due to distrust and risk of exploitation. Thus,
it is not possible for a mediator to know the preferences and
all of the information about the participants in the negotiation.
In our approach, we only consider that a limited amount of
information is transmitted to the mediator.
VI. CONCLUSIONS AND FUTURE WORK
In this article, we have proposed an agent-based negotiation
model for negotiation teams that interact with an opponent
using the bilateral alternating protocol in electronic systems.
A negotiation team is a group of two or more agents that
join together as a single negotiation party because they share
a common goal which is related to the negotiation process.
Thus, as a team, they have to decide which offers are sent
to the opponent and whether or not the offers received from
14
the opponent are acceptable. The main strength of our pro-
posed model lies in the fact that decisions within the team
are unanimous (i.e., the utility reported by the decision is
greater than or equal to the desired utility level by each team
member). The negotiation model relies on a trusted mediator
that coordinates voting processes, regulates an iterated process
for offer construction, and guarantees unanimity.
After describing our proposal, we have shown how the
model is capable of ensuring unanimity regarding team de-
cisions. Then, we theoretically analyzed the robustness of
our model against different types of attacks. The proposed
model is robust against manipulations from the opponent, but
it is sensitive to manipulations coming from agents from the
competition that try to sabotage a possible agreement with the
opponent. We also presented different experiments analyzing
the impact of different parameters of the model such as the
negotiation agenda followed by the mediator to decide which
attributes are set first, and the impact of the number of attribute
decision rights that are handed over by team members prior
to the negotiation. Additionally, we carried out an empirical
evaluation of the robustness of the proposed model against
attacks from agents that represent
the competition, which
reflects our initial concerns in the theoretical analysis. Finally,
we studied whether or not team members have incentives to
deviate from the proposed strategy. Empirical results suggest
that there is not much incentive to deviate from the proposed
strategy since deviations may impact the number of failed
negotiations and, thus, the average utility.
Future work includes the evaluation of the present negoti-
ation model and other models proposed in the literature [4]
in different negotiation scenarios. The rationale behind this
analysis is to determine which strategies are more appropriate
for team members according to different criteria such as
utilitarian measures (minimum, average, maximum) and com-
putational measures (number of messages exchanged, number
of negotiation rounds). Our goal is using this knowledge in
a decision-making mechanism that allows teams to select the
most appropriate negotiation model according to their needs.
We acknowledge that the current model is capable of reach-
ing unanimity given the assumption of monotonic valuation
functions and linear utility functions. Therefore, future work
also includes exploring aggregation mechanisms that reach
consensus/soft consensus for non-monotonic attributes like
colors, brands, etc. In this sense, fuzzy consensus/similarity
measures as the ones proposed in [21], [29] can help to
aggregate agents' opinions over this type of attributes.
Moreover, in the last few years there has been a growing in-
terest in modeling negotiation as a dynamic process [30], [31].
Despite being an interesting approach, its study is still at early
stages and focused on simple negotiations, while our current
model involves a negotiation team, and a negotiation party,
which increments the complexity of the modeling problem. In
this sense, the use of new search and optimization algorithms
like gravitational search algorithms [32], pomdp [31], and
machine learning approaches with efficient data selection [33]
can help to further improve the state of the art in negotiation.
Nevertheless, dynamic modeling of agent negotiation teams is
a topic that should be studied in the future.
REFERENCES
[1] S. Brodt and L. Thompson, "Negotiation within and between groups in
organizations: Levels of analysis," Group Dynam., pp. 208 -- 219, 2001.
[2] N. Halevy, "Team negotiation: Social, epistemic, economic, and psycho-
logical consequences of subgroup conflict," Pers. Soc. Psychol. Bull.,
vol. 34, pp. 1687 -- 1702, 2008.
[3] V. S´anchez-Anguix, V. Julian, V. Botti, and A. Garc´ıa-Fornes, "Towards
agent-based negotiation teams," in Group Decision and Negotiation
2010, 2010, pp. 328 -- 331.
[4] -- -- , "Analyzing Intra-Team Strategies for Agent-Based Negotiation
Teams," in 10th International Conference on Autonomous Agents and
Multiagent Systems, 2011, pp. 929 -- 936.
[5] A. Rubinstein, "Perfect equilibrium in a bargaining model," Economet-
rica, vol. 50, no. 1, pp. 97 -- 109, 1982.
[6] P. Faratin, C. Sierra, and N. R. Jennings, "Negotiation decision functions
for autonomous agents," Robot. Autonom. Syst., vol. 24, no. 3-4, pp.
159 -- 182, 1998.
[7] G. Lai, K. Sycara, and C. Li, "A decentralized model for automated
multi-attribute negotiations with incomplete information and general
utility functions," Multiagent and Grid Systems, vol. 4, no. 1, pp. 45 -- 65,
2008.
[8] J. Sabater and C. Sierra, "Review on computational trust and reputation
models," Artif. Intell. Rev., vol. 24, no. 1, pp. 33 -- 60, 2005.
[9] M. Klein, "Coordination science: Challenges and directions," in Co-
ordination Technology for Collaborative Applications-Organizations,
Processes, and Agents, 1998, pp. 161 -- 176.
[10] C. Sierra, V. Botti, and S. Ossowski, "Agreement computing," KI-
Kunstliche Intelligenz, pp. 1 -- 5, 2011.
[11] F. Dignum, "Autonomous agents with norms," Artif. Intell. Law, vol. 7,
no. 1, pp. 69 -- 79, 1999.
[12] B. Horling and V. Lesser, "A survey of multi-agent organizational
paradigms," Knowl. Eng. Rev., vol. 19, no. 04, pp. 281 -- 316, 2004.
[13] S. Esparcia and E. Argente, "Formalizing Virtual Organizations," in 3rd
International Conference on Agents and Artificial Intelligence, vol. 2,
2011, pp. 84 -- 93.
[14] I. Rahwan, S. Ramchurn, N. Jennings, P. Mcburney, S. Parsons, and
L. Sonenberg, "Argumentation-based negotiation," Knowl. Eng. Rev.,
vol. 18, no. 04, pp. 343 -- 375, 2003.
[15] P. Pardo, S. Pajares, E. Onaind´ıa, L. Godo, and P. Dellunde, "Multiagent
Argumentation for Cooperative Planning in DeLP-POP," in 10th Inter-
national Conference on Autonomous Agents and Multiagent Systems,
2011, pp. 971 -- 978.
[16] N. R. Jennings, P. Faratin, A. R. Lomuscio, S. Parsons, M. J. Wooldridge,
and C. Sierra, "Automated negotiation: Prospects, methods and chal-
lenges," Group Decis. Negot., vol. 10, pp. 199 -- 215, 2001.
[17] V. S´anchez-Anguix, S. Valero, V. Julian, V. Botti, and A. Garc´ıa-
Fornes, "Evolutionary-aided negotiation model for bilateral bargaining
in Ambient Intelligence domains with complex utility functions," Inf.
Sci., 2010.
[18] Y. In and R. Serrano, "Agenda restrictions in multi-issue bargaining (ii):
unrestricted agendas," Econ. Lett., vol. 79, no. 3, pp. 325 -- 331, 2003.
[19] S. S. Fatima, M. Wooldridge, and N. R. Jennings, "Multi-issue negoti-
ation with deadlines," J. Artif. Intell. Res., vol. 27, pp. 381 -- 417, 2006.
[20] C. M. Jonker and J. Treur, "An agent architecture for multi-attribute
negotiation," in 17th International Joint Conference on Artificial Intel-
ligence, 2001, pp. 1195 -- 1201.
[21] P. Faratin, C. Sierra, and N. R. Jennings, "Using similarity criteria to
make issue trade-offs in automated negotiations," Artif. Intell., vol. 142,
pp. 205 -- 237, 2002.
[22] H. Ehtamo, E. Kettunen, and R. P. Hamalainen, "Searching for joint
gains in multi-party negotiations," Eur. J. Oper. Res., vol. 130, no. 1,
pp. 54 -- 69, 2001.
[23] M. Klein, P. Faratin, H. Sayama, and Y. Bar-Yam, "Negotiating complex
contracts," Group Decis. Negot., vol. 12, no. 2, pp. 111 -- 125, 2003.
[24] K. Fujita, T. Ito, and M. Klein, "Secure and efficient protocols for
multiple interdependent issues negotiation," J. Intell. Fuzzy Syst., vol. 21,
no. 3, pp. 175 -- 185, 2010.
[25] P. Stone and M. Veloso, "Task decomposition, dynamic role assignment,
and low-bandwidth communication for real-time strategic teamwork,"
Artif. Intell., vol. 110, no. 2, pp. 241 -- 273, 1999.
[26] H. Kitano and S. Tadokoro, "Robocup rescue: A grand challenge for
multiagent and intelligent systems," AI Mag., vol. 22, no. 1, pp. 39 -- 52,
2001.
[27] N. R. Jennings, "Controlling cooperative problem solving in industrial
multi-agent systems using joint intentions," Artif. Intell., vol. 75, no. 2,
pp. 195 -- 240, 1995.
15
[28] W. D. Cook, M. Kress, and L. M. Seiford, "A general framework for
distance-based consensus in ordinal ranking models," Eur. J. Oper. Res.,
vol. 96, pp. 392 -- 397, 1996.
[29] E. Herrera-Viedma, F. Herrera, and F. Chiclana, "A consensus model
for multiperson decision making with different preference structures,"
IEEE Trans. Syst. Man Cybern. Part A-Syst. Hum., vol. 32, no. 3, pp.
394 -- 402, 2002.
[30] P. C. Da Costa and D. M. Buede, "Dynamic decision making: A
comparison of approaches," J. multi-criteria decis. anal., vol. 9, pp.
243 -- 262, 2000.
[31] P. Parachuri, N. Chakraborty, R. Zivan, K. Sycara, M. Dudik, and
G. Gordon, "Pomdp based negotiation modeling," in Modeling Inter-
cultural Collaboration and Negotiation Workshop, 2009, pp. 66 -- 78.
[32] H. Askari and S. Zahiri, "Decision function estimation using intelligent
gravitational search algorithm," Int. J. Mach. Learn. Cybern., pp. 1 -- 10.
[33] S. Zhang, P. McCullagh, C. Nugent, H. Zheng, and M. Baumgarten,
"Optimal model selection for posture recognition in home-based health-
care," Int. J. Mach. Learn. Cybern., vol. 2, pp. 1 -- 14, 2011.
Victor Sanchez-Anguix was born in Valencia,
Spain. He received the B.S. and M.S. degrees in
Computer Science from Universidad Polit´ecnica de
Valencia, Valencia, Spain, in 2008 and 2010, respec-
tively. He is currently a PhD. student in Computer
Science at Universidad Polit´ecnica de Valencia and
holds a research grant supported by the Spanish
government until 2013. His research interests in-
clude agreement technologies, soft computing, and
multiagent systems.
Vicente Julian was born in Valencia, Spain. He
received his PhD. degree in Computer Science from
Universidad Polit´ecnica de Valencia in 2002. His
PhD. dissertation focused on agent methodologies
for real time systems. He is a researcher with Grupo
Tecnolog´ıa Inform´atica - Inteligencia Artificial at
Universidad Polit´ecnica de Valencia. He also holds
an Associate Professor position. His research inter-
ests are adaptive systems, agreement technologies,
and software engineering.
Vicente Botti was born in Valencia, Spain. He
received his PhD. degree in Computer Science from
Universidad Polit´ecnica de Valencia in 1990. His
PhD. dissertation focused on temporal knowledge
based systems. He is currently head researcher with
Grupo Tecnolog´ıa Inform´atica - Inteligencia Artifi-
cial at Universidad Polit´ecnica de Valencia. He also
holds a Full Professor position. Some of his research
interests include agreement technologies, and real
time systems.
Ana Garcia-Fornes was born in Valencia, Spain.
She received her PhD. degree in Computer Science
from Universidad Polit´ecnica de Valencia in 1996.
Her PhD. dissertation focused on real time architec-
tures for intelligent systems. She is head researcher
with Grupo Tecnolog´ıa Inform´atica - Inteligencia
Artificial at Universidad Polit´ecnica de Valencia. She
also holds an Associate Professor position. Real
time systems, agent infrastructures and agreement
technologies are some of her interests.
|
1911.04146 | 1 | 1911 | 2019-11-11T09:03:06 | Optimal Common Contract with Heterogeneous Agents | [
"cs.MA",
"cs.GT"
] | We consider the principal-agent problem with heterogeneous agents. Previous works assume that the principal signs independent incentive contracts with every agent to make them invest more efforts on the tasks. However, in many circumstances, these contracts need to be identical for the sake of fairness. We investigate the optimal common contract problem. To our knowledge, this is the first attempt to consider this natural and important generalization. We first show this problem is NP-complete. Then we provide a dynamic programming algorithm to compute the optimal contract in $O(n^2m)$ time, where $n,m$ are the number of agents and actions, under the assumption that the agents' cost functions obey increasing difference property. At last, we generalize the setting such that each agent can choose to directly produce a reward in $[0,1]$. We provide an $O(\log n)$-approximate algorithm for this generalization. | cs.MA | cs |
OPTIMAL COMMON CONTRACT WITH
HETEROGENEOUS AGENTS
Shenke Xiao∗ 1, Zihe Wang† 2, Mengjing Chen‡ 1, Pingzhong Tang§ 1, and Xiwang Yang¶3
1Tsinghua University
2Shanghai University of Finance and Economics
3ByteDance
ABSTRACT
We consider the principal-agent problem with heterogeneous agents. Previous works
assume that the principal signs independent incentive contracts with every agent
to make them invest more efforts on the tasks. However, in many circumstances,
these contracts need to be identical for the sake of fairness. We investigate the
optimal common contract problem. To our knowledge, this is the first attempt to
consider this natural and important generalization. We first show this problem is
NP-complete. Then we provide a dynamic programming algorithm to compute the
optimal contract in O(n2m) time, where n, m are the number of agents and actions,
under the assumption that the agents' cost functions obey increasing difference
property. At last, we generalize the setting such that each agent can choose to
directly produce a reward in [0, 1]. We provide an O(log n)-approximate algorithm
for this generalization.
1 Introduction
Principal-agent theory is a subfield of mechanism design theory. The principal hires an agent to
accomplish a task. The agent is able to take actions on behalf of the principal. Agent's different
actions lead to different rewards the principal receives. Moral hazard occurs when the agent acts
in his own interest which may be in conflict with the principal's interest. Therefore the principal
designs an incentive contract with the agent to maximize the principal's utility subject to the agent's
utility being maximized. The contract is a transfer function from the principal to the agent which
could depend on the outcome which is affected by the agent's action.
Many economic interactions fit in the principal-agent model. For example, a firm (principal) hires a
salesman (agent) to sell products. The salesman invests effort on selling products. More efforts he
invests, more products will be sold. To incentivize salesman invest more efforts, the firm can set a
bonus depending on the amount of products a salesman has sold. A salesman wants to maximize his
utility which is defined to be his bonus minus his efforts. The firm's utility is the revenue generated
∗[email protected]
†[email protected]
‡[email protected]
§[email protected]
¶[email protected]
from selling products minus the bonus paid to salesmen. The central question in this research field
asks: What is the principal's optimal contract?
Due to the wide application, principal-agent model has been extensively studied [1, 2, 3, 4, 5]. The
agent takes a hidden action like effort which cannot be observed by the principal directly. The
principal can only observe the outcome of this action and the contract is designed to depend on the
outcome only. Most works focus on the problem with one principal and one agent [6, 7]. When
the agent takes different actions, there is a different distribution over principal's reward. Given the
distribution information, the optimal contract can be computed efficiently through linear programs.
In this paper, we consider the problem when there is one principal and multiple heterogeneous
agents. These agents could be good at different tasks and we do not assume any relationship
between the cost for different tasks among different agents. For sake of the fairness, we do not allow
the principal design personal contracts for different agents. Instead, the principal has to design a
common contract that applies to every agents. We assume the mapping from the action played to
the outcome is deterministic. So the principal knows every agent's action by observing her outcome.
The difficulty in our model stems from the multiple agents. Since the principal can only use a
common contract, he needs to balance the incentivization for every agent.
1.1 Our Contribution
Our contribution can be summarized as follows.
1. We first show that the optimal contract problem with heterogeneous agents is strongly
NP-complete.
2. We then proposes an O(n2m) dynamic programming algorithm, where n is the number of
agents and m is the number of actions, to compute an optimal contract under the assumption
that the agents' costs obey increasing differences.
3. Next, we generalize the discrete-action setting such that each agent can choose to directly
produce a reward in [0, 1]. We shows that this generalization is harder than the original
discrete-action version, and provides an O(log n)-approximate algorithm for this generaliza-
tion.
1.2 Other Related Works
Other works also consider multiple agents but in different angles [8, 9, 10]. They assume the union
of agents' actions together determines the outcome. The contract is personalized and specifies the
payment in every possible outcome. Therefore the payment to an agent depends on both his action
and other agents' actions. In contrast, in our paper, the payment to an agent only depends on his own
action. In their setting, Babaioff et al. consider that each agent only has a binary action space [8],
Babaioff and Winter consider the tradeoff between simplicity of the contract and the performance
of it [10].
Lavi and Shamash study the model with multiple principals and multiple agents [11]. Agents do
not have cost on actions. This model focuses on the competition between principals. McAfee and
McMillan study another totally different problem where multiple agents compete for a principal's
2
contract [12]. A recent work of Azizan et al. studies a model that is almost the same as ours where
each agent can choose to directly produce a real number reward [13]. However, they assume the
designed payment function can be parameterized by a vector in a given set A ⊆ Rd, and their
algorithm explores the whole set A, which is not that efficient.
2 Problem Description
In this paper, we study the Multiple Agents Contract Problem. There is a principal, n agents and m
actions. Each agent can take an action j ∈ [m] and produces a reward ρj ≥ 0 for the principal. The
reward only depends on the action, not on the agent. Each agent i also has a cost ci,j ≥ 0 to take an
action j. This cost depends on both the agent and the action. Besides the m actions, there is always
a zero action with reward 0 such that the cost for each agent to take this action is 0. This action
means it is free for each agent to choose to produce nothing. The principal specifies a payment
profile (t1, t2, . . . , tm): each agent taking action j will earn a payment tj. The utility for agent i
to take action j is tj − ci,j. The agents are self-interested meaning each agent will take an action
that maximizes its utility. W.l.o.g., we assume the agents tie-break in favor of the principal. The
payoff of the principal is the sum of rewards produced by the agents minus the payments given to
i=1(ρi∗ − ti∗). Our goal is
the agents, i.e., if agent i takes action i∗, the payoff of the principal is(cid:80)n
to design the payment profile (t1, t2, . . . , tm) to maximize the payoff of the principal.
Example 1. Suppose there are two agents and two actions. The rewards for the two actions are
8 and 10 respectively. For action 1, agent 1 has a cost 5 and agent 2 has a cost 4. For action 2,
agent 1 has a cost 9 and agent 2 has a cost 2. Without agent 2, we can set the payments for the two
actions to 5 and 0 respectively, which brings a payoff of 3 to the principal. Without agent 1, we
can set the payments for the two actions to 0 and 2 respectively, which brings a payoff of 8 to the
principal. However, when the two agents both exist, no matter how we set the payments, the payoff
of the principal cannot achieve 3 + 8 = 11. It is optimal to the payments for the two actions to 5
and 3 respectively, which brings a payoff of 10 to the principal.
3 Hardness
The problem defined in the previous section is very hard. To see its hardness, let us consider its
decision version, i.e., the problem of determining whether there is a payment profile (t1, t2, . . . , tm)
such that the payoff of the principal is no less than a given number r. For convenience, we call this
decision problem MAC. We will show in the following theorem that MAC is strongly NP-complete.
Theorem 1. MAC is strongly NP-complete.
Proof. MAC obviously belongs to NP. In the following proof, we reduce the well-known NP-
complete problem Not-All-Equal 3-Satisfiability (NAE3SAT) to MAC to show that MAC is strongly
NP-complete.
Given an instance of NAE3SAT with n variables and m clauses (we assume the variables in one
clause are different without loss of generality), we build an instance of MAC as follows. For any
variable x in an instance of NAE3SAT, we define x0 as its negation and define x1 = x.
3
• Agents
-- For each variable xi, we have an agent Ai.
-- For each literal xb
-- For each clause cj, we have 6 agents Vj,1, Vj,2, . . . , Vj,6.
i,j.
i and each clause cj, we have an agent T b
• Actions
-- We have a zero action zero with reward 0.
-- For each literal xb
-- For each clause cj, we have 6 actions clausej,1, . . . , clausej,6 with reward ρ2.
i, we have an action variableb
i with reward ρ1.
• Costs
-- For the zero action zero, each agent has a cost 0.
-- For action variableb
-- For action clausej,k, agent Vj,k has a cost 0.
-- For action clausej,k where cj = xb1
i1
i, agent Ai has a cost δ, and T b
∨ xb3
∨ xb2
i,j has a cost 0 for each j.
i3 , the costs vary for different k's and are
summarized in Table 1. Note there are exactly 3 agents with cost 1 to take this action.
We call the three agents the associated agents of this action.
i2
-- For each action and agent, if we do not mention the cost above, the cost is greater than
the reward of the action.
The parameters ρ1, ρ2, δ satisfy the following constraints.6
ρ1 − δ > m((2n − 3)(ρ2 − ρ1) + nδ + 4),
δ > 3(ρ2 − ρ1 − 1),
ρ2 − ρ1 > 2.
(1)
(2)
(3)
Figure 1 shows an example instance of MAC corresponding to an instance of NAE3SAT with 4
variables and 2 clauses x1 ∨ x2 ∨ ¬x3 and x1 ∨ ¬x2 ∨ x4.
Table 1: Cost Table
T 1−b1
T b1
i1
1
1
-
-
-
1
clausej,1
clausej,2
clausej,3
clausej,4
clausej,5
clausej,6
T b2
i2
1
-
1
-
1
-
T b3
i3
-
1
1
1
-
-
T 1−b2
i2
-
1
-
1
-
1
T 1−b3
i3
1
-
-
-
1
1
i1
-
-
1
1
1
-
Then we ask whether we can set the payments to the agents so that the optimal payoff of the
principal is no less than
n(ρ1 − δ) + m(6ρ2 − 1) + m(n(ρ1 − δ) + (n − 3)ρ1 + 3(ρ2 − 1)).
(4)
6For example, we can set δ = 7, ρ1 = 13mn + 8 and ρ2 = 13mn + 11.
4
ρ1
ρ1
ρ1
ρ1
A1
δ
1,1, T 0
T 0
0
1,2
variable0
ρ1
1
A1
δ
1,1, T 1
T 1
0
1,2
variable1
ρ1
1
A2
δ
2,1, T 0
T 0
0
2,2
variable0
ρ1
2
A2
δ
2,1, T 1
T 1
0
2,2
variable1
ρ1
2
A3
δ
T 0
3,1, T 0
0
3,2
variable0
ρ2
3
A3
δ
T 1
3,1, T 1
0
3,2
variable1
ρ2
3
A4
δ
T 0
4,1, T 0
0
4,2
variable0
4
A4
δ
T 1
4,1, T 1
0
4,2
variable1
ρ2
4
2,1, T 1
3,1
1,1, T 1
T 1
V1,1
1
0
clause1,1
ρ2
2,1, T 0
3,1
1,1, T 0
T 1
V1,2
1
0
clause1,2
ρ2
2,1, T 0
3,1
1,1, T 1
T 0
V1,3
1
0
clause1,3
ρ2
2,1, T 0
3,1
1,1, T 0
T 0
V1,4
1
0
clause1,4
ρ2
2,2, T 0
4,2
T 1
1,2, T 0
V2,1
1
0
clause2,1
ρ2
2,1, T 1
3,1
1,1, T 1
T 0
V1,5
1
0
clause1,5
ρ2
2,1, T 1
3,1
1,1, T 0
T 1
V1,6
1
0
clause1,6
ρ2
2,2, T 1
4,2
T 1
1,2, T 1
V2,2
1
0
clause2,2
ρ2
2,2, T 1
4,2
T 0
1,2, T 0
V2,3
1
0
clause2,3
ρ2
1
0
clause2,4
T 0
1,2, T 1
V2,4
2,2, T 1
4,2
1
0
clause2,5
T 0
1,2, T 0
V2,5
2,2, T 0
4,2
1
0
clause2,6
T 1
1,2, T 1
V2,6
2,2, T 0
4,2
Figure 1: An example instance of MAC corresponding to an instance of NAE3SAT with 4 variables
and 2 clauses x1 ∨ x2 ∨ ¬x3 and x1 ∨ ¬x2 ∨ x4: each rectangular represents an action; each line
(including the bottom line) in a rectangular represents one or multiple agents, whose names are
recorded to the right of the line; the number to the left of a line represents the cost for the agents to
take this action; in particular, the number to the left of the top line of a rectangular represents the
reward for the action.
5
In an optimal solution, the payment for an action will not exceed the reward, so an agent will never
be incentivized to take an action whose cost is greater than the reward. The actions that an agent
will be potentially incentivized to take in an optimal solution is summarized as follows (since all
agents can take the zero action, we omit it in the following list).
• Agent Ai will potentially take action variable0
• Agent T b
i,j will potentially take action variableb
clausej,k for some k.
i or variable1
i .
i or, if variable xi appears in clause cj,
• Agent Vj,k will potentially take action clausej,k.
Suppose the instance of NAE3SAT has a valid solution, then we set the payments in the instance of
MAC as follows.
• For each action variableb
we set the payment to δ.
i, if the value of xb
i is True, we set the payment to 0; otherwise
• For each action clausej,k with associated agents T h1
i1,j, T h2
i2,j, T h3
i3,j, we set the payment to 1 if
the values of xh1
i1
, xh2
i2
, xh3
i3 are all True; otherwise we set the payment to 0.
i . For agent T b
i ; otherwise she will take action variable0
i. Hence the total payoff of the principal is exactly (4).
Under these payments, agent Vj,k will always take action clausej,k. If the value of xi is True, agent
Ai will take action variable1
i,j, if the
i is True and variable xi appears in clause cj, she will take action clausej,k for some k;
value of xb
otherwise she will take action variableb
Now suppose there exists a payment setting such that the optimal payoff of the principal is no
less than (4). We first show that agent Ai will take one of the actions variable0
i and variable1
i .
Otherwise, the payoff of the principal cannot exceed (n − 1)(ρ1 − δ) + 6mρ2 + 2mnρ2, which is
less than (4) by (1).
We define bi such that Ai takes action variablebi
i must
be no less than δ -- the cost for agent Ai to take this action. If the payment is greater than δ, we
can adjust it to δ. After this adjustment, some agent T bi
i before the
adjustment may turn out to take action clausej,k for some k. This is the only possible cause of
payoff loss of the principal. Suppose the payments for variablebi
i (before the adjustment) and
clausej,k are t1, t2 respectively, since agent T bi
i before the
adjustment, we have t1 ≥ t2 − 1, so ρ2 − t2 ≥ ρ2 − t1 − 1 > ρ1 − t1 by (3). This means the
adjustment does not reduce the payoff of the principal, hence we can assume the payment for action
is exactly δ. By an analogous argument, we can also assume the payment for action
variablebi
i
variable1−bi
If there exist some i, j such that agent T bi
clausej,k for some k, and the payment t for action clausej,k must incentivize agent T bi
action clausej,k, i.e., it must satisfy
i , it must take action
i,j to take
i , then the payment for action variablebi
i,j chooses to take action variablebi
i,j does not take action variablebi
i,j that takes action variablebi
is exactly 0.
i
(5)
t − 1 ≥ δ.
6
i,j takes action variablebi
i .
Then we adjust the payment for action clausej,k to 1 so that agent T bi
i,j is incentivized to take
i . After this adjustment, at most three agents that take action clausej,k before
action variablebi
the adjustment deviate to take actions of the form variablek(cid:48)
· . Each of these agent brings a payoff
of ρ2 − t to the principal before the adjustment, and brings a payoff of at least ρ1 − δ after the
adjustment, so the adjustment reduces the payoff of the principal by at most 3(ρ2 − t − ρ1 + δ). On
the other hand, the payoff of the principal increases by t − 1 due to the contribution of agent Vj,k.
As a result, since t − 1 ≥ 3(ρ2 − t − ρ1 + δ) due to (2) and (5), this adjustment does not reduce the
payoff of the principal. Hence, we can assume for any i, j, agent T bi
Suppose there exists some j such that the payment for action clausej,k is less than 1 for each k.
Suppose clause cj contains three variables xi1, xi2, xi3, and action clausej,k0 is an action that the
cost for agent T 1−bi1
to take is 1 (there may exist multiple such k0's, and we arbitrarily choose
one). We then adjust the payment for clausej,k0 to 1. This adjustment attracts T 1−bi1
to take action
clausej,k0, which increases the payoff of the principal by ρ2 − 1 − ρ1. On the other hand, the
payoff of the principal contributed by Vj,k0 is decreased by at most 1, which is the only cause that
reduces the payoff of the principal. As a result, since ρ2 − 1 − ρ1 > 1, the payoff of the principal
increases. Hence, we can assume for any j, there exists at least one k such that the payment for
action clausej,k is no less than 1.
Under the assumptions above, the maximum payoff of the principal is exactly (4). To achieve this
i3 , there exists exactly one k such that the payment
optimal payoff, for any j, say cj = xh1
i1
for clausej,k is 1, and its three associated agents T 1−bi1
take this action. According
to Table 1, (1 − bi1) ⊕ h1, (1 − bi2) ⊕ h2, (1 − bi3) ⊕ h3 do not have the same value. So we can
set variable xi to the value 1 − bi (0 represents False and 1 represents True), then all clauses are
satisfied.
, T 1−bi3
, T 1−bi2
i2,j
i1,j
i1,j
i1,j
∨ xh2
i2
∨ xh3
i3,j
4 Increasing Differences
In this section, we consider the case where agents have different abilities. Roughly speaking, the
agents can be ordered from weak to strong, i1, i2, . . . , in, in the sense that it takes less cost for a
stronger agent to produce a certain amount reward. We have for each j ∈ [m],
(6)
ci1,j > ci2,j > ··· > cin,j,
Additionally, we assume the costs obey increasing differences.
Definition 1. Given an instance of the Multiple Agents Contract Problem, we call the costs obey
increasing differences if there exists a permutation j1, j2, . . . , jm of 1, 2, . . . , m and a permutation
i1, i2, . . . , in of 1, 2, . . . , n such that for any k < k(cid:48), 0 < cik,j1 − cik(cid:48) ,j1 < cik,j2 − cik(cid:48) ,j2 < ··· <
cik,jm − cik(cid:48) ,jm.
Though MAC is proved to be hard, we give a dynamic programming algorithm to solve the Multiple
Agents Contract Problem under the assumption that the costs obey increasing differences.
The permutation i1, i2, . . . , in can be found in O(n log n) time by sorting c1,j, c2,j, . . . , cn,j for
an arbitrary j, then the permutation j1, j2, . . . , jm can be found in O(m log m) time by sorting
ci2,1 − ci1,1, ci2,2 − ci1,2, . . . , ci2,m − ci1,m. For convenience, we assume the actions and agents are
7
already ordered without loss of generality, i.e., ik = jk = k for each k. The zero action is also
considered action 0.
Before describing the algorithm, we first show the following lemmas.
Lemma 1. If the costs obey increasing differences, then for any payment profile (t1, t2, . . . , tm), if
agent i and i(cid:48) take actions j and j(cid:48) respectively, then i < i(cid:48) ⇒ j ≤ j(cid:48).
Proof. Suppose i < i(cid:48) but j > j(cid:48). Since agent i prefers action j to j(cid:48), we have
tj − ci,j ≥ tj(cid:48) − ci,j(cid:48).
Similarly, since agent i(cid:48) prefers action j(cid:48) to j, we have
tj(cid:48) − ci(cid:48),j(cid:48) ≥ tj − ci(cid:48),j.
By combining (7) and (8), we have
(7)
(8)
ci,j − ci(cid:48),j ≤ ci,j(cid:48) − ci(cid:48),j(cid:48).
(9)
However, since i < i(cid:48) and j > j(cid:48), by increasing differences we have ci,j − ci(cid:48),j > ci,j(cid:48) − ci(cid:48),j(cid:48), which
contradicts to (9). Therefore, we must have i < i(cid:48) ⇒ j ≤ j(cid:48).
Lemma 2. Given any 0 ≤ j1 ≤ ··· ≤ jm ≤ m, we have
1. Under the constraint that agent i is incentivized to take action ji, the optimal payoff of the
principal cannot exceed
(ρji − ci,ji − (n − i) (ci,ji − ci+1,ji)) + ρjn − cn,jn.
(10)
n−1(cid:88)
i=1
tj =
0,
2. If the costs obey increasing differences, and we set the payment profile (t1, t2, . . . , tm) such
that
(cid:80)i−1
i(cid:48)=1
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
(cid:1) + ci,ji,
if there ex-
ists i such
that j =
ji
otherwise,
7,
(11)
then the payoff of the principal is no less than (10).
Proof. Since agent i(cid:48) prefers action ji(cid:48) to ji(cid:48)−1, we have
(12)
7If there exist multiple such i's, we arbitrarily choose one, because if, for example, j = jk(cid:48) = jk(cid:48)+1 = ··· = jk,
tji(cid:48) − ci(cid:48),ji(cid:48) ≥ tji(cid:48)−1 − ci(cid:48),ji(cid:48)−1,
(cid:1) + ci,ji is the same for i = k(cid:48), k(cid:48) + 1, . . . , k.
then the value of(cid:80)i−1
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
i(cid:48)=1
8
and for i(cid:48) = 1 we have tj1 − ci,j1 ≥ 0 since agent 1 prefers action j1 to the zero action. By summing
up (12) for i(cid:48) = 1, 2, . . . , i, we have
i(cid:48)=1
tji ≥ i−1(cid:88)
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
(cid:32) i−1(cid:88)
tji ≥ n(cid:88)
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
n−1(cid:88)
i=1
=
(cid:1) + ci,ji.
(cid:1) + ci,ji
(cid:33)
Hence,
n(cid:88)
i=1
i(cid:48)=1
(ci,ji + (n − i) (ci,ji − ci+1,ji)) + cn,jn,
i=1
so the payoff of the principal cannot exceed (10).
On the other hand, suppose the costs obey increasing differences and we set tj according to (11).
For any agent i and any action j, there are three cases.
1. If there does not exist some k such that j = jk, then
tj − ci,j ≤ 0
≤ i−1(cid:88)
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
(cid:1)
i(cid:48)=1
= tji − ci,ji,
where the inequality (13) holds due to (6).
2. If there exists some k ≤ i such that j = jk, we have
i(cid:48)=1
=
=
i(cid:48)=1
tj − ci,j
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
k−1(cid:88)
k−1(cid:88)
≤ k−1(cid:88)
i−1(cid:88)
=
= tji − ci,ji,
i(cid:48)=1
i(cid:48)=1
(cid:1) + ck,jk − ci,jk
i−1(cid:88)
(cid:1) +
i−1(cid:88)
(cid:1) +
(cid:1)
i(cid:48)=k
i(cid:48)=k
(ci(cid:48),jk − ci(cid:48)+1,jk)
(cid:1)
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
(13)
(14)
where the inequality (14) holds due to increasing differences: for any i(cid:48) ≥ k, ci(cid:48),jk−ci(cid:48)+1,jk ≤
ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48) .
9
3. If there exists some k > i such that j = jk, we have
i(cid:48)=1
=
=
tj − ci,j
k−1(cid:88)
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
(cid:1) − (ci,jk − ck,jk)
i−1(cid:88)
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
(cid:1)
k−1(cid:88)
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
≤ i−1(cid:88)
(cid:1)
(cid:0)ci(cid:48),ji(cid:48) − ci(cid:48)+1,ji(cid:48)
(cid:1) − k−1(cid:88)
i(cid:48)=1
+
i(cid:48)=i
i(cid:48)=1
= tji − ci,ji,
(ci(cid:48),jk − ci(cid:48)+1,jk)
i(cid:48)=i
(15)
where the inequality (15) holds due to increasing differences: for any i(cid:48) < k, ci(cid:48),ji(cid:48) −
ci(cid:48)+1,ji(cid:48) ≤ ci(cid:48),jk − ci(cid:48)+1,jk.
Anyway, we have tj − ci,j ≤ tji − ci,ji, which means taking action ji maximizes agent i's utility.
Note if agent i takes action ji for each i, the payoff of the principal is exactly (10). Recall that the
agents tie-break in favor of the principal, so the payoff of the principal is no less than (10).
Lemma 1 and 2 show that we can find 0 ≤ j1 ≤ j2 ≤ ··· ≤ jm ≤ m that maximizes (10), then
an optimal payment profile is given by (11). To find the optimal j1, j2, . . . , jm, we use a dynamic
programming algorithm. For convenience, we define φ(i, j) = ρj − ci,j − (n − i) (ci,j − ci+1,j)
for i = 1, 2, . . . , n − 1, and define φ(n, j) = ρj − cn,j. We define the subproblem OPT(i, j) =
i(cid:48)=1 φ(i(cid:48), ji(cid:48)). We can see the optimal value of (10) is OPT(n, m), and we
max0≤j1≤j2≤···≤ji≤j
have the recursion formula
(cid:80)i
OPT(i, j + 1) =
(cid:32)
max
0≤k≤i
OPT(k, j) +
(cid:33)
i(cid:88)
i(cid:48)=k+1
φ(i(cid:48), j + 1)
with OPT(i, 0) = 0 for each i. Hence, the optimal value of (10), as well as the optimal j1, j2, . . . , jm,
can be computed in O(n2m) time. We conclude the result above as the following theorem.
Theorem 2. If the costs obey increasing differences, there is an O(n2m) algorithm solving the
Multiple Agents Contract Problem.
5 Real Number Actions
In previous sections, we considered the Multiple Agents Contract Problem with discrete actions
(DA). A natural generalization is to consider the problem where each agent can choose to produce
10
an arbitrary reward in [0, 1]. We call this generalization Multiple Agents Contract Problem with
Real Number Actions (RNA), and formalize it as follows.
There is a principal and n agents. Each agent chooses to produce a reward x ∈ [0, 1] for the principal.
To take such an action, each agent has a cost which may differ from each other. We define ci(x) ≥ 0
as the cost for agent i to produce a reward x. We assume without loss of generality that ci(0) = 0
for all i, which means it is free for each agent to choose to produce nothing. To incentivize these
agents to produce rewards, the principal specifies a payment function t(x): each agent taking this
action will earn a payment t(x). The utility for agent i to produce x is t(x) − ci(x). Agents are
self-interested, meaning each agent will produce a reward that maximizes her utility. We assume
agents tie-break in favor of the principal. The payoff of the principal is the sum of the rewards
produced by these agents minus the payments given to the agents, i.e., if agent i produces a reward
i=1(xi − t(xi)). Our goal is to design the payment function to
xi, the payoff of the principal is(cid:80)n
maximize the payoff of the principal.
Note in this paper, the functions t and ci's are not necessarily continuous. To guarantee every agent
has an optimal action we only concern the payment function t where for all i, t(x) − ci(x) and
x − t(x) (in case of tie-breaking) are able to attain their maximums on [0, 1].
5.1 Hardness
We first show that this generalization is harder than our original problem by a reduction from DA to
RNA. Given an instance of DA, we can construct an instance of RNA by letting
0,
ci,1+M
ρm+mM ,
...
ci,j +jM
ρm+mM ,
...
ci,m+mM
ρm+mM ,
ci(x) =
if x = 0,
if 0 < x ≤ ρ1+M
ρm+mM ,
if ρj−1+(j−1)M
ρm+mM < x ≤ ρj +jM
ρm+mM ,
if ρm−1+(m−1)M
ρm+mM
< x ≤ 1,
for each i, where M is a large enough number8. We will show how to construct an optimal payment
profile of the DA instance from an optimal payment function of the RNA instance. For convenience,
we define zj = (ρj + jM )/(ρm + mM ) and z0 = 0.
Given an optimal payment function t(x) of the RNA instance, suppose agent i chooses to produce xi
and define ji such that zji−1 < xi ≤ zji (if xi = 0, then ji = 0). Now consider a fixed i. If xi < zji,
we adjust the value of t(x) at x = zji to t(xi). Before this adjustment, agent i produces xi, and after
this adjustment, agent i has the same utility to produce zji as to produce xi, so agent i will produce
zji after the adjustment (recall the agent tie-breaks in favor of the principal), which increases the
payoff of the principal. On the other hand, for any other agent i(cid:48), t(xi) − ci(cid:48)(zji) ≤ t(xi) − ci(cid:48)(xi)
(since the cost function is weakly increasing), which means the utility of producing zji after the
adjustment does not exceed that of producing xi. Hence, for any agent except i, changing her
produced value to zji due to the adjustment does not decrease the payoff of the principal (recall
8It is sufficient to choose M = maxi,j{ci,j, ρj} + 1.
11
again that the agents tie-break in favor of the principal). As a result, the payoff of the principal is
increased by this adjustment, which contradicts to the fact that t(x) is optimal. Therefore, we can
assume xi = zji. The payoff of the principal under the payment function t(x) in the RNA instance
is pRNA =(cid:80)n
i=1(zji − t(zji)).
Now we construct a payment profile (t1, t2, . . . , tm) of the DA instance where tj = t(zj)(ρm +
mM ) − jM. Under this payment profile, for each agent i and each j, the utility of agent i to
take action j is tj − ci,j, which is exactly (ρm + mM ) times the utility of agent i to produce zj
under the payment function t(x) in the RNA instance. Also, agent i brings a payoff of ρj − tj to
the principal by taking action j, which is exactly (ρm + mM ) times the payoff of the principal
brought by agent i by producing zj under the payment function t(x) in the RNA instance. Since
agent i produces zji under payment function t(x) in the RNA instance, she will take action ji under
payment profile (t1, t2, . . . , tm) in the DA instance. The payoff of the principal under the payment
profile (t1, t2, . . . , tm) in the DA instance is pDA =(cid:80)n
DA =(cid:80)n
To show the payment profile (t1, t2, . . . , tm) is optimal, we compare it to another arbitrary payment
profile (t(cid:48)
m),
then the payoff of the principal under the payment profile (t(cid:48)
m) in the DA instance is
p(cid:48)
Let
i=1(ρji − tji) = (ρm + mM )pRNA.
m). Suppose agent i takes action j(cid:48)
i under the payment profile (t(cid:48)
1, t(cid:48)
2, . . . , t(cid:48)
i=1(ρji − t(cid:48)
1, t(cid:48)
2, . . . , t(cid:48)
1, t(cid:48)
2, . . . , t(cid:48)
).
ji
t(cid:48)(x) =
0,
t(cid:48)
1+M
ρm+mM ,
...
t(cid:48)
j +jM
ρm+mM ,
...
t(cid:48)
m+mM
ρm+mM ,
if x = 0,
if 0 < x ≤ z1,
if zj−1 < x ≤ zj,
if zm−1 < x ≤ zm = 1.
1, t(cid:48)
2, . . . , t(cid:48)
We can see for all i, x − t(cid:48)(x) and t(cid:48)(x) − ci(x) are able to attain their maximum on [0, 1], so
t(cid:48)(x) is a valid payment function. Under this payment function, agent i has the same utility for
producing a reward on (zj, zj+1], thus she will produce zj for some j in favor of the principal.
Observe, again, that under the payment profile (t(cid:48)
m), for each agent i and each j, the
j − ci,j, which is exactly (ρm + mM ) times the utility of
utility of agent i to take action j is t(cid:48)
agent i to produce zj under the payment function t(cid:48)(x) in the RNA instance. Also, agent i brings
a payoff of ρj − t(cid:48)
j to the principal by taking action j, which is exactly (ρm + mM ) times the
payoff of the principal brought by agent i by producing zj under the payment function t(cid:48)(x) in
under the payment function t(cid:48)(x) in the RNA
the RNA instance. Hence, agent i will produce zj(cid:48)
instance. The payoff of the principal under the payment function t(cid:48)(x) in the RNA instance is
p(cid:48)
Hence, p(cid:48)
indeed an optimal payoff profile of the DA instance.
DA/(ρm + mM ).
RNA ≤ (ρm + mM )pRNA = pDA, which means (t1, t2, . . . , tm) is
i=1(zji − t(cid:48)(zji)) = p(cid:48)
DA = (ρm + mM )p(cid:48)
RNA =(cid:80)n
i
12
5.2 An Approximate Contract
Knowing the RNA problem is hard, we are going to design an approximate contract. We assume for
all i, x− ci(x) is able to attain its maximum on [0, 1]. Let xi ∈ arg maxx∈[0,1](x− ci(x)) (if there are
is achieving the maximum value, we arbitrarily choose one), yi = maxx∈[0,1](x − ci(x)),
multiple x(cid:48)
we have immediately
(16)
Let
ti(x) =
0,
x − yi,
if 0 ≤ x ≤ yi,
if yi < x ≤ 1.
yi = xi − ci(xi) ≤ xi.
(cid:40)
We assume without loss of generality that y1 ≤ y2 ≤ ··· ≤ yn. We first show that ti(x) is a valid
payment function, i.e. for all i(cid:48), x − ti(x) and ti(x) − ci(cid:48)(x) are able to attain their maximum on
[0, 1]. The former is trivial. For ti(x) − ci(cid:48)(x), if 0 ≤ x ≤ yi, then ti(x) − ci(cid:48)(x) = −ci(cid:48)(x) ≤ 0; if
yi < x ≤ 1, then ti(x) − ci(cid:48)(x) = x − yi − ci(cid:48)(x) ≤ yi(cid:48) − yi, so ti(x) − ci(cid:48)(x) ≤ max{0, yi(cid:48) − yi}.
In addition, ti(0) − ci(cid:48)(0) = 0 and ti(xi(cid:48)) − ci(cid:48)(xi(cid:48)) ≥ xi(cid:48) − yi − ci(cid:48)(xi(cid:48)) = yi(cid:48) − yi. This means the
maximum value of ti(x) − ci(cid:48)(x) is max{0, yi(cid:48) − yi}, and is achievable at x = 0 or x = xi(cid:48). Hence,
ti(x) is indeed a valid payment function.
Note the argument above also shows that for any i(cid:48) ≥ i, ti(x) − ci(cid:48)(x) attains its maximum at
x = xi(cid:48). By (16), we have xi(cid:48) ≥ yi(cid:48) ≥ yi, so if agent i(cid:48) chooses to produce xi(cid:48), she brings a payoff
of xi(cid:48) − ti(xi(cid:48)) = xi(cid:48) − (xi(cid:48) − yi) = yi to the principal. Recall that the agents tie-break in favor
of the principal, agent i(cid:48) brings a payoff of at least yi to the principal. Hence, under the payment
function ti(x), the payoff of the principal is at least (n − i + 1)yi. Let i∗ ∈ arg maxi(n − i + 1)yi,
then we have for all i,
On the other hand, let OPT denote the optimal payoff of the principal. Since agent i brings a payoff
i=1 yi. Hence,
of at most maxx∈[0,1](x − ci(x)) = yi to the principal, we have OPT ≤(cid:80)n
This means the payment function ti∗(x) is an(cid:80)n
OPT ≤ n(cid:88)
n − i + 1
i=1
.
yi ≤ (n − i∗ + 1)yi∗
n − i + 1
n(cid:88)
yi ≤ (n − i∗ + 1)yi∗
1
.
i=1
O(log n)-approximate solution.
In conclusion, we have the following algorithm.
i=1(1/(n − i + 1))-approximate solution, i.e. an
1. For any i, find yi = maxx∈[0,1](x − ci(x)) and sort them such that y1 ≥ y2 ≥ ··· ≥ yn.
2. Let i∗ ∈ arg max1≤i≤n(n − i∗ + 1)yi∗.
3. Output the payment function
t(x) =
0,
x − yi∗,
if 0 ≤ x ≤ yi∗,
if yi∗ < x ≤ 1.
(cid:40)
13
References
[1] Bengt Holmstrom and Paul Milgrom. Multitask principal-agent analyses: Incentive contracts,
asset ownership, and job design. JL Econ. & Org., 7:24, 1991.
[2] Bengt Holmstrom and Paul Milgrom. Aggregation and linearity in the provision of intertem-
poral incentives. Econometrica: Journal of the Econometric Society, pages 303 -- 328, 1987.
[3] Mengjing Chen, Pingzhong Tang, Zihe Wang, Shenke Xiao, and Xiwang Yang. Optimal
mechanisms with budget for user generated contents. arXiv preprint arXiv:1907.04740, 2019.
[4] Sanford J Grossman and Oliver D Hart. An analysis of the principal-agent problem. In
Foundations of Insurance Economics, pages 302 -- 340. Springer, 1992.
[5] Patrick Bolton, Mathias Dewatripont, et al. Contract theory. MIT press, 2005.
[6] Mark Armstrong and John Vickers. A model of delegated project choice. Econometrica, 78
(1):213 -- 244, 2010.
[7] Jon Kleinberg and Robert Kleinberg. Delegated search approximates efficient search. In
Proceedings of the 2018 ACM Conference on Economics and Computation, pages 287 -- 302.
ACM, 2018.
[8] Moshe Babaioff, Michal Feldman, and Noam Nisan. Combinatorial agency. In Proceedings of
the 7th ACM conference on Electronic commerce, pages 18 -- 28. ACM, 2006.
[9] Moshe Babaioff, Michal Feldman, and Noam Nisan. Mixed strategies in combinatorial agency.
In International Workshop on Internet and Network Economics, pages 353 -- 364. Springer,
2006.
[10] Moshe Babaioff and Eyal Winter. Contract complexity. In EC, page 911, 2014.
[11] Ron Lavi and Elisheva Shamash. Principal-agent vcg contracts. In Proceedings of the 2019
ACM Conference on Economics and Computation, pages 783 -- 783. ACM, 2019.
[12] R Preston McAfee and John McMillan. Bidding for contracts: a principal-agent analysis. The
RAND Journal of Economics, pages 326 -- 338, 1986.
[13] Navid Azizan, Yu Su, Krishnamurthy Dvijotham, and Adam Wierman. Optimal pricing in
markets with non-convex costs. In Proceedings of the 2019 ACM Conference on Economics
and Computation, pages 595 -- 595. ACM, 2019.
14
|
1803.03059 | 1 | 1803 | 2018-03-08T12:14:49 | Hierarchical Heuristic Learning towards Effcient Norm Emergence | [
"cs.MA"
] | Social norms serve as an important mechanism to regulate the behaviors of agents and to facilitate coordination among them in multiagent systems. One important research question is how a norm can rapidly emerge through repeated local interaction within an agent society under different environments when their coordination space becomes large. To address this problem, we propose a Hierarchically Heuristic Learning Strategy (HHLS) under the hierarchical social learning framework, in which subordinate agents report their information to their supervisors, while supervisors can generate instructions (rules and suggestions) based on the information collected from their subordinates. Subordinate agents heuristically update their strategies based on both their own experience and the instructions from their supervisors. Extensive experiment evaluations show that HHLS can support the emergence of desirable social norms more efficiently and is applicable in a much wider range of multiagent interaction scenarios compared with previous work. We also investigate the effectiveness of HHLS by separating out the different components of the HHLS and evaluating the relative importance of those components. The influence of key related factors (e.g., hierarchical factors, non-hierarchical factors, fixed-strategy agents) are investigated as well. | cs.MA | cs | Noname manuscript No.
(will be inserted by the editor)
Hierarchical Heuristic Learning towards Efficient
Norm Emergence
Tianpei Yang · Jianye Hao · Zhaopeng
Meng · Sandip Sen · Sheng Jin
Received: date / Accepted: date
Abstract Social norms serve as an important mechanism to regulate the be-
haviors of agents and to facilitate coordination among them in multiagent
systems. One important research question is how a norm can rapidly emerge
through repeated local interaction within an agent society under different envi-
ronments when their coordination space becomes large. To address this prob-
lem, we propose a Hierarchically Heuristic Learning Strategy (HHLS) under
the hierarchical social learning framework, in which subordinate agents report
their information to their supervisors, while supervisors can generate instruc-
tions (rules and suggestions) based on the information collected from their
subordinates. Subordinate agents heuristically update their strategies based
on both their own experience and the instructions from their supervisors. Ex-
tensive experiment evaluations show that HHLS can support the emergence of
desirable social norms more efficiently and is applicable in a much wider range
of multiagent interaction scenarios compared with previous work. We also in-
vestigate the effectiveness of HHLS by separating out the different components
of the HHLS and evaluating the relative importance of those components. The
Tianpei Yang · Jianye Hao ( ) · Sheng Jin
School of Computer Software, Tianjin University, China
E-mail: [email protected]
Tianpei Yang
E-mail: [email protected]
Zhaopeng Meng
Tianjin University of Traditional Chinese Medicine, China
E-mail: [email protected]
Sandip Sen
University of Tulsa, USA
E-mail: [email protected]
Sheng Jin
E-mail: [email protected]
8
1
0
2
r
a
M
8
]
A
M
.
s
c
[
1
v
9
5
0
3
0
.
3
0
8
1
:
v
i
X
r
a
2
Tianpei Yang et al.
influence of key related factors (e.g., hierarchical factors, non-hierarchical fac-
tors, fixed-strategy agents) are investigated as well.
Keywords Norm emergence · Multiagent social learning · Multiagent
hierarchical learning
1 Introduction
Social norms play an important role in regulating agents' behaviors to ensure
coordination among agents and functioning of agent societies. One commonly
adopted characterization of a norm is to model it as a consistent equilibrium
that all agents follow during interactions where multiple equivalent equilibria
coexist [34]. How social norms can emerge efficiently in agent societies is a key
research problem in the area of normative multiagent systems.
There exist two major approaches for addressing norm emergence problem:
the top-down [2,26] approach and the bottom-up approach [24,38]. The former
approach investigates how to efficiently synthesize a norm for all agents be-
forehand, while the latter one investigates how a norm can emerge through re-
peated local interactions. In distributed multiagent interaction environments,
it is usually difficult to design a norm a priori, and before agents interact,
since a centralized controller might not exist and due to the fact that the
optimal norm may vary frequently as the environment dynamically changes
and therefore, the bottom-up approach of emergent norms via local interac-
tions promises to be more suitable for such kinds of distributed and dynamic
environments.
Until now, significant efforts have been devoted to investigating norm emer-
gence problem from the bottom-up approaches [3,11,38,4,17,12,39,13,19,23,
24,25,28,29,30,35,36,37]. Sen and Airiau [24] investigated the norm emer-
gence problem in a population of agents within randomly connected networks
where each agent is equipped with certain multiagent learning algorithms. The
local interaction among each pair of agents is modeled as two-player normal-
form games, and a norm corresponds to one consistent Nash equilibrium of the
coordination/anti-coordination game. Later, a number of papers [19,30,22,3]
subsequently extended this work by using more realistic and complex networks
(e.g., small-world network and scale-free network) to model the diverse inter-
action patterns among agents. Additionally, different learning strategies and
mechanisms have been proposed to better facilitate norm emergence among
agents within different interaction environments [39,21,17,4,12,36].
Most of the previous works only focus on games of relatively small size.
This simplification does not accurately reflect the practical interaction sce-
narios where the action space of agents can be quite large. With the increase
of action space, most of the existing approaches usually result in very slow
norm emergence or even fail to converge. Recently, Yu et al. [35] proposed a
hierarchical learning strategy to improve the norm emergence rate for very
large action space problems. However, this work only considers the case in
which a norm corresponds to a Nash equilibrium where the interacting agents
Hierarchical Heuristic Learning towards Efficient Norm Emergence
3
Table 1 An example of coordination game.
Agent 2's actions
a
1
-1
b
-1
1
Agent 1's
actions
a
b
Table 2 An example of anti-coordination game.
Agent 2's actions
a
-1
1
b
1
-1
Agent 1's
actions
a
b
select the same action. This usually can be modelled as a two-player n-action
coordination game (CG). One simple example with n=2 is shown in Table 1.
However, in realistic interaction scenarios, a norm may correspond to a Nash
equilibrium where all agents select different actions. One notable example is
considering two drivers arriving at a road intersection from two neighboring
roads. To avoid collision, one possible norm is "yield to the left", i.e., waiting
for the car on the left-hand side to go through the intersection first. This kind
of scenario can naturally be modeled as an anti-coordination game (ACG), as
shown in Table 2, which exist two different norms, i.e., (a, b) and (b, a).
Furthermore, agents may be faced with the challenge of high mis-coordination
cost and stochasticity of the environment. One representative example is shown
in Table 3, which we call fully stochastic coordination game with high penalty
(FSCGHP). In this game, there exist two optimal Nash equilibria marked with
red each of which corresponds to a norm, and one suboptimal Nash equilib-
rium marked with blue. Two major challenges coexist in this game: agents are
vulnerable to converge to the suboptimal Nash equilibrium due to the high
penalty when agents mis-coordinate on the outcomes; agents need to effectively
distinguish between the stochasticity of the environment and the exploration
of other learners. It is not clear, a priori, how a population of agents can ef-
ficiently evolve towards a consistent norm, given the large space of possible
norms in such challenging environments.
To answer this question, in this paper we propose a novel Hierarchical
Heuristic Learning Strategy (HHLS) under the hierarchical social learning
framework to facilitate rapid norm emergence in agent societies. In the hierar-
chical social learning framework, the agent society is separated into a number
of clusters, each of which consists of several subordinate agents. Each cluster's
strategies are monitored and guided by one supervisor agent. In each round,
each supervisor agent collects the interaction information of its subordinate
agents and generates instructions in the forms of behavioral rules and sugges-
4
Tianpei Yang et al.
Table 3 Fully stochastic coordination game with high penalty.
1's payoff
2's payoff
Agent 1's
actions
a
b
c
Agent 2's actions
a
b
c
-5/5
-20/-40
8/12
-5/5
0/14
-20/-40
-5/5
-5/5
8/12
tions for its subordinates. On the other hand, each subordinate agent, apart
from learning from its local interaction, also adjusts its strategy based on the
instructions from its supervisor. The main feature of the proposed framework
is that through hierarchical supervision among agents, an effective compromise
solution can be generated to effectively balance distributed interactions and
centralized control towards efficient and robust norm emergence.
We evaluate the performance of HHLS under a wide range of games and
experimental results show that HHLS can facilitate rapid emergence of norms
compared with the state-of-the-art approaches. We investigate the influence of
hierarchical (e.g., the cluster size, different grouping mechanisms, the number
of disabling supervisors) and non-hierarchical factors (e.g., the population size,
the size of action space, different network topologies, the neighborhood size) on
norm emergence. Lastly, we investigate the influence of fixed-strategy agents
on norm emergence (e.g., initial and late intervention, different placement
strategies) and norm emergence in isolated sub-networks.
The remainder of the paper is organized as follows. Section 2 discusses
related work. Section 3 introduces the hierarchical social learning framework
and the heuristic learning strategy. Section 4 presents experimental evaluation
results compared with two representative state-of-the-art approaches. Finally,
Section 5 concludes the paper and points out future directions.
2 Related Work
2.1 Norm Emergence
Norm emergence problem has received a wide range of attention in Multiagent
Systems (MASs) literature. Shoham and Tennenholtz [25] first investigated the
norm emergence problem in an agent society based on a simple and natural
strategy - the Highest Cumulative Reward (HCR). In this study, they showed
that HCR achieved high efficiency on social conventions in a class of games. Sen
and Airiau [24] investigated the norm emergence problem in MASs where each
agent is equipped with certain existing multiagent learning algorithms. They
firstly proposed the model of social learning, where each agent learns from
repeated interactions with multiple agents in a given scenario. In this study, the
local interaction among each pair of agents is modeled as two-player normal-
Hierarchical Heuristic Learning towards Efficient Norm Emergence
5
form games, and a norm corresponds to one consistent Nash equilibrium of the
game. Later, a number of papers [19,30,23,3] extended this work by leveraging
more realistic and complex networks (e.g., small-world network and scale-free
network) to model the interaction patterns among agents and evaluated the
influence of heterogeneous agent systems and space-constrained interactions
on norm emergence. Savarimuthu [21] recapped the existing mechanisms on
the multiagent-based emergence, and investigated the role of three proactive
learning methods in accelerating norm emergence. The influence of liars on
norm emergence is also considered and simulation results showed that norm
emergence can still be sustained in the presence of liars.
Villatoro et al. [30] proposed a reward learning mechanism based on in-
teraction history. In this study, they investigated the influence of different
network topologies and the effects of memory of past activities on convention
emergence. Later, they [28,29] introduced two rules (i.e., re-wiring links with
neighbors and observation) to overcome the suboptimal norm problems. They
investigated the influence of Self-Reinforcing Substructure (SRS) in the net-
work on impeding full convergence towards society-wide norms, which usually
results in reduced convergence rates. Hao et al. [11] investigated the prob-
lem of coordinating towards optimal joint actions in cooperative games under
the social learning framework by introducing two types of learners (IALs and
JALs). Yu et al. [38] proposed a novel collective learning framework to inves-
tigate the influence of agent local collective behaviors on norm emergence in
different scenarios and defined two strategies (collective learning-l and collec-
tive learning-g) to promote the emergence of norms where agents are allowed
to make collective decisions within networked societies. Later, Hao et al. [12]
proposed two learning strategies under the collective learning framework: col-
lective learning EV-l and collective learning EV-g to address the problem of
high mis-coordination cost and stochasticity in complex and dynamic inter-
action scenarios. More recently, Yu et al. [36] proposed an adaptive learning
framework for efficient norm emergence. Later, Yu et al. [37] proposed a novel
adaptive learning to facilitate consensus formation among agents, in order to
efficiently establish a consistent social norm in agent societies. However, all the
aforementioned works usually focus on relatively small-size games, and do not
address the issue of efficient norm emergence in large action space problems.
Hierarchical learning framework, as a promising solution to accelerate co-
ordination among agents, has been studied in different multiagent applications
(e.g., package routing [39], traffic control [1], p2p network [18] and smart-grid
[33]). For example, Zhang et al. [39,40] studied the package routing prob-
lem and proposed a multi-level organizational structure for automated su-
pervision and a communication protocol for information exchange between
higher-level supervising agents and their subordinate agents. Simulation shows
that the organization-based control framework can significantly increase the
overall package routing efficiency compared to traditional non-hierarchical ap-
proaches. Abdoos et al. [1] proposed a multi-layer organizational controlling
framework to model large traffic networks to improve the coordination between
different car agents and improve the overall traffic efficiency. Recently, Yu et
6
Tianpei Yang et al.
al. [35] proposed a hierarchical learning framework to study the norm emer-
gence problem. In this study, they proposed a two-level hierarchical framework.
Agents in the lower level interact with each other and report information to
their supervisors in the higher level, while agents in the higher level called
supervisors pass down guidance to the lower level. Agents in the lower level
follow guidance in policy update. However, their framework is designed for
coordination game only, where agents only need to coordinate to select the
same action for effective norm emergence.
2.2 Fixed-strategy Agents on Norm Emergence
Given the existence of multiple norms, agents usually do not have preference
over different norms. It has been found that fixed-strategy agents could play a
critical role in influencing the direction of norm emergence and have received
wide range of attention in previous work [24,16,6,10]. Fixed-strategy agents
are those who always select the same action regardless of its efficiency or others'
choices. Previous work has shown that inserting relatively small numbers of
fixed-strategy agents can significantly influence much larger populations when
placed in networked social learning framework.
Sen and Airiau [24,3] firstly investigated the effect of fixed-strategy agents
in affecting norm adoption in social learning framework. Griffiths and Anand
[10] proposed a social learning model where each agent has a fixed length of
memory recording the most recent actions it has selected. They demonstrated
how fixed-strategy agents can manipulate emergence and evaluated strategies
for inserting fixed-strategy agents using placement heuristics such as degree
and betweenness centrality. Marchant et al. [16] investigated the influence of
fixed-strategy agents on norm emergence in a dynamic network. They defined
a new heuristic and proposed late intervention of fixed-strategy agents, where
the whole system has already emerged an existing norm. Experimental results
show that placing small numbers of fixed-strategy agents can disrupt the al-
ready established norm and the whole population eventually converged to the
norm adopted by the fixed-strategy agents. Tiwari et al. [27] investigated the
effect of leader placement in simulated robotic swarms. Their experimental
results showed that the leader placement strategy determines the time it takes
for the swarm to converge: leaders placed in the middle or periphery of the
swarm are better in maneuvering the swarm than leaders placed in the front.
Franks et al. [6] first introduced Influencer Agents (IAs) as a mechanism
to manipulate norm emergence direction. The IAs can be considered as agents
with uniform levels strategies and goals to aid the emergence of high quality
norms in domains characterized by heterogeneous ownership and uniform levels
of agent authority. IAs are those agents with specific norms (fixed strategies),
such that the whole population, through their rational selection of actions, is
guided towards the adoption of the specific norms. They evaluated the influ-
ence of IAs in the language coordination problem domain. Results showed that
small number of IAs can effectively manipulate the emergence of high-quality
Hierarchical Heuristic Learning towards Efficient Norm Emergence
7
conventions. On the other hand, they proved the fragility of convention emer-
gence in the presence of malicious or faulty agents that attempted to propagate
low quality conventions, and confirmed the importance of network topologies
in norm adoption. Franks et al. [5] later proposed a methodology for learning
the influence of the placement of agents in a network and evaluated their ap-
proach in the context of the language coordination domain. They investigated
the influence of IAs in a coordination game with heterogeneous agent learning
mechanisms. They used several topological metrics to measure the influence
of fixed-strategy agents on norm emergence. Experimental results showed that
when placing agents following their methodology, agents can gain much higher
influence power than random placement.
Another notable work related with fixed-strategy agents is by Genter et al.
[9] which investigates the influence of placing fixed-strategy agents in a flock
of agents. The group flocking behavior emerges from simple local control rules,
through which each individual agent adjusts its own trajectory based on those
of its neighbors. They defined several methodologies for placing the influencing
agents into the flock. Experimental results showed that the graph approach
outperforms other approaches when initially placing influencing agents any-
where into the flock and the grid approach performs best when the agents must
travel to their desired positions after being initially placed outside the flock.
Later, Genter and Stone [7] extended their work on considering the placement
heuristic of influencing agents in a flock and how influencing agents should
behave in order to join a flock in motion (i.e., the joining case). They showed
that the placement heuristics of fixed-strategy agents which work well for ini-
tial placement also work well if fixed-strategy agents can place themselves in
an incoming flock within the flight path. Recently, they [8] considered the
situation where robot birds are deployed to influence flocks in nature. They
proposed a hover approach for robot birds to use when joining and leaving a
flock. Finally, they summarized and pointed out the main drawback of their
hover approach: the robot birds are required to hover at desired positions. This
may cause hovering problems because robot birds would be unable to hover
when they are recognized by natural birds as 'one of their own'.
3 Hierarchical Social Learning Framework
3.1 Framework Overview
We consider a population of N agents where each agent is connected by the
underlying network topology. In each round, each agent interacts with one ran-
domly selected agent from its neighborhood. An agent's neighborhood consists
of all agents that are physically connected. We model the interaction between
each pair of agents as a normal-form game. At the beginning of each interac-
tion, one agent is randomly assigned as the row player and the other as the
column player. We assume that each agent is unaware of opponent's action
and payoff information and can only have access to its own action and pay-
8
Tianpei Yang et al.
Fig. 1 An example of the two-level hierarchical network.
off information during interaction. The population of agents are divided into
multiple levels, and the agents in each level supervise the behaviors of agents
in its neighboring lower level.
One example of two-level hierarchical network is shown in Fig 1. Each su-
pervisor agent i in the higher level is in charge of a group of subordinate agents
(denoted as sub(i)) in the bottom level surrounded by dashes lines. For each
subordinate agent j, its supervisor agent is denoted as sup(j). For subordi-
nates, the topological connections between them are determined by the original
network topology; for supervisors, a pair of supervisors are neighboring agents
if the corresponding group of subordinates they supervise are connected. Note
that each supervisor agent is also allowed to communicate with its neighboring
supervisor agents. The way of dividing subordinates into groups and choosing
each group's supervisor will be discussed in details in Section 3.2.
The interaction protocol of agents under the hierarchical social learning
framework is summarized in Algorithm 1. In each round, each agent is paired
with another agent randomly selected from its neighborhood (Line 3), and
their roles are randomly assigned (Line 4). Each agent then chooses an action
following its learning strategy (Line 5), and then updates its strategy based
on its current-round feedback (Line 6). After that, each subordinate agent
reports its action and reward information to its supervisor (Line 7-9). At the
end of each round, each supervisor collects all subordinate agents' information,
generates and issues the instructions to its subordinate agents (Line 12-14).
Finally, each subordinate agent updates its strategy based on the instructions
accordingly (Line 15-17).
3.2 Grouping Mechanism
As described in Section 3.1, the whole population is divided into two levels.
Each agent in the higher level supervises a group of subordinate agents in the
lower level. Each supervisor has a dual role in the network. One role is as
a supervisor of a group of subordinates and the other role is a subordinate
agent which is also allowed to interact with one of its neighbors. How to group
Hierarchical Heuristic Learning towards Efficient Norm Emergence
9
for each agent i ∈ N do
Reporting its experience hsi, ai, rii to sup(i);
Randomly choose a neighboring agent j to interact;
Assign distinct roles randomly i → state si, j → state sj
Select actions ai and aj and get rewards ri and rj;
Update its strategy based on hsi, ai, rii.
if agent i is a subordinate agent then
Algorithm 1 The interaction protocol of the hierarchical framework.
1: for each round of interaction do
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18: end for
Generate instructions based on the information from sub(j);
Provide the instructions to sub(j);
end if
end for
for each supervisor agent j do
end for
for each subordinate agent k do
Update its strategy based on the instructions from sup(k);
end for
subordinate agents and how to choose supervisors are two key aspects that
can influence the performance of our algorithm. One natural way is to allow
each supervisor to supervise its direct neighbors, which, however, may not be
effective due to the sparsity of network topologies. To this end, we propose a
number of grouping mechanisms as follows.
– Random Grouping
We divide subordinate agents equally into k groups randomly. The super-
visors are also selected randomly from each group. This grouping mechanism
is used in our previous work [32].
– Degree Grouping
Given any network topology GhV, Ei, each agent i corresponds to a node
vi ∈ V , N (vi) denotes the set of neighbors of the agent i, and the degree of
agent i is N (vi) . We divided subordinate agents into different groups based
on their degrees. Agents with the same degree are allocated into the same
group. Also, the supervisors are randomly selected from each group. Note that
degree grouping is not an equally division.
– K-means Grouping
We introduce the core idea of K-means clustering algorithm into our group-
ing mechanism [15]. In a network GhV, Ei, we define the distance Disspl(vi, vj)
between two agents i, j in such network as the shortest path between them
(i.e. the minimum number of edges connecting these two agents). We first
randomly select k agents as the k centering points. Each of the remaining
agents is assigned to a group if the distance between it and this center point
is minimum over the k centering points (random assignment in case of a tie).
The centering agent of each group is recomputed after all agents have been
10
Tianpei Yang et al.
assigned to one group. This process repeats until each agent no longer changes
its group. The final centering agent of each group is selected as the supervisor
of the group. Note that K-means grouping is also not an equal division and
it may lead to extreme cases where some groups have too small number of
agents, which causes undesirable results.
To address this issue, we modify the original K-means grouping to suit
our problem domain in which all groups make some adjustments after the
original K-means algorithm terminates. We adopt a threshold as the maximum
distance between the centering agent and other agents in each group. Agents
will be reassigned to other groups if the distance between it and the centering
point is larger than the given threshold. Moreover, to prevent a group from
being too small, a group will no longer compete for agents if the number of
agents in the group size exceeds a maximum threshold.
– Kernighan-Lin Grouping
The Kernighan-Lin algorithm is a heuristic algorithm for the graph par-
titioning problem [14]. Subordinate agents are first divided into two equal
groups randomly. The sum of the costs of edges between node vi ∈ V and
other nodes within the same group is denoted as the internal cost of node
vi ∈ V , and the sum of the costs of edges between node vi ∈ V and nodes
in opposite group is denoted as the external cost of node vi ∈ V . Each agent
makes an exchange from one group to the other if the exchange reduces the
internal cost and increases the external cost. Kernighan-Lin Grouping ends if
no further exchanges are feasible. Each group will be divided into two equal
groups to repeat the process described above if we need two more groups. An
agent is selected as the supervisor if its degree is the highest in the group.
3.3 Information Exchange between Supervisors and Subordinates
In the hierarchical social learning framework, subordinate agents send their
feedback information to their corresponding supervisors, while supervisors pass
down instructions to their corresponding subordinate agents. In details, each
subordinate agent i reports its current-round interaction experience hsi, ai, rii
to its supervisor sup(i). For supervisors, we distinguish two different forms
of instructions that they can provide to their subordinates: suggestion and
rule [39]. Intuitively, a rule is a hard constraint that specifies an action that
subordinate agents are forbidden to select under certain state in the next
round; in contrast, a suggestion is a soft constraint which indirectly affects
the strategies of the subordinate agents next round.
A set F of rules consists of all the forbidden actions for subordinates under
different states. Formally we have,
F = {hs, aia ∈ A, s ∈ S}
(1)
where each element hs, ai denotes that action a is forbidden to take in state s;
A and S are the action space and state space of the subordinates respectively.
Hierarchical Heuristic Learning towards Efficient Norm Emergence
11
A set D of suggestions specifies the recommendation degrees for different
state-action pairs, which can be formally represented as follows,
D = {hs, a, d(s, a)ia ∈ A, s ∈ S}
(2)
where d(s, a) is the recommendation degree of action a under state s. Given
an action and a state hs, ai, if d(s, a) < 0, it indicates that action a is not
recommended for selection in state s; if d(s, a) > 0, it indicates subordinate
agents are encouraged to select action a when they are in state s. The way of
determining rules and suggestions will be covered in detail in Section 3.4.2.
f req(s, a) =
{hsk, ak, rki hsk, ak, rki ∈ RepInf, sk = s, ak = a, rk = rmax(s, a)}
{hsk, ak, rki hsk, ak, rki ∈ RepInf, sk = s, ak = a}
(3)
3.4 Learning Strategy
In this section, we first present the learning strategy of supervisors and how
the rules and suggestions are generated in Section 3.4.1 and 3.4.2 respectively.
Following that, we describe the learning strategy of subordinate agents and
how they utilize the instructions from supervisors in Section 3.4.3. Without
loss of generality, let us assume that there is a set S of supervisors, and each
supervisor i ∈ S, it supervises the set sub(i) of subordinate agents. Each
subordinate agent j has a set neigh(j) of neighbors, and each supervisor agent
i communicates with a set com(i) of other supervisors.
3.4.1 Supervisor's Strategy
We propose that each supervisor i holds a Q-value Qi(s, a) for each action a
under each state s (row or column player). Let us denote the set of informa-
tion from its subordinates as RepInfi = {hsk, ak, rki k ∈ sub(i)}. For each
piece of information hs, a, ri ∈ RepInfi, supervisor agent i updates its Q-value
following the optimistic assumption shown in Equation (4),
Qi(s, a) = (1 − αi) ∗ Qi(s, a) + αi ∗ r
(4)
where αi is its learning rate reflecting its updating degree between using the
past experience and using the current round information.
After that, supervisor i further updates its Q-values based on optimistic
assumption and the frequency information of each action similar to the FMQ
heuristic [13]. Formally we have,
F M Qi(s, a) = Qi(s, a) + f req(s, a) ∗ rmax(s, a) ∗ C
(5)
where rmax(s, a) is the max reward of each action a, f req(s, a) is the frequency
12
Tianpei Yang et al.
of receiving the reward of rmax(s, a) by choosing action a under state s and
C is a weighting factor.
The value of rmax(s, a) is obtained from the reported information of its
subordinate agents sub(i). Specifically, The value of rmax(s, a) is computed
as the maximum reward that all of its subordinates receives under state s by
choosing action a in the current round experience. Formally we have,
R(s, a) = {rk hs, a, rki ∈ Ref Inf }
rmax(s, a) = max{R(s, a)}
(6)
(7)
The frequency information f req(s, a) is calculated as the empirical proba-
bility of receiving the maximum reward rmax(s, a) under state s when action
a is selected based on the reported information RepInf collected from the
subordinates which is shown in Equation (3).
After updating the strategy based on the information collected from its
subordinates, we also allow each supervisor to learn from its neighboring peers
(supervisors). Specifically, each supervisor communicates with a neighboring
supervisor selected based on the neighboring degree between them and im-
itates the neighbor's strategy. The neighboring degree between a supervisor
and its neighbor is related to the correlation of two groups of subordinates
they supervise. A higher neighboring degree means the greater probability of
choosing this neighboring supervisor. The motivation of imitating peers comes
from evolutionary game theory [31], which provides a powerful methodology
to model how strategies evolve over time based on their relative performance.
One of the widely used imitation rules is the proportional imitation [20], which
is adopted here as shown in Equation (8),
p =
1
1 + e−β∗(F M Qj (s,a)−F M Qi(s,a))
(8)
where the parameter β controls the each supervisor i's degree of imitating the
strategy (the F M Q-value) of the neighboring supervisor j.
Finally each supervisor i updates its strategy (denoted as E-value Ei(s, a))
for each action a under state s as the average between the FMQ-values of its
own and its neighbor j weighted by parameter p. Formally we have,
Ei(s, a) = (1 − p) ∗ F M Qi(s, a) + p ∗ F M Qj(s, a)
(9)
3.4.2 Supervisor Instruction Generation
Next we introduce how a supervisor generates instructions for its subordinates
at the end of each round. As previously mentioned, there are two forms of
instructions from a supervisor: rules and suggestions. First each supervisor i
normalizes the E-values, which serves as the basis for generating instructions
for its subordinates. Formally we have,
E′
i(s, a) =
Ei(s, a) − Ei(s)
σ
(10)
Hierarchical Heuristic Learning towards Efficient Norm Emergence
13
where Ei(s) is the mean of E-values averaged over all action estimates in state
s and as shown in Equation (11),
Ei(s) = Pa∈A Ei(s, a)
A
(11)
The parameter σ is the standard deviation of F M Q-value following Equation
(12),
σ =s 1
A Xa∈A
(Ei(s, a) − Ei(s, a))2
(12)
Given a state-action pair hs, ai, if the E′-value E′(s, a) is smaller than a
given threshold, it indicates that selecting action a is not a wise choice under
state s, thus it is encoded as a rule. Formally we have,
F = {hs, aiE′(s, a) < δ}
(13)
where δ is the threshold which is set to the value of -0.5 in this paper.
For each state-action pair (s, a), its recommendation degree d(s, a) is set
i(s, a). Thus the set of suggestions from supervisor i can be
to the value of E′
represented as follows,
D = {hs, a, E′
i(s, a)i a ∈ A, s ∈ S}
(14)
Given a state-action pair (s, a), if E′
a is not recommended under state s; if E′
agents are encouraged to select action a when they are in state s.
i(s, a) < 0, it indicates that selecting action
i(s, a) > 0, it indicates subordinate
3.4.3 Learning Strategy of Subordinates
Similar to the strategies of supervisors, each subordinate agent j also keeps a
record of a Q-value Qj(s, a) for each action a ∈ Aj under each state s. The
Q-value Qj(s, a) indicates the past performance of choosing action a under
state s and serves as the basis for making decisions [3]. For each subordinate
agent j, let us first denote its feedback information received by the end of
round t as F eedInf t
j = {hsm, am, rmim ∈ [1, t]}. At the end of each round
t, subordinate agent j updates its Q-value based on its feedback hst, at, rti as
follows,
Qj(st, at) = (1 − αj) ∗ Qj(st, at) + αj ∗ rt
(15)
where αj is the learning rate modelling its updating degree between using the
previous experience and using the most recent information.
Additionally, each subordinate agent also updates its Q-values by taking
into consideration both the optimistic assumption and the frequency informa-
tion [13]. Formally we have,
F M Qj(s, a) = Qj(s, a) + f req(s, a) ∗ rmax(s, a) ∗ C
(16)
14
Tianpei Yang et al.
where rmax(s, a) is the max reward of each action a based on its own experi-
ence, f req(s, a) is the frequency of getting the payoff of rmax(s, a) until now
for action a and C is a weighting factor defining the trade-off between updat-
ing using Q-values and maximum payoff information. f req(s, a) is calculated
the same as shown in Equation (3).
After receiving supervisor's suggestions, each subordinate further adjusts
its estimation of the goodness of each state-action pair based on the FMQ-
values as follows,
Ej (s, a) = F M Qj(s, a) ∗ (1 + d(s, a) ∗ ρ)
(17)
where d(s, a) is the suggestion degree on the state-action pair (s, a), and ρ is
a weighting factor controlling the influence of the recommendation degree on
the E-values.
Besides, supervisors also influence the subordinate agents' exploration rates.
Let us suppose a subordinate agent j selects action a under current state s. If
the supervisor i's recommendation degree d(s, a) < 0, which indicates subordi-
nate agent j's current choice is not recommended, and agent j should increase
the exploration rate to have more chance to select the recommended actions
next time. On the other hand, if the recommendation degree d(s, a) > 0, it
indicates the subordinates' current choice is recommended. Thus subordinate
agent j decreases its exploration rate to avoid selecting discouraging actions in
the future. For both cases, the adjustment degree varies depending on the ab-
solute value of the state-action pair's recommendation degree. Formally each
subordinate agent updates its exploration rate as follows,
ǫj = ǫj ∗ (1 − d(s, a) ∗ γ)
(18)
where γ is a weighting factor controlling the influence degree of the supervisor's
suggestion on the subordinates' exploration rates.
Finally, given the current state s, each subordinate agent j chooses its
action from those actions whose corresponding state-action pair do not belong
to the set F of rules based on the corresponding set of E-values according to
the ǫ-greedy mechanism. The probability πj (s, a) for choosing action a under
state s is defined in Equation 19. Specifically, each agent chooses its action
with the highest E-value with probability 1 − ǫj to exploit the action with
best performance currently (random selection in case of a tie), and makes
random choices with probability ǫj to explore new actions with potentially
better performance.
1 − ǫj
0
ǫj
Aj−Fj −1
if a = argmaxa′ E(s, a
if a ∈ Fj
otherwise
′
)
(19)
πj (s, a) =
4 Experimental Simulation
In this section, we start with evaluating the norm emergence performance of
our approach HHLS under different types of games by comparing with the
Hierarchical Heuristic Learning towards Efficient Norm Emergence
15
Table 4 The initial value of parameters.
Parameters
α
ǫ
β
γ
ρ
Value
0.99
0.93
0.1
0.05
0.01
d
6
Table 5 The explanations of parameters.
Parameters
Explanations
Learning rate
ǫ-greedy probability
α
ǫ
β
γ
ρ
d
Weighting factor on supervisors' interaction
Weighting factor on subordinates' exploration rates
Weighting factor on subordinates' E-Value
Average connection degree of small-world and scale-free network
state-of-the-art strategies. Following that we explore the influence of some key
parameters on norm emergence. We use degree grouping in the following ex-
perimental simulations, and the influence of different grouping mechanisms are
investigated in Section 4.3. All results are averaged over 1000 runs. The param-
eter settings and explanations are shown in Table 4 and Table 5 respectively.
4.1 Performance Evaluation in Different Games
We compare our approach HHLS with two previous works: hierarchical learn-
ing in [35] and social learning in [3]. All these three learning approaches are
within the same social learning environment, i.e., each agent is allowed to
interact with only one of its neighbors each round. The work in [3] is the
representative state-of-the-art approach tackling norm emergence problem un-
der multiagent social learning framework without considering any hierarchical
organization. The work in [35] is the most recent approach introducing hier-
archical learning into multiagent social learning framework to improve norm
emergence efficiency. Four representative 6-action games are considered shown
from Table 6 to Table 9.
4.1.1 Coordination Game (CG)
We first consider agents playing a 6-action coordination game (Table 6) in
which there exist six norms marked with red. Agents are preferred to choose the
same action. Fig 2 shows the dynamics of the average payoffs of agents with the
number of rounds averaged for the three learning approaches. We can observe
that all learning methods enable agents to achieve an average payoff of 1. Our
16
Tianpei Yang et al.
Table 6 The payoff matrix of coordination game.
Agent 2's actions
Agent 1's
actions
a
1
-1
-1
-1
-1
-1
a
b
c
d
e
f
f
f
o
y
a
p
e
g
a
r
e
v
a
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
0
HHLS (our strategy)
Hierarchical Learning
Social Learning
500
1000
round
1500
2000
b
-1
1
-1
-1
-1
-1
f
f
o
y
a
p
e
g
a
r
e
v
a
c
-1
-1
1
-1
-1
-1
d
-1
-1
-1
1
-1
-1
e
-1
-1
-1
-1
1
-1
f
-1
-1
-1
-1
-1
1
1
0.8
0.6
0.4
0.2
0
−0.2
−0.4
−0.6
−0.8
0
HHLS (our strategy)
Social Learning
Hierarchical Learning
500
1000
round
1500
2000
Fig. 2 The dynamics of the average pay-
offs of agents in coordination games under
different strategies.
Fig. 3 The dynamics of the average payoffs
of agents in anti-coordination games under
different strategies.
hierarchically heuristic learning strategy converges faster than the hierarchical
learning method [35], and the social learning method [3] is the slowest. This is
because HHLS enables the supervisor agents to influence subordinate agents
in a more efficient manner, thus accelerating norm emergence.
4.1.2 Anti-coordination Game (ACG)
Similarly, we consider agents playing a 6-action anti-coordination game (Ta-
ble 7) in which there also exist six equivalently optimal norms marked with
red. However, different from coordination game, each norm requires agents to
choose different actions. Fig 3 shows the dynamics of the average payoffs us-
ing three learning methods. We can observe that both social learning [3] and
HHLS enable agents to achieve an average payoff of 1, while the hierarchical
learning fails. Besides, our HHLS converge faster than the social learning ap-
proach [3], which justifies the efficiency of introducing a hierarchical learning
structure. For the hierarchical learning [35], it does not distinguish the state
Hierarchical Heuristic Learning towards Efficient Norm Emergence
17
Table 7 The payoff matrix of anti-coordination game.
Agent 1's
actions
Agent 2's actions
a
-1
-1
-1
-1
-1
1
b
-1
-1
-1
-1
1
-1
c
-1
-1
-1
1
-1
-1
d
-1
-1
1
-1
-1
-1
e
-1
1
-1
-1
-1
-1
f
1
-1
-1
-1
-1
-1
a
b
c
d
e
f
Table 8 The payoff matrix of coordination game with high penalty.
Agent 1's
actions
Agent 2's actions
a
10
0
-30
-30
0
-30
b
0
7
0
0
0
0
c
d
-30
-30
0
10
-30
0
0
-30
10
0
-30
-30
e
0
0
0
0
7
0
f
-30
0
-30
-30
0
10
a
b
c
d
e
f
information and thus cannot adaptively select different actions for different
states.
4.1.3 Coordination Game with High Penalty (CGHP)
Next, we consider 100 agents playing a 6-action coordination game with high
penalty (Table 8), in which there exist four optimal norms marked with red
and two suboptimal norms marked with blue. In this kind of games, agents
are vulnerable to converge to suboptimal norms due to the existence of high
mis-coordination cost (-30). Fig 4 shows the dynamics of the average payoffs
of agents with the number of rounds for the three learning approaches. We
can see that only HHLS enables agents to achieve an average payoff of 10 (i.e.,
converging to one optimal norm). The other two learning methods converge
to one of the suboptimal norms, and they also converge slower than HHLS.
We hypothesize the superior performance of HHLS is due to the integration of
optimistic assumption during strategy update (to overcome mis-coordination
cost effect) and efficient hierarchical supervision (to accelerate norm emergence
speed).
18
f
f
o
y
a
p
e
g
a
r
e
v
a
10
5
0
−5
−10
−15
0
f
f
o
y
a
p
e
g
a
r
e
v
a
10
5
0
−5
−10
−15
0
Tianpei Yang et al.
HHLS (our strategy)
Hierarchical Learning
Social Learning
500
1000
round
1500
2000
HHLS (our strategy)
Hierarchical Learning
Social Learning
500
1000
round
1500
2000
Fig. 4 The dynamics of the average payoffs
of agents in CGHP under different strate-
gies.
Fig. 5 The dynamics of the average payoffs
of agents in FSCGHP under different strate-
gies.
4.1.4 Fully Stochastic Coordination Game with High Penalty (FSCGHP)
Last, we consider agents playing a 6-action fully stochastic coordination game
with high penalty (Table 9). In FSCGHP, each outcome is associated with
two possible payoffs and the agents receive one of them with probability 0.5,
which models the uncertainty of the interaction results. The expected payoffs
of this game are the same with the CGHP and there also exist four optimal
norms marked with red and two suboptimal norms marked with blue. But
it is more complex and difficult to emerge norms due to the stochasticity of
the environments. Fig 5 shows the dynamics of the average payoffs of agents
against the number of rounds for the three learning strategies. We can observe
that in this challenging game, only HHLS enables agents to achieve an average
payoff of 10 (agents learn to converge to one optimal). In contrast, the other
two learning strategies converge to one of the suboptimal norms with a slower
convergence rate. Finally, it is worth mentioning that if the size of the action,
and norm space is further increased, the social learning method [3] and hier-
archical learning method [35] cannot converge (to a suboptimal norm) within
10000 runs. However, HHLS still can support converging to one optimal norm
within approximately 200 rounds. The influence of action size will be discussed
in more detail in Section 4.4.2.
4.2 Influence of Different Components of HHLS
Next, to justify the importance of the key components (optimistic assumption
and supervisor's guidance) in our HHLS framework, we further investigate the
relative importance of different modified versions of HHLS mechanisms: HHLS,
HHLS without optimistic assumption (i.e., HHLS without FMQ heuristic),
HHLS without Supervisors' guidance (i.e., HHLS without hierarchical struc-
Hierarchical Heuristic Learning towards Efficient Norm Emergence
19
Table 9 The payoff matrix of fully stochastic coordination game with high penalty.
1's payoff
2's payoff
Agent 1's
actions
a
b
c
d
e
f
Agent 2's actions
a
b
c
d
e
f
12/8
5/-5
5/-5
-20/-40
-20/-40
5/-5
-20/-40
14/0
5/-5
12/8
5/-5
5/-5
5/-5
-20/-40
5/-5
-20/-40
-20/-40
5/-5
-20/-40
5/-5
-20/-40
5/-5
5/-5
5/-5
12/8
5/-5
-20/-40
5/-5
-20/-40
-20/-40
5/-5
5/-5
-20/-40
14/0
5/-5
12/8
d
r
a
w
e
r
e
g
a
r
e
v
A
10
5
0
−5
−10
−15
0
10
d
r
a
w
e
R
e
g
a
r
e
v
A
HHLS
HHLS without
FMQ heuristic
HHLS without
Supervisors' guidance
500
1000
1500
Round
2000
2500
3000
5
0
−5
−10
−15
0
HHLS
HHLS without
FMQ Heuristic
HHLS without
Supervisors' guidance
500
1000
1500
Round
2000
2500
3000
Fig. 6 The dynamics of the average pay-
offs of agents in 3-action CGHP under dif-
ferent HHLS mechanisms.
Fig. 7 The dynamics of the average pay-
offs of agents in 6-action CGHP under dif-
ferent strategies.
ture). We use coordination game with high penalty (CGHP) as an illustrating
example and the results are similar under other matrix games.
Fig 6 shows the different average payoffs under three different HHLS mech-
anisms for a small-size CGHP with 3 actions. We can see that HHLS without
optimistic assumption or without supervisors' guidance still enables agents to
converge to an optimal norm. However, the convergence rate under these two
mechanisms is much slower than the full HHLS mechanism. Next, we consider
increasing the game size to 6 actions and the results are given in Fig 7. We
can see that only the full HHLS mechanism can enable fast convergence to the
optimal norm while the other two approaches fail. These results confirm the
hypothesis proposed in Section 4.1.3 that both components are critical and the
integration of the two components leads to superior performance of HHLS.
20
Tianpei Yang et al.
10
f
f
o
y
a
P
e
g
a
r
e
v
A
5
0
−5
0
K−means Clustering
Degree Clustering
Random Clustering
Kernighan Lin Clustering
50
Round
100
150
Fig. 8 The influence of different grouping
mechanisms on the performance of norm
emergence.
4.3 Performance of Different Grouping Mechanisms
In this section, we investigate the performance of norm emergence using dif-
ferent grouping mechanisms described above in Section 3.2.
Fig 8 shows the influence of different grouping mechanisms on the per-
formance of norm emergence. The results are averaged over 1000 iterations
in 3-action coordination game with high penalty with a population of 100
agents. We can see that all these grouping mechanisms enable efficient norm
emergence while degree grouping outperforms the rest of three grouping mech-
anisms. The slight differences among these grouping mechanisms prove to be
statistically significant under t-test.
Intuitively, those subordinate agents with higher degrees play a more im-
portant role in the whole population because they can influence more neigh-
boring agents. Degree grouping allocates agents with the same degree into the
same group. For more important agents (with higher degree), they can reach a
unified opinion more quickly, and in turn influence the rest of agents of lower
degrees to quicker convergence.
4.4 Influence of Non-hierarchical Parameters
In this section, we investigate the influence of key parameters on the perfor-
mance of norm emergence. We present the following results for hierarchically
heuristic learning using a small-world network. Except the parameter whose
effect is being investigated, the rest of parameters follow the same settings as
in Section 4.
Hierarchical Heuristic Learning towards Efficient Norm Emergence
21
f
f
o
y
a
p
e
g
a
r
e
v
a
10
5
0
−5
−10
−15
−20
0
100 agents
200 agents
500 agents
1000 agents
50
100
150
round
200
250
300
f
f
o
y
a
p
e
g
a
r
e
v
a
10
5
0
−5
−10
−15
0
21 actions
18 actions
15 actions
9 actions
6 actions
3 actions
200
400
600
800
1000
round
Fig. 9 The influence of population size.
Fig. 10 The influence of action size.
4.4.1 Influence of Population Size
The influence of population size is shown in Fig 9. For a 6-action coordination
game with high penalty (CGHP), we can clearly observe the norm emergence
efficiency is reduced with the increase of the population size. As the size of
clusters is held constant, the number of clusters increase as the population
size becomes larger. We hypothesize that the increase of clusters impede effi-
cient coordination among agents in the hierarchical framework in the following
two ways. First, the coordination efforts required among cluster supervisors
increase significantly with the number of clusters. Second, the increase of clus-
ters also add difficulty to the efficient guidance of each supervisor to generate
suggestions to guide all of their subordinate agents towards a consistent norm.
4.4.2 Influence of Action Size
Fig 10 shows the dynamics of the average payoffs of agents for different action
sizes on the 6-action coordination game with high penalty (CGHP). We can
see that the convergence rate decreases with increase in the size of the action
space. This is expected because the coordination space becomes larger when
the action size increases. Besides, larger action size usually results in more
chances of mis-coordination cost and suboptimal norms, which additionally
increases the coordination difficulty for agents selecting a consistent norm.
Finally, it is worth noting that with the increase in the action space size, our
framework can still efficiently support norm emergence without significantly
degrading the performance (norm emergence occurs within 1000 rounds for all
cases). In contrast, in previous socially learning framework without utilizing
a hierarchical organization [3], the norm convergence speed is significantly
reduced when the action space is increased.
22
Tianpei Yang et al.
Table 10 The average number of rounds needed before convergence under different network
topologies.
Convergence
Speed
Game type
CG ACG CGHP
FSCGHP
Network
topology
Grid
Ring
Random
Small-world
Scale-free
142
144
146
141
149
141
146
136
141
144
131
135
122
124
129
153
157
162
162
163
Table 11 The influence of neighborhood size.
Neighbor
size
Convergence
2
6
8
10
20
30
50
99
Rate
144
141
143
139
141
141
142
139
4.4.3 Influence of Network Topology
We evaluate the influence of five different networks: random network, grid
network, ring network, small-world and scale-free network on the 6-action
coordination game with high penalty (CGHP). Table 10 shows the average
number of rounds needed before convergence. We find that hierarchical so-
cial learning framework performs robustly under different network topologies.
HHLS enables agents to converge to norms in approximately the same number
of rounds under all the above five network topologies for the different game
types we have experimented with.
4.4.4 Influence of Neighborhood Size
We empirically evaluate the influence of neighborhood size varying it from 2
up to 99 (fully connected) with a population of 100 agents on the 6-action
coordination game with high penalty (CGHP). Table 11 shows the average
number of rounds needed before convergence for different neighborhood sizes.
We can see that the average number of rounds required stabilizes at around
140 rounds. This finding is different from the results usually observed in the
traditional socially learning framework without a hierarchical structure [3].
This is because in hierarchical social learning framework, each supervisor su-
pervises and guides a cluster of subordinate agents, which can overcome the
low connectivity disadvantage when the neighborhood size is small.
Hierarchical Heuristic Learning towards Efficient Norm Emergence
23
f
f
o
y
a
p
e
g
a
r
e
v
a
10
5
0
−5
−10
−15
0
Cluster size = 2
Cluster size = 6
Cluster size = 8
Cluster size = 10
Cluster size >=15
1000
900
800
700
600
500
400
300
200
100
e
c
n
e
g
r
e
v
n
o
c
e
g
a
r
e
v
a
500
1000
round
1500
2000
0
0
20
40
60
cluster size
80
100
Fig. 11 The influence of cluster size.
Fig. 12 The average number of rounds
needed before convergence under different
cluster sizes.
4.5 Influence of Hierarchical Factors
4.5.1 Influence of Cluster Size
One unique feature of the hierarchical social learning framework is the division
of clusters of agents. Fig 11 and 12 show the influence of cluster size on norm
emergence with a population of 100 agents on the 6-action coordination game
with high penalty (CGHP). From Fig 11, we can see that the norm emergence
rate is gradually increased with the increase of cluster sizes, and stabilizes when
the cluster size crosses 15. This phenomenon can be observed more clearly in
Fig 12, which shows the average number of rounds needed before convergence
is reduced with the increase of cluster size and stabilizes around 100 rounds.
When the cluster size is increased to 100 in the extreme case, it is essentially
reduced to centralized control in which only one supervisor agent supervises
all the rest of agents. In this case, all the communication and computation
burden would fall on this single supervisor agent. When the cluster size is 1,
it is essentially equivalent with the case of the traditional social learning with-
out a hierarchical structure. When the cluster size varies between 1 and 100,
with the increase of the cluster size, each supervisor agent can supervise more
subordinate agents and thus it is easier for agents to coordinate among each
other. However, as the cluster size exceeds certain threshold, the advantage of
centralized supervision diminishes. This property is desirable since the same
level performance as fully centralized supervision can be achieved under dis-
tributed supervision, which not only increases the robustness of the HHLS and
the framework itself but reduces the communication and computation burden
of supervisor agents.
24
Tianpei Yang et al.
f
f
o
y
a
p
e
g
a
r
e
v
a
10
8
6
4
2
0
−2
−4
−6
−8
Disable 0 supervisor
Disable 1 supervisor
Disable 2 supervisors
Disable 3 supervisors
Disable 4 supervisors
Disable 5 supervisors
0
200
400
600
800
1000
round
Fig. 13 The influence of the number of dis-
abled supervisors.
4.5.2 Influence of The Number of Disabling Supervisors
Next we further examine the robustness of HHLS in details by investigating
the following questions: whether a consistent norm can still rapidly emerge
and how is the emergence efficiency changed when a number of supervisors is
disabled? Fig 13 shows the dynamics of expected payoffs of agents with some
supervisors disabled, and the results are averaged over 6-action coordination
game with high penalty. We can see that hierarchically heuristic learning still
enables agents to converge to a consistent norm when some of the supervisors
is disabled. Though the convergence rate is gradually decreased with increased
in the number of disabled supervisors, better performance than the traditional
social learning framework can still be achieved. This is expected since those
subordinate agents without supervisors can only learn based on their local
information, and the hierarchical social learning framework would be reduced
to the traditional social learning framework when all supervisors are disabled.
4.6 Fixed-strategy Agents on Norm Emergence
In this section, we investigate the extraneous effects of norm emergence. Specif-
ically, we consider injecting a small number of agents with a fixed-strategy, and
study the influence of those fixed-strategy agents on the population's norm
emergence. We compare the influence of fixed-strategy agents on HHLS with
that on other learning strategies in previous work. Besides, we investigate the
influence of different placement strategies of fixed-strategy agents on the con-
vergence rate. For random networks, the location of each fixed-strategy agent is
randomly selected. For other network topologies, the location of placing fixed-
strategy agents could be critical in influencing the norm emergence in agent
societies. We also consider the influence of late intervention and the influence
Hierarchical Heuristic Learning towards Efficient Norm Emergence
25
norm(b,b)
norm(a,a)
e
c
n
e
u
q
e
r
f
0.5
0.4
0.3
0.2
0.1
0
(a,a)/(b,b)
(a,b)/(b,a)
−0.1
0
20
40
60
round
80
100
m
r
o
n
a
o
t
d
e
g
r
e
v
n
o
c
t
i
e
m
i
t
f
o
t
e
g
a
n
e
c
r
e
P
1
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7 8 9 10111213141516171819
Number of agents playing a fixed strategy b
Fig. 14 The frequency of each joint ac-
tion.
Fig. 15 The influence of fixed-strategy
agents on norm emergence.
of placing fixed-strategy agents on different levels within our hierarchical social
learning framework.
4.6.1 Influence of Fixed-strategy Agents on Norm Adoption
When all agents are employing HHLS under the hierarchical social learning
framework, for a 2-action coordination game, it can be observed that both
norms ((a, a) and (b, b)) can be achieved with the same frequency over multiple
runs as shown in Fig 14. This is understandable because the game itself is
symmetric and the agents do not have any preference towards either of the
two norms.
In contrast, Fig 15 shows the frequency of the two norms being evolved
when the number of agents playing the fixed-strategy b is increased gradu-
ally. Initially both norms ((a, a) and (b, b)) are evolved with roughly equal
frequency. As the number of fixed-strategy agents increases, the frequency of
evolving norm (b,b) gradually increases towards 1. This indicates that small
amount of agents, using a fixed and deterministic strategy, can have significant
influence on the overall population's norm emergence direction.
4.6.2 Comparison with Previous Works
Regarding the influence of fixed-strategy agent on norm adoption, we also
compare our HHLS with social learning [3] and hierarchical learning [35]. For
HHLS and hierarchical learning [35], all fixed-strategy agents are inserted into
groups of subordinate agents randomly. Fig 16 shows the dynamics of the fre-
quency of norm (b,b) emerging in 2-action coordination game as the number
of fixed-strategy agents increases. The coordination game is symmetric: both
norms ((a, a) and (b, b)) can be achieved with the same frequency. However,
we see that hierarchical learning [35] shows a slight bias towards norm (a,a)
initially. Given the same number of fixed-strategy agents, agents are more
26
,
)
b
b
(
m
r
o
n
a
o
t
i
g
n
g
r
e
v
n
o
c
e
m
i
t
f
o
t
e
g
a
n
e
c
r
e
P
1
0.8
0.6
0.4
0.2
0
0
350
e
m
i
t
e
c
n
e
g
r
e
v
n
o
c
e
g
a
r
e
v
A
300
250
200
150
100
50
20
HHLS
Hierarchical
Learning
Social
Learning
5
10
15
The number of agents playing a fixed strategy b
Tianpei Yang et al.
Random
BC
CC
DC
EC
10
20
30
40
Number of agents playing the fixed strategy b
Fig. 16 The influence of fixed-strategy
agents on norm adoption under different
strategies.
Fig. 17 The influence of fixed-strategy
agents on convergence time.
inclined to learn the norm played by fixed-strategy agents under the social
learning framework [3] than HHLS and hierarchical learning [35]. This is be-
cause a hierarchical structure weakens the influence between subordinates and
their neighbors through the supervisor's guidance.
4.6.3 Influence of the Placement Strategies of Fixed-strategy Agents
Previous sections investigate the effects of fixed-strategy agents which are
randomly distributed in the population. Next we consider the influence of
fixed-strategy agent placement strategies on the norm emergence rate. We
are interested in investigating whether placing fixed-strategy agents following
appropriate heuristics can decrease the convergence time.
We consider the following commonly used placement metrics: degree (DC),
betweenness (BC), closeness (CC) and eigenvector (EC) centrality heuristics.
These metrics quantify the structural properties of the underlying network and
previous work has shown these placement heuristics have better efficacy than
random placement [16,6,10]. The DC of an agent is defined as the numbers of
its neighbors. An agent with a higher DC means it can influence more agents
directly in a network. The BC of an agent measures the number of the shortest
paths in the network between other agents that pass through it. An agent with
higher BC can manipulate the information flow more effectively in a network.
The core idea of CC is that an agent is closeness central if it can interact
all other agents easily. The closeness centrality of an agent located at a node
is calculated using the average shortest path between the node and all other
nodes. The EC of an agent is related to the connections of its neighbors. An
agent has a higher EC if its neighbors has more connections with other agents.
This measure takes into account both direct and indirect connections between
agents. EC is calculated using the eigenvector of the largest eigenvalue given
by the adjacency matrix representing the network.
Hierarchical Heuristic Learning towards Efficient Norm Emergence
27
Fig 17 shows the influence of placing fixed-strategy agents following the
above placement metrics on convergence rate. Random placement is also con-
sidered as the baseline strategy. First we can see that the average convergence
time is gradually decreased with the increase of the number of fixed strategy
agents. Besides, BC, CC, DC and EC placement strategies outperform random
placement. Intuitively fixed strategy agents located in nodes with higher cen-
trality values can influence the whole population more quickly and widely, thus
significantly reducing the convergence time compared with random placement.
The performance difference between BC, CC, DC and EC is insignificant for
scale-free network. This is due to the reason that almost the same set of agents
are selected by these metrics in scale-free network since the values of Pearson's
Correlation between these metrics are very high (approximately 0.9).
4.6.4 Late Intervention
We have analyzed the influence of fixed-strategy agents on norm emergence
where we place fixed-strategy agents at the beginning of the population in-
teraction. Now we consider an alternative form of intervention, late interven-
tion, i.e., inserting some number of fixed-strategy agents when a norm already
emerges among the population, to investigate the robustness of our algorithm.
We first allow all agents to employ HHLS without fixed-strategy agents being
inserted. For a 3-action coordination game with high penalty (CGHP) pre-
sented in Table 12, when the whole population has already converged to one
of the optimal norms (e.g. norm (a,a)), we inject a small number of agents
playing a fixed-strategy c. We expect these fixed-strategy agents to be able to
influence the whole population to change to norm (c,c).
Fig 18 shows the influence of small number of fixed-strategy agents on
replacing an existing norm with the 3-action coordination game with high
penalty (CGHP). There exist two norms in CGHP and we can see that the
whole population of 100 agents first converges to norm (a,a) without fixed-
strategy agents at approximately the 190th time step. After this, we manually
replace the top 20 agents with the highest EC values with fixed-strategy agents
playing action c. We can see that the rest of agents are gradually incentivized
to adopt strategy c and the whole population finally converges to norm (c,c)
at approximately the 450th time step.
We further investigate the influence of late intervention using fixed-strategy
agents on influencing an existing norm under different game sizes. The results
are shown in Fig 19, which are averaged over 100 runs. We can see that for a
2-action coordination game, the frequency of norm (b,b) gradually increases
towards 1 when the number of fixed-strategy agents is larger than 50. Com-
pared with the results on initial intervention shown in Section 4.6.1, we can see
that late intervention needs a larger number of fixed-strategy agents to change
the whole population's norm emergence direction. Another observation is that
as the action space increases, late intervention of fixed-strategy agents fails to
influence the whole population. This phenomenon indicates that the emerged
norm becomes more stable with the increase of action space, in which late
28
Tianpei Yang et al.
Table 12 Coordination game with high penalty.
1's payoff
2's payoff
Agent 1's
actions
Agent 2's actions
a
b
c
b
0
7
0
a
10
0
-30
)
b
,
b
(
1
m
r
o
n
o
t
d
e
g
r
e
v
n
o
c
t
i
e
m
0.8
0.6
0.4
i
t
f
0.2
o
e
g
a
t
n
e
c
r
e
P
0
20
c
-30
0
10
2 actions
3 actions
4 actions
5 actions
6 actions
9 actions
10 actions
30
40
Number of agents playing fixed strategy b
50
60
70
80
Fig. 18 The influence of late intervention
of fixed-strategy agents on norm emergence.
Fig. 19 The influence of late intervention
of fixed-strategy agents on norm emer-
gence under different action sizes.
intervention may not exert any influence on norm emergence when the action
space is sufficiently large.
4.6.5 Influence of Fixed-strategy Agents in Different Levels of the
Hierarchical Framework
Finally, we investigate the influence of the location of fixed-strategy agents
when they are placed in different levels (i.e. supervisor level and subordinate
level) of the hierarchical learning framework. Fig 20 shows the influence of the
locations of fixed-strategy agents when they are placed in different levels given
a 2-action coordination game. The whole population of 500 agents are divided
into 20 groups using degree grouping. We can see that placing fixed-strategy
agents in the supervisor level influences the whole population more quickly
than placing fixed-strategy agents in the subordinate level. The result is rea-
sonable because supervisors play a more important role than subordinates in
the system. In addition to interacting with their neighbors as ordinary subor-
dinate agents, supervisors collect all subordinates' information and generate
recommendations for subordinates. With a number of supervisors playing a
fixed-strategy, they can not only can influence their neighbors through inter-
actions, but can pass their influence directly down to the group of subordinate
Hierarchical Heuristic Learning towards Efficient Norm Emergence
29
)
b
,
b
(
i
m
r
o
n
o
t
g
n
g
r
e
v
n
o
c
e
m
i
t
f
o
e
g
a
t
n
e
c
r
e
P
1
0.8
0.6
0.4
0.2
Supervisors as fixed
strategy agents
Subordinates as fixed
strategy agents
0
0
20
The number of agents playing a fixed strategy b
10
5
15
Fig. 20 The influence of the location of
fixed-strategy agents when they are placed
in different levels.
m
r
o
n
a
o
t
e
g
r
e
v
n
o
c
m
e
t
s
y
s
e
m
i
t
f
o
n
o
i
t
c
a
r
f
100
90
80
70
60
50
40
30
20
10
0
norm(a,a)
norm(b,b)
0
0.01 0.05 0.1
0.2
0.3
0.4
0.5
probability of interaction between isolated subnetworks
Fig. 21 Norm emergence in isolated sub-
networks.
agents they supervise through recommendations. Therefore, the whole system
is more responsive towards the influence of fixed-strategy supervisor agents.
4.7 Norm Emergence in Isolated Sub-networks
In human societies, isolated populations can use contradictory norms, e.g.,
driving on the "right" or the "wrong" side of the road. When the frequency
of interaction between the groups is quite small, a different norm can emerge
in each group. To this end, we investigate the degree of isolation required for
divergent norms to emerge in each group.
We consider two groups of equal number of 100 agents and agents from
one group interact with agents from the other group with certain probability.
Results are presented in Fig 21. We observe that when the probability of
interaction is higher than 0.2, a single norm pervades the entire population.
In roughly half of the runs all agents learn to norm (a, a) and the other half
of the runs, agents learn towards norm (b, b). However, when the probability
of interaction becomes less than 0.2, there are runs where divergent norms
emerge in two groups. It indicates that a higher interaction frequency between
two isolated populations could sustain a consistent norm emergence for the
whole system.
5 Conclusion and Future Work
We propose a hierarchically heuristic learning strategy to ensure efficient norm
emergence in different distributed multiagent environments with various inter-
action topologies and several interaction scenarios (games). Extensive simula-
tion shows that the proposed HHLS strategy can enable agents to reach con-
sistent norms more efficiently and in a wider variety of games compared with
30
Tianpei Yang et al.
previous approaches. We introduce different kinds of grouping mechanisms into
our framework and evaluate the performance of these grouping mechanisms
on norm emergence. The influence of different key non-hierarchical parameters
(e.g., population size, action space, neighborhood size and network topology)
are also investigated.
We also evaluate the influence of centralized and decentralized hierarchi-
cally design by examining the effects of different cluster sizes (e.g., different
number of supervisors). We find that certain degree of distributed supervision
(a number of supervisors) can achieve the same performance as fully cen-
tralized supervision (only one supervisor), and thus make the HHLS robust
towards the failure of certain supervisors. Moreover, we investigate the dif-
ferent components of the HHLS scheme and study the effect of the relative
importance of those components. The results confirm that both components
are critical and the integration of the two components leads to superior perfor-
mance of HHLS. Furthermore, the influence of fixed-strategy agents on norm
emergence are extensively investigated from different aspects and valuable
insights are revealed (e.g., initial and late intervention, different placement
strategies, fixed-strategy agents in different levels of the hierarchical struc-
ture). Lastly, we investigate the norm emergence in isolated population and
experiments show that different norms emerge in isolated groups and a higher
interaction frequency between isolated populations could sustain a consistent
norm emergence for the whole system.
As future work, it is worthwhile investigating how to extend HHLS to
multi-level hierarchical structures and how to design different communication
and negotiation mechanisms between supervisors to accelerate norm emer-
gence in larger action and agent spaces. Another worthwhile direction is to
further investigate the influence of fixed-strategy agents in different ways, e.g.,
late intervention with some number of agents randomly selecting actions in
contrast to fixed-strategy agents, mixing times between populations starting
with different norms.
Acknowledgements This work is partially supported by the subproject of the National
Key Technology R&D Program of China (No.: 2015BAH52F01-1), National Natural Science
Foundation of China (No.: 61304262, 61502072) and Tianjin Research Program of Applica-
tion Foundation and Advanced Technology (No.: 16JCQNJC00100).
References
1. Abdoos, M., Mozayani, N., Bazzan, A.L.: Holonic multi-agent system for traffic signals
control. Engineering Applications of Artificial Intelligence 26(5), 1575–1587 (2013)
2. Agotnes, T., Wooldridge, M.: Optimal social laws. In: 9th International Conference on
Autonomous Agents and Multiagent Systems (AAMAS 2010), Toronto, Canada, May
10-14, 2010, Volume 1-3, pp. 667–674 (2010)
3. Airiau, S., Sen, S., Villatoro, D.: Emergence of conventions through social learning.
Autonomous Agents and Multi-Agent Systems 28(5), 779–804 (2014)
4. Bianchi, R.A., Ribeiro, C.H., Costa, A.H.R.: Heuristic selection of actions in multiagent
In: IJCAI 2007, Proceedings of the 20th International Joint
reinforcement learning.
Hierarchical Heuristic Learning towards Efficient Norm Emergence
31
Conference on Artificial Intelligence, Hyderabad, India, January 6-12, 2007, pp. 690–
695 (2007)
5. Franks, H., Griffiths, N., Anand, S.S.: Learning agent influence in mas with complex
social networks. Autonomous Agents and Multi-Agent Systems 28(5), 836–866 (2014)
6. Franks, H., Griffiths, N., Jhumka, A.: Manipulating convention emergence using influ-
encer agents. Autonomous Agents and Multi-Agent Systems 26(3), 315–353 (2013)
7. Genter, K., Stone, P.: Adding influencing agents to a flock. In: Proceedings of the 2016
International Conference on Autonomous Agents & Multiagent Systems, Singapore,
May 9-13, 2016, pp. 615–623 (2016)
8. Genter, K., Stone, P.: Agent behaviors for joining and leaving a flock. In: Proceedings of
the 16th Conference on Autonomous Agents and MultiAgent Systems, AAMAS 2017,
Sao Paulo, Brazil, May 8-12, 2017, pp. 1553–1555 (2017)
9. Genter, K., Zhang, S., Stone, P.: Determining placements of influencing agents in a
flock. In: Proceedings of the 2015 International Conference on Autonomous Agents and
Multiagent Systems, pp. 247–255 (2015)
10. Griffiths, N., Anand, S.S.: The impact of social placement of non-learning agents on
In: Proceedings of the 11th International Conference on Au-
convention emergence.
tonomous Agents and Multiagent Systems-Volume 3, pp. 1367–1368 (2012)
11. Hao, J., Leung, H.f.: The dynamics of reinforcement social learning in cooperative mul-
tiagent systems. In: IJCAI 2013, Proceedings of the 23rd International Joint Conference
on Artificial Intelligence, Beijing, China, August 3-9, 2013, vol. 13, pp. 184–190 (2013)
12. Hao, J., Sun, J., Huang, D., Cai, Y., Yu, C.: Heuristic collective learning for efficient and
robust emergence of social norms. In: Proceedings of the 2015 International Conference
on Autonomous Agents and Multiagent Systems, AAMAS 2015, Istanbul, Turkey, May
4-8, 2015, pp. 1647–1648 (2015)
13. Kapetanakis, S., Kudenko, D.: Reinforcement learning of coordination in heterogeneous
In: Adaptive Agents and Multi-Agent Systems II:
cooperative multi-agent systems.
Adaptation and Multi-Agent Learning, pp. 119–131 (2005)
14. Kernighan, B.W., Lin, S.: An efficient heuristic procedure for partitioning graphs. Bell
system technical journal 49(2), 291–307 (1970)
15. MacQueen, J., et al.: Some methods for classification and analysis of multivariate ob-
servations. In: Proceedings of the fifth Berkeley symposium on mathematical statistics
and probability, pp. 281–297 (1967)
16. Marchant, J., Griffiths, N., Leeke, M.: Manipulating conventions in a particle-based
topology. In: Proceedings of the Coordination, Organizations, Institutions and Norms
in Agent Systems Workshop (2015)
17. Mihaylov, M., Tuyls, K., Now´e, A.: A decentralized approach for convention emergence
in multi-agent systems. Autonomous Agents and Multi-Agent Systems 28(5), 749–778
(2014)
18. Miralles, J.C., Esteva, M., L´opez-S´anchez, M., Morales, J., Salam´o, M.: Organisational
adaptation of multi-agent systems in a peer-to-peer scenario. Computing 91(2), 169–215
(2011)
19. Mukherjee, P., Sen, S., Airiau, S.: Norm emergence under constrained interactions in
diverse societies. In: 7th International Joint Conference on Autonomous Agents and
Multiagent Systems (AAMAS 2008), Estoril, Portugal, May 12-16, 2008, Volume 2, pp.
779–786 (2008)
20. Pacheco, J.M., Traulsen, A., Nowak, M.A.: Coevolution of strategy and structure in
complex networks with dynamical linking. Physical Review Letters 97(25), 258,103
(2006)
21. Savarimuthu, B.T.R., Arulanandam, R., Purvis, M.: Aspects of active norm learning
and the effect of lying on norm emergence in agent societies. In: Agents in Principle,
Agents in Practice, pp. 36–50. Springer (2011)
22. Sen, O., Sen, S.: Effects of social network topology and options on norm emergence.
In: Coordination, Organizations, Institutions and Norms in Agent Systems V, COIN
2009 International Workshops. COIN@AAMAS 2009, Budapest, Hungary, May 2009,
COIN@IJCAI 2009, Pasadena, USA, July 2009, COIN@MALLOW 2009, Turin, Italy,
September 2009. Revised Selected Papers, pp. 211–222 (2009)
32
Tianpei Yang et al.
23. Sen, O., Sen, S.: Effects of social network topology and options on norm emergence. In:
Coordination, Organizations, Institutions and Norms in Agent Systems V, pp. 211–222.
Springer (2010)
24. Sen, S., Airiau, S.: Emergence of norms through social learning. In: International Joint
Conference on Artificial Intelligence, vol. 1507, p. 1512 (2007)
25. Shoham, Y., Tennenholtz, M.: On the emergence of social conventions: modeling, anal-
ysis, and simulations. Artificial Intelligence 94(1), 139–166 (1997)
26. Terroine, E.F., Lvy, M.: Automated synthesis of normative systems. In: Proceedings of
the 2014 international conference on Autonomous agents and Multiagent systems, pp.
109–116 (2013)
27. Tiwari, R., Jain, P., Butail, S., Baliyarasimhuni, S.P., Goodrich, M.A.: Effect of leader
placement on robotic swarm control. In: Proceedings of the 16th Conference on Au-
tonomous Agents and MultiAgent Systems, AAMAS 2017, Sao Paulo, Brazil, May 8-12,
2017, pp. 1387–1394 (2017)
28. Villatoro, D., Sabater-Mir, J., Sen, S.: Social instruments for robust convention emer-
gence. In: International Joint Conference on Artificial Intelligence, vol. 11, pp. 420–425
(2011)
29. Villatoro, D., Sabater-Mir, J., Sen, S.: Robust convention emergence in social networks
through self-reinforcing structures dissolution. ACM Transactions on Autonomous and
Adaptive Systems (TAAS) 8(1), 2 (2013)
30. Villatoro, D., Sen, S., Sabater-Mir, J.: Topology and memory effect on convention emer-
gence. In: Proceedings of the 2009 IEEE/WIC/ACM International Joint Conference on
Web Intelligence and Intelligent Agent Technology-Volume 02, pp. 233–240. IEEE Com-
puter Society (2009)
31. Weibull, J.W.: Evolutionary game theory. MIT press (1997)
32. Yang, T., Meng, Z., Hao, J., Sen, S., Yu, C.: Accelerating norm emergence through
hierarchical heuristic learning. In: ECAI 2016 - 22nd European Conference on Artificial
Intelligence, 29 August-2 September 2016, The Hague, The Netherlands - Including
Prestigious Applications of Artificial Intelligence (PAIS 2016), pp. 1344–1352 (2016)
33. Ye, D., Zhang, M., Sutanto, D.: A hybrid multiagent framework with q-learning for
IEEE Transactions on Power Systems 26(4), 2434–
power grid systems restoration.
2441 (2011)
34. Young, H.P.: The economics of convention. The Journal of Economic Perspectives 10(2),
105–122 (1996)
35. Yu, C., Lv, H., Ren, F., Bao, H., Hao, J.: Hierarchical learning for emergence of social
norms in networked multiagent systems. In: AI 2015: Advances in Artificial Intelligence,
pp. 630–643. Springer (2015)
36. Yu, C., Lv, H., Sen, S., Hao, J., Ren, F., Liu, R.: An adaptive learning framework for
efficient emergence of social norms. In: Proceedings of the 2016 International Conference
on Autonomous Agents & Multiagent Systems, pp. 1307–1308 (2016)
37. Yu, C., Tan, G., Lv, H., Wang, Z., Meng, J., Hao, J., Ren, F.: Modelling adaptive
learning behaviours for consensus formation in human societies. Scientific reports 6
(2016)
38. Yu, C., Zhang, M., Ren, F., Luo, X.: Emergence of social norms through collective learn-
ing in networked agent societies. In: Proceedings of the 2013 international conference
on Autonomous agents and multi-agent systems, pp. 475–482 (2013)
39. Zhang, C., Abdallah, S., Lesser, V.: Integrating organizational control into multi-agent
learning. In: Proceedings of The 8th International Conference on Autonomous Agents
and Multiagent Systems-Volume 2, pp. 757–764 (2009)
40. Zhang, C., Lesser, V., Abdallah, S.: Self-organization for coordinating decentralized
reinforcement learning. In: 9th International Conference on Autonomous Agents and
Multiagent Systems (AAMAS 2010), Toronto, Canada, May 10-14, 2010, Volume 1-3,
pp. 739–746 (2010)
|
1811.08614 | 2 | 1811 | 2018-12-07T06:32:48 | Tetris | [
"cs.MA",
"cs.DC"
] | Tetris is an Asynchronous Byzantine Fault Tolerance consensus algorithm designed for next generation high-throughput permission and permissionless blockchain. The core concept of Tetris is derived from Reasoning About Knowledge, which we believe to be the most appropriate tools for revealing and analyzing the fundamental complexity of distributed systems. By analyzing the states of knowledge that each participant attained in an unreliable system, we can capture some of the basis underlying structure of the system, then help us designing effective & efficient protocols. Plus the adoption of Full Information Protocol (FIP) with the optimized message traffic model, Tetris has finally got high performance, with proved safety. Tetris achieve consensus finality in seconds, means transactions can be confirmed greatly faster than other scheme like Pow/Dpos. Tetris also achieve fairness, which is critically important in some areas such as stock market etc. | cs.MA | cs | Tetris
Jiajun Xu, Sam Huang
{xujiajun, sam.huang}@yeefoundation.com
( draft 0.2 )
Abstract
Tetris is an Asynchronous Byzantine Fault Tolerance consensus algorithm designed for next
generation high-throughput permission and permissionless blockchain. The core concept of Tetris
is derived from Reasoning About Knowledge, which we believe to be the most appropriate tools
for revealing and analyzing the fundamental complexity of distributed systems. By analyzing the
states of knowledge that each participant attained in an unreliable system, we can capture some
of the basis underlying structure of the system, then help us designing effective & efficient
protocols. Plus the adoption of Full Information Protocol (FIP) with the optimized message traffic
model, Tetris has finally got high performance, with proved safety. Tetris achieve consensus finality
in seconds, means transactions can be confirmed greatly faster than other scheme like Pow/Dpos.
Tetris also achieve fairness, which is critically important in some areas such as stock market etc.
1. Introduction
Blockchain technology has attracted great attention as one of the biggest discoveries in the past
decade for its capabilities of create trust among unlimited number of trustless nodes with no
identity authentication. It has emerged as a potentially disruptive technology with great opportunity
to transform our society and therefore be entitled blockchain revolution[23], nearly as important an
innovation as the creation of the Internet.
However, the gap between vision and reality is still huge. Especially at the most important side of
technical，current blockchains still have great room for improvement in order to achieve industrial-
scale application.
For example, the most successful project of blockchain cryptocurrency, Bitcoin has long faced
complaints for its hours long time waiting for confirmation of transactions. And Ethereum, a cute
cat Crypto-Kitties once almost disrupted the whole network - creating network congestion and
traffic for all users of the Ethereum blockchain.
The performance of blockchain significantly relies on the performance of the adopted consensus
mechanisms, e.g., in terms of speed of transaction process, speed of consensus finality,
robustness to arbitrary behaving nodes and network scalability.
After Pow, numerous PoX have been proposed but with few successfully go beyond the high
energy consumptive Pow.
Almost all PoX, and some relative new promising technics like VRF, sound good at first, finally turn
to the direction of add a BFT scheme to guarantee its safety, for example: Pos+BFT in TON,
Dpos+BFT in EOS, VRF+BFT in Algorand... This kind of taking BFT as a patch to rescue the
unreliable scheme comes along with potentially many problems.
Other proved schemes like PBFT, Hashgraph, capable of supporting much higher performance,
but determined by its system model and assumptions, can only apply to permission blockchain.
We here take a different approach.
At first, we are trying to design a full Byzantine Fault Tolerance algorithm with proved safety at the
core with the architecture adaptive to permissionless blockchain context.
BFT algorithm require a set of designated members to communicate with each other and take
collaborative actions. Then, so we'll depend on an upper level protocol to choose members, rotate
members, evaluate them and combine incentive into the system. We can take this protocol as
pluggable. Here traditional PoX can play a role while the most important safety property was
guaranteed by the BFT core.
Create a BFT algorithm is not an easy task, especially in a permissionless blockchain
environment. We try to ensure that our design decision can deduced from the core requirement of
the system, from the fundamental properties of the system model and environment. We found that
the theory of knowledge of distributed system can play an important role in this field and reasoning
about knowledge help us capture the most essential concept of system, protocol and
collaboration.
When construct the protocol and algorithm, we try at every step, not only we propose how/what to
do, more important, we should point out why to do? what's the benefit & shortcoming? For
example, we adopt full-information protocol(FIP) in Tetris, contrary to traditional belief that FIP
cause unrealistic traffic load, we achieve the minimum bandwidth requirement.
At the end, correctness proof will provided for all the key steps. It's crucial in a byzantine
environment.
Although the consensus mechanism is important for the performance of blockchain, it is not all.
Imagine that if you design a consensus algorithm support tens thousands of transactions per
second, but every node in the system need to consume tens of megabytes per second of
bandwidth and most of the cpu capacity. It is still unpractical.
So in Tetris, in order to reach the maximum potential of consensus algorithm, We have carefully
designed the p2p overlay sub-network structure, push/pull traffic model and multilevel node mode
etc. We will describe them later.
2. System Model
Blockchain at its core is a distributed system ，more exactly a decentralized distributed system.
When talking about distributed system, it's critical to define the system model precisely in order to
create formal proofs of various guarantees of the algorithms under investigation.
We focus on asynchronous distributed system with byzantine failure model because we believe
this model most accurately captures the properties of the public internet and trustless blockchain
nodes on the network.
By the term distributed, we mean that the system is composed of a set of nodes that can
communicate only by sending/receiving messages along a set of channels. The network is not
necessarily completely connected.
By the term asynchronous, we mean that there is no global clock in the system, the relative
speeds of nodes are independent, and the delivery time of messages is finite but unbounded.
The network can drop, delay, duplicate, or deliver messages out of order. However, we assume
that messages are eventually delivered, provided that the corresponding senders keep on
retransmitting them.
We assume authenticated channels, where the receiver of a message can always identify its
sender. Each node has a public/private key pair and all nodes know the others' public keys. We
use these keys to implement authenticated channels, and sign messages where needed.
We assume a Byzantine failure model where nodes may deviate arbitrarily from the protocol. We
allow for a strong adversary that can coordinate faulty nodes. The adversary is assumed to be
computationally bounded, meaning he cannot subvert common cryptographic techniques such as
signatures. It is also assumed that secure hash functions exist, for which collisions will never be
found. Typically, we assume the number of faulty nodes cannot exceed 1/3 of total node number.
The description of the system is based on the behaviors of the individual nodes/processes in the
system. (The term node and process are the same meaning and either could be used later
following traditional preferences in the context.) Details are here:
Local State
Local state contain all the local information of a process.
Action
An action is a function from local state to local state. There three types of actions: send actions
denoted send(m) where m is a message, receive actions denoted receive(m), and internal actions.
Event
⟨s, α,
⟩′
An event is a tuple
⟨s, α,
to a state transition. The
⟩s′
jth
consisting of a state, an action, and a state. Each event corresponds
sj−1
event in process i's history,
⟨
i
is denoted
ej
i
αj
i sj
,
i
⟩
,
.
Local History
A local history,
actions, beginning with a distinguished initial state. We write such a sequence as follows.
of process i, is a (possibly infinite) sequence of alternating local states and
hi
hi
=
α1
i
s0
i −→
α2
i
i −→
s1
α3
i
i −→
s2
s3
i
…
If we assumed that the local state includes a description of all past actions in the local history, then
the local history may be equivalently described as following:
hi
Message
s0
= ,
i s1
,
i s2
i s3
, …
i
A message is a triple
and b is the body of the message.
⟨i, j, b⟩
where i is the sender of the message, j is the message recipient,
Asynchronous Runs
We identify a system with its set of possible runs, which is a complete description of all the
relevant events that occur in a system over time. At every step in a run, the system is in some
global state, where a global state is a description of each process's current local state.
Each asynchronous run is a vector of local histories, one per process, indexed by process
identifiers. Thus we use the notation
a =< ,
Asynchronous distributed system
h1 h2 h3
,
, . . .
hN
>.
We model asynchronous distributed system as consists of the following sets.
1. A set P = {
the system.
p1 p2
,
, . . . ,
pN
} of processes/nodes, where N is the total number of processes in
2. A set C
⊆
{(i, j) i, j
∣
can send messages to process j.
P } of channels. The occurrence of (i, j) in C indicates that process i
∈
3. A set
Hi
of possible local histories for each process i in P.
4. A set M of messages.
5. A set A of asynchronous runs.
The model of an asynchronous system does not mention time. However, there is an ordering of
events in the system due to the fact that certain events are known to precede other events. It is
named happens-immediately-before and happens-before relation.
Happens-immediately-before
ex
i
happens-immediately-before event
Event
different events in the history of some process i and
is the sending of a message and
x<y, or (2)
ey
ex
j
i
m such that
= receive(m)
ex
i
= send(m)
and
ey
j
, denoted
↦ex
i
, if and only if (1)
are
occurs earlier in the sequence, i.e. i=j and
is the reception of that message; i.e. there exists
.
and
ey
j
ey
j
ey
j
ex
i
ex
i
Happens-before
ey
j
, then either
The happens-before relation, denoted
→
Thus if
ey
→ex
↦ex
j
i
i
.
ey
↦ez
k
j
Global State
, is the transitive closure of happens-immediately-before.
or there exists an event
and
such that
→ex
i
ez
k
ez
k
A global state at some time of run a is a N-vector of prefixes of local histories of a, one prefix per
process.
Consistent Cut
→ex
i
also in the global state.
A consistent cut of a run is any global state such that if
ex
i
The definition of global state above is not always consistent, while there are chance prefix of one
process's history contain reception of a message but no prefix of others history contain the
sending of the message. This is meanless, so the definition of consistent cut in asynchronous
system is analogous to the global state in synchronous system.
is in the global state, then
and
ey
j
ey
j
Note that a consistent cut is simply a vector of local states, we will use the notation (a,c)[i] to
indicate the local state of process i in cut c of run a.
Most of the description above come from[5], and are conceptually similar to the timed runs model
for describe synchronous system.[4][7]
Protocols
Processes usually perform actions according to some protocol. Intuitively, a protocol for process i
is a description of what actions process i takes as a function of its local state. A system's all
possible runs must follow a protocol P, an initial state, and a failure pattern.
Formally, we define a protocol in terms of a message generation function(which describes the
messages that each process sends as a function of its local state), a state transition function,
and an output function.
Full-information protocols
A protocol is said to be a full-information protocol if each process is required to send all its current
state to all processes at each step. The state of a process in a full-information protocol consists of
the process's name, initial state, message history, etc.
Thus the message generation function and state transition function of a full-information
protocol is predetermined. So the states of processes following a full-information protocol are
completely independent of their decision function; full-information protocols differ only in their
output functions.
Intuitively, the states of the processes in a full-information protocol make the finest possible
distinctions among histories. It gives each process as much information about the operating
environment as any other protocol could. That is why the full-information protocol is particularly
well suited for proving possibility and impossibility of achieving certain goals in distributed
systems, and also for the design and analysis of distributed protocols.
Roughly, it looks as if every process send all the information of its state to all the other processes
at each step can easily result in unrealistically large messages size. But in practice，we will see in
section 5, the total bandwidth requirement can reduce to a minimum with the the help of hash
chain and push/pull traffic mode.
3. Knowledge & Common Knowledge
After a clear definition of the concept of the system model and protocol, we can catch the points
here:
1. The tasks that distributed systems are required to perform are normally stated in terms of the
global behavior of the system. From outside of the system, we view it has a whole.
2. The actions that each individual process performs, based on the protocol, can only depend
on its local information.
Therefore, the most essential problem is how can we try to obtain some level of global information
or group information from the local information.
Reasoning about knowledge has been argued the right tools to analyses such systems. We regard
local information of an individual process as it's knowledge. And we think of communication in the
system as a means of transferring knowledge. Process receive messages to improving its state of
knowledge and send messages to share it's knowledge to others. At any time, what we have is
individual process's accumulating knowledge. What we want is some form of "group knowledge"
derived from process's own knowledge to perform coordinative actions.
Reasoning about knowledge has often been introduced in the classical distributed system
textbook by an interesting example of "the muddy children puzzle"[1][2]. And another similar "blue-
eyed islander puzzle" has also recently attracted the attention of a famous mathematician, Terence
Tao[25]
In [2], J.Y.Halpern defined a hierarchy of states of group knowledge, and described the
relationship between common knowledge and a variety of desirable actions in a distributed
system. The weakest state of group knowledge is distributed knowledge, which corresponds to
knowledge that is distributed among the members of the group, without any individual process
necessarily having it. The strongest state of knowledge in the hierarchy is common knowledge,
which roughly corresponds to "public knowledge".
If process i know a given fact
. Knowledge of processes need to satisfy two
properties. The first is that a process's knowledge at a given time must depend only on its local
history: the information that it started out with combined with the events it has observed since
then. Secondly, we require that only true things be known, or more formally:
, we denote it by
φKi
φ
i.e., if an process i knows
φ
, then
φ
is true.
Ki
φ ⊃ φ
What does it mean to say that a group G of processes knows a fact ? here are some reasonable
possibilities from [2].
φ
(read "the group G has distributed knowledge of
φDG
φ
distributed in G if someone who knew everything that each member of G knows would know
For instance, if one member of G knows
said to have distributed knowledge of
.
φ
"): We say that
holds iff some member of G knows
is
.
φ
, the group G may be
"): We say that knowledge of
(read "someone in G knows
and another knows that
ψ ⊃ φ
φ
ψ
.
φ
φ
φSG
More formally,
φSG
Ki
φ
φ ≡
SG ⋁
i∈G
"): We say that
(read "everyone in G knows
φEG
formally,
φ
φEG
holds iff all members of G know
. More
φ
φEk
G
, for k
≥
1 (read "
is
φ Ek
-knowledge in G"):
φ
Ki
φ ≡
EG ⋀
i∈G
φEk
G
EG
φ, for k ≥ 1.
φ =
E1
G
EGEk
G
φ
is defined by
Ek+1
G
φ =
Ek
is said to be
φ
everyone in G knows that
in the sentence
k
times.
φ
-knowledge in G if "everyone in G knows that everyone in /g knows that ... that
is true" holds, where the phrase "everyone in G knows that" appears
(read "
φ
φCG
if
is
φ Ek
G
-knowledge for all
k ≥
1. In other words,
is common knowledge in G"): The formula
is said to be common knowledge in G
φ
φ ∧ …
(We omit the subscript G when the group G is understood from context.)
φ ∧ ⋯ ∧
φ ≡
Em
G
CG
EG
φ ∧
E2
G
Clearly, the notions of group knowledge introduced above form a hierarchy, with
Cφ ⊃ ⋯ ⊃
Ek+1
φ ⊃ ⋯ ⊃ Eφ ⊃ Sφ ⊃ Dφ ⊃ φ
The analysis of knowledge in distributed system reaches the following conclusions:[2]
1. When communication is not guaranteed it is impossible to attain common knowledge. This
generalizes the impossibility of a solution to the well-known coordinated attack problem. It is
also impossible to reach consensus in a distributed system, so in our system model in
Section 2, we assume that messages should eventually delivered.
2. Common knowledge of strict definition, can not be attained in practical distributed systems,
because that common knowledge can only be attained in systems that support simultaneous
coordinated actions.
3. Introduced with the concept of internal knowledge consistency, it is safe at certain context to
assume that certain facts are common knowledge, even when strictly speaking they are not.
For example, in the muddy children puzzle, or a synchronous protocol that proceeds in
rounds, in which it is guaranteed that no process will ever receive a message out of round.
4. There are weaker variants of common knowledge, which are attainable in practical systems.
-common knowledge(eventual common knowledge),
-common knowledge,
For example,
timestamped common knowledge.
ϵ
◊
4. Revisit Consensus
Since the first introduction of "The byzantine generals problem" by L. Lamport, R. Shostak, and M.
Pease[10], Numerous articles have been published on the problem of consensus under all kinds of
situations. Here we will revisit the most simple Binary Byzantine Consensus.
The consensus problem is typically defined for a set of n known processes, with the maximum
number t of faulty processes. We say that a process is correct if it follows its algorithm until
completion, otherwise it is said to be faulty.
A binary consensus algorithm aims to achieve consensus on a binary value v ∈ {0, 1}. Each
process proposes its initial value (binary) and decides on a value v. The problem can be formally
defined in terms of three properties:
Validity: If all correct processes propose the same value v, then any correct process that
decides, decides v.
Agreement: No two correct processes decide differently.
Termination: Every correct process eventually decides.
The first two properties are safety properties, i.e., properties that say that some 'bad thing' cannot
happen, while the last is a liveness property, i.e., a property that states 'good things' that must
happen.
There are many algorithms have been proposed in the past decades for this kind of Asynchronous
Binary Byzantine Consensus problem. Most of them are complex, unintuitive and relatively hard to
implement. So here we are trying to create such an algorithm that is mainly deduced from some
basic principle and keep simple to be intuitively understood and implemented. We simplify the
things by:
1. Adoption of FIP, help us mainly focus on the decision function
2. Based on the maximum knowledge a single individual process can obtain, to infer the status
.
of other processes in the system.
The decision function take the local history
consistent c of a run a. The output of the decision function in a binary protocol is either 0 or 1 or
meaning no output currently.
is the local history of a
⊥
as parameter,
= (a, c)[i]
hi
hi
The decision function
(
)hi
(
hi
must satisfy:
) = {0, 1, ⊥}
1. With the accumulation of knowledge process attained by receive messages and growth of the
history,
must eventually output 0 or 1. i.e., process must decide eventually.
)hi
(
2. After
output 0 or 1, it should always output the same value later. i.e., process once
decided a value, it can not change then.
3. When
)hi
(
of process i ready to decide a value v, it must be determined that process i
have obtain the eventual common knowledge of the fact that every processes will eventually
decide on the same value. i.e.,
means
= v.
(a, c)[i] ⊨
φC◊
where
φ
(h)
4. When
)hi
(
of process i ready to decide a value v, it must be determined that process i
know there will never has eventual common knowledge of the fact that some processes
decide on the opposite value.
5. Certainly,
(
)hi
property of consensus.
can not always decide a fixed value of 0 or 1, which violate the validity
5. Tetris
< Figure 5.1 >
In a system with unrestricted number of nodes connected via a p2p network. For a certain period
of time, a fixed number of nodes are selected as members of Tetris. We call them validators.
Validators are special nodes that continuously interact with each other to reach consensus on a
set of transactions to be processed and generate new blocks in the system. When a validator join
the Tetris, a unique member id was assigned to it called VID.
How are the validators be selected, and how an old validator quit and how can a new validator
join, will be treated at upper level protocol called member rotation protocol. We'll describe it later.
Each validator continuously accept *transactions and events broadcasted on the p2p network from
all other nodes on the network. Each transaction and event can be accepted by a number of
validators.
Periodically, the validator create an event, and broadcast it to all other validators. An event
contains all the hashes of transactions received during the period, and all the hashes of its parents
which defined as:
Definition 5.1. Parents, Parents of an event comprised of self-parent and other-parents, self-
parent is the last event this validator created and other-parents are all the events received from
other validators during the period.
When an event was created by a validator, a sequence number N was assigned to it.
N = Max(N of all the parents) + 1.
Sequence number N start from 0 and will keep increasing. If N of an event is greater than that of
its self-parent's plus 1, then empty placeholder events were created between them to keep the
sequence number increasing one by one.
So an event is a collection containing the following entries:
VID: event's creator
N: a sequence number start from 0
Hash of all the transactions received during the period.
Hash of all the events received from other validator during the period.
which denoted as
E = {VID, N, {Hash of E0, E1, ...}, {Hash of tx0, tx1, tx2, ...}}
Where we specify the first item E0 must be the self-parent, if it does not exist, then we specify E0
= null and Hash(E0) = 0.
When the new created event to be broadcast, it is also be signed by the validator, so others
received the event can verify it.
When an event is received and verified, it include all the hash of its parent events. The validator
check these hashes, if the corresponding parent event of the hash has not arrived, it send a
request to ask for the event. In our p2p network design, this is done by send a DHT request. Of
course, according to our design, every events and transactions will write to the DHT temporarily
while broadcasting. This is our push/pull model we'll describe at the next section. An event will be
accepted until all its parent accepted.
Thus, in each validator, all the events it created or received, by their parents-children relationship,
form a directed graph like Figure 5.1. This graph is an actual data structure stored in the validator,
which we can call it a tetris and denoted it as
to denote the sub-
tetris under the event
of Figure5.1 look
. For example,
like:
𝒯
𝒯vid,n
𝒯alice,4 𝒯bob,3 𝒯carol,4 𝒯dave,3
Evid,n
. Furthermore, we using
,
,
,
What we do here, is first for the purpose to construct a Full-Information Protocol.
< Figure 5.2 >
Lemma 5.2. Tetris is a Full-Information Protocol
Proof: A protocol is said to be a full-information protocol if each process is required to send all its
current state to all processes at each step. In Tetris, when a validator broadcasting an event, the
event include the hashes of all transactions it received, the event also include hashes of all events
it received which also include all the hashes of it's received transactions and predecessor events
and so on. Because the hash function is secure, this hash chain can guarantee the whole history
is correct when it was received by other validators. other validators using these hashes to retrieve
the original information safely, and can know all the state of the sender.
This kind of FIP make validators in Tetris obtain maximum knowledge about the running of the
system. For example, in Figure 5.2, Alice at Sn+4 must know that Bob know That Carol know
Alice's Sn event. And Carol at Sn+4 must know Alice know that Bob's Sn event. Hashgraph's
gossip of gossip[20] can also be seen as a FIP. But their members need directly connect to each
other to gossip the information. We establish this FIP on a p2p broadcasting network, and
validators even don't know each other, and validators can freely join or quit the system. This is
crucial for a permissionless blockchain.
Events and transactions are broadcasted on the network, only their hashes are packed into the
body of events to keep the relationship information of them. By this kind of hash chain, Tetris is a
FIP with minimum bandwidth requirement.
Any time, the bottom line of events with same sequence number in Figure5.1 are called base
event, all the base events are waiting for a decision for whether it can be committed at this stage.
The decision was made based on the continuous coming new events and transactions fall down
from upside of the figure. The piled up new events with their relations form some kind of structure
that will satisfy certain conditions at a certain level and therefore help making the decision. Until all
the base events are decided YES/NO to commit, this bottom line is confirmed. All the transactions
in the YES decision event will tagged as ready to commit and will later packed into a block. The
block's height is same as the sequence number. Then the bottom line of base events is eliminated,
the above line of events drop down to the bottom line and repeat the above process to produce
the next block. We call this process a stage, the stage number is equal to the block's height or the
sequence number of the base event.
This process is very similar to the world famous Russian game Tetris, where new tetriminos fall
down continuously and the bottom horizontal line of tetriminos are destroyed automatically when
certain conditions are met. This is where the name of our algorithm comes from!
Of course, we are not designing games. What important here is each validator has its own tetris
, each validator should independently make a decision base on its own tetris, their decision
𝒯
must be the same, but their
may not be identical.
𝒯
At each stage, with the accumulation of knowledge of validators by receive events, validators
eventually get enough knowledge to deduce that some form of common knowledge can obtained
and then support coordinated decision. What we need to emphasize is that the computing of
stages are independent. Because at different stage, old validators might quit, new validators might
join. It is another crucial requirement for a permissionless blockchain. For example, at stage n, a
consensus decision is made that one validator should quit, then at next stage, all the events from
that validator need to be ignored.
𝒯
The tetris
of each validator grows up with new events been created and received as time goes
on. For the very recent events, at the upper part of the tetris, every validators may have events
that the others not yet seen. But for the older events at the bottom part, they will roughly be the
same after some time. They may not be identical at particular moment, but they are always
consistent. Here consistent means that if there are tetris A of Alice and B of Bob, if A and B both
contain event e, then they will both contain exactly the same set of ancestors for e.
Definition 5.3. Ancestors, Ancestors of an event e include e itself and all the parents of e, all the
parents of parents of e, and so on.
Lemma 5.4. Tetrises of all validators are consistent
Proof: A and B contain same event e, so the hashes of parents contain in the e must be the same,
and all the parents must be existed because if e to be accepted in a tetris, all his parents
according to the hashes must be received and accepted already. And so does the same of parents
of parents and so on. Because the hash function is assumed to be secure, so A and B must
contain all the same of ancestors of e.
Consistency of tetrises of all validators is the base of our following work when we define functions
on
at different validators must generate the same output.
𝒯
so same
𝒯
Although tetrises are consistent, but according to our assumption of the byzantine system model,
some of validators may be byzantine, typically, we assumed that number of faulty nodes does not
exceed t, and number of all nodes is 3t+1. A byzantine node may maliciously broadcast forked
events, e.g.,create different events and send one to some validators and another to others.
Definition 5.5. Know, An event x know y, if y is the ancestor of x, and ancestors of x do not
include a fork by the creator of y.
Definition 5.6. Know-well, An event x know-well y, if x know y, and there is a set S of events by at
least 2t+1 of the members such that x know every event in S, and every event in S know y.
Lemma 5.7. At a certain position of tetris, an event e has a fork e', then if there is an event x in
one validator tetris know-well e, then there are no event in any validators can know-well e'
Proof: Suppose tetris A of validator Alice and B of Bob, an event x in A know-well e, an event y in
B know-well e'. The x know-well e means there are events of at least 2t+1 members which x know
all know e. So there must exist events from at least t+1 non-faulty members know e. In tetris B,
event y should know events from at least 2t+1 members which know e'. Because tetris A and B
are consistent, so there must exist at least 1 non-faulty member who has an event know both e
and e'. According our definition of know, this is impossible.
This lemma is similar to the Strongly Seeing Lemma in Hashgraph[20] and the Prepared status in
PBFT[12]. All them doing is just like the Reliable Broadcast primitive in Bracha's[13], to restrict the
adversary in sending non-fork events, forces him to follow the protocol for events broadcasting.
In each stage of the running of Tetris, we'll define the concept of witness and round.
Definition 5.8. Round, At each stage of the running of Tetris, the base events are specified round
= 0. For each event e above the base events, We specify its round = r + i, where r is the maximum
round number of the parents of e, and i = 1 if e can know-well at least 2t+1 witnesses in round r, or
i = 0 otherwise.
Definition 5.9 Witness, Witness is the first event created by a validator in a round. The base
events are called round 0 witness.
, and
ECarol,sn
.
For example, in Figure 5.1, event
Ebob,sn
Up to here, we have transformed the chaotic relationship of events like shown in Figure 5.1, into a
relatively ordered structure as:
is a round 1 witness because it know-well
Ebob,sn+4
Ealice,sn
,
< Figure 5.3 >
This structure to some extent is similar to the round concept of synchronous system, where we
can see:
1. At every round r, there must exist at least 2t+1 witnesses suppose a round r+1 witness
existed.
2. Each round r witness must know-well at least 2t+1 round r-1 witnesses.
3. Witnesses at one same round may not have the same sequence number.
4. Witnesses of the Tetris of all validators are consistent according to the Lemmas above.
Having all above things prepared, we are going to discuss how the validators make decision.
As mentioned in Section 4, having a FIP, what we need to do the next step is to designing a
decision function
. With the witness defined, our decision function need only defined on
witness. For each base event e, we define the function as follow:
(𝒯)
function decide(T, e)
e.committable = UNDECIDED
for each witness w in round 1
w.vote = true if w know-well e, false otherwise.
for each witness w in round 2
s = the set of witnesses in round 1 which w know-well
w.vote = true if there are at least t/2+1 witnesses in s vote
true otherwise w.vote = false
for r = 3 to current max round
for each witness w in round r
s = the set of witnesses in round r-1 which w know-well
v = majority vote in s, true for a tie
n = number of events in s with a vote of v
if n >= 2t+1
e.committable = v
w.vote = v
return v as decided
else
w.vote = v
return UNDECIDED
This function run repeatedly whenever a new witness found in the Tetris, and finally all the base
events has their committable property been decided true or false.
Once all the base events been decided, we'll check all the transactions included in these
committable = true base events. If the hash of one transaction has appeared in at least t+1
committable = true base events or their ancestors from different validators, then the transaction
can be tagged as committable. All the committable transactions then validated by all the validators
and used to generate a new block. Each validator will create a block header for it and sign by its
private key and broadcast it. All the nodes on the network once receive t+1 block header signed by
validators can be sure to confirm the new generated block. Then this line of base events drop
down and a new line of base events repeat all above to generate the next block of the next stage.
There are still a few questions remain to be clarified.
First, why we need hash of a transaction appeared in t+1 validators for tagged as committable?
The problem comes from our push/pull traffic mode. Remember that when the validator receive a
new event, included it are hashes of its parents and hashes of txs. If the body of the event/tx of
these hashes have not been received by the validator, we should retrieve them from dht. So if a
byzantine validator send a fake event or txs without store item on the dht network, the receiver will
never got these items. As of events, we specify an event can be accepted only when all of its
ancestors have accepted to solve the problem. But as of txs, if the same scheme is adopted, the
acceptance of event will delay too much, and it is also not necessary when running the consensus
algorithm. So here we need t+1 validators, to ensure that at least 1 honest validator send that
hash to guarantee the tx and the hash of it is not faked.
The second question is what if there are less than t+1 base events can be decided committable =
true? Yes, at Lemma 5.6.13 below, we'll prove that it must exists at least t+1 base events will be
decided committable = true under our algorithm.
The third question, though tx is broadcasted on the network, when it reaches every validators,
great probability not guaranteed that the tx embodied into the events with same sequence number
of all validators. Yes, has mentioned above, we'll check the ancestors of the committable = true
events. these ancestors have sunk down the base line, but we still need them and they are all
consistent for all validators. What matters is we need stipulate a depth of how old ancestors can
take part in our checking of tx. For example depth of 10 means we'll check tx in ancestors whose
sequence number not less than sequence number of base events minus 10. The unlimited depth
bring another question of too much sunk ancestors events need to sync while new validator join
the group.
The fourth question, we require all the base events's committable status been decided before
generate new block. But what if some validators's base events have not received and not
contained in one's tetris? This is possible, image if one of the validators stop broadcasting it's
events, all the tetris of others will not contain it's base event. At Lemma 5.6.12, we'll prove that as
long as one base event decided, base events that has not appeared must decide false.
Now, it's time to prove the correctness of decision function defined above.
Lemma 5.10. For one base event e, if
never change that value when new event created, new witness found later.
)𝒯alice
(
output decision value v, it will keep output v,
Proof: It is obvious from the algorithm above.
Lemma 5.11. For one base event e, if
then an eventual common knowledge
(
φC◊
)𝒯alice
obtained where
φ
output decision value v, i.e., e.committable = v.
means e.committable = v.
(
)𝒯alice
output v at round r by witness w, then w must know-well at least 2t+1
Proof: If
witnesses at round r-1 whose vote = v. For any witness x at round r of other validators, x also
know-well at least 2t+1 witnesses at round r-1. The number of these witnesses whose vote
v
must less than t. Then x.vote must be v. Then at r+1 round, all the witnesses must all receive v
and can decide v. So when Alice decided v, Alice know that all others must decided v eventually，
and Alice also know when other eventually decided v, they know that all others must decided v
eventually, and so on. Therefore,
obtained.
≠
φC◊
Normally, eventual common knowledge can not guarantee consistency according [7]. We should
, there will never exists eventual common knowledge of
also ensure that when Alice obtain
opposite value. In our case, it is obvious. If Bob obtain
means
e.committable = !v. Then eventually Alice will decided on !v, this contradict with Lemma 5.10. as
once decided, the decision value will never change.
earlier than Alice where
C◊φ′
φC◊
φ′
A valid decision function should satisfy another requirement of termination,
must eventually
output a value v. Our algorithm above, theoretically can not guarantee termination according the
FLP theorem[11]. This can avoid by add a coin round periodically on the algorithm above. We can
use a cryptographic pseudorandom number or a shared coin protocol with all the above proof still
correct. For example, modified the algorithm as:
for each witness w in round r
s = the set of witnesses in round r-1 which w know-well
v = majority vote in s, true for a tie
n = number of events in s with a vote of v
c = a constant of interval of coin round, such as 10.
if r mod c > 0
if n >= 2t+1
e.committable = v
w.vote = v
return v as decided
else
w.vote = v
else
if n >= 2t+1
w.vote = v
else
w.vote = middle bit of w.signature
But deeper research reveal that such situation can only happened while adversary maliciously split
the vote at each step and the attacker should have the capability of fully control the internet and
manipulate every message of honest validators. Traditionally, for the sake of theoretical interest,
the proposed adversary models usually assume a strong adversary that completely controls the
scheduling of the network and decides which processes receive which messages and in what
order. In practice, a real adversary usually does not possess this ability, and if it does, it will
probably perform simpler attacks like blocking the communication entirely. Therefore, in practice,
the network scheduling can be random enough and lead to a fast termination. Two papers show
that this is true and that these algorithms can be practical (Moniz et al., 2006a, 2006b)[14].
So, algorithm above can satisfy all the requirements of consensus mentioned above at Section 4.
We still leave several lemma need to be proved here.
Lemma 5.12. Any time as long as one base event be decided, base events that have not
appeared must be decided committable = false.
Proof: Any time when a base event has been decided, either true or false, the decision must be
made by some witness above round 3, then at round 2, at least 2t+1 witnesses must existed. All
these witnesses must can not know base events that have not appeared. So they must vote false
at any tetrises of validators as all tetrises are consistent. Then at next round these base events
must be decided committable = false.
Lemma 5.13. At least t+1 base events will be decided committable = true every stage.
Proof: First, we assume a tetris
with 3t+1 base events and 3t+1 round 1 witness. Then the total
number of votes of round 1 witness is at least (3t+1)*(2t+1). We consider an extreme situation for
𝒯
distribution of these votes to base events, when t+1 base events get 3t+1 votes each, and the
remain 2t base events still has t*(3t+1) votes, on average 1.5t+0.5 for each, so they can not all
below 1.5t+0.5. Then consider any round 2 witness, each of them must vote true to the t+1 base
events with 3t+1 votes and then this t+1 base events must be decided committable = true. for 2t
other base events, 1.5t+0.5 votes is the maximum number they can consume and assure that they
don't be guaranteely decided committable = true.
Then we to reduce the number of base events, or reduce the number of round 1 witness, either
satisfy the condition better than above extreme situation. So at least t+1 base events will be
decided committable = true every stage.
Theorem 5.14. Tetris is a valid asynchronous byzantine fault tolerance consensus algorithm.
Proof: According to the lemmas above, all honest validators will eventually make the same
decision of the base events's committable property under our byzantine system model hypothesis.
And at least t+1 base events will be decided committable = true every stage. If the hash of one
transaction has appeared in at least t+1 committable = true base events or their ancestors from
different validators, then the transaction must be valid and can be retrieved from dht. All honest
validators then can validate these transactions and produce a valid block. Thus, Tetris is a valid
asynchronous byzantine fault tolerance consensus algorithm.
Member rotation protocol
Member rotation protocol is an upper level protocol to change the validators when Tetris running. It
is an indispensable basic requirement of permissionless blockchain.
What important first is Tetris it own has the capability of dynamically remove or add validators one
by one at any time. Because Tetris run by stages, each stage the decision functions run
independently, so when a new stage start, remove a member of Tetris is just ignore all the events
of this member, and add a member into the validators need only the new validator start receive
and broadcast events and transactions.
Furthermore, Tetris can review the performance of validators and suggest which validator to
remove at particular time.
Having these capability of Tetris, upper lever protocol of member rotation can be designed very
flexibly and can easily implemented as a pluggable module. So most of PoX currently existed can
adopt to do the work.
6. Properties
Performance
Tetris has been tested to support tens of thousands tps. Furthermore, Tetris's high performance
has several its own characteristics:
1. Due to the overlay sub-network design of the underlying p2p network of Tetris, the more
nodes in the system, more sub-network could be divided, thus the more tps supported.
2. The greater the volume of transactions per unit time，more frequently the events will created
and transmit, then quicker to reach consensus. The consensus finality will be reached faster
under the heavy load period.
These characteristics different so much from traditional system such as Bitcoin and Ethereum.
There the total throughput of the system is limited to the lowest end of the node participated in the
system.
Scalability
Scalability of a blockchain system is shown in the follow aspects:
1. Maximum number of nodes supported by the system
2. Maximum accounts supported by the system
3. Maximum tps of the system
4. Storage requirement of the system
Blockchain based on Tetris can:
1. Support unlimited nodes
2. Support unlimited accounts
3. Tens of thousands of tps
4. Minimum storage requirement for normal nodes.
Finality and state
In Tetris, once consensus is reached, it's final. So the history is frozen, the state of the system at
this point become immutable. This brings great benefits. For a new added node, it may not
necessary to sync all the huge history of the blockchain, but only sync with the latest state, thus
save a lot of storage space. Furthermore, this finality has the potential for simplify the future cross-
blockchain interaction or sharding scheme.
Fairness
The meaning of fairness here includes two aspects:
1. How the system order the transactions? If Alice and Bob submit a transaction to the system
almost at the same time, who will be the first one?
2. Can anybody in the system manipulate the consensus order of transactions? For example,
deliberately delay or ignore some of specific transactions?
The exact transaction order does not matter much for some applications, but can be critically
important for others. A real world case is in the field of high frequency trading[24], some guys are
willing to pay millions of dollars only to reduce the network latency by microseconds for their
transactions reach stock exchanges faster than others.
The fairness has been detailed discussed in[20]. Unlike in a centralized system, when consider a
distributed peer-to-peer system, the most tricky thing is the definition of "fair".
Obviously, the fair decision of transaction order can not depend on the timestamp since there are
no global time in asynchronous systems and simple timestamps are easily forged.
With the algorithms exist leader, such as PBFT, round-robin system, or Pow system, can "fair" be
defined as reflecting the order in which the transactions reached the current leader? Unfortunately,
It's still a bad idea. The leader could arbitrarily decide to delay or ignore some of transactions, no
single individual can be trusted.
In Tetris, there are two conditions need to be fulfilled for a transaction to be committed. First,
transaction should be received by t+1 validators. Then these validators should actively participate
in the interaction between other validators and its base event have a chance to be decided as
committable = true event. Fairness defined here reflects the speed of transactions reach a certain
number of validators and the activity and performance of these validators in the group. It sounds
better in a distributed environment.
7. Practical Issue
P2P network
Tetris relies heavily on the underlying network. To exert the full potential of Tetris, we have several
special requirements for the underlying p2p network in permissionless environment.
1. Overlay sub-network. It's crucial to have the ability to create and manage Overlay sub-
network in p2p network. In a flatten p2p network, each nodes could receive all the traffic, it
can not support a high throughput system.
2. Erasure code enabled DHT. Huge amounts of data will produce in high throughput system.
The storage efficiency and economy arisen as an important problem, and we think this can
be tackled by some kind of coding.
3. Temp DHT. In Tetris, transactions and events broadcasting on the network, but also need to
retrieve in a certain period of time in our push/pull traffic mode. So this kind of data need
temporarily store on the DHT and can be retrieved by who need it later.
We will develop a new p2p network for Tetris, and the topic may be elaborated in a separate
whitepaper.
Sentry Node
In a permissionless blockchain, validators of Tetris should hide themselves on the network to avoid
possible attacks like DDOS. Although the validators have no direct network connections to other
validators by broadcasting message on a p2p network. The few normal nodes directly connected
to the validators in p2p network might have the chance to analyze the traffic mode and locate the
ip address of the validators. We can avoid this situation by deploy a set of Sentry Nodes which
connect to the p2p network on behalf of the validators and validators only connect to their Sentry
Nodes.
8. Conclusions
A new Asynchronous Byzantine Fault Tolerance consensus algorithm, Tetris is proposed here for
future blockchain infrastructure with its great performance and proved safety.
Tetris algorithm / protocol is constructed mainly under the most challenging permissionless
blockchain context. It is easy to port to permission blockchain or other application areas.
We also provide a demo project at https://github.com/yeeco/tetris_demo implementing the basic
functions of Tetris, and may be of interest.
Reference
[1] R. Fagin, J. Y. Halpern, Y. Moses, and M. Y. Vardi. Reasoning about Knowledge. MIT Press,
Cambridge, Mass., 1995
[2] J. Y. Halpern and Y. Moses. Knowledge and Common Knowledge in a Distributed Environment.
JACM'90.
[3] Ajay D. Kshemkalyani , Mukesh Singhal, Distributed Computing: Principles, Algorithms, and
Systems, Cambridge University Press, New York, NY, 2008
[4] C. Dwork and Y. Moses. Knowledge and common knowledge in a Byzantine environment:
crash failures. Information and Computation, 88(2):156-186, 1990.
[5] P. Panangaden and S. Taylor. Concurrent common knowledge: define agreement for
asynchronous systems. Distributed Computing, 6(2):73-93, 1992
[6] K. M. Chandy and J. Misra. How processes learn. Distributed Computing, 1(1)40-52，1986.
[7] Joseph Y. Halpern, Yoram Moses, and Orli Waarts. A characterization of eventual Byzantine
agreement. In Proceedings of the Ninth ACM Symposium on Principles of Distributed Computing,
pages 333 -- 346, August 1990.
[8] MOSES, Y., AND TUTTLE, M.R. Programming simultaneous actions using common
knowledge. Algorithmica 3 (1988), 121-169.
[9] Gil Neiger. Using knowledge to achieve consistent coordination in distributed systems. In
preparation, July 1990.
[10] L. Lamport, R. Shostak, and M. Pease, "The byzantine generals problem," ACM Trans.
Program. Lang. Syst., vol. 4, pp. 382 -- 401, July 1982.
[11] M. J. Fischer, N. A. Lynch, and M. S. Paterson, "Impossibility of distributed consensus with
one faulty process," J. ACM, vol. 32, pp. 374 -- 382, Apr. 1985.
[12] M. Castro and B. Liskov. Practical Byzantine Fault Tolerance, Proceedings of the Third
Symposium on Operating Systems Design and Implementation. New Orleans, Louisiana, USA,
1999, pp. 173 -- 186.
[13] G. Bracha, Asynchronous byzantine agreement protocols, Inf. Comput., vol. 75, pp. 130 -- 143,
Nov. 1987.
[14] M. Correia, G. S. Veronese, N. F. Neves, and P. Veríssimo, "Byzantine consensus in
asynchronous message-passing systems: a survey," IJCCBS, vol. 2, no. 2, pp. 141 -- 161, 2011.
[15] S. Toueg, Randomized byzantine agreements, in Proceedings of the Third Annual ACM
Symposium on Principles of Distributed Computing, PODC 1984, (New York, NY, USA), pp. 163 --
178, ACM.
[16] A. Clement, E. Wong, L. Alvisi, M. Dahlin, Making Byzantine Fault Tolerant Systems Tolerate
Byzantine Faults. NSDI, 09
[17] S. Nakamoto, Bitcoin: A peer-to-peer electronic cash system, 2008. [Online]. Available:
http://bitcoin.org/bitcoin.pdf
[18] E. Buchman. Tendermint: Byzantine fault tolerance in the age of blockchains. M.Sc. Thesis,
University of Guelph, Canada, June 2016.
[19] S. Popov. The tangle. White paper, available at https://iota.org/IOTA_Whitepaper.pdf, 2016.
[20] Baird L. The Swirlds Hashgraph Consensus Algorithm: Fair, Fast, Byzantine Fault Tolerance,
Swirlds Tech Report SWIRLDS-TR-2016-01(2016)
[21] NXT Whitepaper, 2014. [Online]. Available: https://wiki.nxtcrypto.org/
wiki/Whitepaper:Nxt#ProofofStake
[22] BitShares Delegated Proof of Stake, 2014. [Online]. Available: https:
//github.com/BitShares/bitshares/wiki/Delegated-Proof-of-Stake
[23] Don Tapscott, Alex Tapscott, Blockchain Revolution: How the Technology Behind Bitcoin Is
Changing Money, Business, and the World. Brilliance Audio, 2017
[24] https://www.aldec.com/en/company/blog/169--the-race-to-zero-latency-for-high-frequency-
trading
[25] Terence Tao, https://terrytao.wordpress.com/2011/05/19/epistemic-logic-temporal-epistemic-
logic-and-the-blue-eyed-islander-puzzle-lower-bound/
|
1911.09535 | 3 | 1911 | 2019-12-13T16:10:42 | Agent Probing Interaction Policies | [
"cs.MA",
"cs.AI",
"cs.LG"
] | Reinforcement learning in a multi agent system is difficult because these systems are inherently non-stationary in nature. In such a case, identifying the type of the opposite agent is crucial and can help us address this non-stationary environment. We have investigated if we can employ some probing policies which help us better identify the type of the other agent in the environment. We've made a simplifying assumption that the other agent has a stationary policy that our probing policy is trying to approximate. Our work extends Environmental Probing Interaction Policy framework to handle multi agent environments. | cs.MA | cs | AGENT PROBING INTERACTION POLICIES
A PREPRINT
Siddharth Ghiya1, Oluwafemi Azeez1, Brendan Miller1
1 Carnegie Mellon University
Pittsburgh, PA, USA
9
1
0
2
c
e
D
3
1
]
A
M
.
s
c
[
3
v
5
3
5
9
0
.
1
1
9
1
:
v
i
X
r
a
December 16, 2019
ABSTRACT
Reinforcement learning in a multi agent system is difficult because these systems are inherently
non-stationary in nature. In such a case, identifying the type of the opposing agent is crucial and can
help us address this non-stationary environment. We have investigated if we can employ some probing
policies which help us better identify the type of the other agent in the environment. We've made a
simplifying assumption that the other agent has a stationary policy from a fixed set that our probing
policy is trying to classify. Our work extends Environmental Probing Interaction Policy framework to
handle multi agent environments.
1
Introduction
Real world environments often have multiple agents. In such a case, considering the actions of the other actors in the
environment becomes crucial. In many of the scenarios, the identity of other agents may not be known. Therefore,
identifying the type of other agent is useful so that we can make a more informed decision taking into account the
nature of behaviour of other agents in the environment.
Tremendous progress has been made in the field of multi-agent reinforcement learning. However most of these works
have considered scenarios which are either only collaborative or scenarios where the type of the opposing agent is
already known. In real world scenarios, agents often have to work in environments where there might be other unknown
agents present. In such scenarios, we can perform actions and observe how the other agent in the environment reacts to
our actions to determine the type of the other agent.
In this work, we present a framework which combines a Reinforcement Learning agent and an LSTM classifier to
identify the type of the opposing agent present in the environment. We learn a "probing" policy which our agent
performs, and observe the reactions of the other agent in the environment.
2 Related Work
2.1 Multi Agent reinforcement learning
MARL is one of the more widely studied topics in the field of reinforcement learning. Independent Q-Learning [1]
represents some of the earliest works in this field. In Independent Q-Learning, agents in the environment are trained
with the assumption that the other agent is part of the environment. Naturally, this fails if we increase the number of
agents in the environment due to non-stationarity in the environment.
More recently, there has been some work in MARL which can be classified under the "centralised training and
decentralised execution paradigm" [2] [3] [4][5]. In most of these works, a centralised critic is maintained during the
training process which accounts for the non-stationarity in the environment. In all of these works, no communication
is assumed between the agents. There has been some work in which agents learn to communicate with each other
[6] [7] [8] [9] which helps counter the non-stationarity of the environment. Some of the works don't explicitly make
assumptions about the type of other agents in the environment. In this context, there has been work where an agent
actively tries to modify the learning behaviour of the opposing agent [10] or employs recursive reasoning to reason
A PREPRINT - DECEMBER 16, 2019
about the behaviour of other agents in the environment [11].
Our work falls more under the paradigm of agent modelling [12]. [13] [14] try to reason about the type of the other
agent in the environment by observing their behaviour while carrying out their own policy. As they collect more data
about the opposing agent in the environment, they condition their policy on the belief of the other agent and are able to
achieve better generalization against different kinds of agents. To the best of our knowledge, in none of the available
literature, attempts have been made to learn a policy which helps to identify the type of other agent present in the
environment. To this end, we want to propose a separate policy, which when executed helps identify the type of the other
agent in the environment. We speculate that, having information about the type of the other agent in the environment
might help us in conditioning our policy and generalise better against different types of agents.
2.2 Non stationary environments
Multi-Agent systems are inherently non-stationary in nature. A lot of literature is available for dealing with non-
stationary environments in single agent reinforcement learning. Many of these approaches use meta learning [15] [16]
[17] [18] to adapt to such non stationary environments. [19] explicitly identify the properties of the environments which
make it non-stationary and then use the obtained information to condition their policy. A more recent approach [20] on
dealing with non-stationary environments is learning a probing policy to identify the dynamics of the environment and
then using this information to make an informed decision. We intend to use similar probing policies to help us better
identify the type of other agent present in the environment.
2.3 Classifier
Our work does classification on trajectories, which are sequences of states. Recurrent neural networks are suitable for
working with sequences. Long Short-Term Memory [21] is probably the most popular type of RNN and is used in our
classifier.
Gated Recurrent Units [22] are an improved and more computationally efficient RNN. We may experiment with
replacing our LSTM based classifier with a GRU in future works.
3 Method
Our approach involves iteratively training a probing RL agent and a classifier in alternating phases, as diagrammed in
Fig. 5. First, the RL agent (see Fig. 4) executes a series of steps over several episodes according to a probing policy,
while receiving reward based on a fixed classifier. The reward given to the RL agent during training is the negative cross
entropy loss of the classifier. Then, after the probing policy is updated, a new classifier is trained.
During the classifier (see Fig. 3) training phase, a large batch of trajectories are generated from the probing policy, along
with the ground truth of the opposing agent type. The LSTM based classifier then learns to categorize the trajectories
according to agent type. After the classifier is trained, it is frozen, and the RL agent training phase starts up again. This
whole process repeats for several iterations.
The probing RL agent learns to take steps that can help the classifier detect the opposing agent correctly. Typically, this
involves moving into the opposing agents field of view in order to cause it to take some action that can be used for
classification. The opposing agent itself is controlled by a policy randomly selected at the beginning the episode.
Our current framework has no concept of an early done state. We always train with fixed length episodes.
3.1 Environment
Figure 1: Explanation of the Environment
The environment used is a modification of openAI gym [23] frozen lake environment. The environment was originally
designed for a single agent navigating a grid world. The game ends whenever the agent either reaches a goal or falls into
2
A PREPRINT - DECEMBER 16, 2019
a hole. For the purpose of our experiment we modified the environment to use the holes as obstacles. Also the game has
no concept of goal, it's only goal is to probe its opponent to help classify it properly. We also made the environment
multi agent, i.e the agent and opponent agent take actions simultaneously. The state representation was designed to be
easily to be used as input to the RL agent and the classifier. We used a matrix of size 4x4 with a value of 1 wherever
there's an obstacle, 2 for where the agent is located and 3 for where the opponent is located. Rendering is done on the
command line.
3.2 Opposing Agent
We have two separate sets of experiments. In first set of experiments, we only have two kind of opposing agents in the
environments, but we have a more challenging randomized environment. We also restrict the field of view of opposing
agents. In the second set of experiments, we have many types of opposing agents, but the environment is simplified.
Different settings of the opposing agents have been explained in more details below:
3.2.1 Agents for Difficult Environments
In this set of experiments, we only have two kinds of opposing agents in the environment. These opposing agents either
go towards our API agent (predator) or go in the opposite direction (prey). Also, the opposing agent can only sense the
presence of the API agent in the environment if the API agent is within its line of sight i.e. in the same row or in the
same column with no intervening obstacles.
The starting positions of the agents are randomized but are always in positions where the API agent has to actively try
to carry out a policy to try to get the opposing agent to react. A good example of such initialisation can be found in Fig.
1. In this example, our API agent will have to navigate into the line of sight of the the opposing agent in order to get it
to react.
3.2.2 Agents for Many Agent Environments
In this set of experiments, the opposing agent's policy is dependent on the main agent's actions. An agent can take four
actions in the frozen lake gridworld, which are LEFT, RIGHT, UP and DOWN. The opposing agent can be deterministic
or stochastic. when it's behaviour is deterministic, it can be
1. diagonal: This opposing agent would always take 90 degree anti-clockwise direction of the agent's direction.
So if the agent goes LEFT, this agent would go DOWN.
2. opposite: This opposing agent would always go the opposite direction of the original agent. So if the agent
goes LEFT, then it would go RIGHT.
3. opposite-diagonal: This opposing agent would always take 90 degree clockwise direction of the agent's
direction. So if the agent goes LEFT, this agent would go UP.
4. follower: This opposing agent always follow the original agent. So if the agent goes LEFT, this agent would
always go LEFT.
Figure 2: Opposing agents
The opposing agents can also be stochastic. This means an action can be categorically sampled from a distribution with
more weights on each of the options listed above. i.e, stochastic diagonal would give more weights to the diagonal
3
A PREPRINT - DECEMBER 16, 2019
Figure 3: LSTM Based Classifier architecture
direction in the probability distribution sampled from while an opposing agent would have more weights on the opposite
direction of the agent. We then have a final opposing agent which just behaves randomly. This framework of opposing
agents could be generalized to a scenario where we can always create 2n + 1 opposing agent given n actions the original
agent can take. The conjecture is that stochastic opposing agents would prove very challenging to the agent, it would be
easier to detect deterministic agents than stochastic ones.
3.3 Classifier
The classifier takes in a trajectory specified as a world map W and a sequence of states s1 and s2 for agents 1 and 2.
The world W is an m by m occupancy grid specified as a binary matrix.
The states are sequences of row ri and column ci coordinates:
W =
0
1
1
...
...
0
r11
r21
r12
...
r1n
r22
...
r2n
S1 =
S2 =
0
...
0
···
...
···
c11
c12
...
c1n
c21
c22
...
c2n
(1)
(2)
(3)
(4)
When feeding into the LSTM based classifier the complete state for time i is flattened and fed into the classifier,
along with the world map of that state. Currently, the world map is fed in for each time step in the sequence, which is
somewhat wasteful.
The complete sequence is described in U below:
r11
U =
c21
c11
c21 w11 w12
...
...
...
r1n c1n c2n c2n w11 w12
...
...
...
··· wmm
...
··· wmm
4
A PREPRINT - DECEMBER 16, 2019
Figure 4: API Agent network architecture
Figure 5: Diagram of the overall architecture
3.4 API Agent
We use Proximal Policy Optimization Algorithm [24] for training the agent to follow a probing policy. The reward for
the agent is the negative cross entropy loss of the classifier as shown in Fig. 5. Currently, we are using a sparse reward
formulation in which the agent is rewarded only at the last time step.
We have a simple 2 layer fully connected layer which takes in the current state of the world. We have two separate heads
for policy and value. Please refer to Fig. 4 for the network architecture of the probing agent.
4 Results
4.1 Agents for Difficult Environments
We have a trajectory classifier implemented using an LSTM that runs on the sequence of states. It runs on a fixed world
pictured in Fig. 1. In this experiment, a 4x4 world is used with some blocks in the center which must be navigated
around.
Every episode starts with randomly placed agents. The agents then move according to their policy. In this experiment,
the probing agent moves according to it's policy. It's then given a reward at the last time step of the trajectory. Currently
we have a trajectory of maximum 20 time steps. The reward given to the API agent at the end of it's trajectory is
5
A PREPRINT - DECEMBER 16, 2019
Figure 6: Accuracy of the classifier tested every
200 episodes
Figure 7: Comparison with random probing
agent
negative of the loss of the classifier for that trajectory. Basically, the classifier looks at the sequence of states of the
trajectory and outputs a label of the opposing agent.
We randomly generate 1000 episodes of 20 time steps each. The API agent is trained in every episode. The classifier
is trained only every 200 episodes. Also, whenever the classifier is trained, we generate some validation trajectories
and calculate the accuracy for which the classifier is able to correctly identify the opposing agent in the environment.
The environment which we used in these set of experiments is similar to world pictured in 1 with randomized obstacle
positions and randomized start positions of the two agents. As we see in the results, the classifier is able to correctly
classify trajectories and recognize the opposing agent(Fig. 6).
We have a video of our experiments at : https://www.youtube.com/watch?v=6Y5fpdrQifA. In the video,
it can be seen that the probing agent sometimes wanders off the optimal path to the probing position.
We think that we observe such a performance because currently we are providing reward to the API agent(negative loss
of the classifier) at only the last time-step(which is fixed at 20). This can be improved by giving a reward to the agent at
every time-step or by penalising its movement.
We also did a comparison with a random probing agent. In case of random probing agent, the probing agent is taking
random actions at every time step. We then pass this trajectory through the classifier. As expected, in such a case, the
classifier is not able to properly classify the opposing agent in the environment. Since there is no policy which guides
the probing agent to a spot in the environment such that the opposing agent is able to see it and react to it's presence,
the classifier ends up giving predicting random labels for the opposing agent(Fig. 7).
4.2 Experiments with Many Agent Environment
4.2.1 One State Trajectory
We randomly generated 1000 episodes of 128 time steps each. Then we train our network using the Adam optimizer
with a learning rate of 0.001 for 20 epochs.
Each step was treated as a one trajectory move and the 128 steps as 128 batch. Average reward of the RL agent and Loss
of the classifier of the 128 batch was then plotted against 1000 timestep. An experiment was setup for only deterministic
opposing agents, deterministic and stochastic opposing agents and an agent that runs random policy was used as a
baseline.
The classifier loss result is shown in Fig. 8 and agent reward is shown in Fig. 9. It is obvious that classifier loss
reduces while the agent reward increases and is better than random in both case. The agent could easily classify the
opposing agent whenever it's a simple as being deterministic, however when some of the opposing agent start behaving
stochastically, it begins to get very difficult for the agent to classify. This generally exposes a rare kind of supervised
learning difficulty where ground truth label data is stochastic. The loss of combination of all opponents in Fig. 8 is
6
A PREPRINT - DECEMBER 16, 2019
Figure 8: One state Trajectory Classifier Loss
Classifier Loss against number of iterations.
Figure 9: One state Trajectory Reward
Reward against number of iterations.
Figure 10: 128 state Trajectory Classifier Loss
Classifier Loss against number of iterations.
Figure 11: 128 state Trajectory Reward
Reward against number of iterations.
higher than for deterministic opponents because the classification loss is for 9 opposing agent and would be as high as
ln(9) initially compared to 4 opposing agent which starts at about ln(4).
4.2.2 128 State Trajectories
To verify if adding extra information would provide extra information, we added another experiment in which we
randomly generated 1000 episodes of 128 time steps each. Then we train our network using the Adam optimizer with a
learning rate of 0.001 for 20 epochs.
Each step was treated as a 128 trajectory move and a batch size of 10. Average reward of the RL agent and Loss of
the classifier of the 10 batch was then plotted against 1000 timestep. An experiment was setup for only deterministic
opposing agents, deterministic and stochastic opposing agents and an agent that runs random policy was used as a
baseline.
The classifier loss result is shown in Fig. 10 and agent reward is shown in Fig. 11. The expectation here is that giving
more state information in terms of a full trajectory to the LSTM should significantly improve the classification loss and
agent reward. The behaviour is similar to the one state scenario. An observation is that the extra information seem to
hurt the agent's performance whenever stochastic agents are considered. This can be interpreted as extra information
introducing confusion to the classifier.
7
A PREPRINT - DECEMBER 16, 2019
Figure 12: 128 state Trajectory Classifier Loss
Classifier Loss against number of iterations.
Figure 13: 128 state Trajectory Reward
Reward against number of iterations.
4.2.3 One State vs 128 State Trajectories
The one state and 128 state trajectory approaches were compared and the result presented in fig 12 and fig 13. The
observation is that the 128 state trajectory performed better than one state trajectory when only deterministic opposing
agents are considered, even though variance was generally more. This implies that providing more information to the
agent does not significantly improve the performance which may be because of the simplicity of the environment. There
was no improvement when stochastic opposing agents were added, it is plausible to notice that extra information about
a stochastic opposing agent might prove to be much more confusing.
5 Ethics
The key ethical risk of the use of agent probing policies in a real-world environment is one of instrumentation causing
harm. The probing policy is trained to determine the opposing agents type by any means. The probing policy does not
take into account the welfare of either agent.
To make this more concrete, consider an autonomous driving scenario. In this scenario, we want to know if other drivers
on the road are distracted. A distracted driver could be defined as one who will not react swiftly to a dangerous driving
situation such as an obstacle in its path. In this scenario, the easiest way to determine whether a driver will react to a
dangerous situation is to create a dangerous situation and observe how the driver reacts. Likely, the probing agent will
simply drive directly at the opposing agent and see if the opposing agent swerves to avoid the probing agent.
We can see this behavior in our experiments where we have restricted the field of vision of the opposing agent. The
probing policy simply moves directly into the field of view of the opposing agent to see how it will react. In our
experiments, we have defined predator and prey opposing agents, and the probing policy does not consider the danger
of encountering a predator agent.
To solve this problem, the process for training the probing policy could be modified so that it receives negative reward if
it takes a dangerous action. However, a practical solution to this problem is saved for future work.
A secondary ethical issue of the probing policy relates to regularity and accountability. Consider the autonomous driving
scenario outlined above. For the reasons discussed, it may not be possible to safely determine whether another driver is
distracted or not. For this reason, the regularity of the classifiers ability to determine the opposing agent type is low.
This lack of regularity means that manufacturer of the autonomous driving system can't accept responsibility for the
safety of the system in the presence of distracted drivers. Thus, the manufacturer lacks accountability.
6 Conclusions and Future Work
6.1
Improvements with Current Experiments
We want to do the following improvements with our current experiment setup:
8
A PREPRINT - DECEMBER 16, 2019
1. In the work mentioned in this paper, we were able to show that our agent was indeed able to learn a probing
policy. It displayed behaviour of trying to go to a position in the environment to observe the reaction of the
opposing agent. However, we also noticed that the agent sometimes seemed to wander off the optimal trajectory
and take random steps. This is because we are only giving it a sparse reward at the end of the trajectory. In
our current experiments, the trajectory is of fixed length and therefore the agent does not learn to go to the
probing position as soon as possible. A simple fix to this problem can be providing the negative of the loss of
the classifier at every time step to the agent. This might help the agent learn to go to the probing position in the
environment optimally.
2. In the current experiments, we have a fixed number of time steps to execute our probing policy. If we want
to use the type of the opposing agent in the environment to actually help us in our task policy, we need to
have a done condition for our probing policy. If the classifier is able to classify the opposing agent as a certain
type of agent with certainty above a threshold percentage, we can stop executing our probing policy and start
executing our task policy. We can modify our current experiments to add such an early done condition.
6.2 Further Future Directions
In current experiments, we are giving the label of the opposing agent in the environment as a supervised target for
calculating loss and training the classifier and the API policies. In real world scenarios, we would ideally want our agent
to form a belief about the opposing agent in the environment and use this belief to condition its task policy. We would
want our agent to learn such a belief about the opposing agent without supervision. A promising future direction can be
coming up with an architecture for such end to end training of API policies.
References
[1] Ardi Tampuu, Tambet Matiisen, Dorian Kodelja, Ilya Kuzovkin, Kristjan Korjus, Juhan Aru, Jaan Aru, and Raul
Vicente. Multiagent cooperation and competition with deep reinforcement learning. CoRR, abs/1511.08779, 2015.
[2] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed
cooperative-competitive environments. CoRR, abs/1706.02275, 2017.
[3] Jakob N. Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. Counterfac-
tual multi-agent policy gradients. CoRR, abs/1705.08926, 2017.
[4] Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinícius Flores Zambaldi, Max
Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z. Leibo, Karl Tuyls, and Thore Graepel. Value-decomposition
networks for cooperative multi-agent learning. CoRR, abs/1706.05296, 2017.
[5] Tabish Rashid, Mikayel Samvelyan, Christian Schröder de Witt, Gregory Farquhar, Jakob N. Foerster, and Shimon
Whiteson. QMIX: monotonic value function factorisation for deep multi-agent reinforcement learning. CoRR,
abs/1803.11485, 2018.
[6] Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learning multiagent communication with backpropagation.
2016.
[7] Yedid Hoshen. Vain: Attentional multi-agent predictive modeling, 2017.
[8] Jakob N. Foerster, Yannis M. Assael, Nando de Freitas, and Shimon Whiteson. Learning to communicate with
deep multi-agent reinforcement learning, 2016.
[9] Igor Mordatch and Pieter Abbeel. Emergence of grounded compositional language in multi-agent populations,
2017.
[10] Jakob N. Foerster, Richard Y. Chen, Maruan Al-Shedivat, Shimon Whiteson, Pieter Abbeel, and Igor Mordatch.
Learning with opponent-learning awareness, 2017.
[11] Ying Wen, Yaodong Yang, Rui Luo, Jun Wang, and Wei Pan. Probabilistic recursive reasoning for multi-agent
reinforcement learning. In International Conference on Learning Representations, 2019.
[12] Stefano V. Albrecht and Peter Stone. Autonomous agents modelling other agents: A comprehensive survey and
open problems. Artificial Intelligence, 258:66 -- 95, May 2018.
[13] Aditya Grover, Maruan Al-Shedivat, Jayesh K. Gupta, Yura Burda, and Harrison Edwards. Learning policy
representations in multiagent systems, 2018.
[14] Maruan Al-Shedivat, Trapit Bansal, Yuri Burda, Ilya Sutskever, Igor Mordatch, and Pieter Abbeel. Continuous
adaptation via meta-learning in nonstationary and competitive environments, 2017.
9
A PREPRINT - DECEMBER 16, 2019
[15] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep
networks, 2017.
[16] Jane X Wang, Zeb Kurth-Nelson, Dhruva Tirumala, Hubert Soyer, Joel Z Leibo, Remi Munos, Charles Blundell,
Dharshan Kumaran, and Matt Botvinick. Learning to reinforcement learn, 2016.
[17] Yan Duan, John Schulman, Xi Chen, Peter L. Bartlett, Ilya Sutskever, and Pieter Abbeel. Rl2: Fast reinforcement
learning via slow reinforcement learning, 2016.
[18] Tianhe Yu, Chelsea Finn, Annie Xie, Sudeep Dasari, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot
imitation from observing humans via domain-adaptive meta-learning, 2018.
[19] Wenhao Yu, Jie Tan, C. Karen Liu, and Greg Turk. Preparing for the unknown: Learning a universal policy with
online system identification, 2017.
[20] Wenxuan Zhou, Lerrel Pinto, and Abhinav Gupta. Environment probing interaction policies, 2019.
[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9:1735 -- 1780, 1997.
[22] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk,
and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation.
ArXiv, abs/1406.1078, 2014.
[23] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech
Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016.
[24] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization
algorithms, 2017.
10
|
1205.4220 | 2 | 1205 | 2013-05-05T22:42:36 | Diffusion Adaptation over Networks | [
"cs.MA",
"cs.LG"
] | Adaptive networks are well-suited to perform decentralized information processing and optimization tasks and to model various types of self-organized and complex behavior encountered in nature. Adaptive networks consist of a collection of agents with processing and learning abilities. The agents are linked together through a connection topology, and they cooperate with each other through local interactions to solve distributed optimization, estimation, and inference problems in real-time. The continuous diffusion of information across the network enables agents to adapt their performance in relation to streaming data and network conditions; it also results in improved adaptation and learning performance relative to non-cooperative agents. This article provides an overview of diffusion strategies for adaptation and learning over networks. The article is divided into several sections: 1. Motivation; 2. Mean-Square-Error Estimation; 3. Distributed Optimization via Diffusion Strategies; 4. Adaptive Diffusion Strategies; 5. Performance of Steepest-Descent Diffusion Strategies; 6. Performance of Adaptive Diffusion Strategies; 7. Comparing the Performance of Cooperative Strategies; 8. Selecting the Combination Weights; 9. Diffusion with Noisy Information Exchanges; 10. Extensions and Further Considerations; Appendix A: Properties of Kronecker Products; Appendix B: Graph Laplacian and Network Connectivity; Appendix C: Stochastic Matrices; Appendix D: Block Maximum Norm; Appendix E: Comparison with Consensus Strategies; References. | cs.MA | cs |
DIFFUSION ADAPTATION OVER NETWORKS∗
Ali H. Sayed
Electrical Engineering Department
University of California at Los Angeles
Adaptive networks are well-suited to perform decentralized information processing and optimization tasks
and to model various types of self-organized and complex behavior encountered in nature. Adaptive networks
consist of a collection of agents with processing and learning abilities. The agents are linked together through
a connection topology, and they cooperate with each other through local interactions to solve distributed
optimization, estimation, and inference problems in real-time. The continuous diffusion of information across
the network enables agents to adapt their performance in relation to streaming data and network conditions;
it also results in improved adaptation and learning performance relative to non-cooperative agents. This
article provides an overview of diffusion strategies for adaptation and learning over networks. The article is
divided into several sections:
1. Motivation.
2. Mean-Square-Error Estimation.
3. Distributed Optimization via Diffusion Strategies.
4. Adaptive Diffusion Strategies.
5. Performance of Steepest-Descent Diffusion Strategies.
6. Performance of Adaptive Diffusion Strategies.
7. Comparing the Performance of Cooperative Strategies.
8. Selecting the Combination Weights.
9. Diffusion with Noisy Information Exchanges.
10. Extensions and Further Considerations.
11. Appendix A: Properties of Kronecker Products.
12. Appendix B: Graph Laplacian and Network Connectivity.
13. Appendix C: Stochastic Matrices.
14. Appendix D: Block Maximum Norm.
15. Appendix E: Comparison with Consensus Strategies.
16. References.
∗The cite information for this article is as follows: A. H. Sayed, “Diffusion adaptation over networks,” in E-Reference
The work was supported in part by
Signal Processing, R. Chellapa and S. Theodoridis, editors, Elsevier, 2013.
NSF grants EECS-060126, EECS-0725441, CCF-0942936, and CCF-1011918. The author is with the Electrical Engineering
Department, University of California, Los Angeles, CA 90095, USA. Email: [email protected].
1
1 Motivation
Consider a collection of N agents interested in estimating the same parameter vector, wo , of size M × 1. The
vector is the minimizer of some global cost function, denoted by J glob (w), which the agents seek to optimize,
say,
wo = argmin
w
J glob (w)
(1)
We are interested in situations where the individual agents have access to partial information about the
global cost function.
In this case, cooperation among the agents becomes beneficial. For example, by
cooperating with their neighbors, and by having these neighbors cooperate with their neighbors, procedures
can be devised that would enable all agents in the network to converge towards the global optimum wo
through local interactions. The ob jective of decentralized processing is to allow spatially distributed agents
to achieve a global ob jective by relying solely on local information and on in-network processing. Through a
continuous process of cooperation and information sharing with neighbors, agents in a network can be made
to approach the global performance level despite the localized nature of their interactions.
1.1 Networks and Neighborhoods
In this article we focus mainly on connected networks, although many of the results hold even if the network
graph is separated into disjoint subgraphs.
In a connected network, if we pick any two arbitrary nodes,
then there will exist at least one path connecting them: the nodes may be connected directly by an edge
if they are neighbors, or they may be connected by a path that passes through other intermediate nodes.
Figure 1 provides a graphical representation of a connected network with N = 10 nodes. Nodes that are
able to share information with each other are connected by edges. The sharing of information over these
edges can be unidirectional or bi-directional. The neighborhood of any particular node is defined as the set
of nodes that are connected to it by edges; we include in this set the node itself. The figure illustrates the
neighborhood of node 3, which consists of the following subset of nodes: N3 = {1, 2, 3, 5}. For each node,
the size of its neighborhood defines its degree. For example, node 3 in the figure has degree N3 = 4, while
node 8 has degree N8 = 5. Nodes that are well connected have higher degrees. Note that we are denoting
the neighborhood of an arbitrary node k by Nk and its size by Nk . We shall also use the notation nk to
refer to Nk .
The neighborhood of any node k therefore consists of all nodes with which node k can exchange informa-
tion. We assume a symmetric situation in relation to neighbors so that if node k is a neighbor of node ℓ, then
node ℓ is also a neighbor of node k . This does not necessarily mean that the flow of information between
these two nodes is symmetrical. For instance, in future sections, we shall assign pairs of nonnegative weights
to each edge connecting two neighboring nodes — see Fig. 2. In particular, we will assign the coefficient aℓk
to denote the weight used by node k to scale the data it receives from node ℓ; this scaling can be interpreted
as a measure of trustworthiness or reliability that node k assigns to its interaction with node ℓ. Note that we
are using two subscripts, ℓk , with the first subscript denoting the source node (where information originates
from) and the second subscript denoting the sink node (where information moves to) so that:
aℓk ≡ aℓ → k
(information flowing from node ℓ to node k)
(2)
In this way, the alternative coefficient akℓ will denote the weight used to scale the data sent from node k to
ℓ:
akℓ ≡ ak → ℓ
(information flowing from node k to node ℓ)
(3)
The weights {akℓ , aℓk } can be different, and one or both of them can be zero, so that the exchange of
information over the edge connecting the neighboring nodes (k , ℓ) need not be symmetric. When one of the
2
10
8
1
9
7
5
6
3
2
4
neighborhood
of node 3
Figure 1: A network consists of a collection of cooperating nodes. Nodes that are linked by edges can share
information. The neighborhood of any particular node consists of all nodes that are connected to it by edges
(including the node itself ). The figure illustrates the neighborhood of node 3, which consists of nodes {1, 2, 3, 5}.
Accordingly, node 3 has degree 4, which is the size of its neighborhood.
weights is zero, say, akℓ = 0, then this situation means that even though nodes (k , ℓ) are neighbors, node
ℓ is either not receiving data from node k or the data emanating from node k is being annihilated before
reaching node ℓ. Likewise, when akk > 0, then node k scales its own data, whereas akk = 0 corresponds
to the situation when node k does not use its own data. Usually, in graphical representations like those in
Fig. 1, edges are drawn between neighboring nodes that can share information. And, it is understood that
the actual sharing of information is controlled by the values of the scaling weights that are assigned to the
edge; these values can turn off communication in one or both directions and they can also scale one direction
more heavily than the reverse direction, and so forth.
1.2 Cooperation Among Agents
Now, depending on the application under consideration, the solution vector wo from (1) may admit different
interpretations. For example, the entries of wo may represent the location coordinates of a nutrition source
that the agents are trying to find, or the location of an accident involving a dangerous chemical leak.
The nodes may also be interested in locating a predator and tracking its movements over time. In these
localization applications, the vector wo is usually two or three-dimensional. In other applications, the entries
of wo may represent the parameters of some model that the network wishes to learn, such as identifying
the model parameters of a biological process or the occupied frequency bands in a shared communications
medium. There are also situations where different agents in the network may be interested in estimating
different entries of wo , or even different parameter vectors wo altogether, say, {wo
k } for node k , albeit with
some relation among the different vectors so that cooperation among the nodes can still be rewarding. In
this article, however, we focus exclusively on the special (yet frequent and important) case where all agents
are interested in estimating the same parameter vector wo .
Since the agents have a common ob jective, it is natural to expect cooperation among them to be beneficial
in general. One important question is therefore how to develop cooperation strategies that can lead to
better performance than when each agent attempts to solve the optimization problem individually. Another
important question is how to develop strategies that endow networks with the ability to adapt and learn in
3
akk
k
aℓk
akℓ
ℓ
aℓℓ
aℓ1k
akℓ1
aℓ2k
akℓ2
akk
k
aℓ3k
k
akk
akℓ3
Figure 2: In the top part, and for emphasis purposes, we are representing the edge between nodes k and ℓ by two
separate directed links: one moving from k to ℓ and the other moving from ℓ to k. Two nonnegative weights are used
to scale the sharing of information over these directed links. The scalar akℓ denotes the weight used to scale data
sent from node k to ℓ, while aℓk denotes the weight used to scale data sent from node ℓ to k. The weights {ak,ℓ , aℓk }
can be different, and one or both of them can be zero, so that the exchange of information over the edge connecting
any two neighboring nodes need not be symmetric. The bottom part of the figure illustrates directed links arriving
to node k from its neighbors {ℓ1 , ℓ2 , ℓ3 , . . .} (left) and leaving from node k towards these same neighbors (right).
real-time in response to changes in the statistical properties of the data. This article provides an overview
of results in the area of diffusion adaptation with illustrative examples. Diffusion strategies are powerful
methods that enable adaptive learning and cooperation over networks. There have been other useful works
in the literature on the use of alternative consensus strategies to develop distributed optimization solutions
over networks. Nevertheless, we explain in App. E why diffusion strategies outperform consensus strategies
in terms of their mean-square-error stability and performance. For this reason, we focus in the body of the
chapter on presenting the theoretical foundations for diffusion strategies and their performance.
1.3 Notation
In our treatment, we need to distinguish between random variables and deterministic quantities. For this
reason, we use boldface letters to represent random variables and normal font to represent deterministic
(non-random) quantities. For example, the boldface letter d denotes a random quantity, while the normal
font letter d denotes an observation or realization for it. We also need to distinguish between matrices and
vectors. For this purpose, we use CAPITAL letters to refer to matrices and small letters to refer to both
vectors and scalars; Greek letters always refer to scalars. For example, we write R to denote a covariance
matrix and w to denote a vector of parameters. We also write σ2
v to refer to the variance of a random variable.
To distinguish between a vector d (small letter) and a scalar d (also a small letter), we use parentheses to
index scalar quantities and subscripts to index vector quantities. Thus, we write d(i) to refer to the value of
a scalar quantity d at time i, and di to refer to the value of a vector quantity d at time i. Thus, d(i) denotes
a scalar while di denotes a vector. All vectors in our presentation are column vectors, with the exception
of the regression vector (denoted by the letter u), which will be taken to be a row vector for convenience
of presentation. The symbol T denotes transposition, and the symbol ∗ denotes complex conjugation for
scalars and complex-conjugate transposition for matrices. The notation col{a, b} denotes a column vector
4
with entries a and b stacked on top of each other, and the notation diag{a, b} denotes a diagonal matrix
with entries a and b. Likewise, the notation vec(A) vectorizes its matrix argument and stacks the columns
of A on top of each other. The notation kxk denotes the Euclidean norm of its vector argument, while
kxkb,∞ denotes the block maximum norm of a block vector (defined in App. D). Similarly, the notation
Σ denotes the weighted square value, x∗Σx. Moreover, kAkb,∞ denotes the block maximum norm of a
kxk2
matrix (also defined in App. D), and ρ(A) denotes the spectral radius of the matrix (i.e., the largest absolute
magnitude among its eigenvalues). Finally, IM denotes the identity matrix of size M × M ; sometimes, for
simplicity of notation, we drop the subscript M from IM when the size of the identity matrix is obvious from
the context. Table 1 provides a summary of the symbols used in the article for ease of reference.
Table 1: Summary of notation conventions used in the article.
Boldface notation denotes random variables.
d
Normal font denotes realizations of random variables.
d
Capital letters denote matrices.
A
Small letters denote vectors or scalars.
a
Greek letters denote scalars.
α
Small letters with parenthesis denote scalars.
d(i)
Small letters with subscripts denote vectors.
di
Matrix transposition.
T
Complex conjugation for scalars and complex-conjugate transposition for matrices.
∗
col{a, b}
Column vector with entries a and b.
diag{a, b} Diagonal matrix with entries a and b.
vec(A)
Vectorizes matrix A and stacks its columns on top of each other.
Euclidean norm of its vector argument.
kxk
Weighted square value x∗Σx.
kxk2
Σ
Block maximum norm of a block vector – see App. D.
kxkb,∞
Block maximum norm of a matrix – see App. D.
kAkb,∞
kAk
2−induced norm of matrix A (its largest singular value).
Spectral radius of its matrix argument.
ρ(A)
IM
Identity matrix of size M × M ; sometimes, we drop the subscript M .
2 Mean-Square-Error Estimation
Readers interested in the development of the distributed optimization strategies and their adaptive versions
can move directly to Sec. 3. The purpose of the current section is to motivate the virtues of distributed
in-network processing, and to provide illustrative examples in the context of mean-square-error estimation.
Advanced readers can skip this section on a first reading.
We start our development by associating with each agent k an individual cost (or utility) function,
Jk (w). Although the algorithms presented in this article apply to more general situations, we nevertheless
assume in our presentation that the cost functions Jk (w) are strictly convex so that each one of them has
a unique minimizer. We further assume that, for all costs Jk (w), the minimum occurs at the same value
wo . Obviously, the choice of Jk (w) is limitless and is largely dependent on the application. It is sufficient
for our purposes to illustrate the main concepts underlying diffusion adaptation by focusing on the case of
mean-square-error (MSE) or quadratic cost functions. In the sequel, we provide several examples to illustrate
how such quadratic cost functions arise in applications and how cooperative processing over networks can be
beneficial. At the same time, we note that most of the arguments in this article can be extended beyond MSE
optimization to more general cost functions and to situations where the minimizers of the individual costs
Jk (w) need not agree with each other — as already shown in [1–3]; see also Sec. 10.4 for a brief summary.
In non-cooperative solutions, each agent would operate individually on its own cost function Jk (w) and
5
optimize it to determines wo , without any interaction with the other nodes. However, the analysis and
derivations in future sections will reveal that nodes can benefit from cooperation between them in terms of
better performance (such as converging faster to wo or tracking a changing wo more effectively) — see, e.g.,
Theorems 6.3–6.5 and Sec. 7.3.
2.1 Application: Autoregressive Modeling
dk (i) =
βmdk (i − m) + vk (i)
Our first example relates to identifying the parameters of an auto-regressive (AR) model from noisy data.
Thus, consider a situation where agents are spread over some geographical region and each agent k is
observing realizations {dk (i)} of an AR zero-mean random process {dk (i)}, which satisfies a model of the
form:
MXm=1
The scalars {βm} represent the model parameters that the agents wish to identify, and vk (i) represents an
additive zero-mean white noise process with power:
∆= E vk (i)2
It is customary to assume that the noise process is temporally white and spatially independent so that noise
terms across different nodes are independent of each other and, at the same node, successive noise samples
are also independent of each other with a time-independent variance σ2
v,k :
(cid:26) E vk (i)v∗
k (j ) = 0,
E vk (i)v∗
m (j ) = 0,
The noise process vk (i) is further assumed to be independent of past signals {dℓ (i − m), m ≥ 1} across all
nodes ℓ. Observe that we are allowing the noise power profile, σ2
v,k , to vary with k . In this way, the quality
of the measurements is allowed to vary across the network with some nodes collecting noisier data than other
nodes. Figure 3 illustrates an example of a network consisting of N = 10 nodes spread over a region in space.
The figure shows the neighborhood of node 2, which consists of nodes {1, 2, 3}.
for all i 6= j (temporal whiteness)
for all i, j whenever k 6= m (spatial whiteness)
σ2
v,k
(4)
(5)
(6)
Linear Model
To illustrate the difference between cooperative and non-cooperative estimation strategies, let us first explain
that the data can be represented in terms of a linear model. To do so, we collect the model parameters {βm}
into an M × 1 column vector wo :
wo ∆= col {β1 , β2 , . . . , βM }
and the past data into a 1 × M row regression vector uk,i :
∆= (cid:2) dk (i − 1) dk (i − 2)
. . . dk (i − M ) (cid:3)
Then, we can rewrite the measurement equation (4) at each node k in the equivalent linear model form:
dk (i) = uk,iwo + vk (i)
uk,i
(7)
(8)
(9)
Linear relations of the form (9) are common in applications and arise in many other contexts (as further
illustrated by the next three examples in this section).
We assume the stochastic processes {dk (i), uk,i } in (9) have zero means and are jointly wide-sense sta-
tionary. We denote their second-order moments by:
∆= E dk (i)2
∆= E u∗
k,iuk,i
∆= E dk (i)u∗
k,i
(M × M )
(M × 1)
(scalar)
σ2
d,k
rdu,k
Ru,k
(10)
(11)
(12)
6
amplitude
10
9
8
7
d1 (i)
1
4
d2 (i)
2
6
5
3
d3 (i)
a t e
o r d i n
o
x - c
y
-
c
o
o
r
d
i
n
a
t
e
Figure 3: A collection of nodes, spread over a geographic region, observes realizations of an AR random process
and cooperates to estimate the underlying model parameters {βm } in the presence of measurement noise. The noise
power profile can vary over space.
As was the case with the noise power profile, we are allowing the moments {σ2
d,k , Ru,k , rdu,k } to depend
on the node index k so that these moments can vary with the spatial dimension as well. The covariance
matrix Ru,k is, by definition, always non-negative definite. However, for convenience of presentation, we
shall assume that Ru,k is actually positive-definite (and, hence, invertible):
Ru,k > 0
(13)
Non-Cooperative Mean-Square-Error Solution
One immediate result that follows from the linear model (9) is that the unknown parameter vector wo can
be recovered exactly by each individual node from knowledge of the local moments {rdu,k , Ru,k } alone. To
see this, note that if we multiply both sides of (9) by u∗
k,i and take expectations we obtain
= (cid:0)E u∗
k,iuk,i (cid:1)
wo + E u∗
E u∗
k,i vk (i)
k,idk (i)
{z
}
}
{z
{z
}
rdu,k
=0
Ru,k
rdu,k = Ru,k wo ⇐⇒ wo = R−1
u,k rdu,k
so that
(14)
(15)
It is seen from (15) that wo is the solution to a linear system of equations and that this solution can be
computed by every node directly from its moments {Ru,k , rdu,k }. It is useful to re-interpret construction
7
(15) as the solution to a minimum mean-square-error (MMSE) estimation problem [4, 5].
Indeed, it can
be verified that the quantity R−1
u,k rdu,k that appears in (15) is the unique solution to the following MMSE
problem:
E dk (i) − uk,iw2
min
w
To verify this claim, we denote the cost function that appears in (16) by
Jk (w)
∆= E dk (i) − uk,iw2
and expand it to find that
d,k − w∗ rdu,k − r∗
du,kw + w∗Ru,kw
Jk (w) = σ2
(16)
(17)
(18)
The cost function Jk (w) is quadratic in w and it has a unique minimizer since Ru,k > 0. Differentiating
Jk (w) with respect to w we find its gradient vector:
∇w J (w) = (Ru,kw − rdu,k )∗
(19)
It is seen that this gradient vector is annihilated at the same value w = wo given by (15). Therefore, we can
equivalently state that if each node k solves the MMSE problem (16), then the solution vector coincides with
the desired parameter vector wo . This observation explains why it is often justified to consider mean-square-
error cost functions when dealing with estimation problems that involve data that satisfy linear models
similar to (9).
Besides wo , the solution of the MMSE problem (16) also conveys information about the noise level in
the data. Note that by substituting wo from (15) into expression (16) for Jk (w), the resulting minimum
mean-square-error value that is attained by node k is found to be:
MSEk
∆= Jk (wo )
= E dk (i) − uk,iwo 2
(9)
= E v k (i)2
= σ2
v,k
∆= Jk,min
(20)
We shall use the notation Jk (wo ) and Jk,min interchangeably to denote the minimum cost value of Jk (w).
The above result states that, when each agent k recovers wo from knowledge of its moments {Ru,k , rdu,k }
using expression (15), then the agent attains an MSE performance level that is equal to the noise power at
its location, σ2
v,k . An alternative useful expression for the minimum cost can be obtained by substituting
expression (15) for wo into (18) and simplifying the expression to find that
du,kR−1
d,k − r∗
MSEk = σ2
u,k rdu,k
(21)
This second expression is in terms of the data moments {σ2
d,k , rdu,k , Ru,k }.
Non-Cooperative Adaptive Solution
The optimal MMSE implementation (15) for determining wo requires knowledge of the statistical information
{rdu,k , Ru,k }. This information is usually not available beforehand. Instead, the agents are more likely to have
access to successive time-indexed observations {dk (i), uk,i} of the random processes {dk (i), uk,i } for i ≥ 0.
In this case, it becomes necessary to devise a scheme that would allow each node to use its measurements
to approximate wo . It turns out that an adaptive solution is possible. In this alternative implementation,
8
each node k feeds its observations {dk (i), uk,i } into an adaptive filter and evaluates successive estimates for
wo . As time passes by, the estimates would get closer to wo .
The adaptive solution operates as follows. Let wk,i denote an estimate for wo that is computed by node
k at time i based on all the observations {dk (j ), uk,j , j ≤ i} it has collected up to that time instant. There
are many adaptive algorithms that can be used to compute wk,i ; some filters are more accurate than others
(usually, at the cost of additional complexity) [4–7]. It is sufficient for our purposes to consider one simple
(yet effective) filter structure, while noting that most of the discussion in this article can be extended to
other structures. One of the simplest choices for an adaptive structure is the least-mean-squares (LMS)
filter, where the data are processed by each node k as follows:
ek (i) = dk (i) − uk,iwk,i−1
wk,i = wk,i−1 + µku∗
k,i ek (i),
i ≥ 0
(22)
(23)
Starting from some initial condition, say, wk,−1 = 0, the filter iterates over i ≥ 0. At each time instant, i,
the filter uses the local data {dk (i), uk,i } at node k to compute a local estimation error, ek (i), via (22). The
error is then used to update the existing estimate from wk,i−1 to wk,i via (23). The factor µk that appears in
(23) is a constant positive step-size parameter; usually chosen to be sufficiently small to ensure mean-square
stability and convergence, as discussed further ahead in the article. The step-size parameter can be selected
to vary with time as well; one popular choice is to replace µk in (23) with the following construction:
µk (i) ∆=
µk
ǫ + kuk,ik2
(24)
µ(i) = ∞,
(25)
µ2 (i) < ∞
where ǫ > 0 is a small positive parameter and µk > 0. The resulting filter implementation is known as
normalized LMS [5] since the step-size is normalized by the squared norm of the regression vector. Other
choices include step-size sequences {µ(i) ≥ 0} that satisfy both conditions:
∞Xi=0
∞Xi=0
Such sequences converge slowly towards zero. One example is the choice µk (i) = µk
i+1 . However, by virtue of
the fact that such step-sizes die out as i → ∞, then these choices end up turning off adaptation. As such,
step-size sequences satisfying (25) are not generally suitable for applications that require continuous learning,
especially under non-stationary environments. For this reason, in this article, we shall focus exclusively on
the constant step-size case (23) in order to ensure continuous adaptation and learning.
Equations (22)–(23) are written in terms of the observed quantities {dk (i), uk,i }; these are deterministic
values since they correspond to observations of the random processes {dk (i), uk,i }. Often, when we are
interested in highlighting the random nature of the quantities involved in the adaptation step, especially
when we study the mean-square performance of adaptive filters, it becomes more useful to rewrite the
recursions using the boldface notation to highlight the fact that the quantities that appear in (22)–(23) are
actually realizations of random variables. Thus, we also write:
ek (i) = dk (i) − uk,iwk,i−1
wk,i = wk,i−1 + µk u∗
k,i ek (i),
i ≥ 0
(26)
(27)
where {ek (i), wk,i } will be random variables as well.
The performance of adaptive implementations of this kind are well-understood for both cases of stationary
wo and changing wo [4–7]. For example, in the stationary case, if the adaptive implementation (26)–(27)
were to succeed in having its estimator wk,i tend to wo with probability one as i → ∞, then we would
expect the error signal ek (i) in (26) to tend towards the noise signal vk (i) (by virtue of the linear model (9)).
This means that, under this ideal scenario, the variance of the error signal ek (i) would be expected to tend
towards the noise variance, σ2
v,k , as i → ∞. Recall from (20) that the noise variance is the least cost that
the MSE solution can attain. Therefore, such limiting behavior by the adaptive filter would be desirable.
9
However, it is well-known that there is always some loss in mean-square-error performance when adaptation
is employed due to the effect of gradient noise, which is caused by the algorithm’s reliance on observations
(or realizations) {dk (i), uk,i } rather than on the actual moments {rdu,k , Ru,k }. In particular, it is known
that for LMS filters, the variance of ek (i) in steady-state will be larger than σ2
v,k by a small amount, and
the size of the offset is proportional to the step-size parameter µk (so that smaller step-sizes lead to better
mean-square-error (MSE) performance albeit at the expense of slower convergence). It is easy to see why
the variance of ek (i) will be larger than σ2
v,k from the definition of the error signal in (26). Introduce the
weight-error vector
∆= wo − wk,i
and the so-called a-priori error signal
ewk,i
ea,k (i) ∆= uk,i ewk,i−1
This second error measures the difference between the uncorrupted term uk,iwo and its estimator prior to
adaptation, uk,iwk,i−1 . It then follows from the data model (9) and from the defining expression (26) for
ek (i) that
(29)
(28)
ek (i) = dk (i) − uk,iwk,i−1
= uk,iwo + vk (i) − uk,iwk,i−1
= ea,k (i) + vk (i)
(30)
Since the noise component, v k (i), is assumed to be zero-mean and independent of all other random variables,
we conclude that
E ek (i)2 = E ea,k (i)2 + σ2
v,k
(31)
This relation holds for any time instant i; it shows that the variance of the output error, ek (i), is larger
than σ2
v,k by an amount that is equal to the variance of the a-priori error, ea,k (i). We define the filter
mean-square-error (MSE) and excess-mean-square-error (EMSE) as the following steady-state measures:
MSEk
EMSEk
∆= lim
i→∞
∆= lim
i→∞
E ek (i)2
E ea,k (i)2
so that, for the adaptive implementation (compare with (20)):
MSEk = EMSEk + σ2
v,k
(32)
(33)
(34)
Therefore, the EMSE term quantifies the size of the offset in the MSE performance of the adaptive filter.
We also define the filter mean-square-deviation (MSD) as the steady-state measure:
E k ewk,i k2
which measures how far wk,i is from wo in the mean-square-error sense in steady-state. It is known that the
MSD and EMSE of LMS filters of the form (26)–(27) can be approximated for sufficiently small-step sizes
by the following expressions [4–7]:
∆= lim
i→∞
MSDk
(35)
EMSEk ≈ µk σ2
v,kTr(Ru,k )/2
MSDk ≈ µk σ2
v,kM /2
10
(36)
(37)
It is seen that the smaller the step-size parameter, the better the performance of the adaptive solution.
Cooperative Adaptation through Diffusion
Observe from (36)–(37) that even if all nodes employ the same step-size, µk = µ, and even if the regression
data are spatially uniform so that Ru,k = Ru for all k , the mean-square-error performance across the nodes
still varies in accordance with the variation of the noise power profile, σ2
v,k , across the network. Nodes with
larger noise power will perform worse than nodes with smaller noise power. However, since all nodes are
observing data arising from the same underlying model wo , it is natural to expect cooperation among the
nodes to be beneficial. By cooperation we mean that neighboring nodes can share information (such as
measurements or estimates) with each other as permitted by the network topology. Starting in the next
section, we will motivate and describe algorithms that enable nodes to carry out adaptation and learning in
a cooperative manner to enhance performance.
Specifically, we are going to see that one way to achieve cooperation is by developing adaptive algorithms
that enable the nodes to optimize the following global cost function in a distributed manner:
min
w
E dk (i) − uk,iw2
J glob (w) ∆=
E dk (i) − uk,iw2 =
NXk=1
where the global cost is the aggregate ob jective:
NXk=1
NXk=1
Comparing (38) with (16), we see that we are now adding the individual costs, Jk (w), from across all nodes.
Note that since the desired wo satisfies (15) at every node k , then it also satisfies
MXk=1
Ru,k! wo =
NXk=1
But it can be verified that the optimal solution to (38) is given by the same wo that satisfies (40). Therefore,
solving the global optimization problem (38) also leads to the desired wo . In future sections, we will show
how cooperative and distributed adaptive schemes for solving (38), such as (153) or (154) further ahead, lead
to improved performance in estimating wo (in terms of smaller mean-square-deviation and faster convergence
rate) than the non-cooperative mode (26)–(27), where each agent runs its own individual adaptive filter —
see, e.g., Theorems 6.3–6.5 and Sec. 7.3.
Jk (w)
(38)
(39)
(40)
rdu,k
2.2 Application: Tapped-Delay-Line Models
Our second example to motivate MSE cost functions, Jk (w), and linear models relates to identifying the
parameters of a moving-average (MA) model from noisy data. MA models are also known as finite-impulse-
response (FIR) or tapped-delay-line models. Thus, consider a situation where agents are interested in
estimating the parameters of an FIR model, such as the taps of a communications channel or the parameters
of some (approximate) model of interest in finance or biology. Assume the agents are able to independently
probe the unknown model and observe its response to excitations in the presence of additive noise; this
situation is illustrated in Fig. 4, with the probing operation highlighted for one of the nodes (node 4).
The schematics inside the enlarged diagram in Fig. 4 is meant to convey that each node k probes the
model with an input sequence {uk (i)} and measures the resulting response sequence, {dk (i)}, in the presence
11
10
8
1
9
7
5
6
3
4
2
wo
v4 (i)
noise
u4 (i)
probing
signal
4
d4 (i)
observation
Figure 4: The network is interested in estimating the parameter vector wo that describes an underlying tapped-
delay-line model. The agents are assumed to be able to independently probe the unknown system, and observe its
response to excitations, under noise, as indicated in the highlighted diagram for node 4.
of additive noise. The system dynamics for each agent k is assumed to be described by a MA model of the
form:
M −1Xm=0
In this model, the term vk (i) again represents an additive zero-mean noise process that is assumed to be
temporally white and spatially independent; it is also assumed to be independent of the input process,
{uℓ (j )}, for all i, j and ℓ. The scalars {βm} represent the model parameters that the agents seek to identify.
If we again collect the model parameters into an M × 1 column vector wo :
βmuk (i − m) + vk (i)
dk (i) =
(41)
wo ∆= col {β0 , β1 , . . . , βM −1}
(42)
and the input data into a 1 × M row regression vector:
. . . uk (i − M + 1) (cid:3)
∆= (cid:2) uk (i) uk (i − 1)
then, we can again express the measurement equation (41) at each node k in the same linear model as (9),
namely,
(43)
uk,i
dk (i) = uk,iwo + vk (i)
(44)
As was the case with model (9), we can likewise verify that, in view of (44), the desired parameter vector
wo satisfies the same normal equations (15), i.e.,
rdu,k = Ru,k wo ⇐⇒ wo = R−1
u,k rdu,k
(45)
where the moments {rdu,k , Ru,k } continue to be defined by expressions (11)–(12) with uk,i now defined by
(43). Therefore, each node k can determine wo on its own by solving the same MMSE estimation problem
12
(16). This solution method requires knowledge of the moments {rdu,k , Ru,k } and, according to (20), each
agent k would then attain an MSE level that is equal to the noise power level at its location.
Alternatively, when the statistical information {rdu,k , Ru,k } is not available, each agent k can estimate
wo iteratively by feeding data {dk (i), uk,i } into the adaptive implementation (26)–(27). In this way, each
agent k will achieve the same performance level shown earlier in (36)–(37), with the limiting performance
being again dependent on the local noise power level, σ2
v,k . Therefore, nodes with larger noise power will
perform worse than nodes with smaller noise power. However, since all nodes are observing data arising from
the same underlying model wo , it is natural to expect cooperation among the nodes to be beneficial. As we
are going to see, starting from the next section, one way to achieve cooperation and improve performance
is by developing algorithms that optimize the same global cost function (38) in an adaptive and distributed
manner, such as algorithms (153) and (154) further ahead.
2.3 Application: Target Localization
Our third example relates to the problem of locating a destination of interest (such as the location of a
nutrition source or a chemical leak) or locating and tracking an ob ject of interest (such as a predator or a
pro jectile). In several such localization applications, the agents in the network are allowed to move towards
the target or away from it, in which case we would end up with a mobile adaptive network [8]. Biological
networks behave in this manner such as networks representing fish schools, bird formations, bee swarms,
bacteria motility, and diffusing particles [8–12]. The agents may move towards the target (e.g., when it is a
nutrition source) or away from the target (e.g., when it is a predator). In other applications, the agents may
remain static and are simply interested in locating a target or tracking it (such as tracking a pro jectile).
To motivate mean-square-error estimation in the context of localization problems, we start with the
situation corresponding to a static target and static nodes. Thus, assume that the unknown location of the
target in the cartesian plane is represented by the 2 × 1 vector wo = col{xo , y o}. The agents are spread over
the same region of space and are interested in locating the target. The location of every agent k is denoted
by the 2 × 1 vector pk = col{xk , yk } in terms of its x and y coordinates – see Fig. 5. We assume the agents
are aware of their location vectors. The distance between agent k and the target is denoted by ro
k and is
equal to:
The 1 × 2 unit-norm direction vector pointing from agent k towards the target is denoted by uo
k and is given
by:
ro
k = kwo − pk k
(46)
uo
k =
(wo − pk )T
kwo − pk k
Observe from (46) and (47) that ro
k can be expressed in the following useful inner-product form:
k (wo − pk )
k = uo
ro
(47)
(48)
In practice, agents have noisy observations of both their distance and direction vector towards the target.
We denote the noisy distance measurement collected by node k at time i by:
rk (i) = ro
k + v k (i)
(49)
where vk (i) denotes noise and is assumed to be zero-mean, and temporally white and spatially independent
with variance
σ2
v,k
∆= E vk (i)2
(50)
We also denote the noisy direction vector that is measured by node k at time i by uk,i . This vector is a
perturbed version of uo
k . We assume that uk,i continues to start from the location of the node at pk , but
13
target
(cid:1) xo
yo (cid:2)
distance to target
ro
k
unit direction
vector to target
uo
k
node k
(cid:1) xk
yk (cid:2)
y
(0, 0)
x
Figure 5: The distance from node k to the target is denoted by ro
k and the unit-norm direction vector from the same
node to the target is denoted by uo
k . Node k is assumed to have access to noisy measurements of {ro
k , uo
k }.
that its tip is perturbed slightly either to the left or to the right relative to the tip of uo
k — see Fig. 6. The
perturbation to the tip of uo
k is modeled as being the result of two effects: a small deviation that occurs
along the direction that is perpendicular to uo
k , and a smaller deviation that occurs along the direction of
uo
k . Since we are assuming that the tip of uk,i is only slightly perturbed relative to the tip of uo
k , then it is
reasonable to expect the amount of perturbation along the parallel direction to be small compared to the
amount of perturbation along the perpendicular direction.
Thus, we write
k + αk (i) uo⊥
k + βk (i) uo
uk,i = uo
k
(1 × 2)
(51)
where uo⊥
k denotes a unit-norm row vector that lies in the same plane and whose direction is perpendicular to
uo
k . The variables αk (i) and βk (i) denote zero-mean independent random noises that are temporally white
and spatially independent with variances:
∆= E βk (i)2
We assume the contribution of βk (i) is small compared to the contributions of the other noise sources, αk (i)
and v k (i), so that
∆= E αk (i)2 ,
σ2
α,k
σ2
β ,k
β ,k ≪ σ2
σ2
β ,k ≪ σ2
σ2
α,k ,
v,k
The random noises {vk (i), αk (i), βk (i)} are further assumed to be independent of each other.
Using (48) we find that the noisy measurements {rk (i), uk,i } are related to the unknown wo via:
where the modified noise term zk (i) is defined in terms of the noises in rk (i) and uk,i as follows:
rk (i) = uk,i (wo − pk ) + z k (i)
zk (i)
∆= vk (i) − αk (i) uo⊥
k
= vk (i) − βk (i) · ro
k
≈ vk (i)
· (wo − pk ) − βk (i) · uo
k · (wo − pk )
14
(52)
(53)
(54)
(55)
target
perturbed
direction
vector
unit-radius
circle
uk ,i
uo
k
node k
perpendicular direction
determined by uo⊥
k
actual direction
vector
Figure 6: The tip of the noisy direction vector is modeled as being approximately perturbed away from the actual
direction by two effects: a larger effect caused by a deviation along the direction that is perpendicular to uo
k , and a
smaller deviation along the direction that is parallel to uo
k .
since, by construction,
(56)
uo⊥
k
and the contribution by βk (i) is assumed to be sufficiently small. If we now introduce the adjusted signal:
· (wo − pk ) = 0
then we arrive again from (54) and (55) at the following linear model for the available measurement variables
{dk (i), uk,i } in terms of the target location wo :
dk (i) ∆= rk (i) + uk,i pk
(57)
dk (i) ≈ uk,iwo + vk (i)
(58)
There is one important difference in relation to the earlier linear models (9) and (44), namely, the variables
{dk (i), uk,i } in (58) do not have zero means any longer. It is nevertheless straightforward to determine the
first and second-order moments of the variables {dk (i), uk,i }. First, note from (49), (51), and (57) that
E uk,i = uo
k ,
k + uo
E dk (i) = ro
k pk
(59)
Even in this case of non-zero means, and in view of (58), the desired parameter vector wo can still be shown
to satisfy the same normal equations (15), i.e.,
where the moments {rdu,k , Ru,k } continue to be defined as
rdu,k = Ru,kwo ⇐⇒ wo = R−1
u,k rdu,k
Ru,k
∆= E u∗
k,iuk,i ,
rdu,k
∆= E u∗
k,idk (i)
(60)
(61)
To verify that (60) holds, we simply multiply both sides of (58) by u∗
k,i from the left, compute the expectations
of both sides, and use the fact that vk (i) has zero mean and is assumed to be independent of {uℓ,j } for
15
all times j and nodes ℓ. However, the difference in relation to the earlier normal equations (15) is that the
matrix Ru,k is not the actual covariance matrix of uk,i any longer. When uk,i is not zero mean, its covariance
matrix is instead defined as:
covu,k
∆= E (uk,i − uo
k )∗ (uk,i − uo
k )
= E u∗
k,iuk,i − uo∗
k uo
k
(62)
so that
(63)
where the matrix
Ru,k = covu,k + uo∗
k uo
k
We conclude from this relation that Ru,k is positive-definite (and, hence, invertible) so that expression (60)
is justified. This is because the covariance matrix, covu,k , is itself positive-definite. Indeed, some algebra
applied to the difference uk,i − uo
k from (51) shows that
(cid:21)
β ,k (cid:21) (cid:20) uo⊥
k )∗ (cid:3) (cid:20) σ2
covu,k = (cid:2) (cid:0)uo⊥
k (cid:1)∗
α,k
k
(uo
uo
σ2
k
(cid:20) uo⊥
(cid:21)
k
uo
k
k , uo⊥
is full rank since the rows {uo
k } are linearly independent vectors.
Therefore, each node k can determine wo on its own by solving the same minimum mean-square-error
estimation problem (16). This solution method requires knowledge of the moments {rdu,k , Ru,k } and, ac-
cording to (20), each agent k would then attain an MSE level that is equal to the noise power level, σ2
v,k , at
its location.
Alternatively, when the statistical information {rdu,k , Ru,k } is not available beforehand, each agent k can
estimate wo iteratively by feeding data {dk (i), uk,i } into the adaptive implementation (26)–(27). In this case,
each agent k will achieve the performance level shown earlier in (36)–(37), with the limiting performance
being again dependent on the local noise power level, σ2
v,k . Therefore, nodes with larger noise power will
perform worse than nodes with smaller noise power. However, since all nodes are observing distances and
direction vectors towards the same target location wo , it is natural to expect cooperation among the nodes
to be beneficial. As we are going to see, starting from the next section, one way to achieve cooperation and
improve performance is by developing algorithms that solve the same global cost function (38) in an adaptive
and distributed manner, by using algorithms such as (153) and (154) further ahead.
(64)
(65)
Role of Adaptation
The localization application helps highlight one of the main advantages of adaptation, namely, the ability
of adaptive implementations to learn and track changing statistical conditions. For example, in the context
of mobile networks, where nodes can move closer or further away from a target, the location vector for each
agent k becomes time-dependent, say, pk,i = col{xk (i), yk (i)}. In this case, the actual distance and direction
vector between agent k and the target also vary with time and become:
ro
k (i) = kwo − pk,i k,
uo
k,i =
(wo − pk,i )T
kwo − pk,i k
The noisy distance measurement to the target is then:
where the variance of vk (i) now depends on time as well:
rk (i) = ro
k (i) + vk (i)
σ2
v,k (i)
∆= E v k (i)2
16
(66)
(67)
(68)
In the context of mobile networks, it is reasonable to assume that the variance of vk (i) varies both with time
and with the distance to the target: the closer the node is to the target, the less noisy the measurement of
the distance is expected to be. Similar remarks hold for the variances of the noises αk (i) and βk (i) that
perturb the measurement of the direction vector, say,
where now
α,k (i) ∆= E αk (i)2 ,
σ2
β ,k (i) ∆= E βk (i)2
σ2
k,i + αk (i) uo⊥
uk,i = uo
k,i + βk (i) uo
k,i
(69)
(70)
The same arguments that led to (58) can be repeated to lead to the same model, except that now the means
of the variables {dk (i), uk,i } become time-dependent as well:
E uk,i = uo
k,i ,
E dk (i) = ro
k (i) + uo
k,i pk,i
(71)
Nevertheless, adaptive solutions (whether cooperative or non-cooperative), are able to track such time-
variations because these solutions work directly with the observations {dk (i), uk,i} and the successive ob-
servations will reflect the changing statistical profile of the data. In general, adaptive solutions are able to
track changes in the underlying signal statistics rather well [4, 5], as long as the rate of non-stationarity is
slow enough for the filter to be able to follow the changes.
2.4 Application: Collaborative Spectral Sensing
Our fourth and last example to illustrate the role of mean-square-error estimation and cooperation relates
to spectrum sensing for cognitive radio applications. Cognitive radio systems involve two types of users:
primary users and secondary users. To avoid causing harmful interference to incumbent primary users,
unlicensed cognitive radio devices need to detect unused frequency bands even at low signal-to-noise (SNR)
conditions [13–16]. One way to carry out spectral sensing is for each secondary user to estimate the aggregated
power spectrum that is transmitted by all active primary users, and to locate unused frequency bands within
the estimated spectrum. This step can be performed by the secondary users with or without cooperation.
Thus, consider a communications environment consisting of Q primary users and N secondary users. Let
Sq (ejω ) denote the power spectrum of the signal transmitted by primary user q . To facilitate estimation
of the spectral profile by the secondary users, we assume that each Sq (ejω ) can be represented as a linear
combination of basis functions, {fm(ejω )}, say, B of them [17]:
BXm=1
In this representation, the scalars {βqm} denote the coefficients of the basis expansion for user q . The variable
ω ∈ [−π , π ] denotes the normalized angular frequency measured in radians/sample. The power spectrum is
often symmetric about the vertical axis, ω = 0, and therefore it is sufficient to focus on the interval ω ∈ [0, π ].
There are many ways by which the basis functions, {fm (ejω )}, can be selected. The following is one possible
construction for illustration purposes. We divide the interval [0, π ] into B identical intervals and denote
their center frequencies by {ωm}. We then place a Gaussian pulse at each location ωm and control its width
through the selection of its standard deviation, σm , i.e.,
βqm fm (ejω ),
Sq (ejω ) =
q = 1, 2, . . . , Q
(72)
fm (ejω ) ∆= e
− (ω−ωm )2
σ2
m
(73)
Figure 7 illustrates this construction. The parameters {ωm , σm } are selected by the designer and are assumed
to be known. For a sufficiently large number, B , of basis functions, the representation (72) can approximate
well a large class of power spectra.
17
Sq (ejω )
fm (ejω ) = e
− (ω−ωm )2
σ2
m
Sq (ej ω )
0
ω1
ω2
ω3
ωB
ω
Figure 7: The interval [0, π ] is divided into B sub-intervals of equal width; the center frequencies of the sub-intervals
are denoted by {ωm }. A power spectrum Sq (ejω ) is approximated as a linear combination of Gaussian basis functions
centered on the {ωm }.
We collect the combination coefficients {βqm} for primary user q into a column vector wq :
wq
∆= col {βq1 , βq2 , βq3 , . . . , βqB }
(B × 1)
and collect the basis functions into a row vector:
fB (ejω ) (cid:3)
∆= (cid:2) f1 (ejω )
Then, the power spectrum (72) can be expressed in the alternative inner-product form:
Sq (ejω ) = fω wq
f2 (ejω )
(1 × B )
. . .
fω
(74)
(75)
(76)
Let pqk denote the path loss coefficient from primary user q to secondary user k . When the transmitted
spectrum Sq (ejω ) travels from primary user q to secondary user k , the spectrum that is sensed by node k
is pqk Sq (ejω ). We assume in this example that the path loss factors {pqk } are known and that they have
been determined during a prior training stage involving each of the primary users with each of the secondary
users. The training is usually repeated at regular intervals of time to accommodate the fact that the path
loss coefficients can vary (albeit slowly) over time. Figure 8 depicts a cognitive radio system with 2 primary
users and 10 secondary users. One of the secondary users (user 5) is highlighted and the path loss coefficients
from the primary users to its location are indicated; similar path loss coefficients can be assigned to all other
combinations involving primary and secondary users.
Each user k senses the aggregate effect of the power spectra that are transmitted by all active primary
users. Therefore, adding the effect of all primary users, we find that the power spectrum that arrives at
secondary user k is given by:
Sk (ejω ) =
pqk Sq (ejω ) + σ2
k
QXq=1
QXq=1
∆= uk,ω wo + σ2
k
=
pqk fω wq + σ2
k
(77)
where σ2
k denotes the receiver noise power at node k , and where we introduced the following vector quantities:
∆= col{w1 , w2 , . . . , wQ}
. . . pQk fω (cid:3)
∆= (cid:2) p1k fω
(BQ × 1)
(1 × BQ)
p2k fω
uk,ω
(79)
(78)
wo
18
1
3
2
4
9
7
10
2
PU
p15
1
PU
8
p25
5
SU
6
secondary user (SU)
primary user (PU)
Figure 8: A network of secondary users in the presence of two primary users. One of the secondary users is highlighted
and the path loss coefficients from the primary users to its location are indicated as p15 and p25 .
The vector wo is the collection of all combination coefficients for all Q primary users. The vector uk,ω
contains the path loss coefficients from all primary users to user k . Now, at every time instant i, user k
observers its received power spectrum, Sk (ejω ), over a discrete grid of frequencies, {ωr }, in the interval [0, π ]
in the presence of additive measurement noise. We denote these measurements by:
dk,r (i) = uk,ωr wo + σ2
k + vk,r (i)
r = 1, 2, . . . , R
(80)
The term vk,r (i) denotes sampling noise and is assumed to have zero mean and variance σ2
v,k ; it is also
assumed to be temporally white and spatially independent; and is also independent of all other random
variables. Since the row vectors uk,ω in (79) are defined in terms of the path loss coefficients {pqk }, and since
these coefficients are estimated and sub ject to noisy distortions, we model the uk,ωr as zero-mean random
variables in (80) and use the boldface notation for them.
Observe that in this application, each node k collects R measurements at every time instant i and not
only a single measurement, as was the case with the three examples discussed in the previous sections
(AR modeling, MA modeling, and localization). The implication of this fact is that we now deal with an
estimation problem that involves vector measurements instead of scalar measurements at each node. The
solution structure continues to be the same. We collect the R measurements at node k at time i into vectors
and introduce the R × 1 quantities:
=
=
∆
∆
dk,1 (i) − σ2
k
dk,2 (i) − σ2
k
...
dk,R (i) − σ2
k
v k,1 (i)
v k,2 (i)
...
vk,R (i)
(81)
dk,i
,
vk,i
19
and the regression matrix:
∆=
The time subscript in U k,i is used to model the fact that the path loss coefficients can change over time
due to the possibility of node mobility. With the above notation, expression (80) is equivalent to the linear
model:
dk,i = U k,iwo + vk,i
uk,ω1
uk,ω2
...
uk,ωR
(R × QB )
U k,i
(82)
(83)
Compared to the earlier examples (9), (44), and (58), the main difference now is that each agent k collects
a vector of measurements, dk,i , as opposed to the scalar dk (i), and its regression data are represented by
the matrix quantity, U k,i , as opposed to the row vector uk,i . Nevertheless, the estimation approach will
continue to be the same. In cognitive network applications, the secondary users are interested in estimating
the aggregate power spectrum of the primary users in order for the secondary users to identify vacant
frequency bands that can be used by them. In the context of model (83), this amounts to determining the
parameter vector wo since knowledge of its entries allows each secondary user to reconstruct the aggregate
power spectrum defined by:
QXq=1
where the notation ⊗ denotes the Kronecker product operation, and 1Q denotes a Q × 1 vector whose entries
are all equal to one.
As before, we can again verify that, in view of (83), the desired parameter vector wo satisfies the same
normal equations:
Sq (ejω ) = (1T
Q ⊗ fω )wo
SA (ejω ) ∆=
(84)
RdU,k = RU,kwo ⇐⇒ wo = R−1
U,k RdU,k
where the moments {RdU,k , RU,k } are now defined by
RdU,k
RU,k
∆= E U ∗
k,idk,i
∆= E U ∗
k,iU k,i
(QB × 1)
(QB × QB )
(85)
(86)
(87)
(88)
Therefore, each secondary user k can determine wo on its own by solving the following minimum mean-
square-error estimation problem:
E kdk,i − U k,iwk2
min
w
This solution method requires knowledge of the moments {RdU,k , RU,k } and, in an argument similar to the
one that led to (20), it can be verified that each agent k would attain an MSE performance level that is
equal to the noise power level, σ2
v,k , at its location.
Alternatively, when the statistical information {RdU,k , RU,k } is not available, each secondary user k can
estimate wo iteratively by feeding data {dk,i , U k,i } into an adaptive implementation similar to (26)–(27),
such as the following vector LMS recursion:
ek,i = dk,i − U k,iwk,i−1
wk,i = wk,i−1 + µkU ∗
k,iek,i
(89)
(90)
In this case, each secondary user k will achieve the same performance levels shown earlier in (36)–(37) with
Ru,k replaced by RU,k . The performance will again be dependent on the local noise level, σ2
v,k . As a result,
secondary users with larger noise power will perform worse than secondary users with smaller noise power.
However, since all secondary users are observing data arising from the same underlying model wo , it is
natural to expect cooperation among the users to be beneficial. As we are going to see, starting from the
20
next section, one way to achieve cooperation and improve performance is by developing algorithms that solve
the following global cost function in an adaptive and distributed manner:
NXk=1
3 Distributed Optimization via Diffusion Strategies
E kdk,i − U k,iwk2
min
w
(91)
The examples in the previous section were meant to illustrate how MSE cost functions and linear models are
useful design tools and how they arise frequently in applications. We now return to problem (1) and study
the distributed optimization of global cost functions such as (39), where J glob (w) is assumed to consist of
the sum of individual components. Specifically, we are now interested in solving optimization problems of
the type:
NXk=1
where each Jk (w) is assumed to be differentiable and convex over w. Although the algorithms presented in
this article apply to more general situations, we shall nevertheless focus on mean-square-error cost functions
of the form:
Jk (w)
min
w
(92)
Jk (w) ∆= E dk (i) − uk,iw2
(93)
where w is an M × 1 column vector, and the random processes {dk (i), uk,i } are assumed to be jointly
wide-sense stationary with zero-mean and second-order moments:
σ2
d,k
Ru,k
rdu,k
∆= E dk (i)2
∆= E u∗
k,iuk,i > 0
∆= E dk (i)u∗
k,i
(M × M )
(M × 1)
It is clear that each Jk (w) is quadratic in w since, after expansion, we get
d,k − w∗ rdu,k − r∗
du,k w + w∗Ru,k w
Jk (w) = σ2
(94)
(95)
(96)
(97)
A completion-of-squares argument shows that Jk (w) can be expressed as the sum of two squared terms, i.e.,
u,k rdu,k (cid:17) + (w − wo )∗Ru,k (w − wo )
Jk (w) = (cid:16)σ2
du,kR−1
d,k − r∗
Jk (w) = Jk,min + kw − wo k2
Ru,k
or, more compactly,
(98)
(99)
where wo denotes the minimizer of Jk (w) and is given by
wo ∆= R−1
u,k rdu,k
and Jk,min denotes the minimum value of Jk (w) when evaluated at w = wo :
Jk,min
∆= σ2
du,kR−1
d,k − r∗
u,k rdu,k = Jk (wo )
21
(100)
(101)
Observe that this value is necessarily non-negative since it can be viewed as the Schur complement of the
following covariance matrix:
E (cid:18)(cid:20) d∗
k,i (cid:21) (cid:2) dk (i) uk,i (cid:3)(cid:19) = (cid:20) σ2
rdu,k Ru,k (cid:21)
r∗
k (i)
du,k
d,k
u∗
and covariance matrices are nonnegative-definite.
The choice of the quadratic form (93) or (97) for Jk (w) is useful for many applications, as was already
illustrated in the previous section for examples involving AR modeling, MA modeling, localization, and
spectral sensing. Other choices for Jk (w) are of course possible and these choices can even be different for
different nodes. It is sufficient in this article to illustrate the main concepts underlying diffusion adaptation
by focusing on the useful case of MSE cost functions of the form (97); still, most of the derivations and
arguments in the coming sections can be extended beyond MSE optimization to more general cost functions
— as already shown in [1–3]; see also Sec. 10.4.
The positive-definiteness of the covariance matrices {Ru,k } ensures that each Jk (w) in (97) is strictly
convex, as well as J glob (w) from (39). Moreover, all these cost functions have a unique minimum at the
same wo , which satisfies the normal equations:
(102)
Ru,k wo = rdu,k ,
for every k = 1, 2, . . . , N
(103)
Therefore, given knowledge of {rdu,k , Ru,k }, each node can determine wo on its own by solving (103). One
then wonders about the need to seek distributed cooperative and adaptive solutions. There are a couple of
reasons:
(a) First, even for MSE cost functions, it is often the case that the required moments {rdu,k , Ru,k } are not
known beforehand. In this case, the optimal wo cannot be determined from the solution of the normal
equations (103). The alternative methods that we shall describe will lead to adaptive techniques that
enable each node k to estimate wo directly from data realizations.
(b) Second, since adaptive strategies rely on instantaneous data, these strategies possess powerful tracking
abilities. Even when the moments vary with time due to non-stationary behavior (such as wo changing
with time), these changes will be reflected in the observed data and will in turn influence the behavior
of the adaptive construction. This is one of the key advantages of adaptive strategies: they enable
learning and tracking in real-time.
(c) Third, cooperation among nodes is generally beneficial. When nodes act individually, their performance
is limited by the noise power level at their location. In this way, some nodes can perform significantly
better than other nodes. On the other hand, when nodes cooperate with their neighbors and share
information during the adaptation process, we will see that performance can be improved across the
network.
3.1 Relating the Global Cost to Neighborhood Costs
Let us therefore consider the optimization of the following global cost function:
NXk=1
where Jk (w) is given by (93) or (97). Our strategy to optimize J glob (w) in a distributed manner is based
on two steps, following the developments in [1, 2, 18]. First, using a completion-of-squares argument (or,
equivalently, a second-order Taylor series expansion), we approximate the global cost function (104) by
an alternative local cost that is amenable to distributed optimization. Then, each node will optimize the
alternative cost via a steepest-descent method.
J glob (w) =
Jk (w)
(104)
22
To motivate the distributed diffusion-based approach, we start by introducing a set of nonnegative coef-
ficients {ckℓ } that satisfy two conditions:
for k = 1, 2, . . . , N :
ckℓ ≥ 0,
ckℓ = 1, and ckℓ = 0 if ℓ /∈ Nk
NXℓ=1
where Nk denotes the neighborhood of node k . Condition (105) means that for every node k , the sum of the
coefficients {ckℓ } that relate it to its neighbors is one. The coefficients {ckℓ } are free parameters that are
chosen by the designer; obviously, as shown later in Theorem 6.8, their selection will have a bearing on the
performance of the resulting algorithms. If we collect the entries {ckℓ } into an N × N matrix C , so that the
k−th row of C is formed of {ckℓ , ℓ = 1, 2, . . . , N }, then condition (105) translates into saying that each of
row of C adds up to one, i.e.,
(105)
where the notation 1 denotes an N × 1 column vector with all its entries equal to one:
1
∆= col{1, 1, . . . , 1}
C 1 = 1
(106)
(107)
J loc
ℓ
We say that C is a right stochastic matrix. Using the coefficients {ckℓ } so defined, we associate with each
node ℓ, a local cost function of the following form:
(w) ∆= Xk∈Nℓ
This cost consists of a weighted combination of the individual costs of the neighbors of node ℓ (including ℓ
itself ) — see Fig. 9. Since the {ckℓ } are all nonnegative and each Jk (w) is strictly convex, then J loc
(w) is
ℓ
also strictly convex and its minimizer occurs at the same w = wo . Using the alternative representation (99)
for the individual Jk (w), we can re-express the local cost J loc
(w) as
ℓ
(w) = Xk∈Nℓ
ckℓ Jk,min + Xk∈Nℓ
ckℓ kw − wo k2
Ru,k
ckℓ Jk (w)
(109)
(108)
J loc
ℓ
or, equivalently,
J loc
ℓ
(w) = J loc
ℓ,min + kw − wo k2
Rℓ
(110)
J loc
ℓ,min
ckℓ Jk,min
ℓ,min corresponds to the minimum value of J loc
where J loc
(w) at the minimizer w = wo :
ℓ
∆= Xk∈Nℓ
and Rℓ is a positive-definite weighting matrix defined by:
∆= Xk∈Nℓ
That is, Rℓ is a weighted combination of the covariance matrices in the neighborhood of node ℓ. Equality
(110) amounts to a (second-order) Taylor series expansion of J loc
(w) around w = wo . Note that the right-
ℓ
hand side consists of two terms: the minimum cost and a weighted quadratic term in the difference (w − wo ).
ckℓRu,k
(111)
(112)
Rℓ
23
10
J1 (w)
1
8
c13
9
7
c53
5
J5 (w)
6
3
J3 (w)
J loc
3 (w)
c23
4
2
J2 (w)
N3
c13 + c23 + c33 + c53 = 1
J loc
3 (w) = c13J1 (w) + c23J2 (w) + c33J3 (w) + c53J5 (w)
Figure 9: A network with N = 10 nodes. The nodes in the neighborhood of node 3 are highlighted with their
individual cost functions, and with the combination weights {c13 , c23 , c53 } along the connecting edges; there is also a
combination weight associated with node 3 and is denoted by c33 . The expression for the local cost function, J loc
3 (w),
is also shown in the figure.
(105)
=
J glob (w)
Now note that we can express J glob (w) from (104) as follows:
NXk=1 NXℓ=1
ckℓ! Jk (w)
NXℓ=1 NXk=1
ckℓ Jk (w)!
NXℓ=1
J loc
(w)
ℓ
= J loc
k (w) +
(108)
=
=
(w)
J loc
ℓ
NXℓ 6=k
Substituting (110) into the second term on the right-hand side of the above expression gives:
k (w) + Xℓ 6=k
Rℓ + Xℓ 6=k
J glob (w) = J loc
kw − wo k2
The last term in the above expression does not depend on w. Therefore, minimizing J glob (w) over w is
equivalent to minimizing the following alternative global cost:
k (w) + Xℓ 6=k
(w) = J loc
kw − wo k2
Rℓ
J loc
ℓ,min
J glob′
(113)
(114)
(115)
24
Expression (115) relates the optimization of the original global cost function, J glob (w) or its equivalent
J glob′
(w), to the newly-introduced local cost function J loc
k (w). The relation is through the second term on
the right-hand side of (115), which corresponds to a sum of quadratic factors involving the minimizer wo ;
k (w) can be corrected to the global cost J glob′
this term tells us how the local cost J loc
(w). Obviously, the
minimizer wo that appears in the correction term is not known since the nodes wish to determine its value.
Likewise, not all the weighting matrices Rℓ are available to node k ; only those matrices that originate from
its neighbors can be assumed to be available. Still, expression (115) suggests a useful way to replace J loc
k
by another local cost that is closer to J glob′
(w). This alternative cost will be shown to lead to a powerful
distributed solution to optimize J glob (w) through localized interactions.
Our first step is to limit the summation on the right-hand side of (115) to the neighbors of node k (since
every node k can only have access to information from its neighbors). We thus introduce the modified cost
function at node k :
k (w) + Xℓ∈Nk \{k}
(w) ∆= J loc
k (w) and J glob′
The cost functions J loc
(w) are both associated with node k ; the difference between them is
k
that the expression for the latter is closer to the global cost function (115) that we want to optimize.
The weighting matrices {Rℓ } that appear in (116) may or may not be available because the second-
order moments {Ru,ℓ} may or may not be known beforehand. If these moments are known, then we can
proceed with the analysis by assuming knowledge of the {Rℓ }. However, the more interesting case is when
these moments are not known. This is generally the case in practice, especially in the context of adaptive
solutions and problems involving non-stationary data. Often, nodes can only observe realizations {uℓ,i} of
the regression data {uℓ,i} arising from distributions whose covariance matrices are the unknown {Ru,ℓ}.
One way to address the difficulty is to replace each of the weighted norms kw − wo k2
in (116) by a scaled
Rℓ
multiple of the un-weighted norm, say,
kw − wo k2
Rℓ
J glob′
k
(116)
kw − wo k2
Rℓ ≈ bℓk · kw − wo k2
(117)
where bℓk is some nonnegative coefficient; we are even allowing its value to change with the node index k .
The above substitution amounts to having each node k approximate the {Rℓ} from its neighbors by multiples
of the identity matrix
Rℓ ≈ bℓk IM
(118)
Approximation (117) is reasonable in view of the fact that all vector norms are equivalent [19–21]; this norm
property ensures that we can bound the weighted norm kw − wo k2
by some constants multiplying the
Rℓ
un-weighted norm kw − wo k2 , say, as:
r1kw − wo k2 ≤ kw − wo k2
Rℓ ≤ r2kw − wo k2
(119)
for some positive constants (r1 , r2 ). Using the fact that the {Rℓ} are Hermitian positive-definite matrices,
and calling upon the Rayleigh-Ritz characterization of eigenvalues [19, 20], we can be more specific and
replace the above inequalities by
λmin (Rℓ ) · kw − wo k2 ≤ kw − wo k2
Rℓ ≤ λmax (Rℓ ) · kw − wo k2
(120)
We note that approximations similar to (118) are common in stochastic approximation theory and they
mark the difference between using a Newton’s iterative method or a stochastic gradient method [5, 22]; the
former uses Hessian matrices as approximations for Rℓ and the latter uses multiples of the identity matrix.
Furthermore, as the derivation will reveal, we do not need to worry at this stage about how to select the
scalars {bℓk }; they will end up being embedded into another set of coefficients {aℓk } that will be set by the
designer or adjusted by the algorithm — see (132) further ahead.
Thus, we replace (116) by
k (w) + Xℓ∈Nk \{k}
(w) = J loc
25
bℓk kw − wo k2
J glob′′
k
(121)
The argument so far has suggested how to modify J loc
k (w) from (108) and replace it by the cost (121) that is
closer in form to the global cost function (115). If we replace J loc
k (w) by its definition (108), we can rewrite
(121) as
cℓk Jℓ (w) + Xℓ∈Nk \{k}
(w) = Xℓ∈Nk
With the exception of the variable wo , this approximate cost at node k relies solely on information that is
available to node k from its neighborhood. We will soon explain how to handle the fact that wo is not known
beforehand to node k .
bℓk kw − wo k2
J glob′′
k
(122)
,
i ≥ 0
(123)
3.2 Steepest-Descent Iterations
Node k can apply a steepest-descent iteration to minimize J glob′′
(w). Let wk,i denote the estimate for the
k
minimizer wo that is evaluated by node k at time i. Starting from an initial condition wk,−1 , node k can
compute successive estimates iteratively as follows:
wk,i = wk,i−1 − µk h∇w J glob′′
(wk,i−1 )i∗
k
where µk is a small positive step-size parameter, and the notation ∇w J (a) denotes the gradient vector of the
function J (w) relative to w and evaluated at w = a. The step-size parameter µk can be selected to vary with
time as well. One choice that is common in the optimization literature [5, 22, 52] is to replace µk in (123) by
step-size sequences {µ(i) ≥ 0} that satisfy the two conditions (25). However, such step-size sequences are
not suitable for applications that require continuous learning because they turn off adaptation as i → ∞;
the steepest-descent iteration (123) would stop updating since µk (i) would be tending towards zero. For this
reason, we shall focus mainly on the constant step-size case described by (123) since we are interested in
developing distributed algorithms that will endow networks with continuous adaptation abilities.
Returning to (123) and computing the gradient vector of (122) we get:
wk,i = wk,i−1 − µk Xℓ∈Nk
cℓk [∇w Jℓ (wk,i−1 )]∗ − µk Xℓ∈Nk \{k}
Using the expression for Jℓ (w) from (97) we arrive at
wk,i = wk,i−1 + µk Xℓ∈Nk
cℓk (rdu,ℓ − Ru,ℓ wk,i−1 ) + µk Xℓ∈Nk \{k}
This iteration indicates that the update from wk,i−1 to wk,i involves adding two correction terms to wk,i−1 .
Among many other forms, we can implement the update in two successive steps by adding one correction
term at a time, say, as follows:
ψk,i = wk,i−1 + µk Xℓ∈Nk
cℓk (rdu,ℓ − Ru,ℓ wk,i−1 )
wk,i = ψk,i + µk Xℓ∈Nk \{k}
Step (126) updates wk,i−1 to an intermediate value ψk,i by using local gradient vectors from the neighborhood
of node k . Step (127) further updates ψk,i to wk,i . However, this second step is not realizable since wo is
not known and the nodes are actually trying to estimate it. Two issues stand out from examining (127):
bℓk (wo − wk,i−1 )
bℓk (wk,i−1 − wo )
bℓk (wo − wk,i−1 )
(126)
(127)
(124)
(125)
(a) First, iteration (127) requires knowledge of the minimizer wo . Neither node k nor its neighbors know
the value of the minimizer; each of these nodes is actually performing steps similar to (126) and (127)
26
to estimate the minimizer. However, each node ℓ has a readily available approximation for wo , which
is its local intermediate estimate ψℓ,i . Therefore, we replace wo in (127) by ψℓ,i . This step helps diffuse
information throughout the network. This is because each neighbor of node k determines its estimate
ψℓ,i by processing information from its own neighbors, which process information from their neighbors,
and so forth.
(b) Second, the intermediate value ψk,i at node k is generally a better estimate for wo than wk,i−1 since it
is obtained by incorporating information from the neighbors through the first step (126). Therefore,
we further replace wk,i−1 in (127) by ψk,i . This step is reminiscent of incremental-type approaches to
optimization, which have been widely studied in the literature [23–26].
With the substitutions described in items (a) and (b) above, we replace the second step (127) by
wk,i = ψk,i + µk Xℓ∈Nk \{k}
bℓk (ψℓ,i − ψk,i )
bℓk ψk,i + µk Xℓ∈Nk \{k}
= 1 − µk Xℓ∈Nk \{k}
Introduce the weighting coefficients:
∆= 1 − µk Xℓ∈Nk \{k}
bℓk
akk
∆= µk bℓk ,
ℓ ∈ Nk \{k}
∆= 0,
bℓk ψℓ,i
ℓ /∈ Nk
aℓk
aℓk
(128)
(129)
(130)
(131)
and observe that, for sufficiently small step-sizes µk , these coefficients are nonnegative and, moreover, they
satisfy the conditions:
for k = 1, 2, . . . , N :
aℓk ≥ 0,
aℓk = 1, and aℓk = 0 if ℓ /∈ Nk
NXℓ=1
Condition (132) means that for every node k , the sum of the coefficients {aℓk } that relate it to its neighbors
is one. Just like the {cℓk }, from now on, we will treat the coefficients {aℓk } as free weighting parameters that
are chosen by the designer according to (132); their selection will also have a bearing on the performance of
the resulting algorithms — see Theorem 6.8. If we collect the entries {aℓk } into an N × N matrix A, such
that the k−th column of A consists of {aℓk , ℓ = 1, 2, . . . , N }, then condition (132) translates into saying
that each column of A adds up to one:
(132)
We say that A is a left stochastic matrix.
3.3 Adapt-then-Combine (ATC) Diffusion Strategy
AT 1 = 1
(133)
Using the coefficients {aℓk } so defined, we replace (126) and (128) by the following recursions for i ≥ 0:
ψk,i = wk,i−1 + µk Xℓ∈Nk
wk,i = Xℓ∈Nk
aℓk ψℓ,i
cℓk (rdu,ℓ − Ru,ℓ wk,i−1 )
(ATC strategy)
(134)
27
for some nonnegative coefficients {cℓk , aℓk } that satisfy conditions (106) and (133), namely,
C 1 = 1,
AT 1 = 1
(135)
or, equivalently,
for k = 1, 2, . . . , N :
cℓk ≥ 0,
cℓk = 1,
cℓk = 0 if ℓ /∈ Nk
aℓk ≥ 0,
aℓk = 0 if ℓ /∈ Nk
NXk=1
NXℓ=1
aℓk = 1,
To run algorithm (134), we only need to select the coefficients {aℓk , cℓk } that satisfy (135) or (136); there is
no need to worry about the intermediate coefficients {bℓk } any longer since they have been blended into the
{aℓk }. The scalars {cℓk , aℓk } that appear in (134) correspond to weighting coefficients over the edge linking
node k to its neighbors ℓ ∈ Nk . Note that two sets of coefficients are used to scale the data that are being
received by node k : one set of coefficients, {cℓk }, is used in the first step of (134) to scale the moment data
{rdu,ℓ , Ru,ℓ }, and a second set of coefficients, {aℓk }, is used in the second step of (134) to scale the estimates
{ψℓ,i}. Figure 10 explains what the entries on the columns and rows of the combination matrices {A, C }
stand for using an example with N = 6 and the matrix C for illustration. When the combination matrix is
right-stochastic (as is the case with C ), each of its rows would add up to one. On the other hand, when the
matrix is left-stochastic (as is the case with A), each of its columns would add up to one.
(136)
coefficients used to scale
data sent from node 3
to its neighbors.
× × c13 × × ×
× × c23 × × ×
c32
c33
c34
c35
c36
c31
× × c43 × × ×
× × c53 × × ×
× × c63 × × ×
C =
coefficients used to scale
data received by node 3
from its neighbors.
Figure 10: Interpretation of the columns and rows of combination matrices. The pair of entries {ckℓ , cℓk } correspond
to weighting coefficients used over the edge connecting nodes k and ℓ. When nodes (k, ℓ) are not neighbors, then
these weights are zero.
At every time instant i, the ATC strategy (134) performs two steps. The first step is an information
exchange step where node k receives from its neighbors their moments {Ru,ℓ , rdu,ℓ }. Node k combines this
information and uses it to update its existing estimate wk,i−1 to an intermediate value ψk,i . All other nodes
in the network are performing a similar step and updating their existing estimates {wℓ,i−1} into intermediate
estimates {ψℓ,i} by using information from their neighbors. The second step in (134) is an aggregation or
28
consultation step where node k combines the intermediate estimates of its neighbors to obtain its update
estimate wk,i . Again, all other nodes in the network are simultaneously performing a similar step. The
reason for the name Adapt-then-Combine (ATC) strategy is that the first step in (134) will be shown to lead
to an adaptive step, while the second step in (134) corresponds to a combination step. Hence, strategy (134)
involves adaptation followed by combination or ATC for short. The reason for the qualification “diffusion”
is that the combination step in (134) allows information to diffuse through the network in real time. This is
because each of the estimates ψℓ,i is influenced by data beyond the immediate neighborhood of node k .
In the special case when C = I , so that no information exchange is performed but only the aggregation
step, the ATC strategy (134) reduces to:
(ATC strategy without
information exchange)
ψk,i = wk,i−1 + µk (rdu,k − Ru,k wk,i−1 )
wk,i = Xℓ∈Nk
where the first step relies solely on the information {Ru,k , rdu,k } that is available locally at node k .
Observe in passing that the term that appears in the information exchange step of (134) is related to the
gradient vectors of the local costs {Jℓ(w)} evaluated at wk,i−1 , i.e., it holds that
aℓk ψℓ,i
(137)
rdu,ℓ − Ru,ℓ wk,i−1 = − [∇w Jℓ (wk,i−1 )]∗
(138)
(ATC strategy)
cℓk [∇w Jℓ (wk,i−1 )]∗
so that the ATC strategy (134) can also be written in the following equivalent form:
ψk,i = wk,i−1 − µk Xℓ∈Nk
wk,i = Xℓ∈Nk
aℓk ψℓ,i
The significance of this general form is that it is applicable to optimization problems involving more general
local costs Jℓ (w) that are not necessarily quadratic in w, as detailed in [1–3] — see also Sec. 10.4. The
top part of Fig. 11 illustrates the two steps involved in the ATC procedure for a situation where node k
has three other neighbors labeled {1, 2, ℓ}. In the first step, node k evaluates the gradient vectors of its
neighbors at wk,i−1 , and subsequently aggregates the estimates {ψ1,i , ψ2,i , ψℓ,i } from its neighbors. The
dotted arrows represent flow of information towards node k from its neighbors. The solid arrows represent
flow of information from node k to its neighbors. The CTA diffusion strategy is discussed next.
(139)
3.4 Combine-then-Adapt (CTA) Diffusion Strategy
Similarly, if we return to (125) and add the second correction term first, then (126)–(127) are replaced by:
ψk,i−1 = wk,i−1 + µk Xℓ∈Nk \{k}
wk,i = ψk,i−1 + µk Xℓ∈Nk
cℓk (rdu,ℓ − Ru,ℓ wk,i−1 )
Following similar reasoning to what we did before in the ATC case, we replace wo in step (140) by wℓ,i−1 and
replace wk,i−1 in (141) by ψk,i−1 . We then introduce the same coefficients {aℓk } and arrive at the following
bℓk (wo − wk,i−1 )
(140)
(141)
29
∇w J1 (wk ,i−1 )
1
k
∇w Jk (wk,i−1 )
∇w Jℓ (wk ,i−1 )
ℓ
2
ψ1,i
∇w J2 (wk ,i−1 )
ψk ,i
k
1
ATC
2
ψ2,i
ℓ
CTA
2
∇w J2 (ψk ,i−1 )
ψℓ,i
1
k
1
k
w1,i−1
wk,i−1
2
∇w J1 (ψk ,i−1 )
w2,i−1
∇w Jk (ψk ,i−1 )
wℓ,i−1
ℓ
∇w Jℓ (ψk ,i−1 )
ℓ
Figure 11: Illustration of the ATC and CTA strategies for a node k with three other neighbors {1, 2, ℓ}. The updates
involve two steps:
information exchange followed by aggregation in ATC and aggregation followed by information
exchange in CTA. The dotted blue arrows represent the data received from the neighbors of node k, and the solid
red arrows represent the data sent from node k to its neighbors.
combine-then-adapt (CTA) strategy:
ψk,i−1 = Xℓ∈Nk
aℓk wℓ,i−1
wk,i = ψk,i−1 + µk Xℓ∈Nk
where the nonnegative coefficients {cℓk , aℓk } satisfy the same conditions (106) and (133), namely,
cℓk (rdu,ℓ − Ru,ℓ ψk,i−1 )
(CTA strategy)
or, equivalently,
C 1 = 1,
AT 1 = 1
for k = 1, 2, . . . , N :
(142)
(143)
cℓk ≥ 0,
cℓk = 1,
cℓk = 0 if ℓ /∈ Nk
NXk=1
NXℓ=1
aℓk = 1,
At every time instant i, the CTA strategy (142) also consists of two steps. The first step is an aggregation
step where node k combines the existing estimates of its neighbors to obtain the intermediate estimate ψk,i−1 .
aℓk = 0 if ℓ /∈ Nk
aℓk ≥ 0,
(144)
30
All other nodes in the network are simultaneously performing a similar step and aggregating the estimates
of their neighbors. The second step in (142) is an information exchange step where node k receives from
its neighbors their moments {Rdu,ℓ , rdu,ℓ } and uses this information to update its intermediate estimate to
wk,i . Again, all other nodes in the network are simultaneously performing a similar information exchange
step. The reason for the name Combine-then-Adapt (CTA) strategy is that the first step in (142) involves
a combination step, while the second step will be shown to lead to an adaptive step. Hence, strategy (142)
involves combination followed by adaptation or CTA for short. The reason for the qualification “diffusion”
is that the combination step of (142) allows information to diffuse through the network in real time.
In the special case when C = I , so that no information exchange is performed but only the aggregation
step, the CTA strategy (142) reduces to:
(CTA strategy without
information exchange)
ψk,i−1 = Xℓ∈Nk
wk,i = ψk,i−1 + µk (rdu,k − Ru,k ψk,i−1 )
aℓk wℓ,i−1
(145)
where the second step relies solely on the information {Ru,k , rdu,k } that is available locally at node k . Again,
the CTA strategy (142) can be rewritten in terms of the gradient vectors of the local costs {Jℓ (w)} as follows:
ψk,i−1 = Xℓ∈Nk
aℓk wℓ,i−1
wk,i = ψk,i−1 − µk Xℓ∈Nk
The bottom part of Fig. 11 illustrates the two steps involved in the CTA procedure for a situation where
node k has three other neighbors labeled {1, 2, ℓ}.
In the first step, node k aggregates the estimates
{w1,i−1 , w2,i−1 , wℓ,i−1 } from its neighbors, and subsequently performs information exchange by evaluating
the gradient vectors of its neighbors at ψk,i−1 .
cℓk [∇w Jℓ (ψk,i−1 )]∗
(CTA strategy)
(146)
3.5 Useful Properties of Diffusion Strategies
Note that the structure of the ATC and CTA diffusion strategies (134) and (142) are fundamentally the
same: the difference between the implementations lies in which variable we choose to correspond to the
updated weight estimate wk,i . In the ATC case, we choose the result of the combination step to be wk,i ,
whereas in the CTA case we choose the result of the adaptation step to be wk,i .
For ease of reference, Table 2 lists the steepest-descent diffusion algorithms derived in the previous
sections. The derivation of the ATC and CTA strategies (134) and (142) followed the approach proposed in
[18,27]. CTA estimation schemes were first proposed in the works [35–39], and later extended in [18,27,32,33].
The earlier versions of CTA in [35–37] used the choice C = I . This form of the algorithm with C = I , and
with the additional constraint that the step-sizes µk should be time-dependent and decay towards zero as
time progresses, was later applied by [40, 41] to solve distributed optimization problems that require all nodes
to reach consensus or agreement. Likewise, special cases of the ATC estimation scheme (134), involving an
information exchange step followed by an aggregation step, first appeared in the work [28] on diffusion least-
squares schemes and subsequently in the works [18, 29–33] on distributed mean-square-error and state-space
estimation methods. A special case of the ATC strategy (134) corresponding to the choice C = I with
decaying step-sizes was adopted in [34] to ensure convergence towards a consensus state. Diffusion strategies
of the form (134) and (142) (or, equivalently, (139) and (146)) are general in several respects:
(1) These strategies do not only diffuse the local weight estimates, but they can also diffuse the local
gradient vectors. In other words, two sets of combination coefficients {aℓk , cℓk } are used.
31
Table 2: Summary of steepest-descent diffusion strategies for the distributed optimization of general problems of the
form (92), and their specialization to the case of mean-square-error (MSE) individual cost functions given by (93).
Reference
cℓk [∇w Jℓ (wk,i−1 )]∗
(139)
(134)
(146)
(142)
Algorithm
ATC strategy
(general case)
ATC strategy
(MSE costs)
CTA strategy
(general case)
CTA strategy
(MSE costs)
Recursions
ψk,i = wk,i−1 − µk Xℓ∈Nk
wk,i = Xℓ∈Nk
aℓk ψℓ,i
ψk,i = wk,i−1 + µk Xℓ∈Nk
wk,i = Xℓ∈Nk
aℓk ψℓ,i
ψk,i−1 = Xℓ∈Nk
aℓk wℓ,i−1
wk,i = ψk,i−1 − µk Xℓ∈Nk
ψk,i−1 = Xℓ∈Nk
aℓk wℓ,i−1
wk,i = ψk,i−1 + µk Xℓ∈Nk
cℓk (rdu,ℓ − Ru,ℓ wk,i−1 )
cℓk [∇w Jℓ (ψk,i−1 )]∗
cℓk (rdu,ℓ − Ru,ℓ ψk,i−1 )
(2) In the derivation that led to the diffusion strategies, the combination matrices C and A are only
required to be right-stochastic (for C ) and left-stochastic (for A).
In comparison, it is common in
consensus-type strategies to require the corresponding combination matrix A to be doubly stochastic
(i.e., its rows and columns should add up to one) — see, e.g., App. E and [40, 42–44].
(3) As the analysis in Sec. 6 will reveal, ATC and CTA strategies do not force nodes to converge to an
agreement about the desired parameter vector wo , as is common in consensus-type strategies (see
App. E and [40, 45–51]). Forcing nodes to reach agreement on wo ends up limiting the adaptation
and learning abilities of these nodes, as well as their ability to react to information in real-time.
Nodes in diffusion networks enjoy more flexibility in the learning process, which allows their individual
estimates, {wk,i }, to tend to values that lie within a reasonable mean-square-deviation (MSD) level
from the optimal solution, wo . Multi-agent systems in nature behave in this manner; they do not
require exact agreement among their agents (see, e.g., [8–10]).
(4) The step-size parameters {µk } are not required to depend on the time index i and are not required
to vanish as i → ∞ (as is common in many works on distributed optimization, e.g., [22, 40, 52, 53]).
Instead, the step-sizes can assume constant values, which is a critical property to endow networks
with continuous adaptation and learning abilities. An important contribution in the study of diffusion
strategies is to show that distributed optimization is still possible even for constant step-sizes, in
addition to the ability to perform adaptation, learning, and tracking. Sections 5 and 6 highlight the
convergence properties of the diffusion strategies — see also [1–3] for results pertaining to more general
cost functions.
32
(5) Even the combination weights {aℓk , cℓk } can be adapted, as we shall discuss later in Sec. 8.3. In this
way, diffusion strategies allow multiple layers of adaptation: the nodes perform adaptive processing,
the combination weights can be adapted, and even the topology can be adapted especially for mobile
networks [8].
4 Adaptive Diffusion Strategies
The distributed ATC and CTA steepest-descent strategies (134) and (142) for determining the wo that solves
(92)–(93) require knowledge of the statistical information {Ru,k , rdu,k }. These moments are needed in order
to be able to evaluate the gradient vectors that appear in (134) and (142), namely, the terms:
− [∇w Jℓ (wk,i−1 )]∗ = (rdu,ℓ − Ru,ℓ wk,i−1 )
− [∇w Jℓ (ψk,i−1 )]∗ = (rdu,ℓ − Ru,ℓ ψk,i−1 )
(147)
(148)
for all ℓ ∈ Nk . However, the moments {Ru,ℓ , rdu,ℓ } are often not available beforehand, which means that the
true gradient vectors are generally not available. Instead, the agents have access to observations {dk (i), uk,i}
of the random processes {dk (i), uk,i }. There are many ways by which the true gradient vectors can be
approximated by using these observations. Recall that, by definition,
Ru,ℓ
∆= E u∗
ℓ,iuℓ,i ,
rdu,ℓ
∆= E dℓ (i)u∗
ℓ,i
(149)
One common stochastic approximation method is to drop the expectation operator from the definitions of
{Ru,ℓ , rdu,ℓ } and to use the following instantaneous approximations instead [4–7]:
In this case, the approximate gradient vectors become:
Ru,ℓ ≈ u∗
ℓ,iuℓ,i ,
rdu,ℓ ≈ dℓ (i)u∗
ℓ,i
(rdu,ℓ − Ru,ℓ wk,i−1 ) ≈ u∗
ℓ,i [dℓ (i) − uℓ,i wk,i−1 ]
(rdu,ℓ − Ru,ℓ ψk,i−1 ) ≈ u∗
ℓ,i [dℓ (i) − uℓ,i ψk,i−1 ]
(150)
(151)
(152)
Substituting into the ATC and CTA steepest-descent strategies (134) and (142), we arrive at the following
adaptive implementations of the diffusion strategies for i ≥ 0:
ψk,i = wk,i−1 + µk Xℓ∈Nk
wk,i = Xℓ∈Nk
aℓk ψℓ,i
cℓk u∗
ℓ,i [dℓ (i) − uℓ,iwk,i−1 ]
(adaptive ATC strategy)
(153)
and
(adaptive CTA strategy)
ψk,i−1 = Xℓ∈Nk
aℓk wℓ,i−1
wk,i = ψk,i−1 + µk Xℓ∈Nk
where the coefficients {aℓk , cℓk } are chosen to satisfy:
cℓk u∗
ℓ,i [dℓ (i) − uℓ,i ψk,i−1 ]
for k = 1, 2, . . . , N :
cℓk = 0 if ℓ /∈ Nk
aℓk = 0 if ℓ /∈ Nk
cℓk ≥ 0,
aℓk ≥ 0,
cℓk = 1,
NXk=1
NXℓ=1
aℓk = 1,
33
(154)
(155)
The adaptive implementations usually start from the initial conditions wℓ,−1 = 0 for all ℓ, or from some
other convenient initial values. Clearly, in view of the approximations (151)–(152), the successive iterates
{wk,i , ψk,i , ψk,i−1 } that are generated by the above adaptive implementations are different from the iterates
that result from the steepest-descent implementations (134) and (142). Nevertheless, we shall continue
to use the same notation for these variables for ease of reference. One key advantage of the adaptive
implementations (153)–(154) is that they enable the agents to react to changes in the underlying statistical
information {rdu,ℓ , Ru,ℓ } and to changes in wo . This is because these changes end up being reflected in the
data realizations {dk (i), uk,i}. Therefore, adaptive implementations have an innate tracking and learning
ability that is of paramount significance in practice.
We say that the stochastic gradient approximations (151)–(152) introduce gradient noise into each step
of the recursive updates (153)–(154). This is because the updates (153)–(154) can be interpreted as corre-
sponding to the following forms:
(adaptive ATC strategy)
and
(adaptive CTA strategy)
ψk,i = wk,i−1 − µk Xℓ∈Nk
wk,i = Xℓ∈Nk
aℓk ψℓ,i
cℓk
\[∇w Jℓ (wk,i−1 )]
∗
(156)
ψk,i−1 = Xℓ∈Nk
aℓk wℓ,i−1
wk,i = ψk,i−1 − µk Xℓ∈Nk
cℓk
\[∇w Jℓ (ψk,i−1 )]
∗
(157)
where the true gradient vectors, {∇w Jℓ(·)}, have been replaced by approximations, { \∇w Jℓ (·)} — compare
with (139) and (146). The significance of the alternative forms (156)–(157) is that they are applicable to
optimization problems involving more general local costs Jℓ (w) that are not necessarily quadratic, as detailed
in [2, 3]; see also Sec. 10.4. In the next section, we examine how gradient noise affects the performance of
the diffusion strategies and how close the successive estimates {wk,i } get to the desired optimal solution wo .
Table 3 lists several of the adaptive diffusion algorithms derived in this section.
The operation of the adaptive diffusion strategies is similar to the operation of the steepest-descent
diffusion strategies of the previous section. Thus, note that at every time instant i, the ATC strategy (153)
performs two steps; as illustrated in Fig. 12. The first step is an information exchange step where node
k receives from its neighbors their information {dℓ (i), uℓ,i}. Node k combines this information and uses it
to update its existing estimate wk,i−1 to an intermediate value ψk,i . All other nodes in the network are
performing a similar step and updating their existing estimates {wℓ,i−1} into intermediate estimates {ψℓ,i}
by using information from their neighbors. The second step in (153) is an aggregation or consultation step
where node k combines the intermediate estimates {ψℓ,i} of its neighbors to obtain its update estimate wk,i .
Again, all other nodes in the network are simultaneously performing a similar step. In the special case when
C = I , so that no information exchange is performed but only the aggregation step, the ATC strategy (153)
reduces to:
(adaptive ATC strategy
without information exchange)
ψk,i = wk,i−1 + µk u∗
k,i [dk (i) − uk,iwk,i−1 ]
wk,i = Xℓ∈Nk
aℓk ψℓ,i
(158)
34
Table 3: Summary of adaptive diffusion strategies for the distributed optimization of general problems of the form
(92), and their specialization to the case of mean-square-error (MSE) individual cost functions given by (93). These
adaptive solutions rely on stochastic approximations.
Algorithm
Adaptive ATC strategy
(general case)
Adaptive ATC strategy
(MSE costs)
Adaptive ATC strategy
(MSE costs)
(no information exchange)
Adaptive CTA strategy
(general case)
Adaptive CTA strategy
(MSE costs)
Adaptive CTA strategy
(MSE costs)
(no information exchange)
aℓk ψℓ,i
Recursions
ψk,i = wk,i−1 − µk Xℓ∈Nk
wk,i = Xℓ∈Nk
aℓk ψℓ,i
ψk,i = wk,i−1 + µk Xℓ∈Nk
wk,i = Xℓ∈Nk
aℓk ψℓ,i
ψk,i = wk,i−1 + µk u∗
k,i [dk (i) − uk,iwk,i−1 ]
wk,i = Xℓ∈Nk
ψk,i−1 = Xℓ∈Nk
aℓk wℓ,i−1
wk,i = ψk,i−1 − µk Xℓ∈Nk
ψk,i−1 = Xℓ∈Nk
aℓk wℓ,i−1
wk,i = ψk,i−1 + µk Xℓ∈Nk
ψk,i−1 = Xℓ∈Nk
wk,i = ψk,i−1 + µk u∗
k,i [dk (i) − uk,i ψk,i−1 ]
aℓk wℓ,i−1
∗
\[∇w Jℓ (ψk,i−1 )]
cℓk
∗
\[∇w Jℓ (wk,i−1 )]
cℓk
cℓk u∗
ℓ,i [dℓ (i) − uℓ,iwk,i−1 ]
cℓk u∗
ℓ,i [dℓ (i) − uℓ,i ψk,i−1 ]
Reference
(156)
(153)
(158)
(157)
(154)
(159)
35
Nk
ψ1,i
k
1
ψ2,i
2
ψℓ,i
Adapt-then-Combine (ATC) Strategy
(node k, time i)
(a) Information exchange step:
eℓ (i) = dℓ (i) − uℓ,iwk,i−1
ψk,i = wk,i−1 + µk (cid:1)
ℓ∈Nk
cℓk u∗
ℓ,i eℓ (i)
ℓ
(b) Aggregation/consultation step:
wk,i = (cid:1)
ℓ∈Nk
aℓk ψℓ,i
Figure 12: Illustration of the adaptive ATC strategy, which involves two steps: information exchange followed by
aggregation.
Likewise, at every time instant i, the CTA strategy (154) also consists of two steps – see Fig. 13. The
first step is an aggregation step where node k combines the existing estimates of its neighbors to obtain the
intermediate estimate ψk,i−1 . All other nodes in the network are simultaneously performing a similar step
and aggregating the estimates of their neighbors. The second step in (154) is an information exchange step
where node k receives from its neighbors their information {dℓ (i), uℓ,i} and uses this information to update
its intermediate estimate to wk,i . Again, all other nodes in the network are simultaneously performing a
similar information exchange step.
In the special case when C = I , so that no information exchange is
performed but only the aggregation step, the CTA strategy (154) reduces to:
Nk
w1,i−1
w2,i−1
k
1
2
wℓ,i−1
Combine-then-Adapt (CTA) Strategy
(node k, time i)
(a) Aggregation/consultation step:
ψk ,i−1 = (cid:1)
ℓ∈Nk
aℓk wℓ,i−1
ℓ
(b) Information exchange step:
eℓ (i) = dℓ (i) − uℓ,iψk,i−1
wk ,i = ψk ,i−1 + µk (cid:1)
ℓ∈Nk
cℓk u∗
ℓ,i eℓ (i)
Figure 13: Illustration of the adaptive CTA strategy, which involves two steps: aggregation followed by information
exchange.
36
(adaptive CTA strategy
without information exchange)
ψk,i−1 = Xℓ∈Nk
wk,i = ψk,i−1 + µk u∗
k,i [dk (i) − uk,i ψk,i−1 ]
aℓk wℓ,i−1
(159)
We further note that the adaptive ATC and CTA strategies (153)–(154) reduce to the non-cooperative
adaptive solution (22)–(23), where each node k runs its own individual LMS filter, when the coefficients
{aℓk , cℓk } are selected as
aℓk = δℓk = cℓk
(non-cooperative case)
where δℓk denotes the Kronecker delta function:
= (cid:26) 1,
∆
0,
In terms of the combination matrices A and C , this situation corresponds to setting
ℓ = k
otherwise
δℓk
A = IN = C
(non-cooperative case)
(160)
(161)
(162)
5 Performance of Steepest-Descent Diffusion Strategies
Before studying in some detail the mean-square performance of the adaptive diffusion implementations (153)–
(154), and the influence of gradient noise, we examine first the convergence behavior of the steepest-descent
diffusion strategies (134) and (142), which employ the true gradient vectors. Doing so, will help introduce
the necessary notation and highlight some features of the analysis in preparation for the more challenging
treatment of the adaptive strategies in Sec. 6.
5.1 General Diffusion Model
Rather than study the performance of the ATC and CTA steepest-descent strategies (134) and (142) sep-
arately, it is useful to introduce a more general description that includes the ATC and CTA recursions as
special cases. Thus, consider a distributed steepest-descent diffusion implementation of the following general
form for i ≥ 0:
φk,i−1 = Xℓ∈Nk
a1,ℓk wℓ,i−1
ψk,i = φk,i−1 + µk Xℓ∈Nk
wk,i = Xℓ∈Nk
a2,ℓk ψℓ,i
where the scalars {a1,ℓk , cℓk , a2,ℓk } denote three sets of non-negative real coefficients corresponding to the
(ℓ, k) entries of N × N combination matrices {A1 , C, A2 }, respectively. These matrices are assumed to satisfy
the conditions:
cℓk [rdu,ℓ − Ru,ℓ φk,i−1 ]
(164)
(163)
(165)
AT
1
1 = 1, C 1 = 1, AT
2
1 = 1
(166)
37
so that {A1 , A2} are left stochastic and C is right-stochastic, i.e.,
for k = 1, 2, . . . , N :
cℓk ≥ 0,
cℓk = 1,
a2,ℓk ≥ 0,
(167)
a1,ℓk ≥ 0,
a1,ℓk = 1,
a2,ℓk = 1,
a2,ℓk = 0 if ℓ /∈ Nk
cℓk = 0 if ℓ /∈ Nk
a1,ℓk = 0 if ℓ /∈ Nk
NXk=1
NXℓ=1
NXℓ=1
Different choices for {A1 , C, A2 } correspond to different cooperation modes. For example, the choice A1 =
IN and A2 = A corresponds to the ATC implementation (134), while the choice A1 = A and A2 = IN
corresponds to the CTA implementation (142). Likewise, the choice C = IN corresponds to the case in
which the nodes only share weight estimates and the distributed diffusion recursions (163)–(165) become
φk,i−1 = Xℓ∈Nk
ψk,i = φk,i−1 + µk (rdu,k − Ru,k φk,i−1 )
wk,i = Xℓ∈Nk
Furthermore, the choice A1 = A2 = C = IN corresponds to the non-cooperative mode of operation, in which
case the recursions reduce to the classical (stand-alone) steepest-descent recursion [4–7], where each node
minimizes individually its own quadratic cost Jk (w), defined earlier in (97):
a1,ℓk wℓ,i−1
(168)
(169)
(170)
a2,ℓk ψℓ,i
wk,i = wk,i−1 + µk [rdu,k − Ru,k wk,i−1 ] ,
i ≥ 0
(171)
Table 4: Different choices for the combination matrices {A1 , A2 , C } in (163)–(165) correspond to different cooperation
strategies.
A1 A2
A
IN
A
IN
IN
A
A
IN
IN
IN
C
Cooperation Mode
C
ATC strategy (134).
IN ATC strategy (137) without information exchange.
C
CTA strategy (142).
IN CTA strategy (145) without information exchange.
non-cooperative steepest-descent (171).
IN
5.2 Error Recursions
Our ob jective is to examine whether, and how fast, the weight estimates {wk,i } from the distributed imple-
mentation (163)–(165) converge towards the solution wo of (92)–(93). To do so, we introduce the M × 1
error vectors:
eφk,i
eψk,i
ewk,i
∆= wo − φk,i
∆= wo − ψk,i
∆= wo − wk,i
38
(172)
(173)
(174)
Each of these error vectors measures the residual relative to the desired minimizer wo . Now recall from (100)
that
(176)
(177)
(178)
(175)
rdu,k = Ru,k wo
Then, subtracting wo from both sides of the relations in (163)–(165) we get
eφk,i−1 = Xℓ∈Nk
a1,ℓk ewℓ,i−1
cℓk Ru,ℓ! eφk,i−1
eψk,i = IM − µk Xℓ∈Nk
ewk,i = Xℓ∈Nk
a2,ℓk eψℓ,i
We can describe these relations more compactly by collecting the information from across the network into
block vectors and matrices. We collect the error vectors from across all nodes into the following N × 1 block
vectors, whose individual entries are of size M × 1 each:
∆=
∆=
∆=
eφ1,i
eψ1,i
ew1,i
eφ2,i
eψ2,i
ew2,i
eφi
eψi
ewi
...
...
...
eφN ,i
eψN ,i
ewN ,i
The block quantities { eψi , eφi , ewi } represent the state of the errors across the network at time i. Likewise, we
introduce the following N × N block diagonal matrices, whose individual entries are of size M × M each:
M ∆= diag{ µ1 IM , µ2 IM , . . . , µN IM }
(180)
cℓN Ru,ℓ )
R ∆= diag ( Xℓ∈N1
cℓ2 Ru,ℓ , . . . , Xℓ∈NN
cℓ1 Ru,ℓ , Xℓ∈N2
Each block diagonal entry of R, say, the k -th entry, contains the combination of the covariance matrices in
the neighborhood of node k . We can simplify the notation by denoting these neighborhood combinations as
follows:
∆= Xℓ∈Nk
cℓk Ru,ℓ
(182)
(181)
,
,
(179)
Rk
so that R becomes
R ∆= diag { R1 , R2 , . . . , RN }
(when C 6= I )
In the special case when C = IN , the matrix R reduces to
Ru = diag{Ru,1 , Ru,2 , . . . , Ru,N }
(when C = I )
(183)
(184)
with the individual covariance matrices appearing on its diagonal; we denote R by Ru in this special case.
We further introduce the Kronecker products
A1
∆= A1 ⊗ IM ,
A2
∆= A2 ⊗ IM
(185)
39
The matrix A1 is an N × N block matrix whose (ℓ, k) block is equal to a1,ℓk IM . Likewise, for A2 .
In
other words, the Kronecker transformation defined above simply replaces the matrices {A1 , A2 } by block
matrices {A1 , A2} where each entry {a1,ℓk , a2,ℓk } in the original matrices is replaced by the diagonal matrices
{a1,ℓk IM , a2,ℓk IM }. For ease of reference, Table 5 lists the various symbols that have been defined so far,
and others that will be defined in the sequel.
Table 5: Definitions of network variables used throughout the analysis.
Variable
Equation
cℓk Ru,ℓ
A1 = A1 ⊗ IM
A2 = A2 ⊗ IM
C = C ⊗ IM
Rk = Xℓ∈Nk
R = diag {R1 , R2 , . . . , RN }
Ru = diag{Ru,1 , Ru,2 , . . . , Ru,N }
Rv = diag{σ 2
v,1 , σ 2
v,2 , . . . , σ 2
v,N }
M = diag{ µ1 IM , µ2 IM , . . . , µN IM }
v,2Ru,2 , . . . , σ 2
v,1Ru,1 , σ 2
S = diag{σ 2
v,N Ru,N }
G = AT
2 MC T
2 (INM − MR) AT
B = AT
1
Y = GS G T
F ≈ BT ⊗ B∗
Jk = diag{ 0M , . . . , 0M , IM , 0M , . . . , 0M }
Tk = diag{ 0M , . . . , 0M , Ru,k , 0M , . . . , 0M }
(185)
(185)
(245)
(182)
(183)
(241)
(319)
(180)
(241)
(263)
(264)
(280)
(277)
(294)
(298)
(186)
(187)
(188)
Returning to (176)–(178), we conclude that the following relations hold for the block quantities:
eφi−1 = AT
1 ewi−1
eψi = (IN M − MR) eφi−1
2 eψi
ewi = AT
so that the network weight error vector, ewi , ends up evolving according to the following dynamics:
2 (IN M − MR) AT
ewi = AT
i ≥ 0
1 ewi−1 ,
(diffusion strategy)
For comparison purposes, if each node in the network minimizes its own cost function, Jk (w), separately
from the other nodes and uses the non-cooperative steepest-descent strategy (171), then the weight error
vector across all N nodes would evolve according to the following alternative dynamics:
ewi = (IN M − MRu ) ewi−1 ,
where the matrices A1 and A2 do not appear, and R is replaced by Ru from (184). This recursion is a
special case of (189) when A1 = A2 = C = IN .
(non-cooperative strategy)
(189)
(190)
i ≥ 0
40
5.3 Convergence Behavior
Note from (189) that the evolution of the weight error vector involves block vectors and block matrices; this
will be characteristic of the distributed implementations that we consider in this article. To examine the
stability and convergence properties of recursions that involve such block quantities, it becomes useful to
rely on a certain block vector norm. In App. D, we describe a so-called block maximum block and establish
some of its useful properties. The results of the appendix will be used extensively in our exposition. It is
therefore advisable for the reader to review the properties stated in the appendix at this stage.
Using the result of Lemma D.6, we can establish the following useful statement about the convergence of
the steepest-descent diffusion strategy (163)–(165). The result establishes that all nodes end up converging
to the optimal solution wo if the nodes employ positive step-sizes µk that are small enough; the lemma
provides a sufficient bound on the {µk }.
Theorem 5.1. (Convergence to Optimal Solution) Consider the problem of optimizing the global cost
(92) with the individual cost functions given by (93). Pick a right stochastic matrix C and left stochastic ma-
trices A1 and A2 satisfying (166) or (167); these matrices define the network topology and how information is
shared over neighborhoods. Assume each node in the network runs the (distributed) steepest-descent diffusion
algorithm (163)–(165). Then, al l estimates {wk,i } across the network converge to the optimal solution wo if
the positive step-size parameters {µk } satisfy
µk <
2
λmax (Rk )
(191)
where the neighborhood covariance matrix Rk is defined by (182).
Proof. The weight error vector ewi converges to zero if, and only if, the coefficient matrix AT
2 (INM − MR) AT
1 in
(189) is a stable matrix (meaning that all its eigenvalues lie strictly inside the unit disc). From property (605)
established in App. D, we know that AT
2 (INM − MR) AT
1 is stable if the block diagonal matrix (INM − MR) is
stable. It is now straightforward to verify that condition (191) ensures the stability of (INM − MR). It follows that
ewi −→ 0 as
Observe that the stability condition (191) does not depend on the specific combination matrices A1 and A2 .
Thus, as long as these matrices are chosen to be left-stochastic, the weight-error vectors will converge to
zero under condition (191) no matter what {A1 , A2} are. Only the combination matrix C influences the
condition on the step-size through the neighborhood covariance matrices {Rk }. Observe further that the
statement of the lemma does not require the network to be connected. Moreover, when C = I , in which
case the nodes only share weight estimates and do not share the neighborhood moments {rdu,ℓ , Ru,ℓ}, as in
(168)–(170), condition (191) becomes
i −→ ∞
(192)
µk <
2
λmax (Ru,k )
(cooperation with C = I )
(193)
in terms of the actual covariance matrices {Ru,k }. Results (191) and (193) are reminiscent of a classical
result for stand-alone steepest-descent algorithms, as in the non-cooperative case (171), where it is known
that the estimate by each individual node in this case will converge to wo if, and only if, its positive step-size
satisfies
µk <
2
λmax (Ru,k )
(non-cooperative case (171) with A1 = A2 = C = IN )
(194)
41
This is the same condition as (193) for the case C = I .
The following statement provides a bi-directional statement that ensures convergence of the (distributed)
steepest-descent diffusion strategy (163)–(165) for any choice of left-stochastic combination matrices A1 and
A2 .
Theorem 5.2. (Convergence for Arbitrary Combination Matrices) Consider the problem of opti-
mizing the global cost (92) with the individual cost functions given by (93). Pick a right stochastic matrix
C satisfying (166). Then, the estimates {wk,i } generated by (163)–(165) converge to wo , for al l choices of
left-stochastic matrices A1 and A2 satisfying (166) if, and only if,
µk <
2
λmax (Rk )
(195)
Proof. The result follows from property (b) of Corollary D.1, which is established in App. D.
More importantly, we can verify that under fairly general conditions, employing the steepest-descent
diffusion strategy (163)–(165) enhances the convergence rate of the error vector towards zero relative to the
non-cooperative strategy (171). The next three results establish this fact when C is a doubly stochastic
matrix, i.e., it has non-negative entries and satisfies
C 1 = 1,
C T 1 = 1
(196)
with both its rows and columns adding up to one. Compared to the earlier right-stochastic condition on C
in (105), we are now requiring
Xℓ∈Nk
Xℓ∈Nk
For example, these conditions are satisfied when C is right stochastic and symmetric. They are also satisfied
for C = I , when only weight estimates are shared as in (168)–(170); this latter case covers the ATC and
CTA diffusion strategies (137) and (145), which do not involve information exchange.
ckℓ = 1,
cℓk = 1
(197)
Theorem 5.3. (Convergence Rate is Enhanced: Uniform Step-Sizes) Consider the problem of opti-
mizing the global cost (92) with the individual cost functions given by (93). Pick a doubly stochastic matrix C
satisfying (196) and left stochastic matrices A1 and A2 satisfying (166). Consider two modes of operation. In
one mode, each node in the network runs the (distributed) steepest-descent diffusion algorithm (163)–(165).
In the second mode, each node operates individual ly and runs the non-cooperative steepest-descent algorithm
(171). In both cases, the positive step-sizes used by al l nodes are assumed to be the same, say, µk = µ for
al l k , and the value of µ is chosen to satisfy the required stability conditions (191) and (194), which are met
by selecting
1≤k≤N (cid:26)
λmax (Ru,k ) (cid:27)
2
µ < min
(198)
It then holds that the magnitude of the error vector, k ewi k, in the diffusion case decays to zero more rapid ly
than in the non-cooperative case. In other words, diffusion cooperation enhances convergence rate.
Proof. Let us first establish that any positive step-size µ satisfying (198) will satisfy both stability conditions (191)
and (194). It is obvious that (194) is satisfied. We verify that (191) is also satisfied when C is doubly stochastic. In
this case, each neighborhood covariance matrix, Rk , becomes a convex combination of individual covariance matrices
{Ru,ℓ }, i.e.,
Rk = Xℓ∈Nk
(when C is doubly stochastic)
where now
cℓkRu,ℓ
cℓk = 1
Xℓ∈Nk
42
θmf (Xm )
To proceed, we recall that the spectral norm (maximum singular value) of any matrix X is a convex function of
X [56]. Moreover, for Hermitian matrices X , their spectral norms coincide with their spectral radii (largest eigenvalue
magnitude). Then, Jensen’s inequality [56] states that for any convex function f (·) it holds that
θmXm! ≤ Xm
f Xm
for Hermitian matrices Xm and nonnegative scalars θm that satisfy
Xm
Choosing f (·) as the spectral radius function, and applying it to the definition of Rk above, we get
ρ(Rk ) = ρ Xℓ∈Nk
cℓk Ru,ℓ
≤ Xℓ∈Nk
cℓk · ρ(Ru,ℓ )
ρ(Ru,ℓ )(cid:21)
cℓk · (cid:20) max
≤ Xℓ∈Nk
1≤ℓ≤N
ρ(Ru,ℓ )
= max
1≤ℓ≤N
θm = 1
In other words,
It then follows from (198) that
λmax (Rk ) ≤ max
1≤k≤N
{λmax (Ru,k )}
µ <
2
λmax (Rk )
,
for all k = 1, 2, . . . , N
so that (191) is satisfied as well.
Let us now examine the convergence rate. To begin with, we note that the matrix (INM − MR) that appears in
the weight-error recursion (189) is block diagonal:
(INM − MR) = diag{(IM − µR1 ), (IM − µR2 ), . . . , (IM − µRN )}
cℓk (IM − µRu,ℓ )
and each individual block entry, (IM − µRk ), is a stable matrix since µ satisfies (191). Moreover, each of these entries
can be written as
IM − µRk = Xℓ∈Nk
which expresses (IM − µRk ) as a convex combination of stable terms (IM − µRu,ℓ ). Applying Jensen’s inequality
again we get
ρ Xℓ∈Nk
cℓk (IM − µRu,ℓ ) ≤ Xℓ∈Nk
cℓk ρ(IM − µRu,ℓ )
Now, we know from (189) that the rate of decay of ewi to zero in the diffusion case is determined by the spectral
radius of the coefficient matrix AT
2 (INM − MR) AT
1 . Likewise, we know from (190) that the rate of decay of ewi to
zero in the non-cooperative case is determined by the spectral radius of the coefficient matrix (INM − MRu ). Then,
note that
43
ρ (cid:16)AT
1 (cid:17) (605)
2 (INM − MR) AT
≤
=
=
≤
≤
=
=
=
=
max
1≤k≤N
ρ(INM − MR)
ρ (IM − µRk )
ρ Xℓ∈Nk
cℓk (IM − µRu,ℓ )
max
1≤k≤N
1≤k≤N Xℓ∈Nk
cℓk ρ(IM − µRu,ℓ )
max
ρ(IM − µRu,ℓ )(cid:19)
cℓk (cid:18) max
1≤k≤N Xℓ∈Nk
max
1≤ℓ≤N
1≤k≤N
cℓk
(cid:18) max
ρ(IM − µRu,ℓ )(cid:19) · Xℓ∈Nk
max
1≤ℓ≤N
1≤k≤N (cid:18) max
ρ(IM − µRu,ℓ )(cid:19)
max
1≤ℓ≤N
max
ρ(IM − µRu,ℓ )
1≤ℓ≤N
ρ(INM − MRu )
Therefore, the spectral radius of AT
2 (INM − MR) AT
1 is at most as large as the largest individual spectral radius in
the non-cooperative case.
The argument can be modified to handle different step-sizes across the nodes if we assume uniform covariance
data across the network, as stated below.
Theorem 5.4. (Convergence Rate is Enhanced: Uniform Covariance Data) Consider the same
setting of Theorem 5.3. Assume the covariance data are uniform across al l nodes, say, Ru,k = Ru is
independent of k . Assume further that the nodes in both modes of operation employ steps-sizes µk that are
chosen to satisfy the required stability conditions (191) and (194), which in this case are met by:
2
µk <
,
k = 1, 2, . . . , N
(199)
λmax (Ru )
It then holds that the magnitude of the error vector, k ewi k, in the diffusion case decays to zero more rapid ly
than in the non-cooperative case. In other words, diffusion enhances convergence rate.
Proof. Since Ru,ℓ = Ru for all ℓ and C is doubly stochastic, we get Rk = Ru and INM − MR = INM − MRu . Then,
ρ (cid:16)AT
1 (cid:17) (605)
2 (INM − MR) AT
≤
=
ρ(INM − MRu )
ρ(INM − MR)
The next statement considers the case of ATC and CTA strategies (137) and (145) without information
exchange, which correspond to the case C = IN . The result establishes that these strategies always enhance
the convergence rate over the non-cooperative case, without the need to assume uniform step-sizes or uniform
covariance data.
Theorem 5.5. (Convergence Rate is Enhanced when C = I ) Consider the problem of optimizing the
global cost (92) with the individual cost functions given by (93). Pick left stochastic matrices A1 and A2
satisfying (166) and set C = IN . This situation covers the ATC and CTA strategies (137) and (145), which
44
do not involve information exchange. Consider two modes of operation. In one mode, each node in the
network runs the (distributed) steepest-descent diffusion algorithm (168)–(170). In the second mode, each
node operates individual ly and runs the non-cooperative steepest-descent algorithm (171). In both cases, the
positive step-sizes are chosen to satisfy the required stability conditions (193) and (194), which in this case
are met by
2
,
k = 1, 2, . . . , N
(200)
µk <
λmax (Ru,k )
It then holds that the magnitude of the error vector, k ewi k, in the diffusion case decays to zero more rapid ly
than in the non-cooperative case. In other words, diffusion cooperation enhances convergence rate.
Proof. When C = IN , we get Rk = Ru,k and, therefore, R = Ru and INM − MR = INM − MRu . Then,
ρ (cid:16)AT
1 (cid:17) (605)
2 (INM − MR) AT
≤
=
ρ(INM − MRu )
ρ(INM − MR)
The results of the previous theorems highlight the following important facts about the role of the combination
matrices {A1 , A2 , C } in the convergence behavior of the diffusion strategy (163)–(165):
(a) The matrix C influences the stability of the network through its influence on the bound in (191). This
is because the matrices {Rk } depend on the entries of C . The matrices {A1 , A2} do not influence
network stability.
(b) The matrices {A1 , A2 , C } influence the rate of convergence of the network since they influence the
spectral radius of the matrix AT
2 (IN M − MR) AT
1 , which controls the dynamics of the weight error
vector in (189).
6 Performance of Adaptive Diffusion Strategies
We now move on to examine the behavior of the adaptive diffusion implementations (153)–(154), and the
influence of both gradient noise and measurement noise on convergence and steady-state performance. Due
to the random nature of the perturbations, it becomes necessary to evaluate the behavior of the algorithms
on average, using mean-square convergence analysis. For this reason, we shall study the convergence of the
weight estimates both in the mean and mean-square sense. To do so, we will again consider a general diffusion
structure that includes the ATC and CTA strategies (153)–(154) as special cases. We shall further resort
to the boldface notation to refer to the measurements and weight estimates in order to highlight the fact
that they are now being treated as random variables. In this way, the update equations becomes stochastic
updates. Thus, consider the following general adaptive diffusion strategy for i ≥ 0:
φk,i−1 = Xℓ∈Nk
a1,ℓk wℓ,i−1
ψk,i = φk,i−1 + µk Xℓ∈Nk
wk,i = Xℓ∈Nk
a2,ℓk ψ ℓ,i
As before, the scalars {a1,ℓk , cℓk , a2,ℓk } are non-negative real coefficients corresponding to the (ℓ, k) entries
of N × N combination matrices {A1 , C, A2 }, respectively. These matrices are assumed to satisfy the same
conditions (166) or (167). Again, different choices for {A1 , C, A2 } correspond to different cooperation modes.
For example, the choice A1 = IN and A2 = A corresponds to the adaptive ATC implementation (153), while
the choice A1 = A and A2 = IN corresponds to the adaptive CTA implementation (154). Likewise, the
ℓ,i (cid:2)dℓ (i) − uℓ,i φk,i−1 (cid:3)
cℓk u∗
(202)
(201)
(203)
45
choice C = IN corresponds to the case in which the nodes only share weight estimates and the distributed
diffusion recursions (201)–(203) become
φk,i−1 = Xℓ∈Nk
a1,ℓk wℓ,i−1
k,i (cid:2)dk (i) − uk,i φk,i−1 (cid:3)
ψk,i = φk,i−1 + µk u∗
wk,i = Xℓ∈Nk
a2,ℓk ψ ℓ,i
Furthermore, the choice A1 = A2 = C = IN corresponds to the non-cooperative mode of operation, where
each node runs the classical (stand-alone) least-mean-squares (LMS) filter independently of the other nodes:
[4–7]:
(204)
(205)
(206)
wk,i = wk,i−1 + µk uk,i [dk (i) − uk,i wk,i−1 ] ,
i ≥ 0
(207)
6.1 Data Model
When we studied the performance of the steepest-descent diffusion strategy (163)–(165) we exploited result
(175), which indicated how the moments {rdu,k , Ru,k } that appeared in the recursions related to the optimal
solution wo . Likewise, in order to be able to analyze the performance of the adaptive diffusion strategy
(201)–(203), we need to know how the data {dk (i), uk,i } across the network relate to wo . Motivated by the
several examples presented earlier in Sec. 2, we shall assume that the data satisfy a linear model of the form:
dk (i) = uk,iwo + vk (i)
where vk (i) is measurement noise with variance σ2
v,k :
σ2
v,k
∆= E vk (i)2
(208)
(209)
and where the stochastic processes {dk (i), uk,i } are assumed to be jointly wide-sense stationary with mo-
ments:
σ2
d,k
Ru,k
rdu,k
∆= E dk (i)2
∆= E u∗
k,iuk,i > 0
∆= E dk (i)u∗
k,i
(scalar)
(M × M )
(M × 1)
(210)
(211)
(212)
All variables are assumed to be zero-mean. Furthermore, the noise process {vk (i)} is assumed to be tempo-
rally white and spatially independent, as described earlier by (6), namely,
(cid:26) E vk (i)v∗
k (j ) = 0,
E vk (i)v∗
m (j ) = 0,
The noise process vk (i) is further assumed to be independent of the regression data um,j for all k , m and
i, j so that:
for all i 6= j (temporal whiteness)
for all i, j whenever k 6= m (spatial whiteness)
(213)
E v k (i)u∗
m,j = 0,
We shall also assume that the regression data are temporally white and spatially independent so that:
for all k , m, i, j
(214)
E u∗
k,iuℓ,j = Ru,k δkℓ δij
(215)
Although we are going to derive performance measures for the network under this independence assumption
on the regression data, it turns out that the resulting expressions continue to match well with simulation
results for sufficiently small step-sizes, even when the independence assumption does not hold (in a manner
similar to the behavior of stand-alone adaptive filters) [4, 5].
46
6.2 Performance Measures
Our ob jective is to analyze whether, and how fast, the weight estimates {wk,i } from the adaptive diffusion
implementation (201)–(203) converge towards wo . To do so, we again introduce the M × 1 weight error
vectors:
(216)
(217)
(218)
∆= wo − φk,i
∆
= wo − ψk,i
∆= wo − wk,i
eφk,i
eψk,i
ewk,i
Each of these error vectors measures the residual relative to the desired wo in (208). We further introduce
two scalar error measures:
∆= dk (i) − uk,iwk,i−1
∆= uk,i ewk,i−1
The first error measures how well the term uk,iwk,i−1 approximates the measured data, dk (i); in view of
(208), this error can be interpreted as an estimator for the noise term vk (i). If node k is able to estimate wo
well, then ek (i) would get close to v k (i). Therefore, under ideal conditions, we would expect the variance of
ek (i) to tend towards the variance of vk (i). However, as remarked earlier in (31), there is generally an offset
term for adaptive implementations that is captured by the variance of the a-priori error, ea,k (i). This second
error measures how well uk,iwk,i−1 approximates the uncorrupted term uk,iwo . Using the data model (208),
we can relate {ek (i), ea,k (i)} as
(a-priori error)
(output error)
ea,k (i)
(219)
(220)
ek (i)
ek (i) = ea,k + vk (i)
(221)
Since the noise component, v k (i), is assumed to be zero-mean and independent of all other random variables,
we recover (31):
E ek (i)2 = E ea,k (i)2 + σ2
v,k
(222)
This relation confirms that the variance of the output error, ek (i), is always at least as large as σ2
v,k and
away from it by an amount that is equal to the variance of the a-priori error, ea,k (i). Accordingly, in order
to quantify the performance of any particular node in the network, we define the mean-square-error (MSE)
and excess-mean-square-error (EMSE) for node k as the following steady-state measures:
Then, it holds that
MSEk
EMSEk
∆= lim
i→∞
∆= lim
i→∞
E ek (i)2
E ea,k (i)2
MSEk = EMSEk + σ2
v,k
(223)
(224)
(225)
Therefore, the EMSE term quantifies the size of the offset in the MSE performance of each node. We also
define the mean-square-deviation (MSD) of each node as the steady-state measure:
E k ewk,i k2
which measures how far wk,i is from wo in the mean-square-error sense.
We indicated earlier in (36)–(37) how the MSD and EMSE of stand-alone LMS filters in the non-
cooperative case depend on {µk , σ2
v , Ru,k }. In this section, we examine how cooperation among the nodes
∆= lim
i→∞
MSDk
(226)
47
influences their performance. Since cooperation couples the operation of the nodes, with data originating
from one node influencing the behavior of its neighbors and their neighbors, the study of the network per-
formance requires more effort than in the non-cooperative case. Nevertheless, when all is said and done,
we will arrive at expressions that approximate well the network performance and reveal some interesting
conclusions.
6.3 Error Recursions
(229)
(227)
(228)
cℓk u∗
ℓ,iv ℓ (i)
Using the data model (208) and subtracting wo from both sides of the relations in (201)–(203) we get
eφk,i−1 = Xℓ∈Nk
a1,ℓk ew ℓ,i−1
ℓ,iuℓ,i! eφk,i−1 − µk Xℓ∈Nk
eψk,i = IM − µk Xℓ∈Nk
cℓk u∗
ewk,i = Xℓ∈Nk
a2,ℓk eψ ℓ,i
Comparing the second recursion with the corresponding recursion in the steepest-descent case (176)–(178),
we see that two new effects arise: the effect of gradient noise, which replaces the covariance matrices Ru,ℓ
by the instantaneous approximation u∗
ℓ,iuℓ,i , and the effect of measurement noise, v ℓ(i).
We again describe the above relations more compactly by collecting the information from across the
network in block vectors and matrices. We collect the error vectors from across all nodes into the following
N × 1 block vectors, whose individual entries are of size M × 1 each:
∆=
∆=
∆=
eψ1,i
eφ1,i
ew1,i
eψ2,i
eφ2,i
ew2,i
eψ i
eφi
ewi
...
...
...
eψN ,i
eφN ,i
ewN ,i
The block quantities { eψi , eφi , ewi } represent the state of the errors across the network at time i. Likewise,
we introduce the following N × N block diagonal matrices, whose individual entries are of size M × M each:
M ∆= diag{ µ1 IM , µ2 IM , . . . , µN IM }
(231)
ℓ,iuℓ,i )
∆= diag ( Xℓ∈N1
ℓ,iuℓ,i , . . . , Xℓ∈NN
ℓ,iuℓ,i , Xℓ∈N2
cℓN u∗
cℓ2 u∗
cℓ1 u∗
Each block diagonal entry of Ri , say, the k -th entry, contains a combination of rank-one regression terms
collected from the neighborhood of node k . In this way, the matrix Ri is now stochastic and dependent
on time, in contrast to the matrix R in the steepest-descent case in (181), which was a constant matrix.
Nevertheless, it holds that
(230)
(232)
Ri
,
,
E Ri = R
(233)
so that, on average, Ri agrees with R. We can simplify the notation by denoting the neighborhood combi-
nations as follows:
∆= Xℓ∈Nk
∆= diag { R1,i , R2,i , . . . , RN ,i }
so that Ri becomes
cℓk u∗
ℓ,iuℓ,i
(when C 6= I )
(234)
(235)
Rk,i
Ri
48
Again, compared with the matrix Rk defined in (182), we find that Rk,i is now both stochastic and time-
dependent. Nevertheless, it again holds that
E Rk,i = Rk
In the special case when C = I , the matrix Ri reduces to
Ru,i
∆= diag{u∗
1,iu1,i , u∗
2,iu2,i , . . . , u∗
N ,iuN ,i}
(when C = I )
with
E Ru,i = Ru
(236)
(237)
(238)
where Ru was defined earlier in (184).
We further introduce the following N × 1 block column vector, whose entries are of size M × 1 each:
si
∆= col{ u∗
2,iv 2 (i), . . . , u∗
1,iv1 (i), u∗
N ,ivN (i) }
(239)
Obviously, given that the regression data and measurement noise are zero-mean and independent of each
other, we have
and the covariance matrix of si is N × N block diagonal with blocks of size M × M :
E si = 0
S ∆= E si s∗
v,2Ru,2 , . . . , σ2
v,1Ru,1 , σ2
i = diag{σ2
v,N Ru,N }
(240)
(241)
(242)
where
Returning to (227)–(229), we conclude that the following relations hold for the block quantities:
eφi−1 = AT
1 ew i−1
eψ i = (IN M − MRi ) eφi−1 − MC T si
2 eψ i
ewi = AT
C ∆
= C ⊗ IM
(245)
so that the network weight error vector, ewi , ends up evolving according to the following stochastic recursion:
2 (IN M − MRi ) AT
ew i = AT
2 MC T si ,
1 ew i−1 − AT
i ≥ 0
(diffusion strategy)
(246)
For comparison purposes, if each node operates individually and uses the non-cooperative LMS recursion
(207), then the weight error vector across all N nodes would evolve according to the following stochastic
recursion:
(243)
(244)
ewi = (IN M − MRu,i ) ewi−1 − Msi ,
where the matrices A1 and A2 do not appear, and Ri is replaced by Ru,i from (237).
i ≥ 0
(non-cooperative strategy)
(247)
49
6.4 Convergence in the Mean
i ≥ 0
(248)
(diffusion strategy)
Taking expectations of both sides of (246) we find that:
E ewi = AT
2 (IN M − MR) AT
1 · E ewi−1 ,
where we used the fact that ewi−1 and Ri are independent of each other in view of our earlier assumptions
on the regression data and noise in Sec. 6.1. Comparing with the error recursion (189) in the steepest-
descent case, we find that both recursions are identical with ewi replaced by E ew i . Therefore, the convergence
statements from the steepest-descent case can be extended to the adaptive case to provide conditions on the
step-size to ensure stability in the mean, i.e., to ensure
E ewi −→ 0 as
i −→ ∞
(249)
When (249) is guaranteed, we would say that the adaptive diffusion solution is asymptotically unbiased. The
following statements restate the results of Theorems 5.1–5.5 in the context of mean error analysis.
Theorem 6.1. (Convergence in the Mean) Consider the problem of optimizing the global cost (92) with
the individual cost functions given by (93). Pick a right stochastic matrix C and left stochastic matrices A1
and A2 satisfying (166) or (167). Assume each node in the network measures data that satisfy the conditions
described in Sec. 6.1, and runs the adaptive diffusion algorithm (201)–(203). Then, al l estimators {wk,i }
across the network converge in the mean to the optimal solution wo if the positive step-size parameters {µk }
satisfy
µk <
2
λmax (Rk )
(250)
where the neighborhood covariance matrix Rk is defined by (182). In other words, E wk,i → wo for al l nodes
k as i → ∞.
(cid:3)
Observe again that the mean stability condition (250) does not depend on the specific combination matrices
A1 and A2 that are being used. Only the combination matrix C influences the condition on the step-size
through the neighborhood covariance matrices {Rk }. Observe further that the statement of the lemma does
not require the network to be connected. Moreover, when C = IN , in which case the nodes only share weight
estimators and do not share neighborhood data {dℓ (i), uℓ,i } as in (204)–(206), condition (250) becomes
µk <
2
λmax (Ru,k )
(adaptive cooperation with C = IN )
(251)
Results (250) and (251) are reminiscent of a classical result for the stand-alone LMS algorithm, as in the
non-cooperative case (207), where it is known that the estimator by each individual node in this case would
converge in the mean to wo if, and only if, its step-size satisfies
µk <
2
λmax (Ru,k )
(non-cooperative adaptation)
(252)
The following statement provides a bi-directional result that ensures the mean convergence of the adaptive
diffusion strategy for any choice of left-stochastic combination matrices A1 and A2 .
50
Theorem 6.2. (Mean Convergence for Arbitrary Combination Matrices) Consider the problem of
optimizing the global cost (92) with the individual cost functions given by (93). Pick a right stochastic matrix
C satisfying (166). Assume each node in the network measures data that satisfy the conditions described in
Sec. 6.1. Then, the estimators {wk,i } generated by the adaptive diffusion strategy (201)–(203), converge in
the mean to wo , for al l choices of left stochastic matrices A1 and A2 satisfying (166) if, and only if,
µk <
2
λmax (Rk )
(253)
(cid:3)
As was the case with steepest-descent diffusion strategies, the adaptive diffusion strategy (201)–(203) also
enhances the convergence rate of the mean of the error vector towards zero relative to the non-cooperative
strategy (207). The next results restate Theorems 5.3–5.5; they assume C is a doubly stochastic matrix.
Theorem 6.3. (Mean Convergence Rate is Enhanced: Uniform Step-Sizes) Consider the problem
of optimizing the global cost (92) with the individual cost functions given by (93). Pick a doubly stochastic
matrix C satisfying (196) and left stochastic matrices A1 and A2 satisfying (166). Assume each node in the
network measures data that satisfy the conditions described in Sec. 6.1. Consider two modes of operation.
In one mode, each node in the network runs the adaptive diffusion algorithm (201)–(203). In the second
mode, each node operates individual ly and runs the non-cooperative LMS algorithm (207). In both cases, the
positive step-sizes used by al l nodes are assumed to be the same, say, µk = µ for al l k , and the value of µ is
chosen to satisfy the required mean stability conditions (250) and (252), which are met by selecting
1≤k≤N (cid:26)
λmax (Ru,k ) (cid:27)
2
µ < min
(254)
It then holds that the magnitude of the mean error vector, kE ewi k in the diffusion case decays to zero more
rapid ly than in the non-cooperative case. In other words, diffusion enhances convergence rate.
(cid:3)
Theorem 6.4. (Mean Convergence Rate is Enhanced: Uniform Covariance Data) Consider the
same setting of Theorem 6.3. Assume the covariance data are uniform across al l nodes, say, Ru,k = Ru is
independent of k . Assume further that the nodes in both modes of operation employ steps-sizes µk that are
chosen to satisfy the required stability conditions (250) and (252), which in this case are met by:
2
µk <
,
k = 1, 2, . . . , N
(255)
λmax (Ru )
It then holds that the magnitude of the mean error vector, kE ewik, in the diffusion case also decays to zero
more rapid ly than in the non-cooperative case. In other words, diffusion enhances convergence rate.
(cid:3)
The next statement considers the case of ATC and CTA strategies (204)–(206) without information exchange,
which correspond to the choice C = IN . The result establishes that these strategies always enhance the
convergence rate over the non-cooperative case, without the need to assume uniform step-sizes or uniform
covariance data.
Theorem 6.5. (Mean Convergence Rate is Enhanced when C = I ) Consider the problem of opti-
mizing the global cost (92) with the individual cost functions given by (93). Pick left stochastic matrices A1
and A2 satisfying (166) and set C = IN . This situation covers the ATC and CTA strategies (204)–(206)
that do not involve information exchange. Assume each node in the network measures data that satisfy the
conditions described in Sec. 6.1. Consider two modes of operation. In one mode, each node in the network
51
runs the adaptive diffusion algorithm (204)–(206). In the second mode, each node operates individual ly and
runs the non-cooperative LMS algorithm (207). In both cases, the positive step-sizes are chosen to satisfy
the required stability conditions (251) and (252), which in this case are met by
2
µk <
,
k = 1, 2, . . . , N
(256)
λmax (Ru,k )
It then holds that the magnitude of the mean error vector, kE ewik, in the diffusion case decays to zero more
rapid ly than in the non-cooperative case. In other words, diffusion cooperation enhances convergence rate.
(cid:3)
The results of the previous theorems again highlight the following important facts about the role of the
combination matrices {A1 , A2 , C } in the convergence behavior of the adaptive diffusion strategy (201)–(203):
(a) The matrix C influences the mean stability of the network through its influence on the bound in (250).
This is because the matrices {Rk } depend on the entries of C . The matrices {A1 , A2} do not influence
network mean stability.
(b) The matrices {A1 , A2 , C } influence the rate of convergence of the mean weight-error vector over the
network since they influence the spectral radius of the matrix AT
2 (IN M − MR) AT
1 , which controls the
dynamics of the weight error vector in (248).
6.5 Mean-Square Stability
It is not sufficient to ensure the stability of the weight-error vector in the mean sense. The error vectors,
wk,i , may be converging on average to zero but they may have large fluctuations around the zero value.
We therefore need to examine how small the error vectors get. To do so, we perform a mean-square-error
analysis. The purpose of the analysis is to evaluate how the variances E k ewk,i k2 evolve with time and what
their steady-state values are, for each node k .
In this section, we are particularly interested in evaluating the evolution of two mean-square-errors,
namely,
E ea,k (i)2
E k ewk,i k2
(257)
and
The steady-state values of these quantities determine the MSD and EMSE performance levels at node k
and, therefore, convey critical information about the performance of the network. Under the independence
assumption on the regression data from Sec. 6.1, it can be verified that the EMSE variance can be written
as:
E ea,k (i)2
∆= E uk,i ewk,i−1 2
= E ew∗
k,i−1u∗
k,iuk,i ewk,i−1
k,iuk,i ewk,i−1 ewk,i )(cid:3)
= E (cid:2)E ( ew∗
k,i−1u∗
k,iuk,i (cid:3) ewk,i−1
k,i−1 (cid:2)E u∗
= E ew∗
= E ew∗
k,i−1Ru,k ewk,i−1
= E k ewk,i−1 k2
Ru,k
in terms of a weighted square measure with weighting matrix Ru,k . Here we are using the notation kxk2
Σ to
denote the weighted square quantity x∗Σx, for any column vector x and matrix Σ. Thus, we can evaluate
mean-square-errors of the form (257) by evaluating the means of weighted square quantities of the following
form:
E k ewk,i k2
Σk
for an arbitrary Hermitian nonnegative-definite weighting matrix Σk that we are free to choose. By setting
Σk to different values (say, Σk = I or Σk = Ru,k ), we can extract various types of information about
(258)
(259)
52
the nodes and the network, as the discussion will reveal. The approach we follow is based on the energy
conservation framework of [4, 5, 57].
So, let Σ denote an arbitrary N × N block Hermitian nonnegative-definite matrix that we are free to
choose, with M × M block entries {Σℓk }. Let σ denote the (N M )2 × 1 vector that is obtained by stacking
the columns of Σ on top of each other, written as
σ ∆= vec(Σ)
(260)
In the sequel, it will become more convenient to work with the vector representation σ than with the matrix
Σ itself.
We start from the weight-error vector recursion (246) and re-write it more compactly as:
ewi = Bi ewi−1 − G si ,
where the coefficient matrices Bi and G are short-hand representations for
∆
= AT
2 (IN M − MRi ) AT
1
i ≥ 0
(261)
(262)
Bi
and
G ∆= AT
2 MC T
Note that Bi is stochastic and time-variant, while G is constant. We denote the mean of Bi by
B ∆= E Bi = AT
2 (IN M − MR) AT
1
(263)
(264)
(266)
where R is defined by (181). Now equating weighted square measures on both sides of (261) we get
Σ = kBi ew i−1 − G si k2
k ewik2
Σ
Expanding the right-hand side we find that
Σ = ew∗
B∗
i ΣBi ewi−1 + s∗
i G T ΣG si −
k ewi k2
i−1
ew∗
B∗
i ΣG si − s∗
i G T ΣBi ew i−1
i−1
Under expectation, the last two terms on the right-hand side evaluate to zero so that
i G T ΣG si (cid:1)
i ΣBi ew i−1 (cid:1) + E (cid:0)s∗
Σ = E (cid:0) ew∗
B∗
E k ewi k2
i−1
Let us evaluate each of the expectations on the right-hand side. The last expectation is given by
i (cid:1)
i G T ΣG si (cid:1) = Tr (cid:0)G T ΣG E si s∗
E (cid:0)s∗
= Tr (cid:0)G T ΣGS (cid:1)
(241)
= Tr (cid:0)ΣGS G T (cid:1)
where S is defined by (241) and where we used the fact that Tr(AB ) = Tr(BA) for any two matrices A and
B of compatible dimensions. With regards to the first expectation on the right-hand side of (267), we have
i ΣBi ewi−1 ewi−1 (cid:1)(cid:3)
i ΣBi ewi−1 (cid:1) = E (cid:2)E (cid:0) ew∗
E (cid:0) ew∗
B∗
B∗
i−1
i−1
= E ew∗
i−1 [E (B∗
i ΣBi )] ewi−1
∆= E ew∗
i−1Σ′ ew i−1
= E k ewi−1k2
Σ′
53
(268)
(267)
(265)
(269)
where we introduced the nonnegative-definite weighting matrix
Σ′
∆= E B∗
i ΣBi
(262)
2 (IN M − MRi ) AT
= E A1 (IN M − RiM) A2ΣAT
1
= A1A2ΣAT
2 AT
1 − A1A2ΣAT
2 MRAT
1 − A1RMA2ΣAT
2 AT
1 + O(M2 )
(270)
where R is defined by (181) and the term O(M2 ) denotes the following factor, which depends on the square
of the step-sizes, {µ2
k }:
1 (cid:1)
O(M2 ) = E (cid:0)A1RiMA2ΣAT
2 MRiAT
The evaluation of the above expectation depends on higher-order moments of the regression data. While we
can continue with the analysis by taking this factor into account, as was done in [4, 5, 18, 57], it is sufficient
for the exposition in this article to focus on the case of sufficiently small step-sizes where terms involving
higher powers of the step-sizes can be ignored. Therefore, we continue our discussion by letting
(271)
Σ′ ∆= A1A2ΣAT
2 AT
1 − A1A2ΣAT
2 MRAT
1 − A1RMA2ΣAT
2 AT
1
(272)
The weighting matrix Σ′ is fully defined in terms of the step-size matrix, M, the network topology through
the matrices {A1 , A2 , C }, and the regression statistical profile through R. Expression (272) tells us how to
construct Σ′ from Σ. The expression can be transformed into a more compact and revealing form if we
instead relate the vector forms σ ′ = vec(Σ′ ) and σ = vec(Σ). Using the following equalities for arbitrary
matrices {U, W, Σ} of compatible dimensions [5]:
vec(U ΣW ) = (W T ⊗ U )σ
Tr(ΣW ) = (cid:2)vec(W T )(cid:3)T
and applying the vec operation to both sides of (272) we get
σ ′ = (A1A2 ⊗ A1A2 ) σ − (cid:0)A1RT MA2 ⊗ A1A2 (cid:1) σ − (A1A2 ⊗ A1RMA2 ) σ
σ ′ ∆= F σ
That is,
σ
where we are introducing the coefficient matrix of size (N M )2 × (N M )2 :
F ∆= (A1A2 ⊗ A1A2 ) − (cid:0)A1RT MA2 ⊗ A1A2 (cid:1) − (A1A2 ⊗ A1RMA2 )
A reasonable approximate expression for F for sufficiently small step-sizes is
F ≈ BT ⊗ B ∗
(273)
(274)
(275)
(276)
(277)
Indeed, if we replace B from (264) into (277) and expand terms, we obtain the same factors that appear in
(276) plus an additional term that depends on the square of the step-sizes, {µ2
k }, whose effect can be ignored
for sufficiently small step-sizes.
In this way, using in addition property (274), we find that relation (267) becomes:
Σ′ + (cid:2)vec (cid:0)GS T G T (cid:1)(cid:3)T
E k ewi k2
Σ = E k ewi−1k2
54
(278)
σ
The last term is dependent on the network topology through the matrix G , which is defined in terms of
{A2 , C , M}, and the noise and regression data statistical profile through S . It is convenient to introduce the
alternative notation kxk2
σ to refer to the weighted square quantity kxk2
Σ , where σ = vec(Σ). We shall use
these two notations interchangeably. The convenience of the vector notation is that it allows us to exploit
the simpler linear relation (275) between σ ′ and σ to rewrite (278) as shown in (279) below, with the same
weight vector σ appearing on both sides.
Theorem 6.6. (Variance Relation) Consider the data model of Sec. 6.1 and the independence statistical
conditions imposed on the noise and regression data, including (208)–(215). Assume further sufficiently
smal l step-sizes are used so that terms that depend on higher-powers of the step-sizes can be ignored. Pick
left stochastic matrices A1 and A2 and a right stochastic matrix C satisfying (166). Under these conditions,
the weight-error vector ewi = col{ ewk,i }N
k=1 associated with a network running the adaptive diffusion strategy
(201)–(203) satisfies the fol lowing variance relation
F σ + (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
E k ewi k2
σ = E k ewi−1k2
for any Hermitian nonnegative-definite matrix Σ with σ = vec(Σ), and where {S , G , F } are defined by (241),
(263), and (277), and
(279)
σ
Y ∆= GS G T
(280)
(cid:3)
Note that relation (279) is not an actual recursion; this is because the weighting matrices {σ, F σ} on both
sides of the equality are different. The relation can be transformed into a true recursion by expanding it
into a convenient state-space model; this argument was pursued in [4, 5, 18, 57] and is not necessary for the
exposition here, except to say that stability of the matrix F ensures the mean-square stability of the filter
— this fact is also established further ahead through relation (327). By mean-square stability we mean that
each term E k ewk,i k2 remains bounded over time and converges to a steady-state MSDk value. Moreover, the
spectral radius of F controls the rate of convergence of E k ewi k2 towards its steady-state value.
Theorem 6.7. (Mean-Square Stability) Consider the same setting of Theorem 6.6. The adaptive dif-
fusion strategy (201)–(203) is mean-square stable if, and only if, the matrix F defined by (276), or its
approximation (277), is stable (i.e., al l its eigenvalues lie strictly inside the unit disc). This condition is
satisfied by sufficiently smal l positive step-sizes {µk } that also satisfy:
µk <
2
λmax (Rk )
(281)
where the neighborhood covariance matrix Rk is defined by (182). Moreover, the convergence rate of the
algorithm is determined by the value [ρ(B )]2 (the square of the spectral radius of B ).
Proof. Recall that, for two arbitrary matrices A and B of compatible dimensions, the eigenvalues of the Kronecker
product A ⊗ B is formed of all product combinations λi (A)λj (B ) of the eigenvalues of A and B [19]. Therefore, using
expression (277), we have that ρ(F ) = [ρ(B)]2 . It follows that F is stable if, and only if, B is stable. We already noted
earlier in Theorem 6.1 that condition (281) ensures the stability of B. Therefore, step-sizes that ensure stability in
the mean and are sufficiently small will also ensure mean-square stability.
Remark. More generally, had we not ignored the second-order term (271), the expression for F would have
been the following. Starting from the definition Σ′ = E B∗
i ΣBi , we would get
i (cid:17) σ
σ ′ = (cid:16)E BT
i ⊗ B∗
55
so that
F ∆= E (cid:16)BT
i (cid:17)
i ⊗ B∗
(for general step-sizes)
= (A1 ⊗ A1 ) · nI − (cid:0)RT M ⊗ I (cid:1) − (I ⊗ RM) + E (cid:16)RT
i M ⊗ RiM(cid:17)o · (A2 ⊗ A2 )
Mean-square stability of the filter would then require the step-sizes {µk } to be chosen such that they ensure
the stability of this matrix F (in addition to condition (281) to ensure mean stability).
(282)
(cid:3)
6.6 Network Mean-Square Performance
lim
i→∞
We can now use the variance relation (279) to evaluate the network performance, as well as the performance
of the individual nodes, in steady-state. Since the dynamics is mean-square stable for sufficiently small
step-sizes, we take the limit of (279) as i → ∞ and write:
F σ + (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
E k ewi−1 k2
E k ewi k2
σ = lim
i→∞
Grouping terms leads to the following result.
Corol lary 6.1. (Steady-State Variance Relation) Consider the same setting of Theorem 6.6. The
weight-error vector, ewi = col{ ewk,i }N
k=1 , of the adaptive diffusion strategy (201)–(203) satisfies the fol lowing
relation in steady-state:
(I−F )σ = (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
E k ewik2
lim
i→∞
for any Hermitian nonnegative-definite matrix Σ with σ = vec(Σ), and where {F , Y } are defined by (277)
and (280).
(283)
(284)
σ
σ
Expression (284) is a very useful relation; it allows us to evaluate the network MSD and EMSE through
proper selection of the weighting vector σ (or, equivalently, the weighting matrix Σ). For example, the
network MSD is defined as the average value:
(cid:3)
1
N
1
N
MSDnetwork = lim
i→∞
MSDnetwork ∆= lim
i→∞
NXk=1
E k ewk,i k2
which amounts to averaging the MSDs of the individual nodes. Therefore,
E k ewik2
E k ewik2 = lim
1/N
i→∞
This means that in order to recover the network MSD from relation (284), we should select the weighting
vector σ such that
1
(I − F )σ =
vec (IN M )
N
Solving for σ and substituting back into (284) we arrive at the following expression for the network MSD:
· (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
· (I − F )−1 · vec (IN M )
MSDnetwork =
(287)
1
N
(285)
(286)
56
1
N
E ea,k (i)2
EMSEnetwork
∆= lim
i→∞
Likewise, the network EMSE is defined as the average value
NXk=1
NXk=1
E k ewk,i k2
Ru,k
which amounts to averaging the EMSEs of the individual nodes. Therefore,
E k ewik2
E k ewi k2
diag{Ru,1 ,Ru,2 ,...,Ru,N } = lim
Ru
i→∞
where Ru is the matrix defined earlier by (184), and which we repeat below for ease of reference:
Ru = diag{Ru,1 , Ru,2 , . . . , Ru,N }
EMSEnetwork = lim
i→∞
= lim
i→∞
1
N
1
N
1
N
(288)
(289)
(290)
(291)
(292)
This means that in order to recover the network EMSE from relation (284), we should select the weighting
vector σ such that
vec (Ru )
(I − F )σ =
1
N
Solving for σ and substituting into (284) we arrive at the following expression for the network EMSE:
· (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
6.7 Mean-Square Performance of Individual Nodes
· (I − F )−1 · vec (Ru )
EMSEnetwork =
1
N
We can also assess the mean-square performance of the individual nodes in the network from (284). For
instance, the MSD of any particular node k is defined by
E k ewk,i k2
Introduce the N × N block diagonal matrix with blocks of size M × M , where all blocks on the diagonal are
zero except for an identity matrix on the diagonal block of index k , i.e.,
∆= lim
i→∞
MSDk
(293)
Then, we can express the node MSD as follows:
Jk
∆= diag{ 0M , . . . , 0M , IM , 0M , . . . , 0M }
∆= lim
i→∞
MSDk
E k ewik2
Jk
The same argument that was used to obtain the network MSD then leads to
MSDk = (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
· (I − F )−1 · vec (Jk )
Likewise, the EMSE of node k is defined by
EMSEk
∆= lim
i→∞
= lim
i→∞
57
E ea,k (i)2
E k ewk,i k2
Ru,k
(294)
(295)
(296)
(297)
Introduce the N × N block diagonal matrix with blocks of size M × M , where all blocks on the diagonal are
zero except for the diagonal block of index k whose value is Ru,k , i.e.,
Tk
∆
= diag{ 0M , . . . , 0M , Ru,k , 0M , . . . , 0M }
(298)
Then, we can express the node EMSE as follows:
∆= lim
i→∞
(299)
EMSEk
E k ewi k2
Tk
The same argument that was used to obtain the network EMSE then leads to
EMSEk = (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
· (I − F )−1 · vec (Tk )
We summarize the results in the following statement.
Theorem 6.8. (Network Mean-Square Performance) Consider the same setting of Theorem 6.6. In-
troduce the 1 × (N M )2 row vector hT defined by
hT ∆= (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
where {F , Y } are defined by (277) and (280). Then the network MSD and EMSE and the individual node
performance measures are given by
· (I − F )−1
(300)
(301)
MSDnetwork = hT · vec (IN M ) /N
EMSEnetwork = hT · vec (Ru ) /N
MSDk = hT · vec (Jk )
EMSEk = hT · vec (Tk )
(302)
(303)
(304)
(305)
where {Jk , Tk } are defined by (294) and (298).
(cid:3)
We can obviously recover from the above expressions the performance of the nodes in the non-cooperative
implementation (207), where each node performs its adaptation individually, by setting A1 = A2 = C = IN .
We can express the network MSD, and its EMSE if desired, in an alternative useful form involving a
series representation.
Corol lary 6.2. (Series Representation for Network MSD) Consider the same setting of Theorem 6.6.
The network MSD can be expressed in the fol lowing alternative series expansion form:
∞Xj=0
Tr (cid:0)B j Y B ∗j (cid:1)
MSDnetwork =
(306)
1
N
where
Y = GS G T
G = AT
2 MC T
2 (I − MR)AT
B = AT
1
Proof. Since F is stable when the filter is mean-square stable, we can expand (I − F )−1 as
(I − F )−1
I + F + F 2 + . . .
I + (cid:16)BT ⊗ B∗(cid:17) + (cid:16)BT ⊗ B∗ (cid:17)2
Substituting into (287) and using property (274), we obtain the desired result.
(277)
=
+ . . .
=
(307)
(308)
(309)
58
6.8 Uniform Data Profile
We can simplify expressions (307)–(309) for {Y , G , B} in the case when the regression covariance matrices
are uniform across the network and all nodes employ the same step-size, i.e., when
Ru,k = Ru ,
µk = µ,
for all k
for all k
(uniform covariance profile)
(uniform step-sizes)
and when the combination matrix C is doubly stochastic, so that
C 1 = 1,
C T 1 = 1
(310)
(311)
(312)
We refer to conditions (310)–(312) as corresponding to a uniform data profile environment. The noise
variances, {σ2
v,k }, do not need to be uniform so that the signal-to-noise ratio (SNR) across the network can
still vary from node to node. The simplified expressions derived in the sequel will be useful in Sec. 7 when
we compare the performance of various cooperation strategies.
Thus, under conditions (310)–(312), expressions (180), (181), and (263) for {M, R, G } simplify to
M = µIN M
R = IN ⊗ Ru
G = µAT
2 C T
Substituting these values into expression (309) for B we get
B = AT
2 (I − MR)AT
1
= (AT
2 ⊗ I ) · (I − µ(I ⊗ Ru )) · (AT
1 ⊗ I )
2 ⊗ I )(I ⊗ Ru )(AT
= (AT
2 ⊗ I )(AT
1 ⊗ I ) − µ(AT
1 ⊗ I )
= (AT
2 AT
1 ⊗ I ) − µ(AT
2 AT
1 ⊗ Ru )
= AT
2 AT
1 ⊗ (I − µRu )
where we used the useful Kronecker product identities:
(X + Y ) ⊗ Z = (X ⊗ Z ) + (Y ⊗ Z )
(X ⊗ Y )(W ⊗ Z ) = (XW ⊗ Y Z )
(313)
(314)
(315)
(316)
(317)
(318)
for any matrices {X, Y , Z, W } of compatible dimensions. Likewise, introduce the N × N diagonal matrix
with noise variances:
Rv
∆= diag{σ2
v,2 , . . . , σ2
v,1 , σ2
v,N }
Then, expression (241) for S becomes
v,2Ru , . . . , σ2
v,1Ru , σ2
S = diag{σ2
v,N Ru}
= Rv ⊗ Ru
It then follows that we can simplify expression (307) for Y as:
Y = µ2AT
2 C T S CA2
= µ2 · (AT
2 ⊗ I ) · (C T ⊗ I ) ⊗ (Rv ⊗ Ru ) · (C ⊗ I ) · (A2 ⊗ I )
= µ2 (AT
2 C T RvCA2 ⊗ Ru )
59
(319)
(320)
(321)
Corol lary 6.3. (Network MSD for Uniform Data Profile) Consider the same setting of Theorem 6.6
with the additional requirement that conditions (310)–(312) for a uniform data profile hold. The network
MSD is stil l given by the same series representation (306) where now
Y = µ2 (AT
2 C T RvCA2 ⊗ Ru )
B = AT
2 AT
1 ⊗ (I − µRu )
(322)
(323)
Using these expressions, we can decouple the network MSD expression (306) into two separate factors: one
is dependent on the step-size and data covariance {µ, Ru}, and the other is dependent on the combination
matrices and noise profile {A1 , A2 , C, Rv }:
MSDnetwork =
µ2
N
∞Xj=0
Tr (cid:18)(cid:20)(cid:16)AT
2 C T Rv CA2(cid:17) (A1A2 )j (cid:21) ⊗ h(I − µRu )j Ru (I − µRu )j i(cid:19)
1 (cid:17)j (cid:16)AT
2 AT
(324)
Proof. Using (306) and the given expressions (322)–(323) for {Y , B}, we get
Tr (cid:18)(cid:20)(cid:16)AT
⊗ (I − µRu )j (cid:21) (AT
2 C T Rv CA2 ⊗ Ru ) h(A1A2 )j ⊗ (I − µRu )j i(cid:19)
1 (cid:17)j
∞Xj=0
2 AT
Result (324) follows from property (317).
MSDnetwork =
µ2
N
6.9 Transient Mean-Square Performance
(325)
Before comparing the mean-square performance of various cooperation strategies, we pause to comment that
the variance relation (279) can also be used to characterize the transient behavior of the network, and not
just its steady-state performance. To see this, iterating (279) starting from i = 0, we find that
·
F j σ
iXj=0
F i+1σ + (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
σ = E k ew−1k2
E k ewi k2
where
∆= wo − w−1
ew−1
(326)
in terms of the initial condition, w−1 . If this initial condition happens to be w−1 = 0, then ew−1 = wo .
Comparing expression (325) at time instants i and i − 1 we can relate E k ewi k2
σ and E k ewi−1 k2
σ as follows:
σ + (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
· F iσ − E k ew−1k2
σ = E k ewi−1k2
E k ewi k2
(327)
(I−F )F iσ
This recursion relates the same weighted square measures of the error vectors { ewi , ewi−1 }.
It therefore
describes how these weighted square measures evolve over time. It is clear from this relation that, for mean-
square stability, the matrix F needs to be stable so that the terms involving F i do not grow unbounded.
The learning curve of the network is the curve that describes the evolution of the network EMSE over
time. At any time i, the network EMSE is denoted by ζ (i) and measured as:
NXk=1
NXk=1
60
E k ewk,i k2
Ru,k
E ea,k (i)2
(328)
1
N
1
N
ζ (i)
∆=
=
The above expression indicates that ζ (i) is obtained by averaging the EMSE of the individual nodes at time
i. Therefore,
1
1
E k ewi k2
E k ewi k2
diag{Ru,1 ,Ru,2 ,...,Ru,N } =
Ru
N
N
where Ru is the matrix defined by (290). This means that in order to evaluate the evolution of the network
EMSE from relation (327), we simply select the weighting vector σ such that
ζ (i) =
(329)
Substituting into (327) we arrive at the learning curve for the network.
σ =
1
N
vec (Ru )
(330)
Corol lary 6.4. (Network Learning Curve) Consider the same setting of Theorem 6.6. Let ζ (i) denote
the network EMSE at time i, as defined by (328). Then, the learning curve of the network corresponds to
the evolution of ζ (i) with time and is described by the fol lowing recursion over i ≥ 0:
N (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
1
where {F , Y , Ru } are defined by (277), (280), and (290).
E k w−1k2
(I−F )F ivec(Ru )
· F i · vec (Ru ) −
1
N
ζ (i) = ζ (i − 1) +
(331)
(cid:3)
7 Comparing the Performance of Cooperative Strategies
Using the expressions just derived for the MSD of the network, we can compare the performance of various
cooperative and non-cooperative strategies. Table 6 further ahead summarizes the results derived in this
section and the conditions under which they hold.
7.1 Comparing ATC and CTA Strategies
We first compare the performance of the adaptive ATC and CTA diffusion strategies (153) and (154) when
they employ a doubly stochastic combination matrix A. That is, let us consider the two scenarios:
C, A1 = A, A2 = IN
C, A1 = IN , A2 = A
(adaptive CTA strategy)
(adaptive ATC strategy)
where A is now assumed to be doubly stochastic, i.e.,
A1 = 1,
AT 1 = 1
(332)
(333)
(334)
with its rows and columns adding up to one. For example, these conditions are satisfied when A is left
stochastic and symmetric. Then, expressions (307) and (309) give:
Bcta = (I − MR)AT ,
Batc = AT (I − MR),
Ycta = MC T S CM
Yatc = AT MC T S CMA
where
A = A ⊗ IM
Following [18], introduce the auxiliary nonnegative-definite matrix
∆= (cid:2)(I − MR)AT (cid:3)j
· MC T S CM · (cid:2)(I − MR)AT (cid:3)∗j
61
Hj
(335)
(336)
(337)
(338)
Then, it is immediate to verify from (306) that
MSDnetwork
cta
=
MSDnetwork
atc
=
1
N
1
N
∞Xj=0
∞Xj=0
Tr(Hj )
Tr(AT Hj A)
(339)
(340)
so that
MSDnetwork
cta
∞Xj=0
Tr (cid:0)Hj − AT Hj A(cid:1)
Now, since A is doubly stochastic, it also holds that the enlarged matrix A is doubly stochastic. Moreover,
for any doubly stochastic matrix A and any nonnegative-definite matrix H of compatible dimensions, it
holds that (see part (f ) of Theorem C.3):
− MSDnetwork
atc
(341)
=
1
N
Applying result (342) to (341) we conclude that
Tr(AT HA) ≤ Tr(H)
MSDnetwork
atc
≤ MSDnetwork
cta
(doubly stochastic A)
(342)
(343)
so that the adaptive ATC strategy (153) outperforms the adaptive CTA strategy (154) for doubly stochastic
combination matrices A.
7.2 Comparing Strategies with and without Information Exchange
We now examine the effect of information exchange (C 6= I ) on the performance of the adaptive ATC and
CTA diffusion strategies (153)–(154) under conditions (310)–(312) for uniform data profile.
CTA Strategies
We start with the adaptive CTA strategy (154), and consider two scenarios with and without information
exchange. These scenarios correspond to the following selections in the general description (201)–(203):
C 6= I , A1 = A, A2 = IN
C = I , A1 = A, A2 = IN
(adaptive CTA with information exchange )
(adaptive CTA without information exchange)
Then, expressions (322) and (323) give:
Bcta,C6=I = AT ⊗ (I − µRu ), Ycta,C6=I = µ2 (C T RvC ⊗ Ru )
Bcta,C=I = AT ⊗ (I − µRu ), Ycta,C=I = µ2 (Rv ⊗ Ru )
(344)
(345)
(346)
(347)
where the matrix Rv is defined by (319). Note that Bcta,C6=I = Bcta,C=I , so we denote them simply by B in
the derivation that follows. Then, from expression (306) for the network MSD we get:
∞Xj=0
Tr (cid:0)B j (cid:2)(Rv − C T RvC ) ⊗ Ru (cid:3) B ∗j (cid:1)
It follows that the difference in performance between both CTA implementations depends on how the matrices
Rv and C T Rv C compare to each other:
cta,C=I − MSDnetwork
MSDnetwork
cta,C 6=I =
µ2
N
(348)
62
(1) When Rv − C T Rv C ≥ 0, we obtain
cta,C=I ≥ MSDnetwork
MSDnetwork
cta,C 6=I
(when C T RvC ≤ Rv )
(349)
so that a CTA implementation with information exchange performs better than a CTA implementation
without information exchange. Note that the condition on {Rv , C } corresponds to requiring
C T Rv C ≤ Rv
(350)
which can be interpreted to mean that the cooperation matrix C should be such that it does not
amplify the effect of measurement noise. For example, this situation occurs when the noise profile is
uniform across the network, in which case Rv = σ2
v IM . This is because it would then hold that
Rv − C T Rv C = σ2
v (I − C T C ) ≥ 0
in view of the fact that (I − C T C ) ≥ 0 since C is doubly stochastic (cf. property (e) in Lemma C.3).
(351)
(2) When Rv − C T Rv C ≤ 0, we obtain
cta,C=I ≤ MSDnetwork
MSDnetwork
cta,C 6=I
(when C T RvC ≥ Rv )
(352)
so that a CTA implementation without information exchange performs better than a CTA implementa-
tion with information exchange. In this case, the condition on {Rv , C } indicates that the combination
matrix C ends up amplifying the effect of noise.
ATC Strategies
We can repeat the argument for the adaptive ATC strategy (153), and consider two scenarios with and with-
out information exchange. These scenarios correspond to the following selections in the general description
(201)–(203):
C 6= I , A1 = IN , A2 = A
C = I , A1 = IN , A2 = A
(adaptive ATC with information exchange )
(adaptive ATC without information exchange)
Then, expressions (322) and (323) give:
Batc,C6=I = AT ⊗ (I − µRu ), Yatc,C6=I = µ2 (AT C T Rv CA ⊗ Ru )
Batc,C=I = AT ⊗ (I − µRu ), Yatc,C=I = µ2 (AT Rv A ⊗ Ru )
(353)
(354)
(355)
(356)
Note again that Batc,C6=I = Batc,C=I , so we denote them simply by B . Then,
∞Xj=0
Tr (cid:0)B j (cid:2)AT (Rv − C T RvC )A ⊗ Ru (cid:3) B ∗j (cid:1)
It again follows that the difference in performance between both ATC implementations depends on how the
matrices Rv and C T RvC compare to each other and we obtain:
atc,C=I − MSDnetwork
MSDnetwork
atc,C 6=I =
µ2
N
(357)
atc,C=I ≥ MSDnetwork
MSDnetwork
atc,C 6=I
(when C T RvC ≤ Rv )
and
atc,C=I ≤ MSDnetwork
MSDnetwork
atc,C 6=I
(when C T RvC ≥ Rv )
(358)
(359)
63
Table 6: Comparison of the MSD performance of various cooperative strategies.
Comparison
Conditions
MSDnetwork
atc
≤ MSDnetwork
cta
A doubly stochastic, C right stochastic.
MSDnetwork
cta,C 6=I ≤ MSDnetwork
cta,C=I
C T Rv C ≤ Rv , C doubly stochastic, Ru,k = Ru , µk = µ.
cta,C=I ≤ MSDnetwork
MSDnetwork
cta,C 6=I
C T Rv C ≥ Rv , C doubly stochastic, Ru,k = Ru , µk = µ.
MSDnetwork
atc,C 6=I ≤ MSDnetwork
atc,C=I
C T Rv C ≤ Rv , C doubly stochastic, Ru,k = Ru , µk = µ.
MSDnetwork
atc,C=I ≤ MSDnetwork
atc,C 6=I
C T Rv C ≥ Rv , C doubly stochastic, Ru,k = Ru , µk = µ.
MSDnetwork
atc
≤ MSDnetwork
cta
≤ MSDnetwork
lms
{A, C } doubly stochastic, Ru,k = Ru , µk = µ.
7.3 Comparing Diffusion Strategies with the Non-Cooperative Strategy
We now compare the performance of the adaptive CTA strategy (154) to the non-cooperative LMS strategy
(207) assuming conditions (310)–(312) for uniform data profile. These scenarios correspond to the following
selections in the general description (201)–(203):
C, A1 = A, A2 = I
C = I , A1 = I , A2 = I
(adaptive CTA)
(non-cooperative LMS)
where A is further assumed to be doubly stochastic (along with C ) so that
A1 = 1,
AT 1 = 1
Then, expressions (322) and (323) give:
Bcta = AT ⊗ (I − µRu ),
Blms = I ⊗ (I − µRu ),
Ycta = µ2 (C T Rv C ⊗ Ru )
Ylms = µ2 (Rv ⊗ Ru )
Now recall that
C = C ⊗ IM
so that, using the Kronecker product property (317),
Ycta = µ2 (C T RvC ⊗ Ru )
= µ2 (C T ⊗ IM )(Rv ⊗ Ru )(C ⊗ IM )
= µ2C T (Rv ⊗ Ru )C
= C T Ylms C
Then,
MSDnetwork
lms
− MSDnetwork
cta
=
=
=
1
N
1
N
1
N
∞Xj=0
∞Xj=0
∞Xj=0
1
N
cta(cid:17)
Tr (cid:16)B j
lms(cid:17) −
Tr (cid:16)B j
∞Xj=0
ctaC T Ylms CB ∗j
lmsYlmsB ∗j
lmsYlms(cid:17) −
ctaC T Ylms(cid:17)
Tr (cid:16)CB ∗j
Tr (cid:16)B ∗j
∞Xj=0
lmsB j
ctaB j
ctaC T (cid:17) Ylmsi
Tr h(cid:16)B ∗j
lms − CB ∗j
lmsB j
ctaB j
64
1
N
(360)
(361)
(362)
(363)
(364)
(365)
(366)
(367)
Let us examine the difference:
B ∗j
lms − CB ∗j
lmsB j
ctaB j
ctaC T
= (cid:0)I ⊗ (I − µRu )2j (cid:1) − (cid:0)CAj ⊗ (I − µRu )j (cid:1) (cid:0)AjT C T ⊗ (I − µRu )j (cid:1)
= (cid:0)I ⊗ (I − µRu )2j (cid:1) − (cid:0)CAj AjT C T ⊗ (I − µRu )2j (cid:1)
(317)
= (I − CAj AjT C T ) ⊗ (I − µRu )2j
(368)
Now, due to the even power, it always holds that (I − µRu )2j ≥ 0. Moreover, since Aj and C are doubly
stochastic, it follows that CAj AjT C T is also doubly stochastic. Therefore, the matrix (I − CAj AjT C T ) is
nonnegative-definite as well (cf. property (e) of Lemma C.3). It follows that
B ∗j
lms − CB ∗j
lmsB j
ctaB j
ctaC T ≥ 0
But since Ylms ≥ 0, we conclude from (367) that
(369)
MSDnetwork
lms
≥ MSDnetwork
cta
(370)
This is because for any two Hermitian nonnegative-definite matrices A and B of compatible dimensions, it
holds that Tr(AB ) ≥ 0; indeed if we factor B = X X ∗ with X full rank, then Tr(AB ) = Tr(X ∗AX ) ≥ 0. We
conclude from this analysis that adaptive CTA diffusion performs better than non-cooperative LMS under
uniform data profile conditions and doubly stochastic A. If we refer to the earlier result (343), we conclude
that the following relation holds:
MSDnetwork
atc
≤ MSDnetwork
cta
≤ MSDnetwork
lms
(371)
Table 6 lists the comparison results derived in this section and lists the conditions under which the conclusions
hold.
8 Selecting the Combination Weights
The adaptive diffusion strategy (201)–(203) employs combination weights {a1,ℓk , a2,ℓk , cℓk } or, equivalently,
combination matrices {A1 , A2 , C }, where A1 and A2 are left-stochastic matrices and C is a right-stochastic
matrix. There are several ways by which these matrices can be selected.
In this section, we describe
constructions that result in left-stochastic or doubly-stochastic combination matrices, A. When a right-
stochastic combination matrix is needed, such as C , then it can be obtained by transposition of the left-
stochastic constructions shown below.
8.1 Constant Combination Weights
Table 7 lists a couple of common choices for selecting constant combination weights for a network with N
nodes. Several of these constructions appeared originally in the literature on graph theory. In the table, the
symbol nk denotes the degree of node k , which refers to the size of its neighborhood. Likewise, the symbol
nmax refers to the maximum degree across the network, i.e.,
nmax = max
1≤k≤N
{nk }
(372)
The Laplacian rule, which appears in the second line of the table, relies on the use of the Laplacian matrix
L of the network and a positive scalar γ . The Laplacian matrix is defined by (574) in App. B, namely, it is
a symmetric matrix whose entries are constructed as follows [64–66]:
[L]kℓ =
nk − 1,
if k = ℓ
−1,
if k 6= ℓ and nodes k and ℓ are neighbors
0, otherwise
(373)
65
The Laplacian rule can be reduced to other forms through the selection of the positive parameter γ . One
choice is γ = 1/nmax, while another choice is γ = 1/N and leads to the maximum-degree rule. Obviously,
it always holds that nmax ≤ N so that 1/nmax ≥ 1/N . Therefore, the choice γ = 1/nmax ends up assigning
larger weights to neighbors than the choice γ = 1/N . The averaging rule in the first row of the table is one
of the simplest combination rules whereby nodes simply average data from their neighbors.
Table 7: Selections for combination matrices A = [aℓk ].
Entries of Combination Matrix A
Type of A
1. Averaging rule [68]:
aℓk = (cid:26) 1/nk ,
0,
2. Laplacian rule [49, 69]:
if k 6= ℓ are neighbors or k = ℓ
otherwise
A = IN − γL, γ > 0
left-stochastic
symmetric and
doubly-stochastic
symmetric and
doubly-stochastic
symmetric and
doubly-stochastic
symmetric and
doubly-stochastic
1/nmax ,
1 − (nk − 1)/nmax ,
0,
if k 6= ℓ are neighbors
k = ℓ
otherwise
if k 6= ℓ are neighbors
k = ℓ
otherwise
1/ max{nk , nℓ },
1 − Xℓ∈Nk \{k}
aℓk ,
0,
3. Laplacian rule using γ = 1/nmax :
aℓk =
4. Laplacian rule using γ = 1/N (maximum-degree rule [50]) :
aℓk =
1/N ,
1 − (nk − 1)/N ,
0,
5. Metropolis rule [49, 70, 71]:
aℓk =
6. Relative-degree rule [29]:
aℓk =
nm ,
nℓ / Xm∈Nk
0,
In the constructions in Table 7, the values of the weights {aℓk } are largely dependent on the degree of the
nodes. In this way, the number of connections that each node has influences the combination weights with
its neighbors. While such selections may be appropriate in some applications, they can nevertheless degrade
the performance of adaptation over networks [54]. This is because such weighting schemes ignore the noise
profile across the network. And since some nodes can be noisier than others, it is not sufficient to rely solely
on the amount of connectivity that nodes have to determine the combination weights to their neighbors. It is
important to take into account the amount of noise that is present at the nodes as well. Therefore, designing
combination rules that are aware of the variation in noise profile across the network is an important task.
if k and ℓ are neighbors or k = ℓ
left-stochastic
if k 6= ℓ are neighbors
k = ℓ
otherwise
otherwise
66
It is also important to devise strategies that are able to adapt these combination weights in response to
variations in network topology and data statistical profile. For this reason, following [58, 62], we describe in
the next subsection one adaptive procedure for adjusting the combination weights. This procedure allows
the network to assign more or less relevance to nodes according to the quality of their data.
8.2 Optimizing the Combination Weights
Ideally, we would like to select N × N combination matrices {A1 , A2 , C } in order to minimize the network
MSD given by (302) or (306). In [18], the selection of the combination weights was formulated as the following
optimization problem:
min
{A1 ,A2 ,C }
MSDnetwork given by (302) or (306)
1 = 1,
over left and right-stochastic matrices with nonnegative entries:
AT
1
AT
1 = 1,
2
C 1 = 1,
a2,ℓk = 0 if ℓ /∈ Nk
cℓk = 0 if ℓ /∈ Nk
a1,ℓk = 0 if ℓ /∈ Nk
(374)
We can pursue a numerical solution to (374) in order to search for optimal combination matrices, as was
done in [18]. Here, however, we are interested in an adaptive solution that becomes part of the learning
process so that the network can adapt the weights on the fly in response to network conditions. We illustrate
an approximate approach from [58, 62] that leads to one adaptive solution that performs reasonably well in
practice.
We illustrate the construction by considering the ATC strategy (158) without information exchange where
A1 = IN , A2 = A, and C = I . In this case, recursions (204)–(206) take the form:
ψk,i = wk,i−1 + µk u∗
k,i [dk (i) − uk,iwk,i−1 ]
wk,i = Xℓ∈Nk
aℓkψ ℓ,i
and, from (306), the corresponding network MSD performance is:
Tr (cid:16)B j
atc(cid:17)
∞Xj=0
atcYatcB ∗j
MSDnetwork
atc
1
N
=
where
Batc = AT (I − MRu )
Yatc = AT MSMA
Ru = diag{Ru,1 , Ru,2 , . . . , Ru,N }
v,2Ru,2 , . . . , σ2
v,1Ru,1 , σ2
S = diag{σ2
v,N Ru,N }
M = diag{µ1IM , µ2 IM , . . . , µN IM }
A = A ⊗ IM
(375)
(376)
(377)
(378)
(379)
(380)
(381)
(382)
(383)
Minimizing the MSD expression (377) over left-stochastic matrices A is generally non-trivial. We pursue an
approximate solution.
To begin with, for compactness of notation, let r denote the spectral radius of the N × N block matrix
I − MRu :
r ∆= ρ(I − MRu )
67
(384)
We already know, in view of the mean and mean-square stability of the network, that r < 1. Now, consider
the series that appears in (377) and whose trace we wish to minimize over A. Note that its block maximum
norm can be bounded as follows:
atc(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)b,∞
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
∞Xj=0 (cid:13)(cid:13)(cid:13)B j
atc(cid:13)(cid:13)(cid:13)b,∞
· kYatc kb,∞ · (cid:13)(cid:13)(cid:13)B ∗j
atc(cid:13)(cid:13)(cid:13)b,∞
∞Xj=0
atcYatcB ∗j
B j
≤
≤ N ·
· kYatc kb,∞
∞Xj=0 (cid:13)(cid:13)(cid:13)B j
atc(cid:13)(cid:13)(cid:13)
(a)
2
b,∞
≤ N ·
b,∞ · kYatc kb,∞
∞Xj=0
kBatc k2j
≤ N ·
r2j · kYatc kb,∞
∞Xj=0
(b)
N
1 − r2 · kYatc kb,∞
=
where for step (b) we use result (602) to conclude that
(385)
kBatckb,∞ = kAT (I − MRu )kb,∞
≤ kAT kb,∞ · kI − MRukb,∞
= kI − MRukb,∞
(602)
= r
atc and its complex conjugate, hB j
atci∗
To justify step (a), we use result (584) to relate the norms of B j
kB ∗j
atckb,∞ ≤ N · kB j
atckb,∞
(386)
, as
(387)
Expression (385) then shows that the norm of the series appearing in (377) is bounded by a scaled multiple
of the norm of Yatc , and the scaling constant is independent of A. Using property (586) we conclude that
there exists a positive constant c, also independent of A, such that
atc ≤ c · Tr(Yatc )
Tr
∞Xj=0
atcYatcB ∗j
B j
Therefore, instead of attempting to minimize the trace of the series, the above result motivates us to minimize
an upper bound to the trace. Thus, we consider the alternative problem of minimizing the first term of the
series (377), namely,
(388)
Tr(Yatc )
min
A
sub ject to AT 1 = 1, aℓk ≥ 0, aℓk = 0 if ℓ /∈ Nk
Using (379), the trace of Yatc can be expressed in terms of the combination coefficients as follows:
NXℓ=1
NXk=1
ℓk σ2
ℓ a2
µ2
v,ℓ Tr(Ru,ℓ )
Tr(Yatc ) =
(389)
(390)
68
so that problem (389) can be decoupled into N separate optimization problems of the form:
NXℓ=1
min
{aℓk }N
ℓ=1
sub ject to
µ2
ℓ a2
ℓk σ2
v,ℓ Tr(Ru,ℓ ),
k = 1, . . . , N
NXℓ=1
With each node ℓ, we associate the following nonnegative noise-data-dependent measure:
aℓk = 0 if ℓ /∈ Nk
aℓk ≥ 0,
aℓk = 1,
γ 2
ℓ
∆= µ2
ℓ · σ2
v,ℓ · Tr(Ru,ℓ )
(391)
(392)
(393)
,
if ℓ ∈ Nk
γ−2
m
k
otherwise
(relative-variance rule)
γ−2
ℓPm∈N
0,
This measure amounts to scaling the noise variance at node ℓ by µ2
ℓ and by the power of the regression
data (measured through the trace of its covariance matrix). We shall refer to γ 2
ℓ as the noise-data variance
product (or variance product, for simplicity) at node ℓ. Then, the solution of (391) is given by:
aℓk =
We refer to this combination rule as the relative-variance combination rule [58]; it leads to a left-stochastic
matrix A. In this construction, node k combines the intermediate estimates {ψ ℓ,i} from its neighbors in (376)
in proportion to the inverses of their variance products, {γ−2
m }. The result is physically meaningful. Nodes
with smaller variance products will generally be given larger weights. In comparison, the following relative-
degree-variance rule was proposed in [18] (a typo appears in Table III in [18], where the noise variances
appear written in the table instead of their inverses):
aℓk =
This second form also leads to a left-stochastic combination matrix A. However, rule (394) does not take into
account the covariance matrices of the regression data across the network. Observe that in the special case
when step-sizes, the regression covariance matrices, and the noise variances are uniform across the network,
i.e., µk = µ, Ru,k = Ru , and σ2
v,k = σ2
v for all k , expression (393) reduces to the simple averaging rule (first
line of Table 7). In contrast, expression (394) reduces the relative degree rule (last line of Table 7).
nℓ σ−2
v,ℓ
Pm∈N
nm σ−2
v,m
0,
(relative degree-variance rule)
if ℓ ∈ Nk
(394)
,
k
otherwise
8.3 Adaptive Combination Weights
To evaluate the combination weights (393), the nodes need to know the variance products, {γ 2
m}, of their
neighbors. According to (392), the factors {γ 2
m} are defined in terms of the noise variances, {σ2
v,m}, and the
regression covariance matrices, {Tr(Ru,m )}, and these quantities are not known beforehand. The nodes only
have access to realizations of {dm (i), um,i}. We now describe one procedure that allows every node k to
learn the variance products of its neighbors in an adaptive manner. Note that if a particular node ℓ happens
to belong to two neighborhoods, say, the neighborhood of node k1 and the neighborhood of node k2 , then
69
lim
i→∞
each of k1 and k2 need to evaluate the variance product, γ 2
ℓ , of node ℓ. The procedure described below allows
each node in the network to estimate the variance products of its neighbors in a recursive manner.
To motivate the algorithm, we refer to the ATC recursion (375)–(376) and use the data model (208) to
write for node ℓ:
ψ ℓ,i = wℓ,i−1 + µℓu∗
ℓ,i [uℓ,i ewℓ,i−1 + v ℓ (i)]
so that, in view of our earlier assumptions on the regression data and noise in Sec. 6.1, we obtain in the limit
as i → ∞:
ℓ,i kuℓ,i k2 uℓ,i )(cid:19) + µ2
ℓ · (cid:18) lim
E (cid:13)(cid:13)ψ ℓ,i − wℓ,i−1(cid:13)(cid:13)2
ℓ · σ2
= µ2
E k ewi−1k2
v,ℓ · Tr(Ru,ℓ )
E (u
∗
i→∞
We can evaluate the limit on the right-hand side by using the steady-state result (284). Indeed, we select
the vector σ in (284) to satisfy
ℓ,ikuℓ,ik2uℓ,i(cid:1)(cid:3)
(I − F )σ = vec (cid:2)E (cid:0)u∗
Then, from (284),
ℓ,i kuℓ,i k2uℓ,i ) = (cid:2)vec (cid:0)Y T (cid:1)(cid:3)T
ℓ,ikuℓ,i k2uℓ,i(cid:1)(cid:3)
· (I − F )−1 · vec (cid:2)E (cid:0)u∗
E k ewi−1 k2
E (u
∗
Now recall from expression (379) for Y that for the ATC algorithm under consideration we have
Y = AT MSMA
lim
i→∞
(399)
(397)
(398)
(395)
(396)
so that the entries of Y depend on combinations of the squared step-sizes, {µ2
m , m = 1, 2, . . . , N }. This fact
implies that the first term on the right-hand side of (396) depends on products of the form {µ2
ℓ µ2
m}; these
fourth-order factors can be ignored in comparison to the second-order factor µ2
ℓ for small step-sizes so that
E (cid:13)(cid:13)ψ ℓ,i − wℓ,i−1(cid:13)(cid:13)2
in terms of the desired variance product, γ 2
ℓ . Using the following instantaneous approximation at node k
(where wℓ,i−1 is replaced by wk,i−1 ):
ℓ · σ2
≈ µ2
v,ℓ · Tr(Ru,ℓ )
= γ 2
ℓ
lim
i→∞
(400)
E kψ ℓ,i − w ℓ,i−1k2 ≈ kψℓ,i − wk,i−1 k2
(401)
we can motivate an algorithm that enables node k to estimate the variance product, γ 2
ℓ , of its neighbor ℓ.
Thus, let γ 2
ℓk (i) denote an estimate for γ 2
ℓ that is computed by node k at time i. Then, one way to evaluate
γ 2
ℓk (i) is through the recursion:
ℓk (i) = (1 − νk ) · γ 2
γ 2
ℓk (i − 1) + νk · kψℓ,i − wk,i−1 k2
(402)
where 0 < νk ≪ 1 is a positive coefficient smaller than one. Note that under expectation, expression (402)
becomes
ℓk (i) = (1 − νk ) · E γ 2
E γ 2
ℓk (i − 1) + νk · E kψ ℓ,i − wk,i−1 k2
so that in steady-state, as i → ∞,
lim
i→∞
E γ 2
ℓk (i) ≈ (1 − νk ) · lim
i→∞
ℓk (i − 1) + νk · γ 2
E γ 2
ℓ
Hence, we obtain
lim
i→∞
ℓk (i) ≈ γ 2
E γ 2
ℓ
70
(403)
(404)
(405)
ℓk (i) converges on average close to the desired variance product γ 2
That is, the estimator γ 2
ℓ . In this way, we
can replace the optimal weights (393) by the adaptive construction:
aℓk (i) =
Equations (402) and (406) provide one adaptive construction for the combination weights {aℓk }.
9 Diffusion with Noisy Information Exchanges
γ−2
ℓk (i)
Pm∈N
γ−2
mk (i)
0,
otherwise
(406)
,
if ℓ ∈ Nk
k
The adaptive diffusion strategy (201)–(203) relies on the fusion of local information collected from neighbor-
hoods through the use of combination matrices {A1 , A2 , C }. In the previous section, we described several
constructions for selecting such combination matrices. We also motivated and developed an adaptive scheme
for the ATC mode of operation (375)–(376) that computes combination weights in a manner that is aware
of the variation of the variance-product profile across the network. Nevertheless, in addition to the mea-
surement noises {vk (i)} at the individual nodes, we also need to consider the effect of perturbations that
are introduced during the exchange of information among neighboring nodes. Noise over the communication
links can be due to various factors including thermal noise and imperfect channel information. Studying
the degradation in mean-square performance that results from these noisy exchanges can be pursued by
straightforward extension of the mean-square analysis of Sec. 6, as we proceed to illustrate. Subsequently,
we shall use the results to show how the combination weights can also be adapted in the presence of noisy
exchange links.
9.1 Noise Sources over Exchange Links
To model noisy links, we introduce an additive noise component into each of the steps of the diffusion strategy
(201)–(203) during the operations of information exchange among the nodes. The notation becomes a bit
cumbersome because we need to account for both the source and destination of the information that is being
exchanged. For example, the same signal dℓ (i) that is generated by node ℓ will be broadcast to all the
neighbors of node ℓ. When this is done, a different noise will interfere with the exchange of dℓ (i) over each
of the edges that link node ℓ to its neighbors. Thus, we will need to use a notation of the form dℓk (i), with
two subscripts ℓ and k , to indicate that this is the noisy version of dℓ (i) that is received by node k from node
ℓ. The subscript ℓk indicates that ℓ is the source and k is the sink, i.e., information is moving from ℓ to k .
For the reverse situation where information flows from node k to ℓ, we would use instead the subscript kℓ.
With this notation in mind, we model the noisy data received by node k from its neighbor ℓ as follows:
(w)
wℓk,i−1 = w ℓ,i−1 + v
ℓk,i−1
ψ ℓk,i = ψ ℓ,i + v (ψ)
ℓk,i
uℓk,i = uℓ,i + v (u)
ℓk,i
dℓk (i) = dℓ (i) + v (d)
ℓk (i)
(407)
(408)
(409)
(410)
where v (w)
ℓk,i (1 × M ) are vector noise signals, and v (d)
ℓk,i (M × 1), and v (u)
ℓk,i−1 (M × 1), v (ψ)
ℓk (i) is a scalar
noise signal. These are the noise signals that perturb exchanges over the edge linking source ℓ to sink k
(i.e., for data sent from node ℓ to node k). The superscripts {(w), (ψ), (u), (d)} in each case refer to the
variable that these noises perturb. Figure 14 illustrates the various noise sources that perturb the exchange
of information from node ℓ to node k . The figure also shows the measurement noises {v ℓ (i), vk (i)} that exist
locally at the nodes in view of the data model (208).
71
v ℓ (i)
source
ℓ
vk (i)
k
sink
v
(u)
ℓk ,i
uℓ,i
uℓk,i
v
(d)
ℓk (i)
dℓ (i)
dℓk (i)
ψℓ,i
ψ ℓk ,i
v
(ψ)
ℓk,i
wℓ,i−1
wℓk,i−1
v
(w)
ℓk ,i−1
Figure 14: Additive noise sources perturb the exchange of information from node ℓ to node k. The subscript ℓk in
this illustration indicates that ℓ is the source node and k is the sink node so that information is flowing from ℓ to k.
We assume that the following noise signals, which influence the data received by node k ,
ℓk,io
nv k (i), v (d)
ℓk (i), v (w)
ℓk,i−1 , v (ψ)
ℓk,i , v (u)
are temporally white and spatially independent random processes with zero mean and variances or covariances
given by
nσ2
v,ℓko
v,ℓk , R(u)
v,ℓk , R(ψ)
v,ℓk , R(w)
v,k , σ2
nσ2
v,ℓk o
v,ℓk , R(u)
v,ℓk , R(ψ)
v,ℓk , R(w)
are all zero if ℓ /∈ Nk or when ℓ = k . We further assume that the noise processes (411) are independent of
each other and of the regression data um,j for all k , ℓ, m and i, j .
Obviously, the quantities
(411)
(412)
(413)
9.2 Error Recursion
Using the perturbed data (407)–(410), the adaptive diffusion strategy (201)–(203) becomes
φk,i−1 = Xℓ∈Nk
a1,ℓk w ℓk,i−1
ψk,i = φk,i−1 + µk Xℓ∈Nk
wk,i = Xℓ∈Nk
a2,ℓk ψ ℓk,i
cℓk u∗
ℓk,i [dℓk (i) − uℓk,iφk,i−1 ]
(414)
(415)
(416)
72
(417)
(418)
Observe that the perturbed quantities {wℓk,i−1 , uℓk,i , dℓk (i), ψ ℓk,i}, with subscripts ℓk , appear in (414)–
(416) in place of the original quantities {wℓ,i−1 , uℓ,i , dℓ (i), ψ ℓ,i} that appear in (201)–(203). This is because
these quantities are now sub ject to exchange noises. As before, we are still interested in examining the
evolution of the weight-error vectors:
∆= wo − wk,i ,
ewk,i
k = 1, 2, . . . , N
For this purpose, we again introduce the following N × 1 block vector, whose entries are of size M × 1 each:
∆=
ew1,i
ew2,i
ew i
...
ewN ,i
and proceed to determine a recursion for its evolution over time. The arguments are largely similar to what
we already did before in Sec. 6.3 and, therefore, we shall emphasize the differences that arise. The main
deviation is that we now need to account for the presence of the new noise signals; they will contribute
additional terms to the recursion for ewi — see (442) further ahead. We may note that some studies on
the effect of imperfect data exchanges on the performance of adaptive diffusion algorithms are considered
in [59–61]. However, these earlier investigations were limited to particular cases in which only noise in the
exchange of wℓ,i−1 was considered (as in (407)), in addition to setting C = I (in which case there is no
exchange of {dℓ (i), uℓ,i}), and by focusing on the CTA case for which A2 = I . Here, we consider instead the
general case that accounts for the additional sources of imperfections shown in (408)–(410), in addition to
the general diffusion strategy (201)–(203) with combination matrices {A1 , A2 , C }.
To begin with, we introduce the aggregate M × 1 zero-mean noise signals:
∆= Xℓ∈Nk
∆= Xℓ∈Nk
These noises represent the aggregate effect on node k of all exchange noises from the neighbors of node k
while exchanging the estimates {wℓ,i−1 , ψ ℓ,i} during the two combination steps (201) and (203). The M ×M
covariance matrices of these noises are given by
∆= Xℓ∈Nk
∆= Xℓ∈Nk
These expressions aggregate the exchange noise covariances in the neighborhood of node k ; the covariances
are scaled by the squared coefficients {a2
1,ℓk , a2
2,ℓk }. We collect these noise signals, and their covariances,
from across the network into N × 1 block vectors and N × N block diagonal matrices as follows:
N ,i−1o
∆= col nv (w)
2,i−1 , . . . , v (w)
1,i−1 , v (w)
N ,io
∆= col nv (ψ)
2,i , . . . , v (ψ)
1,i , v (ψ)
v,N o
∆= diag nR(w)
v,1 , R(w)
v,2 , . . . , R(w)
∆= diag nR(ψ)
v,N o
v,1 , R(ψ)
v,2 , . . . , R(ψ)
We further introduce the following scalar zero-mean noise signal:
a1,ℓk v (w)
ℓk,i−1 ,
1,ℓk R(w)
a2
v,ℓk ,
v (w)
i−1
v (ψ)
i
R(w)
v
R(ψ)
v
(421)
(422)
(423)
(424)
R(ψ)
v,k
2,ℓk R(ψ)
a2
v,ℓk
v (ψ)
k,i
a2,ℓk v (ψ)
ℓk,i
(419)
(420)
v (w)
k,i−1
R(w)
v,k
v ℓk (i) ∆= v ℓ(i) + v (d)
ℓk (i) − v (u)
ℓk,i wo
(425)
73
whose variance is
v,ℓk + wo∗R(u)
σ2
ℓk = σ2
v,ℓ + σ2
v,ℓkwo
(426)
In the absence of exchange noises for the data {dℓ(i), uℓ,i}, the signal v ℓk (i) would coincide with the measure-
ment noise v ℓ (i). Expression (425) is simply a reflection of the aggregate effect of the noises in exchanging
{dℓ (i), uℓ,i} on node k . Indeed, starting from the data model (208) and using (409)–(410), we can easily
verify that the noisy data {dℓk (i), uℓk,i } are related via:
dℓk (i) = uℓk,iwo + v ℓk (i)
(427)
We also define (compare with (234)–(235) and note that we are now using the perturbed regression vectors
{uℓk,i }):
∆= Xℓ∈Nk
cℓku∗
ℓk,iuℓk,i
N ,i(cid:9)
∆= diag (cid:8)R′
2,i , . . . , R′
1,i , R′
E R′
k,i = R′
k
It holds that
R′
k,i
(428)
(430)
(429)
R′
i
where
R′
k
cℓk hRu,ℓ + R(u)
v,ℓk i
∆= Xℓ∈Nk
When there is no noise during the exchange of the regression data, i.e., when R(u)
v,ℓk = 0, the expressions for
{R′
k,i , R′
i , R′
k } reduce to expressions (234)–(235) and (182) for {Rk,i , Ri , Rk }.
Likewise, we introduce (compare with (239)):
∆= Xℓ∈Nk
∆= col {z 1,i , z 2,i , . . . , zN ,i}
cℓk u∗
ℓk,iv ℓk (i)
(432)
(431)
(433)
zk,i
z i
Compared with the earlier definition for si in (239) when there is no noise over the exchange links, we see
that we now need to account for the various noisy versions of the same regression vector uℓ,i and the same
signal dℓ(i). For instance, the vectors uℓk,i and uℓm,i would denote two noisy versions received by nodes k
and m for the same regression vector uℓ,i transmitted from node ℓ. Likewise, the scalars dℓk (i) and dℓm (i)
would denote two noisy versions received by nodes k and m for the same scalar dℓ (i) transmitted from node
ℓ. As a result, the quantity z i is not zero mean any longer (in contrast to si , which had zero mean). Indeed,
note that
E z k,i = Xℓ∈Nk
cℓk E u∗
ℓk,iv ℓk (i)
cℓk E (cid:16)huℓ,i + v (u)
ℓk,i i∗
= Xℓ∈Nk
= − Xℓ∈Nk
v,ℓk! wo
cℓk R(u)
ℓk,i wo i(cid:17)
· hv ℓ (i) + v (d)
ℓk (i) − v (u)
(434)
74
It follows that
wo
(435)
E z i = −
cℓ1 R(u)
v,ℓ1
cℓ2 R(u)
v,ℓ2
Xℓ∈N1
Xℓ∈N2
Xℓ∈NN
Although we can continue our analysis by studying this general case in which the vectors z i do not have
zero-mean (see [62, 63]), we shall nevertheless limit our discussion in the sequel to the case in which there is
no noise during the exchange of the regression data, i.e., we henceforth assume that:
ℓk,i = 0, R(u)
v (u)
(assumption from this point onwards)
v,ℓk = 0, uℓk,i = uℓ,i
cℓN R(u)
v,ℓN
(436)
...
We maintain all other noise sources, which occur during the exchange of the weight estimates {wℓ,i−1 , ψ ℓ,i}
and the data {dℓ (i)}. Under condition (436), we obtain
E z i = 0
v,ℓ + σ2
ℓk = σ2
σ2
v,ℓk
k = Xℓ∈Nk
R′
Then, the covariance matrix of each term z k,i is given by
∆= Xℓ∈Nk
and the covariance matrix of z i is N × N block diagonal with blocks of size M × M :
ℓk σ2
c2
ℓk Ru,ℓ
(182)
= Rk
cℓk Ru,ℓ
Rz ,k
Z ∆
= E z iz ∗
i = diag{Rz ,1 , Rz ,2 , . . . , Rz ,N }
(437)
(438)
(439)
(440)
(441)
Now repeating the argument that led to (246) we arrive at the following recursion for the weight-error vector:
i (cid:1) v (w)
2 (cid:0)IN M − MR′
i(cid:1) AT
2 (cid:0)IN M − MR′
i−1 − v (ψ)
2 Mz i − AT
1 ewi−1 − AT
ew i = AT
i
For comparison purposes, we repeat recursion (246) here (recall that this recursion corresponds to the case
when the exchanges over the links are not sub ject to noise):
2 (IN M − MRi ) AT
ewi = AT
2 MC T si
1 ewi−1 − AT
Comparing (442) and (443) we find that:
(1) The covariance matrix Ri in (443) is replaced by R′
i . Recall from (429) that R′
i contains the influence
of the noises that arise during the exchange of the regression data, i.e., the {R(u)
v,ℓk }. But since we are
now assuming that R(u)
v,ℓk = 0, then R′
i = Ri .
(perfect links)
(noisy links)
(442)
(443)
75
(2) The term C T si in (443) is replaced by z i . Recall from (432) that z i contains the influence of the noises
v,ℓk , R(u)
that arise during the exchange of the measurement data and the regression data, i.e., the {σ2
v,ℓk }.
i−1 and v (ψ)
(3) Two new driving terms appear involving v (w)
. These terms reflect the influence of the noises
i
during the exchange of the weight estimates {wℓ,i−1 , ψℓ,i }.
(4) Observe further that:
(4a) The term involving v
before adaptation.
(w)
i−1 accounts for noise introduced at the information-exchange step (414)
(4b) The term involving z i accounts for noise introduced during the adaptation step (415).
(4c) The term involving v (ψ)
i
after adaptation.
accounts for noise introduced at the information-exchange step (416)
Therefore, since we are not considering noise during the exchange of the regression data, the weight-error
recursion (442) simplifies to:
i−1 − v (ψ)
2 (IN M − MRi ) v (w)
ewi = AT
2 (IN M − MRi ) AT
1 ewi−1 − AT
2 Mz i − AT
i
where we used the fact that R′
i = Ri under these conditions.
(noisy links)
(444)
9.3 Convergence in the Mean
Taking expectations of both sides of (444) we find that the mean error vector evolves according to the
following recursion:
E ew i = AT
2 (IN M − MR) AT
1 · E ewi−1 ,
with R defined by (181). This is the same recursion encountered earlier in (248) during perfect data
exchanges. Note that had we considered noises during the exchange of the regression data, then the vector
z i in (444) would not be zero mean and the matrix R′
i will have to be used instead of Ri . In that case, the
recursion for E ewi will be different from (445); i.e., the presence of noise during the exchange of regression
data alters the dynamics of the mean error vector in an important way — see [62, 63] for details on how to
extend the arguments to this general case with a driving non-zero bias term. We can now extend Theorem 6.1
to the current scenario.
(445)
i ≥ 0
Theorem 9.1. (Convergence in the Mean) Consider the problem of optimizing the global cost (92) with
the individual cost functions given by (93). Pick a right stochastic matrix C and left stochastic matrices A1
and A2 satisfying (166). Assume each node in the network runs the perturbed adaptive diffusion algorithm
(414)–(416). Assume further that the exchange of the variables {wℓ,i−1 , ψ ℓ,i , dℓ (i)} is subject to additive
noises as in (407), (408), and (410). We assume that the regressors are exchanged unperturbed. Then, al l
estimators {wk,i } across the network wil l stil l converge in the mean to the optimal solution wo if the step-size
parameters {µk } satisfy
where the neighborhood covariance matrix Rk is defined by (182). That is, E wk,i → wo as i → ∞.
µk <
2
λmax (Rk )
(446)
(cid:3)
76
9.4 Mean-Square Convergence
Recall from (264) that we introduced the matrix:
We further introduce the N × N block matrix with blocks of size M × M each:
B ∆= AT
2 (IN M − MR) AT
1
H ∆= AT
2 (IN M − MR)
(447)
(448)
Then, starting from (444) and repeating the argument that led to (279) we can establish the validity of the
following variance relation:
(cid:17)(cid:21)T
F σ + (cid:20)vec (cid:0)AT
2 MZ T MA2 (cid:1) + vec (cid:18)(cid:16)HR(w)T
v H∗(cid:17)T (cid:19) + vec (cid:16)R(ψ)T
E k ewi k2
σ = E k ewi−1 k2
v
for an arbitrary nonnegative-definite weighting matrix Σ with σ = vec(Σ), and where F is the same matrix
defined earlier either by (276) or (277). We can therefore extend the statement of Theorem 6.7 to the present
scenario.
(449)
σ
Theorem 9.2. (Mean-Square Stability) Consider the same setting of Theorem 9.1. Assume sufficiently
smal l step-sizes to justify ignoring terms that depend on higher powers of the step-sizes. The perturbed
adaptive diffusion algorithm (414)–(416) is mean-square stable if, and only if, the matrix F defined by
(276), or its approximation (277), is stable (i.e., al l its eigenvalues are strictly inside the unit disc). This
condition is satisfied by sufficiently smal l step-sizes {µk } that are also smal ler than:
µk <
2
λmax (Rk )
(450)
where the neighborhood covariance matrix Rk is defined by (182). Moreover, the convergence rate of the
algorithm is determined by [ρ(B )]2 .
(cid:3)
We conclude from the previous two theorems that the conditions for the mean and mean-square convergence
of the adaptive diffusion strategy are not affected by the presence of noises over the exchange links (under
the assumption that the regression data are exchanged without perturbation; otherwise, the convergence
conditions would be affected). The mean-square performance, on the other hand, is affected as follows.
Introduce the N × N block matrix:
Yimperfect
∆= AT
v H∗ + R(ψ)
2 MZMA2 + HR(w)
v
(imperfect exchanges)
(451)
which should be compared with the corresponding quantity defined by (280) for the perfect exchanges case,
namely,
2 MC T S CMA2
Yperfect = AT
When perfect exchanges occur, the matrix Z reduces to C T S C . We can relate Yimperfect and Yperfect as
follows. Let
v,ℓN Ru,ℓ)
R(du) ∆= diag ( Xℓ∈N1
ℓN σ2
c2
v,ℓ2Ru,ℓ , . . . Xℓ∈NN
ℓ2σ2
c2
(perfect exchanges)
v,ℓ1Ru,ℓ , Xℓ∈N2
ℓ1σ2
c2
77
(452)
(453)
Then, using (438) and (441), it is straightforward to verify that
Z = C T S C + R(du)
and it follows that:
v H∗ + R(ψ)
Yimperfect = Yperfect + AT
2 MR(du)MA2 + HR(w)
v
∆= Yperfect + ∆Y
(454)
(455)
Expression (455) reflects the influence of the noises {R(w)
, R(ψ)
, σ2
v,ℓk }. Substituting the definition (451)
v
v
into (449), and taking the limit as i → ∞, we obtain from the latter expression that:
imperfect(cid:1)(cid:3)T
(I−F )σ = (cid:2)vec (cid:0)Y T
E k ewik2
which has the same form as (284); therefore, we can proceed analogously to obtain:
imperfect(cid:1)(cid:3)T
· (cid:2)vec (cid:0)Y T
· (I − F )−1 · vec (IN M )
MSDnetwork
imperfect =
lim
i→∞
(456)
σ
1
N
(457)
and
1
N
EMSEnetwork
imperfect =
· (I − F )−1 · vec (Ru )
imperfect(cid:1)(cid:3)T
· (cid:2)vec (cid:0)Y T
Using (455), we see that the network MSD and EMSE deteriorate as follows:
· (cid:2)vec (cid:0)∆Y T (cid:1)(cid:3)T
1
· (I − F )−1 · vec (IN M )
N
· (cid:2)vec (cid:0)∆Y T (cid:1)(cid:3)T
1
N
imperfect = EMSEnetwork
EMSEnetwork
perfect +
imperfect = MSDnetwork
MSDnetwork
perfect +
· (I − F )−1 · vec (Ru )
9.5 Adaptive Combination Weights
(458)
(459)
(460)
We can repeat the discussion from Secs. 8.2 and 8.3 to devise one adaptive scheme to adjust the combination
coefficients in the noisy exchange case. We illustrate the construction by considering the ATC strategy
corresponding to A1 = IN , A2 = A, C = IN , so that only weight estimates are exchanged and the update
recursions are of the form:
where from (408):
ψk,i = wk,i−1 + µk u∗
k,i [dk (i) − uk,iwk,i−1 ]
wk,i = Xℓ∈Nk
aℓk ψ ℓk,i
ψ ℓk,i = ψ ℓ,i + v (ψ)
ℓk,i
In this case, the network MSD performance (457) becomes
Tr (cid:16)B j
atc,C=I(cid:17)
∞Xj=0
atc,C=IYatc,imperfectB ∗j
MSDnetwork
atc,C=I,imperfect =
1
N
(461)
(462)
(463)
(464)
78
where, since now Z = S and R(w)
v = 0, we have
Batc,C=I = AT (I − MRu )
Yatc,imperfect = AT MSMA + R(ψ)
v
v,N o
= diag nR(ψ)
v,2 , . . . , R(ψ)
v,1 , R(ψ)
R(ψ)
v
v,k = Xℓ∈Nk
R(ψ)
ℓk R(ψ)
a2
v,ℓk
Ru = diag{Ru,1 , Ru,2 , . . . , Ru,N }
S = diag{σ2
v,1Ru,1 , σ2
v,2Ru,2 , . . . , σ2
v,N Ru,N }
M = diag{µ1 IM , µ2 IM , . . . , µN IM }
A = A ⊗ IM
To proceed, as was the case with (389), we consider the following simplified optimization problem:
Tr(Yatc,imperfect)
min
A
sub ject to AT 1 = 1, aℓk ≥ 0, aℓk = 0 if ℓ /∈ Nk
(465)
(466)
(467)
(468)
(469)
(470)
(471)
(472)
(473)
Using (466), the trace of Yatc,imperfect can be expressed in terms of the combination coefficients as follows:
ℓk hµ2
v,ℓ Tr(Ru,ℓ ) + Tr (cid:16)R(ψ)
v,ℓk(cid:17)i
NXk=1
NXℓ=1
a2
ℓ σ2
so that problem (473) can be decoupled into N separate optimization problems of the form:
Tr(Yatc,imperfect) =
(474)
k = 1, . . . , N
NXℓ=1
sub ject to
aℓk ≥ 0,
aℓk = 1,
min
{aℓk }N
ℓ=1
aℓk = 0 if ℓ /∈ Nk
v,ℓ Tr(Ru,ℓ ) + Tr (cid:16)R(ψ)
ℓk hµ2
v,ℓk (cid:17)i ,
ℓ σ2
a2
NXℓ=1
With each node ℓ, we associate the following nonnegative variance products:
v,ℓ · Tr(Ru,ℓ ) + Tr (cid:16)R(ψ)
v,ℓk (cid:17) , k ∈ Nℓ
∆= µ2
ℓ · σ2
This measure now incorporates information about the exchange noise covariances {R(ψ)
v,ℓk }. Then, the solution
of (475) is given by:
aℓk =
We continue to refer to this combination rule as the relative-variance combination rule [58]; it leads to a
left-stochastic matrix A. To evaluate the combination weights (477), the nodes need to know the variance
products, {γ 2
mk }, of their neighbors. As before, we can motivate one adaptive construction as follows.
γ−2
ℓkPm∈N
k
0,
(relative-variance rule)
,
if ℓ ∈ Nk
γ−2
mk
(475)
(476)
(477)
γ 2
ℓk
otherwise
79
(478)
lim
i→∞
We refer to the ATC recursion (461)–(463) and use the data model (208) to write for node ℓ:
ℓ,i [uℓ,i ewℓ,i−1 + v ℓ (i)] + v (ψ)
ψ ℓk,i = w ℓ,i−1 + µℓu∗
ℓk,i
so that, in view of our earlier assumptions on the regression data and noise in Secs. 6.1 and 9.1, we obtain
in the limit as i → ∞:
ℓ,i kuℓ,i k2 uℓ,i )(cid:19) + µ2
ℓ · (cid:18) lim
v,ℓ · Tr(Ru,ℓ ) + Tr (cid:16)R(ψ)
v,ℓk (cid:17) (479)
E (cid:13)(cid:13)ψ ℓk,i − wℓ,i−1(cid:13)(cid:13)2
ℓ · σ2
E k ewi−1 k2
= µ2
E (u
∗
i→∞
In a manner similar to what was done before for (396), we can evaluate the limit on the right-hand side by
using the corresponding steady-state result (456). We select the vector σ in (456) to satisfy:
ℓ,ikuℓ,ik2uℓ,i(cid:1)(cid:3)
(I − F )σ = vec (cid:2)E (cid:0)u∗
Then, from (456),
atc,imperfect(cid:1)(cid:3)T
ℓ,ikuℓ,ik2uℓ,i(cid:1)(cid:3)
· (I − F )−1 · vec (cid:2)E (cid:0)u∗
ℓ,i kuℓ,i k2 uℓ,i ) = (cid:2)vec (cid:0)Y T
E k ewi−1k2
(481)
E (u
∗
Now recall from expression (466) that the entries of Yatc,imperfect depend on combinations of the squared
m , m = 1, 2, . . . , N }, and on terms involving nTr (cid:16)R(ψ)
v,m(cid:17)o. This fact implies that the first
step-sizes, {µ2
term on the right-hand side of (479) depends on products of the form {µ2
ℓ µ2
m}; these fourth-order factors can
be ignored in comparison to the second-order factor µ2
ℓ for small step-sizes. Moreover, the same first term
on the right-hand side of (479) depends on products of the form nµ2
ℓ Tr (cid:16)R(ψ)
v,m(cid:17)o, which can be ignored in
comparison to the last term, Tr (cid:16)R(ψ)
v,ℓk (cid:17) , in (479), which does not appear multiplied by a squared step-size.
Therefore, we can approximate:
v,ℓ · Tr(Ru,ℓ ) + Tr (cid:16)R(ψ)
v,ℓk (cid:17)
E (cid:13)(cid:13)ψ ℓk,i − wℓ,i−1(cid:13)(cid:13)2
≈ µ2
ℓ · σ2
lim
i→∞
= γ 2
ℓk
in terms of the desired variance product, γ 2
ℓk . Using the following instantaneous approximation at node k
(where wℓ,i−1 is replaced by wk,i−1 ):
lim
i→∞
(480)
(482)
E kψ ℓk,i − w ℓ,i−1k2 ≈ kψℓk,i − wk,i−1 k2
(483)
ℓk . Thus, let bγ 2
we can motivate an algorithm that enables node k to estimate the variance products γ 2
ℓk (i)
ℓk that is computed by node k at time i. Then, one way to evaluate bγ 2
denote an estimate for γ 2
ℓk (i) is through
the recursion:
ℓk (i − 1) + νk · kψℓk,i − wk,i−1 k2
ℓk (i) = (1 − νk ) · bγ 2
bγ 2
(484)
where νk is a positive coefficient smaller than one. Indeed, it can be verified that
E bγ 2
ℓk (i) ≈ γ 2
lim
ℓk
i→∞
so that the estimator bγ 2
ℓk (i) converges on average close to the desired variance product γ 2
ℓk . In this way, we
can replace the weights (477) by the adaptive construction:
aℓk (i) =
Equations (484) and (486) provide one adaptive construction for the combination weights {aℓk }.
bγ−2
ℓk (i)
Pm∈N
k bγ−2
mk (i)
0,
if ℓ ∈ Nk
otherwise
(486)
(485)
,
80
10 Extensions and Further Considerations
Several extensions and variations of diffusion strategies are possible. Among those variations we mention
strategies that endow nodes with temporal processing abilities, in addition to their spatial cooperation
abilities. We can also apply diffusion strategies to solve recursive least-squares and state-space estimation
problems in a distributed manner. In this section, we highlight select contributions in these and related
areas.
10.1 Adaptive Diffusion Strategies with Smoothing Mechanisms
so that
Jk (w) =
qko > 0,
J glob (w) =
In the ATC and CTA adaptive diffusion strategies (153)–(154), each node in the network shares information
locally with its neighbors through a process of spatial cooperation or combination. In this section, we describe
briefly an extension that adds a temporal dimension to the processing at the nodes. For example, in the
ATC implementation (153), rather than have each node k rely solely on current data, {dℓ(i), uℓ,i , ℓ ∈ Nk },
and on current weight estimates received from the neighbors, {ψℓ,i , ℓ ∈ Nk }, node k can be allowed to
store and process present and past weight estimates, say, P of them as in {ψℓ,j , j = i, i − 1, . . . , i − P + 1}.
In this way, previous weight estimates can be smoothed and used more effectively to help enhance the
mean-square-deviation performance especially in the presence of noise over the communication links.
To motivate diffusion strategies with smoothing mechanisms, we continue to assume that the random
data {dk (i), uk,i } satisfy the modeling assumptions of Sec. 6.1. The global cost (92) continues to be the
same but the individual cost functions (93) are now replaced by:
P −1Xj=0
qkj E dk (i − j ) − uk,i−j w2
qkj E dk (i − j ) − uk,i−j w2
P −1Xj=0
NXk=1
where each coefficient qkj is a non-negative scalar representing the weight that node k assigns to data from
time instant i − j . The coefficients {qkj } are assumed to satisfy the normalization condition:
P −1Xj=0
When the random processes dk (i) and uk,i are jointly wide-sense stationary, as was assumed in Sec. 6.1, the
optimal solution wo that minimizes (488) is still given by the same normal equations (40). We can extend the
arguments of Secs. 3 and 4 to (488) and arrive at the following version of a diffusion strategy incorporating
temporal processing (or smoothing) of the intermediate weight estimates [73, 74]:
φk,i = wk,i−1 + µk Xℓ∈Nk
P −1Xj=0
(temporal processing or smoothing)
fkj φk,i−j
ψk,i =
wk,i = Xℓ∈Nk
cℓk qℓo u∗
ℓ,i [dℓ (i)− uℓ,iwk,i−1 ]
qkj = 1, k = 1, 2, . . . , N
aℓk ψℓ,i
(spatial processing)
(491)
(492)
(adaptation)
(490)
(487)
(488)
(489)
81
where the nonnegative coefficients {cℓk , aℓk , fkj , qlo } satisfy:
for k = 1, 2, . . . , N :
cℓk = 0 if ℓ /∈ Nk
(493)
cℓk ≥ 0,
cℓk = 1,
NXk=1
NXℓ=1
aℓk = 1,
aℓk ≥ 0,
aℓk = 0 if ℓ /∈ Nk
P −1Xj=0
0 < qℓo ≤ 1
Since only the coefficients {qℓo} are needed, we alternatively denote them by the simpler notation {qℓ} in
the listing in Table 8. These are simply chosen as nonnegative coefficients:
fkj ≥ 0,
fkj = 1
(494)
(495)
(496)
0 < qℓ ≤ 1,
ℓ = 1, 2, . . . , N
(497)
Note that algorithm (490)-(492) involves three steps: (a) an adaptation step (A) represented by (490); (b) a
temporal filtering or smoothing step (T) represented by (491), and a spatial cooperation step (S) represented
by (492). These steps are illustrated in Fig. 15. We use the letters (A,T,S) to label these steps; and we use the
sequence of letters (A,T,S) to designate the order of the steps. According to this convention, algorithm (490)-
(492) is referred to as the ATS diffusion strategy since adaptation is followed by temporal processing, which
is followed by spatial processing. In total, we can obtain six different combinations of diffusion algorithms
by changing the order by which the temporal and spatial combination steps are performed in relation to the
adaptation step. The resulting variations are summarized in Table 8. When we use only the most recent
weight vector in the temporal filtering step (i.e., set ψk,i = φk,i ), which corresponds to the case P = 1, the
algorithms of Table 8 reduce to the ATC and CTA diffusion algorithms (153) and (154). Specifically, the
variants TSA, STA, and SAT (where spatial processing S precedes adaptation A) reduce to CTA, while the
variants TAS, ATS, and AST (where adaptation A precedes spatial processing S) reduce to ATC.
Adaptation (A)
Temporal Processing (T)
Spatial Processing (S)
{d1 (i), u1,i }
k
{dk (i), uk,i}
{dℓ (i), uℓ,i}
wk,i−1
k
φk,i
φk,i−1
ψ1,i
ψk,i
φk,i−P +1
{φk ,i−j }
ψℓ,i
{ψℓ,i}
{cℓk , qℓ}
{dℓ (i), uℓ,i }
adaptation
{fkj }
filtering
{aℓ,k }
aggregation
φk,i
ψk,i
wk,i
Figure 15: Illustration of the three steps involved in an ATS diffusion strategy: adaptation, followed by temporal
processing or smoothing, followed by spatial processing.
The mean-square performance analysis of the smoothed diffusion strategies can be pursued by extending
the arguments of Sec. 6. This step is carried out in [73, 74] for doubly stochastic combination matrices A
82
when the filtering coefficients {fkj } do not change with k . For instance, it is shown in [74] that whether
temporal processing is performed before or after adaptation, the strategy that performs adaptation before
spatial cooperation is always better. Specifically, the six diffusion variants can be divided into two groups
with the respective network MSDs satisfying the following relations:
Group #1 :
MSDnetwork
TSA
= MSDnetwork
STA
≥ MSDnetwork
TAS
Group #2 :
MSDnetwork
SAT
≥ MSDnetwork
ATS
= MSDnetwork
AST
(498)
(499)
Note that within groups 1 and 2, the order of the A and T operations is the same: in group 1, T precedes A
and in group 2, A precedes T. Moreover, within each group, the order of the A and S operations determines
performance; the strategy that performs A before S is better. Note further that when P = 1, so that
temporal processing is not performed, then TAS reduces to ATC and TSA reduces to CTA. This conclusion
is consistent with our earlier result (343) that ATC outperforms CTA.
Table 8: Six diffusion strategies with temporal smoothing steps.
TSA diffusion:
P −1Xj=0
fkj wk,i−j−1
φk,i−1 =
ψk,i−1 = Xℓ∈Nk
aℓk φℓ,i−1
wk,i = ψk,i−1 + µk Xℓ∈Nk
STA diffusion:
φk,i−1 = Xℓ∈Nk
aℓk wℓ,i−1
P −1Xj=0
fkj φk,i−j−1
ψk,i−1 =
wk,i = ψk,i−1 + µk Xℓ∈Nk
SAT diffusion:
φk,i−1 = Xℓ∈Nk
aℓkwℓ,i−1
ψk,i = φk,i−1 + µk Xℓ∈Nk
P −1Xj=0
fkj ψk,i−j
wk,i =
TAS diffusion:
P −1Xj=0
fkj wk,i−j−1
φk,i−1 =
ψk,i = φk,i−1 + µk Xℓ∈Nk
ℓ,i [dℓ (i) − uℓ,i ψk,i−1 ] wk,i = Xℓ∈Nk
qℓ cℓk u∗
aℓ,kψℓ,i
ATS diffusion:
φk,i = wk,i−1 + µk Xℓ∈Nk
P −1Xj=0
fkj φk,i−j
ψk,i =
wk,i = Xℓ∈Nk
AST diffusion:
φk,i = wk,i−1 + µk Xℓ∈Nk
ψk,i = Xℓ∈Nk
aℓkφℓ,i
P −1Xj=0
fkj ψk,i−j
wk,i =
qℓ cℓk u∗
ℓ,i [dℓ (i) − uℓ,iφk,i−1 ]
qℓ cℓk u∗
ℓ,i [dℓ (i) − uℓ,iψk,i−1 ]
aℓk ψℓ,i
qℓ cℓk u∗
ℓ,i [dℓ (i) − uℓ,iφk,i−1 ]
qℓ cℓk u∗
ℓ,i [dℓ (i) − uℓ,iwk,i−1 ]
qℓ cℓk u∗
ℓ,i [dℓ (i) − uℓ,iwk,i−1 ]
In related work, reference [75] started from the CTA algorithm (159) without information exchange and
added a useful pro jection step to it between the combination step and the adaptation step; i.e., the work
considered an algorithm with an STA structure (with spatial combination occurring first, followed by a
pro jection step, and then adaptation). The pro jection step uses the current weight estimate, φk,i , at node
k and pro jects it onto hyperslabs defined by the current and past raw data. Specifically, the algorithm
83
from [75] has the following form:
φk,i−1 = Xℓ∈Nk
aℓk wℓ,i−1
′
ψk,i−1 = P
k,i [φk,i−1 ]
fkj · Pk,i−j [φk,i−1 ]
wk,i = ψk,i−1 − µk
P −1Xj=0
ψk,i−1 −
where the notation ψ = Pk,i [φ] refers to the act of pro jecting the vector φ onto the hyperslab Pk,i that
′
consists of all M × 1 vectors z satisfying (similarly for the pro jection P
k,i ):
∆= { z such that dk (i) − uk,iz ≤ ǫk }
Pk,i
′
∆= { z such that dk (i) − uk,iz ≤ ǫ′
k }
(500)
(501)
(502)
P
k,i
(503)
(504)
if d − ǫ > uφ
P [φ] = φ +
u∗
kuk2 [d − ǫ − uφ] ,
k } are positive (tolerance) parameters chosen by the designer to satisfy ǫ′
where {ǫk , ǫ′
k > ǫk . For generic
values {d, u, ǫ}, where d is a scalar and u is a row vector, the pro jection operator is described analytically
by the following expression [76]:
The pro jections that appear in (501)–(502) can be regarded as another example of a temporal processing
step. Observe from the middle plot in Fig. 15 that the temporal step that we are considering in the
algorithms listed in Table 8 is based on each node k using its current and past weight estimates, such as
{φk,i , φk,i−1 , . . . , φk,i−P +1 }, rather than only φk,i and current and past raw data {dk (i), dk (i − 1), . . . , dk (i −
P + 1), uk,i , uk,i−1 , . . . , uk,i−P +1 }. For this reason, the temporal processing steps in Table 8 tend to exploit
information from across the network more broadly and the resulting mean-square error performance is
generally improved relative to (500)–(502).
u∗
kuk2 [d + ǫ − uφ] ,
if d − uφ ≤ ǫ
if d + ǫ < uφ
(505)
0,
10.2 Diffusion Recursive Least-Squares
Diffusion strategies can also be applied to recursive least-squares problems to enable distributed solutions of
least-squares designs [28,29]; see also [72]. Thus, consider again a set of N nodes that are spatially distributed
over some domain. The ob jective of the network is to collectively estimate some unknown column vector of
length M , denoted by wo , using a least-squares criterion. At every time instant i, each node k collects a
scalar measurement, dk (i), which is assumed to be related to the unknown vector wo via the linear model:
dk (i) = uk,iwo + vk (i)
(506)
In the above relation, the vector uk,i denotes a row regression vector of length M , and vk (i) denotes measure-
ment noise. A snapshot of the data in the network at time i can be captured by collecting the measurements
and noise samples, {dk (i), vk (i)}, from across all nodes into column vectors yi and vi of sizes N × 1 each,
and the regressors {uk,i} into a matrix Hi of size N × M :
yi =
vi =
(N × 1), Hi =
d1 (i)
d2 (i)
...
dN (i)
v1 (i)
v2 (i)
...
vN (i)
u1,i
u2,i
...
uN ,i
(N × M )
(N × 1),
(507)
84
Likewise, the history of the data across the network up to time i can be collected into vector quantities as
follows:
Yi =
Vi =
Hi =
yi
vi
Hi
yi−1
vi−1
Hi−1
...
...
...
y0
v0
H0
Then, one way to estimate wo is by formulating a global least-squares optimization problem of the form:
Πi + kYi − Hiwk2
kwk2
min
Wi
w
(508)
(509)
,
,
where Πi > 0 represents a Hermitian regularization matrix and Wi ≥ 0 represents a Hermitian weighting
matrix. Common choices for Πi and Wi are
Wi = diag{IN , λIN , . . . , λi IN }
Πi = λi+1 δ−1
(510)
(511)
where δ > 0 is usually a large number and 0 ≪ λ ≤ 1 is a forgetting factor whose value is generally very
close to one. In this case, the global cost function (509) can be written in the equivalent form:
dk (j ) − uk,j w2!
λi−j NXk=1
iXj=0
which is an exponentially weighted least-squares problem. We see that, for every time instant j , the squared
errors, dk (j ) − uk,j w2 , are summed across the network and scaled by the exponential weighting factor λi−j .
The index i denotes current time and the index j denotes a time instant in the past.
In this way, data
occurring in the remote past are scaled more heavily than data occurring closer to present time. The global
solution of (509) is given by [5]:
λi+1 kwk2 +
min
w
(512)
wi = [Πi + HiWiHi ]−1 H∗
i WiYi
(513)
and the notation wi , with a subscript i, is meant to indicate that the estimate wi is based on all data
collected from across the network up to time i. Therefore, the wi that is computed via (513) amounts to a
global construction.
In [28, 29] a diffusion strategy was developed that allows nodes to approach the global solution wi by
relying solely on local interactions. Let wk,i denote a local estimate for wo that is computed by node k at
time i. The diffusion recursive-least-squares (RLS) algorithm takes the following form. For every node k ,
we start with the initial conditions wk,−1 = 0 and Pk,−1 = δIM , where Pk,−1 is an M × M matrix. Then,
for every time instant i, each node k performs an incremental step followed by a diffusion step as follows:
85
Diffusion RLS.
Step 1 (incremental update)
ψk,i ← wk,i−1
Pk,i ← λ−1Pk,i−1
for every neighboring node ℓ ∈ Nk , update :
cℓkPk,i u∗
ℓ,i
1 + cℓkuℓ,iPk,iu∗
ℓ,i
ψk,i ← ψk,i +
[dℓ,i − uℓ,iψk,i ]
Pk,i ← Pk,i −
cℓkPk,i u∗
ℓ,iuℓ,iPk,i
1 + cℓk uℓ,iPk,i u∗
ℓ,i
end
Step 2 (diffusion update)
wk,i = Xℓ∈Nk
aℓkψℓ,i
(514)
where the symbol ← denotes a sequential assignment, and where the scalars {aℓk , cℓk } are nonnegative
coefficients satisfying:
for k = 1, 2, . . . , N :
aℓk ≥ 0,
(515)
(516)
cℓk ≥ 0,
cℓk = 1,
cℓk = 0 if ℓ /∈ Nk
aℓk = 0 if ℓ /∈ Nk
NXk=1
NXℓ=1
aℓk = 1,
The above algorithm requires that at every instant i, nodes communicate to their neighbors their mea-
surements {dℓ (i), uℓ,i} for the incremental update, and the intermediate estimates {ψℓ,i} for the diffusion
update. During the incremental update, node k cycles through its neighbors and incorporates their data
contributions represented by {dℓ (i), uℓ,i} into {ψk,i , Pk,i }. Every other node in the network is performing
similar steps. At the end of the incremental step, neighboring nodes share their intermediate estimates
{ψℓ,i} to undergo diffusion. Thus, at the end of both steps, each node k would have updated the quantities
{wk,i−1 , Pk,i−1 } to {wk,i , Pk,i }. The quantities Pk,i are matrices of size M × M each. Observe that the
diffusion RLS implementation (514) does not require the nodes to share their matrices {Pℓ,i}, which would
amount to a substantial burden in terms of communications resources since each of these matrices has M 2
entries. Only the quantities {dℓ (i), uℓ,i , ψℓ,i} are shared. The mean-square performance and convergence of
the diffusion RLS strategy are studied in some detail in [29].
The incremental step of the diffusion RLS strategy (514) corresponds to performing a number of Nk
successive least-squares updates starting from the initial conditions {wk,i−1 , Pk,i−1 } and ending with the
values {ψk,i , Pk,i } that move on to the diffusion update step.
It can be verified from the properties of
recursive least-squares solutions [4, 5] that these variables satisfy the following equations at the end of the
incremental stage (step 1):
k,i−1 + Xℓ∈Nk
P −1
k,i = λP −1
cℓk u∗
ℓ,iuℓ,i
k,i−1wk,i−1 + Xℓ∈Nk
P −1
k,i ψk,i = λP −1
86
cℓku∗
ℓ,idℓ (i)
(517)
(518)
Introduce the auxiliary M × 1 variable:
∆= P −1
k,i ψk,i
Then, the above expressions lead to the following alternative form of the diffusion RLS strategy (514).
qk,i
(519)
Alternative form of diffusion RLS.
wk,i−1 = Xℓ∈Nk
aℓkψℓ,i−1
k,i−1 + Xℓ∈Nk
P −1
k,i = λP −1
cℓku∗
ℓ,iuℓ,i
k,i−1wk,i−1 + Xℓ∈Nk
qk,i = λP −1
ψk,i = Pk,i qk,i
cℓk u∗
ℓ,idℓ (i)
(520)
Under some approximations, and for the special choices A = C and λ = 1, the diffusion RLS strategy
(520) can be reduced to a form given in [79] and which is described by the following equations:
cℓk hP −1
ℓ,iuℓ,ii
k,i = Xℓ∈Nk
P −1
ℓ,i−1 + u∗
qk,i = Xℓ∈Nk
cℓk (cid:2)qℓ,i−1 + u∗
ℓ,idℓ (i)(cid:3)
ψk,i = Pk,i qk,i
(522)
(523)
(521)
Algorithm (521)–(523) is motivated in [79] by using consensus-type arguments. Observe that the algorithm
requires the nodes to share the variables {dℓ (i), uℓ,i , qℓ,i−1 , Pℓ,i−1 }, which corresponds to more communica-
tions overburden than required by diffusion RLS; the latter only requires that nodes share {dℓ (i), uℓ,i , ψℓ,i−1}.
In order to illustrate how a special case of diffusion RLS (520) can be related to this scheme, let us set
A = C and λ = 1
(524)
Then, equations (520) give:
Special form of diffusion RLS when A = C and λ = 1.
wk,i−1 = Xℓ∈Nk
cℓk ψℓ,i−1
k,i−1 + Xℓ∈Nk
P −1
k,i = P −1
cℓku∗
ℓ,iuℓ,i
k,i−1wk,i−1 + Xℓ∈Nk
qk,i = P −1
ψk,i = Pk,i qk,i
cℓk u∗
ℓ,idℓ (i)
87
(525)
Comparing these equations with (521)–(523), we find that algorithm (521)–(523) of [79] would relate to the
diffusion RLS algorithm (520) when the following approximations are justified:
Xℓ∈Nk
Xℓ∈Nk
(526)
cℓkP −1
ℓ,i−1ψℓ,i−1
ℓ,i−1 ≈ P −1
cℓk P −1
k,i−1
cℓk qℓ,i−1 = Xℓ∈Nk
≈ Xℓ∈Nk
cℓkP −1
k,i−1 ψℓ,i−1
k,i−1 Xℓ∈Nk
= P −1
= P −1
k,i−1wk,i−1
cℓk ψℓ,i−1
(527)
(528)
It was indicated in [29] that the diffusion RLS implementation (514) or (520) leads to enhanced performance
in comparison to the consensus-based update (521)–(523).
10.3 Diffusion Kalman Filtering
Diffusion strategies can also be applied to the solution of distributed state-space filtering and smoothing
problems [30, 31, 33]. Here, we describe briefly the diffusion version of the Kalman filter; other variants and
smoothing filters can be found in [33]. We assume that some system of interest is evolving according to linear
state-space dynamics, and that every node in the network collects measurements that are linearly related to
the unobserved state vector. The ob jective is for every node to track the state of the system over time based
solely on local observations and on neighborhood interactions.
Thus, consider a network consisting of N nodes observing the state vector, xi , of size n × 1 of a linear
state-space model. At every time i, every node k collects a measurement vector yk,i of size p × 1, which is
related to the state vector as follows:
xi+1 = Fixi + Gini
yk,i = Hk,ixi + v k,i , k = 1, 2, . . . , N
(529)
(530)
The signals ni and vk,i denote state and measurement noises of sizes n × 1 and p × 1, respectively, and they
are assumed to be zero-mean, uncorrelated and white, with covariance matrices denoted by
0 Rk,i (cid:21) δij
vk,j (cid:21)∗
vk,i (cid:21) (cid:20) nj
∆= (cid:20) Qi
E (cid:20) ni
0
The initial state vector, xo , is assumed to be zero-mean with covariance matrix
(531)
E xox∗
o = Πo > 0
(532)
and is uncorrelated with ni and vk,i , for all i and k . We further assume that Rk,i > 0. The parameter
matrices {Fi , Gi , Hk,i , Qi , Rk,i , Πo } are assumed to be known by node k .
Let bxk,ij denote a local estimator for xi that is computed by node k at time i based solely on local
observations and on neighborhood data up to time j . The following diffusion strategy was proposed in
[30, 31, 33] to evaluate approximate predicted and filtered versions of these local estimators in a distributed
manner for data satisfying model (529)–(532). For every node k , we start with bxk,0−1 = 0 and Pk,0−1 = Πo ,
where Pk,0−1 is an M × M matrix. At every time instant i, every node k performs an incremental step
88
followed by a diffusion step:
Time and measurement-form of the diffusion Kalman filter.
Step 1 (incremental update)
ψk,i ← bxk,ii−1
Pk,i ← Pk,ii−1
for every neighboring node ℓ ∈ Nk , update :
Re ← Rℓ,i + Hℓ,iPk,iH ∗
(cid:2)y ℓ,i − Hℓ,iψk,i (cid:3)
ℓ,i
ℓ,iR−1
ψk,i ← ψk,i + Pk,iH ∗
e
Pk,i ← Pk,i − Pk,iH ∗
ℓ,iR−1
e Hℓ,iPk,i
end
Step 2 (diffusion update)
bxk,ii = Xℓ∈Nk
aℓkψ ℓ,i
Pk,ii = Pk,i
bxk,i+1i = Fi bxk,ii
i + GiQiG∗
Pk,i+1i = FiPk,ii F ∗
i .
(533)
where the symbol ← denotes a sequential assignment, and where the scalars {aℓk } are nonnegative coefficients
satisfying:
for k = 1, 2, . . . , N :
aℓk ≥ 0,
aℓk = 1,
(534)
aℓk = 0 if ℓ /∈ Nk
NXℓ=1
The above algorithm requires that at every instant i, nodes communicate to their neighbors their mea-
surement matrices Hℓ,i , the noise covariance matrices Rℓ,i , and the measurements y ℓ,i for the incremental
update, and the intermediate estimators ψ ℓ,i for the diffusion update. During the incremental update, node
k cycles through its neighbors and incorporates their data contributions represented by {y ℓ,i , Hℓ,i , Rℓ,i} into
{ψk,i , Pk,i }. Every other node in the network is performing similar steps. At the end of the incremental step,
neighboring nodes share their updated intermediate estimators {ψ ℓ,i} to undergo diffusion. Thus, at the
end of both steps, each node k would have updated the quantities {bxk,ii−1 , Pk,ii−1 } to {bxk,i+1i , Pk,i+1i }.
The quantities Pk,ii−1 are n × n matrices.
It is important to note that even though the notation Pk,ii
and Pk,ii−1 has been retained for these variables, as in the standard Kalman filtering notation [5, 77], these
matrices do not represent any longer the covariances of the state estimation errors, xk,ii−1 = xi − bxk,ii−1 ,
but can be related to them [33].
An alternative representation of the diffusion Kalman filter may be obtained in information form by
further assuming that Pk,ii−1 > 0 for all k and i; a sufficient condition for this fact to hold is to requires
the matrices {Fi} to be invertible [77]. Thus, consider again data satisfying model (529)–(532). For every
node k , we start with bxk,0−1 = 0 and P −1
k,0−1 = Π−1
o . At every time instant i, every node k performs an
incremental step followed by a diffusion step:
89
Information form of the diffusion Kalman filter.
ℓ,iR−1
H ∗
ℓ,i y ℓ,i
Step 1 (incremental update)
Sk,i = Xℓ∈Nk
ℓ,iR−1
H ∗
ℓ,i Hℓ,i
qk,i = Xℓ∈Nk
k,ii = P −1
P −1
k,ii−1 + Sk,i
ψk,i = bxk,ii−1 + Pk,ii (cid:2)qk,i − Sk,i bxk,ii−1 (cid:3)
Step 2: (diffusion update)
bxk,ii = Xℓ∈Nk
aℓkψ ℓ,i
bxk,i+1i = Fi bxk,ii
i + GiQiG∗
Pk,i+1i = FiPk,iiF ∗
i
(535)
ǫ · ψ ℓ,i
(ψ ℓ,i − ψk,i )
The incremental update in (535) is similar to the update used in the distributed Kalman filter derived
in [48]. An important difference in the algorithms is in the diffusion step. Reference [48] starts from a
continuous-time consensus implementation and discretizes it to arrive at the following update relation:
bxk,ii = ψ k,i + ǫ Xℓ∈Nk
which, in order to facilitate comparison with (535), can be equivalently rewritten as:
bxk,ii = (1 + ǫ − nk ǫ) · ψk,i + Xℓ∈Nk \{k}
where nk denotes the degree of node k (i.e., the size of its neighborhood, Nk ). In comparison, the diffusion
step in (535) can be written as:
bxk,ii = akk · ψk,i + Xℓ∈Nk \{k}
Observe that the weights used in (537) are (1 + ǫ − nk ǫ) for the node’s estimator, ψk,i , and ǫ for all other
estimators, {ψℓ,i}, arriving from the neighbors of node k . In contrast, the diffusion step (538) employs a
convex combination of the estimators {ψ ℓ,i} with generally different weights {aℓk } for different neighbors;
this choice is motivated by the desire to employ combination coefficients that enhance the fusion of infor-
mation at node k , as suggested by the discussion in App. D of [33]. It was verified in [33] that the diffusion
implementation (538) leads to enhanced performance in comparison to the consensus-based update (537).
Moreover, the weights {aℓk } in (538) can also be adjusted over time in order to further enhance performance,
as discussed in [78]. The mean-square performance and convergence of the diffusion Kalman filtering imple-
mentations are studied in some detail in [33], along with other diffusion strategies for smoothing problems
including fixed-point and fixed-lag smoothing.
aℓk · ψ ℓ,i
(536)
(537)
(538)
90
10.4 Diffusion Distributed Optimization
The ATC and CTA steepest-descent diffusion strategies (134) and (142) derived earlier in Sec. 3 provide
distributed mechanisms for the solution of global optimization problems of the form:
NXk=1
where the individual costs, Jk (w), were assumed to be quadratic in w, namely,
Jk (w)
min
w
(539)
d,k − w∗ rdu,k − r∗
du,k w + w∗Ru,k w
Jk (w) = σ2
(540)
for given parameters {σ2
d,k , rdu,k , Ru,k }. Nevertheless, we remarked in that section that similar diffusion
strategies can be applied to more general cases involving individual cost functions, Jk (w), that are not
necessarily quadratic in w [1–3]. We restate below, for ease of reference, the general ATC and CTA diffusion
strategies (139) and (146) that can be used for the distributed solution of global optimization problems of
the form (539) for more general convex functions Jk (w):
ψk,i = wk,i−1 − µk Xℓ∈Nk
wk,i = Xℓ∈Nk
aℓk ψℓ,i
ψk,i−1 = Xℓ∈Nk
aℓk wℓ,i−1
wk,i = ψk,i−1 − µk Xℓ∈Nk
cℓk [∇w Jℓ (ψk,i−1 )]∗
cℓk [∇w Jℓ (wk,i−1 )]∗
(ATC strategy)
(CTA strategy)
(542)
(541)
and
for positive step-sizes {µk } and for nonnegative coefficients {cℓk , aℓk } that satisfy:
for k = 1, 2, . . . , N :
cℓk ≥ 0,
NXk=1
NXℓ=1
aℓk = 1,
That is, the matrix A = [aℓk ] is left-stochastic while the matrix C = [cℓk ] is right-stochastic:
aℓk = 0 if ℓ /∈ Nk
cℓk = 0 if ℓ /∈ Nk
aℓk ≥ 0,
cℓk = 1,
C 1 = 1,
AT 1 = 1
(543)
(544)
We can again regard the above ATC and CTA strategies as special cases of the following general diffusion
scheme:
φk,i−1 = Xℓ∈Nk
a1,ℓk wℓ,i−1
ψk,i = φk,i−1 − µk Xℓ∈Nk
wk,i = Xℓ∈Nk
a2,ℓk ψℓ,i
cℓk [∇w Jℓ (φk,i−1 )]∗
(545)
(546)
(547)
91
where the coefficients {a1,ℓk , a2,ℓk , cℓk } are nonnegative coefficients corresponding to the (l, k)-th entries of
combination matrices {A1 , A2 , C } that satisfy:
AT
1
1 = 1, AT
2
1 = 1, C 1 = 1
(548)
The convergence behavior of these diffusion strategies can be examined under both conditions of noiseless
updates (when the gradient vectors are available) and noisy updates (when the gradient vectors are sub ject
to gradient noise). The following properties can be proven for the diffusion strategies (545)–(547) [1–3]. The
statements that follow assume, for convenience of presentation, that all data are real-valued; the conditions
would need to be adjusted for complex-valued data.
Noiseless Updates
Let
Jk (w)
J glob (w) =
NXk=1
denote the global cost function that we wish to minimize. Assume J glob (w) is strictly convex so that its
minimizer wo is unique. Assume further that each individual cost function Jk (w) is convex and has a mini-
mizer at the same wo . This case is common in practice; situations abound where nodes in a network need to
work cooperatively to attain a common ob jective (such as tracking a target, locating the source of chemical
leak, estimating a physical model, or identifying a statistical distribution). The case where the {Jk (w)} have
different individual minimizers is studied in [1, 3], where it is shown that the same diffusion strategies of this
section are still applicable and nodes would converge instead to a Pareto-optimal solution.
(549)
Theorem 10.1. (Convergence to Optimal Solution: Noise-Free Case) Consider the problem of
minimizing the strictly convex global cost (549), with the individual cost functions {Jk (w)} assumed to be
convex with each having a minimizer at the same wo . Assume that al l data are real-valued and suppose the
Hessian matrices of the individual costs are bounded from below and from above as fol lows:
λℓ,min IM ≤ ∇2
w Jℓ (w) ≤ λℓ,maxIM ,
ℓ = 1, 2, . . . , N
for some positive constants {λℓ,min , λℓ,max}. Let
∆= Xℓ∈Nk
∆= Xℓ∈Nk
Assume further that σk,min > 0 and that the positive step-sizes are chosen such that:
cℓk λℓ,min ,
cℓk λℓ,max
σk,max
σk,min
µk ≤
2
σk,max
,
k = 1, . . . , N
(550)
(551)
(552)
Then, it holds that wk,i → wo as i → ∞. That is, the weight estimates generated by (545)–(547) at al l nodes
wil l tend towards the desired global minimizer.
(cid:3)
We note that in works on distributed sub-gradient methods (e.g., [40, 80]), the norms of the sub-gradients
are usually required to be uniformly bounded. Such a requirement is restrictive in the unconstrained op-
timization of differentiable functions. Condition (550) is more relaxed since it allows the gradient vector
∇w Jℓ (w) to have unbounded norm. This extension is important because requiring bounded gradient norms,
as opposed to bounded Hessian matrices, would exclude the possibility of using quadratic costs for the Jℓ (w)
(since the gradient vectors would then be unbounded). And, as we saw in the body of the chapter, quadratic
costs play a critical role in adaptation and learning over networks.
92
Updates with Gradient Noise
It is often the case that we do not have access to the exact gradient vectors to use in (546), but to noisy
versions of them, say,
\∇w Jℓ (φk,i−1 ) ∆= ∇w Jℓ (φk,i−1 ) + v ℓ ( eφk,i−1 )
where the random vector variable v ℓ (·) refers to gradient noise; its value is generally dependent on the
weight-error vector realization,
∆= wo − φk,i−1
eφk,i−1
at which the gradient vector is being evaluated. In the presence of gradient noise, the weight estimates at the
various nodes become random quantities and we denote them by the boldface notation {wk,i }. We assume
that, conditioned on the past history of the weight estimators at all nodes, namely,
(553)
(554)
Fi−1
∆= {wm,j , m = 1, 2, . . . , N , j < i}
(555)
(557)
(556)
the gradient noise has zero mean and its variance is upper bounded as follows:
E nv ℓ ( eφk,i−1 ) Fi−1o = 0
E nkv ℓ ( eφk,i−1 )k2 Fi−1o ≤ αk eφk,i−1 k2 + σ2
v
for some α > 0 and σ2
v ≥ 0. Condition (557) allows the variance of the gradient noise to be time-variant,
so long as it does not grow faster than E k eφk,i−1 k2 . This condition on the noise is more general than the
“uniform-bounded assumption” that appears in [40], which required instead:
E nkv ℓ ( eφk,i−1 )k2 Fi−1o ≤ σ2
E nkvℓ ( eφk,i−1 )k2o ≤ σ2
(558)
v ,
v
These two requirements are special cases of (557) for α = 0. Furthermore, condition (557) is similar to
condition (4.3) in [81], which requires the noise variance to satisfy:
E nkv ℓ ( eφk,i−1 )k2 Fi−1o ≤ α (cid:2)k∇w Jℓ (φk,i−1 )k2 + 1(cid:3)
This requirement can be verified to be a combination of the “relative random noise” and the “absolute
random noise” conditions defined in [22] — see [2].
Now, introduce the column vector:
NXℓ=1
col {cℓ1v ℓ (wo ), cℓ2v ℓ (wo ), . . . , cℓN v ℓ (wo )}
(560)
(559)
∆=
z i
and let
where
Let further
Z ∆= E z iz ∗
i
∆= col { ewi,1 , ewi,2 , . . . , ewi,N }
ew i
∆= wo − wk,i
ewk,i
Then, the following result can be established [2]; it characterizes the network mean-square deviation in
steady-state, which is defined as
i→∞ 1
E k ewk,i k2!
MSDnetwork ∆= lim
N
NXk=1
(564)
(563)
(561)
(562)
93
Theorem 10.2. (Mean-Square Stability: Noisy Case) Consider the problem of minimizing the strictly
convex global cost (549), with the individual cost functions {Jk (w)} assumed to be convex with each having a
minimizer at the same wo . Assume al l data are real-valued and suppose the Hessian matrices of the individual
costs are bounded from below and from above as stated in (550). Assume further that the diffusion strategy
(545)–(547) employs noisy gradient vectors, where the noise terms are zero mean and satisfy conditions
(557) and (561). We select the positive step-sizes to be sufficiently smal l and to satisfy:
1 )
µk < min (
2σk,min
k,min + αkC k2
σ2
for k = 1, 2, . . . , N . Then, the diffusion strategy (545)–(547) is mean-square stable and the mean-square-
deviation of the network is given by:
2σk,max
k,max + αkC k2
σ2
1
(565)
,
where
MSDnetwork ≈
· (I − F )−1 · vec(IN M )
2 MZ T MA2 (cid:1)(cid:3)T
N (cid:2)vec (cid:0)AT
1
A2 = A2 ⊗ IM
M = diag{µ1 IM , µ2 IM , . . . , µN IM }
F ≈ BT ⊗ B ∗
2 (I − MR) AT
B = AT
1
NXℓ=1
w Jℓ (wo )(cid:9)
diag (cid:8)cℓ1∇2
w Jℓ (wo ), . . . , cℓN ∇2
w Jℓ (wo ), cℓ2∇2
R =
(566)
(567)
(568)
(569)
(570)
(571)
(cid:3)
Acknowledgements
The development of the theory and applications of diffusion adaptation over networks has benefited greatly
from the insights and contributions of several UCLA PhD students, and several visiting graduate students
to the UCLA Adaptive Systems Laboratory (http://www.ee.ucla.edu/asl). The assistance and contributions
of all students are hereby gratefully acknowledged, including Cassio G. Lopes, Federico S. Cattivelli, Sheng-
Yuan Tu, Jianshu Chen, Xiaochuan Zhao, Zaid Towfic, Chung-Kai Yu, Noriyuki Takahashi, Jae-Woo Lee,
Alexander Bertrand, and Paolo Di Lorenzo. The author is also particularly thankful to S.-Y. Tu, J. Chen,
X. Zhao, Z. Towfic, and C.-K. Yu for their assistance in reviewing an earlier draft of this article.
94
A Appendix: Properties of Kronecker Products
For ease of reference, we collect in this appendix some useful properties of Kronecker products. All matrices
are assumed to be of compatible dimensions; all inverses are assumed to exist whenever necessary. Let
E = [eij ]n
i,j=1 and B = [bij ]m
i,j=1 be n × n and m × m matrices, respectively. Their Kronecker product is
denoted by E ⊗ B and is defined as the nm × nm matrix whose entries are given by [20]:
E ⊗ B =
In other words, each entry of E is replaced by a scaled multiple of B . Let {λi (E ), i = 1, . . . , n} and
{λj (B ), j = 1, . . . , m} denote the eigenvalues of E and B , respectively. Then, the eigenvalues of E ⊗ B will
consist of all nm product combinations {λi (E )λj (B )}. Table 9 lists some well-known properties of Kronecker
products.
e11B e12B . . .
e21B e22B . . .
...
...
en1B en2B . . .
(572)
e1nB
e2nB
ennB
Table 9: Properties of Kronecker products.
(E + B ) ⊗ C = (E ⊗ C ) + (B ⊗ C )
(E ⊗ B )(C ⊗ D) = (EC ⊗ BD)
(E ⊗ B )T = E T ⊗ B T
(E ⊗ B )∗ = E ∗ ⊗ B ∗
(E ⊗ B )−1 = E−1 ⊗ B−1
(E ⊗ B )ℓ = E ℓ ⊗ B ℓ
{λ(E ⊗ B )} = {λi (E )λj (B )}n,m
i=1,j=1
det(E ⊗ B ) = (det E )m (det B )n
Tr(E ⊗ B ) = Tr(E )Tr(B )
Tr(EB ) = (cid:2)vec(B T )(cid:3)T
vec(E )
vec(ECB ) = (B T ⊗ E )vec(C )
B Appendix: Graph Laplacian and Network Connectivity
Consider a network consisting of N nodes and L edges connecting the nodes to each other. In the construc-
tions below, we only need to consider the edges that connect distinct nodes to each other; these edges do
not contain any self-loops that may exist in the graph and which connect nodes to themselves directly. In
other words, when we refer to the L edges of a graph, we are excluding self-loops from this set; but we are
still allowing loops of at least length 2 (i.e., loops generated by paths covering at least 2 edges).
The neighborhood of any node k is denoted by Nk and it consists of all nodes that node k can share
information with; these are the nodes that are connected to k through edges, in addition to node k itself.
The degree of node k , which we denote by nk , is defined as the positive integer that is equal to the size of
its neighborhood:
nk
∆= Nk
(573)
Since k ∈ Nk , we always have nk ≥ 1. We further associate with the network an N × N Laplacian matrix,
95
(574)
nk − 1,
if k = ℓ
−1,
if k 6= ℓ and nodes k and ℓ are neighbors
0, otherwise
denoted by L. The matrix L is symmetric and its entries are defined as follows [64–66]:
[L]kℓ =
Note that the term nk − 1 measures the number of edges that are incident on node k , and the locations
of the −1′s on row k indicate the nodes that are connected to node k . We also associate with the graph
an N × L incidence matrix, denoted by I . The entries of I are defined as follows. Every column of I
represents one edge in the graph. Each edge connects two nodes and its column will display two nonzero
entries at the rows corresponding to these nodes: one entry will be +1 and the other entry will be −1. For
directed graphs, the choice of which entry is positive or negative can be used to identify the nodes from
which edges emanate (source nodes) and the nodes at which edges arrive (sink nodes). Since we are dealing
with undirected graphs, we shall simply assign positive values to lower indexed nodes and negative values to
higher indexed nodes:
[I ]ke =
Figure 16 shows the example of a network with N = 6 nodes and L = 8 edges. Its Laplacian and incidence
matrices are also shown and these have sizes 6 × 6 and 6 × 8, respectively. Consider, for example, column
6 in the incidence matrix. This column corresponds to edge 6, which links nodes 3 and 5. Therefore, at
location I36 we have a +1 and at location I56 we have −1. The other columns of I are constructed in a
similar manner.
if node k is the lower-indexed node connected to edge e
+1,
−1,
if node k is the higher-indexed node connected to edge e
0, otherwise
(575)
nodes
1 2 3 4 5 6
1
1
L =
2 −1 −1
−1
3 −1 −1
0
0
0
−1 −1
4 −1 −1
0
0
0
0 −1 −1
0 −1
3
0
0 −1
2 −1
0
0 −1 −1
2
0
0
6
3
5
4
8
7
5
6
3
4
2
2
I =
1
1
0 −1
0
1
−1
0 −1
0
1
0
0
0
1
0 −1 −1
0
0
1
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0 −1
0
0 −1 −1
1
3
4
5
2
6
nodes
1
2
3
4
5
6
edges
1 2 3 4 5 6
7 8
Figure 16: A network with N = 6 nodes and L = 8 edges. The nodes are marked 1 through 6 and the edges are
marked 1 through 8. The corresponding Laplacian and incidence matrices L and I are 6 × 6 and 6 × 8.
Observe that the Laplacian and incidence matrices of a graph are related as follows:
L = I I T
96
(576)
The Laplacian matrix conveys useful information about the topology of the graph. The following is a classical
result from graph theory [64–67].
Lemma B.1. (Laplacian and Network Connectivity) Let
θ1 ≥ θ2 ≥ . . . ≥ θN
(577)
denote the ordered eigenvalues of L. Then the fol lowing properties hold:
(a) L is symmetric nonnegative-definite so that θi ≥ 0.
(b) The rows of L add up to zero so that L1 = 0. This means that 1 is a right eigenvector of L corresponding
to the eigenvalue zero.
(c) The smal lest eigenvalue is always zero, θN = 0. The second smal lest eigenvalue, θN −1 , is cal led the
algebraic connectivity of the graph.
(d) The number of times that zero is an eigenvalue of L (i.e., its multiplicity) is equal to the number of
connected subgraphs.
(e) The algebraic connectivity of a connected graph is nonzero, i.e., θN −1 6= 0. In other words, a graph is
connected if, and only if, its algebraic connectivity is nonzero.
Proof. Property (a) follows from the identity L = I I T . Property (b) follows from the definition of L. Note that
for each row of L, the entries on the row add up to zero. Property (c) follows from properties (a) and (b) since
L1 = 0 implies that zero is an eigenvalue of L. For part (d), assume the network consists of two separate connected
subgraphs. Then, the Laplacian matrix would have a block diagonal structure, say, of the form L = diag{L1 , L2},
where L1 and L2 are the Laplacian matrices of the smaller subgraphs. The smallest eigenvalue of each of these
Laplacian matrices would in turn be zero and unique by property (e). More generally, if the graph consists of m
connected subgraphs, then the multiplicity of zero as an eigenvalue of L must be m. To establish property (e), first
observe that if the algebraic connectivity is nonzero then it is obvious that the graph must be connected. Otherwise,
if the graph were disconnected, then its Laplacian matrix would be block diagonal and the algebraic multiplicity of
zero as an eigenvalue of L would be larger than one so that θN −1 would be zero, which is a contradiction. For the
converse statement, assume the graph is connected and let x denote an arbitrary eigenvector of L corresponding to
the eigenvalue at zero, i.e., Lx = 0. We already know that L1 = 1 from property (b). Let us verify that x must
be proportional to the vector 1 so that the algebraic multiplicity of the eigenvalue at zero is one. Thus note that
xT Lx = 0. If we denote the individual entries of x by xk , then this identity implies that for each node k:
Xℓ∈Nk
It follows that the entries of x within each neighborhood have equal values. But since the graph is connected, we
conclude that all entries of x must be equal. It follows that the eigenvector x is proportional to the vector 1, as
desired.
(xk − xℓ )2 = 0
C Appendix: Stochastic Matrices
Consider N × N matrices A with nonnegative entries, {aℓk ≥ 0}. The matrix A = [aℓk ] is said to be
right-stochastic if it satisfies
in which case each row of A adds up to one. The matrix A is said to be left-stochastic if it satisfies
A1 = 1
(right-stochastic)
AT 1 = 1
(left-stochastic)
97
(578)
(579)
in which case each column of A adds up to one. And the matrix is said to be doubly stochastic if both
conditions hold so that both its columns and rows add up to one:
A1 = 1,
AT 1 = 1
(doubly-stochastic)
(580)
Stochastic matrices arise frequently in the study of adaptation over networks. This appendix lists some of
their properties.
Lemma C.1. (Spectral Norm of Stochastic Matrices) Let A be an N × N right or left or doubly
stochastic matrix. Then, ρ(A) = 1 and, therefore, al l eigenvalues of A lie inside the unit disc, i.e., λ(A) ≤ 1.
Proof. We prove the result for right stochastic matrices; a similar argument applies to left or doubly stochastic
matrices. Let A be a right-stochastic matrix. Then, A1 = 1, so that λ = 1 is one of the eigenvalues of A. Moreover,
for any matrix A, it holds that ρ(A) ≤ kAk∞ , where k · k∞ denotes the maximum absolute row sum of its matrix
argument. But since all rows of A add up to one, we have kAk∞ = 1. Therefore, ρ(A) ≤ 1. And since we already
know that A has an eigenvalue at λ = 1, we conclude that ρ(A) = 1.
The above result asserts that the spectral radius of a stochastic matrix is unity and that A has an eigenvalue
at λ = 1. The result, however, does not rule out the possibility of multiple eigenvalues at λ = 1, or even
other eigenvalues with magnitude equal to one. Assume, in addition, that the stochastic matrix A is regular.
This means that there exists an integer power jo such that all entries of Ajo are strictly positive, i.e.,
for all (ℓ, k), it holds that (cid:2)Ajo (cid:3)ℓk > 0, for some jo > 0
Then a result in matrix theory known as the Perron-Frobenius Theorem [20] leads to the following stronger
characterization of the eigen-structure of A.
(581)
Lemma C.2. (Spectral Norm of Regular Stochastic Matrices) Let A be an N × N right stochastic
and regular matrix. Then:
(a) ρ(A) = 1.
(b) Al l other eigenvalues of A are strictly inside the unit circle (and, hence, have magnitude strictly less
than one).
(c) The eigenvalue at λ = 1 is simple, i.e., it has multiplicity one. Moreover, with proper sign scaling, al l
entries of the corresponding eigenvector are positive. For a right-stochastic A, this eigenvector is the
vector 1 since A1 = 1.
(d) Al l other eigenvectors associated with the other eigenvalues wil l have at least one negative or complex
entry.
Proof. Part (a) follows from Lemma C.1. Parts (b)-(d) follow from the Perron-Frobenius Theorem when A is
regular [20].
Lemma C.3. (Useful Properties of Doubly Stochastic Matrices) Let A be an N ×N doubly stochastic
matrix. Then the fol lowing properties hold:
(a) ρ(A) = 1.
(b) AAT and AT A are doubly stochastic as wel l.
(c) ρ(AAT ) = ρ(AT A) = 1.
(d) The eigenvalues of AAT or AT A are real and lie inside the interval [0, 1].
98
(e) I − AAT ≥ 0 and I − AT A ≥ 0.
(f ) Tr(AT H A) ≤ Tr(H ), for any N × N nonnegative-definite Hermitian matrix H .
Proof. Part (a) follows from Lemma C.1. For part (b), note that AAT is symmetric and AAT 1 = A1 = 1. Therefore,
AAT is doubly stochastic. Likewise for AT A. Part (c) follows from part (a) since AAT and AT A are themselves
doubly stochastic matrices. For part (d), note that AAT is symmetric and nonnegative-definite. Therefore, its
eigenvalues are real and nonnegative. But since ρ(AAT ) = 1, we must have λ(AAT ) ∈ [0, 1]. Likewise for the matrix
AT A. Part (e) follows from part (d). For part (f ), since AAT ≥ 0 and its eigenvalues lie within [0, 1], the matrix
AAT admits an eigen-decomposition of the form:
where U is orthogonal (i.e., U −1 = U T ) and Λ is diagonal with entries in the range [0, 1]. It then follows that
AAT = U ΛU T
Tr(AT H A) = Tr(AAT H )
= Tr(U ΛU T H )
= Tr(ΛU T H U )
(∗)
≤ Tr(U T H U )
= Tr(U U T H )
= Tr(H )
where step (∗) is because U T H U = U −1H U and, by similarity, the matrix U −1H U has the same eigenvalues as H .
Therefore, U T H U ≥ 0. This means that the diagonal entries of U T H U are nonnegative. Multiplying U T H U by Λ
ends up scaling the nonnegative diagonal entries to smaller values so that (∗) is justified.
D Appendix: Block Maximum Norm
Let x = col{x1 , x2 , . . . , xN } denote an N × 1 block column vector whose individual entries are of size M × 1
each. Following [52, 54, 55], the block maximum norm of x is denoted by kxkb,∞ and is defined as
kxkb,∞
∆= max
1≤k≤N
kxk k
(582)
where k · k denotes the Euclidean norm of its vector argument. Correspondingly, the induced block maximum
norm of an arbitrary N × N block matrix A, whose individual block entries are of size M × M each, is defined
as
kAkb,∞
∆= max
x 6=0
kAxkb,∞
kxkb,∞
(583)
The block maximum norm inherits the unitary invariance property of the Euclidean norm, as the following
result indicates [54].
Lemma D.1. (Unitary Invariance) Let U = diag{U1 , U2 , . . . , UN } be an N × N block diagonal matrix
with M × M unitary blocks {Uk }. Then, the fol lowing properties hold:
(a) kU xkb,∞ = kxkb,∞
(b) kU AU ∗kb,∞ = kAkb,∞
for al l block vectors x and block matrices A of appropriate dimensions.
(cid:3)
99
The next result provides useful bounds for the block maximum norm of a block matrix.
Lemma D.2. (Useful Bounds) Let A be an arbitrary N × N block matrix with blocks Aℓk of size M × M
each. Then, the fol lowing results hold:
(a) The norms of A and its complex conjugate are related as fol lows:
kA∗kb,∞ ≤ N · kAkb,∞
(b) The norm of A is bounded as fol lows:
kAℓk k(cid:19)
kAℓk k ≤ kAkb,∞ ≤ N · (cid:18) max
1≤ℓ,k≤N
where k · k denotes the 2−induced norm (or maximum singular value) of its matrix argument.
max
1≤ℓ,k≤N
(584)
(585)
(c) If A is Hermitian and nonnegative-definite (A ≥ 0), then there exist finite positive constants c1 and c2
such that
c1 · Tr(A) ≤ kAkb,∞ ≤ c2 · Tr(A)
(586)
Proof. Part (a) follows directly from part (b) by noting that
kA∗ kb,∞ ≤ N · (cid:18) max
1≤ℓ,k≤N
= N · (cid:18) max
1≤ℓ,k≤N
≤ N · kAkb,∞
ℓk k(cid:19)
kA∗
kAℓk k(cid:19)
where the equality in the second step is because kA∗
ℓk k = kAℓk k; i.e., complex conjugation does not alter the 2−induced
norm of a matrix.
To establish part (b), we consider arbitrary N × 1 block vectors x with entries x = col{x1 , x2 , . . . , xN } and where
each xk is M × 1. Then, note that
Aℓk xk (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
1≤ℓ≤N (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
NXk=1
kAxkb,∞ = max
kAℓk k · kxk k!
1≤ℓ≤N NXk=1
≤ max
kAℓk k! · max
≤ max
NXk=1
kxk k
1≤k≤N
1≤ℓ≤N
kAℓk k! · kxkb,∞
≤ max
NXk=1
max
1≤k≤N
1≤ℓ≤N
kAℓk k(cid:19) · kxkb,∞
≤ N · (cid:18) max
1≤ℓ,k≤N
kAℓk k(cid:19)
≤ N · (cid:18) max
1≤ℓ,k≤N
which establishes the upper bound in (585).
To establish the lower bound, we assume without loss of generality that max1≤ℓ,k≤N kAℓk k is attained at ℓ = 1
and k = 1. Let σ1 denote the largest singular value of A11 and let {v1 , u1 } denote the corresponding M × 1 right and
left singular vectors. That is,
kAxkb,∞
kxkb,∞
so that
kAkb,∞
∆= max
x 6=0
kA11 k = σ1 ,
A11 v1 = σ1 u1
(587)
100
where v1 and u1 have unit norms. We now construct an N × 1 block vector xo as follows:
Then, obviously,
and
It follows that
xo ∆= col{v1 , 0M , 0M , . . . , 0M }
kxo kb,∞ = 1
Axo = col{A11 v1 , A21 v1 , . . . , AN 1 v1 }
kAxo kb,∞ = max { kA11 v1 k, kA21 v1 k, . . . , kAN 1 v1 k }
≥ kA11 v1 k
= kσ1 u1 k
= σ1
= kA11 k
=
max
1≤ℓ,k≤N
kAℓk k
Therefore, by the definition of the block maximum norm,
x 6=0 (cid:18) kAxkb,∞
kxkb,∞ (cid:19)
∆= max
kAxo kb,∞
kxo kb,∞
= kAxo kb,∞
kAℓk k
max
≥
1≤ℓ,k≤N
kAkb,∞
≥
(588)
(589)
(590)
(591)
(592)
which establishes the lower bound in (585).
To establish part (c), we start by recalling that all norms on finite-dimensional vector spaces are equivalent [20,21].
This means that if k · ka and k · kd denote two different matrix norms, then there exist positive constants c1 and c2
such that for any matrix X ,
c1 · kX ka ≤ kX kd ≤ c2 · kX ka
Now, let kAk∗ denote the nuclear norm of the square matrix A. It is defined as the sum of its singular values:
∆= Xm
Since A is Hermitian and nonnegative-definite, its eigenvalues coincide with its singular values and, therefore,
kAk∗ = Xm
Now applying result (593) to the two norms kAkb,∞ and kAk∗ we conclude that
λm (A) = Tr(A)
σm (A)
kAk∗
(593)
(594)
as desired.
c1 · Tr(A) ≤ kAkb,∞ ≤ c2 · Tr(A)
(595)
The next result relates the block maximum norm of an extended matrix to the ∞−norm (i.e., maximum
absolute row sum) of the originating matrix. Specifically, let A be an N × N matrix with bounded entries
and introduce the block matrix
A ∆= A ⊗ IM
(596)
The extended matrix A has blocks of size M × M each.
101
Lemma D.3. (Relation to Maximum Absolute Row Sum) Let A and A be related as in (596). Then,
the fol lowing properties hold:
(a) kAkb,∞ = kAk∞ , where the notation k · k∞ denotes the maximum absolute row sum of its argument.
(b) kA∗kb,∞ ≤ N · kAkb,∞ .
Proof. The results are obvious for a zero matrix A. So assume A is nonzero. Let x = col{x1 , x2 , . . . , xN } denote an
arbitrary N × 1 block vector whose individual entries {xk } are vectors of size M × 1 each. Then,
1≤k≤N (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
akℓxℓ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
NXℓ=1
kAxkb,∞ = max
akℓ · kxℓ k!
1≤k≤N NXℓ=1
≤ max
akℓ ! · max
≤ max
NXℓ=1
1≤ℓ≤N
1≤k≤N
= kAk∞ · kxkb,∞
(597)
kxℓ k
so that
kAkb,∞
∆= max
x 6=0
kAxkb,∞
kxkb,∞
≤ kAk∞
(598)
The argument so far establishes that kAkb,∞ ≤ kAk∞ . Now, let ko denote the row index that corresponds to the
maximum absolute row sum of A, i.e.,
NXℓ=1
We construct an N × 1 block vector z = col{z1 , z2 , . . . , zN }, whose M × 1 entries {zℓ } are chosen as follows:
kAk∞ =
ako ℓ
where e1 is the M × 1 basis vector:
zℓ = sign(ako ℓ ) · e1
and the sign function is defined as
e1 = col{1, 0, 0, . . . , 0}
sign(a) = (cid:26) 1,
−1,
Then, note that z 6= 0 for any nonzero matrix A, and
if a ≥ 0
otherwise
kzkb,∞ = max
1≤ℓ≤N
kzℓ k = 1
102
Moreover,
kAkb,∞
≥
∆= max
x 6=0
kAxkb,∞
kxkb,∞
kAzkb,∞
kzkb,∞
= kAzkb,∞
1≤k≤N (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
akℓ zℓ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
NXℓ=1
max
=
≥ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ako ℓ zℓ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
NXℓ=1
= (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
ako ℓ · sign(ako ℓ )e1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
NXℓ=1
NXℓ=1
ako ℓ · ke1k
=
NXℓ=1
ako ℓ
=
= kAk∞
Combining this result with (598) we conclude that kAkb,∞ = kAk∞ , which establishes part (a). Part (b) follows from
the statement of part (a) in Lemma D.2.
(599)
The next result establishes a useful property for the block maximum norm of right or left stochastic matrices;
such matrices arise as combination matrices for distributed processing over networks as in (166) and (185).
Lemma D.4. (Right and Left Stochastic Matrices) Let C be an N × N right stochastic matrix, i.e.,
its entries are nonnegative and it satisfies C 1 = 1. Let A be an N × N left stochastic matrix, i.e., its entries
are nonnegative and it satisfies AT 1 = 1. Introduce the block matrices
The matrices A and C have blocks of size M × M each. It holds that
AT ∆= AT ⊗ IM ,
C ∆= C ⊗ IM
kAT kb,∞ = 1,
kC kb,∞ = 1
(600)
(601)
Proof. Since AT and C are right stochastic matrices, it holds that kAT k∞ = 1 and kC k∞ = 1. The desired result
then follows from part (a) of Lemma D.3.
The next two results establish useful properties for the block maximum norm of a block diagonal matrix
transformed by stochastic matrices; such transformations arise as coefficient matrices that control the evo-
lution of weight error vectors over networks, as in (189).
Lemma D.5. (Block Diagonal Hermitian Matrices) Consider an N × N block diagonal Hermitian
matrix D = diag{D1 , D2 , . . . , DN }, where each Dk is M × M Hermitian. It holds that
ρ(D) = max
1≤k≤N
ρ(Dk ) = kDkb,∞
(602)
103
where ρ(·) denotes the spectral radius (largest eigenvalue magnitude) of its argument. That is, the spectral
radius of D agrees with the block maximum norm of D, which in turn agrees with the largest spectral radius
of its block components.
Proof. We already know that the spectral radius of any matrix X satisfies ρ(X ) ≤ kX k, for any induced matrix
norm [19, 20]. Applying this result to D we readily get that ρ(D) ≤ kDkb,∞ . We now establish the reverse inequality,
namely, kDkb,∞ ≤ ρ(D). Thus, pick an arbitrary N × 1 block vector x with entries {x1 , x2 , . . . , xN }, where each xk
is M × 1. From definition (583) we have
kDkb,∞
· max
1≤k≤N
kDxkb,∞
∆= max
kxkb,∞
x 6=0
x 6=0 (cid:18)
kDk xk k (cid:19)
1
= max
kxkb,∞
x 6=0 (cid:18)
(kDk k · kxk k) (cid:19)
1
≤ max
· max
kxkb,∞
1≤k≤N
kxkb,∞ (cid:19)
1≤k≤N (cid:18)kDk k ·
kxk k
max
= max
x 6=0
max
kDk k
1≤k≤N
≤
where the notation kDk k denotes the 2−induced norm of Dk (i.e., its largest singular value). But since Dk is assumed
to be Hermitian, its 2−induced norm agrees with its spectral radius, which explains the last equality.
=
max
1≤k≤N
ρ(Dk )
(603)
Lemma D.6. (Block Diagonal Matrix Transformed by Left Stochastic Matrices) Consider an
N × N block diagonal Hermitian matrix D = diag{D1 , D2 , . . . , DN }, where each Dk is M × M Hermitian.
Let A1 and A2 be N × N left stochastic matrices, i.e., their entries are nonnegative and they satisfy AT
1 = 1
1
and AT
1 = 1. Introduce the block matrices
2
AT
1 = AT
1 ⊗ IM ,
AT
2
∆= AT
2 ⊗ IM
(604)
The matrices A1 and A2 have blocks of size M × M each. Then it holds that
1 (cid:1) ≤ ρ(D)
ρ (cid:0)AT
2 · D · AT
Proof. Since the spectral radius of any matrix never exceeds any induced norm of the same matrix, we have that
1 (cid:13)(cid:13)(cid:13)b,∞
(cid:13)(cid:13)(cid:13) AT
ρ (cid:16)AT
1 (cid:17)
2 · D · AT
2 · D · AT
1 (cid:13)(cid:13)(cid:13)b,∞
2 (cid:13)(cid:13)(cid:13)b,∞
· kDkb,∞ · (cid:13)(cid:13)(cid:13)AT
(cid:13)(cid:13)(cid:13)AT
kDkb,∞
ρ(D)
(601)
=
(602)
=
≤
≤
(605)
(606)
In view of the result of Lemma D.5, we also conclude from (605) that
1 (cid:1) ≤ max
ρ (cid:0)AT
2 · D · AT
1≤k≤N
ρ(Dk )
(607)
104
It is worth noting that there are choices for the matrices {A1 , A2 , D} that would result in strict inequality
in (605). Indeed, consider the special case:
D = (cid:20) 2 0
3 (cid:21)
2 = (cid:20) 1
3 (cid:21) , AT
1 = (cid:20) 1
0 1 (cid:21) , AT
2
2
3
3
3
3
1
1
2
2
3
3
This case corresponds to N = 2 and M = 1 (scalar blocks). Then,
1 = (cid:20) 2
1 (cid:21)
2
AT
2 DAT
3
3
2
3
ρ(AT
2 DAT
1 ) ≈ 1.52
The following conclusions follow as corollaries to the statement of Lemma D.6, where by a stable matrix X
we mean one whose eigenvalues lie strictly inside the unit circle.
and it is easy to verify that
ρ(D) = 2,
Corol lary D.1. (Stability Properties) Under the same setting of Lemma D.6, the fol lowing conclusions
hold:
2 DAT
(a) The matrix AT
1 is stable whenever D is stable.
2 DAT
(b) The matrix AT
1 is stable for al l possible choices of left stochastic matrices A1 and A2 if, and only
if, D is stable.
Proof. Since D is block diagonal, part (a) follows immediately from (605) by noting that ρ(D) < 1 whenever D is
stable.
[This statement fixes the argument that appeared in App. I of [18] and Lemma 2 of [33]. Since the matrix
X in App. I of [18] and the matrix M in Lemma 2 of [33] are block diagonal, the k · kb,∞ norm should replace the
k · kρ norm used there, as in the proof that led to (606) and as already done in [54].] For part (b), assume first that
2 DAT
D is stable, then AT
1 will also be stable by part (a) for any left-stochastic matrices A1 and A2 . To prove the
converse, assume that AT
2 DAT
1 is stable for any choice of left stochastic matrices A1 and A2 . Then, AT
2 DAT
1 is stable
for the particular choice A1 = I = A2 and it follows that D must be stable.
E Appendix: Comparison with Consensus Strategies
Consider a connected network consisting of N nodes. Each node has a state or measurement value xk ,
possibly a vector of size M × 1. All nodes in the network are interested in evaluating the average value of
their states, which we denote by
NXk=1
A centralized solution to this problem would require each node to transmit its measurement xk to a fusion
center. The central processor would then compute wo using (608) and transmit it back to all nodes. This
centralized mode of operation suffers from at least two limitations. First, it requires communications and
power resources to transmit the data back and forth between the nodes and the central processor; this
problem is compounded if the fusion center is stationed at a remote location. Second, the architecture has a
critical point of failure represented by the central processor; if it fails, then operations would need to be halted.
wo ∆=
(608)
1
N
xk
Consensus Recursion
The consensus strategy provides an elegant distributed solution to the same problem, whereby nodes interact
locally with their neighbors and are able to converge to wo through these interactions. Thus, consider an
105
arbitrary node k and assign nonnegative weights {aℓk } to the edges linking k to its neighbors ℓ ∈ Nk . For
each node k , the weights {aℓk } are assumed to add up to one so that
for k = 1, 2, . . . , N :
aℓk = 1,
aℓk ≥ 0,
aℓk = 0 if ℓ /∈ Nk
NXℓ=1
The resulting combination matrix is denoted by A and its k−th column consists of the entries {aℓk , ℓ =
1, 2, . . . , N }. In view of (609), the combination matrix A is seen to satisfy AT 1 = 1. That is, A is left-
stochastic. The consensus strategy can be described as follows. Each node k operates repeatedly on the data
from its neighbors and updates its state iteratively according to the rule:
wk,n = Xℓ∈Nk
where wℓ,n−1 denotes the state of node ℓ at iteration n − 1, and wk,n denotes the updated state of node k
after iteration n. The initial conditions are
aℓk wℓ,n−1 , n > 0
(609)
(610)
If we collect the states of all nodes at iteration n into a column vector, say,
wk,o = xk ,
k = 1, 2, . . . , N
zn
∆= col{w1,n , w2,n , . . . , wN ,n}
Then, the consensus iteration (610) can be equivalently rewritten in vector form as follows:
zn = AT zn−1 , n > 0
AT = AT ⊗ IM
zo
∆= col{x1 , x2 , . . . , xN }
where
The initial condition is
Error Recursion
Note that we can express the average value, wo , from (608) in the form
wo =
1
N
· (1T ⊗ IM ) · zo
(611)
(612)
(613)
(614)
(615)
(616)
where 1 is the vector of size M × 1 and whose entries are all equal to one. Let
ewk,n = wo − wk,n
denote the weight error vector for node k at iteration n; it measures how far the iterated state is from the
desired average value wo . We collect all error vectors across the network into an N × 1 block column vector
whose entries are of size M × 1 each:
∆=
ew1,n
ew2,n
...
ewN ,n
106
ewn
(618)
(617)
Then,
ewn = (1 ⊗ IM )wo − zn
Convergence Conditions
The following result is a classical result on consensus strategies [42–44]. It provides conditions under which
the state of all nodes will converge to the desired average, wo , so that ewn will tend to zero.
Theorem E.1. (Convergence to Consensus) For any initial states {xk }, the successive iterates wk,n
generated by the consensus iteration (610) converge to the network average value wo as n → ∞ if, and only
if, the fol lowing three conditions are met:
(619)
AT 1 = 1
A1 = 1
11T (cid:19) < 1
ρ (cid:18)AT −
That is, the combination matrix A needs to be doubly stochastic, and the matrix AT − 1
N
stable.
1
N
(620)
(621)
(622)
11T needs to be
Proof. (Sufficiency). Assume first that the three conditions stated in the theorem hold. Since A is doubly stochastic,
then so is any power of A, say, An for any n ≥ 0, so that
[An ]T 1 = 1,
An1 = 1
Using this fact, it is straightforward to verify by induction the validity of the following equality:
11T (cid:19)n
(cid:18)AT −
Likewise, using the Kronecker product identities
= [An ]T −
1
N
1
N
11T
(E + B ) ⊗ C = (E ⊗ C ) + (B ⊗ C )
(E ⊗ B )(C ⊗ D) = (EC ⊗ BD)
(E ⊗ B )n = E n ⊗ B n
1
N
(An )T −
· (1 ⊗ IM ) · (1T ⊗ IM )
for matrices {E , B , C, D} of compatible dimensions, we observe that
h(An )T ⊗ IM i −
1
· (11T ⊗ IM )
N
· 11T (cid:21) ⊗ IM
(cid:20)(An )T −
1
N
(cid:18)AT −
11T (cid:19)n
1
⊗ IM
N
11T (cid:19) ⊗ IM (cid:21)n
(cid:20)(cid:18)AT −
1
N
zn = [An ]T zo
Iterating (613) we find that
(624)
=
=
=
=
and, hence, from (616) and (619),
ewn
=
(628)
=
− (cid:20)(An )T −
− (cid:20)(cid:18)AT −
1
N
· (1 ⊗ IM ) · (1T ⊗ IM )(cid:21) · zo
1
N
11T (cid:19) ⊗ IM (cid:21)n
· zo
107
(623)
(624)
(625)
(626)
(627)
(628)
(629)
(630)
Now recall that, for two arbitrary matrices C and D of compatible dimensions, the eigenvalues of the Kronecker
product C ⊗ D is formed of all product combinations λi (C )λj (D) of the eigenvalues of C and D [19]. We conclude
from this property, and from the fact that AT − 1
11T is stable, that the coefficient matrix
N
(cid:18)AT −
· 11T (cid:19) ⊗ IM
ewn → 0 as n → ∞
(Necessity). In order for zn in (629) to converge to (1 ⊗ IM )wo , for any initial state zo , it must hold that
1
(An )T · zo =
· (1 ⊗ IM ) · (1T ⊗ IM ) · zo
N
is also stable. Therefore,
lim
n→∞
(631)
(632)
1
N
for any zo . This implies that we must have
or, equivalently,
This in turn implies that we must have
lim
n→∞
(An )T =
1
N
· (11T ⊗ IM )
lim
n→∞
(An )T =
11T
1
N
(633)
(634)
11T
1
N
(An )T
lim
n→∞
But since
= lim
n→∞
AT · (An )T = AT ·
n→∞ (cid:0)An+1 (cid:1)T
AT · (An )T = lim
lim
n→∞
we conclude from (634) and (635) that it must hold that
1
1
11T =
AT · 11T
N
N
N (cid:16)AT 1 − 1(cid:17) · 1T = 0
1
from which we conclude that we must have AT 1 = 1. Similarly, we can show that A1 = 1 by studying the limit
of (An )T AT . Therefore, A must be a doubly stochastic matrix. Now using the fact that A is doubly stochastic, we
know that (624) holds. It follows that in order for condition (634) to be satisfied, we must have
ρ (cid:18)AT −
11T (cid:19) < 1
That is,
(639)
(635)
(636)
(637)
(638)
1
N
Rate of Convergence
From (630) we conclude that the rate of convergence of the error vectors { ewk,n} to zero is determined by
the spectrum of the matrix
1
11T
AT −
(640)
N
Now since A is a doubly stochastic matrix, we know that it has an eigenvalue at λ = 1. Let us denote the
eigenvalues of A by λk (A) and let us order them in terms of their magnitudes as follows:
where λ1 (A) = 1. Then, the eigenvalues of the coefficient matrix (AT − 1
N
11T ) are equal to
0 ≤ λM (A) ≤ . . . ≤ λ3 (A) ≤ λ2 (A) ≤ 1
{ λM (A), , . . . , λ3 (A), λ2 (A), 0 }
(641)
(642)
It follows that the magnitude of λ2 (A) becomes the spectral radius of AT − 1
N
ensures that λ2 (A) < 1. We therefore arrive at the following conclusion.
11T . Then condition (639)
108
Corol lary E.1. (Rate of Convergence of Consensus) Under conditions (620)–(622), the rate of con-
vergence of the successive iterates {wk,n } towards the network average wo in the consensus strategy (610) is
determined by the second largest eigenvalue magnitude of A, i.e., by λ2 (A) as defined in (641).
It is worth noting that doubly stochastic matrices A that are also regular satisfy conditions (620)–(622).
This is because, as we already know from Lemma C.2, the eigenvalues of such matrices satisfy λm (A) < 1,
for m = 2, 3, . . . , N , so that condition (622) is automatically satisfied.
(cid:3)
Corol lary E.2. (Convergence for Regular Combination Matrices) Any doubly-stochastic and regular
matrix A satisfies the three conditions (620)–(622) and, therefore, ensures the convergence of the consensus
iterates {wk,n} generated by (610) towards wo as n → ∞.
(cid:3)
A regular combination matrix A would result when the two conditions listed below are satisfied by the graph
connecting the nodes over which the consensus iteration is applied.
Corol lary E.3. (Sufficient Condition for Regularity) Assume the combination matrix A is doubly
stochastic and that the graph over which the consensus iteration (610) is applied satisfies the fol lowing two
conditions:
(a) The graph is connected. This means that there exists a path connecting any two arbitrary nodes in the
network. In terms of the Laplacian matrix that is associated with the graph (see Lemma B.1), this
means that the second smal lest eigenvalue of the Laplacian is nonzero.
(b) aℓk = 0 if, and only if, ℓ /∈ Nk . That is, the combination weights are strictly positive between any two
neighbors, including akk > 0.
Then, the corresponding matrix A wil l be regular and, therefore, the consensus iterates {wk,n } generated by
(610) wil l converge towards wo as n → ∞.
Proof. We first establish that conditions (a) and (b) imply that A is a regular matrix, namely, that there should
exist an integer jo > 0 such that
hAjo iℓk
for all (ℓ, k). To begin with, by the rules of matrix multiplication, the (ℓ, k) entry of the i−th power of A is given by:
hAi iℓk
NXmi−1 =1
NXm2 =1
NXm1 =1
The summand in (644) is nonzero if, and only if, there is some sequence of indices (ℓ, m1 , . . . , mi−1 , k) that forms
a path from node ℓ to node k. Since the network is assumed to be connected, there exists a minimum (and finite)
integer value iℓk such that a path exists from node ℓ to node k using iℓk edges and that
hAiℓk iℓk
aℓm1 am1m2 . . . ami−1 k
(643)
=
. . .
> 0
> 0
(644)
In addition, by induction, if (cid:2)Aiℓk (cid:3)ℓk > 0, then
hAiℓk+1 iℓk
amk
NXm=1 hAiℓk iℓm
=
≥ hAiℓk iℓk
akk
> 0
Let
jo = max
1≤k,ℓ≤N
{iℓk }
109
Then, property (643) holds for all (ℓ, k). And we conclude from (581) that A is a regular matrix. It then follows from
Corollary E.2 that the consensus iterates {wk,n} converge to the average network value wo .
Comparison with Diffusion Strategies
Observe that in comparison to diffusion strategies, such as the ATC strategy (153), the consensus iteration
(610) employs the same quantities wk,· on both sides of the iteration. In other words, the consensus con-
struction keeps iterating on the same set of vectors until they converge to the average value wo . Moreover,
the index n in the consensus algorithm is an iteration index. In contrast, diffusion strategies employ different
quantities on both sides of the combination step in (153), namely, wk,i and {ψℓ,i}; the latter variables have
been processed through an information exchange step and are updated (or filtered) versions of the wℓ,i−1 . In
addition, each step of the diffusion strategy (153) can incorporate new data, {dℓ(i), uℓ,i}, that are collected
by the nodes at every time instant. Moreover, the index i in the diffusion implementation is a time index
(and not an iteration index); this is because diffusion strategies are inherently adaptive and perform online
learning. Data keeps streaming in and diffusion incorporates the new data into the update equations at
every time instant. As a result, diffusion strategies are able to respond to data in an adaptive manner, and
they are also able to solve general optimization problems: the vector wo in adaptive diffusion iterations is
the minimizer of a global cost function (cf. (92)), while the vector wo in consensus iterations is the average
value of the initial states of the nodes (cf. (608)).
Moreover, it turns out that diffusion strategies influence the evolution of the network dynamics in an
interesting and advantageous manner in comparison to consensus strategies. We illustrate this point by
means of an example. Consider initially the ATC strategy (158) without information exchange, whose
update equation we repeat below for ease of reference:
(645)
(646)
(ATC diffusion)
(647)
(ATC diffusion )
ψk,i = wk,i−1 + µk u∗
k,i [dk (i) − uk,iwk,i−1 ]
wk,i = Xℓ∈Nk
aℓk ψ ℓ,i
These recursions were derived in the body of the article as an effective distributed solution for optimizing
(92)–(93). Note that they involve two steps, where the weight estimator wk,i−1 is first updated to the
intermediate estimator ψk,i , before the intermediate estimators from across the neighborhood are combined
to obtain wk,i . Both steps of ATC diffusion (645)–(646) can be combined into a single update as follows:
wk,i = Xℓ∈Nk
aℓk (cid:2)wℓ,i−1 + µℓu∗
ℓ,i (dℓ (i) − uℓ,iwℓ,i−1 )(cid:3)
Likewise, consider the CTA strategy (159) without information exchange, whose update equation we also
repeat below:
ψk,i−1 = Xℓ∈Nk
aℓk w ℓ,i−1
k,i (cid:2)dk (i) − uk,iψk,i−1 (cid:3)
wk,i = ψk,i−1 + µk u∗
Again, the CTA strategy involves two steps: the weight estimators {wℓ,i−1} from the neighborhood of node
k are first combined to yield the intermediate estimator ψk,i−1 , which is subsequently updated to wk,i . Both
steps of CTA diffusion can also be combined into a single update as follows:
aℓkw ℓ,i−1#
k,i "dk (i) − uk,i Xℓ∈Nk
aℓk wℓ,i−1 + µk u∗
110
wk,i = Xℓ∈Nk
(CTA diffusion )
(CTA diffusion)
(650)
(648)
(649)
(consensus strategy)
(651)
aℓk wℓ,i−1 + µku∗
k,i [dk (i) − uk,iwk,i−1 ]
Now, motivated by the consensus iteration (610), and based on a procedure for distributed optimization
suggested in [52] (see expression (7.1) in that reference), some works in the literature (e.g., [45, 53, 82–88])
considered distributed strategies that correspond to the following form for the optimization problem under
consideration (see, e.g., expression (1.20) in [53] and expression (9) in [87]):
wk,i = Xℓ∈Nk
This strategy can be derived by following the same argument we employed earlier in Secs. 3.2 and 4 to
arrive at the diffusion strategies, namely, we replace wo in (127) by wℓ,i−1 and then apply the instantaneous
approximations (150). Note that the same variable wk,· appears on both sides of the equality in (651).
Thus, compared with the ATC diffusion strategy (647), the update from wk,i−1 to wk,i in the consensus
implementation (651) is only influenced by data {dk (i), uk,i } from node k . In contrast, the ATC diffusion
structure (645)–(646) helps incorporate the influence of the data {dℓ (i), uℓ,i} from across the neighborhood
of node k into the update of wk,i , since these data are reflected in the intermediate estimators {ψℓ,i}.
Likewise, the contrast with the CTA diffusion strategy (650) is clear, where the right-most term in (650)
relies on a combination of all estimators from across the neighborhood of node k , and not only on wk,i−1 as
in the consensus strategy (651). These facts have desirable implications on the evolution of the weight-error
vectors across diffusion networks. Some simple algebra, similar to what we did in Sec. 6, will show that the
mean of the extended error vector for the consensus strategy (651) evolves according to the recursion:
E ew i = (cid:0)AT − MRu(cid:1) · E ew i−1 ,
where Ru is the block diagonal covariance matrix defined by (184) and ew i is the aggregate error vector
defined by (230). We can compare the above mean error dynamics with the ones that correspond to the
ATC and CTA diffusion strategies (645)–(646) and (648)–(650); their error dynamics follow as special cases
from (248) by setting A1 = I = C and A2 = A for ATC and A2 = I = C and A1 = A for CTA:
E ewi = AT (IN M − MRu ) · E ew i−1 ,
E ewi = (IN M − MRu ) AT · E ew i−1 ,
i ≥ 0
(CTA diffusion)
(654)
We observe that the coefficient matrices that control the evolution of E ew i are different in all three cases. In
particular,
consensus strategy (652) is stable in the mean ⇐⇒ ρ (cid:0)AT − MRu(cid:1) < 1
(655)
ATC diffusion (653) is stable in the mean ⇐⇒ ρ (cid:2)AT (IN M − MRu )(cid:3) < 1
(656)
CTA diffusion (654) is stable in the mean ⇐⇒ ρ (cid:2)(IN M − MRu ) AT (cid:3) < 1
It follows that the mean stability of the consensus network is sensitive to the choice of the combination
matrix A. This is not the case for the diffusion strategies. This is because from property (605) established in
App. D, we know that the matrices AT (IN M − MRu ) and (IN M − MRu ) AT are stable if (IN M − MRu )
is stable. Therefore, we can select the step-sizes to satisfy µk < 2/λmax(Ru,k ) for the ATC or CTA diffusion
strategies and ensure their mean stability regardless of the combination matrix A. This also means that
the diffusion networks will be mean stable whenever the individual nodes are mean stable, regardless of the
topology defined by A. In contrast, for consensus networks, the network can exhibit unstable mean behavior
(consensus strategy)
(657)
i ≥ 0
(ATC diffusion)
(653)
i ≥ 0
(652)
and
111
even if all its individual nodes are stable in the mean. For further details and other results on the mean-
square performance of diffusion networks in relation to consensus networks, the reader is referred to [89, 90].
Acknowledgement. The development of the theory and applications of diffusion adaptation over networks
has benefited greatly from the insights and contributions of several UCLA Ph.D. students, and several visiting
graduate students to the UCLA Adaptive Systems Laboratory (http://www.ee.ucla.edu/asl). The assistance
and contributions of all students are hereby gratefully acknowledged, including Cassio G. Lopes, Federico S.
Cattivelli, Sheng-Yuan Tu, Jianshu Chen, Xiaochuan Zhao, Zaid Towfic, Chung-Kai Yu, Noriyuki Takahashi,
Jae-Woo Lee, Alexander Bertrand, and Paolo Di Lorenzo. The author is also particularly thankful to S.-Y.
Tu, J. Chen, X. Zhao, Z. Towfic, and C.-K. Yu for their assistance in reviewing an earlier draft of this chapter.
References
[1] J. Chen and A. H. Sayed, “On the limiting behavior of distributed optimization strategies,” Proc. 50th Annual
Al lerton Conference on Communication, Control, and Computing, pp. 1–8, Allerton, IL, October 2012.
[2] J. Chen and A. H. Sayed, “Diffusion adaptation strategies for distributed optimization and learning over net-
works,” IEEE Trans. Signal Processing, vol. 60, no. 8, pp. 4289–4305, August 2012.
[3] J. Chen and A. H. Sayed, “Distributed Pareto optimization via diffusion strategies,” IEEE J. Selected Topics in
Signal Processing, vol. 7, no. 2, pp. 205–220, April 2013.
[4] A. H. Sayed, Fundamentals of Adaptive Filtering, Wiley, NJ, 2003.
[5] A. H. Sayed, Adaptive Filters, Wiley, NJ, 2008.
[6] S. Haykin, Adaptive Filter Theory, 4th edition, Prentice Hall, NJ, 2002.
[7] B. Widrow and S. D. Stearns, Adaptive Signal Processing, Prentice Hall, NJ, 1985.
[8] S.-Y. Tu and A. H. Sayed, “Mobile adaptive networks,” IEEE J. Sel. Topics. Signal Process., vol. 5, no. 4,
pp. 649–664, Aug. 2011.
[9] F. Cattivelli and A. H. Sayed, “Modeling bird flight formations using diffusion adaptation,” IEEE Transactions
on Signal Processing, vol. 59, no. 5, pp. 2038–2051, May 2011.
[10] J. Li and A. H. Sayed, “Modeling bee swarming behavior through diffusion adaptation with asymmetric infor-
mation sharing,” EURASIP Journal on Advances in Signal Processing, 2012:18, doi:10.1186/1687-6180-2012-18,
2012.
[11] J. Chen and A. H. Sayed, “Bio-inspired cooperative optimization with application to bacteria motility,” Proc.
ICASSP, Prague, Czech Republic, pp. 5788–5791, May 2011.
[12] A. H. Sayed and F. A. Sayed, “Diffusion adaptation over networks of particles sub ject to Brownian fluctuations,”
Proc. Asilomar Conference on Signals, Systems, and Computers, pp. 685–690, Pacific Grove, CA, November 2011.
[13] J. Mitola and G. Q. Maguire, “Cognitive radio: Making software radios more personal,”IEEE Personal Commun.,
vol. 6, pp. 13-18, 1999.
[14] S. Haykin, “Cognitive radio: Brain-empowered wireless communications,” IEEE J. Sel. Areas Commun., vol.
23, no. 2, pp. 201-220, Feb. 2005.
[15] Z. Quan, W. Zhang, S. J. Shellhammer, and A. H. Sayed, “Optimal spectral feature detection for spectrum
sensing at very low SNR,” IEEE Transactions on Communications, vol. 59, no. 1, pp. 201-212, January 2011.
[16] Q. Zou, S. Zheng, and A. H. Sayed, “Cooperative sensing via sequential detection,” IEEE Transactions on Signal
Processing, vol. 58, no. 12, pp. 6266-6283, December 2010.
[17] P. Di Lorenzo, S. Barbarossa, and A. H. Sayed, “Bio-inspired swarming for dynamic radio access based on
diffusion adaptation,” Proc. EUSIPCO, pp. 402-406, Barcelona, Spain, August 2011.
[18] F. S. Cattivelli and A. H. Sayed, “Diffusion LMS strategies for distributed estimation,” IEEE Trans. Signal
Process., vol. 58, no. 3, pp. 1035–1048, March 2010.
[19] Golub, G. H. and C. F. Van Loan (1996), Matrix Computations, 3rd edition, The John Hopkins University Press,
Baltimore.
112
[20] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University Press, 2003.
[21] E. Kreyszig, Introductory Functional Analysis with Applications, Wiley, NY, 1989.
[22] B. Poljak, Introduction to Optimization, Optimization Software, NY, 1987.
[23] D. P. Bertsekas, “A new class of incremental gradient methods for least squares problems,” SIAM J. Optim.,
vol. 7, no. 4, pp. 913–926, 1997.
[24] A. Nedic and D. P. Bertsekas, “Incremental subgradient methods for nondifferentiable optimization,” SIAM J.
Optim., vol. 12, no. 1, pp. 109–138, 2001.
[25] M. G. Rabbat and R. D. Nowak, “Quantized incremental algorithms for distributed optimization,” IEEE J. Sel.
Areas Commun., vol. 23, no. 4, pp. 798–808, 2005.
[26] C. G. Lopes and A. H. Sayed, “Incremental adaptive strategies over distributed networks,” IEEE Trans. Signal
Process., vol. 55, no. 8, pp. 4064–4077, Aug. 2007.
[27] F. S. Cattivelli and A. H. Sayed, “Diffusion LMS algorithms with information exchange,” Proc. Asilomar Conf.
Signals, Syst. Comput., Pacific Grove, CA, pp. 251–255, Nov. 2008.
[28] F. S. Cattivelli, C. G. Lopes, and A. H. Sayed, “A diffusion RLS scheme for distributed estimation over adaptive
networks,” Proc. IEEE Workshop on Signal Process. Advances Wireless Comm. (SPAWC), Helsinki, Finland,
pp. 1–5, June 2007.
[29] F. S. Cattivelli, C. G. Lopes, and A. H. Sayed, “Diffusion recursive least-squares for distributed estimation over
adaptive networks,” IEEE Trans. Signal Process., vol. 56, no. 5, pp. 1865–1877, May 2008.
[30] F. S. Cattivelli, C. G. Lopes, and A. H. Sayed, “Diffusion strategies for distributed Kalman filtering: Formulation
and performance analysis,” Proc. IAPR Workshop on Cognitive Inf. Process.(CIP), Santorini, Greece, pp. 36–41,
June 2008.
[31] F. S. Cattivelli and A. H. Sayed, “Diffusion mechanisms for fixed-point distributed Kalman smoothing,” Proc.
EUSIPCO, Lausanne, Switzerland, pp. 1–4, Aug. 2008.
[32] A. H. Sayed and F. Cattivelli, “Distributed adaptive learning mechanisms,” Handbook on Array Processing and
Sensor Networks, S. Haykin and K. J. Ray Liu, Eds., pp. 695–722, Wiley, NJ, 2009.
[33] F. Cattivelli and A. H. Sayed, “Diffusion strategies for distributed Kalman filtering and smoothing,” IEEE
Transactions on Automatic Control, vol. 55, no. 9, pp. 2069–2084, Sep. 2010.
[34] S. S. Stankovic, M. S. Stankovic, and D. S. Stipanovic, “Decentralized parameter estimation by consensus based
stochastic approximation,” IEEE Trans. on Autom. Control, vol. 56, no. 3, pp. 531–543, Mar. 2011.
[35] C. G. Lopes and A. H. Sayed, “Distributed processing over adaptive networks,” in Proc. Adaptive Sensor Array
Processing Workshop, MIT Lincoln Laboratory, MA, pp.1–5, June 2006.
[36] A. H. Sayed and C. G. Lopes, “Adaptive processing over distributed networks,” IEICE Trans. Fund. of Electron.,
Commun. and Comput. Sci., vol. E90-A, no. 8, pp. 1504–1510, 2007.
[37] C. G. Lopes and A. H. Sayed, “Diffusion least-mean-squares over adaptive networks,” Proc. IEEE ICASSP,
Honolulu, Hawaii, vol. 3, pp. 917-920, April 2007.
[38] C. G. Lopes and A. H. Sayed, “Steady-state performance of adaptive diffusion least-mean squares,” Proc. IEEE
Workshop on Statistical Signal Processing (SSP), pp. 136-140, Madison, WI, August 2007.
[39] C. G. Lopes and A. H. Sayed, “Diffusion least-mean squares over adaptive networks: Formulation and perfor-
mance analysis,” IEEE Trans. Signal Process., vol. 56, no. 7, pp. 3122–3136, July 2008.
[40] S. S. Ram, A. Nedic, and V. V. Veeravalli, “Distributed stochastic subgradient pro jection algorithms for convex
optimization,” J. Optim. Theory Appl., vol. 147, no. 3, pp. 516–545, 2010.
[41] P. Bianchi, G. Fort, W. Hachem, and J. Jakubowicz, “Convergence of a distributed parameter estimator for
sensor networks with local averaging of the estimates,” Proc. IEEE ICASSP, Prague, Czech, pp. 3764–3767, May
2011.
[42] M. H. DeGroot, “Reaching a consensus,” Journal of the American Statistical Association, vol. 69, no. 345, pp.
118–121, 1974.
[43] R. L. Berger, “A necessary and sufficient condition for reaching a consensus using DeGroot’s method,” Journal
of the American Statistical Association, vol. 76, no. 374, pp. 415-418, Jun. 1981.
113
[44] J. Tsitsiklis and M. Athans, “Convergence and asymptotic agreement in distributed decision problems,” IEEE
Trans. Autom. Control, vol. 29, no. 1, pp. 42–50, Jan. 1984.
[45] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups of mobile autonomous agents using nearest
neighbor rules,” IEEE Trans. Autom. Control, vol. 48, no. 6, pp. 988–1001, Jun. 2003.
[46] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of agents with switching topology and
time-delays,” IEEE Trans. Autom. Control, vol. 49, pp. 1520–1533, Sep. 2004.
[47] R. Olfati-Saber, “Distributed Kalman filter with embedded consensus filters,” Proc. 44th IEEE Conf. Decision
Control, pp. 8179–8184, Sevilla, Spain, Dec. 2005.
[48] R. Olfati-Saber, “Distributed Kalman filtering for sensor networks,” Proc. 46th IEEE Conf. Decision Control,
pp. 5492–5498, New Orleans, LA, Dec. 2007.
[49] L. Xiao and S. Boyd, “Fast linear iterations for distributed averaging,” Syst. Control Lett., vol. 53, no. 1, pp.
65–78, Sep. 2004.
[50] L. Xiao, S. Boyd, and S. Lall, “A scheme for robust distributed sensor fusion based on average consensus,” Proc.
IPSN, 2005, pp. 63–70, Los Angeles, CA, April 2005.
[51] U. A. Khan and J. M. F. Moura, “Distributing the Kalman filter for large-scale systems,” IEEE Trans. Signal
Processing, vol. 56, no. 10, pp. 4919–4935, Oct. 2008.
[52] D. P. Bertsekas and J. N. Tsitsiklis, Paral lel and Distributed Computation: Numerical Methods, 1st edition,
Athena Scientific, Singapore, 1997.
[53] A. Nedic and A. Ozdaglar, “Cooperative distributed multi-agent optimization,” in Convex Optimization in Signal
Processing and Communications, Y. Eldar and D. Palomar (Eds.), Cambridge University Press, pp. 340-386, 2010.
[54] N. Takahashi, I. Yamada, and A. H. Sayed, “Diffusion least-mean-squares with adaptive combiners: Formulation
and performance analysis,” IEEE Trans. on Signal Processing, vol. 9, pp. 4795-4810, Sep. 2010.
[55] N. Takahashi and I. Yamada, “Parallel algorithms for variational inequalities over the cartesian product of the
intersections of the fixed point sets of nonexpansive mappings,” J. Approx. Theory, vol. 153, no. 2, pp. 139–160,
Aug. 2008.
[56] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004.
[57] T. Y. Al-Naffouri and A. H. Sayed, “Transient analysis of data-normalized adaptive filters,”IEEE Transactions
on Signal Processing, vol. 51, no. 3, pp. 639–652, March 2003.
[58] S-Y. Tu and A. H. Sayed, “Optimal combination rules for adaptation and learning over networks,” Proc. IEEE
International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP), San Juan,
Puerto Rico, pp. 317–320, December 2011.
[59] R. Abdolee and B. Champagne, “Diffusion LMS algorithms for sensor networks over non-ideal inter-sensor
wireless channels,” Proc. IEEE Int. Conf. Dist. Comput. Sensor Systems (DCOSS), pp. 1–6, Barcelona, Spain,
June 2011.
[60] A. Khalili, M. A. Tinati, A. Rastegarnia, and J. A. Chambers, “Steady state analysis of diffusion LMS adaptive
networks with noisy links,” IEEE Trans. Signal Processing, vol. 60, no. 2, pp. 974–979, Feb. 2012.
[61] S-Y. Tu and A. H. Sayed, “Adaptive networks with noisy links,” Proc. IEEE Globecom, pp. 1–5, Houston, TX,
December 2011.
[62] X. Zhao, S-Y. Tu, and A. H. Sayed, “Diffusion adaptation over networks under imperfect information exchange
and non-stationary data,” IEEE Transactions on Signal Processing, vol. 60, no. 7, pp. 3460–3475, July 2012.
[63] X. Zhao and A. H. Sayed, “Combination weights for diffusion strategies with imperfect information exchange,”
Proc. IEEE ICC, pp. 1–5, Ottawa, Canada, June 2012.
[64] D. M. Cvetkovi´c, M. Doob, and H. Sachs, Spectra of Graphs: Theory and Applications, Wiley, NY, 1998.
[65] B. Bollobas, Modern Graph Theory, Springer, 1998.
[66] W. Kocay and D. L. Kreher, Graphs, Algorithms and Optimization, Chapman & Hall/CRC Press, Boca Raton,
2005.
[67] M. Fiedler, “Algebraic connectivity of graphs,” Czech. Math. J., vol. 23, pp. 298–305, 1973.
114
[68] V. D. Blondel, J. M. Hendrickx, A. Olshevsky, and J. N. Tsitsiklis, “Convergence in multiagent coordination,
consensus, and flocking,” Proc. Joint 44th IEEE Conf. on Decision and Control and European Control Conf.
(CDC-ECC), pp. 2996-3000, Seville, Spain, Dec. 2005.
[69] D. S. Scherber and H. C. Papadopoulos, “Locally constructed algorithms for distributed computations in ad-hoc
networks,” Proc. Information Processing in Sensor Networks (IPSN), pp. 11-19, Berkeley, CA, April 2004.
[70] N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, “Equations of state calculations
by fast computing machines,” Journal of Chemical Physics, vol. 21, no. 6, pp. 1087-1092, 1953.
[71] W. K. Hastings, “Monte Carlo sampling methods using Markov chains and their applications,” Biometrika, vol.
57, no. 1, pp. 97-109, 1970.
[72] A. H. Sayed and C. Lopes, “Distributed recursive least-squares strategies over adaptive networks,” Proc. 40th
Asilomar Conference on Signals, Systems and Computers, Pacific Grove, CA, pp. 233-237, October-November,
2006.
[73] J-W. Lee, S-E. Kim, W-J. Song, and A. H. Sayed, “Spatio-temporal diffusion mechanisms for adaptation over
networks,” Proc. EUSIPCO, pp. 1040-1044, Barcelona, Spain, August-September 2011.
[74] J-W. Lee, S-E. Kim, W-J. Song, and A. H. Sayed, “Spatio-temporal diffusion strategies for estimation and
detection over networks,” IEEE Trans. Signal Processing, vol. 60, no. 8, pp. 4017–4034, August 2012.
[75] S. Chouvardas, K. Slavakis, and S. Theodoridis, “Adaptive robust distributed learning in diffusion sensor net-
works,” IEEE Trans. on Signal Processing, vol. 59, no. 10, pp. 4692–4707, Oct. 2011.
[76] K. Slavakis, Y. Kopsinis, and S. Theodoridis, “Adaptive algorithm for sparse system identification using pro jec-
tions onto weighted ℓ1 balls,” Proc. IEEE ICASSP, pp. 3742–3745, Dallas, TX, March 2010.
[77] T. Kailath, A. H. Sayed, and B. Hassibi, Linear Estimation, Prentice Hall, NJ, 2000.
[78] F. Cattivelli and A. H. Sayed, “Diffusion distributed Kalman filtering with adaptive weights,” Proc. Asilomar
Conference on Signals, Systems and Computers, pp. 908–912, Pacific Grove, CA, November 2009.
[79] L. Xiao, S. Boyd and S. Lall, “A space-time diffusion scheme peer-to-peer least-squares-estimation,” Proc.
Information Processing in Sensor Networks (IPSN), pp. 168–176, Nashville, TN, April 2006.
[80] A. Nedic and A. Ozdaglar, “Distributed subgradient methods for multi-agent optimization,” IEEE Trans. Autom.
Control, vol. 54, no. 1, pp. 48–61, Jan. 2009.
[81] D. P. Bertsekas and J. N. Tsitsiklis, “Gradient convergence in gradient methods with errors,” SIAM J. Optim.,
vol. 10, no. 3, pp. 627–642, 2000.
[82] S. Barbarossa, and G. Scutari, “Bio-inspired sensor network design,” IEEE Signal Processing Magazine, vol. 24,
no. 3, pp. 26–35, May 2007.
[83] R. Olfati-Saber, “Kalman-consensus filter: Optimality, stability, and performance,” Proc. IEEE CDC, pp. 7036–
7042, Shangai, China, 2009.
[84] I. D. Schizas, G. Mateos, and G. B. Giannakis, “Distributed LMS for consensus-based in-network adaptive
processing,” IEEE Transactions on Signal Processing, vol. 57, no. 6, pp. 2365–2382, June 2009.
[85] G. Mateos, I. D. Schizas, and G. B. Giannakis, “Performance analysis of the consensus-based distributed LMS
algorithm,” EURASIP J. Adv. Signal Process., pp. 1–19, 2009.
[86] S. Kar and J. M. F. Moura, “Distributed consensus algorithms in sensor networks: Link failures and channel
noise,” IEEE Trans. Signal Process., vol. 57, no. 1, pp. 355–369, Jan. 2009.
[87] S. Kar and J. M. F. Moura, “Convergence rate analysis of distributed gossip (linear parameter) estimation:
Fundamental limits and tradeoffs,” IEEE Journal on Selected Topics on Signal Processing, vol. 5, no. 4, pp. 674–
690, August 2011.
[88] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione, “Gossip algorithms for distributed
signal processing,” Proc. IEEE, vol. 98, no. 11, pp. 1847–1864, November 2010.
[89] S-Y. Tu and A. H. Sayed, “Diffusion networks outperform consensus networks,” Proc. IEEE Statistical Signal
Processing Workshop, pp. 313–316, Ann Arbor, Michigan, August 2012.
[90] S.-Y. Tu and A. H. Sayed, “Diffusion strategies outperform consensus strategies for distributed estimation over
adaptive networks,” IEEE Trans. Signal Processing, vol. 60, no. 12, pp. 6217–6234, Dec. 2012.
115
|
1707.06444 | 1 | 1707 | 2017-07-20T11:08:15 | Consistent Tomography under Partial Observations over Adaptive Networks | [
"cs.MA",
"cs.IT",
"cs.IT"
] | This work studies the problem of inferring whether an agent is directly influenced by another agent over an adaptive diffusion network. Agent i influences agent j if they are connected (according to the network topology), and if agent j uses the data from agent i to update its online statistic. The solution of this inference task is challenging for two main reasons. First, only the output of the diffusion learning algorithm is available to the external observer that must perform the inference based on these indirect measurements. Second, only output measurements from a fraction of the network agents is available, with the total number of agents itself being also unknown. The main focus of this article is ascertaining under these demanding conditions whether consistent tomography is possible, namely, whether it is possible to reconstruct the interaction profile of the observable portion of the network, with negligible error as the network size increases. We establish a critical achievability result, namely, that for symmetric combination policies and for any given fraction of observable agents, the interacting and non-interacting agent pairs split into two separate clusters as the network size increases. This remarkable property then enables the application of clustering algorithms to identify the interacting agents influencing the observations. We provide a set of numerical experiments that verify the results for finite network sizes and time horizons. The numerical experiments show that the results hold for asymmetric combination policies as well, which is particularly relevant in the context of causation. | cs.MA | cs | Consistent Tomography under Partial
Observations over Adaptive Networks
Vincenzo Matta and Ali H. Sayed
7
1
0
2
l
u
J
0
2
]
A
M
.
s
c
[
1
v
4
4
4
6
0
.
7
0
7
1
:
v
i
X
r
a
Abstract
This work studies the problem of inferring whether an agent is directly influenced by another agent
over an adaptive diffusion network. Agent i influences agent j if they are connected (according to the
network topology), and if agent j uses the data from agent i to update its online statistic. The solution
of this inference task is challenging for two main reasons. First, only the output of the diffusion learning
algorithm is available to the external observer that must perform the inference based on these indirect
measurements. Second, only output measurements from a fraction of the network agents is available, with
the total number of agents itself being also unknown. The main focus of this article is ascertaining under
these demanding conditions whether consistent tomography is possible, namely, whether it is possible to
reconstruct the interaction profile of the observable portion of the network, with negligible error as the
network size increases. We establish a critical achievability result, namely, that for symmetric combination
policies and for any given fraction of observable agents, the interacting and non-interacting agent pairs
split into two separate clusters as the network size increases. This remarkable property then enables the
application of clustering algorithms to identify the interacting agents influencing the observations. We
provide a set of numerical experiments that verify the results for finite network sizes and time horizons.
The numerical experiments show that the results hold for asymmetric combination policies as well, which
is particularly relevant in the context of causation.
Index Terms
Diffusion networks, network tomography, causation, combination policy, Erdos-Renyi model.
I. INTRODUCTION
A significant number of complex real-world systems are modeled well by networks. At an abstract
level, a network is an ensemble of interconnected agents. The interactions among neighboring agents
V. Matta is with DIEM, University of Salerno, via Giovanni Paolo II, I-84084, Fisciano (SA), Italy (e-mail: [email protected]).
A. H. Sayed is with the ´Ecole Polytechnique F´ed´erale de Lausanne EPFL, School of Engineering, CH-1015 Lausanne,
Switzerland (e-mail: [email protected]). His work was supported in part by NSF grants CCF-1524250 and ECCS-1407712.
June 26, 2021
DRAFT
2
enable the flow of information across the graph, and give rise to interesting and complex patterns of
coordinated behavior.
One problem of immense value in network science is the inverse modeling problem. In this problem,
the network structure (topology) is assumed to be largely unknown and one is interested in inferring
relationships between agents based on some dataset arising from the agents' operations. Formulations of
this type are of great interest in several application domains, such as communications, computer science,
cyber-security, control, physics, biology, economics, and social sciences.
Inverse problems over networks are challenging because, in the vast majority of applications, direct
access to the data exchanged between agents is often unavailable, and the inference of inter-agent relations
must be based on indirect observations. Another source of difficulty in such inverse problems is that the
access to observations is generally limited to a subset of the network agents. The process of discovering
inter-agent interactions from indirect measurements is broadly referred to as network tomography. Some
useful applications of network tomography include, among other possibilities: tracing the routes of
clandestine information flows across communications networks [1] -- [6]; revealing agent interactions over
social networks, where disclosing commonalities within groups of agents might be useful for commercial
as well as security purposes [7], [8]; inferential problems related to group testing and identification of
defective items [9]; brain networks, where interactions among neurons might be of paramount importance
to the understanding of particular diseases [10], [11]; anomaly detection in communications networks,
where one tries to reveal the activities of intermediate nodes through destination-only measurements [12].
One useful type of networks is the class of adaptive networks [13], [14]. These networks consist of
spatially dispersed agents that continually exchange information through diffusion mechanisms, and which
are specifically designed to enable simultaneous adaptation and learning from streaming data (such as
tracking targets moving in formation from streaming spatial observations; modeling the prey-predator
behavior of animal groups on the move; allocating frequency resources over cognitive communication
networks) [13], [14]. Using a powerful form of agent cooperation, adaptive diffusion networks are able
to solve rather sophisticated inferential tasks, including: estimation tasks [14], detection tasks [15], [16],
optimization and online learning tasks [17], [18]. This work is focused on tomography over partially
observed diffusion networks.
A. Related Work
From a merely theoretical perspective, the aforementioned problems fall under the umbrella of signal
processing over graphs [13], [14], [19] -- [21]. They deal with the objective of retrieving a graph topology
(a connection topology, or an "effective" topology corresponding to the exchange of information) from
June 26, 2021
DRAFT
3
a set of indirect measurements taken at some accessible network locations. There are several works
addressing a similar construction, albeit with different specific goals. With no pretence of exhaustiveness,
we give a brief summary of the works that we believe are most related to the present article.
In [22], an unknown network topology is reconstructed by taking advantage of the locality properties
of the Wiener filter. Exact reconstruction results are provided for self-kin networks, while reconstruction
of the smallest self-kin network embracing the true network topology is guaranteed for general networks.
In [23], the authors introduce the concept of directed information graphs, which is used to capture the
dependencies in networks of interacting processes linked by causal dynamics. The setting is further
enriched in [24], where a new metric to learn causal relationships by means of functional dependencies,
in a possibly nonlinear dynamical network, is proposed.
More closely related to the network model considered in our work is the problem of estimating a
graph when the relationships are encoded into an autoregressive model -- see, e.g., [25], where several
methods to address causal inference in such a context are reviewed. Causal graph processes are also
exploited in [26], where a computationally tractable algorithm for graph structure recovery is proposed,
along with a detailed convergence analysis. In [27], the authors examine a causal inference problem that
is modeled through a vector autoregressive process. The objective is that of reconstructing the important
parts of the transition matrix through observation of a subset of the random process. Special technical
conditions for exact reconstruction are provided. In [28], new methods are proposed to learning directed
information polytrees where samples are available from only a subset of processes.
Some recent works exploit spectral graph properties, in conjunction with sparsity constraints. In [29],
[30], the problem of inferring the graph topology from observations of a random signal diffusing over
the network is addressed. It is shown that the space of feasible matrices is a convex polytope, and
two inferential strategies are proposed to select a point in this space as the topology estimate. In [31],
the authors exploit convex optimization and sparsity in order to reconstruct an unknown graph from
observable indirect relationships generated by diffusive signals defined on the graph nodes. The main
idea is identifying a graph shift operator given only its eigenvectors, with the corresponding spectral
templates being obtained from the sample covariance of independent graph signals diffused over the
network.
B. Main Result
This work complements the previous efforts: it establishes an important identifiability condition and
clarifies the asymptotic behavior of the recovery error as a function of the network size. An identification
procedure is also developed to carry out the tomography calculations using a standard clustering technique.
June 26, 2021
DRAFT
4
We summarize the main contributions of this work as follows. We consider a network of agents running
a diffusion strategy to solve some inference task of interest (such as a distributed detection problem).
The overall network size is unknown, and the outputs of the diffusion algorithm are available from
only a limited fraction of agents. The goal is determining the relationships among these agents, namely,
establishing whether the datum of an individual agent is influencing another individual agent. We show
that, under some regularity conditions, the interaction relationships existing within the observable portion
of the network can be recovered, with negligible error for sufficiently large network sizes.
More specifically, we consider the class of Erdos-R´enyi random graphs, where the probability of two
agents being connected follows a Bernoulli distribution. For these graphs we discover that, for any fraction
of observable agents, the group of interacting agent pairs and the group of non-interacting agent pairs split
into two well-defined clusters. These clusters emerge as clearly separate for sufficiently large network
sizes. This result is established for diffusion strategies employing symmetric combination matrices,
a feature that enables analytical tractability of the problem. However, from a "physical" viewpoint,
separability of the clusters does not appear to be limited to the symmetry assumption. Accordingly, we
examine also the relevant case of asymmetric matrices, which is of interest especially in the context
of causal networks. With reference to a typical form of asymmetric combination matrices, numerical
simulations show that separability of the clusters can be preserved.
One distinctive feature of our work is that we establish theoretical achievability results for consistent
tomography from partially observed networks. In answering these particular questions, in addition to
what has been answered before, our work complements well existing results and recent progresses in the
field of network tomography.
Notation. We use boldface letters to denote random variables, and normal font letters for their real-
izations. Capital letters refer to matrices, small letters to both vectors and scalars. Sometimes we violate
the latter convention, for instance, we denote the total number of network agents by N.
For i, j = 1, 2, . . . , N, the (i, j)-th entry of an N × N matrix Z will be denoted by zij, or alternatively
by [Z]ij. Moreover, the sub-matrix that lies in the rows of Z indexed by the set S1 ⊆ {1, 2, . . . , N} and
in the columns indexed by the set S2 ⊆ {1, 2, . . . , N}, will be denoted by ZS1S2, or alternatively by
[Z]S1S2. If S1 = S2 = S, the sub-matrix ZS1S2 will be abbreviated as ZS, or as [Z]S.
The symbols P, E, and V are used to denote the probability, expectation, and variance operators,
respectively. The notation p−→ denotes convergence in probability as N → ∞.
June 26, 2021
DRAFT
5
A. The Adaptive Diffusion Network
II. THE PROBLEM
A network of N agents collects streaming data from the environment. The datum collected by the i-th
agent at time n is denoted by xi(n), and the global sequence of data is assumed to be formed by spatially
(i.e., w.r.t. index i) and temporally (i.e., w.r.t. index n) independent and identically distributed random
variables. Without loss of generality, we assume that the variables have zero mean and unit variance.
In order to track drifts in the phenomenon that the network is monitoring, the agents implement a
distributed adaptive strategy, where each agent relies on sharing information with local neighbors. In
this work we focus on the Combine-Then-Adapt (CTA) diffusion mechanism, whose useful properties in
terms of estimation and online learning performance have been already studied in some revealing detail
in [13], [14], [17], [18]. The CTA algorithm can be described as follows.
First, during the combination step, agent i mixes the data received from its neighbors by using a
sequence of convex (i.e., nonnegative and adding up to 1) combination weights wi(cid:96), for (cid:96) = 1, 2, . . . , N,
giving rise to the following intermediate variable:
vi(n − 1) =
wi(cid:96) y(cid:96)(n − 1).
(1)
(cid:96)=1
Then, during the adaptation step, agent i updates its output variable by comparison with the incoming
streaming data xi(n), employing a (typically small) step-size µ ∈ (0, 1):
yi(n) = vi(n − 1) + µ[xi(n) − vi(n − 1)].
(2)
N(cid:88)
N(cid:88)
Equations (1) and (2) can be merged into a single step as:
yi(n) = (1 − µ)
wi(cid:96) y(cid:96)(n − 1) + µ xi(n).
(3)
For later use, it is convenient to introduce the scaled combination matrix A, whose entries are defined
(cid:96)=1
as:
aij (cid:44) (1 − µ)wij,
(4)
which, in the sequel, will be simply referred to as the combination matrix. We remark that, since we use
a sequence of convex combination weights, the matrix A/(1 − µ) is a right-stochastic matrix.
If we now stack the observations across the network at time n into the N × 1 vector xn, and the state
variables at time n into the N × 1 vector yn, Eq. (3) can be compactly rewritten as:
yn = A yn−1 + µ xn,
(5)
DRAFT
June 26, 2021
6
Fig. 1. Conceptual sketch of the tomography problem addressed in this article.
which, by iteration, allows us to express the diffusion output vector as a function of the streaming data:
n(cid:88)
yn = µ
An−mxm, n ≥ 1
(6)
starting from state y0 = 0, i.e., neglecting the transient term.
m=1
B. Network Tomography
An illustration of the setting considered in this work is given in Fig. 1. An entity external to the
network (hereafter named Tomography Center, TC) is interested in reconstructing the interaction profile
of the network, namely, it is interested in ascertaining which agent is influencing which other agent.
The TC is assumed to have access to a subset of the network agents, and is able to collect the streams
of outputs exchanged by such agents during their communication activity. Letting Ω ⊂ {1, 2, . . . , N} be
the observable subnet, the data available to the TC at time n are {yi(1), yi(2), . . . , yi(n)}i∈Ω. We shall
focus on the asymptotic regime of large networks (N → ∞), for the meaningful case where the fraction
of observed agents does not vanish. Letting K = Ω, such regime is defined in terms of the following
condition:
where ξ ∈ (0, 1) takes on the meaning of the asymptotic fraction of observed agents.
→ ξ
K
N
June 26, 2021
(7)
DRAFT
2 a12>0 a21>0 a11>0 Observable Subnet !"#1 j i k ℓdenotethetotalnumberofnetworkagentsbyI.THEPROBLEMEquationsforfigureyi(1),...,yi(n)(1),...,yj(n)(Sometimesweviolatethislatterconvention,forinstance,denotethetotalnumberofnetworkagentsbyareusedtodenotetheprobabilityandexpectationI.Tyi(1),...,yi(n)yj(1),...,yj(n)ℓ(1),...,yℓ(n),...,yk(nagentscollectsstreamingdatafromthe,andtheglobalsequenceofdataisyℓ(1),...,yℓ(n),...,yj(nyℓ(1),...,yℓ(n)yk(1),...,yk(n)A.TheAdaptiveDiffusionNetworkagentscollectsstreamingdatafromtheenvironment.Thedatumcollectedbytheassumedtobeformedbyspatially(i.e.,w.r.t.indexdistributedrandomvariables,withvarianceofgenerality,weassumethatthevariableshavezeromean. 7
In our setting, the TC does not know the overall number of agents in the network. Accordingly, the
main goal of the TC is producing an estimate of the interaction profile for the observed agents. This
problem is challenging for the following reasons. Let us ignore for a moment the fact that the network
is partially observed, and assume that the TC is able to collect all output sequences from all agents at
all times. Using such a dataset, there exist several well-established strategies to make inference about the
influence that one agent has on another agent. The most intuitive is an estimate of the correlation between
the outputs of two agents, which, however, is problematic for directed flows of information (where agent
i can influence agent j but not vice versa). Such asymmetry would be reflected into the (i, j)-th and
(j, i)-th entries of the combination matrix, yielding aij = 0 while aji > 0. The case of asymmetric
influence is well studied within the framework of causal inference, or causation. Many solutions exist
for disclosing causal relationships [23] -- [28].
There is, however, a challenging problem that is peculiar to the network setting of this work due to the
streaming nature of the data. In general, when the TC starts collecting data, the network would have been
in operation since some time already. Therefore, the output signals at the agents would have benefited
from sufficient exchanges of information. While this exchange of information is beneficial for solving
inference tasks, it nevertheless can become detrimental for reconstructing the network tomography. This
is because, over a strongly connected network and after a "transient" phase, all agents would have become
"correlated!"
In order to overcome this difficulty, we exploit knowledge of the diffusion mechanism. To this aim,
let us introduce the correlation matrix of the diffusion output vector:
which, from (6), admits the following closed-form representation:
Ai(Ai)T n→∞−→ R0 = µ2
R0(n) = µ2
n−1(cid:88)
R0(n) (cid:44) E[ynyT
n ],
(8)
(9)
Ai(Ai)T ,
∞(cid:88)
where the latter series is guaranteed to converge since all eigenvalues of A are strictly inside the unit
disc. The limiting correlation matrix, R0, can be interpreted as the (unique) solution to the discrete-time
i=0
i=0
Lyapunov equation [13]:
(10)
where IN denotes the N × N identity matrix. We note that R0(n) is positive definite for each n, and so
is R0, due to the stability of A [33]. We also introduce the one-lag correlation matrix, which, in view
R0 − AR0AT = µ2IN ,
of (5), takes on the form:
R1(n) (cid:44) E[ynyT
n−1] = AR0(n − 1)
n→∞−→ R1 = AR0.
June 26, 2021
(11)
DRAFT
Therefore, we obtain the following relationship:
A = R1(n)(R0(n − 1))
−1, n ≥ 2,
8
(12)
and, at steady-state,
(13)
In principle, since there exist many ways to estimate R0 and R1 consistently as n → ∞, expression (13)
reveals one possible strategy to estimate A from the output of the diffusion process. However, the approach
−1
A = R1R
0
described so far suffers from a critical problem: the network is only partially observed and, hence, not
all entries in the matrices R0 and R1 can be estimated. This in turn means that the evaluation of (13) is
in general problematic.
In order to estimate the combination (sub-)matrix corresponding to the observable agents, i.e.,
A(obs) (cid:44) AΩ,
(14)
one might be tempted to simply replace the matrices involved in (13) by their observable counterparts,
R(obs)
(cid:44) [R1]Ω, obtaining the rough estimate:1
0
(cid:44) [R0]Ω, and R(obs)
1
A(obs) = R(obs)
1
(R(obs)
0
)
−1
(15)
Needless to say, calculation on the right-hand side of (15) does not lead to the true A(obs), except in some
special cases. For this reason, we are denoting the result of (15) by using the hat notation. If we could
recover A(obs) exactly, then we could deduce the desired influence relations. However, given that we only
have the estimate A(obs), it is not clear at all whether the mutual influence relationships existing between
the observed nodes can be consistently retrieved from A(obs). Answering this nontrivial question in the
affirmative is one of the main contributions of this work.
In order to highlight the key ideas without added complexity, we focus in this article on the case of
symmetric combinations matrices. The symmetry assumption is made because, if A is symmetric, the
autocorrelation matrix of the diffusion output takes on the following convenient form -- see (9):
n−1(cid:88)
R0(n) = µ2
A2i ⇒ R0 = µ2(IN − A2)
−1,
(16)
which is exploited in the proof of Theorem 1. The extension of the results to the asymmetric case
requires adjustments in the arguments used in the proof of Theorem 1. For example, this can be pursued by
i=0
1Actually, the matrices computed by the TC will generally be renumbered versions of [R0]Ω and [R1]Ω, since the labeling
used by the TC need not correspond to the ordering in [R0]Ω and [R1]Ω. This is because the TC does not know the original
labeling of the agents in the network, nor the network size. However, our focus is on evaluating the interaction between pairs
of agents, irrespective of their labels, and, hence, we can safely keep the notation used in (15).
June 26, 2021
DRAFT
9
appealing instead to the Kronecker representation of the solution of discrete-time Lyapunov equations [34].
We note in passing that, since the matrix A/(1− µ) is always right-stochastic, the symmetry assumption
makes A/(1 − µ) doubly-stochastic.
III. ERROR CAUSED BY PARTIAL OBSERVATIONS
Using the estimate (15), we can write:
A(obs) = A(obs) + E,
(17)
where E denotes the error matrix. In terms of the individual entries, we can write, for i, j = 1, 2, . . . , K:2
a(obs)
ij = a(obs)
ij + eij.
(18)
In this work we are interested in establishing whether the estimated values a(obs)
allow us to identify the
condition a(obs)
ij = 0, which would reveal whether the agents i and j influence each other. In
order to enable this determination, we start by characterizing the behavior of the error terms eij in (18).
ij > 0 or a(obs)
ij
Theorem 1 (Concentration of errors): For a symmetric combination matrix, the entries of the error
matrix defined in (17) are nonnegative, and satisfy for all i = 1, 2, . . . , K:
K(cid:88)
eij ≤ 1 − µ
(19)
Proof: See Appendix A.
j=1
We have the following corollary, which is particularly tailored to our application.
Corollary 1 (Number of errors exceeding a threshold): For a symmetric combination matrix, we have,
for all i = 1, 2, . . . , K, and for any > 0:
K(cid:88)
j=1
I{eij > } ≤ 1 − µ
(20)
where I{·} is the indicator function (which is equal to one when its argument is true and zero otherwise).
Proof: Suppose that (20) is false. Then, we would have (recall that eij ≥ 0 for all i, j = 1, 2, . . . , K):
K(cid:88)
j=1
eij =
(cid:88)
K(cid:88)
j=1
j:eij>
>
(cid:88)
eij ≥ (cid:88)
eij
eij +
j:eij≤
I{eij > } >
j:eij>
1 − µ
× = 1 − µ,
2A pair (i, j) refers to agents ωi and ωj, where ω1 < ω2 < ··· < ωK span the observable subnet Ω.
June 26, 2021
(21)
DRAFT
10
which contradicts (19).
Theorem 1 and its corollary provide useful information about the concentration of the entries in the
error matrix. In particular, Eq. (19) reveals that the row-sum of the entries in the error matrix cannot
exceed 1− µ, whereas Eq. (20) places an upper bound on the number of entries that exceed any positive
threshold.
Consider now a small threshold , and the fraction of off-diagonal (because we are interested in inter-
agent interactions) entries that exceed :
1
K(K − 1)
K(cid:88)
(cid:88)
i=1
j(cid:54)=i
I{eij > }.
(22)
In the regime dictated by (7), where K → ∞ as N → ∞, we see from (20) that such fraction vanishes,
namely, that most entries of the error matrix will be small in the asymptotic regime of large networks.
Therefore, in view of definition (18), it will hold for large networks that, for i (cid:54)= j,
a(obs)
ij =
a(obs)
ij + small quantity,
if a(obs)
ij > 0,
small quantity,
if a(obs)
ij = 0.
(23)
This useful dichotomy suggests that the nonzero entries of A(obs) will make the estimated entries a(obs)
stand out above the error floor as N increases. As a result, we should be able to decide whether a(obs)
or a(obs)
ij > 0 by comparing the estimated value, a(obs)
However, and unfortunately, the behavior of the error matrix alone is not sufficient to conclude that
, against some threshold.
ij = 0
ij
ij
ij
this inferential procedure is feasible, for one crucial reason. For most typical combination matrices, the
nonzero entries a(obs)
decrease with N as well, so that the nonzero entries appearing in the first line
of (23) vanish for large network sizes, along with the errors. This means that the estimated entries, a(obs)
,
would vanish even when agents i and j are interacting (i.e., even when a(obs)
ij > 0). For this reason, a
closer examination of the behavior of the error quantities is necessary before we can conclude that this
ij
inference procedure is viable. Specifically, it is necessary to assess how fast the error signals eij decay
to zero in relation to the desired entries a(obs)
assumptions about the network structure. We will select some typical and popular random models, as
. In order to carry out this analysis, we need to make some
ij
explained next.
June 26, 2021
DRAFT
IV. BEHAVIOR OF INTERACTING AGENTS (aij > 0)
11
The interaction profile can be conveniently described in terms of a symmetric interaction matrix G,
defined by the following conditions:
gij =
1, if aij > 0,
0, if aij = 0.
(24)
We shall set gii = 1 for all i = 1, 2, . . . , N, which corresponds to the assumption that an agent does
always use its own state variable in the combination step.
In our analysis, the combination matrix, A, will be constructed through the following two-step proce-
dure. First, an interaction matrix G is generated according to a random graph model [35], [36]. Then,
A is determined by a combination policy, γ(G), which sets the values of the combination weights
corresponding to the nonzero entries of G. Formally:
A = γ(G)
(25)
Note that γ(G) must always assign positive weights at the locations corresponding to nonzero entries
of G, otherwise the interaction matrix related to A would be different from G. Moreover, since in this
article we focus on symmetric combination matrices, we must have that [γ(G)]ij = [γ(G)]ji (a condition
that would not be directly implied by the symmetry of G).
A. Random Model for the Interaction Profile
As we have stated, in this work we examine the useful case where the interaction profile of the network
is generated according to a random graph model [35], [36]. In particular, we consider the classic Erdos-
R´enyi model. This model, which we denote by G (N, pN ), is an undirected (i.e., symmetric) graph, where
the presence/absence of the N (N − 1)/2 edges is determined by a sequence of N (N − 1)/2 independent
Bernoulli random variables with success (i.e., interaction) probability pN . Accordingly, the variables gij,
for i = 1, 2, . . . , N and j > i, are independent Bernoulli random variables with P[gij = 1] = pN , and
the matrix G is symmetric.
One meaningful regime to examine the random graph properties is the regime where the probability
pN decreases as N increases [36]. Examining the asymptotic regime where pN is kept constant while N
increases would generally produce networks that are too much connected with respect to what happens in
typical applications. For instance, in the regime of constant pN , the network diameter (i.e., the maximum
of the shortest distance between any pair of nodes) is equal to 2 with high probability [37]. On the other
hand, pN could not vanish in an arbitrary fast way, otherwise the number of neighbors of an agent would
June 26, 2021
DRAFT
be too small, and the network would become very scarcely connected. A well-known result that holds
true for the Erdos-R´enyi graph is that the following scaling law:
12
(26)
with cN → ∞ (in an arbitrarily slow fashion, i.e., even with pN → 0), ensures that the graph remains
connected with probability tending to 1 as N diverges [36]. In the following, we shall focus on the regime
pN =
N
ln N + cN
of connected Erdos-R´enyi graph with vanishing pN , namely, on the regime where (26) is satisfied with
cN → ∞ and pN → 0. Such a regime will be denoted by the symbol G (cid:63)(N, pN ).
B. Statistical Properties of Fundamental Graph Descriptors
Some useful descriptors of the network graph can be conveniently represented in terms of the matrix
G in (24). In particular, the (interaction) neighborhood of agent i (which includes i itself) is defined as:
Ni = {(cid:96) ∈ {1, 2, . . . , N} : gi(cid:96) = 1},
while the degree of agent i is:
di = Ni =
N(cid:88)
gi(cid:96) = 1 +
(cid:96)=1
We define the network maximal degree as:
(cid:88)
(cid:96)(cid:54)=i
gi(cid:96) = 1 +
(cid:88)
(cid:96)(cid:54)=i
g(cid:96)i.
dmax (cid:44) max
i=1,2,...N
di.
(27)
(28)
(29)
Let us now highlight some useful statistical properties of the degree and maximal degree variables for
the Erdos-Renyi model.
We observe from (28) that, for an Erdos-Renyi graph, the random variable di− 1 is a binomial random
variable with parameters N − 1 and pN , which shall be denoted by β(N − 1, pN ). Note also that di and
dj, for i (cid:54)= j, are not independent, because of the implied graph symmetry. Now, for a binomial random
variable β(N, pN ), we have:
(cid:20) β(N, pN )
(cid:21)
(cid:20) β(N, pN )
(cid:21)
E
= 1, V
(30)
Under the G (cid:63)(N, pN ) model, we see from (26) that the product N pN diverges as N → ∞, and, hence, the
variance in (30) vanishes as N → ∞, implying in particular the following convergence in probability [38]:
(N pN )2
N pN
N pN
=
.
N pN (1 − pN )
which reveals that the degrees of the nodes scale as N pN .
β(N, pN )
N pN
p−→ 1
June 26, 2021
(31)
DRAFT
It is also of interest to characterize the asymptotic behavior of dmax, which, by being the maximum
of N degrees, is expected to grow faster than N pN . However, and interestingly, the following lemma
shows that it cannot grow much faster.
13
Lemma 1 (Behavior of maximal degree): Under the G (cid:63)(N, pN ) model we have, for all i, j = 1, 2, . . . , N,
with i (cid:54)= j:
P[dmax ≥ N pN e gij = 1] ≤
where e is Euler's number.
Proof: See Appendix B.
(cid:19)
(cid:18)
e +
2e2
N
−cN N→∞−→ 0
e
(32)
According to (32), the maximal degree exceeds the level N pN e with negligible probability, i.e., it cannot
grow substantially faster than N pN . We note that the probability in (32) is computed conditionally on
the event that two agents interact. This choice is made because, in the following, we need to know the
behavior of the maximal degree in relation to interacting agent pairs.
C. Stable Combination Policies
We can now use (32) to characterize the asymptotic behavior of the (off-diagonal) nonzero entries in
the combination matrix. In order to illustrate the main idea, we start by examining a popular combination
policy, known as the Laplacian rule, which is given by [13]:
gij (1 − µ) λ/dmax,
aij =
(1 − µ) −(cid:80)
(cid:96)(cid:54)=i ai(cid:96)
for i (cid:54)= j,
for i = j.
for some λ, with 0 < λ ≤ 1. Therefore, from Lemma 1 we can write, for all i (cid:54)= j:
(33)
(34)
P[N pN aij > (1 − µ)λ/e gij = 1]
= P[dmax < N pN e gij = 1]
≥ 1 −
−cN N→∞−→ 1.
(cid:18)
(cid:19)
e +
e
2e2
N
Equation (34) has the following important implication: for large enough N, any nonzero entry of the
Laplacian combination matrix, scaled by the factor N pN , stays "almost always" above a certain threshold
(namely, the value (1 − µ)λ/e). Therefore, multiplying aij by the scaling factor N pN would keep the
nonzero entries stable (in the sense that they will not vanish) as N diverges. The same scaling factor
is relevant for other combination policies. Therefore, it makes sense to introduce the following general
class of combination policies.
June 26, 2021
DRAFT
14
Combination-policy class Cτ . A combination policy belongs to class Cτ if there exists τ > 0 such
that, for all i, j = 1, 2, . . . , N, with i (cid:54)= j:
(35)
where N goes to zero as N → ∞, and where the probability is evaluated under the G (cid:63)(N, pN ) model.
(cid:3)
P[N pN aij > τ gij = 1] ≥ 1 − N
Let us now examine the physical meaning of (35) in connection to network tomography applications.
For a policy belonging to class Cτ , we can rephrase (23) as:
(cid:124) (cid:123)(cid:122) (cid:125)
+ N pN eij
small quantity?
,
if a(obs)
ij > 0,
ij
(cid:124)
not vanishing
N pN a(obs)
(cid:125)
(cid:123)(cid:122)
(cid:124) (cid:123)(cid:122) (cid:125)
N pN eij
small quantity?
N pN a(obs)
ij =
,
if a(obs)
ij = 0,
(36)
where the qualification of being "not vanishing" is a consequence of (35). In light of (36), if we will
able to show that N pN eij is still a small quantity, then a(obs)
ij would be effectively useful for tomography
purposes, because the nonzero entry N pN a(obs)
ij would stand out from the error floor as N gets large.
Actually, this heuristic argument will be made rigorous in the proof of Theorem 2, which is presented
in the next section.
Before ending this section, we would like to introduce another useful combination policy that belongs
to class Cτ , namely, the Metropolis rule, which is given by [13]:
gij (1 − µ)/ max(di, dj),
aij =
(1 − µ) −(cid:80)
(cid:96)(cid:54)=i ai(cid:96),
for i = j.
for i (cid:54)= j,
(37)
Since max(di, dj) ≤ dmax, for the Metropolis rule we have the following implication, for all i (cid:54)= j:
{dmax < N pN e gij = 1}
⇒ {max(di, dj) < N pN e gij = 1}
= {N pN aij > (1 − µ)/e gij = 1},
(38)
where the latter equality comes from (37). Now, since for two events E1 and E2, the condition E1 ⇒ E2
implies that P[E2] ≥ P[E1], from (38) we can write:
P[N pN aij > (1 − µ)/e gij = 1]
≥ P[dmax < N pN e gij = 1]
≥ 1 −
(cid:18)
(cid:19)
−cN N→∞−→ 1,
e +
e
2e2
N
(39)
DRAFT
June 26, 2021
15
where the latter inequality follows by Lemma 1. Equation (39) reveals that the Metropolis rule satis-
fies (35) with the choice τ = (1 − µ)/e.
V. CONSISTENT TOMOGRAPHY
In order to establish whether the estimated matrix, A(obs), can be used to infer the interaction pattern
contained in A(obs), we still need to provide a statistical characterization for the entries of A(obs). We
pursue this goal by characterizing the asymptotic behavior of two conditional distributions: the empirical
distribution given that two agents do not interact, and the empirical distribution given that two agents
interact. In particular, owing to the N pN scaling factor that we have obtained in the previous section,
we shall focus on the scaled matrix N pN A(obs).
Let us preliminarily introduce the number of non-interacting (N0) and the number of interacting (N1)
agent pairs over the observed set, defined respectively as:
K(cid:88)
(cid:88)
i=1
j(cid:54)=i
N0 (cid:44)
(1 − g(obs)
ij
),
N1 (cid:44)
K(cid:88)
(cid:88)
i=1
j(cid:54)=i
g(obs)
ij
,
(40)
ij = I{a(obs)
ij > 0} is the (i, j)-the entry of the interaction sub-matrix corresponding to the
where g(obs)
observable agents. Next we introduce the number of entries in N pN A(obs) that stay below some positive
level α, for the case of non-interacting and interacting agent pairs, respectively:
N0(α) (cid:44)
N1(α) (cid:44)
I{N pN a(obs)
ij ≤ α, g(obs)
ij = 0},
I{N pN a(obs)
ij ≤ α, g(obs)
ij = 1}.
(41)
(42)
K(cid:88)
K(cid:88)
i=1
(cid:88)
(cid:88)
j(cid:54)=i
i=1
j(cid:54)=i
Using (40), (41) and (42), we can compute the conditional empirical distributions given that the agents
are not interacting and given that they are interacting, defined respectively as:
F0(α) =
N0(α)
N0
,
F1(α) =
N1(α)
N1
,
(43)
where F0(α) (resp., F1(α)) are conventionally set to 1/2 when N0 = 0 (resp., N1 = 0). It is also
convenient to introduce the complementary distribution ¯F1(α) = 1 − F1(α). Accordingly, the quantity
F0(α) represents the fraction of entries in N pN A(obs) that correspond to non-interacting agent pairs and
that stay below α. In contrast, the quantity ¯F1(α) represents the fraction of entries in N pN A(obs) that
correspond to interacting agent-pairs and that stay above α.
The next theorem establishes achievability of consistent tomography through the asymptotic charac-
terization of the aforementioned empirical distributions.
June 26, 2021
DRAFT
16
Fig. 2. A pictorial illustration of Theorem 2, which establishes that the scaled estimated entries a(obs)
depending on whether agents i and j are interacting or not.
ij
cluster in two groups
Theorem 2 (Achievability of consistent tomography): Let the network interaction profile obey a G (cid:63)(N, pN )
model, and let the combination policy belong to class Cτ . Then, for any nonzero fraction of observable
agents satisfying (7), we have in the limit as the network size increases (N → ∞):
F0()
p−→ 1 ∀ > 0,
¯F1(τ )
p−→ 1
(44)
Proof: See Appendix C.
Theorem 2 reveals that, from the knowledge of the estimated combination matrix, A(obs), it is possible
to separate the zero/nonzero entries of the true combination matrix, A(obs). In fact, we see from Theorem 2
that the following dichotomous behavior is observed, asymptotically with N: i) when agents i and j are
not interacting, most of the (scaled) estimated matrix entries, N pN a(obs)
hence, also below an arbitrarily small value < τ); ii) when agents i and j are interacting, most of
the (scaled) estimated matrix entries, N pN a(obs)
between the two classes of non-interacting and interacting agent pairs arises, which translates into the
ij ≤ , and the other one
emergence of two separate clusters, one corresponding to the region N pN a(obs)
corresponding to the region N pN a(obs)
ij > τ. This situation is illustrated in Fig. 2.
, stay above a positive value τ. Therefore, a separation
ij
, stay below any level (and,
ij
Theorem 2 establishes the separability of the two classes based upon knowledge of A(obs), whose
computation requires knowledge of the exact correlation sub-matrices, [R0]Ω and [R1]Ω -- see (15).
Since, in principle, such matrices can be estimated with arbitrarily large precision as the steady-state is
June 26, 2021
DRAFT
12012000.050.10.150.20.250.30.35 indexiindexj NpNa(obs)ij(i,j):g(obs)ij=0(i,j):g(obs)ij=1levelτlevelǫ 17
approached (i.e., as the number of collected diffusion output samples increases) the result in Theorem 2
is an achievability result. In addition, for many of the methods available to estimate correlation matrices,
the behavior of the estimation error is known, at least for n sufficiently large. A useful extension of the
present treatment would be examining the interplay between the two sources of error, namely, the error
caused by partial observations, and the error caused by estimation of the correlation sub-matrices.
When some prior knowledge about the value of τ and the typical number of neighbors, N pN , is
available, Theorem 2 provides a constructive recipe to perform the reconstruction of the interaction
profile across the observed network. In fact, the separation between the two classes of interacting and
non-interacting agent pairs can be performed by comparing each entry a(obs)
lying between 0 and τ /(N pN ).
to an intermediate threshold
ij
On the other hand, in many contexts it is unrealistic to assume precise knowledge of the parameters
τ and N pN . When such knowledge is not available, the aforementioned reconstruction strategy is not
applicable, but the achievability result in Theorem 2 still has a fundamental implication in that it guarantees
the existence of the clustering structure! The existence of two separate thresholds, < τ, and the related
ij ≤ and
(asymptotic) separation of the scaled estimated entries in two separate clusters (N pN a(obs)
N pN a(obs)
ij > τ), opens up the possibility of employing universal and non-parametric pattern recognition
strategies to perform cluster separation. In particular, in our numerical experiments, we shall verify the
validity of this argument by employing the k-means clustering algorithm.
VI. HIGHER ORDER ASYMPTOTIC ANALYSIS
As illustrated in the previous section, and as we can infer from Fig. 2, it is undesirable to have
N pN a(obs)
ij > when gij = 0. For the sake of brevity, let us refer to the agent pair for which such event
occurs as being identified as "mistakenly-interacting". Examining (40), (41) and (43), we deduce that the
number of mistakenly-interacting pairs is given by:
where we used the fact that, under the G (cid:63)(N, pN ) model, N0 scales as N (N − 1)(1 − pN ), and pN
vanishes as N → ∞.
On the other hand, for the number of truly-interacting agent pairs we have that:
N1 ≈ N 2pN .
(46)
The fact that pN vanishes as N increases causes the following issue. According to (45) -- (46) and also
to (84) -- (85), without any information about the speed of convergence of 1 − F0() (in comparison to
June 26, 2021
DRAFT
N0 − N0() = N0[1 − F0()] ≈ N 2 [1 − F0()]
,
→0, see (44)
(cid:125)
(cid:124)
(cid:123)(cid:122)
(45)
18
pN ), we cannot exclude that the number of mistakenly-interacting pairs is larger than the number of
truly-interacting pairs. In order to ward off the presence of such unpleasant feature, we should prove that
the quantity 1 − F0() goes to zero faster than pN , formally:
p−→ 0,
(47)
a condition that will be abbreviated as 1 − F0() = op(pN ). Verification of such desired requirement
is addressed in the forthcoming Proposition 1. In order to prove Proposition 1, we require that the
1 − F0()
pN
combination policy possesses two additional regularity properties, which are again automatically satisfied
by the Laplacian and Metropolis rules.
Property P1. There exists κ > 0 such that, for all i, j = 1, 2, . . . , N, with i (cid:54)= j:
with the inequality being trivially an equality for gij = 0.
aij ≤ κ
di
gij
(48)
(cid:3)
It is useful to make the following comparison between (35) and (48). Equation (35) means that the
nonzero entries in the combination matrix, scaled by N pN , do not collapse to zero, i.e., they are stable
from below. On the other hand, in view of (31), the upper bound in (48) implies that the nonzero entries
in the combination matrix, scaled by N pN , are asymptotically stable from above. This is because di
approaches N pN asymptotically as N → ∞ so that (48) translates into the asymptotic condition that
N pN aij ≤ κ.
Moreover, it is straightforward to verify that (48) holds for the Laplacian rule, with κ = (1 − µ)λ, as
well as for the Metropolis rule, with κ = 1 − µ.
Property P2. Consider a certain permutation of the agents, i.e., a renumbering of the rows and columns
of the interaction matrix G, which leads to the interaction matrix G. Such operation can be represented
by means of a permutation matrix P (see Appendix E), as G = P GP T . Property P2 holds if:
(49)
namely, if applying the combination policy γ(·) to the renumbered interaction matrix is equivalent to
(cid:3)
renumbering (through the same P ) the initial combination matrix A = γ(G).
γ(P GP T ) = P γ(G) P T ,
The meaning of property P2 is illustrated in Fig. 3. In the leftmost panel, we represent a network
graph, along with the corresponding combination matrix obtained by applying the Metropolis rule. In the
June 26, 2021
DRAFT
19
(50)
Fig. 3.
Illustration of property P2: permuting the agents, leads to a permuted version of the combination policy.
rightmost panel, the agents are exchanged as detailed in the figure, using the permutation matrix:
P =
0 0 0 1
0 1 0 0
1 0 0 0
0 0 1 0
Then, the Metropolis rule is applied to the new (i.e., renumbered) interaction matrix. It is seen that
the resulting combination matrix corresponds to renumbering the original combination matrix using the
permutation (50).
Property P2 is particularly relevant for the following reasons. First, under the Erdos-Renyi model,
the statistical distribution of the interaction matrix is invariant to agents' permutations. Owing to (49),
such invariance is automatically inherited by the combination matrix. Moreover, property P2 is satisfied
by typical combination policies, such as the Laplacian rule and the Metropolis rule. It is also useful to
provide a counterexample that shows why property P2 is not always verified. Consider a network with
three agents, and with the following interaction matrix:
1 1 0
1 1 1
G =
0 1 1
(51)
The combination policy of our counterexample works as follows. First, a Metropolis rule is applied.
Second, the resulting self-combination weight assigned to agent 1 is slightly increased by adding a small
June 26, 2021
DRAFT
1 ! 3 2 ! 2 3 ! 4 4 ! 1 3 2 4 1 1 2 3 4 1/4 1/4 1/4 1/4 1/4 3/4 0 0 1/4 0 3/4 0 1/4 0 0 3/4 A=γ(G) 3/4 0 1/4 0 0 3/4 1/4 0 1/4 1/4 1/4 1/4 0 0 1/4 3/4 1 1 1 1 1 1 0 0 1 0 1 0 1 0 0 1 G G=PGPT !""#$%&'()*+,%&-.*&(/0#1(~(1 0 1 0 0 1 1 0 1 1 1 1 0 0 1 1 21/+0.&'(-'1&34()*%&)%514(6%37(PAPT((γ(G) ~((1-µ) (1-µ) extra-weight. The extra-weight assigned to agent 1 is then subtracted, in equal parts, from the other
(cid:96)(cid:54)=1 a1(cid:96) = 1− µ. Finally, the other entries
of the combination matrix are adjusted so as to make A/(1 − µ) symmetric and right-stochastic. The
final result is (with (cid:28) 1):
0
1/3 − 1/3 + 1/3
2/3
1/3
0
A = (1 − µ)
nonzero weights of agent 1, in order to meet the requirement(cid:80)
2/3 + 1/3 −
1 1 1
1/3 +
1/3 − /2 2/3 + /2
1/3 − /2
A = (1 − µ)
matrix:
G =
1 1 0
1 0 1
1/3 − /2 1/3 − /2
0
2/3 + /2
0
20
(52)
(53)
(54)
Assume now that agents 1 and 2 are exchanged, which would yield the following interaction matrix:
Applying the combination policy described before, we would end up with the following combination
We see that: i) the interaction matrix in (53) is a renumbered version of the interaction matrix in (51),
which corresponds to exchanging agents 1 and 2, while ii) the combination matrix in (54) is not obtained
by applying the same renumbering to the combination matrix in (52). Therefore, property P2 is violated.
The presented counterexample shows that, while in practice constructing a combination policy that violates
property P2 might look rather artificial, from a purely theoretical viewpoint it must be stated that property
P2 is not always verified.
A combination policy satisfying the additional properties P1 and P2 describes the following class of
combination policies.
Combination-policy class C (cid:48)
possesses properties P1 and P2.
τ . A combination policy belongs to class C (cid:48)
τ if it belongs to class Cτ and
(cid:3)
Proposition 1, which is stated below, will be proved by resorting to an approximation that will be
referred to as the independence approximation. More specifically, the proof of the proposition will rely
on characterizing the variance of the entries in the error matrix (see Appendix D). Unfortunately, the
exact evaluation of the error variance is generally a formidable task. In order to gain insight into the
asymptotic behavior of the variance, in the proof of Proposition 1 we make a simplified evaluation by
treating the entries of the various matrices involved in the calculations as independent random variables.
The accuracy of the results arising from this approximation is tested by means of numerical experiments
June 26, 2021
DRAFT
21
-- see Fig. 8 in Appendix D. A more rigorous analysis would require estimating the order of the error
introduced by the independence approximation.
Proposition 1 (Rate of convergence of 1−F0()): Let the network interaction profile obey a G (cid:63)(N, pN )
model, and let the combination policy belong to class C (cid:48)
τ . Then, the result in (44) holds true because
τ ⊂ Cτ . In addition, for any nonzero fraction of observable agents satisfying (7), and for all > 0, we
C (cid:48)
have:
1 − F0() ≈ op(pN )
(55)
where the approximation in (55) arises from the independence approximation used in the proof.
Proof: See Appendix D.
VII. ILLUSTRATIVE EXAMPLES
We now examine the practical significance of the asymptotic results derived in the previous sections,
with reference to three combination matrices that are rather popular in the literature of adaptive networks.
The first two strategies lead to symmetric combination matrices, which therefore match the hypotheses
of our theorems. The third strategy corresponds to an asymmetric combination matrix. Even if the
asymmetric case is not contemplated by our theorems, it is relevant in practice and, as we shall see from
the forthcoming experiments, consistent tomography works also for such a relevant case. The presentation
of the examples is organized through the following steps.
S1) We consider first the case that the projections of the correlation matrices, R0 and R1, onto the
observable part of the network, are available without error. For this case, we compute the inversion
of the observable part, which leads to the matrix A(obs), through:
A(obs) = R(obs)
1
(R(obs)
0
)
−1
(56)
S2) We use the off-diagonal entries of A(obs), and apply a k-means clustering algorithm in order to
split these entries into two clusters. The cluster with smallest arithmetic average is labeled as "non-
interacting", while the other cluster is labeled as "interacting". We remark that such classification
strategy is implemented in a universal, fully non-parametric way.
S3) Then, we enlarge the setting to the case that the projections of the correlation matrices are estimated
from the diffusion outputs. In the simulations, the observations fed into the diffusion algorithm,
{xi(n)}i,n, follow a standard normal distribution. As an estimator for R(obs)
(cid:44) [R0]Ω, we use the
empirical correlation, namely:
0
June 26, 2021
R(obs)
0 =
Y Y T ,
1
n
(57)
DRAFT
where, for i = 1, 2, . . . , K, the i-th row of the K × n matrix Y is given by:
yωi(1), yωi(2), . . . , yωi(n),
22
(58)
with the indices ω1 < ω2 < ··· < ωK spanning the observable set Ω. The estimate R(obs)
similarly. We remark that in this work we do not focus on optimizing the performance of the
is computed
1
correlation matrix estimators, since our focus is on ascertaining the fundamental limits of tomography.
There are already considerable works in the literature on perfecting correlation estimation from
ensemble data. One challenge that arises in the network case is the interplay between the matrix
size and the number of samples used for the estimation.
Now, using (56) with the true correlation matrices replaced by their estimated counterparts, we get
the following estimate:
A(obs) = R(obs)
1
( R(obs)
0
)
−1
(59)
S4) We run the k-means clustering algorithm over the entries of A(obs) in (59).
A. Laplacian Combination Rule
Under the Laplacian combination rule, the off-diagonal combination weights are zero when the agents
are not interacting, and are otherwise equal to a constant across the network. Therefore, we see that
the weights reflect perfectly the structure of the underlying network graph. In fact, several important
properties of the graph are encoded in the properties of the Laplacian matrix [21].
In Fig. 4, leftmost panel, we display the off-diagonal entries of the (scaled) true combination matrix,
N pN A(obs), corresponding to the observable network portion. For clarity of presentation, the matrix
has been vectorized by means of column-major ordering, and the (vectorized) (i, j) pairs have been
rearranged in such a way that the zero entries appear before the nonzero entries. The same ordering used
for A(obs) will be then applied to the matrices displayed in the remaining panels, i.e., the horizontal axis
is homogeneous across the different panels, so as to allow a fair comparison. If agents i and j do not
interact, the pertinent matrix entry is marked with a blue circle, whereas if they do interact, a red square
is used. The observed step-function behavior comes from the fact that, for the Laplacian combination
rule, the nonzero weights are constant across the network.
In the middle panel we focus on steps S1) and S2): we display the scaled estimated matrix, N pN A(obs),
, and we display the classification performed by
computed under perfect knowledge of R(obs)
the k-means algorithm. In the rightmost panel we focus instead on steps S3) and S4): we display the
scaled estimated matrix, N pN A(obs), computed with the empirical estimates R(obs)
display the classification performed by the k-means algorithm. In the latter two panels, matrix entries are
and R(obs)
, and we
1
and R(obs)
1
0
0
June 26, 2021
DRAFT
Laplacian combination rule
23
Fig. 4. Network tomography for the case of a Laplacian combination rule with parameter λ = 0.5 -- see Sec. VII-A. The
network consists of N = 100 agents, where only K = 20 agents are observable (ξ = 0.2). The interaction probability is
pN = 2 (ln N )/N ≈ 0.092, and the value of the step-size is µ = 0.1. Leftmost panel: the true combination matrix, A(obs), is
vectorized with column-major ordering, and the (vectorized) (i, j) pairs are rearranged in such a way that the zero entries come
before the nonzero entries (the same ordering is applied to the estimated matrix, A(obs), in the other two panels). The different
markers highlight the true interaction profile of the observable network portion. Middle panel: matrix A(obs), computed under
perfect knowledge of the steady-state correlation matrices of the observable diffusion output, see (56). The different markers
highlight the interaction profile as reconstructed by the k-means algorithm. Rightmost panel: same as middle panel, with matrix
A(obs) computed using the correlation matrices estimated empirically with n = 2 × 104 samples, see (57), (58), and (59). Inset
plots: interaction profiles represented through the corresponding network graphs. In the inset plots of the middle and of the
rightmost panels, erroneously detected edges (the edge is not present but it is "seen" by the tomography algorithm) are marked
in magenta, while missed edged (the edge is present but the tomography algorithm misses it) are marked in cyan.
marked with different colors and symbols, depending on the results of the k-means clustering algorithm:
blue-circle markers if agents i and j have been classified as non-interacting, and red-square markers if
agents i and j have been classified as interacting.
The network considered in Fig. 4 consists of N = 100 agents, where only K = 20 agents are observable
(ξ = 0.2). According to the connection properties of the Erdos-R´enyi model, see (26), the interaction
probability is chosen as pN = 2 (ln N )/N ≈ 0.092. The parameter of the Laplacian combination matrix
is λ = 0.5. As regards the diffusion algorithm, we choose a typical value for the step-size, i.e., µ = 0.1.
Let us begin with examining the output of steps S1) and S2), middle panel. As we can see, the
experiments confirm in toto what is expected from the theoretical analysis: the entries of the matrix
A(obs) appear to be well-separable, since: i) the unavoidable error induced by partial observation of the
network, is relatively small, implying that zero entries of A(obs) are concentrated around zero; ii) the
nonzero entries of the true combination matrix (leftmost panel) are bounded from below, and, since the
error eij is positive, this fact keeps the nonzero entries of A(obs) (middle panel) sufficiently far away from
the origin.
June 26, 2021
DRAFT
50100150200250300350−0.200.20.40.60.81(i,j)(vectorized-rearranged)NpNa(obs)ijTrue interaction profile non-interactinginteracting50100150200250300350−0.200.20.40.60.81(i,j)(vectorized-rearranged)NpNa(obs)ijClustering with known corr. mat. classifiedasnon-interactingclassifiedasinteracting50100150200250300350−0.200.20.40.60.81(i,j)(vectorized-rearranged)NpNa(obs)ijClustering with unknown corr. mat. classifiedasnon-interactingclassifiedasinteracting Next we move on to steps S3) and S4), namely, we move on to examining the behavior in the presence
of imperfect knowledge of the correlation matrices. In particular, we use n = 2× 104 samples to perform
estimation of the correlation matrix from the diffusion output. This situation is examined in the rightmost
24
panel of Fig. 4. By comparison with the middle panel, we see that the estimated clusters are more "noisy",
which makes sense since the procedure must be affected by the error in estimating R(obs)
notwithstanding, tomography is still effective, meaning that the number of observations collected to
and R(obs)
0
. This
1
estimating the correlation matrices is sufficiently high. Few classification errors are committed: the red-
square markers appearing among the blue-circle markers, and the blue-circle marker appearing among
the red-square markers (approximately at position 350), represent mistakenly classified agent pairs.
The results arising from the above example are collected, perhaps in a more revealing form, in the inset
plots of Fig. 4. The displayed network graphs (corresponding only to the observable subnet) are drawn
with the following rules. An edge drawn from j to i means that agent i is influenced (leftmost panel)
or is estimated to be influenced (middle and rightmost panels) by agent j. When an edge is erroneously
detected (i.e., the edge is in fact not present but the tomography algorithm "sees" it), it is marked in
magenta. Likewise, when an edge is not detected (i.e., the edge is present but the tomography algorithm
misses it), it is marked in cyan. The impact of imperfect knowledge of the correlation matrices can be
noticed in the inset plot, where we see that some errors are committed in reconstructing the profile of
the observed subnet.
B. Metropolis Combination Rule
In Fig. 5, the analysis presented in the previous section is applied to a different combination matrix,
namely, the Metropolis combination matrix. This example is useful because, in a Metropolis rule,
the nonzero weights are no longer constant, and exhibit a certain dynamics related to the different
neighborhood sizes corresponding to the different agents. It is important to examine the impact of such
dynamics on the performance of the tomography algorithm. From the leftmost panel of Fig. 5, we see
that the nonzero entries of the true combination matrix exhibit a certain variability. This notwithstanding,
since the Metropolis rule belongs to class Cτ , with τ = (1 − µ)/e, the scaled nonzero entries are still
lower bounded in probability. The separability of the clusters is preserved, and considerations similar to
those drawn in the previous section as regards the Laplacian rule apply.
Still, there is an important difference between the two kinds of combination policies. By inspecting
carefully the rightmost panel in Fig. 5 and the related inset, we see that, differently from what happened in
the rightmost panel of Fig. 4, the network profile is reconstructed perfectly. This suggests that estimation
of the correlation matrices is easier for Metropolis rules than for Laplacian rules. The latter effect
June 26, 2021
DRAFT
Metropolis combination rule
25
Fig. 5. Network tomography for the case of a Metropolis combination rule addressed in Sec. VII-B. The network consists of
N = 100 agents, where only K = 20 agents are observable (ξ = 0.2). The interaction probability is pN = 2 (ln N )/N ≈ 0.092,
and the value of the step-size is µ = 0.1. Leftmost panel: the true combination matrix, A(obs), is vectorized with column-major
ordering, and the (vectorized) (i, j) pairs are rearranged in such a way that the zero entries come before the nonzero entries
(the same ordering is applied to the estimated matrix, A(obs), in the other two panels). The different markers highlight the true
interaction profile of the observable network portion. Middle panel: matrix A(obs), computed under perfect knowledge of the
steady-state correlation matrices of the observable diffusion output, see (56). The different markers highlight the interaction
profile as reconstructed by the k-means algorithm. Rightmost panel: same as middle panel, with matrix A(obs) computed using
the correlation matrices estimated empirically with n = 2×104 samples, see (57), (58), and (59). Inset plots: interaction profiles
represented through the corresponding network graphs. In the inset plots of the middle and of the rightmost panels, erroneously
detected edges (the edge is not present but it is "seen" by the tomography algorithm) are marked in magenta, while missed
edged (the edge is present but the tomography algorithm misses it) are marked in cyan.
could be probably ascribed to the improved convergence properties of Metropolis combination matrices
(with respect to Laplacian combination matrices), which should allow a more accurate estimation of the
correlation matrices for a given number of samples.
C. Uniform Averaging Rule
One of the most important goals of causation is ascertaining whether one agent influences another agent.
Such influence is not necessarily symmetrical. Although we have obtained an analytical characterization
of the error matrix for the case of symmetric combination matrices, it is useful to examine whether the
conclusions drawn for the symmetric case find some correspondence in the asymmetric case. To this aim,
the network interaction profile is now generated according to a binomial random graph model, B(N, pN ),
where the variables gij, for i = 1, 2, . . . , N and j (cid:54)= i, are independent Bernoulli random variables with
P[gij = 1] = pN -- see [37]. Such a procedure leads (in general) to an asymmetric matrix G, i.e., to a
directed graph. As regards the nonzero entries of the combination matrix, we consider the well-known
uniform averaging rule (where the weight that an agent assigns to the data of its interacting neighbors
is inversely proportional to the size of its neighborhood), which leads in general to an asymmetric
June 26, 2021
DRAFT
5010015020025030035000.511.52(i,j)(vectorized-rearranged)NpNa(obs)ijTrue interaction profile non-interactinginteracting5010015020025030035000.511.52(i,j)(vectorized-rearranged)NpNa(obs)ijClustering with known corr. mat. classifiedasnon-interactingclassifiedasinteracting5010015020025030035000.511.52(i,j)(vectorized-rearranged)NpNa(obs)ijClustering with unknown corr. mat. classifiedasnon-interactingclassifiedasinteracting Uniform averaging combination rule
26
Fig. 6. Network tomography for the case of a uniform averaging combination rule addressed in Sec. VII-C. The network consists
of N = 100 agents, where only K = 20 agents are observable (ξ = 0.2). The interaction probability is pN = 2 (ln N )/N ≈
0.092, and the value of the step-size is µ = 0.1. Leftmost panel: the true combination matrix, A(obs), is vectorized with
column-major ordering, and the (vectorized) (i, j) pairs are rearranged in such a way that the zero entries come before the
nonzero entries (the same ordering is applied to the estimated matrix, A(obs), in the other two panels). The different markers
highlight the true interaction profile of the observable network portion. Middle panel: matrix A(obs), computed under perfect
knowledge of the steady-state correlation matrices of the observable diffusion output, see (56). The different markers highlight
the interaction profile as reconstructed by the k-means algorithm. Rightmost panel: same as middle panel, with matrix A(obs)
computed using the correlation matrices estimated empirically with n = 2 × 104 samples, see (57), (58), and (59). Inset plots:
interaction profiles represented through the corresponding network graphs. In the inset plots of the middle and of the rightmost
panels, erroneously detected edges (the edge is not present but it is "seen" by the tomography algorithm) are marked in magenta,
while missed edged (the edge is present but the tomography algorithm misses it) are marked in cyan.
combination matrix. The results of our experiments are reported in Fig. 6, with reference to the case of
perfect and imperfect knowledge of the correlation matrices, respectively. Remarkably, the tomography
algorithm exhibits proper operation even in the asymmetric case. From the inset plots, we can appreciate
that, differently from the previous examples, it is now possible that the influence of an agent over another
agent is in general unidirectional.
D. Role of the Step-Size
The step-size, µ, is a fundamental parameter ruling the performance of the adaptive diffusion algorithm
implemented by the network agents. As examined in detail in recent works [17], [18], the following
adaptation/learning trade-off arises: small values of µ enhance learning precision, while relatively high
values of µ enhance adaptation.
We now examine the impact of the step-size on the tomography algorithm. For the sake of concreteness,
we focus on one of the combination rules examined in the previous sections, namely, the Metropolis rule,
and consider the noiseless case, namely, steps S1) and S2). In the previous simulations concerning the
Metropolis rule (see Fig. 5), we have used µ = 0.1. In the new simulation, reported in Fig. 7 (leftmost
June 26, 2021
DRAFT
5010015020025030035000.511.522.53(i,j)(vectorized-rearranged)NpNa(obs)ijTrue interaction profile non-interactinginteracting5010015020025030035000.511.522.53(i,j)(vectorized-rearranged)NpNa(obs)ijClustering with known corr. mat. classifiedasnon-interactingclassifiedasinteracting5010015020025030035000.511.522.53(i,j)(vectorized-rearranged)NpNa(obs)ijClustering with unknown corr. mat. classifiedasnon-interactingclassifiedasinteracting The effect of different system parameters
Parameter µ
Parameter ξ
Parameter pN
27
Fig. 7. Network tomography for the case of a Metropolis combination rule addressed in Sec. VII-B, and for a network with
N = 100 agents. Leftmost panel: the case of "high" values for the step-size addressed in Sec. VII-D. The relevant parameters
are: µ = 0.5, ξ = 0.2, and pN = 2 (ln N )/N ≈ 0.092. Middle panel: the case of "small" fractions of observed agents
addressed in Sec. VII-E. The relevant parameters are: ξ = 0.1, pN ≈ 0.092, and µ = 0.1. Rightmost panel: the case of "high"
values for the interaction probability addressed in Sec. VII-F. The relevant parameters are: pN ≈ 0.46, ξ = 0.2, and µ = 0.1.
Lowermost panels: In all panels, the matrix A(obs) (displayed with the same ordering criterion used in the previous figures)
is computed under perfect knowledge of the steady-state correlation matrices of the observable diffusion output, see (56). The
different markers highlight the interaction profile as reconstructed by the k-means algorithm. Inset plots: interaction profiles
represented through the corresponding network graphs. In the inset plots of the lowermost panels, erroneously detected edges
(the edge is not present but it is "seen" by the tomography algorithm) are marked in magenta, while missed edged (the edge is
present but the tomography algorithm misses it) are marked in cyan.
panel), we run the diffusion algorithm with µ = 0.5, which turns out to be a relatively high value for
the step-size in practical applications.
By joint inspection of Fig. 5 (middle panel) and Fig. 7 (leftmost panel), we can appreciate the presence
of different effects. The first effect relates to the amplitude of the nonzero entries. We see that increasing
µ corresponds to diminishing the amplitude of the nonzero entries in the combination matrix. This effect
is rather obvious, since our definition of A embodies a scaling factor 1− µ, see (4). The reduction of the
distance between zero and nonzero entries could have a negative impact on tomography performance,
especially when the correlation matrices must be estimated from the diffusion output, and the additional
error associated to such estimation reduces further the separation between the two classes.
The second effect relates to the amplitude of the entries in the error matrix. We see that increasing µ
corresponds to diminishing the amplitude of the errors, which could be beneficial for tomography. This
effect can be grasped by Theorem 1, since we see that the (columnwise) sum of errors is upper bounded
by 1− µ. However, Theorem 1 provides only an upper bound, and, hence, the exact behavior of the error
as a function of µ is not known.
June 26, 2021
DRAFT
5010015020025030035000.511.52(i,j)(vectorized-rearranged)NpNa(obs)ijClustering with known corr. mat. classifiedasnon-interactingclassifiedasinteracting10203040506070809000.511.52(i,j)(vectorized-rearranged)NpNa(obs)ijClustering with known corr. mat. classifiedasnon-interactingclassifiedasinteracting5010015020025030035000.511.522.533.5(i,j)(vectorized-rearranged)NpNa(obs)ijClustering with known corr. mat. classifiedasnon-interactingclassifiedasinteracting 28
As a result, it is not easy to anticipate which of the aforementioned two effects will prevail, and
elucidating the trade-offs related to the effect of the step-size on the tomography performance, in
connection to the adaptation/learning trade-off, is an interesting issue that deserves further investigation.
E. Degree of Observability
It is not unreasonable to presume that the error induced by partial observation of the network depends
inversely on the observable fraction of agents, ξ. Accordingly, it is expected that the error grows when the
size of the monitored portion of the network decreases, i.e., as ξ decreases. This effect is confirmed by our
numerical experiments. A sample of such experiments is reported in Fig. 7 (middle panel). In particular,
by comparison with Fig. 5 (middle panel), it is observed that the averages of the errors shift upward for
both the zero and the nonzero entries, and that the error's amplitude is generally increased. However,
even in the presence of higher errors, separability between the two clusters seems to be preserved.
F. "High" Interaction Probability
The theoretical analysis conducted in this work focused on a classic asymptotic regime for random
graphs, namely, the regime where the interaction probability goes to zero as the network size increases.
We were able to conclude that, as N → ∞, the (scaled) error, N pN eij, is driven to zero in probability.
On the other hand, in practice one obviously operates with fixed (finite) values of N and nonzero
values of pN , and, hence, the error will be never exactly equal to zero. In particular, for a fixed value
of N, a higher value of pN (which, in the asymptotic setting described by (26), would correspond to
select a sequence cN diverging faster), is expected to imply higher values of N pN eij. Such features are
observed in Fig. 7, rightmost panel. Comparison with the middle panel of Fig. 5 (where the interaction
probability was smaller) reveals that the error in Fig. 7 is, on average, higher.
However, there is an important aspect that emerges here. We see clearly that the average increase of the
error does not impair the capability of the system to discriminate between interacting and non-interacting
agent pairs. Clustering is still effective, since the two clusters are merely shifted upward. This suggests
the following important observation. In this case, what matters is the variance of the error, which should
be sufficiently small in order to guarantee separability of the clusters.
VIII. CONCLUDING REMARKS AND OPEN ISSUES
We considered the problem of tomography over partially observed diffusion networks, where the goal
is to infer whether an agent is directly influenced by another agent. Under the challenging scenario where
only the output of the diffusion algorithm is observable (i.e., there is no direct information about who's
June 26, 2021
DRAFT
29
talking to whom) and only a fraction of the agents is accessible, we prove that consistent tomography
is possible: as the network size increases, the interacting and non-interacting agent pairs split in two
separate clusters, for any given fraction of accessible agents. The theoretical results are proved under
some technical assumptions: i) the network graph is an Erdos-R´enyi graph; ii) the combination matrices
belong to a certain symmetric class; iii) perfect knowledge is assumed for the correlation matrices
of the observable diffusion output (steady-state regime). As another contribution, we show by means
of numerical experiments that the results hold for asymmetric matrices as well (which is useful in the
context of causal inference), and that the theoretical correlation matrices can be replaced by their empirical
counterparts, estimated from the diffusion outputs observed for a sufficiently long time.
There are many interesting questions and open issues that deserve future investigation. First, it would
be useful to generalize the theoretical results to broader settings, including: the case of asymmetric
combination matrices; graph generative models different from the Erdos-R´enyi model; different classes
of combination matrices.
We see from the numerical analysis that the separation between the agents' clusters is very marked,
suggesting that vanishing of the error (which, in this work, has been established in terms of the fraction of
erroneously classified agents) might hold under more restrictive criteria, e.g., a maximum error criterion,
where we require that the largest entry in the error matrix vanishes. It would be also of interest to
see whether the characterization of the result in Proposition 1 can be obtained without resorting to the
independence approximation. Moreover, we notice that the additive structure of the entries in the error
matrix encourages to pursue an asymptotic normality characterization.
From a more practical perspective, an interesting problem is optimizing the algorithms for tomography
(exploiting, when possible, further structural properties, such as sparsity), which requires careful managing
of the interplay between network size and number of samples collected for the inferential task.
Proof of Theorem 1:
In view of (11) and (16), the steady-state zero-lag and one-lag correlation
matrices can be written as:
APPENDIX A
R0 = µ2Z
−1, R1 = µ2AZ
−1,
(60)
where we let Z = IN − A2. Applying the rules for multiplication between partitioned matrices [32,
p. 17], the observable part of the one-lag correlation matrix is given by:
−1]Ω(cid:48)Ω,
−1]Ω + µ2AΩΩ(cid:48)[Z
[R1]Ω = µ2AΩ[Z
(61)
June 26, 2021
DRAFT
where Ω(cid:48) = {1, 2, . . . , N} \ Ω denotes the complement of Ω. Therefore, using (60) and (61), the matrix
A(obs) in (15) can be written as:
30
A(obs) = (AΩ[Z
−1]Ω + AΩΩ(cid:48)[Z
−1]Ω(cid:48)Ω([Z
−1]Ω(cid:48)Ω)([Z
−1.
−1]Ω)
−1]Ω)
−1
= AΩ + AΩΩ(cid:48)[Z
Now, from the matrix inversion lemma we have that [32, p. 18]:
−1ZΩ(cid:48)Ω[Z
−1]Ω(cid:48)Ω = −(ZΩ(cid:48))
[Z
−1]Ω.
Substituting (63) into (62), and making explicit the definition of Z, we get:
A(obs) = AΩ − AΩΩ(cid:48)(ZΩ(cid:48))
−1ZΩ(cid:48)Ω
= AΩ − AΩΩ(cid:48)( [IN − A2]Ω(cid:48)
IN−K−[A2]Ω(cid:48)
(cid:123)(cid:122)
(cid:124)
(cid:125)
−1 [IN − A2]Ω(cid:48)Ω
)
(cid:124)
(cid:123)(cid:122)
−[A2]Ω(cid:48) Ω
(62)
(63)
(64)
,
(cid:125)
which, in view of (17), and noting that AΩ = A(obs), gives the following expression for the error matrix
E:
E = AΩΩ(cid:48)(IN−K − [A2]Ω(cid:48))
−1[A2]Ω(cid:48)Ω.
By introducing the following matrices:
B (cid:44) A2, H (cid:44) (IN−K − BΩ(cid:48))
−1,
the error matrix can be represented as:
E = AΩΩ(cid:48)HBΩ(cid:48)Ω.
Accordingly, the entries of the error matrix, for i, j = 1, 2, . . . , K, take on the form:
N−K(cid:88)
(cid:96),m=1
eij =
aωiω(cid:48)
(cid:96) h(cid:96)m bω(cid:48)
mωj ,
(65)
(66)
(67)
(68)
where indices ω1 < ω2 < ··· < ωK span the observable set Ω, while indices ω(cid:48)
N−K
span the complement set Ω(cid:48). We remark that the error matrix E is a nonnegative matrix (i.e., eij ≥ 0
for all i, j = 1, 2, . . . , K), because all terms in the summation (68) are nonnegative. In fact, AΩΩ(cid:48) and
B = A2 are nonnegative matrices since A is nonnegative. With regard to H, from (66) it can be expressed
2 < ··· < ω(cid:48)
1 < ω(cid:48)
∞(cid:88)
as:
H = (IN−K − BΩ(cid:48))
−1 =
and, hence, H is nonnegative because so is BΩ(cid:48).
k=0
June 26, 2021
(BΩ(cid:48))k,
(69)
DRAFT
Let us now consider the (columnwise) sum of the errors, which, in view of (68), can be written as:
K(cid:88)
N−K(cid:88)
K(cid:88)
eij =
aωiω(cid:48)
(cid:96) h(cid:96)m
bω(cid:48)
mωj .
j=1
(cid:96),m=1
j=1
It is convenient to introduce the following auxiliary matrix:
F (cid:44) HBΩ(cid:48)Ω,
whose entries, for (cid:96) = 1, 2, . . . , N − K, and j = 1, 2, . . . , K, are accordingly:
With such definition, the matrix entry eij in (68) becomes:
f(cid:96)j =
h(cid:96)mbω(cid:48)
mωj .
eij =
aωiω(cid:48)
(cid:96)f(cid:96)j.
m=1
N−K(cid:88)
N−K(cid:88)
K(cid:88)
(cid:96)=1
31
(70)
(71)
(72)
(73)
(74)
We start by proving that:
Recalling the definition of the matrix B in (66), we have that(cid:80)N
j=1
f(cid:96)j ≤ 1
write:
j=1 bmj = (1 − µ)2. Therefore, we can
K(cid:88)
j=1
bω(cid:48)
mω(cid:48)
j
bω(cid:48)
mωj = (1 − µ)2 − N−K(cid:88)
= 1 − µ(2 − µ) − N−K(cid:88)
N−K(cid:88)
(cid:124)
= −µ(2 − µ) +
j=1
j=1
j=1
bω(cid:48)
mω(cid:48)
j
(cid:123)(cid:122)
(δmj − {BΩ(cid:48)}mj)
={H−1}mj
(cid:125)
,
(75)
and, hence, from (72) we get:
K(cid:88)
j=1
f(cid:96)j = −µ(2 − µ)
N−K(cid:88)
j=1
N−K(cid:88)
(cid:124)
m=1
h(cid:96)m +
h(cid:96)m{H
(cid:123)(cid:122)
δ(cid:96)j
−1}mj
(cid:125)
N−K(cid:88)
N−K(cid:88)
m=1
= −µ(2 − µ)
h(cid:96)m + 1 ≤ 1,
(76)
where the inequality holds since H is a nonnegative matrix, and since µ(2− µ) > 0 because 0 < µ < 1.
Using now (73), we can write:
m=1
K(cid:88)
N−K(cid:88)
eij =
aωiω(cid:48)
(cid:96)
N−K(cid:88)
f(cid:96)j ≤ N−K(cid:88)
j=1
(cid:96)=1
j=1
(cid:96)=1
June 26, 2021
≤ 1 − µ,
aωiω(cid:48)
(cid:96)
(77)
DRAFT
where i) the first inequality comes from (74) (since A is a nonnegative matrix), and ii) the second
inequality holds since(cid:80)N
(cid:96)=1 ai(cid:96) = 1 − µ.
32
APPENDIX B
Proof of Lemma 1: We observe that for any nonnegative real number d and i (cid:54)= j, it holds that:
P[d(cid:96) ≥ d gij = 1]
P[dmax ≥ d gij = 1] ≤ N(cid:88)
(cid:96)=1
(cid:88)
=
P[d(cid:96) ≥ d gij = 1] +
(cid:96)(cid:54)=i,j
P[di ≥ d gij = 1] + P[dj ≥ d gij = 1],
(78)
where the inequality is an application of the union bound. Now, for (cid:96) (cid:54)= i, j, we have that d(cid:96) is independent
of gij, and that d(cid:96) = 1 + β(N − 1, pN ), implying that P[d(cid:96) ≥ d gij = 1] = P [1 + β(N − 1, pN ) ≥ d].
On the other hand, for (cid:96) = i or (cid:96) = j, we have that, conditioned on the event gij = 1, the degree d(cid:96) is
equal to 2 (because gii = 1 and gij = 1) plus the number of neighbors arising from the remaining N − 2
agents, which yields d(cid:96) = 2 + β(N − 2, pN ). Therefore, it is legitimate to write the following chain of
inequalities, for η ∈ R and t > 0:
P[dmax ≥ ηN pN gij = 1]
≤ (N − 2) P [1 + β(N − 1, pN ) ≥ ηN pN ] +
≤ N e
2 P [2 + β(N − 2, pN ) ≥ ηN pN ]
−ηN pN t E[e(1+β(N−1,pN ))t] +
−ηN pN t E[e(2+β(N−2,pN ))t]
2 e
= N et e
2e2t e
−ηN pN t[1 + pN (et − 1)]N−1 +
−ηN pN t[1 + pN (et − 1)]N−2
≤ (N et + 2e2t) e
−N pN (ηt+1−et),
(79)
where: i) the first inequality follows from (78); ii) the second inequality is a classic Chernoff bound3; iii)
the equality follows by evaluating explicitly the moment generating function, E[eβ(N,p)t], of the binomial
3For a random variable x, and a positive t, the Chernoff bound is given by P[x ≥ η] ≤ E[etx]/etη.
June 26, 2021
DRAFT
random variable4; iv) the latter inequality follows from (1 + x)N ≤ eN x (which holds true for x > 0),
having further replaced the exponents N − 1 and N − 2 with the (higher) exponent N. Now, maximizing
the exponent ηt − et (w.r.t. t) yields the solution et = η, which in turn implies:
−N pN (η ln η+1−η).
P[dmax ≥ ηN pN gij = 1] ≤ (ηN + 2η2)e
(80)
33
Using now the particular value η = e yields:
(cid:19)
P[dmax ≥ eN pN gij = 1] ≤ (e N + 2e2)e
(cid:18)
≤
e +
2e2
N
−N pN
−cN ,
e
where, in the latter inequality, we used the fact that pN = (ln N + cN )/N.
Proof of Theorem 2: The empirical distributions F0(α) and F1(α) in (43) can be written as:
APPENDIX C
F0(α) =
F1(α) =
N0(α)
N0 + I{N0 = 0} +
N1(α)
N1 + I{N1 = 0} +
1
2
1
2
I{N0 = 0},
I{N1 = 0},
(81)
(82)
(83)
where we have simply used a formal representation to state that F0(α)) (resp., F1(α)) are set to 1/2
when N0 = 0 (resp., N1 = 0).
We start by proving that F0()
p−→ 1 for all > 0. From (82) we can write:
1 − F0() =
N0 − N0()
N0 + I{N0 = 0} +
I{N0 = 0}.
1
2
(84)
Using (41), and since a(obs)
ij = a(obs)
ij + eij, from (84) we can write:
N0 − N0()
N0 + I{N0 = 0}
I{N pN eij > , g(obs)
j(cid:54)=i
N0 + I{N0 = 0}
I{N pN eij > }
j(cid:54)=i
(cid:80)K
(cid:80)K
(cid:80)
(cid:80)
i=1
i=1
=
ij = 0}
≤
≤
=
N0 + I{N0 = 0}
1 − µ
KN pN
(cid:18)
N0 + I{N0 = 0}
K2
N0 + I{N0 = 0}
(cid:19) N
1 − µ
K
pN ,
(85)
4The moment generating function of a Bernoulli random variable with success probability p is given by pet + (1 − p). Since
a binomial random variable β(N, p) is the sum of N independent Bernoulli variables with success probability p, we have
E[eβ(N,p)t] = [pet + (1 − p)]N .
June 26, 2021
DRAFT
34
where the latter inequality follows by applying Corollary 1. Now, the asymptotic properties of N0 can be
deduced from the asymptotic properties of binomial random variables. Indeed, since for the Erdos-R´enyi
model, gij = gji, we have:
N0 = 2
K(cid:88)
(cid:88)
(1 − g(obs)
ij
).
(86)
i=1
j<i
form, for j < i, a collection of L = K(K − 1)/2 independent
Moreover, since the variables 1 − g(obs)
Bernoulli random variables with parameter 1 − pN , we have that N0 is distributed as 2β(L, 1 − pN ).
Now, in view of (7), K (and, hence, L) diverges as N goes to infinity. Since by assumption pN vanishes
as N goes to infinity, the product L(1 − pN ) diverges as well, and we can invoke (31) to conclude that:
ij
N0
L(1 − pN )
(87)
Since I{N0 = 0} ∈ {0, 1}, using (87) and (7) into (85), and recalling that pN vanishes by assumption,
we have in fact proved that the first term on the RHS in (84) goes to zero as N → ∞. With regard to
the second term on the RHS in (84), it converges to zero in mean (and, hence, in probability [38]), since
p−→ 2 ⇒ N0
K2
p−→ 1.
we have:
E[I{N0 = 0}] = P[all observed agents are interacting]
= (pN )L N→∞−→ 0.
We have in fact proved that F0()
p−→ 1 for all > 0.
We now show that ¯F1(τ )
(cid:80)
p−→ 1. Using (42), from (83) we can write:
ij ≤ τ, g(obs)
(cid:123)(cid:122)
I{N pN a(obs)
K(K − 1)pN
ij = 1}
(cid:125)
j(cid:54)=i
i=1
×
F1(τ ) =
K(K − 1)pN
N1 + I{N1 = 0}
T1
1
2
+
(cid:125)
I{N1 = 0}.
(cid:80)K
(cid:124)
(cid:124)
(cid:80)K
(cid:123)(cid:122)
(cid:80)
T2
(88)
(89)
With regard to the term T1, we have:
where we used the fact that P[g(obs)
we have that:
E[T1] =
i=1
j(cid:54)=i
ij ≤ τg(obs)
ij = 1]
P[N pN a(obs)
K(K − 1)
ij = 1] = pN . Moreover, since a(obs)
ij = a(obs)
(90)
,
ij + eij, and since eij ≥ 0,
ij ≤ τg(obs)
P[N pN a(obs)
≤ P[N pN a(obs)
ij = 1]
ij ≤ τg(obs)
ij = 1].
(91)
DRAFT
June 26, 2021
Combining (90) and (91), in view of (35) we can write:
E[T1] ≤ N
N→∞−→ 0.
35
(92)
Thus, we have shown that the term T1 appearing in (89) converges to zero in mean and, hence, in
probability [38]. Moreover, the term T2 in (89) goes to one in probability in view of (31), because: i) for
the G (cid:63)(N, pN ) model, N1 is distributed as 2β(L, pN ), where L = K(K − 1)/2; ii) the product LpN
diverges as N goes to infinity in view of (7), since by assumption the product N pN diverges; iii) the
term I{N1 = 0} is either 0 or 1. Finally, as regards the latter term in (89), it is straightforward to see
that it converges to zero in mean (and, hence, in probability [38]), since we have:
E[I{N1 = 0}] = P[all observed agents are not interacting]
= (1 − pN )L ≤ e
−LpN N→∞−→ 0,
(93)
where the inequality holds because pN ≤ 1.
APPENDIX D
The main result of this appendix is the proof of Proposition 1. We start by proving two auxiliary
lemmas.
Lemma 2: Let β(N, p) be a binomial random variable of parameters N and p. We have, for m = 1, 2:
E
1
(1 + β(N, p))m
≤ m
(N p)m .
(94)
Proof: We prove the claim for the case m = 2, since the proof for the case m = 1 follows similar
(cid:21)
(cid:21)
(1 + β(N, pN ))2
(cid:20)
(cid:19)
(cid:19)
(cid:19)
1
k
(cid:18)N
(cid:18)N
(cid:18)N
k
k
(cid:20)
E
steps. We have that:
=
k=0
N(cid:88)
N(cid:88)
≤ N(cid:88)
(cid:19)
Now, since we have (cid:18)N + 2
k=0
k=0
=
k + 2
June 26, 2021
pk(1 − p)N−k
1
(k + 1)2
pk(1 − p)N−k
1
(k + 1)(k + 2)
pk(1 − p)N−k
2
k + 2
k + 1
(cid:124)(cid:123)(cid:122)(cid:125)≤2
.
(k + 1)(k + 2)
(cid:18)N
(cid:19) (N + 1)(N + 2)
k
(k + 1)(k + 2)
=
(N + 2)!
(k + 2)!(N − k)!
=
,
(95)
(96)
DRAFT
we can write:
(cid:18)N
(cid:19)
k
N(cid:88)
k=0
pk(1 − p)N−k
(k + 1)(k + 2)
2
(cid:18)N + 2
(cid:19)
(N + 1)(N + 2)
k + 2
pk(1 − p)N−k
≤
≤
=
2
(N p)2
2
(N p)2
2
k=0
N(cid:88)
(cid:18)N + 2
(cid:19)
N(cid:88)
(cid:19)
(cid:18)N + 2
N +2(cid:88)
(cid:124)
k + 2
k=0
k=2
pk+2(1 − p)(N +2)−(k+2)
pk(1 − p)(N +2)−k
(cid:123)(cid:122)
k
P[β(N +2,p)≥2]<1
(cid:125)
<
2
(N p)2 ,
36
(97)
and the proof is complete.
We recall that E and V denote the expectation and variance operators, respectively. In the following, the
expectation, the variance and the mean-square value of a random variable z will be denoted respectively
by:
¯z (cid:44) E[z], σ2
z
(cid:44) V[z] = E[(z − ¯z)2], z2 (cid:44) E[z2] = σ2
z + ¯z2.
(98)
Moreover, given a sequence of identically distributed random variables, z1, z2, . . . , zL, we shall still
write:
¯z (cid:44) E[z(cid:96)], σ2
z
(cid:44) V[z(cid:96)] = E[(z(cid:96) − ¯z)2], z2 (cid:44) E[z2
(cid:96) ] = σ2
z + ¯z2,
(99)
since, owing to identical distribution across (cid:96), both the expectation and the variance do not depend upon
(cid:96).
Lemma 3: Let [u1, u2, . . . , uL] and [z1, z2, . . . , zL] be mutually independent random vectors, with the
variables z(cid:96), for (cid:96) = 1, 2, . . . , L, being mutually uncorrelated and identically distributed scalar random
variables across (cid:96). Let also:
Then:
Proof: We have:
(cid:96)=1
June 26, 2021
L(cid:88)
(cid:96)=1
s (cid:44)
u(cid:96)z(cid:96).
(cid:96)=1
L(cid:88)
L(cid:88)
L(cid:88)
L(cid:88)
(cid:96)=1
m=1
σ2
s = σ2
z
E[u2
(cid:96) ] + ¯z2 V
s2 =
=
u(cid:96)umz(cid:96)zm
L(cid:88)
(cid:88)
m(cid:54)=(cid:96)
(cid:96)=1
u2
(cid:96) z2
(cid:96) +
(cid:35)
u(cid:96)
(cid:34) L(cid:88)
(cid:96)=1
u(cid:96)umz(cid:96)zm.
(100)
(101)
(102)
DRAFT
Now, observing that:
we can rearrange (104) as:
E[s2] = (σ2
z + ¯z2)
E[u2
(cid:96) ] + ¯z2
E[u(cid:96)um].
L(cid:88)
L(cid:88)
(cid:96)=1
=
(cid:32) L(cid:88)
(cid:33)2
u(cid:96)
(cid:96)=1
(cid:96)=1
z
L(cid:88)
(cid:35)
(cid:34) L(cid:88)
(cid:96)=1
u(cid:96)
u2
(cid:96) +
u(cid:96)um,
(cid:88)
m(cid:54)=(cid:96)
(cid:96)=1
(cid:96)=1
m(cid:54)=(cid:96)
L(cid:88)
L(cid:88)
(cid:88)
(cid:32) L(cid:88)
(cid:34) L(cid:88)
(cid:32)
(cid:96)=1
E
u(cid:96)
(cid:33)2 .
(cid:35)(cid:33)2
u(cid:96)
,
E[s2] = σ2
E[u2
(cid:96) ] + ¯z2 E
37
(103)
(104)
(105)
(106)
(107)
(108)
(109)
DRAFT
Since [u1, u2, . . . , uL] and [z1, z2, . . . , zL] are mutually independent random vectors, and since the
variables z(cid:96) are mutually uncorrelated, by taking expectations we get:
L(cid:88)
(cid:96)=1
L(cid:88)
(cid:88)
m(cid:54)=(cid:96)
(cid:96)=1
E[s2] =
E[u2
(cid:96) ]E[z2
(cid:96) ] +
E[u(cid:96)um]E[z(cid:96)]E[zm].
Moreover, since the variables z(cid:96) are identically distributed, Eq. (103) yields:
On the other hand, from (100)we have:
¯s = ¯z E
⇒ ¯s2 = ¯z2
implying, in view of (106):
(cid:96)=1
(cid:96)=1
(cid:96) ] +
= σ2
z
s = E[s2] − ¯s2
L(cid:88)
σ2
E
E[u2
L(cid:88)
(cid:32) L(cid:88)
¯z2
(cid:96)=1
(cid:96)=1
= σ2
z
u(cid:96)
E[u2
(cid:96) ] + ¯z2 V
(cid:33)2 −
(cid:32)
(cid:34) L(cid:88)
E
u(cid:96)
(cid:35)(cid:33)2
u(cid:96)
(cid:34) L(cid:88)
(cid:35)
(cid:96)=1
.
(cid:96)=1
(cid:96)=1
Proof of Proposition 1: We start by observing that, using (84), we can write:
1 − F0()
pN
N0 − N0()
N0 + I{N0 = 0}
1
pN
+
1
2
=
I{N0 = 0}
pN
.
June 26, 2021
In view of (88), the second term appearing on the RHS in (109) converges to zero in mean, and, hence,
in probability [38]. With regard to the first term appearing on the RHS in (109), from the first inequality
in (85) we have:
38
.
(110)
(cid:80)K
i=1
The fact that T4 converges to 1 in probability has been already ascertained in the proof of Theorem 2
-- see the rightmost relationship in (87). With regard to T3, by taking expectations we get:
P[N pN eij > ]
j(cid:54)=i
K(K − 1)pN
E[T3] =
(111)
where we used the fact that, in view of (151), the random variables eij, for i (cid:54)= j, are identically
distributed. This is because any renumbering leaves unaltered the statistical distribution of matrix E. In
addition, identical distribution across the K − 1 agents distinct from agent i, implies in particular the
following identity for expectations:
pN
=
,
P[N pN eij > ]
(112)
j(cid:54)=i
Multiplying the expected error by the factor N pN , we get:
E[eij] =
1
K − 1
E[eij].
N pNE[eij] = N pN
N0 − N0()
N0 + I{N0 = 0}
j(cid:54)=i
≤
i=1
(cid:80)
(cid:80)
i=1
=
1
pN
I{N pN eij > }
N0 + I{N0 = 0}
I{N pN eij > }
j(cid:54)=i
K(K − 1)pN
(cid:123)(cid:122)
T3
1
pN
×
(cid:125)
K(K − 1)
N0 + I{N0 = 0}
(cid:125)
(cid:123)(cid:122)
T4
(cid:80)K
(cid:80)K
(cid:124)
(cid:124)
(cid:80)
(cid:88)
1
K − 1
(cid:88)
(cid:88)
j(cid:54)=i
eij
E[eij]
=
N pN
K − 1
E
≤ (1 − µ)
j(cid:54)=i
N
K − 1
N→∞−→ 0,
pN
(113)
where the inequality follows from Theorem 1, while the convergence follows from (7) and from the fact
June 26, 2021
DRAFT
that pN vanishes. Now, we can write, for i, j = 1, 2, . . . , K, and i (cid:54)= j:
P[N pN eij > ]
= P[(N pN eij − N pNE[eij]) > ( − N pNE[eij])]
≤ P[(N pN )2(eij − E[eij])2 > ( − N pNE[eij])2]
≤
( − N pNE[eij])2 ≤ 4
2 (N pN )2σ2
e ,
(N pN )2σ2
e
39
(114)
where we have applied Chebyshev's inequality, and we have assumed that N is sufficiently large to have
N pNE[eij] ≤ /2 (a condition guaranteed by (113)). Therefore, in view of (111) and (114), the claim
of the proposition will be proved if we are able to show that
N→∞−→ 0
(N pN )2σ2
e
(115)
pN
Accordingly, we now focus on the evaluation of σ2
in (73), and apply Lemma 3 to eij, for i (cid:54)= j, with the choices L = N − K, and:
e. To this end, we use for eij the representation given
[u1, u2, . . . , uL] → [aωi1, aωi2, . . . , aωiL],
[z1, z2, . . . , zL] → [f1j, f2j, . . . , fLj],
s → eij.
(116)
We remark that all the matrices are random, as a consequence of the randomness implied by the underlying
Erdos-Renyi graph. Preliminarily, it is necessary to verify that the choices in (116) meet the assumptions
of Lemma 3. First, observe that, in view of (152), row-permutations leave unaltered the statistical
distribution of F . This property implies in particular that the random variables f1j, f2j, . . . , fLj are
identically distributed. Moreover, under the independence approximation, such entries are approximated
as mutually independent (and, hence, uncorrelated). Under the same approximation, the random vectors
[aωi1, aωi2, . . . , aωiL] and [f1j, f2j, . . . , fLj] are approximated as mutually independent. Therefore, we
conclude that the hypotheses of Lemma 3 are met, and we can write:
N−K(cid:88)
(cid:34)N−K(cid:88)
(cid:35)
V[eij] = σ2
e ≈ σ2
f
E[a2
ωiω(cid:48)
(cid:96)
] + ¯f 2 V
aωiω(cid:48)
(cid:96)
,
(117)
The approximation symbol in (117) comes from the fact that the pertinent calculations are done under
(cid:96)=1
(cid:96)=1
the independence approximation.
Now, in view of (150), any renumbering leaves unaltered the statistical distribution of A. This property
implies that the random variables aij, for all i, j = 1, 2, . . . , N, with i (cid:54)= j, are identically distributed,
June 26, 2021
DRAFT
ij] = a2. Since ωi (cid:54)= ω(cid:48)
(cid:35)
(cid:34)N−K(cid:88)
and, in particular, that they share the common mean-square value E[a2
Ω ∩ Ω(cid:48) = ∅), Eq. (117) can be rewritten as:
e ≈ (N − K)σ2
σ2
f a2 + ¯f 2 V
aωiω(cid:48)
.
Likewise, we apply Lemma 3 to f(cid:96)j, with the choices L = N − K, and:
[u1, u2, . . . , uL] → [h(cid:96)1, h(cid:96)2, . . . , h(cid:96)L],
[z1, z2, . . . , zL] → [b1ωj , b2ωj , . . . , bLωj ],
s → f(cid:96)j.
(cid:96)=1
(cid:96)
40
(cid:96) (because
(118)
(119)
In view of (164), any renumbering leaves unaltered the statistical distribution of B. This property implies
in particular that the random variables b1ωj , b2ωj , . . . , bLωj are identically distributed. Moreover, under the
independence approximation, such entries are approximated as mutually independent (and, hence, uncor-
related). Under the same approximation, the random vectors [h(cid:96)1, h(cid:96)2, . . . , h(cid:96)L] and [b1ωj , b2ωj , . . . , bLωj ]
are approximated as mutually independent. Therefore, we conclude that also in this case the hypotheses
of Lemma 3 are met, which allows writing:
N−K(cid:88)
Substituting (120) into (118) gives:
m=1
e ≈ (N − K) a2 σ2
σ2
b
f ≈ σ2
σ2
b
E[h2
(cid:96)m] + ¯b2 V
h(cid:96)m
.
(120)
(cid:35)
m=1
(cid:34)N−K(cid:88)
N−K(cid:88)
(cid:34)N−K(cid:88)
(cid:35)
E[h2
m=1
m=1
(cid:96)m] +
(cid:35)
h(cid:96)m
+
(N − K) a2 ¯b2 V
(cid:34)N−K(cid:88)
aωiω(cid:48)
(cid:96)
¯f 2 V
(cid:44) σ2
e .
(cid:96)=1
(121)
A comparison between the true variance, σ2
e, is given in Fig. 8.
We see that, for the considered examples, the agreement between the two quantities is satisfying. Note
e, and the approximate variance, σ2
that, in certain cases, the true curve and the approximate curve (pertaining to the same example) exhibit
slightly different slopes, suggesting that the approximation does not always capture the precise asymptotic
behavior.
We now compute an upper bound for the approximate variance in (121). First, we find an upper
bound for ¯f. In view of (153), column-permutations leave unaltered the statistical distribution of F .
June 26, 2021
DRAFT
41
Fig. 8. Comparison between the true error variance, V[eij] (evaluated through 103 Monte Carlo runs), and the error variance
arising from the independence approximation, see (121). The different curves are obtained by varying the type of combination
matrix (Laplacian or Metropolis), and the fraction of observable agents (ξ = 0.2 or ξ = 0.5).
This property implies in particular that the random variables f(cid:96)1, f(cid:96)2, . . . , f(cid:96)K are identically distributed,
yielding:
(cid:34)(cid:80)K
(cid:35)
¯f =
1
K
E[f(cid:96)j] = E
j=1 f(cid:96)j
K
≤ 1
K
,
(122)
K(cid:88)
j=1
where the inequality comes from (74).
Second, we observe that:
N−K(cid:88)
h(cid:96)m
m=1
(a)
=
(b)≤
(c)
=
(d)≤
(e)
=
k=0
k=0
∞(cid:88)
∞(cid:88)
∞(cid:88)
∞(cid:88)
∞(cid:88)
k=0
k=0
k=0
m=1
N−K(cid:88)
N−K(cid:88)
N−K(cid:88)
N(cid:88)
m=1
m=1
[(BΩ(cid:48))k](cid:96)m
[[Bk]Ω(cid:48)](cid:96)m
[Bk]ω(cid:48)
(cid:96)ω(cid:48)
m
[Bk]ω(cid:48)
(cid:96)j
j=1
(1 − µ)2k =
1
µ(2 − µ)
(cid:44) √
C,
(123)
where (a) follows from (69); (b) holds since for any two nonnegative matrices P and Q we have that
PΩ(cid:48)QΩ(cid:48) ≤ [P Q]Ω(cid:48) (with entrywise inequality), and, hence, by iteration we get (BΩ(cid:48))k ≤ [Bk]Ω(cid:48) because
June 26, 2021
DRAFT
10210−610−510−4NError variance ξ=0.2ξ=0.5ξ=0.2ξ=0.5trueapprox.Metropolis ruleLaplacian rule 42
B and its powers are nonnegative; (c) holds because indices ω(cid:48)
N−K span the set Ω(cid:48);
(d) holds because Bk is nonnegative, and, hence, adding more terms can only increase the value of the
j=1[Bk]ij = (1 − µ)2k, for all
j=1 bij = (1 − µ)2 ⇒(cid:80)N
2 < ··· < ω(cid:48)
1 < ω(cid:48)
i = 1, 2, . . . , N. Exploiting now the bounds (122) and (123), from (121) we get:
sum; and (e) holds because B = A2, yielding(cid:80)N
(cid:34)N−K(cid:88)
e ≤ (N − K) a2 σ2
σ2
where we used the following known bounds:
b C + (N − K) a2 ¯b2 C +
,
(cid:33)2 .
(124)
(125)
(126)
(cid:35)
.
V
1
K2
aωiω(cid:48)
(cid:96)
,
(cid:96)=1
N−K(cid:88)
(cid:34)N−K(cid:88)
m=1
m=1
(cid:96)m ≤
(cid:35)
h2
V
h(cid:96)m
≤ E
m=1
h(cid:96)m
(cid:33)2
(cid:35)
(cid:32)N−K(cid:88)
(cid:32)N−K(cid:88)
(cid:34)N−K(cid:88)
1
1
K2
(N − 1)
h(cid:96)m
m=1
κ2
V
(cid:96)=1
E
(cid:20)
e ≤ (N − K) a2 b2 C +
σ2
≤
a2
i(cid:96)
N(cid:88)
(cid:21)
(cid:20) 1
(cid:96)(cid:54)=i
di
a2 =
1
N − 1
E
≤
κ2
N − 1
E
aωiω(cid:48)
(cid:96)
N(cid:88)
(cid:96)(cid:54)=i
gi(cid:96)
(cid:21)
d2
i
1
=
κ2
N − 1
E
1 + β(N − 1, pN )
,
Since σ2
b + ¯b2 = b2, inequality (124) can be also written as:
We must now examine the asymptotic behavior of the terms a2 and b2. We start with a2. We can write:
follows because(cid:80)
(127)
where the first equality holds because, as already observed, the random variables ai(cid:96), for (cid:96) (cid:54)= i, are
identically distributed in view of (150). The first inequality follows from (48), the second inequality
(cid:96)(cid:54)=i gi(cid:96) = di − 1, and the last equality follows from the fact that di − 1 is distributed
as a binomial random variable of parameters N − 1 and pN . In view of Lemma 2, Eq. (127) implies the
existence of an upper bound,
(1)
N , such that:
a2 ≤
(1)
N ∼ 1
N 2pN
,
(128)
June 26, 2021
DRAFT
where the notation aN ∼ bN means that aN is on the same order of bN , namely, that the limit
limN→∞ aN /bN exists and is finite. It remains to examine the asymptotic behavior of the quantity
b2. Recalling that B = A2 and that A is symmetric, we have, for i (cid:54)= j:
43
bij =
ai(cid:96)a(cid:96)j = (aii + ajj)aij +
N(cid:88)
(cid:96)=1
≤ 2 aij + κ2
(cid:80)
(cid:124)
(cid:96)(cid:54)=i,j gi(cid:96)gj(cid:96)
(cid:123)(cid:122)
didj
(cid:44)zij
,
(cid:125)
(cid:88)
(cid:96)(cid:54)=i,j
ai(cid:96)aj(cid:96)
(129)
(130)
where, in the last step, we used (48). We can thus write:
ij ≤ 4 a2
b2
ij + κ4 z2
ij + 4κ2 aijzij.
Let us first focus on the term z2
ij. We observe that:
(cid:88)
(cid:88)
gi(cid:96)gj(cid:96)gimgjm
gi(cid:96)gj(cid:96) +
(cid:96)(cid:54)=i,j
(cid:96)(cid:54)=i,j
m(cid:54)=i,j
(cid:96)(cid:54)=m
(di − gi(cid:96))2(dj − gj(cid:96))2
gi(cid:96)gj(cid:96)
(cid:123)(cid:122)
Z1
+
(cid:125)
gi(cid:96)gj(cid:96)gimgjm
(di − gi(cid:96) − gim)2(dj − gj(cid:96) − gjm)2
,
(cid:123)(cid:122)
Z2
(cid:125)
z2
ij =
1
d2
i d2
j
(cid:96)(cid:54)=i,j
≤ (cid:88)
(cid:124)
(cid:88)
(cid:124)
(cid:96)(cid:54)=i,j
m(cid:54)=i,j
(cid:96)(cid:54)=m
(131)
where the inequality holds because, in view of (28), we have di > di − gi(cid:96) > 0 for (cid:96) (cid:54)= i and di >
di − gi(cid:96) − gim > 0 for (cid:96) (cid:54)= i and m (cid:54)= i. Using again (28), the terms Z1 and Z2 in (131) can be rewritten
as, respectively:
Z1 =
Z2 =
(cid:88)
(cid:88)
(cid:96)(cid:54)=i,j
(cid:96)(cid:54)=i,j
m(cid:54)=i,j
(cid:96)(cid:54)=m
(1 +(cid:80)
(1 +(cid:80)
gi(cid:96)gj(cid:96)
k(cid:54)=i,(cid:96) gik)2(1 +(cid:80)
k(cid:54)=i,(cid:96),m gik)2(1 +(cid:80)
gi(cid:96)gj(cid:96)gimgjm
k(cid:54)=j,(cid:96) gjk)2
k(cid:54)=j,(cid:96),m gjk)2 .
We are now interested in evaluating the expectation of Z1 and Z2. To this aim, two observations are useful.
First, we note that the random variables gi(cid:96), gj(cid:96), gim, gjm are, for the indices considered in the pertinent
(132)
June 26, 2021
DRAFT
summations, mutually independent Bernoulli variables. Second, the terms in the denominator are binomial
example, the quantity(cid:80)
E[Z1] =
random variables that are independent from the variables present in the numerator. This is because, for
k(cid:54)=i,(cid:96),m gik appearing at the denominator of the second ratio in (132), does not
contain any of the random variables gi(cid:96), gj(cid:96), gim, gjm (nor their symmetric counterparts g(cid:96)i, g(cid:96)j, gmi, gmj)
that appear at the numerator. Therefore, taking the expectation of Z1 in (131) yields:
(cid:88)
(1 +(cid:80)
Now, since E[gi(cid:96)] = E[gj(cid:96)] = pN , since (cid:80)
(133)
k(cid:54)=i,(cid:96) gik is a binomial random variable β(N − 2, pN ), and
(cid:20)
(cid:18)
since the summation in (133) contains no more than N terms, we have that:
(1 +(cid:80)
1
k(cid:54)=j,(cid:96) gjk)2
E[gi(cid:96)]E[gj(cid:96)]×
1
k(cid:54)=i,(cid:96) gik)2
(cid:21)(cid:19)2
(cid:34)
(cid:35)
(cid:34)
(cid:35)
(cid:96)(cid:54)=i,j
E
E
.
E[Z1] ≤ N p2
N
E
1
(1 + β(N − 2, pN ))2
.
(134)
Likewise, taking the expectation of Z2 in (131) yields:
E[Z2] =
E[gi(cid:96)]E[gj(cid:96)]E[gim]E[gjm]×
(cid:88)
(cid:34)
E
(cid:96)(cid:54)=i,j
m(cid:54)=i,j
(cid:96)(cid:54)=m
(1 +(cid:80)
(cid:35)
(cid:34)
E
(1 +(cid:80)
1
k(cid:54)=j,(cid:96),m gjk)2
(cid:35)
.
1
k(cid:54)=i,(cid:96),m gik)2
Since (cid:80)
(135)
k(cid:54)=i,(cid:96),m gik is a binomial random variable β(N − 3, pN ), and since the summation in (135)
contains no more than N 2 terms, we have that:
44
(136)
(137)
(138)
DRAFT
Finally, using (134) and (136) into (131), we have:
E[Z2] ≤ N 2p4
N
(1 + β(N − 3, pN ))2
E[z2
ij] ≤ N p2
N
(1 + β(N − 2, pN ))2
1
(1 + β(N − 3, pN ))2
N 2p4
N
.
(cid:21)(cid:19)2
(cid:21)(cid:19)2
(cid:21)(cid:19)2
+
.
E
(cid:18)
(cid:18)
(cid:18)
E
(cid:20)
(cid:20)
(cid:20)
E
1
1
In view of Lemma 2, we conclude that:
E[z2
ij] ≤
(2)
N ∼ 1
N 3p2
N
+
1
N 2 .
June 26, 2021
Let us move on to examining the term aijzij in (130). Using (48) and the definition of zij in (129), we
can write:
aijzij ≤ κ
di
gijzij = κ gij
(cid:96)(cid:54)=i,j gi(cid:96)gj(cid:96)
d2
i dj
Working along the same lines as done to obtain (135) and (136), we can write:
.
(139)
(cid:80)
45
(cid:21)
(cid:96)(cid:54)=i,j gi(cid:96)gj(cid:96)
d2
i dj
E[aijzij]
≤ κ E
≤ κ E
gij
(cid:80)
(cid:20)
(cid:88)
(cid:88)
(cid:34)
(1 +(cid:80)
(cid:20)
N ×
(cid:96)(cid:54)=i,j
= κ
E
≤ κ N p3
(cid:35)
gijgi(cid:96)gj(cid:96)
(di − gi(cid:96) − gij)2(dj − gj(cid:96) − gij)
(cid:96)(cid:54)=i,j
E[gij]E[gi(cid:96)]E[gj(cid:96)] ×
1
k(cid:54)=i,j,(cid:96) gik)2
1
k(cid:54)=i,j,(cid:96) gjk)
(cid:35)
(cid:34)
E
E
(cid:21)
(1 +(cid:80)
(cid:20)
E
1
(1 + β(N − 3, pN ))2
1
1 + β(N − 3, pN )
(cid:21)
,
(140)
(141)
which, using again Lemma 2, yields:
E[aijzij] ≤
(3)
N ∼ 1
N 2 .
We can now apply to (130) the upper bounds obtained in (128), (138) and (141), getting:
E[b2
ij] ≤ 4 E[a2
ij] + κ4 E[z2
ij] + 4κ2 E[aijzij]
≤ 4
∼
(1)
N + κ4
1
N 2pN
(3)
(2)
N + 4κ2
N
1
1
N 2
+
+
(cid:18)
N 3p2
N
1
1 +
N pN
=
1
N 2pN
(cid:19)
∼ 1
N 2pN
,
+ pN
(142)
where the last estimate follows because, under the G (cid:63)(N, pN ) model, N pN = ln N + cN → ∞, and
pN → 0, as N → ∞.
It remains to use the results obtained as regards the decaying rate of a2 and b2 into (126). To this end,
it is useful to recast (126) in the following form:
e ≤ C(1 − K/N ) N a2 b2 +
σ2
1/N 2
(K/N )2 (1 − µ)2,
(143)
DRAFT
June 26, 2021
where we used the fact that(cid:80)N−K
≤ 1− µ, and that the variance is upper bounded by the mean-
square value. Using now (128) and (142) into (143), and noticing further that K/N → ξ in view of (7),
we conclude that:
(cid:96)=1 aωiω(cid:48)
(cid:96)
e ≤
σ2
(4)
N ∼ 1
N 3p2
N
+
1
N 2 ,
implying:
(N pN )2σ2
e
pN
≤ N 2pN
(4)
N ∼ 1
N pN
N→∞−→ 0,
+ pN
which, in view of (115), establishes the proposition.
(144)
(145)
We start by introducing the operations of matrix permutation and renumbering. A matrix P is a
permutation matrix if exactly one entry in each row and column is equal to 1, and all other entries are
APPENDIX E
0 -- see [32]. Left multiplication of a matrix Z by a permutation matrix P permutes the rows of Z,
whereas right multiplication permutes the columns. For example, if:
46
(146)
P =
0 0 0 1
0 1 0 0
1 0 0 0
0 0 1 0
then Z = P Z sends the first row of Z to the third row of Z, leaves the second row of Z in the second
row of Z, sends the third row of Z to the fourth row of Z, and sends the fourth row of Z to the first
row of Z.
Permutation matrices have the following property [32]:
−1.
P T = P
(147)
Since P T permutes columns in the same way that P permutes rows, the transformation Z → P ZP T
permutes the rows and columns of Z. In our network setting, where the (i, j)-th matrix entry is associated
to a property of agents i and j, the latter transformation corresponds to renumbering the agents.
A random matrix Z will be said to be statistically invariant to renumbering if, for any permutation
matrix P :
P ZP T (d)
= Z,
(148)
where the symbol (d)
permutation, if we have, for any permutation matrix P :
= denotes equality in distribution. Likewise, Z is invariant to row (resp., to column)
June 26, 2021
P Z
(d)
= Z
(resp., ZP T (d)
= Z).
(149)
DRAFT
47
Lemma 4 (Renumbered combination matrix): Under the Erdos-Renyi model, if the combination policy
possesses property P2, the N × N combination matrix A and the K × K error matrix E in (67) are
statistically invariant to renumbering. Likewise, the (N − K) × K matrix F in (71) is invariant to row
and to column permutation. Formally, letting PM be the ensemble of all M × M permutation matrices,
we have that:
P AP T (d)
P EP T (d)
(d)
= F ,
P F
F P T (d)
= F ,
= A,
= E,
∀P ∈ PN ,
∀P ∈ PK,
∀P ∈ PN−K,
∀P ∈ PK.
(150)
(151)
(152)
(153)
Proof: Under the Erdos-Renyi model, the variables gij, for i = 1, 2, . . . , N and j > i, are independent
Bernoulli random variables with P[gij = 1] = pN , and the matrix G is a symmetric matrix. Therefore,
exchanging the agents does not alter the statistical properties of G, namely, any renumbered version of
G has the same probability of occurrence:
P GP T (d)
= G ⇒ γ(P GP T )
(d)
= γ(G),
(154)
where the latter equality in distribution holds because γ(·) is a deterministic function. Moreover, by
property P2 we have that:
γ(P GP T ) = P γ(G)P T .
(155)
Since A = γ(G), Eqs. (154) and (155) immediately imply (150).
Next we prove (151). Using (67) and (66), the matrix E can be formally written as a function of A
as follows:
E = AΩΩ(cid:48)HBΩ(cid:48)Ω
= AΩΩ(cid:48)(IN−K − [A2]Ω(cid:48))
(cid:44) ψ(A).
−1[A2]Ω(cid:48)Ω
(156)
Let now A and E denote arbitrarily renumbered versions of A and E, respectively. Since we know that
A and A share the same distribution, and since ψ(·) is a deterministic function, also ψ(A) and ψ( A) will
have the same distribution. Therefore, in view of (156), claim (151) will be proved if we show that any
renumbered error matrix, E, can be always written as E = ψ( A), for a certain renumbered combination
matrix, A. To this end, let us introduce an N × N permutation matrix P that permutes rows belonging to
June 26, 2021
DRAFT
the index set Ω only with rows belonging to Ω, and rows belonging to the complement set Ω(cid:48) only with
rows belonging to Ω(cid:48). Formally, this assumption implies that the principal sub-matrix PΩ is a K × K
permutation matrix, that PΩ(cid:48) is an (N − K)× (N − K) permutation matrix, and that PΩΩ(cid:48) and PΩ(cid:48)Ω are
matrices containing only zero entries. Therefore, since any renumbering of matrix E is accomplished by
using a certain K × K permutation matrix, we can always write, without losing generality:
48
E = PΩEP T
Ω = PΩAΩΩ(cid:48)P T
Ω(cid:48)PΩ(cid:48)HP T
Ω(cid:48)PΩ(cid:48)BΩ(cid:48)ΩP T
Ω ,
(157)
where we used (147). From the rules for multiplication of partitioned matrices (and since PΩΩ(cid:48) and PΩ(cid:48)Ω
are matrices with all zeros), we have, for a generic N × N matrix Z:
[P ZP T ]Ω = PΩZΩP T
Ω ,
[P ZP T ]Ω(cid:48) = PΩ(cid:48)ZΩ(cid:48)P T
Ω(cid:48),
[P ZP T ]ΩΩ(cid:48) = PΩZΩΩ(cid:48)P T
Ω(cid:48).
[P ZP T ]Ω(cid:48)Ω = PΩ(cid:48)ZΩ(cid:48)ΩP T
Ω ,
Therefore, applying (160) and (161) to (157), we get:
E = PΩEP T
Ω = [P AP T ]ΩΩ(cid:48) PΩ(cid:48)HP T
Ω(cid:48) [P BP T ]Ω(cid:48)Ω.
Moreover, we observe that:
PΩ(cid:48)HP T
Ω(cid:48) = PΩ(cid:48)(IN−K − BΩ(cid:48))
−1P T
Ω(cid:48)
= (PΩ(cid:48)(IN−K − BΩ(cid:48))P T
−1
Ω(cid:48))
= (IN−K − PΩ(cid:48)BΩ(cid:48)P T
−1
Ω(cid:48))
= (IN−K − [P BP T ]Ω(cid:48))
−1
where we used again (147). On the other hand, we have that:
P BP T = P A2P T = (P AP T )(P AP T ) = (P AP T )2.
Finally, letting A = P AP T , and using (163) and (164) into (162), we get:
−1[ A2]Ω(cid:48)Ω = ψ( A),
E = AΩΩ(cid:48)(IN−K − [ A2]Ω(cid:48))
(158)
(159)
(160)
(161)
(162)
(163)
(164)
(165)
which completes the proof of (151).
Let us now focus on proving (152). Using (71) and (66), the matrix F can be formally written as a
function of A as follows:
F = HBΩ(cid:48)Ω = (IN−K − [A2]Ω(cid:48))
−1[A2]Ω(cid:48)Ω (cid:44) ϕ(A).
(166)
DRAFT
June 26, 2021
49
Let now F denote an arbitrarily row-permuted version of F . Claim (152) will be proved if we show that
any row-permuted matrix F can be always written as F = ϕ( A), for a certain renumbered combination
matrix, A. To this end, let us introduce an N × N permutation matrix P that permutes rows belonging
to the complement set Ω(cid:48) only with rows belonging to Ω(cid:48), while leaving unaltered the rows belonging to
the index set Ω. Such assumption implies that the principal sub-matrix PΩ(cid:48) is an (N − K) × (N − K)
permutation matrix, that PΩ = IK, and that PΩΩ(cid:48) and PΩ(cid:48)Ω are matrices containing only zero entries.
Thus, an arbitrary row-permutation of matrix F can be represented as:
F = PΩ(cid:48)F = PΩ(cid:48)HP T
Ω(cid:48)PΩ(cid:48)BΩ(cid:48)ΩP T
Ω
= PΩ(cid:48)HP T
Ω(cid:48) [P BP T ]Ω(cid:48)Ω,
(167)
where we used (147), the equality PΩ = IK, and (161). Letting now A = P AP T , and using (163)
and (164) into (167), we get:
F = (IN−K − [ A2]Ω(cid:48))
−1[ A2]Ω(cid:48)Ω = ϕ( A),
(168)
which completes the proof of (152). The proof of (153) is similar to the proof of (152), and is accordingly
omitted.
REFERENCES
[1] P. Venkitasubramaniam, T. He, and L. Tong, "Anonymous networking amidst eavesdroppers," IEEE Trans. Inf. Theory,
vol. 54, no. 6, Jun. 2008.
[2] T. He and L. Tong, "Distributed detection of information flows," IEEE Trans. Inf. Forensics and Security, vol. 3, no. 3,
pp. 390 -- 403, Sep. 2008.
[3] T. He and L. Tong, "Detection of information flows," IEEE Trans. Inf. Theory, vol. 54, no. 11, pp. 4925 -- 4945, Nov. 2008.
[4] T. He, A. Agaskar, and L. Tong, "Distributed detection of multi-hop information flows with fusion capacity constraints,"
IEEE Trans. Signal Processing, vol. 58, no. 6, pp. 3373 -- 3383, Jun. 2010.
[5] J. Kim and L. Tong, "Unsupervised and nonparametric detection of information flows," Signal Processing, vol. 92, no. 11,
pp. 2577 -- 2593, Nov. 2012.
[6] S. Marano, V. Matta, T. He, and L. Tong, "The embedding capacity of information flows under renewal traffic," IEEE
Trans. Inf. Theory, vol. 59, no. 3, pp. 1724 -- 1739, Mar. 2013.
[7] S. Mahdizadehaghdam, H. Wang, H. Krim and L. Dai, "Information diffusion of topic propagation in social media," IEEE
Trans. Signal Inf. Process. Netw., vol. 2, no. 4, pp. 569 -- 581, Dec. 2016.
[8] S. Marano, V. Matta, and P. Willett, "The importance of being Earnest: social sensing with unknown agent quality," IEEE
Trans. Signal Inf. Process. Netw., vol. 2, no. 3, pp. 306 -- 320, Sep. 2016.
[9] M. Cheraghchi, A. Karbasi, S. Mohajer, and V. Saligrama, "Graph-constrained group testing," IEEE Trans. Inf. Theory,
vol. 58, no. 1, pp. 248 -- 262, Jan. 2012.
[10] H-J. Park and K. Friston, "Structural and functional brain networks: from connections to cognition," Science, vol. 342,
no. 1658, Nov. 2013.
June 26, 2021
DRAFT
50
[11] C. Quinn, T. Coleman, N. Kiyavash, and N. Hatsopoulos, "Estimating the directed information to infer causal relationships
in ensemble neural spike train recordings," Journal of Computational Neuroscience, vol. 30, no. 1, pp. 17 -- 44, Feb. 2011.
[12] M. Mardani, G. Mateos, and G. B. Giannakis, "Dynamic anomalography: tracking network anomalies via sparsity and low
rank," IEEE J. Sel. Topics Signal Process., vol. 7, no. 1, pp. 50 -- 66, Feb. 2013.
[13] A. H. Sayed, "Adaptation, learning, and optimization over networks," Foundations and Trends in Machine Learning, vol. 7,
no. 4-5, pp. 311 -- 801, NOW Publishers, Boston-Delft, 2014.
[14] A. H. Sayed, "Adaptive networks," Proceedings of the IEEE, vol. 102, no. 4, pp. 460 -- 497, Apr. 2014.
[15] V. Matta, P. Braca, S. Marano, and A. H. Sayed, "Diffusion-based adaptive distributed detection: steady-state performance
in the slow adaptation regime," IEEE Trans. Inf. Theory, vol. 62, no. 8, pp. 4710 -- 4732, Aug. 2016.
[16] V. Matta, P. Braca, S. Marano, and A. H. Sayed, "Distributed detection over adaptive networks: refined asymptotics and
the role of connectivity," IEEE Trans. Signal Inf. Process. Netw., vol. 2, no. 4, pp. 442 -- 460, Dec. 2016.
[17] J. Chen and A. H. Sayed, "On the learning behavior of adaptive networks -- Part I: Transient analysis," IEEE Trans. Inf.
Theory, vol. 61, no. 6, pp. 3487-3517, Jun. 2015.
[18] J. Chen and A. H. Sayed, "On the learning behavior of adaptive networks -- Part II: Performance analysis," IEEE Trans.
Inf. Theory, vol. 61, no. 6, pp. 3518-3548, Jun. 2015.
[19] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, "The emerging field of signal processing on
graphs: Extending high-dimensional data analysis to networks and other irregular domains," IEEE Signal Process. Mag.,
vol. 30, no. 3, pp. 83 -- 98, May 2013.
[20] A. Anis, A. Gadde, and A. Ortega, "Efficient sampling set selection for bandlimited graph signals using graph spectral
proxies," IEEE Trans. Signal Process., vol. 64, no. 14, pp. 3775 -- 3789, Jul. 2016.
[21] M. Tsitsvero, S. Barbarossa, and P. Di Lorenzo, "Signals on graphs: Uncertainty principle and sampling," IEEE Trans.
Signal Process., vol. 64, no. 18, pp. 4846 -- 4861, Sep. 2016.
[22] D. Materassi and M. V. Salapaka, "On the problem of reconstructing an unknown topology via locality properties of the
Wiener filter," IEEE Trans. Autom. Control, vol. 57, no. 7, pp. 1765 -- 1777, Jul. 2012.
[23] C. Quinn, N. Kiyavash, and T. Coleman, "Directed information graphs," IEEE Trans. Inf. Theory, vol. 61, no. 12, pp.
6887 -- 6909, Dec. 2015.
[24] J. Etesami and N. Kiyavash, "Measuring causal relationships in dynamical systems through recovery of functional
dependencies," IEEE Trans. Signal Inf. Process. Netw., to appear, DOI: 10.1109/TSIPN.2016.2639439, 2016.
[25] A. Moneta, N. Chlass, D. Entner, and P. Hoyer, "Causal search in structural vector autoregressive models," in JMLR
Workshop and Conference Proceedings: Causality in Time Series, vol. 12, pp. 95 -- 114, 2011.
[26] J. Mei and J. Moura, "Signal processing on graphs: causal modeling of unstructured data," IEEE Trans. Signal Process.,
to appear, DOI: 10.1109/TSP.2016.2634543, 2016.
[27] P. Geiger, K. Zhang, M. Gong, D. Janzing, and B. Scholkopf, "Causal inference by identification of vector autoregressive
processes with hidden components," in Proc. of the 32nd International Conference on Machine Learning, Lille, France,
Jul. 2015, vol. 37, pp. 1917 -- 1925.
[28] J. Etesami, N. Kiyavash, and T. Coleman, "Learning minimal latent directed information polytrees," Neural Computation,
vol. 28, no. 9, pp. 1723 -- 1768, Aug. 2016.
[29] B. Pasdeloup, V. Gripon, G. Mercier, D. Pastor, and M. G. Rabbat, "Characterization and inference of graph diffusion
processes from observations of stationary signals," avaliable as: arXiv:1605.02569v3, Mar. 2017.
[30] B. Pasdeloup, M. Rabbat, V. Gripon, D. Pastor, and G. Mercier, "Graph reconstruction from the observation of diffused
June 26, 2021
DRAFT
51
signals," in Proc. of the 53rd Annual Allerton Conference on Communication, Control, and Computing, Monticello, IL,
USA, Sep. 29 - Oct. 2, 2015, pp. 1386 -- 1390.
[31] S. Segarra, A. G. Marques, G. Mateos, and A. Ribeiro, "Network topology identification from imperfect spectral templates,"
in Proc. of the Asilomar Conf. on Signals, Systems, and Computers, Pacific Grove, CA, USA, Nov. 2016, pp. 1465 -- 1469.
[32] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge University Press, New York, 1985.
[33] T. Kailath, A. H. Sayed, and B. Hassibi Linear Estimation. Prentice Hall, 2000.
[34] A. H. Sayed, Adaptive Filters. Wiley, NJ, 2008.
[35] P. Erdos and N. R´enyi, "On random graphs, I.," Publ. Math. Debrecen, vol. 6, pp. 290 -- 297, 1959.
[36] B. Bollob´as, Random Graphs. Cambridge University Press, 2001.
[37] A. Frieze and M. Kar´onski, Introduction to Random Graphs. Cambridge University Press, 2016.
[38] H. Shao, Mathematical Statistics, 2nd ed., Springer, 2003.
June 26, 2021
DRAFT
|
cs/0603126 | 1 | 0603 | 2006-03-30T22:53:15 | Open at the Top; Open at the Bottom; and Continually (but Slowly) Evolving | [
"cs.MA"
] | Systems of systems differ from traditional systems in that they are open at the top, open at the bottom, and continually (but slowly) evolving. "Open at the top" means that there is no pre-defined top level application. New applications may be created at any time. "Open at the bottom" means that the system primitives are defined functionally rather than concretely. This allows the implementation of these primitives to be modified as technology changes. "Continually (but slowly) evolving" means that the system's functionality is stable enough to be useful but is understood to be subject to modification. Systems with these properties tend to be environments within which other systems operate--and hence are systems of systems. It is also important to understand the larger environment within which a system of systems exists. | cs.MA | cs | Russ Abbott
Department of Computer Science
California State University, Los Angeles
and The Aerospace Corporation
Los Angeles, Ca, USA
[email protected]
Abstract – Systems of systems differ from traditional sys-
tems in that they are open at the top, open at the bottom,
and continually (but slowly) evolving. “Open at the top”
means that there is no pre-defined top level application.
New applications may be created at any time. “Open at
the bottom” means that the system primitives are defined
functionally rather than concretely. This allows the imple-
mentation of these primitives to be modified as technology
changes. “Continually (but slowly) evolving” means that
the system’s functionality is stable enough to be useful but
is understood to be subject to modification. Systems with
these properties tend to be environments within which
other systems operate—and hence are systems of systems.
It is also important to understand the larger environment
within which a system of systems exists.
Keywords: System of systems, open, evolving, environ-
ment.
Introduction
1
The term system of systems has been in widespread
use for at least a decade. As long ago as 1996 Maier [1]
offered “Architecting Principles for Systems-of-Systems.”
In last year’s IEEE Conference on Systems of Systems
Engineering (October 2005) Jamshidi [2] listed six other
definitions of system of systems, which he asserted were
among the most common. Three date from 1998 or earlier.
Even though the term system of systems has been
around for quite a while, we still seem to be struggling with
the concept. Jamshidi quoted approvingly from the claim in
Sage and Cuppan [3] that “there is no universally accepted
definition of systems of systems.” Most definitions of sys-
tem of systems are not very helpful. Some are harmful.
An example of a not very helpful definition is Kotov’s
[4] (from 1997), which Jamshidi describes as his “favorite.”
It reads, “Systems of systems are large scale concurrent and
distributed systems that are comprised of complex sys-
tems.”
Open at the Top; Open at the Bottom; and
Continually (but Slowly) Evolving
This definition does not offer much insight into either
what a system of systems is or what distinguishes a system
of systems from, say, a collection of (large scale concurrent
and distributed) systems. The definition is flawed also in
that it attempts to distinguish systems of systems from just
plain systems—which themselves may be composed of
subsidiary systems—on the grounds that the component
systems that make up a system of systems must be complex.
Not only is it unclear how a complex system is defined for
the purpose of this definition, it is also unclear how this
definition distinguishes a system of system from a plain old
system whose (non-system) subsystems are complex.
The Defense Acquisition University’s Defense Acqui-
sition Guidebook [5] includes a discussion of system of
systems engineering (section 4.2.6).
System of systems engineering deals with plan-
ning, analyzing, organizing, and integrating the
capabilities of a m ix of existing and new systems
into a system of systems capability greater than
the sum of the capabilities of the constituent
parts.
What is unfortunate about this definition is its focus
on the capability provided by the system of systems as a
functional entity. In doing so, it treats a system of systems
as similar to any other system—defined in terms of a set of
capabilities. According to this definition, the primary dif-
ference between a system of systems and a regular system is
apparently that the systems of systems has systems (rather
than subsystems?) as components. From this perspective,
systems of systems are not qualitatively different from plain
old systems. If there are no significant qualitative or struc-
tural differences between traditional systems and systems of
systems, what really is all the fuss about?
In this paper we argue that properly understood a sys-
tem of systems is qualitatively and structurally different
from a traditional system and that the term system of sys-
tems should be reserved for this new perspective. As we
elaborate below, a system of systems is best viewed not as a
hierarchy built of component systems but as an environment
within which other systems operate and which can support
the addition of new systems that build on systems already in
the environment. Furthermore to fully understand a system
of systems not only must it be viewed as an environment for
other systems, it must itself be understood in terms of the
larger environment within which it and its participating
systems exist. In other words, a system of systems perspec-
tive requires one to look outward from a system rather than
inwards towards the system’s hierarchical components.
2 Systems of systems are different
Our objective here is to describe how systems of sys-
tems are qualitatively and structurally different from tradi-
tional system and are not just a larger version of the same
old hierarchical structure. We claim that a system of sys-
tems—when properly understood as an environment along
with the systems operating within it—is open at the top,
open at the bottom, and continually evolving—but slowly
enough to be stable.
1. Open at the top. A system of systems is not defined
in terms of some fixed top-level application. This
distinguishes it from most systems as traditionally
understood. Instead, a system of system enables the
continual introduction of new applications. For ex-
ample, although often quite primitive, the so-called
mashups combine information from two or more
websites. Many combine Google Maps with some
other source. (See [6], which publishes a matrix of
mashups, updated daily.) Burke [7] documented such
connections more generally three decades ago.
2. Open at the bottom. There is no fixed bottom level
for a system of systems. The lowest level of a system
of system may be changed out from under it at any
time. As an example, consider any communication
stack. (As we shall discuss below, communication
systems are prototypical systems of systems.) Signal
transport is typically the lowest layer. Yet consider
how wireless networks are replacing wired connec-
tions. If one were committed to wires as a fixed con-
crete bottom layer, this would not be possible. It is
frequently the case that changes at the lowest level
reflect influences from the environment within which
the system of system is itself embedded—in this case
improved technology developed in the world outside
the communication system.
3. Continually evolving, but slowly. A system of sys-
tems is never finished. It evolves continually as the
environment within which it operates changes. Sys-
tems of systems evolve in at least three ways.
(a) Technology changes: wireless replaces wires.
(b) Usage changes. New features are added, and ex-
isting features are modified. A system of systems
must be able to support new uses of existing capa-
bilities as well as the addition of new capabilities
built on top of existing capabilities. (c) Standards
and interfaces change. Even though stable standards
and interfaces are important, it must be possible to
change them as needs change and as our understand-
ing of what we are attempting to standardize im-
proves. Consider virtually any standard. One will
find that it is continually (although not too fre-
quently) re-issued as the purposes for which it is
used evolve and as we understand better what it was
intended to do.
Systems with these properties do not lend themselves
to easy hierarchical control. On the other hand, systems of
this sort are not completely formless. Any system, to be
useful, must be able to perform specific functions at par-
ticular times. Systems of systems achieve this goal in that at
any given time (a) they include a collection of (relatively
stable) participating systems and (b) they implement a (rela-
tively stable) set of standards and interfaces. But neither the
set of participating systems nor the standards and interfaces
are fixed forever. They evolve—but slowly. Thus the best
way to think of a system of systems is as a collection of
participating systems along with a set of shared standards
and interfaces all of which evolve slowly enough so that
making an investment in them is likely to be worthwhile.
What do we mean by a standard that is continually
evolving but evolving slowly enough so that making an
investment in it is worthwhile? Think of a person. Each of
us changes physically from day to day. The atoms that
make up our body one day are not the same atoms that
make up our body the next. Moreover, we change our struc-
tures as well. We deteriorate with age, and we improve with
learning. We are not isomorphic to our ourselves of a cer-
tain number of years ago—although there is generally a
significant resemblance. Although we change continually,
we change slowly enough that we find it worthwhile to
invest in ourselves. There is a pace of change that we as
human beings (and as organizations composed of human
beings) can tolerate. Systems of systems change continu-
ally—often in fundamental ways. But the rate at which a
systems of systems changes is for the most part slow
enough that we can adjust and adapt.
Most changes to a system of systems will be slow—
although an aggregation of small changes sometimes leads
to a phase change and to punctuated equilibrium effects.
But fast or slow, there will be changes. It is wishful think-
ing (or foolish wishing) to imagine that a system (or a sys-
tem of system) (a) can be built according to a specification
that is so perfect and complete and (b) will have an archi-
tecture that is so abstract and general that the system’s
fundamental framework will be eternal. Other than nature
itself, no system has a permanent set of rules. And nature,
the ultimate system of systems, is at the same time both
eternal and always evolving.
2.1 Communication systems: the USPS
In this section we consider wide area communication
networks—such as
the United States Postal Service
(USPS), the telephone infrastructure, and the internet—as
prototypical systems of systems. They all exhibit the prop-
erties of being open at the top, open at the bottom, and
always evolving, albeit slowly. Of the three systems men-
tioned we will examine the United States Postal Service
since it is both the most longstanding and the least exam-
ined—at least in this context.
No one has ever accused the USPS (the people who
bring us “snail mail”) of being a high tech system. It was
originally created under authorization by the US constitu-
tion, and it has existed for as long as the country. A pre-
constitutional postal service existed in colonial times. The
constitution authorized Congress to continue providing
postal services. (See [8].) The USPS has been remarkable
in both its stability and its adaptability.
The primary function of the USPS is to deliver mail
(envelopes and packages) to addressees in the United
States. One may think of this as defining its user interface:
give the USPS an item (an envelop or package) with a valid
address (and sufficient postage), and that item will be de-
livered to the indicated address. (Strangely no explicit for-
mal statement of USPS services was found anywhere on the
USPS website: http://www.usps.com/. It may be that the
fundamental mission of the USPS is defined in the same
way that English common law is defined, by tradition.)
Conceptually, what the USPS offers is a remarkably
simple and intuitive service. All one needs to know are
(a) how to address an item (the semantics), (b) how to enter
an addressed item into the USPS system (the mechanism for
providing input to the system), and (c) how to retrieve items
from the system (the mechanism for retrieving output). The
rest of the system may be understood as a black box. Noth-
ing need be said about how the USPS goes about getting
items from where they are received to their destinations.
Open at the bottom. When we say that a system of
systems is open at the bottom, we refer to the system’s
ability to change its lowest level of functionality. In the case
of the USPS, the system was able to re-implement its inter-
nal mail sorting system to make use of technologies (optical
character recognition devices, for example) that didn’t exist
when the system came into existence.
Another illustration of the way in which the USPS is
open at the bottom is its use of scheduled commercial air-
liners to transport mail across the country. Basic transporta-
tion of postal items is the lowest level of the “functionality
stack” upon which all other USPS services are built. Yet
the USPS has been able to change that lowest level with
advances in technology. And that change has occurred
without requiring a modification in the user interface, which
has stayed relatively, if not absolutely, stable.
As these two examples illustrate, openness at the bot-
tom typically reflects influences from the larger environ-
ment within which a given system is embedded—in this
case from the worlds of technology and economics.
Open at the top. When we say that a system of sys-
tems is open at the top, we refer to the system’s ability to
serve as a basis for an unlimited range of services. In the
case of the USPS, consider the following three activities
that depend on its services: (a) the mail order industry,
(b) stamp collecting, and (c) chain letters
Just as web-based companies like Amazon.com exist
only because the internet exists, the giant mail order houses
such as Sears Roebuck and Montgomery Ward and more
modern companies such as Land’s End, LL Bean, SkyMall,
Sharper Image, Harry and David, etc. grew up as catalog
mail order companies. Each of these companies provides a
service that rides on top of the service provided by the
USPS. (Some even order from each other.) USPS services
are essential (a) for delivering catalogs, (b) for receiving
orders, and often (c) for delivering ordered products. Just as
there are now many internet-only businesses, there were
once many catalog-only mail-order businesses, which oper-
ated within the environment provided by the USPS.
Stamp collecting and the businesses that support it ex-
ist only because (a) the USPS issues stamps and (b) people
decided that stamps have a “collectable” value outside their
use to pay for the delivery of items. This is an entirely new
and creative application of services offered by the USPS.
Of course stamp collecting is not limited to USPS-issued
stamps. But the idea here is that stamp collecting is an
application that was built on top of and was tangential to
the reason the USPS issued stamps. Illustrative of how one
service builds on another note that many stamp collecting
companies are mail-order businesses and that the USPS
(and the postal services of many countries) now issue
stamps with the express intent of selling them to collectors.
Chain letters also illustrates how new phenomena may
be built on top of an existing USPS service. VanArdsdale
[9] has written a comprehensive but apparently otherwise
unpublished study of (paper) chain letters. It discusses the
wide range of motivations that drive chain letters, e.g.,
profit, luck, etc. It also describes the historical rates at
which chain letters have propagated. (They seem to be
making a comeback as the number of email chain letters has
declined.) Underneath all this, a feature common to all
successful chain letters is the reproduction and distribution
of the chain letter contents. Thus one of the emergent func-
tions of the USPS is replication and distribution of content.
A chain letter is one way to achieve this end.
The flourishing ecology of services built on top of the
services offered by the USPS “lives” in the environment
created by the USPS—much as an even more flourishing
ecology of services “lives” in the environment created by
the internet. It is in this sense that the environment provided
by the USPS is open at the top and allows new systems to
be developed which offer additional services.
Always evolving—but slowly. When we say that a sys-
tem is always evolving, albeit slowly, we refer not only to
its continual evolution as an ecology of services (as just
discussed) but also to its relative but not absolute stability
with respect to its fundamental service interfaces. Zip codes
(and then Zip+4 codes) and fixed 2-character state abbre-
viations were added to the semantics of postal address in
the mid-20th century. Because the semantics of addressing
has changed relatively slowly (one might say glacially),
users of USPS services were able to accommodate.
One may think of the allowable format for addresses
as comparable to the standards and protocols in today’s
internet environment. The USPS addressing standard has
been stable enough to allow users to learn how to use it and
to adopt it economically. But the standard is not eternal. It
has changed as new demands were placed on the system.
2.2 Other communication systems
We have focused on the USPS. Similar (and more
dramatic) stories can be told about other wide-area commu-
nication systems. Not only are the telephone system and the
internet much more open at the top and the bottom, these
systems too are in a constant state of change. But they too
are stable enough to be useful.
These other systems differ from the USPS in that con-
trol over them is far more distributed. The USPS is, after
all, a monolithic system owned and operated by an agency
of the US government.. Control of the telephone system and
the internet is much broader. And they are not run by the
government.
We chose to focus on the USPS to illustrate that even
a traditional, familiar, and low tech system may be seen as a
successful systems of systems. We also thought it worth
noting that it is possible for the government to nurture and
sustain a viable and long lived system of systems. (Would
anyone ever have imagined that the USPS would be held up
as a positive model in a technical paper?)
2.3 The economic system as a system of sys-
tems
Consider the laws and rules regulating the economic
system. Focus specifically on laws regulating the formation
of corporations and laws regulating how corporations and
individuals interact with each other as economic entities.
The system framed by those laws and rules is an envi-
ronment within which there has been (and continues to be)
enormous activity and creativity. This environment, along
with the activity that occurs within it, may be seen as a
system of systems in the sense described above.
• It is open at the top. There are few limits to the prod-
ucts and services that can be created and offered.
• It is open at the bottom. There is no lowest economic
level of abstraction. Technology at any level down to
quantum physics can be pressed into economic service.
• It evolves continually, but slowly enough to be useful.
The laws and regulations that govern economic activi-
ties are not fixed. Old laws are revised, new laws are
created, and new categories of economic interests are
brought under the law. Intellectual property law is in
particular ferment these days.
As the overseer of the economic system of systems the
government provides relatively minimal services. It pro-
vides a court system, and it provides a monetary system. It
also provides certain regulatory services. But for the most
part, the economic system consists of non-governmental
participants interacting with each other—but within the
framework of the system’s laws and regulations.
The economic system is enormous in terms of the ac-
tivities that take place within it. In comparison, the gov-
ernment’s participation is relatively small. Even consider-
ing the court system, the central banking system, and all
governmental economic-related bureaucracies, the cost to
operate these is barely visible when compared to the eco-
nomic system itself. This illustrates the fact that it does not
require a massive infrastructure for a successful system of
systems to flourish.
2.4 Systems of systems exist within larger
systems of systems
Communication systems exist within the economic
system. Much of their reason for being is economic. In
saying this we do not mean to diminish the importance of
person-to-person non-economic communication. But the
point to be made here is that many of the applications that
grew up within the various communication systems did so
for economic reasons. Amazon.com the internet entity is
but one face of Amazon.com the economic entity.
If communication systems exist within an economic
system of systems environment, within which environment
does the economic system exist? One might want to con-
struct a nested structure of environments, but in the end
there is one final environment: reality. Every system of
systems environment is (a) open at the bottom to the extent
allowed by the laws of nature and (b) open at the top to
whatever we can imagine and find a way to implement.
Open at the bottom. One can make money in the eco-
nomic system. But if one is not successful in the economic
system, one can operate in the real world and simply steal.
The economic system is not isolated from the envi-
ronment in which it functions. Burglary and armed robbery
may not exist as valid operations given the rules and regula-
tions of the economic system. But they exist in the larger
system. No system (or system of systems) can be com-
pletely isolated from its environment. The economic system
is open at the bottom in ways we would prefer did not exist.
A more charming illustration of how activity in one
environment can be influenced by forces from its contain-
ing environment is the development of markets (e.g., on
eBay) for virtual assets from online games. (See Castronova
[10].) Multiplayer games—in which players interact in
virtual worlds and accumulate virtual assets that have value
in those virtual worlds—have developed into widely used
services within the internet environment. eBay, also a ser-
vice within the internet environment, allows participants to
buy and sell assets for real money. Given this combination,
along with the fact that users of both services also exist in
the larger economic world, the buying and selling of virtual
assets for real money was inevitable. Some game compa-
nies discourage this sort of activity, preferring to keep their
virtual worlds closed. Others, acknowledging the virtual
impossibility of closing off their virtual worlds from the
larger, real world, tolerate it. Some even encourage it and
make it a feature of their game.
Open at the top. Human beings are remarkably crea-
tive. We enjoy what has become a flood of new products
and services. This is one way in which the economic system
is open at the top. It is also open at the top in ways we
would prefer did not exist. The 9/11 hijackers took advan-
tage of the fact that the economic system offers classes in
how to fly airplanes. They also took advantage of the fact
that airplanes are in effect flying bombs. The hijackers
didn’t have to manufacture their own bombs. They didn’t
have to transport them to the sites of attack. The system
provided bombs and bomb transport if one only looked at it
the right way. The 9/11 hijackers made very creative use of
the economic system. To accomplish their objective re-
quired only that they bring a few box cutters as weapons
from the real world into the economic world. Their use of
the economic system is one which we would have preferred
they hadn’t thought of. But we can’t eliminate creative
thinking. The economic system is open at the top even for
applications we would prefer did not exist.
2.5 Service oriented architectures
Frequently one has an intuition about something be-
fore one is able to formulate it. The same thing often hap-
pens in social and business contexts: businesses drift in a
direction which only later is understood to be important.
That, we believe, is the situation with service oriented ar-
chitectures. They are increasingly popular, but no com-
pletely satisfying rationale for them has been formulated.
Ort [11] offers as good (and current) a definition of service
oriented architecture (SOA) as any.
A service-oriented architecture is an information
technology approach or strategy in which appli-
cations make use of (perhaps more accurately,
rely on) services available in a network such as
the W orld W ide W eb
If we extend this definition beyond information tech-
nology and consider the examples discussed above it’s clear
that these systems have all been implementations of service
oriented architectures. Even the economic system as a
whole implements an SOA. The framework within which
businesses offer their services to each other may quite read-
ily be understood as a network. Unlike many business fads,
the growing popularity of SOA reflects an increasingly
clear intuition about how the world works.
2.6 A caution for the government
We indicated earlier that systems of systems evolve
continually, although slowly. The reason this works, and the
reason that the systems that participate in a systems of sys-
tems adapt to its changes is that it is worthwhile for them to
do so. When the telephone system went to 10 (really 11)
digit dialing, everyone adapted because it was easier and
cheaper to learn how to dial 10 (or 11) digits than it was to
continue to call an operator for “long distance” calls.
Similarly, imagine what would happen if a new and
more powerful language for expressing web pages (e.g., a
major extension and modification of HTML) were adopted
as a formal recommendation by the Internet Engineering
Task Force (IETF), the advisory committee for web stan-
dards, and supported by the major web browsers. We sus-
pect that within a surprisingly short time, most web pages
would be converted to the new standard—and that the con-
version would occur with very little fuss.
Many government systems consist of elements that
themselves are owned and operated by and for the govern-
ment. Imagine how difficult it would be to make a similar
change in a government owned and operated system of
systems. Each participating system would have to be modi-
fied to use the new standard. For the government to make
all these changes (or pay contractors to make those
changes) would require an enormous amount of coordina-
tion. Very likely it would also be quite expensive.
In contrast, if the participating systems were owned
and operated by entities that were motivated to participate
in the overall environment, one wouldn’t have to formalize
all that coordination. The owners of the participating sys-
tems would see it in their interest to ensure that their sys-
tems continued to function under the new standards.
ton, MA, (1996), pp. 567- 574. Available as of Jan 2006 at:
http://www.infoed.com/Open/PAPERS/systems.htm.
[2] Jamshidi, M. “System-of-Systems Engineering - a
Definition,” IEEE SMC (October 2005). Available as of
Jan 2006 at: http://ieeesmc2005.unm.edu/SoSE_Defn.htm.
[3] Sage, A.P. and C.D. Cuppan, "On the Systems Engi-
neering and Management of Systems of Systems and Fed-
erations of Systems," Information, Knowledge, Systems
Management, Vol. 2, No. 4 (2001), pp. 325-345.
[4] Kotov, V., "Systems of Systems as Communicating
Structures," Hewlett Packard Computer Systems Labora-
tory Paper HPL-97-124, (1997).
[5] Department of Defense, Defense Acquisition Guide-
book. (December 2004). Available as of Jan 2006 at
http://akss.dau.mil/dag/DoD5000.asp?view=framework.
[6] programmableweb. Available as of Jan 2006 at
http://www.programmableweb.com/.
[7] Burke, J. Connections, Little Brown, (1978).
[8] United States Postal Service, History of the United
States Postal Service, (undated). Available as of January
2006 at: http://www.usps.com/history/history/his1.htm.
[9] VanArdsdale, D. W., “Chain Letter Evolution,” (De-
cember 2005). Available as of January 2006 at:
http://www.silcom.com/~barnowl/chain-
letter/evolution.html.
[10] Castronova, E., Synthetic Worlds: The Business and
Culture of Online Games, University Of Chicago Press
(November 2005).
[11] Ort, E., “Service-Oriented Architecture and Web
Services: Concepts, Technologies, and Tools,” Sun
Microsystems, (April 2005). Available as of Jan 2006 at:
http://java.sun.com/developer/technicalArticles/WebServic
es/soa2/.
[12] Abbott, R., “Emergence Explained,” submitted for
publication. Available
at:
2006
Jan
of
as
http://cs.calstatela.edu/~wiki/images/9/90/Emergence_Expl
ained.pdf.
Thus when the government develops a system of sys-
tems environment, it is important to ensure that the partici-
pating systems are owned and operated by entities that have
enough of a stake in their participation that they will adapt
as the environment evolves. Otherwise, the entire system of
systems environment will drift into obsolescence.
3 Conclusions
The term system of systems should not be used to refer
to larger versions of hierarchically organized traditional
systems. It should be reserved for environments within
which systems may interact and serve as services for each
other. Service oriented architectures are thus a valid ap-
proach to building systems of systems.
Systems of systems have the properties that (a) they
are open at the top (to ever new applications), (b) they are
open at the bottom (to new ways of implementing the low-
est level functions), and (c) they evolve continually—albeit
slowly enough to make it worthwhile for the participating
systems to adopt to the changes.
When building a system of systems it is important to
ensure that the participating systems participate because it
is in their interest to do so. In understanding why it may or
may not be advantageous for a system to participate in a
system of systems, one must keep in mind the larger envi-
ronment within which the system of system and its partici-
pating systems is embedded. The larger environment is also
important for understanding the ways in which the system
of system is open (at both the top and the bottom) and sub-
ject to pressures to change. A system of systems perspective
requires one to look outward from a system to its environ-
ment and not just inward from a system to its components.
Although the term system of systems should be used to
refer to an ensemble consisting of a framework along with
some participating systems, we suggest that the term itself
is distracting and somewhat unfortunate. By referring to the
ensemble as a whole, the term fails to distinguish between
the framework and the participating elements. What is most
important when building a system of systems is that the
framework enable the addition of new systems in such a
way that they can build on services offered by other partici-
pating systems. An extended discussion of these and related
issues may be found in Abbott [12].
References
[1] Maier, M. “Architecting Principles for Systems of
Systems,” Proc. of the Sixth Annual International Sympo-
sium, International Council on Systems Engineering, Bos-
|
1709.05142 | 1 | 1709 | 2017-09-15T10:17:56 | Open Multi-Agent Systems: Gossiping with Random Arrivals and Departures | [
"cs.MA",
"math.OC"
] | We consider open multi-agent systems. Unlike the systems usually studied in the literature, here agents may join or leave while the process studied takes place. The system composition and size evolve thus with time. We focus here on systems where the interactions between agents lead to pairwise gossip averages, and where agents either arrive or are replaced at random times. These events prevent any convergence of the system. Instead, we describe the expected system behavior by showing that the evolution of scaled moments of the state can be characterized by a 2-dimensional (possibly time-varying) linear dynamical system. We apply this technique to two cases : (i) systems with fixed size where leaving agents are immediately replaced, and (ii) systems where new agents keep arriving without ever leaving, and whose size grows thus unbounded. | cs.MA | cs | Open Multi-Agent Systems:
Gossiping with Random Arrivals and Departures
Julien M. Hendrickx and Samuel Martin
7
1
0
2
p
e
S
5
1
]
A
M
.
s
c
[
1
v
2
4
1
5
0
.
9
0
7
1
:
v
i
X
r
a
Abstract- We consider open multi-agent systems. Unlike the
systems usually studied in the literature, here agents may join
or leave while the process studied takes place. The system
composition and size evolve thus with time. We focus here on
systems where the interactions between agents lead to pairwise
gossip averages, and where agents either arrive or are replaced
at random times. These events prevent any convergence of the
system. Instead, we describe the expected system behavior by
showing that the evolution of scaled moments of the state can be
characterized by a 2-dimensional (possibly time-varying) linear
dynamical system. We apply this technique to two cases : (i)
systems with fixed size where leaving agents are immediately
replaced, and (ii) systems where new agents keep arriving
without ever leaving, and whose size grows thus unbounded.
I. INTRODUCTION
Two of the most important features of multi-agent sys-
tems are their flexibility and scalability. Accordingly, these
systems are expected to cope with agent failures and new
agent arrivals. Real life examples of the multi-agent systems
with such properties include flock of birds, ad-hoc networks
of mobile devices, or the Internet. Social systems of various
scales have these properties : work teams in companies
or laboratories subject to important turnovers, companies
themselves, or even entire countries and their cultural norms.
However, the framework typically used to study formal
models of multi-agent systems supposes that, while the
system may be complex, its composition remains unchanged
over time (albeit the interaction topology can evolve). Under
this assumption, researchers are able to characterize the long
term behavior of the multi-agent system such as convergence
and synchronization.
This apparent contradiction is justified when agents ar-
rivals and departures are sufficiently rare as compared to the
time-scale of the process taking place in the system. In such
cases, it makes indeed sense to assume that the composition
of the systems remains unchanged while the process takes
place.
Nevertheless, the probability of a node failure is expected
to grow with the number of agents. As a consequence,
Hendrickx
catholique
is
de
Julien
the
with
Louvain,
Uni-
versit´e
Belgium.
[email protected],
is
with Universit´e de Lorraine
and CNRS, CRAN, UMR 7039, 2
Avenue de la Foret et de Haye, 54518 Vandoeuvre-l`es-Nancy, France
[email protected]
ICTEAM institute,
Louvain-la-Neuve,
Samuel Martin
This work was supported by the Belgian Network DYSCO (Dynamical
Systems, Control, and Optimization), funded by the Interuniversity Attrac-
tion Poles Program, initiated by the Belgian Science Policy Office, and by
the ANR project COMPACS ANR-13-BS03-0004 and by projects PEPS
MoMIS MADRES and PEPS INS2I CONAS funded by the CNRS.
for large systems, this constant size assumption no longer
holds. Similarly, in some systems such as living systems
with birth processes for instance, the probability of a node
arrival increases with the system size, so that the constant
composition assumption also stops being relevant when the
system size is large. Companies or human societies are
instances of such systems where the system's growth is
proportional to its size. This assumption may also appear
unsuitable in extreme environments, where communication
is difficult and infrequent, leading to slow convergence rate,
relative to which the agent failure rate may be important.
Hence we consider here open multi-agent systems, where
agents keep arriving and/or leaving during the execution of
the process considered, an example of which is illustrated in
Figure 1.
Repeated arrivals and departures result
in important dif-
ferences in the analysis or the design of open multi-agent
systems and cause several challenges:
State dimension: Every arrival results in an increase of the
system state dimension, and every departure in a decrease
of the system state dimension. Analyzing the evolution of
the system state is therefore much more challenging than in
"closed systems".
Absence of usual "convergence:" Being continuously per-
turbed by departures and arrivals, open systems will never
asymptotically converge to a specific state (this is clear
from Figure 1). Rather, they may approach some form of
steady state behavior, which can be characterized by some
relevant descriptive quantities. As in classical control in the
presence of perturbations, the choice of the measures is not
neutral, and different descriptive quantities may behave in
very different way.
Robustness and quality of the algorithms: Although this
is not treated here, departures and arrivals also have a fun-
damental impact on the design of decentralized algorithms
over open systems. These events will often imply a loss
of information or a change in the algorithm desired result
(information held by new agents may indeed affect a value
that the algorithm should compute). Hence the algorithms
should be robust to departures and arrivals. On the other
hand, algorithms over open systems cannot be expected to
be "exact": When the system composition keeps changing,
algorithms able to maintain an approximate answer most of
the time may be preferable to those that would eventually
provide an exact answer if the system composition were to
remain constant.
2
Fig. 1. Example of dynamics of an open multi-agent system with random
agent replacements and pairwise average gossips. The evolution with time
of the agents values are represented by black continuous lines. Red circles
highlight the departing agents while the blue circles correspond to the newly
arrived agents. Vertical black dashed lines depict the non uniform random
replacement instants. The repeated replacements prevent convergence to
consensus. See Section IV for a precise description of the system.
A. Contribution
We provide results which are part of an ongoing study on
open multi-agent systems. Here, we study systems where
the interactions among agents occur in discrete-time and
take the form of an average pairwise gossip algorithm [2].
We assume all-to-all (possible) communications, focusing on
systems where departures and arrivals take place at random
times, see Section II for a complete definition.
We analyze the system evolution in terms of two "scale-
independent" quantities. We find that the first two expected
moments form such relevant quantities : namely we study the
expected square mean E(¯x2) and the expected mean square
E(x2) of the system state x. These quantities also provide
the evolution of the expected variance E(x2 − ¯x2). We
show in Section III that these quantities can be characterized
exactly, and that they evolve according to an associated 2-
dimensional linear system.
In Section IV, we analyze in detail the case of systems
with replacements taking place at random time : a departure
is immediately followed by an arrival. Each time an event
occurs, it is either a replacement or a gossip step with a
certain probability. In other words, between two consecutive
replacements, K gossip steps occur, where K is a non-
negative integer random variable. As part of the study, we
characterize the variable K. We then focus in Section V
on growing systems that agent keep joining without ever
leaving. It will in particular be shown that random arrivals or
replacements can result in a significant performance decrease
in terms of variance.
Results on simplified versions of the systems considered
here were presented at the Allerton Conference on Commu-
nication Control and Computing [6]. The main differences
with [6] are (i) the arrivals and replacements are probabilistic
events in this paper, while they followed a deterministic (and
generally periodic) sequence in [6], (ii) a study of the system
convergence rate and the interpretation of the corresponding
eigenvectors, and (iii) a different choice of the moments
studied, allowing for simpler proofs.
B. Other works on open multi-agent systems
The possibility of agents joining or leaving the system has
been recognized in computer science, and specific architec-
tures have for example been proposed to deploy large-scale
open multi-agent systems, see e.g. THOMAS project [3].
There also exist mechanisms allowing distributed computa-
tion processes to cope with the shut down of certain nodes
or to take advantage of the arrival of new nodes.
Frameworks similar to open multi-agent arrivals have
also been considered in the context of trust and reputation
computation, motivated by the need to determine which
arriving agents may be considered reliable, see e.g. the model
FIRE [7]. However, the study of these algorithms's behavior
is mostly empirical.
Varying compositions were also studied in the context of
self-stabilizing population protocols [1], [4], where inter-
acting agents (typically finite-state machines) can undergo
temporary or permanent failures, which can respectively
represent
the replacement or the departure of an agent.
The objective in those works is to design algorithms that
eventually stabilize on the desired answer if the system
composition stops changing, i.e. once the system has become
"closed".
Opinion dynamics models with arrivals and departures
have also been empirically studied in [8], [9].
II. SYSTEM DESCRIPTION
We consider a multi-agent system whose composition
evolves with time. We use integers to label the agents, denote
by N (t) ⊂ N the set of agents present in the system at time
t, and by n(t) the number of agents present at time t, i.e. the
cardinality of N (t). Each agent i holds a value xi(t) ∈ (cid:60),
and we make no assumptions about the values held at t = 0
by the agents initially present in the system.
We consider a discrete evolution of the time t ∈ N. It
is possible to interpret the discrete time t as a sampling
of a continuous time variable. Samples then correspond to
instants where an event occurred. We will comment later
on this interpretation and on its implication on the scaling
of different parameters. At each time t, one of three events
may occur:
(a) Gossip: Two agents i, j ∈ N (t) are uniformly ran-
domly and independently selected among the n(t) agents
present in the system (with in particular the possibility of
selecting twice the same agent), and they update their values
xi, xj by performing a pairwise average:
xi(t) + xj(t)
.
2
xi(t + 1) = xj(t + 1) =
(1)
(b) Departure: One uniformly randomly selected agent i ∈
N (t) leaves the system, so that N (t + 1) = N (t) \ {i} and
n(t + 1) = n(t)− 1. This event may only occur if n(t) > 0.
(c) Arrival: One "new" agent i (cid:54)∈ N (s), ∀s ≤ t, joins the
system, so that N (t+1) = N (t)∪{i} and n(t+1) = n(t)+1.
The initial value xi(t + 1) ∈ (cid:60) of the arriving agent is drawn
060t-0.50.5 independently from a constant distribution D with mean 0
and variance σ2. (The results can immediately be adapted to
systems where the mean of the arriving agents states is an
arbitrary constant).
In addition, we will sometimes consider for simplicity
a "replacement" event, which consists of the instantaneous
combination of a departure and an arrival: an agent leaves
the system and is instantaneously replaced.
Note that all the random events above are assumed inde-
pendent of each other.
Scale-independent quantities of interest
The aim of the study is to characterize the disagreement
among agents, i.e. the distance to consensus. We say that
consensus is reached asymptotically when
lim
t→∞ max
(i,j)∈N (t)2
xi(t) − xj(t) = 0.
(2)
If the system dynamics does not include agent departures
or arrivals, it is known that the gossip process we consider
leads to consensus, see e.g. [2], [5]. The objective here
is to understand how agent arrivals and departures impact
the disagreement among agents. To do so, we study several
quantities of interest. Because the system size may change
significantly with time, we focus on scale-independent quan-
tities, i.e. quantities whose values is independent of the size
of the system. We consider in particular the empirical mean
of the squares and the variance defined as
x2 = 1
n
i∈N
Var(x) = 1
n
x2
i ,
(cid:88)
i∈N
(xi − ¯x)2 = x2 − (¯x)2,
(3)
(cid:88)
respectively, where references to time were removed to
lighten the notation. Our study will focus on the evolution
of EVar(x), which will also require monitoring E(¯x)2 and
Ex2. When new agents keep arriving it is impossible to
achieve asymptotic consensus in the sense of (2), because
the new agent's value will with high probability be different
from the value of the agents already present in the system.
The study of EVar(x) will allow us to see how "far"
the system will be from consensus. But we will see that
in certain systems whose sizes grow unbounded, we may
have limt→∞ EVar(x) = 0, corresponding to a form of
"almost consensus". The expected mean E¯x could also have
been monitored. It evolves following an independent one-
dimensional linear system. However, we skip this part of the
study due to space limitations.
III. EFFECT OF DIFFERENT OPERATIONS
In the sequel we will show that
the evolution of the
expected moments E(¯x2) and Ex2 is governed by an affine
system from which we will derive the evolution of EVar(x).
For notational simplicity we denote by X the vector con-
taining ¯x2 and x2 so that Var(x) = (−1, 1)X.
Lemma 1 (Gossip step): Suppose that a randomly se-
lected pair of agents engage in a gossip averaging according
to equation (1). Let x be the state of the system before that
3
(cid:19)
and as a consequence
interaction, x(cid:48) its state after the interaction and n the number
of agents. There holds
EX(cid:48) = AgEX, where Ag =
0
1 − 1
n
,
(4)
(cid:18) 1
(cid:19)
1
n
(cid:18)
1 − 1
n
EVar(x(cid:48)) =
EVar(x).
Proof:
i + x(cid:48)
j = 2 xi+xj
(5)
Notice that this result or its variations are available in many
previous works, but its proof is presented for the sake of
completeness.
let us first fix the nodes i, j involved in the
gossip. Observe that x(cid:48)
2 = xi + xj, and that
k = xk for all k (cid:54)= i, j. Hence ¯x(cid:48) = ¯x, which establishes
x(cid:48)
the fist line of (4). For the second line, since xk = x(cid:48)
(cid:33)
k for
every k (cid:54)= i, j, there holds
1
n
i − 1
xixj − 1
(6)
x2
2
2
jx) = x2 and E(xixjx) =
ix) = E(x2
Observe that E(x2
(cid:19)
(cid:18)
¯x2. Taking the expectation with respect to i and j in (6)
yields
(cid:18) xi + xj
(cid:19)
(cid:19)2 − x2
k=1
1
n
n(cid:88)
i − x2
k = x2 +
x(cid:48)2 =
= x2 +
(cid:32)
x(cid:48)2
(cid:18)
1
n
x2
j
2
2
j
E(x(cid:48)2x) =
1
n
from which the second line of (4) follows.
1 − 1
n
x2 +
¯x2,
Lemma 2 (Arrival of the n + 1-th agent): Suppose
that
an agent arrives into the system bringing the number of
agents from n to n + 1. Denote x the state before arrival,
x(cid:48) the state after arrival. Then, there holds
EX(cid:48) = AaEX + ba
(7)
(cid:19)
(cid:18)
(n+1)2
1
1
n+1
and ba = σ2
where
Aa =
(cid:32) n2
(n+1)2
0
so that x(cid:48)
new average :
(cid:32)
(cid:33)
(cid:33)
0
n
n+1
n(cid:88)
k=1
Proof: We label n + 1 the arriving agent for simplicity,
k = xk for all k ≤ n. We begin by computing the
¯x(cid:48) =
1
n + 1
x(cid:48)
n+1 +
xk
=
n
n + 1
¯x +
1
n + 1
x(cid:48)
n+1.
n+1 = 0, we have E(¯x(cid:48)x) = n
Since Ex(cid:48)
same reasoning but using Ex(cid:48)
n
n+1
E(x(cid:48)2x) =
(8)
n+1 ¯x. By exactly the
2 = σ2 we also obtain
x2 +
1
n + 1
σ2,
(9)
n + 1
from which the second line of (7) follows. Turning to the
first line, we obtain from (8)
E((¯x(cid:48))2x) =
(cid:0)n2(¯x)2 + n¯xEx(cid:48)
n+1)2(cid:1)
n+1 + E(x(cid:48)
1
(n + 1)2
(n + 1)2 (¯x)2 + 0 +
n2
=
1
(n + 1)2 σ2.
4
Lemma 3 (Departure): Suppose that a randomly selected
agent departs from the system. Denote x the state before
departure, x(cid:48) the state after departure and n ≥ 2 the number
of agents before departure. Then, there holds
1
(n−1)2
1
EX.
(10)
Proof: Let j be the randomly selected agent that leaves
the system. The mean is modified as
(n−1)2
(cid:32) n2−2n
(cid:33)
0
(cid:33)
− xj
xk
=
(cid:33)
(cid:17)
EX(cid:48) =
(cid:32)(cid:32) n(cid:88)
(cid:17)
k=1
¯x(cid:48) =
1
n − 1
(cid:16)
1
(n¯x − xj) . (11)
n − 1
there holds x(cid:48)2 =
jx) =
By exactly the same reasoning,
1
n−1
x2. Hence,
nx2 − x2
j
. Since j is randomly selected, E(x2
E(x(cid:48)2x) =
1
n − 1
nEx2 − Ex2
= Ex2,
(cid:16)
which implies the second line of (10). For the first line, taking
into account E(xjx) = ¯x, it follows from (11) that
E((¯x(cid:48))2x) =
(cid:0)n2(¯x)2 − 2n(¯x)E(xjx) + E(x2
jx)(cid:1)
1
(n − 1)2
n2 − 2n
(n − 1)2 (¯x)2 +
=
1
(n − 1)2 x2.
We now consider the replacement of an agent, which
consists of a departure immediately followed by an arrival.
The next result follows from a combination of Lemma 3 and
2, the latter applied to a system of size n − 1 joined by a
nth agent.
Lemma 4 (Replacement): Suppose that a randomly se-
lected agent departs from the system and is immediately
replaced by a new agent, leaving the size of the system
unchanged. Denote x the state before replacement, x(cid:48) the
state after replacement and n the constant number of agents.
Then, there holds
EX(cid:48) = ArEX + br,
where
Ar =
(cid:18) n−2
n
0
(cid:19)
1
n−1
n2
n
and br =
(12)
(cid:33)
.
(cid:32) σ2
n2
σ2
n
IV. FIXED-SIZE SYSTEM WITH RANDOM REPLACEMENT
A. Model and fixed points
We assume now that at each time step, there is a prob-
ability p that an agent is replaced, and a probability 1 − p
that a gossip step takes place. After giving an exact value
for the fixed point of the expected system behavior, we
will provide asymptotic results when the system size n is
large. Note that when considering results for large n, it is
natural to keep p constant, although other forms of scaling
could be considered. Suppose indeed that our discrete times
correspond to the sampling of a continuous-time process at
those times at which an event occurs. We suppose that the
rates of gossip and of departure (leading to a replacement) of
a single agent are both independent of the system size, which
would be natural for large systems. As a result, the rates
of gossip and replacements at the system level both scale
linearly with n, so that the probability p that a randomly
selected event is a replacement remains independent of n.
The following result, based on Lemma 4 and equation (4),
describes the expected evolution of the system.
Theorem 5: Suppose that an event occurs. This event is an
agent replacement with probability p or gossip with 1 − p.
Denote x the state before the event, x(cid:48) the state after. Then,
there holds
EX(cid:48) =
(13)
Proof: Since the probability of the events are inde-
pendent of x, the conditional expected value is computed as
follows :
(cid:18) 1 − 2p
(cid:18) p
EX + σ2
1 − 1
1−p
n
n
(cid:19)
(cid:19)
n2
p
n
p
n2
n
.
E(X(cid:48)x) = (1 − p)E(X(cid:48)gossip, x) + pE(X(cid:48)repl., x)
= (1 − p)AgX + p(ArX + br)
= ((1 − p)Ag + pAr)X + pbr.
Therefore, there holds EX(cid:48) = ((1− p)Ag + pAr)EX + pbr,
which yields (13).
One can verify that the fixed point of (13) is
E(¯x)2eq =
Ex2eq =
p + 1
σ2
p + 2n − 1
1 + p(2n − 1)
p + 2n − 1
σ2
(14)
leading to a variance EVar(x)eq = σ2 2p(n−1)
p+2n−1 .
The asymptotic values of these expressions admit some
interpretation. Suppose first
that p = 1, meaning that
no gossip ever takes place. We obtain then a variance
EVar(x)eq = σ2(1 − 1
n ), and an expected square of the
average E(x)2eq = σ2
n consistently with a process where
agents are just replaced, i.e., a system eventually consisting
of agents with n random i.i.d. values with mean 0 and
variance σ2. (This is also the fixed point of the affine
equation in Lemma 4). For p → 0, the number of gossips
steps between two replacements tends to infinity, so that a
perfect averaging takes place before any replacement. We
obtain in that case a variance EVar(x)eq = 0, and an
expected square average E(x)2eq = σ2
2n−1. This latter
number is lower than what would be obtained by averaging
n i.i.d. values. This is because it actually results from a
weighted average of the values of all agents having been
part of the system the system at some present or past time.
See Section IV.c of [6] for a detailed computation of this
value in a system with deterministic replacements.
For large n and constant p, which we have argued above
is a natural scaling, the expected square E(x)2eq goes to
0, while the variance EVar(x)eq goes to σ2p. This result is
parallel with that obtained in [6] for periodic replacement,
taking into account that the average number of gossip steps
p − 1. We will come back to
between two replacements is 1
this value σ2p in Section V.
1
5
A realization of trajectories
2 , 1
To illustrate Theorem 5, we consider an open system with
random events (replacements or gossip steps) with n = 25
agents and replacements occur with probability p = 0.05
while gossip steps occur with probability 1 − p. The system
has evolved until it has reached 100 replacements. Arriving
agent values are drawn uniformly in [− 1
2 ] so that σ2 = 1
12.
Figure 2 displays a realization of the trajectories along with
the expected dynamics for the scale-invariant quantities. In
the top plot, it appears that agents leaving the system (in
red) tend to have a more moderate state compared to agents
arriving in the system (in blue). This is due to the gossip
steps. Despite not leading to a consensus, the open system
still presents a contracting tendency. Besides, as seen in the
bottom plot, for a sufficient number of agents (here 25), the
expected variance rather well approximates the dynamics
of empirical variance realization. Also for this number of
the square mean remains small (of order 10−3)
agents,
compared to the mean square, as a consequence, the variance
is mainly due to x2. This would not be the case for n = 5
agents for instance.
The illustration provided in the introduction (Figure 1)
was obtained for an open multi-agent system of the same
kind with n = 4 agents and a replacement probability
p = 0.1 where the dynamics consider the system up to 10
replacements.
B. Convergence rate
We now study the rate at which the expected moments will
converge to the fixed points described above. Eigenvalues of
the matrix in (13) were computed on Mathematica :
(cid:16)
2n − 2p − 1±
√
∆
(cid:17)
,
r+,− =
2n
with ∆ = (1 − 2p)2 + 4p(1−p)
∆ or −√
√
. (The choice of notation
√
r+, r− comes from the use of either +
∆ in the
∆ = 1−2p+o(1),
root expression.) For large n and, since
(cid:18) 1
the eigenvalues are of order
(cid:18) 1
(cid:19)
(cid:19)
n
r+ = 1 − 2p
n
+ o
, r− = 1 − 1
n
n
+ o
,
n
which happen to be the diagonal elements of the matrix, as
if neglecting the 1/n2 term on the upper right hand side.
The corresponding eigenvectors are
(cid:19)T
+ o(1), 1
,
(cid:33)T
∆
, 1
=
(cid:32)
2p − 1 − √
2(p − 1)
(cid:32)
√
2p − 1 +
2(p − 1)
∆
(cid:18) 2p − 1
(cid:33)T
p − 1
v− =
, 1
= (o(1), 1)T ,
v+ =
and
unless p = 1/2 in which case higher order term needs to be
taken into account.
Interpretation:
The couple (r−, v−) is independent of p for large n. The
eigenvector v− has a vanishingly small component in E(x)2
Fig. 2.
Illustration of an open system with random events. (top) shows the
evolution with time of the agents values (in black) for a typical realization.
Red circles highlight the departing agents while the blue circles correspond
to the newly arrived agents. Vertical black dashed lines depict the non
uniform random replacement instants. (middle) shows the evolution of the
square mean value (realization in continuous blue line, expectation in dashed
red line). (bottom) shows the evolution of the expected variance (realization
in continuous blue line, expectation in dashed red line). Expectation were
computed using Theorem 5. The asymptotic values given in equations (14)
are provided for the expected square mean and the expected variance in red
circle at final time in the middle and bottom plots.
01860t-0.50.501860t05.5431#10-37x2ExpectationRealization01860t00.0777Var(x)ExpectationRealization 6
and concerns asymptotically exclusively Ex2 which is then
essentially equivalent to EVar(x). It follows from Lemma
1 that every gossip iteration will contract this quantity by
1 − 1
n, and from Lemma 4 that every replacement will also
contract that quantity by 1 − 1
n (when neglecting E(x)2)
before adding a constant term. If the constant term was set
to 0, Ex2 would thus contract at a rate 1 − 1
n (cid:39) r− at every
iteration, independently of whether a gossip or replacement
takes place. For large n, the couple (r−, v−) can thus be
related to the contraction of Ex2
n = (1 − 2
The couple r+, v+, on the other hand, does strongly
depend on p. In particular, r+ = 1 if p = 0. Remember
that
the square average remains unchanged when gossip
iterations occur. But when a replacement occurs, it follows
from Lemma 4 that the new square average E(x(cid:48))2 contains
a contribution n−2
n ) of the previous value, some
constant contribution related to the arriving agent, and a
vanishingly small contribution related to the average square
value (assuming that E(x)2 and Ex2 are of the same order
of magnitude, which is the case in v+ when p = 0). Hence,
if we were to set the independent term br at 0 in (12)
corresponding to σ2 = 0, the square average would be
multiplied by (1− 2
n ) at each replacement and by 1 otherwise.
Since replacement occur with a probability p, this yields a
rate p(1 − 2
n ) + (1 − p) = 1 − 2p
n (cid:39) r+.
Observe now that there is a transition at p = 1
2. Indeed, for
2, corresponding to frequent replacements, the largest
p > 1
eigenvalue is r−. The rate of convergence to the steady state
depends thus on the rate of variance reduction, which is
independent of p (for large n). On the other hand, for p <
2, the convergence is generally dominated by phenomena
1
related to the convergence of the the square average, which
does depend on p as this quantity only changes when a
replacement takes place.
V. GROWING SYSTEM WITHOUT DEPARTURE
We focus now on systems whose sizes grow unbounded
because new agents keep joining while no agent ever leaves:
Similarly to the fixed-size model described in Section IV-A,
at each time, a new agent joins the system with probability
pn or a gossip step occurs with probability 1 − pn, where
n is the number of agent in the system. The discussion of
the dependence of pn on n is deferred to the end of the
section. When no ambiguity is possible, we will simply use
p. We assume again that the initial value of every agent when
joining the system is a random variable with zero mean and
variance σ2. We denote by tn the time just after the arrival of
the n-th agent, and we let Kn := tn+1−tn−1 be the number
of gossip steps taking place between the arrival of agents n
and n + 1. Both the tn and Kn are random variables, and
the Kn follow a geometric distribution with parameter pn so
that P(Kn = k) = (1−pn)kpn. Since we are only interested
in expected quantities and that the set of sequences tn which
are bounded has probability 0, we assume in the sequel that
sequence tn is unbounded.
We will focus on the values of the expected average of
the square E(x2) and the expected square of the average
+∞(cid:88)
+∞(cid:88)
k=0
E(X(cid:48)x) =
E(X(cid:48)Kn = k, x)P(Kn = k)
+∞(cid:88)
E(¯x2) at the times tn, just after the arrivals of the n-th
agents. We will see that the evolution of these values can be
described by a two-dimensional linear system. This system is
here time-varying because n is not constant, but the absence
of departures makes it triangular, and hence easier to analyze.
We first provide the evolution of the expectation of vector
X = (¯x2, x2)T between two agent arrivals.
Lemma 6: Let x be the state of the system after the arrival
of the nth agent, and suppose that at every time-step, an
arrival takes place with probability p and a gossip with
probability 1 − p. Let then x(cid:48) be the state after the arrival
of agent n + 1, so after the first arrival event. There holds
EX(cid:48) = AGEX where
AG =
1 − γ
0
γ
with γ =
n
n − 1 + 1
.
p
Proof: To obtain the expectation of vector X(cid:48) after
the gossip, we condition with regards to Kn which is by
definition the number of gossip steps having taken place. By
Lemma 1, there holds
(cid:18) 1
(cid:19)
g X(1 − p)kp = p
Ak
=
= p(I − (1 − p)Ag)−1X = AGX.
k=0
k=0
(1 − p)kAk
g X
Recalling that E(E(X(cid:48)x)) = E(X(cid:48)), the expectation of the
previous equation provides the result.
A consequence of this lemma is that between two agent
arrivals, the expected variance evolves autonomously as
EVar(x(cid:48)) = γEVar(x).
(15)
Applying Lemma 2, we obtain the evolution of the variance
after agent arrivals.
Lemma 7: Recall that x(tn) is the state of the system just
after the arrival of the n-th agent and x(tn+1) the state just
after the arrival of the n + 1-th agent, so that x(tn+1) is
obtained starting from state x(tn) and applying Kn gossip
steps and an arrival. Then, there holds
E(¯x2(tn)) =
σ2
n
,
(16)
(n + 1)EVar(x(tn+1)) = γnEVar(x(tn)) + σ2.
(17)
Proof: Applying first Lemma 6 and then Lemma 2, we
obtain
EX(tn+1) = AaAGEX(tn) + ba.
Since Aa is diagonal and AG is lower triangular, the first
line of the previous equation provides
E(¯x2(tn+1)) =
n2
(n + 1)2
E(¯x2(tn)) +
σ2
(1 + n)2 ,
which, recalling that initially E(¯x2(0)) = 0, grants equa-
tion (16) by induction on n. To obtain equation (17), recall
that Var(x) = (−1, 1)X so that EVar(x(tn+1))
=(−1, 1)(AaAGEX(tn) + ba)
n
(cid:19)
n
, γ
n + 1
n + 1
EX(tn)
(cid:18)
(cid:18)
=
=
=
− n2
(n + 1)2 + (1 − γ)
(cid:19)
+ (−1, 1)ba
− n2
(n + 1)2 +
+ (−1, 1)ba
n + 1
n
E(¯x2(tn)) +
nγ
n + 1
EVar(x(tn))
n
(n + 1)2
E(¯x2(tn)) +
nγ
n + 1
EVar(x(tn)) +
n
(n + 1)2 σ2.
We conclude using equation (16).
This recursion allows obtaining the following theorem
characterizing the asymptotic variance, and proved in Ap-
pendix A.
Theorem 8: Consider the growing system without depar-
ture, and remember that when the system with n agents
undergoes an event, pn describes the probability of the n+1-
th agent arrival rather than a gossip step.
(i) If pn = p > 0 for all n ≥ n0 for some n0, then
EVar(x(tn)) = pσ2.
lim
n→∞
(ii) If limn→∞ pn = 0, then limn→∞ EVar(x(tn)) = 0.
pn
(cid:80)n
E(Km), which grows unbounded.
Theorem 8(ii) shows that the system essentially converges
to a consensus as soon as pn goes to zero, even if this
convergence is very slow. This can also be interpreted
in terms of E(Kn), the expected number of gossip steps
between two arrivals. Since Kn follows a geometric law of
which diverges to +∞ when
parameter pn, E(Kn) = 1−pn
pn goes to zero. As a consequence, the system converges
to a consensus as soon as the expected number of gossip
steps between two consecutive arrivals diverges, and even
if the number E(Kn)/n of gossips per agent between two
consecutive arrivals tends to 0. Note, however, that each
agent gets involved (with probability 1) in infinitely many
gossips when E(Kn) → ∞. The expected number of gossips
in which an agent has been involved at time tn is indeed
2 1
n
By contrast, in the case of a fixed probability pn = p
corresponding to a fixed E(Kn) =: K with p = 1
K+1, agents
E(Km) = 2K
have on average been involved in 2 1
n
gossips after any given arrival, which intuitively explains
why the variance stays bounded away from 0. But the actual
asymptotic value pσ2 = σ2
K+1 obtained in Theorem 8(i) is
remarkably high, and is actually the same as that obtained
for the fixed-size system in Section IV-A for large n. As a
basis for comparison, suppose we had first waited until the n
independent agents were present in the system, which would
yield an expected variance σ2 n−1
n , and then performed the
same number nK of gossip averaging operations between
randomly selected pairs of nodes. It follows from application
of equation (5) that the expected variance would then have
been
(cid:80)n
m=1
m=1
(cid:19)nK →n→∞ σ2e−K,
(cid:18)
n − 1
n
σ2
1 − 1
n
7
which is significantly lower than σ2/(K + 1) (for K = 5,
1/6 (cid:39) 0.04). The dynamics
the ratio of variance would be e−5
of the system composition deteriorates thus considerably the
performances in terms of variance reduction.
We now propose possible interpretations of the evolutions
of pn and E(Kn) with the number of agents n. Suppose
that we interpret our discrete t as the sampling of a real
continuous time variable τ at those times τt at which an
event occurs. It is again reasonable to assume the interaction
rate of an agent to be independent of the system size, so that
the total number of gossips per unit of time τ would grow
linearly with n, as say λgn. Suppose first that the agents
arrive at a fixed rate λa. In that case, the probability of agent
arrival is λa/(λa + λgn) and the number of gossips between
two arrivals would be linearly growing with n and E(Kn) =
nλg/λa. Theorem 8(ii) shows then that the variance would
converge to 0.
But one could also imagine a linearly growing rate of ar-
rivals λrn. This would for example be the case if the system
attraction were growing with its size or if the arrivals resulted
from some form of reproduction process. The probability of
agent arrival would then be constant pn = p = λa/(λa +λg)
and so will be the number of gossip iterations between two
arrivals EKn = K = (λgn)/(λrn), leading to a finite
variance pσ2.
VI. CONCLUSIONS
We have made first steps in the analysis of open multi-
agent systems, where agents can leave and arrive. We have
focused on analysing open systems subject to a classical
multi-agent algorithm : on all-to-all pairwise gossips. We
have shown that these systems could be characterized by
fixed-size linear systems in terms of some of the moments.
Interestingly, we have also observed that the open character
of the system may result in a significant performance reduc-
tion in terms of variance reduction.
Ongoing works include the generalization of this approach
to systems where arrivals and departure follow more complex
patterns, or to more complex interactions, such as gossips
restricted to a graph.
Another challenge left untackled so far is the charac-
terization of the variability for individual realizations. Our
results characterize for instance the expected value of the
disagreement in the system EVar(x(t)) where Var(x(t)) =
i∈N (t)(xi(t) − ¯x(t))2, but do not directly allow de-
n(t)
ducing the width of the probability distribution of Var(x(t)).
(cid:80)
1
REFERENCES
[1] Dana Angluin, James Aspnes, Michael J Fischer, and Hong Jiang. Self-
stabilizing population protocols. ACM Transactions on Autonomous and
Adaptive Systems (TAAS), 3(4):13, 2008.
[2] Stephen Boyd, Arpita Ghosh, Balaji Prabhakar, and Devavrat Shah.
Randomized gossip algorithms. IEEE/ACM Transactions on Networking
(TON), 14(SI):2508–2530, 2006.
[3] Carlos Carrascosa, Adriana Giret, Vicente Julian, Miguel Rebollo,
Service oriented mas: an open
Estefania Argente, and V Botti.
In Proceedings of The 8th International Conference on
architecture.
Autonomous Agents and Multiagent Systems-Volume 2, pages 1291–
1292. International Foundation for Autonomous Agents and Multiagent
Systems, 2009.
and hence, noticing Wn ≥ 0, (19) implies
Wn ≤ Wn0
(cid:18) n0 + q
(cid:19)q
n(cid:88)
σ2
+
n + q
(n + q)q
(s + q)q.
(21)
By a change of variable, the sum in the last term can be is
rewritten as
s=n0+1
sq ≤ n+q(cid:88)
s=0
sq
xqdx ≤ (n + q + 1)q+1
q + 1
n(cid:88)
s=n0+1
s=n0+q+1
n+q(cid:88)
(cid:90) n+q+1
(cid:19)q
0
(s + q)q =
≤
(cid:18) n0 + q
n + q
Introducing this in equation (21), we obtain
Wn ≤ Wn0
σ2
+
(n + q + 1)q+1
.
(22)
(n + q)q
q + 1
The first term in (22) decays to 0 when n grows. Hence
lim sup Wn
log(1 − 1
Part (b) follows a parallel reasoning using the upper bound
n ≤ σ2
n ) ≥ − 1
n−1.
q+1 = pσ2.
8
[4] Carole Delporte-Gallet, Hugues Fauconnier, Rachid Guerraoui, and Eric
Ruppert. When birds die: Making population protocols fault-tolerant. In
International Conference on Distributed Computing in Sensor Systems,
pages 51–66. Springer, 2006.
[5] Fabio Fagnani and Sandro Zampieri. Randomized consensus algorithms
over large scale networks. IEEE Journal on Selected Areas in Commu-
nications, 26(4):634–649, May 2008.
[6] Julien M. Hendrickx and Samuel Martin. Open multi-agent systems:
54th Annual
Gossiping with deterministic arrivals and departures.
Allerton Conference on Communication, Control, and Computing, 2016.
[7] Trung Dong Huynh, Nicholas R Jennings, and Nigel R Shadbolt. An
integrated trust and reputation model for open multi-agent systems.
Autonomous Agents and Multi-Agent Systems, 13(2):119–154, 2006.
[8] Gerardo Iniguez, J´anos Torok, Taha Yasseri, Kimmo Kaski, and J´anos
Kert´esz. Modeling social dynamics in a collaborative environment. EPJ
Data Science, 3(1):1, 2014.
[9] J´anos Torok, Gerardo Iniguez, Taha Yasseri, Maxi San Miguel, Kimmo
Kaski, and J´anos Kert´esz. Opinions, conflicts, and consensus: modeling
social dynamics in a collaborative environment. Physical review letters,
110(8):088701, 2013.
APPENDIX
A. Proof of Theorem 8
We prove the following proposition, which implies Theo-
Proposition 9: Let n0 ≥ 2. Let γn =
rem 8 when applied to Wn = nEVar(X(tn)).
n−1+ 1
with pn ∈
(0, 1). Let p ∈ (0, 1) and consider the sequence defined by
W1 = 0 and
pn
n
Wn+1 = Wnγn + σ2.
a) If pn ≤ p for all n ≥ n0, then lim supn→∞ Wn
b) If pn ≥ p for all n ≥ n0, then lim inf n→∞ Wn
(18)
n ≤ pσ2
n ≥ pσ2.
Proof: We prove the statement (a) of the proposition.
Statement (b) can be obtained in a similar way. It follows
from (18) that
Wn = Wn0 Πn−1
m=n0
γm + σ2
Πn−1
m=sγm,
(19)
n(cid:88)
s=n0+1
with the convention Πn−1
We use notation q = 1
m=nγm = 1.
γn = n
n+qn
there holds
= 1 − qn
n+qn
p − 1 and qn = 1
− 1 so that
. Since pn ≤ p, qn ≥ q > 0 and
pn
γn ≤ 1 − n
n + q
.
n+q . Using log(1 − xn) ≤ −xn (for xn < 1),
n+q−1(cid:88)
xm = −q
1
m
.
(20)
m=s
m=s+q
1
m is an upper approximation of the
≥ log(n+q)−log(s+q) ≥ log
n + q
s + q
.
(cid:1) ≤ − n−1(cid:88)
Denote xn = n
we obtain for s ≥ n0
log(cid:0)Πn−1
Observe that (cid:80)n+q−1
integral(cid:82) n+q
m=sγm
m=s+q
1
x=s+q
1
s + q
+···+
n + q − 1
x dx, and hence
1
Reintroducing this in (20) yields
(cid:18) s + q
(cid:19)q
n + q
m=sγm ≤
Πn−1
∀s ≥ n0,
,
|
1708.02357 | 1 | 1708 | 2017-08-08T02:38:18 | Towards A Novel Unified Framework for Developing Formal, Network and Validated Agent-Based Simulation Models of Complex Adaptive Systems | [
"cs.MA",
"cs.AI",
"cs.NI",
"cs.SI",
"nlin.AO"
] | Literature on the modeling and simulation of complex adaptive systems (cas) has primarily advanced vertically in different scientific domains with scientists developing a variety of domain-specific approaches and applications. However, while cas researchers are inher-ently interested in an interdisciplinary comparison of models, to the best of our knowledge, there is currently no single unified framework for facilitating the development, comparison, communication and validation of models across different scientific domains. In this thesis, we propose first steps towards such a unified framework using a combination of agent-based and complex network-based modeling approaches and guidelines formulated in the form of a set of four levels of usage, which allow multidisciplinary researchers to adopt a suitable framework level on the basis of available data types, their research study objectives and expected outcomes, thus allowing them to better plan and conduct their respective re-search case studies. | cs.MA | cs | Towards A Novel Unified Framework for Developing
Formal, Network and Validated Agent-Based
Simulation Models of Complex Adaptive Systems
Muaz Ahmed Khan Niazi
Computing Science and Mathematics
School of Natural Sciences
University of Stirling
Scotland UK
This thesis has been submitted to the University of Stirling
In partial fulfillment for the degree of Doctor of Philosophy
2011
Abstract
Literature on the modeling and simulation of complex adaptive systems (cas) has pri-
marily advanced vertically in different scientific domains with scientists developing a
variety of domain-specific approaches and applications. However, while cas researchers
are inherently interested in an interdisciplinary comparison of models, to the best of our
knowledge, there is currently no single unified framework for facilitating the development,
comparison, communication and validation of models across different scientific domains.
In this thesis, we propose first steps towards such a unified framework using a combination
of agent-based and complex network-based modeling approaches and guidelines formulat-
ed in the form of a set of four levels of usage, which allow multidisciplinary researchers to
adopt a suitable framework level on the basis of available data types, their research study
objectives and expected outcomes, thus allowing them to better plan and conduct their re-
spective research case studies.
Firstly, the complex network modeling level of the proposed framework entails the de-
velopment of appropriate complex network models for the case where interaction data of
cas components is available, with the aim of detecting emergent patterns in the cas under
study. The exploratory agent-based modeling level of the proposed framework allows for
the development of proof-of-concept models for the cas system, primarily for purposes of
exploring feasibility of further research. Descriptive agent-based modeling level of the
proposed framework allows for the use of a formal step-by-step approach for developing
agent-based models coupled with a quantitative complex network and pseudocode-based
specification of the model, which will, in turn, facilitate interdisciplinary cas model com-
parison and knowledge transfer. Finally, the validated agent-based modeling level of the
proposed framework is concerned with the building of in-simulation verification and vali-
dation of agent-based models using a proposed Virtual Overlay Multiagent System
approach for use in a systematic team-oriented approach to developing models. The pro-
posed framework is evaluated and validated using seven detailed case study examples
selected from various scientific domains including ecology, social sciences and a range of
complex adaptive communication networks. The successful case studies demonstrate the
potential of the framework in appealing to multidisciplinary researchers as a methodologi-
cal approach to the modeling and simulation of cas by facilitating effective communication
- ii -
and knowledge transfer across scientific disciplines without the requirement of extensive
learning curves.
- iii -
Declaration
I, Muaz Niazi hereby declare that this work has not been submitted for any other degree
at this University or any other institution and that, except where reference is made to the
work of other authors, the material presented is original. Some portions of the thesis chap-
ters have been published as follows:
• Parts of Chapter 3 in [2], [3].
• Parts of Chapter 4 in [4].
• Parts of Chapter 5 in [5], [6].
• Parts of Chapter 6 in [7-9]. Qasim Siddique helped in some of the simulation
experiments. Some ideas of research simulation were formulated based on ad-
vice from Dr. Saeed Bhatti and Dr. Abdul Rauf Baig.
Muaz Niazi
- iv -
Acknowledgements
This thesis would not have been possible without the help and support of a large number of
individuals. First and foremost, I would like to thank my family members, especially my
beloved parents, who have endured my absence during my research and helped me tre-
mendously in all ways possible. Without your continued help, support and guidance, this
would never have been possible. Thank you.
I am deeply indebted to my principal supervisor and founding Head of our COSIPRA
Lab, Dr. Amir Hussain who provided stimulating advice, guidance and encouragement to
me every step of the way. I would like to thank Kamran, Dr. Erfu, Thomas, Rozniza, Erik
and Andrew from the COSIPRA lab. I would also like to thank Kerstin Rosee who, on be-
half of Dr. Amir, would keep me on my toes by requiring regular updates. Writing updates
always pushed me, by forcing me to quantify my progress and obtain critical feedback on a
periodic basis. I would also like to thank Ian MacLellan from the International Office. Fi-
nally, I would especially like to thank my Examiners Dr. Keshav Dahal (from Bradford
University) and Dr. Savitri Maharaj for their constructive criticism and helpful comments,
which helped me considerably in shaping up the thesis in a much better way than I had
originally planned.
- v -
Table of Contents
Abstract ................................................................................................................................. ii
Declaration ........................................................................................................................... iv
Acknowledgements ............................................................................................................... v
Table of Contents ................................................................................................................. vi
List of Figures ..................................................................................................................... xii
List of Tables ...................................................................................................................... xvi
List of Abbreviations ......................................................................................................... xvii
1 Introduction ...................................................................................................................... 1
1.1 Modeling cas ............................................................................................................ 2
1.2 Motivation ................................................................................................................ 9
1.3 Aims and Objectives .............................................................................................. 10
1.4 Original Contributions ........................................................................................... 11
1.5 Proposed framework for the Modeling and Simulation of cas ............................... 13
1.5.1 Overview of the proposed framework .............................................................. 13
1.5.2 Proposed framework levels formulated in terms of cas study objectives ......... 15
1.5.3 Proposed framework levels formulated in relation to available data types ....... 18
1.6 Publications ............................................................................................................ 19
1.6.1 Refereed Conferences and Workshops .............................................................. 19
1.6.2 Published Journal papers ................................................................................... 19
1.7 Overview of the thesis ............................................................................................ 20
1.7.1 Overview of case studies .................................................................................. 20
1.7.2 Outline of the Thesis ......................................................................................... 21
2 Complex Adaptive Systems: Background and Related Work ........................................ 23
2.1 Overview ................................................................................................................ 23
2.2 Complex Adaptive Systems (cas)........................................................................... 23
2.2.1 The Seven Basics of cas .................................................................................... 26
2.2.2 Emergence ......................................................................................................... 29
2.3 Examples of cas ..................................................................................................... 30
2.3.1 Natural cas example 1: Cas in plants ................................................................ 30
2.3.2 Natural cas example 2: Cas in social systems ................................................... 31
2.3.3 Artificial cas example 1: Complex adaptive communication networks ............ 34
2.3.4 Artificial cas example 2: Simulation of flocking boids .................................... 36
- vi -
2.4 Modeling cas .......................................................................................................... 37
2.4.1 Complex Network Modeling ............................................................................ 38
2.4.1.1 Complex Network Methods ...................................................................... 39
2.4.1.2 Theoretical basis ........................................................................................ 40
2.4.1.3 Centralities and other Network measures .................................................. 41
2.4.1.4 Software tools for Complex Networks ...................................................... 44
2.4.2 Agent-based Modeling and Agent-based Computing ....................................... 45
2.4.2.1 Agent-Oriented Programming ................................................................... 45
2.4.2.2 Multi-agent oriented Programming ........................................................... 45
2.4.2.3 Agent-based Or Massively multiagent Modeling ..................................... 46
2.4.2.4 Benefits of Agent-based Thinking ............................................................. 47
2.5 A Review of Agent-based Tools ............................................................................ 48
2.5.1 NetLogo Simulation: An overview ................................................................... 48
2.5.2 Overview of NetLogo for Modeling Complex interaction protocols ................ 53
2.5.3 Capabilities in handling a range of input values ............................................... 55
2.5.4 Range of statistics and complex metrics ........................................................... 55
2.6 Verification and Validation of simulation models .................................................. 56
2.6.1 Overview ........................................................................................................... 56
2.6.2 Verification and Validation of ABMs ................................................................ 56
2.6.3 Related work on V&V of ABM ........................................................................ 58
2.7 Overview of communication network simulators .................................................. 58
2.7.1 Simulation of WSNs ......................................................................................... 58
2.7.2 Simulation of P2P networks .............................................................................. 59
2.7.3 Simulation of Robotic Swarms ......................................................................... 59
2.7.4 ABM for complex communication networks simulation .................................. 59
2.8 Conclusions ............................................................................................................ 59
3 Proposed Framework: Complex Network Modeling Level ........................................... 60
3.1 Overview ................................................................................................................ 60
3.2 Generalized Methodology ...................................................................................... 61
3.3 Case Study I: Agent-based Computing .................................................................. 63
3.3.1 Problem Statement: ........................................................................................... 64
3.3.2 Methodology ..................................................................................................... 65
3.3.2.1 Data collection .......................................................................................... 66
3.3.3 Results ............................................................................................................... 66
3.3.3.1 Network Workbench (NWB) ..................................................................... 68
3.3.3.2 CiteScape ................................................................................................... 70
3.4 Case Study II: Consumer Electronics..................................................................... 86
- vii -
3.4.1 Problem Statement ............................................................................................ 86
3.4.2 Methodology ..................................................................................................... 87
3.4.3 Results ............................................................................................................... 87
3.5 Summary of results ................................................................................................ 91
3.5.1 Case Study I ...................................................................................................... 92
3.5.2 Case Study II ..................................................................................................... 93
3.6 Broad applicability of the proposed framework level ............................................ 93
3.7 Conclusions ............................................................................................................ 94
4 Proposed Framework: Exploratory Agent-based Modeling Level ................................ 95
4.1 Research Methodology .......................................................................................... 95
4.2 Case Study: Overview, Experimental design and Implementation ........................ 98
4.2.1 Goals definition ............................................................................................... 103
4.2.2 Agent design ................................................................................................... 105
4.2.2.1 Computing device agents ........................................................................ 106
4.2.2.2 Message Agent ........................................................................................ 108
4.2.3 Learning and adaptation in the system ............................................................ 109
4.2.4 Implementation description ............................................................................ 110
4.2.4.1 Global variables....................................................................................... 110
4.2.4.2 Breeds ...................................................................................................... 110
4.2.4.3 Setup ........................................................................................................ 112
4.2.4.4 Make Computing Agents ......................................................................... 113
4.2.4.5 Make Gateway nodes .............................................................................. 113
4.2.4.6 Adjust locations ....................................................................................... 113
4.2.4.7 Searches ................................................................................................... 114
4.2.4.8 Make Search ............................................................................................ 115
4.2.4.9 Make-s-Node ........................................................................................... 115
4.2.4.10 Setup-query ............................................................................................. 115
4.2.4.11 Goals ....................................................................................................... 116
4.2.4.12 Make-Goal ............................................................................................... 116
4.2.4.13 Make-g-Node .......................................................................................... 116
4.2.4.14 Setup-sacs ................................................................................................ 116
4.2.4.15 Setup-sacs-d ............................................................................................ 117
4.2.4.16 Move ....................................................................................................... 119
4.2.4.17 Move-rw .................................................................................................. 119
4.2.4.18 Check-Goal ............................................................................................. 120
4.2.4.19 Move-Devices ......................................................................................... 121
4.3 Simulation Results and discussion ....................................................................... 121
- viii -
4.3.1 Metrics used .................................................................................................... 122
4.3.1.1 Simulation parameters: ............................................................................ 122
4.3.1.2 Output metrics: ........................................................................................ 122
4.3.2 Details of experimentation .............................................................................. 123
4.3.2.1 Discussion of results................................................................................ 125
4.3.3 Discussion of Intelligence in cas Agents ......................................................... 135
4.4 Conclusions .......................................................................................................... 136
5 Proposed Framework: Descriptive Agent-based Modeling Level ............................... 137
5.1 Research Methodology ........................................................................................ 141
5.1.1 An overview of DREAM ................................................................................ 142
5.1.2 Design of the Complex Network Model ......................................................... 148
5.1.3 Analysis of the network sub-model ................................................................ 150
5.1.4 Goals of a pseudocode-based specification model.......................................... 152
5.1.5 Templates for the specification model ............................................................ 153
5.2 Case study: Overview, Experimental design and Implementation ....................... 156
5.2.1 Overview ......................................................................................................... 156
5.2.1.1 Problem statement ................................................................................... 157
5.2.2 Agent design ................................................................................................... 158
5.2.2.1 Boid agents .............................................................................................. 159
5.2.2.2 Node agents ............................................................................................. 160
5.2.3 Network Model ............................................................................................... 161
5.2.4 Pseudocode-based specification ...................................................................... 163
5.2.4.1 Agents and Breeds ................................................................................... 163
5.2.4.2 Globals .................................................................................................... 165
5.2.4.3 Procedures ............................................................................................... 166
5.2.5 Experiments .................................................................................................... 173
5.3 Results and Discussion ......................................................................................... 174
5.3.1 Complex Network Analysis of the Network sub-model ................................. 174
5.3.2 Simulation Measurements ............................................................................... 179
5.3.3 Results from Simulation Visualization............................................................ 180
5.3.4 Quantitative results ......................................................................................... 182
5.3.5 Graphs ............................................................................................................. 184
5.3.6 Significance of findings .................................................................................. 189
5.3.7 Critical Comparison of DREAM with other approaches ................................ 190
5.3.7.1 Template Software approach ................................................................... 190
5.3.7.2 ODD approach......................................................................................... 191
5.4 Conclusions .......................................................................................................... 193
- ix -
6 Proposed Framework: Validated Agent-based Modeling Level ................................... 194
6.1 Problem statement ................................................................................................ 195
6.2 Proposed validate agent-based modeling methodology ....................................... 196
6.2.1 Detailed overview of origins of VOMAS concepts ........................................ 198
6.2.1.1 VOMAS basis in Software Engineering Team concepts ......................... 199
6.2.1.2 VOMAS basis in software engineering concepts .................................... 200
6.2.1.3 VOMAS basis in multiagent system concepts ........................................ 202
6.2.2 A Taxonomy of Agent-Based Validation techniques using VOMAS .............. 204
6.3 Case studies: Experimental Design and Implementation ..................................... 206
6.3.1 Case Study I: Ecological modeling of forest fire spread and forest re-growth207
6.3.1.1 Basic simulation Model ........................................................................... 207
6.3.1.2 Advanced Forest Fire model version 1.................................................... 210
6.3.1.3 Advanced Forest Fire model version 2.................................................... 212
6.3.1.4 Validation Question ................................................................................. 214
6.3.1.5 Software Contract .................................................................................... 216
6.3.1.6 VOMAS agent design ............................................................................. 216
6.3.2 Case Study II: Environmental Monitoring using Wireless Sensor Networks . 221
6.3.2.1 Validation Question ................................................................................. 223
6.3.2.2 Software Contract .................................................................................... 224
6.3.2.3 VOMAS agent design ............................................................................. 224
6.3.3 Case Study III: Researcher Growth Simulation .............................................. 224
6.3.3.1 Research Process I: Publication .............................................................. 225
6.3.3.2 Research Process II: Citations ................................................................. 226
6.3.3.3 Agent Description ................................................................................... 228
6.3.3.4 Validation Question ................................................................................. 229
6.3.3.5 VOMAS agent design ............................................................................. 230
6.4 Results and Discussion ......................................................................................... 230
6.4.1 Case Study I .................................................................................................... 231
6.4.1.1 Simulation Parameters............................................................................. 231
6.4.1.2 Validation discussion ............................................................................... 232
6.4.2 Case Study II ................................................................................................... 236
6.4.2.1 Simulation Parameters............................................................................. 237
6.4.2.2 Validation discussion ............................................................................... 238
6.4.3 Case Study III ................................................................................................. 243
6.4.3.1 Simulation parameters ............................................................................. 243
6.4.3.2 Validation of the representational abilities of TCN models .................... 244
6.4.3.3 Validation of Hirsch Index Calculation ................................................... 245
- x -
6.4.4 Critical comparison of VOMAS with other approaches ................................. 248
6.4.4.1 Empirical Validation ................................................................................ 248
6.4.4.2 Validation using Companion Modeling Approach .................................. 249
6.5 Conclusions .......................................................................................................... 250
7 Conclusions and Future Work ...................................................................................... 251
7.1 Overview .............................................................................................................. 251
7.2 Research Contributions ........................................................................................ 252
7.2.1 Complex Network Modeling Level for Interaction data ................................. 252
7.2.2 Exploratory agent-based modeling level ......................................................... 253
7.2.3 Descriptive agent-based modeling level ......................................................... 253
7.2.4 Validated agent-based modeling level ............................................................. 254
7.3 Critical review of the proposed levels .................................................................. 255
7.3.1 Critical Comparison of DREAM with other approaches ................................ 255
7.3.1.1 Template Software approach ................................................................... 256
7.3.1.2 ODD approach......................................................................................... 256
7.3.2 Critical comparison of VOMAS with other approaches ................................. 257
7.3.2.1 Empirical Validation ................................................................................ 257
7.3.2.2 Validation using Companion Modeling Approach .................................. 258
7.4 Limitation of the Methods .................................................................................... 258
7.4.1 Limitations and proposed extensions of the framework levels ....................... 259
7.4.2 Selection of Case studies ................................................................................ 259
7.4.3 Selection of Agent-based Modeling Tool ........................................................ 260
7.5 Proposed Future directions ................................................................................... 261
7.6 Final Words .......................................................................................................... 263
Appendix 1 ........................................................................................................................ 264
References ......................................................................................................................... 265
- xi -
List of Figures
Figure 1: An overview of the decision-making process for choosing framework levels in relation to
cas research study objectives ................................................................................................... 17
Figure 2: Data Driven View of Proposed Framework ...................................................................... 18
Figure 3: Peculiar shape and striations in a melon ........................................................................... 27
Figure 4: Tree stem adapting structure temporally........................................................................... 30
Figure 5: Complex Network model showing different Journals publishing articles related to agent-
based modeling. Bigger caption represents a higher centrality measure explained later in the
chapter. ..................................................................................................................................... 34
Figure 6: Artificial Global cas formed by P2P clients observed using utorrent software ................ 36
Figure 7 Code for setup .................................................................................................................... 52
Figure 8 Code for go ........................................................................................................................ 52
Figure 9: Mobile Nodes ................................................................................................................... 53
Figure 10 Proposed Methodology for the complex network modeling level (for the discovery of
emergent patterns based on available interaction data of cas components) ............................. 62
Figure 11: Articles published yearly ................................................................................................ 67
Figure 12: Citations per year ............................................................................................................ 67
Figure 13: Top papers with citations > 20 ........................................................................................ 70
Figure 14: Largest Cluster based on indexing terms ........................................................................ 71
Figure 15: Timeline view of terms and clusters of index terms (based on centrality) also showing
citation bursts ........................................................................................................................... 73
Figure 16: Key Journals based on centrality .................................................................................... 74
Figure 17: Complex network model of the category data ................................................................ 77
Figure 18: Co-Author network visualization ................................................................................... 80
Figure 19: Authors in terms of frequency ........................................................................................ 82
Figure 20: Countries with respect to centrality ................................................................................ 83
Figure 21: Timeline based visualization of institutes ....................................................................... 84
Figure 22: Frequency of papers in Consumer Electronics domain .................................................. 87
Figure 23: Visualization of the citation network of the top Journals of the domain ........................ 88
Figure 24: Network of top papers in Consumer Electronics ............................................................ 89
Figure 25: Co-citation network in Consumer Electronics ................................................................ 91
Figure 26 Exploratory agent-based modeling level of the proposed framework (to assess research
feasibility and proof-of-concept) .............................................................................................. 97
Figure 27: Examples of Internet of Things .................................................................................... 102
- xii -
Figure 28: Challenges in hybrid network design for Internet of things by using unstructured P2P
algorithms for searching content in personal computing devices .......................................... 103
Figure 29 Establishment of a gradient by SACS sources............................................................... 104
Figure 30 Query algorithm flowchart ............................................................................................ 105
Figure 31 State chart for Computing device agent ......................................................................... 107
Figure 32: User interface showing various configurable values for the simulation ........................ 111
Figure 33: Description of the setup function ................................................................................... 112
Figure 34: Effects of the execution of the setup function with 2500 computing devices ............... 114
Figure 35: View of the screen after the execution of the setup, Searches and the Goals functions 117
Figure 36: Result of execution of SACS setup algorithm ............................................................... 118
Figure 37 Flowchart for move function. ......................................................................................... 119
Figure 38: Flow chart for the move-rw function ............................................................................ 121
Figure 39: 95% Confidence interval error plot for successful queries plotted against the increasing
sacs-radius .............................................................................................................................. 126
Figure 40: 95% Confidence interval plot for hop count (cost) vs an increase of the SACS-radius
................................................................................................................................................ 127
Figure 41: Scatter graph showing the variation of the cost in terms of hop count with a change in
max-ttl and sacs-radius ........................................................................................................... 129
Figure 42: Mean successful queries with a variation in max-ttl value and a variation in colors based
on different sacs-radius values ............................................................................................... 130
Figure 43: 95% confidence interval plot showing the effects of increasing sacs-radius on the
number of successful queries in the case of node mobility .................................................... 131
Figure 44: 95% confidence interval plot for the cost in terms of hop count with a variation of the
sacs-radius value in mobility of computing devices scenario ................................................ 132
Figure 45: Scatter plot of hop count cost with a variation in max-ttl value color coded according to
sacs-radius in mobility modeling ........................................................................................... 133
Figure 46: Average values of successful queries with respect to a variation in max-ttl value color
coded according to sacs-radius in the case of using mobility modeling ................................ 134
Figure 47: Descriptive agent-based modeling level ....................................................................... 142
Figure 48: An overview of ABM constructs in relation with the specification model constructs.. 144
Figure 49: Pictorial representation of the DREAM Methodology ................................................. 147
Figure 50: Baseline Network template Model of a complete ABM ............................................... 148
Figure 51: State chart diagram for sensor node agents .................................................................. 161
Figure 52: Initial view of imported network .................................................................................. 162
Figure 53: Network view after manipulation algorithms ............................................................... 163
Figure 54: Complex Network Model of Boids and WSN simulation nodes resized and colorized
according to degree centrality ................................................................................................ 175
- xiii -
Figure 55: Plot showing the Degree Centrality .............................................................................. 177
Figure 56: Plot showing the eccentricity centrality ........................................................................ 178
Figure 57: Plot showing the betweenness centrality measure ........................................................ 179
Figure 58:(a) shows unflocked "boids" while (b) shows flocked "boids" and the sensor nodes
sensing the boids, show up as black on the screen. ................................................................ 181
Figure 59: Sensed Motion tracks of "boids" showing emergence of flocking behavior (a) Unflocked
(b) Flocked ............................................................................................................................. 182
Figure 60: 95% Confidence Interval plot of sensed vs. number of boids ...................................... 185
Figure 61: Mean value of sensed over time with different number of boids ................................. 186
Figure 62: 95% Confidence Interval graph of Sensed with a variation in the number of sensors . 187
Figure 63: Mean Sensed value vs. simulation time with a variation in the number of sensors ..... 188
Figure 64: A view of the modeling process, figure adapted from Sargent [203] ........................... 197
Figure 65 Validated agent-based modeling framework level ......................................................... 198
Figure 66: Design and testing of VOMAS Invariants by means of manipulation of the pre-
conditions ............................................................................................................................... 202
Figure 67: Methodology of building In-Simulation Validated ABMs ........................................... 203
Figure 68: A Taxonomy of Agent Based Validation techniques ..................................................... 205
Figure 69: View of the forest before the fire .................................................................................. 209
Figure 70: Forest fire spread .......................................................................................................... 210
Figure 71 Baseline forest fire model with two consecutive fires .................................................... 211
Figure 72 Advanced forest fire model version 1 with tree re-growth ............................................ 212
Figure 73 Advanced forest fire model version 2 with rainfall in part of the forest ........................ 213
Figure 74 Advanced forest fire model version 2 with snowfall effects .......................................... 214
Figure 75: FWI calculation, figure adapted from Wagner[1] ......................................................... 217
Figure 76: WSN network simulation as QUDG with n=200 sensor nodes .................................... 223
Figure 77: Research Process I: Generalized research publication Workflow ................................ 226
Figure 78: TCN with black nodes depicting researchers placed according to their Hirsch index and
blue nodes depicting individual research papers and citations............................................... 229
Figure 79: Relationship between (a) Fire Intensity and temperature, (b) Probability of fire ignition
and fire fuel moisture code ..................................................................................................... 233
Figure 80: Relation of FFMC with an increase in FWI ................................................................. 234
Figure 81: Tree VOMAS agents detecting fire vs. Fire area .......................................................... 235
Figure 82: Maximum FWI detected by VOMAS tree agents plotted vs. the area of fires ............. 236
Figure 83: Basic simulation model of multi-hop WSN modeled as a QUDG monitoring forest
conditions ............................................................................................................................... 238
Figure 84: WSN monitoring the local environment displaying measured parmeters .................... 239
Figure 85: Wireless Sensors recording forest temperature and calculating FWI ........................... 240
- xiv -
Figure 86: Configuration Options of simulation models of combining Forest Fire simulations with
monitoring simulation by a WSN........................................................................................... 241
Figure 87: Effects of monitoring large set of fires and FWI value ................................................ 241
Figure 88 :Earliest detection of FWI spike vs. average distance of sensors from fire incident ..... 242
Figure 89 :Simulation view of n = 60 random researchers with max-init-papers = 10 .................. 244
Figure 90: h-Index plot for twenty years for "Victor Lesser" obtained using Publish or Perish
program .................................................................................................................................. 245
Figure 91: a, b Validation exercise 1: Evolution of a researcher's Hirsch-index (Victor Lesser) .. 246
Figure 92: Validation Exercise 2: Comparison of number of nodes needed to display a author-paper
citation network vs. a TCN for Victor Lesser's papers. ......................................................... 248
Figure 93: Companion modeling cycle, figure credit: Bousquet and Trébuil [222] ...................... 250
- xv -
List of Tables
Table 1 Basic Network Analysis of Extracted Paper Citation Network ........................................... 69
Table 2 Top Journals based on Centrality ........................................................................................ 75
Table 3 Core Journals based on frequency ....................................................................................... 76
Table 4 Key categories based on Centrality ..................................................................................... 77
Table 5 Subject Categories according to frequency ......................................................................... 78
Table 6 Key bursts in subject categories .......................................................................................... 79
Table 7 Authors in terms of centrality .............................................................................................. 81
Table 8 Top authors based on frequency .......................................................................................... 82
Table 9 Core Institutes based on frequency ..................................................................................... 85
Table 10 Top Journals in Consumer Electronics .............................................................................. 89
Table 11 Top papers of IEEE Transactions on Consumer Electronics ............................................. 90
Table 12 Simulation parameters ..................................................................................................... 123
Table 13 SACS varying with constant TTL ................................................................................... 124
Table 14 SACS and TTL varying ................................................................................................... 124
Table 15 SACS vary with constant TTL (mobility) ....................................................................... 125
Table 16 Experiment with TTL and SACS varying (mobility) ...................................................... 125
Table 17 Categorical Sampling of Netlogo-users messages for a recent two weeks period (Aug
2011) ...................................................................................................................................... 139
Table 18 Table showing the format for creation of complex network using Cytoscape ................ 151
Table 19 Table of Eccentricity, Betweenness and Degree Centrality measures of the complex
network ................................................................................................................................... 175
Table 20 Descriptive Statistics of vary-boids experiment .............................................................. 183
Table 21 Descriptive statistics of experiment vary-sensors ........................................................... 183
Table 22 Table of Fire Danger Index values .................................................................................. 234
Table 23 Data for validation. Calc hI = Calculated h-Index and h-Index is the h-Index from Publish
or Perish software which obtains data via Google Scholar (Current as of July, 2011) .......... 247
- xvi -
List of Abbreviations
ABM
Agent-based Model
CNA
Complex Network Analysis
ABC
MAS
Agent-based Computing
Multiagent System
VOMAS
Virtual Overlay Multiagent System
SECAS
Sensing Emergence in Complex Adaptive Systems
CAS or cas Complex Adaptive System
SME
SS
UML
V&V
WSN
CSN
P2P
Subject Matter Expert
Simulation Specialist
Unified Modeling Language
Verification and Validation
Wireless Sensor Network
Cognitive Sensor Network
Peer to Peer
MANET
Mobile Ad-hoc Networks
MASON
Multi-agent Simulator of neighborhoods/ networks
JADE
Java Agent Development Framework
WOS
JCR
LISP
Web of Science
Journal Citation Reports
List Processing (Language)
- xvii -
1 Introduction
"Framework: A structure composed of parts framed together, esp. one designed for inclos-
ing or supporting anything; a frame or skeleton." - Oxford English Dictionary
Complex adaptive systems (cas) [10] is a set of special type of natural and artificial
complex systems. It is representative of the notion of a system where "The whole is more
than the parts". In other words, these are systems where numerous and perhaps quite sim-
ple components interact in a nonlinear fashion to give rise to global unprecedented and
often unpredictable behaviors visible and detectable at a higher level of abstraction. Phe-
nomena and mechanisms associated with cas include "emergence"1, self-organization[11]
and self-assembly[12]. These phenomena are typically observable and comprehensible by
intelligent observers at higher levels of abstraction[13] but not easily quantifiable at the
micro level. At times, while emergent patterns have been noted to be both interesting as
well as unprecedented, it has been observed by cas researchers that there is no known way
of predicting these patterns based solely on an understanding of the individual components.
Thus, an understanding of the individual components of a cas is not enough to develop an
understanding of the entire system [14] because it is the actual interactions of the compo-
nents which plays a key role in the observed global behaviors.
With a strong presence and association with the natural sciences from the physical
world, cas are well-known to transgress disciplinary boundaries. Therefore, it is quite
common to find an interest in different types of cas as demonstrated by the existence of a
large number and variety of models in literature. Multidisciplinary researchers from a wide
1 According to the Oxford Dictionary, the earliest use of the term emergence dates back to 1755 by Brooke
University "Beauty i. 10 From the deep thy [Venus'] bright emergence sprung." And formally defined as
"The process of coming forth, issuing from concealment, obscurity, or confinement. lit. and fig. (Cf. emerge
v. 3, 4.) Also said of the result of an evolutionary process"
- 1 -
and diverse set of backgrounds ranging from social sciences, ecological sciences, computa-
tional economics, system biology and others have been studying cas and related
phenomena. However, a study of this complex behavior can typically turn out to be a non-
trivial exercise; firstly because a research study can involve a set of extensive data collec-
tion exercises coupled with model development exercises for the cas components as well
as interactions. Subsequently these models need to be correlated with their observed data to
develop an understanding of the underlying dynamics of the cas system under study.
Technically, developing an understanding of a cas is associated with the formation of
"explicit" models of the system. While "implicit" models are of a mental cognitive
nature[15], "explicit" models are more effective for communication[16].
In terms of effort and commitment, these models can range from exploratory, requiring
a small time to develop, to inquisitory, requiring perhaps man-years of effort and resources.
These projects can also range from smaller to larger teams with different team structures of
personnel for gathering real-world data and for developing agent-based or complex net-
work-based models of various aspects of the system of interest. While cas researchers
conduct research in parallel disciplines and have an inherent interest in examining, com-
paring and contrasting models across disciplines, currently such comparisons are
performed informally. As such, to the best of our knowledge, there is no unified framework
for the multidisciplinary modeling and simulation of cas. As such, most researchers resort
to either performing informal comparisons[17] or else evolving domain-specific methods
of developing models[18-22].
1.1 Modeling cas
Epstein defines modeling as the development of a "simplified" representation of "some-
thing". Epstein also clarifies misconceptions about modeling and simulation giving a list of
- 2 -
different reasons for modeling [16]. He notes that modeling can be "implicit" (a mental
model), in which case, "assumptions are hidden, internal consistency is untested, logical
consequences and relation to data are unknown". Thus these types of models are not effec-
tive in the communication of ideas. The alternative is "explicit" models, which are more
effective in communicating ideas and concepts associated with the model. In "explicit"
models, the assumptions are carefully laid out and are coupled with details such as the out-
comes of the modeling exercise. Another important feature of developing explicit models
is that these models facilitate the replication of results obtained from the model by the sci-
entific community.
For any cas, the ability to come up with explicit models thus demonstrates the attain-
ment of a level of understanding. Although cas are extremely commonplace since all living
systems themselves tend to be cas at several levels of hierarchy, they do not surrender
themselves easily to modeling. As such, in the absence of a single comprehensive frame-
work governing various aspects of modeling and simulation of cas, these systems have
more or less been loosely modeled using a number of different paradigms with closely re-
lated roots. While most initial modeling of complex systems has traditionally been used
with the goal of simplification[23] such as using differential equations, system dynamics or
Monte Carlo simulations, more recently multidisciplinary literature has noted[24] a prefer-
ence of cas researchers to the use of one of the two modeling approaches as follows :
1. Agent-based (or Individual-based) modeling approaches[25].
2. Complex network-based approaches[26].
To specifically comprehend the cas modeling problem, a closer examination reveals that
the development of a unified framework for modeling cas would entail answering several
- 3 -
open research questions. These questions, ordered from the less abstract to the more ab-
stract, can be phrased as follows:
1. How to better develop models of interaction data?
2. How to describe cas to facilitate communication across scientific and disciplinary
boundaries?
3. In the case of a dearth of real-world data, how can cas simulation models be vali-
dated primarily using meta-data or concepts?
4. In general, how can multidisciplinary cas research projects be structured and exe-
cuted based on availability of resources and commitment?
Next, these four questions are examined in further detail.
1. How to better develop models of interaction data?
As part of a search for suitable data for cas modeling research, it is common to first
discover that while there are a considerable number of existing data sources, not many of
them might be useful for modeling. One key problem lies in the fact that most available
data typically consists of statistical summarized data, which is often not very helpful in de-
veloping models of the exact nature of the underlying dynamics in cas [27]. This approach
does not give sufficient details to develop more advanced data-driven models. As such, cas
researchers need to figure out exactly how to structure their data collection exercise during
their case study to be able to come up with models which actually give some useful infor-
mation.
A particularly useful paradigm for developing models of interaction data is the Com-
plex Networks Analysis (CNA) approach. Complex networks are an extensions of the
traditional graphs[28]. The idea is to use interactions and other relationships of various as-
- 4 -
pects and components in cas to develop network models. Once the network models have
been developed, cas researchers can use network manipulation to extract useful quantita-
tive as well as visual information about the topological structure of the system. Networks
can be trimmed and particular types of nodes or links can be manipulated to highlight use-
ful visual information about the topological aspects of the cas. In addition, various
quantitative measures such as network diameter, clustering coefficients, matching indices
and various centrality measures such as eccentricity, betweenness, degree etc. can be calcu-
lated for the network. These measurements allow making recommendations such as
classification of a particular type of network (Such as small-World, Scale-Free or Random
network models) or else for highlighting important components or interactions inside the
cas.
While there are numerous available software tools for developing complex network
models, selecting the right type and quantity of data can pose as a difficult research chal-
lenge. Currently this is not an automatic process and thus existing data mining techniques
cannot be directly applied to implicitly select important data from large number of data
types and columns. Specifically what cas researchers might be looking for, are specific as-
pects of cas, tied closely with emergence and other complex phenomena.
Taking a specific example of biological networks, currently a large amount of data
sources are available in online repositories such as nucleic acid sequences and bioinformat-
ics analysis clusters which can be used to execute complex algorithms. However, to
actually develop networks, cas researchers have to first understand and develop an implicit
mental model of exactly what should be a suitable node (e.g. a gene) and what would con-
stitute a suitable link (e.g. co-expression) to ensure that the resulting network is useful for
their domain. Some of these networks can themselves be quite complex and hierarchically
structured. An example of these is the gene regulatory networks, which are themselves
- 5 -
well-known as a system consisting of many sub-networks[29]. Other examples include
protein interaction networks which are developed based on protein molecules in undirected
graphs, signal transduction networks which are directed networks demonstrating various
biochemical reactions and social networks such as friendship networks where the weighted
links depend upon the perceived level of friendship of the subjects (persons) involved in
the study[30] etc.
2. How to describe cas to facilitate communication across scientific and disciplinary
boundaries?
Multidisciplinary researchers exploring cas can come from a variety of possible back-
grounds. While they are experts in their subject and can have significant domain
knowledge, it is possible that they might not be comfortable with the nomenclature fol-
lowed by other disciplines. As such, to describe cas, there needs to be a common easily
accessible description format which ties in closely with the cas model but is not specific to
a particular cas scientific discipline. While it should not be highly technical, it should still
allow the construction of descriptive models which should be comparable across case stud-
ies and disciplines. There are many inherent benefits of having such a description; firstly it
would allow for a comparative study of cas models across scientific disciplines and do-
mains. Secondly, it would allow high fidelity of simulation models with the specification
models. Thirdly, it can be used for learning cas concepts from models of other domains.
3. In the case of a dearth of real-world data, how can cas simulation models be validated
primarily using meta-data or concepts?
Being able to develop simulation models of cas is one thing and being able to validate
them is another. Unlike traditional simulation models, the concepts related to cas are typi-
cally quite abstract in nature and not easy to describe verbally. While researchers have
attempted to describe some aspects of a cas, it has traditionally been quite difficult to give
- 6 -
generalized definitions of cas concepts such as emergence or self-organization etc. using
terms globally acceptable by all disciplines. As such, the acceptability of the results of any
cas simulation model is tied closely with how valid these results appear to the researchers.
While for social scientists, models might be considered valid if the results of a simulation
appear similar to what they observe from population studies, for biologists, simulation
models might need a much higher level of fidelity with the actual components of the sys-
tem such as bio-chemical molecules etc. As such, what might seem valid to scientists from
one discipline might not be an acceptable validation for another set of scientists from an-
other discipline. As such, validation needs to be customizable. The problem with this
approach is that currently developing validation can be a fairly nontrivial exercise and
there is currently no way of structure the efforts of the Subject Matter Experts (SME) and
the Simulation Specialists (SS). As such, validation of cas simulation models varies from
case study to another without any basic common validation techniques unlike more tradi-
tional simulation models of complicated systems. These engineered systems are often
better describable in terms of mathematical equations and mostly lend themselves better to
generating suitable data for validation of models.
Another problem associated with validation of cas models is that at times, instead of be-
ing able to validate using data, the validation of complex concepts and emergent behavior
of the cas under study is required so as to ensure that the simulated system behaves close
enough to the cas being modeled. Such behavior is typically hard to put in an explicit mod-
el. Traditional validation techniques being typically data-driven either use various
attributes from the results of a simulation model and compare them with "actual" results
from the systems of interest or compare the results with another model. In the absence of
either of these, SS need to rely on "Turing or face validation"[31], where the observer
compares output of the model visually based on previous experience with the system. In a
- 7 -
cas modeling scenario, at times, it can be difficult to acquire the exact data required for
validation and thus, even though there might be some data, it might not always be a good
candidate for traditional validation schemes. In other words, some points regarding valida-
tion of simulation models of cas can be noted as follows:
a. Cas are extremely interactive in nature. A simple change in the composition can
massively change the global behavior. Having aggregate data in the form of
purely statistical tables might not be a real representation of the actual cas and
might only represent a certain propagation of states of the cas. If validation
were performed similar to traditional simulation models, it might not be valid
itself technically as it is similar to taking a zoomed picture of a bird's yellow
beak in a simulation and validating it with a yellow life jacket simply because
the colors match. Thus, even if the "graphs" or "plots" appear to coincide, the
actual models may be totally different and technically speaking, the validation
would not be valid by itself.
b. Secondly, validation of cas is tied closely with the interaction (run-time behav-
ior) of the various "agents" inside the model and not just the output data. For a
good cas validation scheme, this important fact needs to be taken into consider-
ation.
4. In general, how can multidisciplinary cas research projects be structured and executed
based on availability of resources and commitment?
Cas researchers develop models in a number of different ways, with numerous goals
and for various cas aspects [32]. One problem very often faced by researchers (as can be
observed on agent-based modeling mailing lists such as netlogo-users and repast) is when
researchers are planning research on cas but there is currently no clear way of identifying
- 8 -
which types of models to develop or how to move further, based on the expected outcomes
of the research. As such, researchers can find it difficult to structure their time and resource
commitments to better develop cas research projects. Occasionally, researchers might have
a few days to develop models to demonstrate future feasibility. Subsequently, if they suc-
ceed in this phase and are able to secure funding for further research, they might have
access to more resources to perform modeling and data collection exercises with a larger
team.
To the best of our knowledge, there is no existing multidisciplinary framework allow-
ing the structuring of cas case studies and model selection based on commitment levels or
goals in addition to allowing a combination of complex network and agent-based modeling
methods for multidisciplinary cas researchers. Having such a single unified framework for
multidisciplinary cas research would thus assist considerably in developing and communi-
cating cas models across scientific disciplines and structuring cas research projects.
1.2 Motivation
As noted in the previous section, there are several open questions associated with
developing various types of "explicit" cas models. The key problem here is the multidisci-
plinary nature of cas systems. Researchers from diverse backgrounds such as Life
Sciences, Social Sciences and Communication Engineers need to work with cas. As such,
there is neither a common methodology nor a set of concrete guidelines for developing cas
models spanning multiple scientific disciplines. Researchers can be unaware of exactly
how to proceed and what to expect when developing cas models. In addition, this problem
is compounded by the fact that most cas researchers are non-specialist in Computer Sci-
ences and therefore in spite of being experts in their particular domains, they can tend to be
neither interested in nor are often able to develop advanced highly technical models. How-
ever, it can be noted from an examination of multidisciplinary cas literature that they still
- 9 -
feel comfortable developing various types of explicit models with visual components (such
as agent-based and complex network-based models).
This thesis has been motivated by the lack of a single unified framework for the model-
ing of complex adaptive systems. Although many different models have previously been
developed for various types of cas, in general, cas modeling has evolved vertically with
little cross-flow of ideas between various application domains. As such, although there are
numerous examples of applied modeling and simulation in literature, to the best of our
knowledge, there does not exist a common guiding framework for interested multidiscipli-
nary cas researchers providing concrete guidance on how to approach cas problems, how to
develop different types of models and how to decide which type of data would be most
suitable for developing models, how to describe simulation models with a high fidelity to
the actual model, how to develop model descriptions allowing for visual and quantitative
comparisons of models and last, but not the least, how to structure validation studies of cas
simulations. Such a framework might also assist in the removal of ambiguities in the usage
of terms associated with cas nomenclature2 prevalent due to a parallel evolution of model-
ing practices in different scientific disciplines, enabling different cas models from different
case studies and scientific disciplines to be comparable with each other.
1.3 Aims and Objectives
The aim of this research is to work towards a set of unified framework guidelines allow-
ing researchers interested in multidisciplinary and inter-disciplinary cas studies to explore
the development of explicit models of cas by using a combination of complex network and
agent-based modeling approaches based on their research goals and level of commitment.
2 Examples include the terms individual-based modeling, agent-based modeling and multiagent systems.
More details are given in Chapters 2 and 3.
- 10 -
1.4 Original Contributions
Original research conducted and reported in this thesis is aimed at developing first steps
towards a unified framework in the form of concrete guidelines coupled with detailed case
study examples for use by multidisciplinary cas researchers. Some of the key original con-
tributions in the modeling and simulation of cas are summarized as follows:
A. The key contribution is a unified framework allowing multidisciplinary researchers
to plan their cas research case studies formulated based on levels of research goals
and commitment. The proposed framework uses a combination of agent-based and
complex network models to allow for interaction-based, exploratory, descriptive
and validated models of cas.
B. The proposed framework is structured in the form of different levels composed of a
set of methodological guidelines. Each of the framework levels allows the planning
and execution of a specific type of cas research study based on the availability and
type of data, the objectives of the case study and the expected levels of commit-
ment.
C. The first two levels of the proposed framework are structured specifically to en-
compass existing modeling and simulation research which mainly uses complex
network and agent-based simulation modeling techniques whereas the rest of the
two framework levels allow for more advanced model development using a combi-
nation of these approaches.
D. The complex network modeling level of the proposed framework is structured and
linked with the availability of suitable interaction data from cas components. Using
interaction data, complex network models can thus be developed for cas explora-
tion. These models can be manipulated and subsequently visualized using various
- 11 -
mathematical and software tools giving qualitative as well as quantitative inference
capability to the cas researchers.
E. Using the exploratory agent-based modeling level, researchers can use agent-based
modeling as an exploratory tool to develop proof-of-concept cas models to explore
feasibility of future research thus paving the way for more sophisticated techniques.
F. The descriptive agent-based modeling level of the proposed framework is useful for
researchers who are primarily interested in cross-disciplinary communication and
comparison of models. Descriptive modeling approach is being proposed which us-
es a combination of pseduocode-based specification and complex network
modeling as a means of modeling agent-based models. There are several benefits of
this approach; firstly it allows the description of cas models in a way such that there
is a high degree of fidelity of the model with the ABM. Secondly, it allows for
quantitative, visual and non-code based comparison of cas models developed in
multiple disciplines. Thirdly, it allows the exploitation of learning opportunities for
researchers by allowing the examination of models across scientific disciplines thus
facilitating the creation of heterogeneous multi-domain ABMs.
G. The validated agent-based modeling level of the proposed framework is based on a
step-by-step methodology for the development of in-simulation validation for
agent-based models by means of an interactive collaborative effort involving both
Subject Matter Experts (SME) as well as the Simulation Specialists (SS). This ap-
proach is based on concepts from multiagent systems, software engineering and
social sciences. Using a systematic approach, the outcome of the methodology is an
agent-based model of the cas validated by means of design-by-contract invariants
in the simulation model where the contract is enforced by means of in-simulation
- 12 -
cooperative agents termed as the Virtual Overlay Multiagent System (VOMAS).
Building a VOMAS allows SME and SS to collaboratively develop custom in-
simulation verification and validation schemes for the cas application case study.
H. The viability of the proposed framework is demonstrated with the help of various
case studies spanning different individual scientific disciplines and some case stud-
ies spanning multiple scientific disciplines.
1.5 Proposed framework for the Modeling and Simulation of cas
In this section, firstly an overview of the proposed framework is presented. Next the
framework is described from two different perspectives firstly in terms of study objectives
of conducting a cas research case study and the expected level of commitment. Secondly,
the framework is described in correlation with the availability and access to specific data
types.
1.5.1 Overview of the proposed framework
For the sake of practicality, the framework guidelines have been developed in the form
of levels of abstraction. Thus, cas researchers can opt for modeling at a particular level de-
pending on factors such as availability of data, meta-data as well as the level of interest and
how much time cas researchers can invest in pursuing a research project. In addition, ex-
ample case studies are presented to demonstrate the usage of various framework levels.
Based on the proposed framework, research in cas can be conducted by choosing and sub-
sequently following one of the following four proposed levels for developing cas models:
• The complex network modeling level of the framework is useful if interaction data is
readily available. In this level, complex network models of cas can be developed using
this interaction data and subsequently Complex Network Analysis (CNA) can be per-
- 13 -
formed for network classification as well as determination of various global and local
quantitative measures from the network for the extraction of useful information. Such
information can give details of emergent behavior and patterns which would otherwise
have not been evident using statistical or other more traditional mathematical methods.
• The exploratory agent-based modeling level of the framework extends existing ideas of
agent-based modeling prevalent in multidisciplinary literature which focus on the de-
velopment of exploratory agent-based models of cas to examine and extricate possible
emergent trends in the cas. Building exploratory models allows cas researchers to de-
velop experimental models which help lay foundation for further research. These
proof-of-concept models also assist researchers in determining the feasibility of future
research in the domain using the selected model design.
• Developing models at the descriptive agent-based modeling level of the framework en-
tails developing concrete DescRiptivE Agent-based Models (DREAM) by using a
combination of pseudocode-based specification, a complex network model and a quan-
titative model "fingerprint" based on centrality measures of the agent-based model
which are all associated closely with the ABM. The pseudocode-based specification is
developed in the form of non-code template schemas and has several benefits; firstly it
is close to an executable specification but is not tied with any single programming lan-
guage thus allowing use by cas researchers for developing agent-based models based
on the specification using tools of their own liking. Secondly, this specific type of spec-
ification allows a one-to-one correspondence of ABM concepts with the descriptive
model. Thirdly this specification allows communication and comparison of models in
multidisciplinary studies by using visual as well as quantitative methods.
• The validated agent-based modeling level of the proposed framework is concerned
with developing verified and validated agent-based models. This level allows perform-
- 14 -
ing in-simulation verification and validation of the agent-based models using a Virtual
Overlay Multiagent System (VOMAS) based on a cooperative set of agents inside the
simulation allowing the verification and validation of the cas model by means of de-
sign-by-contract invariants. These invariants are developed as a result of collaboration
of the Subject Matter Expert (SME) and the Simulation Specialist (SS). In this level,
concepts originating from software engineering, multiagent Systems and social scienc-
es are all used in tandem to propose a systematic methodology for ABM verification
and validation.
1.5.2 Proposed framework levels formulated in terms of cas study objectives
In Figure 1, it can be noted how different framework levels can be used by multidisci-
plinary cas researchers to develop models based on their particular study objectives and
expected outcomes.
If there is sufficient interaction data available then the cas research study can proceed
by using the complex network modeling level of the proposed framework. In this level,
researchers first analyze the data columns, extract suitable data, develop complex network
models and subsequently perform network manipulation and complex network analysis for
the discovery of emergent patterns.
However, as is often the case, if such data is unavailable and the goal of the research
study is to determine the feasibility of future research, then it might be possible to feasible
to proceed in their research study by using the exploratory ABM level of the proposed
framework.
These two framework levels essentially can also be used to encompass existing cas re-
search studies which have primarily used either complex network modeling and analysis or
else agent-based modeling or both in their analyses.
- 15 -
If however, the goal of the research is to perform an inter-disciplinary comparative case
study then the descriptive agent-based modeling level of the proposed framework allowing
for developing a DREAM model can be chosen. This particular framework level has the
benefit of allowing for inter-disciplinary model comparison, knowledge transfer and learn-
ing.
Finally, if the goal of the study is develop simulations with a high degree of correlation
with real-world systems such as in the development of decision support systems, then the
appropriate framework level for usage would be the validated agent-based modeling level
based on the development of an in-simulation validation scheme using the VOMAS ap-
proach. This framework level is also more suitable for large-scale team oriented projects
and requires adherence to team-oriented protocols with the goal of a verified and validated
agent-based model of the cas system under study.
- 16 -
Figure 1: An overview of the decision-making process for choosing framework levels in relation to cas
research study objectives
- 17 -
1.5.3 Proposed framework levels formulated in relation to available data types
In the previous section, an overview of the different framework levels in relation to the
cas research study objectives. Here, in this section, we shall describe the framework in
terms of available data or knowledge about the cas. As we can note from Figure 2, a de-
scriptive specification of the cas model can be developed based on the metadata or
knowledge about the cas. An ABM can be developed either from this specification or di-
rectly from the knowledge of cas. The ABM can be verified and validated using in-
simulation validation (which has been developed as a result of extensive meetings between
SMEs and SSs) that is performed by building a VOMAS model. By the help of invariant
constraints enforced by the cooperative agents forming the VOMAS, the simulation can be
verified and validated using in-simulation methods. In addition, if interaction data is avail-
able for the development of network models, complex network models can be developed
and CNA can be used to manipulate and analyze various structural topological features of
the interactions using various information visualization-based and mathematical tools.
cas knowledge
Complex Network
Modeling level
Descriptive
ABM
development
ABM
development
Interaction Data
Validated ABM
development
Figure 2: Data Driven View of Proposed Framework
- 18 -
1.6 Publications
Some parts of the chapters in this thesis have been published in various peer-reviewed
venues. Because of the peculiar inter-disciplinary nature of cas research, some of the work
performed was in collaboration with different researchers and the specific contributions of
the co-authors have been noted earlier in the declaration at the start of the thesis.
1.6.1 Refereed Conferences and Workshops
1.
Niazi MA, Hussain A. Social Network Analysis of trends in the consumer electron-
ics domain. In: Proc Consumer Electronics (ICCE), 2011 IEEE International
Conference on, Las Vegas, NV, 9-12 Jan. 2011, 2011. pp 219-220. (Chapter 3)
2.
Niazi MA, Hussain A, Baig AR, Bhatti S. Simulation of the research process. 40th
Conference on Winter Simulation. Miami, FL: Winter Simulation Conference;
2008. pp 1326-1334. (Chapter 6)
3.
Niazi MA, Hussain A, Kolberg M. Verification &Validation of Agent Based Simu-
lations using the VOMAS (Virtual Overlay Multi-agent System) approach.
MAS&S 09 at Multi-Agent Logics, Languages, and Organisations Federated
Workshops. Torino, Italy; 2009. pp 1-7. (Chapter 6)
4.
Niazi MA, Siddique Q, Hussain A, Kolberg M. Verification and Validation of an
Agent-Based Forest Fire Simulation Model. SCS Spring Simulation Conference.
Orlando, FL, USA: ACM; 2010. pp 142-149. (Chapter 6)
1.6.2 Published Journal papers
1.
Niazi MA, Hussain A. Agent-based Computing from Multi-agent Systems to
Agent-Based Models: A Visual Survey. Springer Scientometrics 2011;In-
press.(Chapter 3)
- 19 -
2.
Niazi MA, Hussain A. Agent based Tools for Modeling and Simulation of Self-
Organization in Peer-to-Peer, Ad-Hoc and other Complex Networks. IEEE Com-
munications Magazine 2009; 47(3):163 - 173. (Chapter 4)
3.
Niazi MA, Hussain A. A Novel Agent-Based Simulation Framework for Sensing in
Complex Adaptive Environments. IEEE Sensors Journal 2011; 11(2):404-412.
(Chapter 5)
4.
Niazi MA, Hussain A. Sensing Emergence in Complex Systems. IEEE Sensors
Journal 2011; 11(10): 2479-2480. (Chapter 5)
1.7 Overview of the thesis
Here, first an overview of the different explored case studies is provided. This is fol-
lowed by an overview of the chapters.
1.7.1 Overview of case studies
To ensure that the research was in line with the norms of various cas, we have
worked in tandem with teams of cas researchers and domain experts from life sciences,
social sciences and telecommunications. The following list gives details of some of the
case studies discussed in the thesis as a means of examples of the application of the pro-
posed methods associated with various levels of framework along in correlation with the
thesis chapters:
-
-
In chapter 3, the proposed framework level is applied on two different case studies in
the domain of scientometric data of agent-based computing and consumer electronics
domains.
In chapter 4, a comprehensive case study on the use of unstructured search algorithms
from the domain of P2P networks in the domain of "Cyber-physical systems"[33] by
the development of an "Internet of things"[34] has been presented.
- 20 -
-
-
In chapter 5 a case study on the development of a heterogeneous ABM of sensing sin-
gle-hop Wireless Sensor Network for sensing complex behaviors of flocking "boids" is
presented.
In chapter 6, three different case studies are presented. The first case study is in the
domain of ecological modeling and models forest fire simulations. The second is in the
domain of multi-hop Wireless Sensor Networks modeled in the form of a Quasi Unit
Disk Graph (QUDG). The third case study is in the domain of simulation of the evolu-
tion of researchers on the basis of their Hirsch index.
1.7.2 Outline of the Thesis
The rest of the thesis is organized as follows:
Chapter 2: This chapter presents background and related work required to com-
prehend the rest of the thesis.
Chapter 3: This chapter presents complex network modeling level of the proposed
framework. These methods are further applied to two different domains; Agent-based
Computing and Consumer Electronics. The first case study is given in considerable detail
and various Scientometric indicators such as key papers, important authors, highly cited
Institutions, key journals and a number of other indicators are identified using complex
network modeling. The second case study is given as a means of validation of the proposed
methods in another separate domain and identifies key Journals, authors and research pa-
pers from the consumer electronics domain.
Chapter 4: This chapter presents exploratory Agent-based modeling level of the
proposed framework. As a demonstration of the proposed methods, a comprehensive ex-
ploratory agent-based model in the domain of Cyber-Physical Systems is developed
- 21 -
demonstrating a combination of unstructured P2P search methods to locate content in static
and mobile physical computing devices.
Chapter 5: In this chapter, descriptive agent-based modeling level of the proposed
framework is presented. Descriptive modeling entails the development of a DescRiptivE
Agent-Based Modeling (DREAM) model by using a combination of a complex network
model, a quantitative centrality-based fingerprint and a pseudocode-based specification
model with a high degree of fidelity with the actual agent-based model. As a means of
demonstration of the proposed framework level, the DREAM approach is applied in a
comprehensive case study of a heterogeneous cas ABM of a WSN observing a set of flock-
ing "boids".
Chapter 6: In this chapter, the validated agent-based modeling level of the pro-
posed framework is proposed. The proposed methodology based a team-oriented approach
of in-simulation validation is demonstrated using three different application case studies
from three different scientific disciplines of ecology, telecommunications and social simu-
lation allowing for a proof of concept of the generalized and broad applicability of the
proposed methods.
Chapter7: In this chapter, the thesis is concluded. First the key research contribu-
tions are presented and are followed by a critical review of the framework levels in relation
with existing approaches in literature. This is followed by limitations and proposed future
enhancements proposed in the framework levels. Finally a set of possible future applica-
tion case studies ideas which can effectively employ the different framework levels is
presented.
Appendix 1: The appendix contains a description of the exact keywords associated
with the Scientometric study of agent-based computing.
- 22 -
2 Complex Adaptive Systems: Background and
Related Work
In the previous chapter, an overview of the entire thesis and the key contributions were
presented. In this chapter, basic background and a literature review of various concepts
needed for the understanding of the thesis are presented.
2.1 Overview
We start by first giving an overview of cas and their key characteristics. Next we give
specific examples of cas from natural and artificial systems. Subsequently we give an
overview of modeling of cas. Next, we give a review of agent-based modeling tools. This
is followed by a review of verification and validation of simulation models. Finally an
overview of different communication network simulators is presented.
2.2 Complex Adaptive Systems (cas)
Cas are a special type of complex systems which arise due to nonlinear interactions of
smaller components or agents[17]. While it is difficult to exactly define cas, Holland [35]
notes that:
"Even though these complex systems differ in detail, the question of coherence under change is
the central enigma for each. This common factor is so important that at the Santa Fe Institute we
collect these systems under a common heading, referring to them as complex adaptive systems
(cas). This is more than terminology. It signals our intuition that general principles rule cas behav-
ior, principles that point to ways of solving the attendant problems."
The modern approach of cas has evolved from a series of attempts by researchers to de-
velop a comprehensive and general understanding of the world around us. Our world is
burgeoning with interactive entities. Most times, such interactions manifest themselves as
- 23 -
change of some type either internal to the entities (change of internal state) or else external
(change of external state or behavior). In other words, these entities adapt in networks of
interactions spread nonlinearly across the entire system spatially as well as temporally[36].
In the science of complexity, it is the "small" and "numerous" that govern changes in
the "large" and "few". The interaction of small (micro) and perhaps simple components
gives rise to structures and behaviors, which are amazingly complex when observed from
the macro perspective[37]. Comprehension of the intricacies of these systems is so hard
that it has lead to not just one but a set of theories based on different observations and dif-
ferent types of modeling methodologies targeting different aspects of the system.
A simple example of such systems is that of life. Although not easily quantifiable, living
systems exhibit elegance absent in the monotonicity of complex but relatively inadaptable
inorganic and chaotic systems. Life surrounds us and embosoms us. Every living system
and life form known to us asserts itself with a display of a dualistic nature which on one
hand is decrepit and frail and on the other it has its own tinge of sturdiness and resilience,
hard to imagine in any engineered systems[38]. The frailty of complex living systems is
apparent because all complex life forms seem stamped with a "certain" programmed expi-
ration date because the growth and aging of most multi-cellular living organisms is
governed by the complex behavior encoded in the genes. On the other hand, a closer exam-
ination reveals the Complexity and resilience working inside the systems. From apparent
humble beginnings and small components (DNA, RNA, genomes, amino acids, proteins
and cells etc.), the complexity of each life form emerges to give complex features such as
muscles, tissues, systems, of which the most important is cognition and "self" for higher
life forms.
A human embryo which starts as just a single cell replicates and forms a complete hu-
man being which is not just a blob of identical cells but has an extremely intricate and
- 24 -
complex structure, with organs, systems and dynamics. Looking at the single cell embryo,
it is hard to imagine how it could end up in forming this amazingly complex life form[39].
If we examine the cell closely, it is hard not to think as to how exactly does this one cell,
which divides into two and then four and so on, apparently suddenly bloom into complex
structures without any apparent set of "physical" guiding forces or advanced sensory abili-
ties such as one which coordinates all cells to form global structures. And when the
structures do mature into systems, it makes one wonder how really do the multiple cells
synchronize to gradually start the dynamic phenomenon associated with complex life
forms, such as the blood flow, breathing, digestion, clotting, immune systems and emo-
tions, thoughts, family, social systems and so on? The interesting part of all this is that the
guiding principles of the entire life of the organism or life form are coded inside the genes,
part of the cell's nucleus.
Another example of similar complex systems is human and animal social systems,
where interactions are once again the key to understanding the larger perspective of things.
The concept of "country", "social or ethnic groups", "families", "clans", "universities",
"civilizations", "religion" and so on are extremely powerful and govern the life of every
human being. However, these phenomena are in one way or the other, rooted in the con-
cepts of interactions. All this makes one thing very obvious. Our current science is known
to miss the big picture of complexity. Developing this understanding further would assist
us in understanding life, the universe and everything that exists[37].
This absence of the "complete picture" is at the heart of research in cas and the closely
related movements such as the "Complexity theory"[17] and the "Systems Biology theo-
ry"[14]. In contrast to reductionism which is rooted in simplifications and thus gives a
misleading confidence that understanding the parts will somehow "ensure" that we will be
able to understand the whole, the complex systems approach focuses instead on a specific
- 25 -
meaning of the phrase "The whole is more than the parts". Here, literature notes that the
key factor in understanding cas is to understand the "interactions" of the parts which can-
not be quantified easily.
2.2.1 The Seven Basics of cas
Holland maintains that there are seven basics of cas. Four of these are properties and
three are mechanisms. These have been termed basics because he notes that other cas-
related ideas can be considered as derivates of these. The following discussion is based on
Holland's description of properties and mechanisms of cas [35]:
1. Aggregation (property)
The word "Aggregation" follows from "aggregate", which, according to the Oxford
English Dictionary, means "Collected into one body". Aggregation is useful in two ways,
firstly as a generalization where different items can be categorized in a single big and oft-
reused umbrella, e.g. animals, plants, bags, baskets etc. In other words, this is a way of ab-
straction as well since it allows focus on the important details to be modeled and ignores
those which can be ignored. This type of aggregation is perhaps classification and is more
related to cognition rather than actual physical containment, to which the second type of
aggregation may be attributed. Holland points out the formation of complex agents called
meta-agents in living and social systems (whether natural or artificial) based on complex
behavior of smaller, simpler agents.
An example can be seen in Figure 3, where we notice that seemingly without any global
intelligence, the shape of the melon is an example of emergence based on the complex in-
teractions of its constituent components, such as skin, seeds and other parts, which are
made up of other structures, eventually coming down to cells and sub-cellular structures.
- 26 -
Figure 3: Peculiar shape and striations in a melon
2. Tagging (Mechanism)
Tagging is a mechanism which is frequently observed in cas. Tagging allows for the
formation of aggregates. Tagging is exhibited in cas in the same manner as flags are used
to identify troops or messages are given IP addresses to reach the correct destination in a
network etc.
3. Nonlinearity (Property)
Nonlinear interactions are the norm in cas and are one of the reasons in the emergent
global behaviors which indicate that the system is a cas.
4. Flows (Property)
Another property of cas is the formation of dynamic networks and flows. As such there
are numerous attributes such as seen in Economics e.g. The Multiplier Effect when an ex-
tra amount of a resource is added to a flow causing it to transform as well as transmit
between different nodes. Another important behavioral property that can be observed in
flows is the recycling effect where resources are recycled over the flows via the network
nodes thus enriching the emergent behavior. An example of the effects of recycling is the
- 27 -
emergence of large number of insect and animal species in undisturbed forests as discussed
by Nair [40]. Similar example is in the case of plants; Lowman notes the patterns of emer-
gence, growth, mortality and herbivory of five rain-forest tree species[41].
5. Diversity (Property)
The species of a rainforest, people living in large towns, vendors in malls, structure of
the mammalian brain all exhibit extreme diversity. So, it is rare to see the same type of
components if any of these cas are explored in depth.
Being a dynamic property, diversity also has self-repairing and self-healing charac-
teristics. Disturbing the population of a particular predator in a forest can result in an
increase of numbers of several prey species which can then result in an increase of num-
bers of the same or other predators. Lima[42] notes that continuous predation can lead to
major impact on entire ecosystems. Hein and Gillooly [43] have recently demonstrated that
dispersal and resource limitation can jointly regulate assembly dynamics of predators and
preys.
6. Internal Models (Mechanism)
Internal models are a mechanism by which agents inside a cas keep models of other in-
dividuals. On first looks, such a model would seem to be the property of only intelligent
mammals. However, a closer look reveals that even the seemingly extremely simple life
forms such as bacteria and viruses need to keep models of a certain kind of their environ-
ment and their hosts. These models are important for the survival of the species in general
and thus the fact that a particular species still exists is a testimony that the species can de-
fend its own in the particular environment (as has been happening for perhaps millions of
years). Berg[44] notes that E. coli exhibit model behavior because they exhibit motility
towards regions in their environment that are deemed favorable.
7. Building Blocks (Mechanism)
- 28 -
Internal models as described in the previous section need to be enhanced by realistic
mechanisms. These realistic mechanisms are termed as the "Building blocks". As an ex-
ample, a diverse numbers of human beings form the 6+ Billion human population and they
can each be basically differentiated from each other based on only a few set of building
blocks such as eyes, nose, lips, hair, face etc. Variations in these blocks include changes in
constituent attributes such as their color, shape etc. and eventually form the basis of the
internal models.
In this section, we have noted how complex adaptive systems are structured and how
they have a set of complex features which need to be examined closely to see why the sys-
tems behave in the peculiar manner at the global scale. We next talk a bit about
"emergence".
2.2.2 Emergence
Emergent behavior in cas has traditionally been considered difficult to define and hard
to detect. Yaneer Bar Yam[45] defines emergence as follows:
"Emergence refers to the existence or formation of collective behaviors - what parts
of a system do together that they would not do alone."
Yaneer also notes that emergence can be considered as a means of explaining collective
behavior or properties of a group of components or agents or else it can also be used to
describe a system in relation to the environment. Boschetti and Gray [13] describe three
levels of emergence:
1.
2.
Pattern Formation and detection such as oscillating reactions etc.
Intrinsic Emergence such as flocking behavior
3. Causal Emergence such as human behavior using messaging.
- 29 -
2.3 Examples of cas
Cas can be both natural as well as artificial. Here we give some examples of different
types of cas. Next, we give specific examples of natural and artificial cas.
Figure 4: Tree stem adapting structure temporally
2.3.1 Natural cas example 1: Cas in plants
An example of adaptation in a natural system can be seen in Figure 4 where we see a
tree stem adapting according to the peculiar shape of an artificial metallic structure. One
cas which can be noted here is the complete plant. The plant itself is made up of numerous
cells interacting and performing tasks based on programs from their genomes. However
each cell is unaware of the large scale features of the plant or the way it interacts with the
environment such as the emergent behavior of adaptation that can be observed in the figure
in response to the position of the metallic structure as well as the path followed by sun for
- 30 -
a better chance at getting sunlight. This emergent behavior can be seen as essentially a trait
which has helped plants and animal species to survive over millions of years.
2.3.2 Natural cas example 2: Cas in social systems
While humans themselves are made up of several cas originating from living cells and a
variety of bacteria, viruses and other biochemical organic molecules. All these play an im-
portant role in the individual as well as social lives of living beings. An example is given
here in the domain of the research process, which can, by far, be considered as one of the
most complex social interaction, based on intelligent behavior.
A research paper published in a peer-reviewed venue represents the culmination of
efforts of a large number of interactions. Some of the entities involved in a research paper
are as follows:
Authors and their Papers which were read during the study by the researcher
Discussions and meetings with colleagues
Advice by advisors
Advice by conference or journal referees during the peer-review process
And so on . . .
One way of understanding this all is to notice that this entire process reflects a highly
dynamic cas with multiple intelligent and inorganic (or even cyber-) entities such as pa-
pers, reviews and Journals. Several emergent behaviors can be observed here such as the
propagation of research, formation of research groups, rise and fall of academic journals
and their impact factors, emerging trends and these are all based on an enormous commu-
nity of people spread globally, which do not even understand or even, at times, do not need
to understand the exact structure and dynamics of how research leads to emergent behav-
ior.
- 31 -
A quantitative study of scientific communication is technically termed as scientomet-
rics [46]. Scientometrics requires the use of a multitude of sophisticated techniques
including Citation Analysis, Social Network Analysis and other quantitative techniques for
mapping and measurement of relationships and flows between people, groups, organiza-
tions, computers, research papers or any other knowledge-based entities.
Scientometrics sometimes involves the use of domain visualization, a relatively newer
research front. The idea is to use information visualization to represent large amounts of
data in research fronts[47]. This allows the viewer to look at a large corpus and develop
deeper insights based on a high level view of the map[48]. Visualization using various
network modeling tools has been performed considerably for social network analysis of
citation and other complex networks[49]. Various types of Scientometric analyses have
previously been performed for domains such as HIV/AIDS [50], Scientometrics [51], Mex-
ican Astronomers[52], scientific collaborations [53] and engineers in South Africa [54].
Extensive work on research policy has been performed by Leydesdroff [46]. Some of the
recent studies in this direction include visualization of the HCI domain[55], identification
of the proximity clusters in dentistry research[56], visualization of the pervasive computing
domain[57], visualization of international innovation Journals[58] as well as identification
of trends in the Consumer Electronics Domain[2].
Scientometric studies which combine co-citation analysis with visualizations greatly
enhance the utility of the study. They allow the readers to quickly delve into the deeper as-
pects of the analysis. Co-citation analysis measures the proximity of documents in various
scientific categories and domain. These analyses include primarily the author and the jour-
nal co-citation analyses. Journal co-citation analysis identifies the role of the key journals
in a scientific domain. In contrast, the author co-citation analysis[59] especially by using
visualization[60] offers a view of the structures inside a domain. The idea is based on the
- 32 -
hypothesis that the higher the frequency of co-citation of two authors, the stronger the sci-
entific relation between them. Whereas document co-citation maps co-citations of
individual documents [61-63], author co-citation focuses on body of literature of individual
authors. In addition, co-citation cluster analysis of documents is useful for the display of
the examination of scientific knowledge evolution structure at the macro level. In initial
cluster analysis, techniques involved clustering highly cited documents using clustering
based on single-links. Subsequently clustering is also performed on the resultant clusters
numerous times[64]. Recent techniques involve the use of computational algorithms to per-
form this clustering. These clusters are then color coded to reflect that.
If we were to take some of these entities, and analyze the data, we can come up with
quite interesting networks e.g. as shown in Figure 5. This network represents top papers,
in terms of citations, with topics related to "agent-based modeling" from the years 1990 –
2010. As can be seen here, the network visualization is clearly allowing the separation of
the top journals in this field in terms of "centrality" measures such as degree, betweenness,
eccentricity or indices or clustering coefficients as described in previous section. Com-
plex Network are not only formed from social interactions. Rather they can be developed
from any cas system interaction such as systems inside living beings[29].
Journal Citation Reports (JCR) has been considered in literature as the key indicator of
a Journal's scientific repute[65]. It is pertinent to note here that previously researchers such
as Amin and Mabe [66] have critically reviewed the way Journal impact factor might be
used by authors and journals. While alternatively other researchers have proposed new al-
ternative impact factors such as by Braun [67] and Fersht [68]. However, the fact remains
that, in general, the scientific contributions listed in Thomson Reuters Web of Science are
considered highly authentic by the overall scientific community[69], [70].
- 33 -
Figure 5: Complex Network model showing different Journals publishing articles related to
agent-based modeling. Bigger caption represents a higher centrality measure explained later in the
chapter.
2.3.3 Artificial cas example 1: Complex adaptive communication networks
In this section, we introduce complex adaptive communication networks. Complex
adaptive communication networks are a recent advancement in artificial cas. Primarily
these are communication networks which have a large number of components resulting in
behavior associated with cas. These networks arise due to a recent rapid advancements in
communication technology because today's communication networks such as those formed
by wireless sensor, ad-hoc, Peer-Peer (P2P), multiagent, nano-Communication and mobile
robot communication networks, are all expected to grow larger and more complex than
ever previously anticipated. Thus, these networks, at times, can possibly give rise to com-
plex global behaviors similar to natural cas. As a result, network designers can expect to
observe unprecedented emergent patterns. Such patterns can be important to understand
since, at times, they can have considerable effect on various aspects in a communication
network such as unanticipated traffic congestion, unprecedented increase in communica-
- 34 -
tion cost or perhaps a complete network/grid shutdown as a result of emergent behavior.
Some well-known examples include the emergence of cascading faults in Message Queue-
based financial transactions after New Year holidays[71], recent cascading failures reported
in the Amazon.com cloud[72], effects of viral and worm infections in large networks, ef-
fects of torrent and other complex traffic in Internet Service Providers and corporate
networks, multi-player gaming and other similar P2P traffic in company intranets, self-
organization and self-assembly related effects in sensor and robotic networks[73].
The torrent protocol is an example of an artificial cas since it exhibits a number of inter-
esting properties associated with cas. It relies on software called the "torrent clients"
which, as can be seen in Figure 6 allow autonomous and almost anonymous interaction
with other clients around the globe. The different peers autonomously self-organize and
adapt to both download as well as upload files. Here, the emergent phenomenon is the
downloading of the file. The interactions, allocations of bandwidth, uploading and down-
loading of chunks are part of the nonlinear interactions.
The way this entire process the process of uploading and downloading (sharing) files
works is outlined by Cohen in [74] as follows:
1. The peer computer which starts to share a file breaks the file into a number of iden-
tically sized pieces with byte sizes of a power of 2 anywhere from 32 KB to 16 MB
each. Each piece is hashed using SHA-1 hash function and this information is rec-
orded in the .torrent file.
2. Peers with a complete file copy are called seeders and those providing the initial
copy are called the initial seeders.
3. The .torrent is a collective set of information about the file coupled with the infor-
mation about tracker servers allowing peer discovery of downloaders.
4. The trackers give random lists of peers to other peers.
- 35 -
5. The different peers automatically self-adapt to download the complete file from dis-
tributed chunks of file from all over the globe.
6. Peers who only download and do not upload are punished by lesser bandwidth while
peers who upload as well are given higher bandwidths.
7. The emergent behaviors here include the file sharing (upload and download) and re-
duction of peers who only download.
Figure 6: Artificial Global cas formed by P2P clients observed using utorrent software
2.3.4 Artificial cas example 2: Simulation of flocking boids
In 1986, a computational model of flocking animal motion was presented by Craig
Reynolds [75] called the "boids". The model is based on three separate but simple rules.
The first rule is "separation" which involves steering to avoid crowding. The second rule is
"alignment", which steers the boid towards the average heading of local flockmates. And
the third rule is "cohesion", which steers towards the average position of local flockmates.
The "boids" model has been considered a model of realistic simulations ranging from
SIGGRAPH presentations to Hollywood. This model offers an example of complex adap-
tive behavior based on apparently local rules of the mobile agents or boids.
- 36 -
2.4 Modeling cas
The word "Model" is defined by Merriam-Webster in 13 different ways:
1 obsolete : a set of plans for a building
2 dialect British : copy, image
3 : structural design *a home on the model of an old farmhouse*
4 : a usually miniature representation of something; also : a pattern of something to be made
5 : an example for imitation or emulation
6 : a person or thing that serves as a pattern for an artist; especially : one who poses for an
artist
7 : Archetype
8 : an organism whose appearance a mimic imitates
9 : one who is employed to display clothes or other merchandise : Mannequin
10 a : a type or design of clothing b : a type or design of product (as a car)
11 : a description or analogy used to help visualize something (as an atom) that cannot be di-
rectly observed
12 : a system of postulates, data, and inferences presented as a mathematical description of an
entity or state of affairs
13 : Version
Critically speaking, modeling a system is important for the survival of the human race.
The entire concept of cognition appears to be based on modeling. We start our learning by
experiencing this world, which leads us to the development of cognitive "mental" models.
Then we use these models to associate with other experiences or knowledge to keep learn-
ing. As an example a baby develops a first implicit "model" of heat if s/he touches a hot
bottle of milk. Later on, in her life, she keeps learning more and more about temperature
and heat while whether she consciously realizes it or not, all new models are rooted essen-
tially with the basic cognitive model. This kind of models have been termed as "implicit"
models[76].
Implicit models, however, are hard to express and are not very communicable. What
one person perceives of a certain concept, might mean a totally different thing to another.
- 37 -
This is the reason for the development of what Epstein, an authority in modeling, mentions
as "explicit models" [16].
Modeling implies a certain level of understanding and prowess over the system and re-
quires a deep study of certain aspects of the system. When it is not just one system, rather a
group of systems, modeling implies developing abstractions, which can cover different
domains.
Previous techniques have included focus on physics as a means of modeling. Such for-
mal models typically are simplified models representing the system as e.g. Differential
equations. Obviously by limiting an entire system to only a few equations implies that
most aspects of the system including the components have been reduced to simple num-
bers, e.g. quantities such as change of numbers etc. However, more recently modeling cas
often entails using one of two basic methods of developing computational models i.e. Ei-
ther use of Agent based modeling to develop simulation models or else development of
complex network models for analysis of interactions using real world or simulation data as
shall be discussed in detail in the next section. Although these models prevail in literature,
to the best of our knowledge, there is no unified framework which outlines this set of ideas
and couples them along with agent-based models.
2.4.1 Complex Network Modeling
In this section, an overview of complex network methods is presented. The basic idea of
Complex Network Analysis originates from graph theory. A complex network is essentially
an advanced graph, where unlike theoretical graphs, each node and edge is loaded with
more information. As an example, if we were to develop a social network of friends, the
network edges could represent the level of emotions which the friends have for each other.
So, if a person Alice considers three people (say Bob, Cassandra and David) as friends,
- 38 -
then this could be represented as a network. Suppose Alice considers Cassandra as her best
friend so this could be modeled in the network by giving a higher attribute value to the
link/edge between Alice and Cassandra. To make things more interesting, the network
could be directed so it is possible that Bob might not even consider Alice as a friend. So,
each friendship relation should be represented by a directed edge.
Using various Mathematical tools and models, complex networks can be subsequently
analyzed and compared with other models from the same or different domains. As an ex-
ample Barabasi [77] compares Biological networks with models of the world wide web.
Besides classification, networks can also be used to discover important features of the nu-
merous nodes (or components) of the cas. While Complex Networks are a general set of
methods, the growth of literature of complex network usage can be identified across vari-
ous cas literature ranging from Social Network Analysis[30] to Biological Networks [29]
and in Citation Networks[78].
2.4.1.1 Complex Network Methods
In a cas, interactions can be modeled as networks. The key difficulty in network design
is to understand which data needs to be captured to develop the network model(s). Once
the data has been converted to a network, there exist a large number of complex network
analysis Mathematical and software tools for performing the analysis. However, the hard
part here is the selection of data, the design and extraction of the complex network models.
While application of the actual centrality measures appears to be deceptively easy due to
the existence of a large set of software tools for assistance, it can be very difficult to actual-
ly come up with data which is relevant as well as generates useful networks, which can
represent the entire cas properly. Development of partial networks can result in completely
incorrect results as complete networks are required for the accurate calculation of centrality
- 39 -
measures (as discussed in the next section). Calculation of centrality measures requires the
use of all possible nodes in a connected network.
2.4.1.2 Theoretical basis
Graph theory has a long history. One of the earliest known representation of graph prob-
lem is the Leonard Euler's "Konigsberg bridge problem" in 1735which involved
representing a map of the river Pregel in the form of a graph[79].
A graph G = (V, E) is essentially formed by a set of vertices or nodes V and a set of
edges or arcs E. Each edge is assigned to two vertices, where the vertices might not be dis-
junct. Graph nodes as well as edges are given other alpha-numeric attributes including the
physical coordinates for display purposes. Networks are typically analyzed using the fol-
lowing methods:
1. Global network properties: Since real-world cas differ from purely randomly gen-
erated networks, they can be distinctly categorized using global properties. Three
distinct types of network models have been used in literature. The first one is the
Erdős-Rényi random network model[80]. The second model was proposed by Watts
and Strogatz by analysis of many real-world networks which they discovered to
have a smaller average shortest path length coupled with a high clustering coeffi-
cient as compared with random networks. These networks are labeled the Watts and
Strogatz small-world model[81] after their discoverers. Another model, called the
scale-free network model was proposed by Barabasi and Albert [82] based on ideas
of complex networks from real-world cas domains. These networks exhibit some
degree of a power law distribution in the "degree centrality". This implies that they
contain very few nodes with a higher degree and a large number of nodes with a
lower degree.
- 40 -
2. Network centralities: Unlike global network properties, centralities deal with par-
ticular nodes. As such, using a centrality analysis, different nodes can be ranked
according to their importance in Centrality indices. It is pertinent to note that not all
centrality measures might give substantial results for a particular network. As such,
it is typical to develop numerous networks as well as perform considerable analysis
by trial and error before a particular extraction of network as well as a centrality
measure is discovered showing significant results.
3. Network motifs: Motifs are commonly used network architectural patterns. They
are often used in cas modeling to discover particular patterns of local interactions.
Examples
include signal
transduction[83] and gene regulatory biological
networks[84]. Recent work includes Davidson [85] who has demonstrated emerg-
ing properties in Gene regulatory networks.
4. Network clustering: Biological networks have been studied in the form of hierar-
chical structures with motifs and pathways at lower levels to functional molecules
for large scale organization of networks by Oltvai and Barabasi[86].
2.4.1.3 Centralities and other Network measures
In this section, a description of some of the commonly used quantitative measures in com-
plex networks is provided.
2.4.1.3.1 Clustering Coefficient
The clustering coefficient is a means of quantitative measurement of the local cohe-
siveness in a complex network. It is a measure of the probability that two vertices with a
common neighbor are connected.
- 41 -
The formula for calculation of the clustering coefficient is given as follows:
Where
(2.1)
Ei = number of edges between neighbors of a node i
ki = degree of the node i
The global/mean clustering coefficient C = <Ci> is found by taking an average of all the
clustering coefficients.
2.4.1.3.2 Matching Index
Matching index gives a quantitative means of measuring the number of common
neighbors between any two vertices. In other words, for empirical networks, two function-
ally similar vertices need not be directly connected with each other. It is formally defined
as following:
(2.2)
Where the numerator represents the common neighbors while the denominator rep-
resents the total number of neighbors of the two nodes i and j.
2.4.1.3.3 Centrality measures
A centrality is formally defined as a function
for a directed or undirected
graph
[29]. A centrality is a real number assigned to each vertex allowing for a
pair-wise comparison across the entire network.
- 42 -
21iiiiECkk,,klikjlijijklikjlNAAMkkNAA:CVR(,)GVE 2.4.1.3.3.1 Degree Centrality
Degree centrality or simply "degree" of a node is a basic centrality measure which
identifies the connections of a node with other nodes. For directed networks, there can be
two kinds of degree centrality measures. The in-degrees and the out-degrees can be found
by the summation of the total in-links and the total out-links respectively. Degree centrality
is regarded as a local centrality measure however nodes having a higher degree can be
more important as compared with nodes with a lower degree. As such, deletion of high de-
gree nodes can disrupt the network structure and flow of interactions.
2.4.1.3.3.2 Eccentricity Centrality
Eccentricity of a node v is defined as the maximum distance between v and u for all
nodes u. Eccentricity centrality assists in computation of the maximum distance, which can
be defined as the length/distance of the longest shortest path to all other vertices present in
the network. Nodes which are easily reachable from other nodes receive the highest value.
Eccentricity identifies the node that can be reached within equal distance from all other
nodes in the network. Mathematically, eccentricity can be defined as:
Where
(2.3)
s and t are the nodes belonging to the set of vertices V
2.4.1.3.3.3 Closeness Centrality
Closeness centrality is similar to eccentricity but utilizes the minimum distance from a
target node to all other nodes within the network.
- 43 -
1()max{(,)}eccCsdistst The formula for the calculation of the closeness centrality is of a node s for all nodes t is
given as follows:
(2.4)
2.4.1.3.3.4 Shortest Path Betweenness Centrality
It is calculated by measuring the number of times a particular node comes in the shortest
path between any two nodes. In other words, it identifies the ability of a node to monitor
communication across the network. A higher value implies a better ability of a particular
node to monitor network communication. Mathematically it can be defined as:
Where
(2.5)
is the number of the shortest paths between nodes s and t containing the ver-
tex v between the nodes
is the total number of the shortest paths present between two nodes s and t
2.4.1.4 Software tools for Complex Networks
While Complex Network Analysis uses a number of mathematical tools, some of which
have been described in this chapter, for practical reasons, network scientists use a number
of software tools for performing network operations. These include network construction,
extraction, simulation, visualization and analysis tools such as Network Workbench[87],
Pajek[88] and Citespace[89], Cytoscape[90], Visone[91] and others.
- 44 -
1()(,)cloCsdistst()()stspbstvCv()stvst 2.4.2 Agent-based Modeling and Agent-based Computing
Agent-based Computing ranges from building intelligent agents to multiagent systems
and agent-based simulation models. There has been considerable focus in the Artificial In-
telligence and other research communities on something which has been termed agent-
based computing[92], [93]. Although prevalent across a number of domains, there is con-
siderable confusion in the literature regarding clearly a differentiation between its various
flavors. Here we shall attempt to disambiguate its various levels by providing a hopefully
clear set of definitions of the various terms and their uses.
2.4.2.1 Agent-Oriented Programming
The first of the agent-oriented paradigm which we see is focused more on the increased
use of artificially intelligent agents such as proposed by Russell and Norvig [94]. The fo-
cus here is on developing more complexity in an individual agent rather than on a large set
of agents.
2.4.2.2 Multi-agent oriented Programming
While agent-oriented programming focused on individual agents, it was difficult to ac-
tually design, implement and run multiple agents with such intricate internals because of
the need for massively parallel computational resources for even simple problems. Howev-
er, multiagent-oriented programming typically attempts to add at least some intelligence to
agents and observe their interactions. One thing to note here is that typically such models
have either few agents with a focus on certain aspects only, or else do not implement a
whole lot of complex behaviors[94], [95]. The primary reason for not implementing all in-
telligent paradigms is the inability to perform all such calculations in real-time in any
currently known artificial computational system. That is one reason why online algorithms
in AI typically use simplistic mechanisms for searching such as heuristic functions other
than purely reactive simple or table-driven agents (which can just perform an if else type of
- 45 -
actions). Another key problem is the knowledge-engineering bottleneck[96], where the
agents had difficulty communicating with each other. Although partially solved by lan-
guages such as Knowledge Query Manipulation Language (KQML)[97] and DARPA
Agent Markup Language (DAML)[98] and DAML+OIL[99] language released in 2000-01
time frame, practically large-scale intelligent agents have not seen the daylight of most
commercial or real-life scenarios (such as earlier predicted by agent-technology evange-
lists).
2.4.2.3 Agent-based Or Massively multiagent Modeling
One particular application of agents, which actually has found extensive use is what is
termed as individual or agent-based models. The key idea in these models is the focus on
use of "simple" simulated agents interacting with a large population of other agents with
emphasis on observing the global behavior. Here instead of the individual agents them-
selves, the interest of researchers is on the use of the interaction at the micro-level to
observe features at the global values. The focus of main-stream agent-based modeling and
simulation has been either social systems [20], [100] or else ecosystems[101], [102] .
. In this section, we shall review these paradigms further. Agent-based models [103] (or
individual-based models as termed in some scientific disciplines) have been used for the
modeling of a wide variety of complex systems[104]. Their use ranges from as diverse as
Biological Systems[105-120] to Social systems[121], [122], from Financial systems [123]
to supply chains [124], from the modeling of honey bees[125] to modeling of traffic
lights[126]. With such an impressive set of applications the strength of agent-based model-
ing[127] is quite apparent. However, with such a parallel evolution of ideas, there also
exists a set of domain-specific and perhaps conflicting concepts. These emerge primarily
due to a lack of communication or standardization between different scientific domains.
- 46 -
One possible explanation of the prevalence of agent-based modeling in such a diverse
set of scientific domains could be the models is their similarity to the human Cognition be-
cause perhaps this is how most humans perceive the surrounding world. In general, agent-
based models are particularly useful for developing models of Complex Adaptive Systems.
Typical examples of software used to develop agent-based modeling and simulation are
NetLogo[128], StarLogo[129], Repast [19], MASON[95], Swarm [130] and others .
2.4.2.4 Benefits of Agent-based Thinking
For system modeling involving a large number of interacting entities, at times it is more
appropriate to be able to access parameters of each individual "agent". In emerging com-
munication paradigms, network modelers need to model a variety of concepts such as life
forms, sensors and robots. Agent-based thinking allows for direct addressability of individ-
ual entities or agents. The concept of breeds allows system designer to freely address
various system entities. In agent-based modeling and simulation, agents are designed to be
addressed for performing a certain action.
Using a type name, agents or "turtles" can be asked to perform actions or change
their attributes. Agent-based models are developed with a high level of abstraction in mind.
In contrast to programming abstracts such as loops, designers ask agents without worrying
about the low-level animation or interaction details necessary for execution. As such, this
results in the production of compact programs with a high degree of functionality.
An important benefit of small programs is their ability to greatly reduce the tweak-test-
analyze cycle. Thus it is more likely to model complex paradigms within a short time
without worrying about the lower layers of other parameters unless they are important for
the particular application. As we shall see in this chapter, it also has great expressive power
and most of all, is fun to use. Although being used very frequently to model complex and
- 47 -
self-organizing systems it has not previously been used extensively to model computer
networks.
2.5 A Review of Agent-based Tools
In the arena of agent-based tools, a number of popular tools are available. These range
from Java based tools such as Mason to Logo based tools such as: StarLogo, NetLogo
[128] etc. Each of these tools has different strengths and weaknesses. In the rest of the
chapter, we focus on just one of these tools: NetLogo as a representative of this set. Build-
ing on the experience of previous tools based on the Logo language, NetLogo has been
developed from grounds up for complex systems research. Historically, NetLogo has been
used for modeling and simulation of complex systems including multi-agent systems, so-
cial simulations, biological systems etc, on account of its ability to model problems based
on human abstractions rather than purely technical aspects. However, it has not been wide-
ly used to model computer networks, to the best of our knowledge.
2.5.1 NetLogo Simulation: An overview
In this section, we introduce NetLogo and demonstrate its usefulness using a number of
modeling and simulation experiments. NetLogo is a popular tool based on the Logo lan-
guage with a strong user base and an active publicly supported mailing list. It provides
visual simulation and is freely available for download and has been used considerably in
multi-agent systems literature. It has also been used considerably in social simulation and
complex adaptive networks[131]. One thing which distinguishes NetLogo from other tools
is its very strong user support community. Most times, you can get a reply from someone
in the community in less than a day. NetLogo also contains a considerable number of code
samples and examples. Most of the time, it is rare to find a problem for which there is no
- 48 -
similar sample freely available either within NetLogo's model library or elsewhere from
the NetLogo M&S community.
Based on the Logo language, the NetLogo world consists of an interface which is made
up of "patches". The inhabitants of this world can range from turtles to links. In addition,
one can have user interface elements such as buttons, sliders, monitors, plots and so on.
NetLogo is a visual tool and is extremely suitable for interactive simulations. When one
first opens up a NetLogo screen, an interface with a black screen is visible. There are three
main tabs and a box called the command center. Briefly, the interface tab is used to work
on the user interface manually and the "Information" tab is used to write the documenta-
tion for the model. And finally the "procedures" tab is where the code is actually written.
The "command center" is a sort of an interactive tool for working directly with the simula-
tion. It can be used for debugging as well as trying out commands similar to the interpreter
model which, if successful, can be incorporated in one's program. The commands of a
NetLogo procedure can be executed in the following main contexts:
The key inhabitants of the Logo world are the turtles which can be used to easily model
network nodes. The concept of agents/turtles is to provide a layer of abstraction. In short,
the simulation can address much more complex paradigms which include pervasive mod-
els, environment or terrain models or indeed any model the M&S specialist can conceive
of without requiring much additional add-on modules. However, the tool is extensible and
can be directly connected to Java based modules. By writing modules using Java, the tool
can potentially be used as the front end of a real-time monitoring or interacting simulation.
For example, we could have a java based distributed file synchronization system, which
reports results to the NetLogo interface and vice versa, the NetLogo interface could be
used by the user to setup the simulation at the backend (e.g. how many machines, how
many files to synchronize and subsequently with the help of the simulation, the user could
- 49 -
simply monitor the results). Although the same can be done with a lot of other tools and
technologies, the difference is that NetLogo offers these facilities almost out of the box and
requires minimal coding besides being non-commercial, free and easy-to-install. A single
place where the turtle exists is a patch. Observer is a context, which can be used in general
without relating to either a patch or a turtle. The NetLogo user manual, which comes pre-
packaged with NetLogo, says: "Only the observer can ask all turtles or all patches. This
prevents you from inadvertently having all turtles ask all turtles or all patches ask all
patches, which is a common mistake to make if you're not careful about which agents will
run the code you are writing."
Inside the NetLogo world, we have the concept of agents. Coming from the domain of
complex systems, all agents inside the world can be addressed in any conceivable way,
which the user can think of, e.g., if we want to change the color of all nodes with commu-
nication power less than 0.5W, a user can simply say: ask nodes with [power < 0.5] [set
color green] or if a user wants to check nodes with two link neighbors only, this can be
done easily too and so on.
The context of each command is one of three. Observer object is the context, when the
context is neither turtle nor the patch. It is called the observer because this can be used in
an interactive simulation where the simulation user can interact in this or other context us-
ing the command window.
Although there are no real rules to creating a NetLogo program, one could design a pro-
gram to have a set of procedures which can be called directly from the command center.
However, in most cases, it suffices to have user interface buttons to call procedures. For the
sake of this chapter, we shall use the standard technique of buttons.
- 50 -
In this program there will be two key buttons; Setup and the Go buttons. The idea is that
the "setup" button is called only once and the "go" button is to be called multiple times
(automatically). These can be inserted easily by right clicking anywhere on the interface
screen and selecting buttons. So, just to start with NetLogo, the user will need to insert
these two buttons in his or her model, remembering to write the names of the buttons in the
commands. For the go, we shall make it a forever button. A forever button is a button
which calls the code behind itself repeatedly. Now, the buttons show up with a red text.
This is actually NetLogo's way of telling us that the commands here do not yet have any
code associated with them. So, let us create two procedures by the name of "setup" and
"go". The procedures are written, as shown in Figure 7, in the procedures tab and the
comments (which come after a semi-colon on any line in NetLogo) explain the actions that
are performed by the commands. This code creates 100 turtles (nodes in this case) on the
screen. However the shape is a peculiar triangle by default and colors are assigned at ran-
dom. Note that we have written code here to have the patches colored randomly.
- 51 -
1. to setup
2. ca ; Clears everything so if we call setup again, it won't
make a mess
3. crt 100 ;This means we are creating 100 turtles
4. [
5. setxy random-pxcor random-pycor ;These 100 turtles, we
want them to be spaced out at random
6. ; patch x and y co-ordinates
7. ]
8. let mycolor random 140 ; Randomly select a color value from
0 to 139
9. ask patches
10. [
11. set pcolor mycolor ; Ask all patches to set their color to
this random color
Figure 7 Code for setup
12. ]
13. end
To create the procedure for the go, the code can be written as listed in Figure 8.
to go
ask turtles
[
fd 0.001; ask each turtle to move a small step
]
end
Figure 8 Code for go
Now, if we press setup followed by go, we see turtles walking slowly in a forward di-
rection on the screen, a snapshot of which is shown in Figure 9.
- 52 -
Figure 9: Mobile Nodes
2.5.2 Overview of NetLogo for Modeling Complex interaction protocols
NetLogo allows for the modeling of various complex protocols. The model, however,
does not have to be limited to the simulation of only networks; it can readily be used to
model human users, intelligent agents, and mobile robots interacting with the system or
virtually any concept that the M&S designer feels worthwhile having in the model.
NetLogo, in particular, has the advantage of LISP-like[132] list-processing features. Thus
modeled entities can interact with computer networks. Alternatively, the simulation spe-
cialist can interact and create run-time agents to interact with the network to experiment
with complex protocols, which are not otherwise straightforward to conceive in terms of
programs.
- 53 -
As an example, let us suppose, if we were to model the number of human network man-
agers (e.g. from 10 to 100) attempting to manage a network of 10000 nodes by working on
workgroups the size of n nodes (e.g. ranging from 5 to 100) at one time while giving a total
of 8 hour shifts with network attacks coming in as a Poisson distribution; this can be mod-
eled in less than a few hours with only a little experience in NetLogo per se. The
simulation can then be used to evaluate policies of shifts to minimize network attacks.
Another example could be the modeling and simulation of link backup policies in case
of communication link failures in a complex network of 10,000 nodes along with network
management policies based on part-time network managers carrying mobile phones for
notification and management vs. full-time network managers working in shifts etc. all in
one simulation model. And to really make things interesting, we could try these out in re-
ality by connecting the NetLogo model to an actual J2ME based application in Symbian
phones using a Java extension; so the J2ME device sends updates using GPRS to a web
server which is polled by the NetLogo program to get updates while the simulation is up-
dated in a user interface provided by NetLogo. Again, although the same could be done by
a team of developers in a man-year or so of effort using different technologies, NetLogo
provides for coding these almost right out of the box and the learning curve is not steep
either.
This expressive nature of NetLogo allows modeling non-network concepts such as per-
vasive computing alongside human mobile users (e.g. in the formation of ad-hoc networks
for location of injured humans) or Body Area Networks come in play along with the net-
work. Now, it is important to note here that simulation would have been incomplete
without effective modeling of all related concepts which come into play. Depending upon
the application, these could vary from ambulances, doctors, nurses to concepts such as lap-
- 54 -
tops, injured humans etc. in addition to readily available connectivity to GIS data provided
by NetLogo extensions.
2.5.3 Capabilities in handling a range of input values
Being a general purpose tool, by design, the abstraction level of NetLogo is considered
higher. As such, the concepts of nodes, antenna patterns and propagation modeling are all
user-dependent. On one hand, this may look burdensome to the user accustomed to using
these on a regular basis, as it might appear that he or she will be working a little extra to
code these in NetLogo modeling. On the other hand, NetLogo allows for the creation of
completely new paradigms of network modeling, wherein the M&S specialist can focus on,
for example, purely self-organization aspects or on developing antenna patterns and propa-
gation modeling directly.
2.5.4 Range of statistics and complex metrics
NetLogo is a flexible tool in terms of using statistics and measurements. Any variable of
interest can be added as a global variable and statistics can be generated based on single or
multiple-run. Plots can be automatically generated for these variables as well.
Measurements of complex terms in NetLogo programs are very easy to perform. As an
example, if it is required to have complex statistics such as network assembly time, global
counters can be used easily for this. Similarly, statistics such as network throughput, net-
work configuration time, throughput delay can be easily modeled by means of similar
counters (which need not be integral). By default, NetLogo provides for real-time analysis.
Variables or reporters (functions which return values) can be used to measure real-time pa-
rameters and the analyst can actually have an interactive session with the modeled system
without modifying the code using the "Command Window".
- 55 -
2.6 Verification and Validation of simulation models
2.6.1 Overview
Validation of a simulation model is a crucial task[31], [133]. Simulations, however
well-designed, are always only an approximation of the system and if it was so easy to
build the actual system, the simulation approach would never have been used [134]. In tra-
ditional modeling and simulation practices, a standard approach to verification and
validation (V&V) is the three step approach given by Naylor et al. in [135] as follows:
1. The first step is to develop a model that can intrinsically be tested for a higher
level of face validity.
2. The next step is to validate the assumptions made in the model
3. Finally, the transformations of the input to output from the model need to be
compared with those for the actual real-world system.
2.6.2 Verification and Validation of ABMs
While verification and validation are important issues in any simulation model devel-
opment, in this section, we discuss the peculiarities associated with ABMs in the domain of
modeling and simulation of cas.
Firstly ABMs can be difficult to validate because there is a high tendency of errors and
un-wanted artifacts to appear during the development of an ABM as noted by Galan et al.
[136]. In addition, since the number of parameters in an ABM model can tend to be quite
high, Lucas et al. have noted that it is possible to fall into the trap of tweaking the variables
[137].
Bianchi et al. note that validation of agent based models can be quite challenging [138].
Another problem noted by Hodges and Dewar is as to how to ensure that the observed be-
- 56 -
havior is a true representative of the actual system[139]. Fagiolo et al. mention four set of
issues with Agent-Based models [140]:
1. A "lack of robustness".
2. Absence of a high degree of comparability between developed agent-based models.
3. A dire shortage of standard techniques for the construction and analysis of agent-
based models.
4. Difficulties in correlation of the models with empirical data.
Keeping these issues in mind, while validation techniques of agent-based models have
been mentioned in some domains such as computational economics and social simulation,
there are still a number of open issues:
A. There is no standard way of building Agent-based Models.
B. There is no standard and formal way of validation of Agent-based Models.
C. Agent-Based Modeling and Agent-Based Simulations are considered in the same
manner because there is no formal methodology of agent-based modeling [141].
D. Agent-Based Models are primarily pieces of Software however no software process
is available for development of such models.
E. All validation paradigms for agent-based models are based on quantifying and
measurable values but none caters for emergent behavior [142], [143] such as traf-
fic jams, structure formation or self-assembly as these complex behaviors cannot be
quantified easily in the form of single or a vector of numbers.
F. Agent-based models are occasionally confused with multi-agent systems (MAS)
even though they are developed for completely different objectives; ABM are pri-
marily built as simulations of cas, whereas MAS are typically actual software or at
time, robotic systems. Although MASs may themselves be simulated in the form of
an ABM but that does not change their inherent nature. We would like to note here
- 57 -
that differences between MAS and ABMs have been discussed earlier in the chapter
and a further exposition to this effect will be performed in Chapter 3.
2.6.3 Related work on V&V of ABM
In social sciences literature such as in the case of Agents in Computation Economics
(ACE), empirical validation of agent-based models has been described by Fagiolo et al. in
[140]. Alternate approaches to empirical validation have been noted by Moss in [144]. Val-
idation of models is closely related to model replication as noted by Wilensky and Rand in
[145]. An approach of validation based on philosophical truth theories in simulations has
been discussed by Schmid in [146]. Another approach called "companion modeling", an
iterative participatory approach where multidisciplinary researchers and stakeholders work
together throughout a four-stage cycle has been proposed by Barreteau et al. in [147].
A different point of view also exists in literature which uses agent-based simulation as a
means of validation and calibration of other models such as by Makowsky in [148]. In ad-
dition, agent-based simulation has also been shown to be useful in the validation of multi-
agent systems by Cossentino et al. in [149].
2.7 Overview of communication network simulators
In this section, an overview of different types of simulators used for the simulation of
communication networks is provided.
2.7.1 Simulation of WSNs
Current sensor network simulation toolkits are typically based on simulators such as
NS-2, OPNET[150], J-Sim[151], TOSSIM[152] etc. SensorSim is a simulation framework
[153]. Amongst traditional network simulations, J-Sim and NS2 have been compared in
[151] and J-Sim has been shown to be more scalable. Atemu [154] on the other hand, fo-
cuses on simulation of particular sensors. Likewise, TOSSIM [152] is a TinyOS Simulator.
- 58 -
2.7.2 Simulation of P2P networks
In terms of simulation of P2P networks, on one hand, there are actual implementations
and on the other, there are simulators of the P2P protocols. These include Oversim [155]
which is based on the OMNET++[156] simulator. Another simulator for P2P networks is
the PeerSim[157] simulator.
2.7.3 Simulation of Robotic Swarms
In the case of swarm robotics, simulators include Swarm-Bot simulator [158] and
WebotsTM robot simulator[159]. Other simulators include LaRoSim which allows for ro-
botic simulations of large scales[160].
2.7.4 ABM for complex communication networks simulation
While we have noted above that individual sub-areas of this domain such as WSNs,
Swarm robots as well as P2P networks all have dedicated as well as general purpose simu-
lators but all of them are limited to their specific domain, the benefit of using ABM in the
modeling and simulation of complex communication networks is that it is a general pur-
pose methodology and thus it can be used to simulate any combination of these in addition
to being able to simulate other entities such as humans, plants and animals in the simula-
tion.
2.8 Conclusions
In this chapter, we have discussed the background and related work required for an
understanding of the subsequent chapters. We have discussed modeling and simulation and
its relation to cas research. In the next chapter, we present complex network modeling level
of the proposed framework, which is concerned with developing Complex Network models
and performing Complex Network Analysis of large amounts of cas interaction data.
- 59 -
3 Proposed Framework: Complex Network Mod-
eling Level
In this chapter, we propose the complex network modeling level of the unified frame-
work. This idea is to develop a level which encompasses existing complex network
modeling studies as well as allow other studies to be conducted. This level is applicable
when suitable interaction data is available from a given cas. The basic idea is to develop
graphs in the form of complex networks using interaction data of different cas components
and characteristics. Subsequently these complex networks can be used to perform a CNA
for the discovery of emergent patterns in the data; patterns which would otherwise not have
been apparent using either legacy statistical or other mathematical methods. While this
chapter can be considered a generalized methodology, it can also be considered as an ex-
tension of the previous chapter since the data used for analysis of the case studies includes
a case study about of agent-based computing. Whereas the second case study is in the do-
main of consumer electronics selected specifically to demonstrate the generalized
applicability of the proposed framework level.
3.1 Overview
This chapter demonstrates the application of CNA on two separate case studies. The
idea is to use Information visualization and Complex Network Analysis to discern complex
patterns in large corpora of reliable scientific interaction data.
Both case studies use different corpora but are based on the well-defined structured data
from the standard source of The Thomson Reuters Web of Science (WOS) which is well-
known as a standardized indexing database for scientific research papers. The corpora was
- 60 -
selected as the WOS since it is a stable index and has a well-defined methodology for add-
ing new data. While it is always possible to find some indexing errors in any database, this
particular dataset presents a suitable candidate better than most other manually collected
datasets. The first case study of the application of the proposed methods is on "Agent-
based Computing" while the second is on "Consumer Electronics".
3.2 Generalized Methodology
In Figure 10, we give details of how complex network modeling level of the proposed
framework can be useful in developing complex network models of a cas and subsequently
performing CNA to determine emergent behavior and patterns in the data.
The first step is the acquisition or retrieval of interaction data. In other words, only data
which allows the development of interaction models would be needed in this framework
level. Typical summarized statistical data or records would not be sufficient for the devel-
opment of these models. Typically such data must demonstrate important relations of some
kind between various components to prove useful for complex network development.
After data retrieval, the next step is to develop complex network models. While it might
appear a trivial exercise to develop these models, it can actually take significant time and
effort before suitable networks can be developed. One especially difficult task is the selec-
tion of the most suitable columns for network extraction and construction. This exercise
can require extensive experimentation using trial and error before the most suitable col-
umns can be discovered for network construction. Network manipulation is often applied
to develop and extract suitable networks in an iterative manner. This can involve selective
removal, insertion, merging and splitting of various nodes and edges followed by refine-
ment of the network using various colors and sizes of the nodes and edges based on loaded
data from the cas components.
- 61 -
Figure 10 Proposed Methodology for the complex network modeling level (for the discovery of emer-
gent patterns based on available interaction data of cas components)
Gradually the networks start to shape up. Eventually this process can result in giving
useful information about the topological structure of the entire domain. This approach is
particularly useful when other basic approaches to modeling fail (such as using statistical
tests, machine learning, data mining etc.). The next step in this level is to perform CNA.
There are essentially two different types of CNA, which have been used in previous litera-
ture (and described earlier in Chapter 2). One type of CNA is useful for classification of
- 62 -
the entire network using global properties such as degree distribution. The other type of
CNA can often prove to be more useful and involves the calculation of various quantitative
measures such as Centralities, Clustering coefficients and indices. Once these measures
have been calculated, quite often, the analysis is still far from being over. Next, we demon-
strate the application of the complex network methods on two different case studies.
3.3 Case Study I: Agent-based Computing
The area of interest of the first case study is agent-based computing, a diverse research
domain concerned with the building of intelligent software based on the concept of
"agents". In the first case study, we perform Scientometric analysis to analyze all sub-
domains of agent-based computing. The data consists of 1,064 journal articles indexed in
the ISI web of knowledge published during a twenty year period: 1990-2010. These were
retrieved using a topic search with various keywords commonly used in sub-domains of
agent-based computing. The proposed approach employs a combination of two applica-
tions for analysis, namely Network Workbench and CiteSpace as mentioned earlier in
chapter 2. The idea is that Network Workbench allowed for the analysis of complex net-
work aspects of the domain whereas a detailed visualization-based analysis of the
bibliographic data was performed using CiteSpace. Results include the identification of the
largest cluster based on keywords, the timeline of publication of index terms, the core
journals and key subject categories. We also identify the core authors, top countries of
origin of the manuscripts along with core research institutes. Finally, the results have re-
vealed the strong presence of agent-based computing in a number of non-computing
related scientific domains including life sciences, ecological sciences and social sciences.
- 63 -
3.3.1 Problem Statement:
As discussed earlier in the previous chapter, agent-based computing encompasses
both multiagent systems as well as agent-based modeling. Agent design and simulation go
hand in hand but in completely different ways in different sub-domains of agent-based
computing. So, e.g. on one hand, there are researchers whose research goals revolve
around the design of various types of agents where the role of simulation is closely linked
to validation of the future operation of actual or physical agents[161]. On the other, there is
another group of researchers whose goal is not agent-design but rather the agent-design is a
means of developing simulations which can lead to better understanding of global or emer-
gent phenomena associated with complex adaptive systems[6], [141]. This broad base of
applications of this research area thus often leads to confusions regarding the exact seman-
tics of various terms in the literature. This is tied closely to the evolution of "agent-based
computing" into a wide assortment of communities. These communities have at times, per-
haps nothing other than the notion of an "agent" in common with each other.
The goal of the CNA here is to use network development, analysis and visualization
to give a Scientometric survey of the diversity and spread of the domain across its various
sub-domains, in general, and to identify key concepts, which are mutual to the various sub-
domains, in particular. These includes identifying such visual and Scientometric indicators
as the core journals, the key subject categories, some of the most highly cited authors, most
central authors, institutes of highly cited authors and the top countries of manuscript origin.
Our goal is to use a combination of two different tools based on their relative
strengths; Network Work Bench [87], [162] for performing a Complex Network Analysis
using Network Analysis Toolkit and CiteSpace[89].
The objectives of this study can be summarized as follows:
- 64 -
• To identify the largest clusters of inter-connected papers for the discovery of com-
plex interrelations of various sub-areas of agent-based computing.
• To discover "bursts"[163], [164] of topics along with a timeline of publication of
the index terms used.
• To identify the core journals for the whole of agent-based computing ranging from
agents, multi-agent systems to agent-based and individual-based simulations, in
terms of citations of articles.
• To identify the key subject categories.
• To identify and study the most productive authors, institutes and countries of manu-
script origins.
3.3.2 Methodology
In this section, we give an overview of the research methodology for the first case study.
While the overall methodology of any CNA is similar, there are always peculiarities of ap-
plication in each case study as well as each cas domain. As such this involved data retrieval
as well as subsequent CNA performed using the two separate tools. It is pertinent to note
here that this analysis might appear trivial and a simple application of the complex network
methods and software, however, the actual case is far from that. Simply finding and under-
standing the correct data tags for development of complex networks requires extensive
understanding of the data types and transformations. Thus the analyses in this chapter re-
flect the last iteration of the application of these methods rather than a simple and naïve
application of these tools. In addition, the datasets were time and again verified manually
and the data was collected initially by hand and consisted of several tens of thousands of
records. Subsequently the networks which are demonstrated here were developed however
again these are not the first networks which were developed. Network manipulation was
also applied extensively. In short, developing complex network models and pruning net-
- 65 -
works for custom application of Complex Network methods is a nontrivial task and takes
considerable time, effort and learning for each application. It is best learnt by practicing
and observing existing case studies, which is the reason we are demonstrating two different
case studies as a means of how other case studies can be applied using the proposed com-
plex network modeling level of the framework.
3.3.2.1 Data collection
The input data was retrieved using searches from the WOS [165]. A thorough topic
search for data was devised to cater for various aspects and keywords used in agent-based
computing in the following three sub-domains:
1. Agent-based, multi-agent based or individual-based models
2. Agent-oriented software engineering
3. Agents and multi-agent systems in AI
The search was performed in all four databases of Web of Science namely SCI-
EXPANDED, SSCI, A&HCI, CPCI-S for all years. Details of the search have been provid-
ed in the Appendix 1. For the sake of analysis, the range of years 1990-2010 was selected
with search limited to only Journal articles. Bibliographic records retrieved from SCI in-
clude information such as authors, titles, abstracts as well as references. The addition of
cited references resulted in a total of 32, 576 nodes. The details of the search keywords
along with reasoning for the selection are given in the Appendix 1.
3.3.3 Results
The results are presented starting with a basic look at the overall picture of articles re-
trieved from the Web of Science. As can be noted in Figure 11, the articles in this domain
start primarily during the early 1990's and gradually keep rising and as such reach a total
of 148 articles published merely in the last year 2009.
- 66 -
Figure 11: Articles published yearly
In addition, since the popularity of a domain is known to be based on its number of cita-
tions, we need to observe this phenomenon closely. It can be observed in the graph using
data from the Web of Science in Figure 12. Thus, starting from a very small number of ci-
tations, Agent-based computing has risen to almost 1630 citations alone during the year
2009.
Figure 12: Citations per year
- 67 -
3.3.3.1 Network Workbench (NWB)
Next, to get the big picture of the citation network, a network (paper co-citation net-
work) was extracted from the ISI records using Network Workbench tool. In addition,
analysis was performed for the extracted network using NWB based Network Analysis
Toolkit [166]. The extracted global properties of the network are shown in Table 1. The
first thing to note here is the large number of nodes i.e. 32, 576 which appears much larger
than reported above from the Web of Science data. This is primarily because it includes the
cited references as well as the original nodes. Now, we see here that there are no isolated
nodes, which is obvious because every paper will at least cite some other papers at the very
least. This is followed by the 39, 096 edges. Another interesting feature is the average de-
gree, which we see is 2.4. The fact that the average degree is not significantly higher shows
actually that a large number of papers have been co-cited in this corpus otherwise the de-
gree would have been much higher.
The graph itself is not weakly connected but it consists of 78 weakly connected compo-
nents with a largest component of size 31,104 nodes. The weakly connected components
are formed based on articles connected with a small number of other articles identifying
the importance of these papers (Which shall be investigated in depth later on in this chap-
ter). Finally, we note that the density is 0.00004 however knowing the density does not
give much structural information about the Scientific domain per se. These properties are a
characteristic of the retrieved empirical data. However, we note here that while interesting,
they do not provide in depth insights of the data. So we perform a further set of analysis
using NWB
- 68 -
Table 1 Basic Network Analysis of Extracted Paper Citation Network
Attribute
Nodes:
Isolated nodes:
Edges:
Self Loops
Parallel Edges
Edge Attributes
Valued Network
Average total degree:
Weakly Connected
Weakly connected components
Value
32, 576
None
39, 096
None
None
None
No
2.400
No
78
Largest connected component
31, 104
Size
nodes
Density (disregarding weights):
0.00004
Here the structure of the overall agent-based computing domain can be examined fur-
ther using NWB. To allow for the examination of the domain using the specific strengths,
central to the NWB tool, the top nodes were extracted using a local citation count > 20.
Subsequently the network was visualized using GUESS[167] and nodes were resized ac-
cording to the values of the citation count. The result can be observed in Figure 13. Here
we can note the peculiar relation of some of the top papers connected with Volker Grimm's
paper in the center. Apart from Grimm's own papers, these include Robert Axelrod's as
well as Parker's and Deangelis's works. In addition, we can note that here Kreft's and Hus-
ton's work in Microbiology and Biosciences as well as Wooldridge's and Ferber's work in
the domain of multiagent systems are also showing up. For further analysis, we move on to
the more advanced Information Visualization tool, namely CiteScape.
- 69 -
Figure 13: Top papers with citations > 20
3.3.3.2 CiteScape
The CiteScape tool allows for a variety of different analysis. CiteScape directly operates
on the downloaded ISI data and builds various networks using time slicing. Subsequently
using the various options selected by the user, the network can then be viewed in different
ways and parameters can be analyzed based on centrality (betweenness) as well as fre-
quency.
3.3.3.2.1 Identification of the largest cluster
The goal of the first analysis was to observe the big picture and identify the most im-
portant indexing terms. Based on a time slice of one year, here in Figure 14, we see the
largest cluster.
These clusters are formed using CiteSpace, which analyzes clusters based on time slic-
es. Here the links between items show the particular time slices.
- 70 -
Grimm's Ecological Modeling
paper
Figure 14: Largest Cluster based on indexing terms
In this figure, we first note the top where the year slices from 1990 to 2010 show up in 4
year slices. Using different colors in CiteSpace allows us to clearly identify turning points
and other interesting phenomena in literature. We notice here that one of Grimm's papers is
actually the key turning point from Agent-based Modeling to Multiagent-systems cluster.
Another thing that can be noted here is the role of this paper in connecting a large number
of papers in 2010 to the other papers in 2002-2006 era showing a somewhat delayed ac-
ceptance of the concepts.
3.3.3.2.2 Timeline and bursts of index terms
The next analysis performed was to observe the clusters on a timeline as shown in Fig-
ure 15.
In complex networks, various types of centrality measures such as degree centrality, ec-
centricity, closeness and shortest path betweenness centralities etc. Citespace, in particular,
uses betweenness centrality[89]. As discussed in previous chapter, this particular centrality
is known to note the communication monitoring ability of a vertex for other network verti-
- 71 -
ces. In other words, a higher centrality ensures that the vertex is between more of the
shortest paths between other nodes relative to other nodes with lower centrality.
Using a time line especially helps identify the growth of the field. Please note here that
these include papers which are based on agent-based computing as well as papers which
are cited by these papers. Here, the red items are the "bursts". Bursts identify sudden inter-
est in a domain exhibited by the number of citations.
We can see that there are a lot of bursts in the domain of agent-based model. In addi-
tion, even the preliminary analytics and visualization here confirms the hypothesis that
agent-based computing has a very wide spread article base across numerous sub-domains.
This is obvious here as we see clusters ranging from "drosophila" and "yellow perch bio-
energetics" to those based on "group decision support systems", all ranging from different
domains. Further analysis strengthens this initial point of view based on an examination of
the various clusters.
Here, the results demonstrate the effects of the semantic vagueness discussion earlier.
So, e.g. Where concepts such as group decision support system, rule-based reasoning and
coordination are concepts tied closely with developing intelligent agents, they also show
up in the domain right alongside agent-based modeling.
- 72 -
Figure 15: Timeline view of terms and clusters of index terms (based on centrality) also showing cita-
tion bursts
3.3.3.2.3 Analysis of Journals
Our next analysis was to identify the key publications of the domain3. This can be seen
in Figure 16. Here the key journals are identified based on their centrality.
Once again, we can note here that the vagueness in the terms of use again shows up in
the set of mainstream Journals of the domain with "Artificial Intelligence" and "Communi-
cations of the ACM" being relevant mostly to Agents associated with the concepts of
"Intelligent agents" and "Multiagent Systems" whereas "Econometrica", "Ecological
Modeling" and the journal "Science" representing the "agent-based modeling" perspective.
3 It is pertinent to note here that we faced one peculiar problem in the analysis of the retrieved ISI data. The
Web of Science data identified a Journal named "Individual-based model". However extensive searches
online did not find any such journal.
- 73 -
Figure 16: Key Journals based on centrality
In Table 2, we give details of these top journals based on centrality shown in the figure.
This represents the centrality of the top ten key journals. In terms of centrality, the "ECOL
MODEL" Journal has the highest value of centrality among all the journals. In addition,
here we observe that "ANN NY ACAD SCI", "CAN J FISH AQUAT SCI", "NATURE"
and the "ANN NY ACAD SCI" are also some of the top Journals of this domain in terms
of Centrality.
- 74 -
Table 2 Top Journals based on Centrality
Centrality Title
Abbreviated Title
0.47
Ecological Modelling
ECOL MODEL
0.29
0.21
0.32
0.15
0.09
0.08
0.08
0.08
Science
SCIENCE
Communications of the ACM
COMMUN ACM
Artificial Intelligence
ARTIF INTELL and
ARTIFICIAL INTELLIGE
Econometrica
ECONOMETRICA
Journal of Theoretical Biology
J THEOR BIOL
Canadian Journal of Fisheries
CAN J FISH AQUAT
and Aquatic Sciences
SCI
Nature
NATURE
Annals of the New York Acad-
ANN NY ACAD SCI
emy of Sciences
Next, we analyze the publications in terms of their frequencies of publication as given
in Table 3. Now, interestingly, the table sorted in terms of article frequency, gives a slightly
different set of core journals. Through frequency analysis of the title words of 240 journals,
it can be seen that "ECOL MODEL" is still at the top with a frequency of 295 articles.
"NATURE" and "SCIENCE" follow with 231 and 216 published articles respectively. "J
THEO BIO" is next with 167 articles. Next "ECOLOGY" has published 145 articles. This
is followed by "AM NAT", "TRENDS ECOL EVOL", "LECT NOTES ARTIF INT" and
"P NATL ACAD SCI USA".
- 75 -
Table 3 Core Journals based on frequency
Frequency
Title
Abbreviated Title
295
231
216
167
145
123
121
121
121
Ecological Modelling
Nature
Science
ECOL MODEL
NATURE
SCIENCE
Journal of Theoretical Biology
J THEOR BIOL
Ecology
The American Naturalist
ECOLOGY
AM NAT
Trends in Ecology & Evolution
TRENDS ECOL EVOL
Lecture Notes in Artificial Intelligence
LECT NOTES ARTIF INT
Proceedings of the National Academy of Scienc-
P NATL ACAD SCI USA
es
3.3.3.2.4 Analysis of Categories
Our next analysis was to discover the prevalence of various agent-based computing arti-
cles in various subject categories. This visualization is shown in Figure 17. The detailed
analysis of the subject category based on centrality follows in Table 4.
- 76 -
Figure 17: Complex network model of the category data
Table 4 Key categories based on Centrality
Centrality
Category
0.44
0.4
0.36
0.34
0.31
0.14
0.13
0.13
0.12
Engineering
Computer Science
Mathematics4
Ecology
Environmental Sciences
Biotechnology & Applied Microbiology
Biology
Economics
Psychology
4 Shown as two categories erroneously by CiteScape
- 77 -
The table represents the centrality based ordering of the key subject categories. It is im-
portant to note here is that this table shows top categories from a total of 75 categories.
Here, it can be observed that in terms of centrality, the "Engineering" category leads other
categories. It is however, closely followed by computer science, mathematics, ecology and
environmental sciences. It appears however that the "Psychology" category has the lowest
value of centrality among all other categories.
For comparative analysis, we also analyze these categories using the publication fre-
quency of articles. The results of this analysis are presented in Table 5.
Table 5 Subject Categories according to frequency
Frequency
Category
287
195
145
92
66
57
57
53
52
52
Computer Science
Ecology
Engineering
Social Sciences
Biology
Environmental Sciences
Mathematics
Environmental Studies
Operations Research & Management Science
Fisheries
The table represents the frequency of the top ten key categories. Through frequency
analysis of the title words of 75 categories, we interestingly come up with a slightly differ-
ent set of results. Here, "Computer Science" with a frequency of 287 articles leads the rest
and is followed closely by ecology and engineering. An interesting observation based on
the two tables is that there are certain categories which have relatively low frequency but
are still central (in terms of having more citations) such as Mathematics. Amongst the top
categories, we can also see categories with a relatively lower frequency such as 53 for
- 78 -
"Environmental Sciences" and 52 for "Operations Research & Management Science" as
well as "Fisheries". This detailed analysis shows that prevalence of agent-based computing
is not based on a few sporadic articles in a variety of subject categories. Instead, there are
well-established journals and researchers with interest in and publishing a considerable
number of papers in this domain, especially agent-based modeling and simulation.
3.3.3.2.5 Bursts in subject categories
Next, we analyze how various subject categories have exhibited bursts. This is shown in
the Table 6. Here we can see that fisheries have the largest burst associated with the year
1991. Next are two closely related categories "Marine and Freshwater Biology" and
"Ecology" in the same time frame. One very interesting finding here is that there are a lot
of bursts in non-computational categories.
Burst
13.09
10.3
9.36
5.58
4.25
3.76
3.3
3.2
Table 6 Key bursts in subject categories
Category
Fisheries
Marine & Freshwater Biology
Ecology
Economics
Evolutionary Biology
Mathematics
Genetics & Heredity
Oceanography
Year
1991
1991
1991
1996
1993
1990
1993
1995
3.3.3.2.6 Analysis of Author Networks
In this section, we analyze the author co-citation networks. Figure 18 shows the visuali-
zation of the core authors of this domain. Here red color exhibits burst of articles and
- 79 -
concentric circles identify separation of years of publications. The size of these circles
based on the centrality values of the author. The blue is color represents the older papers
and the green gives not very old papers. Yellowish and reddish colors are for relatively
more recent publications. Here, the descriptions are based on the color figures, which can
be viewed in the online version of the paper, since Citespace, being a visualization tool,
relies extensively on the use of colors to depict styles.
Figure 18: Co-Author network visualization
Although this visualization perhaps gives a broad picture of the various authors, we also
present a detailed analysis of this data. This can be seen in a tabular form as shown in Ta-
ble 7. Here, we can observe that the top cited (most central) author is Don DeAngelis, a
Biologist. Don is followed by another Biologist Michael Huston. Next is Volker Grimm, an
expert in agent-based and individual-based modeling and Kenneth A Rose, an ecologist.
Next, we have Robert Axelrod, a Political Scientist. Nigel Gilbert, a Sociologist and Mike
Wooldridge, a Computer Scientist is next in the list. Finally, we have François Bousquet
from the field of Ecological Modeling (Agriculture) and Steven F. Railsback, an Ecologist.
This is quite an interesting result because agents and agent-based computing in general was
- 80 -
supposed to be primarily from Computer Science/AI and have very specific meanings.
However, the results show that it is actually quite prevalent and flourishing in an uninhibit-
ed manner in various other fields in terms of renowned (ISI-indexed) archival Journal
articles.
Table 7 Authors in terms of centrality
Centrality Author
0.5
0.33
0.16
0.08
0.08
0.07
0.07
0.06
0.06
DEANGELIS DL
HUSTON M
GRIMM V
ROSE KA
AXELROD R
GILBERT N
WOOLDRIDGE M
BOUSQUET F
RAILSBACK SF
For further comparative analysis, we plotted the top authors in terms of frequency of
publications. This is shown in Figure 19.
- 81 -
Figure 19: Authors in terms of frequency
The detailed analysis of the metrics of the key authors in this domain is shown in a
tabular form in Table 8 in terms of their frequency of publication in the selected pruned
network.
Frequency
128
118
101
99
90
83
76
68
64
59
Table 8 Top authors based on frequency
Author
GRIMM V
DEANGELIS DL
EPSTEIN JM
WOOLDRIDGE M
AXELROD R
GILBERT N
BOUSQUET F
HUSTON M
FERBER J
PARKER DC
It is important to note here that while the results are based on a total of 387 cited au-
thors, new names appear in this table such as Joshua M. Epstein, a Professor of Emergency
- 82 -
Medicine. Prof. Epstein is an expert in human behavior and disease spread modeling.
Jacques Ferber, a Computer Scientist is another new name in the list along with Dawn C.
Parker, an agricultural Economist.
3.3.3.2.7 Country-wise Distribution
Next, we present an analysis of the spread of research in agent-based computing origi-
nating from different countries based on centrality. Here, in Figure 20, we can see the
visualization of various countries. Please note here that the concentric circles of different
colors/shades here represent papers in various time slices (we have selected one year as
one time slice). The diameter of the largest circle thus represents the centrality of the coun-
try. Thus the visualization identifies the key publications in the domain to have originated
from the United States of America. This is followed by papers originating from countries
such as England, China, Germany, France and Canada.
Figure 20: Countries with respect to centrality
3.3.3.2.8 Analysis of Institutes
- 83 -
In this sub-section, we present visualization for a detailed analysis of the role of various
Institutes. We can see the temporal visualization of various institutes in the domain and the
assortment of popular keywords associated with them on the right in Figure 21.
Figure 21: Timeline based visualization of institutes
Here, the prevalence of manuscripts originating from the Oak Ridge National Laborato-
ries from the early 1990's can be observed here. Next, we see papers from US Fish and
Wildlife services. Then, we can see papers from Kyushu University, Japan, University of
Birmingham, UK and French National Institute of Agricultural Research (INRA) close to
1998. From this time to 2000, a prevalent institute is Chinese University, Hong Kong. In
around 2002, the French institute Centre de coopération Internationale en recherche
agronomique pour le développement (CIRAD), an Institute of Agriculture Research for
Development is prevalent followed closely by a sister institute, the Institut de Recherche
pour le Développement (IRD). More recent newcomers to the field of agents include the
MIT and the Massachusetts General Hospital, associated with the Harvard University.
In the following Table 9, we perform an alternative analysis which is based instead
on the frequency of articles.
- 84 -
Table 9 Core Institutes based on frequency
Frequency
Institute
11
10
9
9
8
8
8
8
7
7
7
7
7
University of Illinois, USA
INRA, France
University of Michigan, USA
University of Minnesota, USA
University of Sheffield, UK
Nanyang Technological University, Singapore
Italian National Research Council (CNR), Italy
Oak Ridge National Labs, USA
Harvard University, USA
MIT, USA
University of Washington, USA
University of Hong Kong, P. R. China
IRD, France
The Table 9 listed represents the top institutes in terms of frequency. It should be noted
that the frequency analysis was performed based on title words of a total of 328 institutes.
"University of Illinois"5 in general has a top ranking with a frequency of 11 articles. It is
followed closely by INRA (France) with a frequency of 10 articles. University of Michigan
and University of Minnesota, both from the US follow next with 9 articles each. With 8
articles each, next we have University of Sheffield (UK), Nanyang Technological Universi-
ty (Singapore), Italian National Research Council (CNR) and the Oak Ridge National Labs
(USA).
5 ISI data extracted using CiteSpace does not differentiate further as to which exact campus of the University
of Illinois is considered here (of the primarily three campuses i.e. UIUC, UIC, UIS.)
- 85 -
3.4 Case Study II: Consumer Electronics
The second case study demonstrates the application of the Complex Network methods
to scientific literature in the consumer electronics domain. As mentioned earlier at the start
of the chapter, while the first case study can be considered as an extension of the back-
ground of the previous chapter, the second case study can be considered as a proof-of-
concept of the generalized application of the complex network methods.
3.4.1 Problem Statement
Consumer electronics as a domain has evolved with the evolution of electronics tech-
nology [168]. The guiding principles for this gradual evolution range from consumer
demand[169] in general and the drive from the industry in particular[170]. Thus the state
of the domain, as it is today, reflects a number of trends from the past. To demonstrate the
variety of effects which can dictate trends, trends have been noted to be governed by
seemingly unrelated interactions such as a company's internal management policies[171].
The evolution of the domain is perhaps reflected by its growth in terms of sales which if
aggregated, have shown to be almost exponential [172]. As evident from the analysis of the
Thomson Reuters citation data, a major evolution in the domain occurred during the last 25
years as the citation corpus does not reflect the indexing term of "consumer electronics" in
the WOS before this period.
Our research questions include the identification of the highly central journals in con-
sumer electronics. In addition, the analyses give results ranging from the identification of
the top authors and the top papers.
- 86 -
3.4.2 Methodology
Similar to the previous study on agent-based computing, citation data used here for the
identification of trends and interest of the general Consumer Electronics Community was
retrieved from the WOS using the topic "Consumer Electronics". All available data sources
were used to get a complete picture of trends. These included the SCI-EXPANDED, SSCI,
A&HCI and the CPCI-S databases. The data was retrieved based on search on the "Con-
sumer Electronics" topic (which translates to a search including keywords, abstract, title
etc.) in addition to all records of the IEEE Transactions on Consumer Electronics. Subse-
quently, the data obtained from the ISI Web of Knowledge was imported and analyzed
primarily using various tools such as CiteSpace. First the data was spliced into yearly data.
Next, different co-citation networks were formed such as for Journals, authors and so on.
These were then analyzed based on both the frequency of publication as well as centrality
(in terms of being cited by others).
3.4.3 Results
The citation data was collected from all years including the current year (2010). The da-
ta included a total of 5,372 papers and including cited references, it amounted to 26,860
records.
Figure 22: Frequency of papers in Consumer Electronics domain
- 87 -
Figure 23 visualizes the top journals of the domain. Different colors (or scales of grey in
print) identify the year slices with each slice made up of ten years. The node sizes identify
the frequency of articles, which in essence is an indicator of trends or community interest.
The number of articles listed hailing from IEEE Transactions on Consumer Electronics
(CE) is 1665. The second Journal in the list is IEEE Transactions on Communications with
around one third number of the IEEE Transactions on CE articles (i.e. 560). Here the cen-
trality of the IEEE Transactions on Consumer Electronics is apparent here with the yellow
color representing the papers published in the nineties. The results for the top 10 Journals
are summarized in tabular form in Table 10.
Figure 23: Visualization of the citation network of the top Journals of the domain
- 88 -
Table 10 Top Journals in Consumer Electronics
Serial
Journal Name
1.
2.
3.
IEEE T CONSUM ELECTR
IEEE T COMMUN
IEEE T CIRC SYST VID
4.
P IEEE
5.
6.
7.
8.
9.
10.
IEEE T IMAGE PROCESS
IEEE J SEL AREA COMM
IEEE COMMUN MAG
IEEE T CIRCUITS SYST
IEEE T BROADCAST
ELECTRON LETT
Here, we can note that Figure 24 gives the overall picture of the top papers of the do-
main/ Here, the size of the nodes and the font is scaled according to the higher value of the
centrality of the paper.
Figure 24: Network of top papers in Consumer Electronics
- 89 -
The Table 11 summarizes the top 10 papers of the IEEE Transactions on CE.
Serial
1.
2.
3.
4.
Table 11 Top papers of IEEE Transactions on Consumer Electronics
Paper
Citations
The JPEG2000 still image coding system: An overview (2000)
Pilot tone selection for channel estimation in a mobile OFDM system
(1998)
A new remote user authentication scheme using smart cards (2000)
An efficient remote use authentication scheme using smart cards
(2000)
398
242
192
147
5.
Contrast enhancement using brightness preserving bi-histogram equal-
127
ization (1997)
6.
The Trustworthy Digital Camera - Restoring Credibility To The Pho-
118
tographic Image (1993)
7.
Channel estimation for OFDM systems based on comb-type pilot ar-
116
rangement in frequency selective fading channels (1998)
Digital Sound Broadcasting To Mobile Receivers (1989)
Multidirectional Interpolation For Spatial Error Concealment (1993)
A modified remote user authentication scheme using smart cards
(2003)
90
87
86
8.
9.
10.
Subsequently we can note the co-citation network of the key authors in the consumer
electronics domain in Figure 25. Note that while we have discussed and analyzed similar
results in the previous case study in considerably more details, here we only briefly discuss
them because this case study is a demonstration of the generalized applicability of the
methodology.
- 90 -
Figure 25: Co-citation network in Consumer Electronics
3.5 Summary of results
In this chapter, we presented the generalized ideas related to complex network modeling
followed by two case studied of different types of complex interaction data. The goal of the
case studies was to demonstrate how complex network modeling level of the proposed
framework can be exploited to discover patterns of emergence in different entities consti-
tuting a cas. The two case studies were chosen based on the appropriate availability of
interaction data from a well-structured data source i.e. the WOS. In this section, results of
the two case studies are correlated with the general theme of complex network modeling
level in the form of a set of summarized results.
- 91 -
3.5.1 Case Study I
In this case study, analysis was performed using different types of entities and their in-
teractions in the agent-based computing domain.
Firstly, during the analysis conducted using Network Workbench, the top cited papers of
the domain was identified from both the multiagent side of the domain (based in Computer
Sciences) in addition to the agent-based modeling ideas (from ecology/social sciences).
Next, using a clustering of index terms, the ideas were identified to originate from the
agent-based modeling area. A major recent "turning point" signifying a trend was next
identified based on some of Grimm's papers. An older identified turning point was based
on Goldberg's 1989 paper on Genetic Algorithms, which tied in with the ecological cluster
labeled as "Cannibalism". In addition, the prevalence of the "catchment" cluster was noted
identifying the social aspects coupled with the "increasing returns" cluster identifying eco-
nomics. Thus one single extracted network was able to demonstrate how these diverse
areas are tied in together by means of identification of emergent patterns. The "persis-
tence" cluster again identifies a "multiagent" related set of documents. Here, the oldest
cluster is the "life history" cluster with roots in Biosciences. This strong intermingling of
ideas can be noted to have autonomously guided the evolution of the agent-based compu-
ting domain. Whereas subsequent analyses only prove the point that there is further
diversity in each of subject categories, authors, institutions and countries. However, while
the domain has strong footing, being able to use visualization has allowed the examination
of various complex interactions of the sub-domains, which would not have otherwise been
possible to monitor without the use and application of network extraction, modeling and
manipulation techniques.
- 92 -
3.5.2 Case Study II
In this case study, a complex network modeling and analysis of the consumer electron-
ics domain has been performed. Using centrality measures has allowed a visual and
quantitative identification of the top journals, the top authors as well as the top research
papers in the consumer electronics scientific domain.
3.6 Broad applicability of the proposed framework level
In this section, an overview of how the proposed framework level can be used to con-
duct different complex network modeling based cas research studies is presented. It can be
noted from the two case studies conducted in this chapter that each application domain in
complex network modeling can have slightly different idiosyncrasies. However, complex
network methods are suitable only in the case where specific target data is either readily
available or else such data can be extracted by some means. This interaction data must
primarily allow the creation of complex networks. In other words, the data columns must
point out relationships between individual columns. Using these relationships, complex
networks need to be designed and extracted. At times, the extracted networks contain con-
siderably more information than needed. As such, it can be difficult to study the emergent
patterns. As such, network nodes and links might need to be selectively eliminated or else
new nodes and links might need to be developed inside the network. In addition, different
network measures might need to be calculated for loading on to the different nodes and
links. Based on these manipulations, the network nodes might also have to be selectively
configured to display difference in terms of size, color or shape etc. In addition, different
clustering algorithms might be applied to select clusters of nodes based on a common set
of criteria. Subsequently network analysis might have to be repeatedly performed till a
suitable network emerges allowing the studying of interactions of the various cas compo-
nents.
- 93 -
While the cas methods in this chapter have been applied in the domain of citation data,
the primary reason for this selection has been the stable and verified nature of this data. In
other words, the methods used for CNA are equally applicable to different types of cas data
ranging from social sciences (using survey instruments to develop relationship data etc.),
biological sciences (such as using biological genomic databases) and telecommunication
complex networks (using topological connectivity models of different nodes and peers in
the communication network).
3.7 Conclusions
In this chapter, we have demonstrated the use of the proposed complex network model-
ing level of the framework using two different case studies. The use of Complex Network
Analysis is shown to allow for the identification of emergent behavioral patterns in large
amounts of citation data. Using two separate case studies in the domain of Agent-based
computing and Consumer Electronics, the results demonstrate the effectiveness of the pro-
posed approach. Results included identifying the key patterns and Scientometric indicators
in research such as key authors, papers, Journals and institutions as well as the temporal
aspects underlying the evolution of the entire agent-based computing scientific domain and
sub-domains. In the next chapter, we present the exploratory agent-based modeling level of
the proposed framework i.e. exploratory agent-based modeling for performing feasibility
studies in multidisciplinary cas research.
- 94 -
4 Proposed Framework: Exploratory Agent-based
Modeling Level
In the previous chapter, we gave an overview of the development of complex network
models for cas based on the availability of real-world data by using two separate case stud-
ies. We discovered that by the use of complex networks, we could perform CNA of the
data and discover interesting trends in the data such as finding emergent patterns. While in
the ideal world, we would have all types of interaction data for cas, practically speaking
most cas case studies suffer from a lack of suitable data for developing complex network
models. As such, the rest of the framework modeling levels are concerned with the devel-
opment of agent-based models of cas. While agent-based models of cas are simulation
models, they do not fit in line with the traditional simulation models of engineered sys-
tems. As noted earlier in previous chapters, similar to the complex network modeling level
of the proposed framework, exploratory ABM level is also an encompassing level. In other
words, while on one hand, it is a part of the proposed framework, it allows other existing
studies of legacy ABMs to be tied in with the proposed framework. In the next section, we
develop a proposed methodology of performing exploratory agent-based modeling.
4.1 Research Methodology
Exploratory agent-based models are models which are developed by researchers to
explore the concepts underlying the cas under study. The goals of developing these models
are typically different from traditional computer simulation models such as involving
queuing, Monte Carlo methods etc. Some of the possible goals of building exploratory
agent-based models can be listed as follows:
- 95 -
1. The researchers want to explore if agent-based modeling is a suitable paradigm for
modeling their particular problem.
2. The researchers want to develop a proof of concept that any further research in the
chosen direction would be a viable possibility.
3. The researchers want to develop agent-based models as a means of demonstrating
the applicability of the chosen model design of a cas system to prospective funding
bodies.
4. The researchers want to develop agent-based models to explore exactly what kind
of data would be needed for the sake of validation of these models without spend-
ing considerable amount of time on these models.
As we can note here, these models are at times not meant to be comprehensive or ex-
pected to give high quality research results. However, these models do allow the
researchers to test the water and thus increase the possibility of proceeding on a larger-
scale future project.
The most basic type of agent-based modeling can be considered as the exploratory
agent-based modeling level as shown in Figure 26. This framework level is suitable for ex-
ploratory agent-based model design and development. While a number of researchers are
known to follow exploratory agent-based modeling, previously the concept has not been
formalized in a unified framework. The idea originated based on extensive observation of
published cas models by researchers who like to first test the water and develop proof of
concept models for further research. These models can be useful for assessing feasibility as
well as for demonstrating how future research might be conducted. At this stage, these
models do not have to be developed according to any detailed specification or even lead to
any interesting results. Their goal is often to pave the way for future research by develop-
ing a proof of concept model. This frees the researcher to explore the cas domain and
- 96 -
attempt to develop different models without any pressure. A large number of existing cas
models can be considered as exploratory and thus come under exploratory agent-based
modeling level of the proposed framework.
Figure 26 Exploratory agent-based modeling level of the proposed framework (to assess research fea-
sibility and proof-of-concept)
One possible way of evaluating the success of exploratory ABMs and their associated
problems is to examine the communication of various cas researchers. With the wide avail-
ability of open mailing lists for various agent-based modeling toolkits such as NetLogo,
Repast-S, Mason and Swarm, such an observational approach is quite possible. While
there are several mailing lists and tools, specifically we focus on NetLogo. There are sev-
- 97 -
eral reasons for selecting NetLogo. Among other things, Railsback et al. notes NetLogo as
the highest documented tool from the rest of the ABM tools[173]. They also highly rec-
ommend it for prototyping (exploratory) complex models as follows:
"NetLogo is the highest-level platform, providing a simple yet powerful programming
language, built-in graphical interfaces, and comprehensive documentation. It is designed
primarily for ABMs of mobile individuals with local interactions in a grid space, but not
necessarily clumsy for others. NetLogo is highly recommended, even for prototyping com-
plex models."
In this chapter, using a case study from the domain of complex communication net-
works, the application of exploratory agent-based modeling has been demonstrated. This
proposed real-world application is based on coupling a set of P2P unstructured network
search algorithms with household devices communicating with each other for intelligent
discovery of user content in an "Internet of things".
The rest of the chapter is structured as follows:
We first give an overview of the selected case study along with a problem definition.
Next, we present the design and implementation details of the exploratory agent-based
models. Subsequently we discuss the results of the simulation experiments. Finally we
conclude the chapter.
4.2 Case Study: Overview, Experimental design and Implementation
The chosen case study is the modeling and simulation of computing devices self-
organizing themselves to develop a complex communication network infrastructure. Tech-
nically this is called the "Internet of things" [174] and the idea was initially introduced in
1999 by Kevin Ashton[34]. The basic idea is given as follows in Kevin's own words:
"We need to empower computers with their own means of gathering infor-
mation, so they can see, hear and smell the world for themselves, in all its random
- 98 -
glory. RFID and sensor technology enable computers to observe, identify and un-
derstand the world-without the limitations of human-entered data."
The reason we are exploring this case study is that, to the best of our knowledge, there
is no previous case study exploring the use of agent-based modeling in the domain of "In-
ternet of things". Here, the goal is to firstly understand how agent-based modeling might
be able to perform exploratory studies for the Internet of things. Secondly, this case study
will also allow the use of extensive modeling and simulation demonstrating how explorato-
ry case studies might be effective in the domain of studying new types of cas.
While internet of things is a very interesting concept in theory, till now, there have not
been many practical applications. One possible reason of a lack of such applications is the
inability of current modeling and simulation tools to model and combine different types of
paradigms. As an example, as discussed in Chapter 2, while simulators are available to
model peer-to-peer networks such as Oversim[155] and there are other simulators such as
NS2 and OMNET++ suitable for the modeling and simulation of wireless sensor networks,
practically it can tend to be very difficult, if not impossible, to combine two or more differ-
ent paradigms in a single simulator without massive re-factoring of the simulator source
code. In other words, to the best of our knowledge, existing simulators do not effectively
allow the combination and application of protocols, ideas and entities from different com-
plex communication networks domain. In addition, what is even harder to model is
mobility because none of the previously well-known simulators handle mobility well. As
an example, in traditional simulators, the locations of nodes are typically defined in the
start of the simulation experiments. As such it is difficult to model the effects of Brownian
random motion and other mobility models.
As such, here the goal is to explore the use of agent-based modeling and simulation in
this domain and combine concepts from P2P networks along with the concepts of Wireless
- 99 -
Sensor Networks to develop a working simulation of the internet of things. The specific
problem that we attempt to solve using agent-based modeling and simulation is defined as
follows:
Existing computing media devices both have internet capabilities as well as file storage
capabilities for various types of contents and media. However, a user who keeps a set of
devices can have difficulty in locating e.g. a particular file from their various computing
devices (such as Smart phones, Laptops, and other devices). In other words, we want to
explore further how P2P search algorithms can be exploited for exploring these media de-
vices for locating user files. Here are a few possible use case scenarios for such algorithms:
A. In an ambient assisted living initiative, consumer mobile phones can communicate
with each other to discover internet access which then provides access to content
from customized web services.
B. Julie is late for her meeting, gets in her car and realizes that she has forgotten her
important files on one of her home computers. Using her mobile phone, she is able
to perform a P2P search of her personal devices which connect with each other se-
curely and she is able to access content from the computer in which she had left her
files. This autonomous self-adapting hybrid network does not require prior configu-
ration except any security measures that Julie has previously configured (such as
the use of digital certificates for authentication and authorization of the devices).
C. Joe is in a meeting and Rick, his boss wants him to disseminate the meeting agenda
to all the meeting participants on their local trusted phones. Joe's phone is able to
connect to other local phones without using the internet, which Rick rightfully con-
siders not secure enough for confidential company internal documents. As a
company policy, Rick and all participants are using mobile phones which are not al-
- 100 -
lowed internet access on company facilities. In other words, the devices form an in-
ternet of things to disseminate information.
D. In a media company, it has taken months to arrange a meeting for the busy ad de-
signers with their clients in a nice comfy mountain resort. They are all sitting in
various rooms discussing agendas. Mary has an important clip in her phone but the
size of the clip is very large for uploading and downloading from the internet. She
wants to share the clip with other designers. However, these other attendees are nei-
ther tech savvy enough, nor have high speed internet for fast sharing. She needs to
disseminate the clip inside the room as well as in another room. Her phone recog-
nizes other devices as she gives the phone the names of the trusted phones and she
is able to send it. People in the same room receive the clip using faster transfer over
local wireless network autonomously developed by the P2P applications while de-
vices in the other room receive the clip both via the internet of things by using
phones of meetings organizers in the hallways without requiring extensive configu-
ration settings or using the internet.
Other use cases showing the utility of P2P algorithms coupled with wireless sensor
networks in the domain of internet of things can be observed in Figure 27. As can be seen,
applications can range from real-time weather monitoring, patient monitoring, business
meetings, currency rates, condition monitoring of all of the family cars etc.
- 101 -
Figure 27: Examples of Internet of Things
In Figure 28, we can note some of the challenges in the design of hybrid networks
which involve local as well as remote content access. Content can be accessed using two
primary models. One is the Push model where content is automatically accessed on the de-
vices by being "pushed" form other devices while in the "pull" model, the content is
accessed by performing queries. P2P access is based on unstructured and structured over-
lay methods. In structured overlays, there is typically a Distributed Hash Table (DHT)
which allows for indexing of content to a certain degree whereas unstructured overlay net-
works involve the use of algorithms which can access content without a structured
network.
- 102 -
Consumer net-
work content
Content Retrieval
(Pull)
Content or in-
formation
Dissemination (Push)
From local net-
Retrieval Via the
To local net-
Via the web
work
web
work
To a peer on the
web
To a peer in anoth-
er consumer network
Unstructured or struc-
tured overlay network
From a P2P
Shared network
Unstructured or struc-
tured overlay network
Unstructured
overlay
Structured over-
lay
From another con-
sumer network
Figure 28: Challenges in hybrid network design for Internet of things by using unstructured P2P algorithms
for searching content in personal computing devices
4.2.1 Goals definition
The key goal in this exploratory agent-based modeling case study is to develop an ex-
ploratory model for studying the feasibility of the proposed design of hybrid complex
communication networks. By means of proposing an algorithm Self-Advertising Content
Sources (SACS) which is similar to dropping bread crumbs around a house for locating
content sources, each content source sends a message with a certain numeric value in its
immediate vicinity. This gradient establishment process is demonstrated in Figure 29. Here
S represents the content source. This could be a personal computer, a mobile phone or
some other device. The other square tiles represent other computing devices located near-
by. Now, how it can be physically implemented is by means of having S first locate other
wireless communication devices within its immediate access based on its communication
- 103 -
radius. These devices are then sent the gradient establishment message. Once the message
reaches these devices, being the first in the vicinity, these devices keep the counter of the
message and are shown by a 1 in the figure. These devices next search for other devices
and send the messages but with a lower SACS value. This process continues till the SACS
distance limit expires for the message in which case, it is not re-sent or else there is no oth-
er communication device within the communication radius.
Figure 29 Establishment of a gradient by SACS sources
Subsequently after the establishment of a gradient, devices can query other devices in
the P2P Internet of thing network. This is possible with the help of a TTL based algorithm
as depicted in Figure 30. As can be seen, to avoid unnecessary flooding of the surround-
ings, the queries keep moving forward based on the gradient value. So, essentially the
queries are able to locate content sources because they always move to a computing device
with a lower gradient value.
- 104 -
Figure 30 Query algorithm flowchart
4.2.2 Agent design
As discussed earlier in the previous chapters, Agent-based models have been used in a
large variety of application domains. While some application domains involving social
simulations, such Agents in Computational Economics (ACE) depict reproducibility, eco-
nomic benefits and social interests, these descriptions are not inherent to agent-based
design per se and are domain specific. As such, here we give detailed descriptions of agent-
based models using various detailed design diagrams and implementation details in the
complex adaptive communication networks domain using a case study of the Internet of
things.
- 105 -
The agents in this simulation exploration are of two primary types:
1. Computing devices.
2. Message agents which are of two types
a. Query Message agents.
b. SACS setup message agents.
The detailed design description of these agents is given below:
4.2.2.1 Computing device agents
These agents are designed to simulate various types of computing devices. The devices
can range from mobile phones, handheld games as well as laptop computers to stationary
computers. As such, there are the basic design requirements for these agents that can be
listed as follows:
A. The agents must have a certain amount of memory to hold certain information.
B. In addition, these agents must be able to communicate with other agents in a certain
communication radius.
C. The agents must be able to perform basic computations (such as retrieval of data
from storage or using data structures).
The state chart diagram describes the Computing device agents. It is pertinent to note
here that some of the computing device agents need to be considered as gateway agents.
The idea of gateway agents is that in case the internet of things based algorithm (SACS) is
unable to locate certain content or information locally, then it should be possible for mes-
sage agents to find specific devices which are configured as gateway devices. These
computing devices are essentially useful for allowing the local P2P network to extend its
reach to the larger set of devices. As an example a person can have multiple devices at
home as well as office. If the person requires a certain file, the search query can execute
first locally in the vicinity of the house and if the query cannot locate the file, the search
- 106 -
messages need to be routed on through a gateway node to the office gateway machine
which will look for the required content/file in the office devices.
Figure 31 State chart for Computing device agent
The second type of computing device agents is the SACS source agents. The key idea is
that each content source self-organizes its surroundings by dropping bread-crumbs for en-
ticing query agents. The way it actually would be implemented in physical devices can be
described as follows:
1. First the device sends a message within its communication radius. This message is
basically a Hello message for finding out what is the list of active neighbor compu-
ting devices within communication range.
2. Next, these devices respond back to the sender giving their IDs.
3. The IDs of the neighbor devices are stored by the sender computing device.
4. Next, the sender device broadcasts the SACS setup message agent after assigning a
TTL value to it. The reason for first finding the neighbor devices instead of an ini-
- 107 -
tial broadcast is that the device must be sure that there are any active listening
computing devices before it repeatedly sends these messages.
5. On receiving the message agent, the neighbor devices can subsequently forward the
message based on the policy of the message agent, which will be described below.
Next, we discuss the design of the Message agents.
4.2.2.2 Message Agent
Here we discuss the design of the message agents. Message agents are of two types. The
first type is the SACS setup message agents and the second type is the query agents. Un-
like the computing device agent, it has been primarily designed for forwarding the message
agents, message agents themselves need to act more intelligently. Next, a description of
both these types of message agents is given as follows:
1. SACS setup message agents are responsible for setting up a gradient in the vicinity
of the SACS source nodes. As described above, these agents help establish a gradi-
ent starting from the initial communication radius of the SACS source node agents.
SACS setup agents have a simple task to perform. Once they are on a specific de-
vice, they use the algorithm described in Figure 30 to hop from one device to the
next. This is done by means of counting the distance (radius) by means of a hop
count. As a result when the hop value reaches the end of the SACS radius, it im-
plies the end of the line for the SACS setup process. However, during each hop, the
agents also communicate their respective hop count value to the computing device
agent, on which they are currently located. As such, the computing device agent
stores this information locally for future queries. It is important to note here that
SACS gradient can be established from a smaller radius to a large radius, where the
radius here would be used in terms of communication hop counts. In case of small-
er radii, there are chances that subsequently when queries are looking for a
- 108 -
particular content source, they might spend a long time hopping randomly before
they can locate a particular SACS gradient "scent" i.e. the hop value bread-crumbs
left over by the SACS setup message agent in the initial phase.
2. The second type of message agents is the Query agent type. The goals of the query
message agents are different from the SACS setup message agents. The query
agents are initiated by a user logged on to one of the computing devices. As such,
the goal of the query agent is to look for information. Now, the interesting thing to
note here is that according to the small world complex network theory, a large
number of real-world networks have nodes which can locate most other nodes us-
ing a small number of hops. As such, when an owner of a set of devices is looking
for specific information, chances are that the query would be able to locate this
content on one of the computers locally or within a few hops away from the person.
Query agents are thus responsible for content location discovery based on the
"smell" of the content discovered by means of the bread-crumb values dropped by
the previous SACS setup messages during the initial setup phase.
4.2.3 Learning and adaptation in the system
As discussed in detail in the last two chapters, unlike traditional agents involved in mul-
tiagent systems, the agents in the simulation design here do not individually have
intelligence and adaptation capabilities. However, the system as a whole has significant
adaptation capabilities. Thus the nonlinear interactions of the different agents allow the en-
tire system to emanate emergent behavior. The emergent behavior in this case is the
discovery of content using the bread-crumbs in content vicinity spread earlier during the
course of the execution of the distributed SACS setup algorithm. Here it can be noted that
the bread crumbs are established here in a completely self-organized manner. Thus every
particular assembly of computing devices will establish a completely new gradient and
- 109 -
each case of randomly walking queries will behave differently. However irrespective of
any variation in the configuration SACS algorithms need to be to be able to locate the re-
quired content sources.
4.2.4
Implementation description
Here, we describe the implementation details in the NetLogo tool. The model has been
implemented in a modular fashion and is described as follows:
4.2.4.1 Global variables
Firstly, the global variables depict some of the metrics which need to be collected over
the due course for the simulation experiments. There are three key desired global output
variables here:
1. The number of successful queries.
2. The total number of queries.
3. The total communication cost
In addition, there are several other input variables from the user interface which can be
adjusted by the user either manually or else by using behavior space experiments as shown
in Figure 32.
4.2.4.2 Breeds
The "aggregation" basic of cas, as discussed earlier in chapter 2, can be noted here in
this model. With the help of using two different types of agent breeds developed in the
netlogo model, aggregation can be noted. These breeds can be described as follows:
1. Nodes breed
This breed depicts the computing devices in the simulation. It has its own set of
internal variables. These variables are of two types. The first ones are Boolean vari-
ables which determine the type of the node. These include whether or not, it is a
gateway node, a query node, a goal node (i.e. a content source) and whether or not
- 110 -
it has been previously explored till now. The other type of internal variable for the
node is the SACS distance. This is the variable which is updated based on the hop
distance i.e. gradient from the SACS nodes. SACS nodes are the nodes with the
value of the Boolean variable goal? being true.
Figure 32: User interface showing various configurable values for the simulation
2. The second breed is for the messaging agents. These have been implemented as the
"Query" breed. These agents have three internal variables for storing their state.
Firstly, they have a variable which aggregates the current node agent, on which
they are located. The third variable inside the query agents gives the concept of se-
lective lookup by noting the particular type of content which this agent is searching
for. As can be noted, both of these concepts represent the cas internal model basic
concept as discussed earlier in Chapter 2.
- 111 -
After a discussion of the breeds developed in the agents, we next move to the imple-
mentation of the various procedures:
4.2.4.3 Setup
The first function is the "setup" and allows the creation of the entire scenario for further
simulation of the model. The overall setup function is shown in following Figure 33.
Figure 33: Description of the setup function
Here, we can note that the first part of the setup function is to clear all variables, agents
and the environment of the simulation. This is important especially when multiple simula-
tions are to be executed with the passage of time. The second part of the setup is related to
making the computing agents. This function is related to constructing the computing de-
vice agents in the simulation. After the creation of the computing devices, the gateway
nodes are next selected based on the input parameters. Finally, these gateway nodes adjust
- 112 -
in terms of their locations in the Simulation world. The detailed description of each of
these functions follows along with other functions.
4.2.4.4 Make Computing Agents
This function is essentially the base for developing the computing agents according to
the input variables. This procedure is based on invoking the patches in the NetLogo world.
First of all, each patch creates a random number between 0 and 100. Next, this number is
compared with the probability assigned via a global input variable. Based on a comparison
of these two numbers, the patch might sprout an agent at this location. Next, the agent is
initialized with certain values. Initially, the agent is given an unexplored status. The sacs
distance is assigned equal to the sacs radius input global variable. Initially all nodes are
given "false" as the Boolean value for both the start as well as the goal and the gateway
variables. The shapes of the computing devices are next adjusted to be squares and they are
slightly randomly moved to ensure that most devices will not cover up over previous de-
vices.
4.2.4.5 Make Gateway nodes
This function is concerned with creating the gateway nodes from the previously created
computing device node agents. The working is based on a random selection of agents from
these devices. The number of agents which are to be created as the gateway nodes is based
on a global input variable. After making the node as gateway, its color is changed to yellow
so that it is visible in the overall landscape.
4.2.4.6 Adjust locations
This function asks each gateway node agent to perform certain tasks. First of all, the
agent looks in its immediate location and notes if there are any other agents there. If it
finds other agents located at the exact spot, the gateway node agent starts adjusting its loca-
- 113 -
tion slightly. It executes a conditional loop and keeps taking a random turn and moving
very slight till it is not immediately over another computing node. This exercise is im-
portant for both visualization as well as verification of the correct behavior in the case of
execution of large scale simulations. The effects of the setup give a screen view as shown
in Figure 34.
Figure 34: Effects of the execution of the setup function with 2500 computing devices
Note here that the square computing device agents have self-organized themselves
based on simple rules resulting in a more realistic pattern. In addition, all gateway nodes
are now completely visible on the screen.
4.2.4.7 Searches
The "Searches" function has primarily got the task to repeatedly call make-search func-
tion. The number of times this function is called is again configurable from the User
- 114 -
interface. This is a global variable slide which takes input from the user in terms of manual
execution for the number of searches to be created randomly.
4.2.4.8 Make Search
"Make search" function has two basic tasks. Firstly it randomly selects a computing de-
vice node agent. However, this selection is based on the criteria that the selected random
node agent must not previously be a gateway node. It then calls the make-s-node function.
4.2.4.9 Make-s-Node
This function is the actual function which creates the search node. The first thing that
this function does is to change the "start?" Boolean variable to true. It thus tags the node as
the location of the starting point of the queries. It also changes the color of the node to blue
and its shape to a circle so that these nodes are visually recognizable. The location of the
computing node is then stored in a temporary value. Afterwards, it executes the hatching of
k query agents. These query agents are needed for the "k-random walker" algorithm from
the domain of unstructured P2P networks. For each of the query agents, it calls a setup-
query function.
4.2.4.10 Setup-query
This function is called from the Make-s-node function. The goal of this function is to
properly assign values to the query agents. After making the query agents a circle and as-
signing them green color, the function increments the total query count global variable.
The query variable "loc" is next assigned an agentset made up of the agent where it is cur-
rently located.
- 115 -
4.2.4.11 Goals
The "Goals" function repeatedly executes the Make-goal function. Here the n-cs global
variable essentially reflects the total number of content sources which is another global in-
put variable.
4.2.4.12 Make-Goal
This function has got two main tasks. First, it selects nodes which are neither gateways,
nor start nodes for queries. Secondly, it asks each of these nodes to execute the make-g-
node function.
4.2.4.13 Make-g-Node
This function is called only to have a centralized location for changing the attributes to
each goal node i.e. a content source. Unlike most other functions in the simulation, this
function is quite simple and only sets the goal Boolean variable to true in addition to
changing the color of the node to red.
After calling setup, Searches and Goals functions, the screen now looks as shown in
Figure 35.
4.2.4.14 Setup-sacs
"Setup-sacs" is the key function in the SACS setup algorithm. It is used to firstly locate
each of the goal nodes. Once all the nodes have been located, they are all asked to execute
setup-sacs-d function in the turtle context.
- 116 -
Figure 35: View of the screen after the execution of the setup, Searches and the Goals functions
4.2.4.15 Setup-sacs-d
"Setup-sacs-d" simulates the entire SACS setup algorithm. It is also a recursive algo-
rithm but unlike traditional recursive algorithms, it is an agent-based distributed recursive
algorithm. In other words, it executes on other agents. So, it essentially calls itself repeat-
edly till the SACS gradient is properly setup in the communication radius of each content
source. Like all recursive algorithms, it needs to have a stopping condition. The stopping
condition here is a complex stopping condition. It can be considered as a disjunction of two
separate conditions. Firstly the algorithm will stop if there are no other "unexplored" com-
puting nodes in its vicinity. Secondly, it will also stop if the calling argument "d" is greater
than the global variable "sacs-radius". The "sacs-radius" global input variable is again con-
figurable by means of a slider. The way this algorithm works is that first of all it equates
the sacs-distance value to the minimum one as compared to the previous value. It then
changes the label of the current computing node agent to reflect this sacs-distance value. It
- 117 -
also makes the explored? Boolean variable true. If the sacs-distance is non-zero, it changes
the color of the node to grey. This way, the content sources are the only ones which will
visually stand out from the other nodes. After setting these basic attributes, it compares the
current argument "d" with the sacs-radius global variable. If d is less than or equal to this
value, it creates a new agentset. This agentset is formed of all other nodes in a certain
communication radius but based on a condition of being unexplored till now. The commu-
nication radius is again a configurable value "sens-radius". Now, this agentset is not
guaranteed to be non-empty so it is tested for emptiness. If non-empty, each of these agents
is asked to execute the same function again but with an incremented "d" value. Thus this
process can continue till the SACS gradient is properly setup around all SACS content
sources reflecting how the content sources can be Self-advertising their contents.
Afterwards, the simulation screen can be observed to reflect this setup as shown in Fig-
ure 36.
Figure 36: Result of execution of SACS setup algorithm
- 118 -
4.2.4.16 Move
This is the main function for the execution of the algorithm for queries as shown in Fig-
ure 37. The function itself is focused purely on the execution based on the existence of
query agents. If any queries are there, it will continue by asking them to execute move-rw
function otherwise it will terminate.
Figure 37 Flowchart for move function.
4.2.4.17 Move-rw
This is the main function for the execution of the algorithm for queries as shown in Fig-
ure 38. Here, we can note that the function executes by first creating a list of all nodes in
the given sensing radius. This radius is basically equal in physical terms to the communica-
tion radius. As an example, in case of Bluetooth networks, it will be different as compared
- 119 -
to Wi-Fi networks. In case there are no other devices in range, the query agent will simply
die. If there are other devices, the next step is choosing one of the nodes as the next loca-
tion. This is performed differently based on what is the selected mode of movement of
queries from the user interface. Based on the Boolean variable "sacs?" the next node is ei-
ther a random node in the case of a false value for this variable, or else one of the nodes
with minimum of "sacs-dist" variable. Note that this represents the gradient previously es-
tablished by the SACS nodes. As such, once this node is selected, the query agent can
firstly move to the selected node. Next, it can update its internal variables such as TTL val-
ue by decrementing it and also storing the tagged location node inside the "loc" variable.
After moving and updating these values, it needs to calculate the cost associated with this
move. If the location is a gateway node, then the move will incur cost equal to "gw-cost", a
user input configurable variable otherwise the cost will simply be incremented by 1. Final-
ly before this function terminates, a call is made to the checkgoal function, which is
explained next.
4.2.4.18 Check-Goal
This function is primarily for goal verification for queries. In this function, the query
checks whether it has reached either one of a gateway node or else a content source (goal
node). If the current location agent satisfies either of these conditions, then the overall
number of successful queries is incremented and then the query agent terminates itself. If it
has not reached the goal nodes, then again it checks its TTL value. In case, the TTL value
has reached zero, then again it dies.
- 120 -
4.2.4.19 Move-Devices
This function is to test the effects of mobility of devices on sacs. It randomly moves a
percentage of the devices over time and when devices are out of range, it calls setup-sacs
again to reset the sacs setup messages.
Figure 38: Flow chart for the move-rw function
4.3 Simulation Results and discussion
The goal here is to use agent-based modeling to develop working proof of concept
models for modeling these concepts. Thus while small scale computing devices can be
hand-simulated and might appear to be a trivial exercise which can be performed on a
piece of paper, the specific goal of using agent-based modeling here is to be able to exam-
ine how this concept will work in the case of say hundreds of computing devices in say, a
small community or a small town. The proof of concept models will allow us to examine
the results based on hundreds of simulation runs. Thus it will allow us to discover the
strengths and weaknesses, if any, of the SACS mechanism. By using extensive parameter
sweeping, the simulation experiments should give a strong case for developing further
- 121 -
models of SACS or else to abandon the concept if it does not scale for larger and more re-
alistic settings of large-scale communication networks.
4.3.1 Metrics used
The input parameters and the outcomes of the simulation will be measured using the
key metrics given as follows:
4.3.1.1 Simulation parameters:
The input metrics to the simulation are as follows:
1. The number of computing devices used in the simulation
2. The number of content sources ncs
3. The number of gateway nodes ngw
4. The number of query source nodes nsrc
5. Number of random walkers per source node krw
6. Cost in terms of hop counts for using a gateway node for locating a content
source: cgw
4.3.1.2 Output metrics:
The key metrics in a P2P network are based on the number of lost queries as compared
to the number of successful queries. The output parameters that are of interest in the evalu-
ation of SACS are as follows:
1. Number of successful queries Stot
2. Total cost of execution Ctot
3. Total number of queries Ntot
However, here the total number of queries will be constant and instead the cost of exe-
cution as well as the number of successful queries will be evaluated in depth.
- 122 -
4.3.2 Details of experimentation
The simulation execution is based on the execution of four separate Behavior Space ex-
periments. The details of the experiment parameters and the reasoning are given below:
The Behavior space parameters for the experiments are given as follows in Table 12. We
can note that some parameters are the same across all simulations while two key parame-
ters are varied i.e. max-ttl and the sacs-radius parameters. Each of these variables in the
simulations starts with a certain value. Subsequently these variables are incremented by the
next value listed up until it reaches the third listed value as shown in the table. In addition,
the last two experiments are conducted to evaluate the mobility of computing node devices.
Table 12 Simulation parameters
Params
Exp I
Exp II
Exp III
Exp IV
gw-cost
n-srchs
sens-radius
n-gw
n-cs
max-ttl
10
100
2
10
10
10
100
2
10
10
10
100
2
10
10
10
100
2
10
10
10
[5 5 20]
10
[5 5 20]
sacs-radius
[0 5 20]
[0 5 20]
[0 5 20]
[0 5 20]
k
5
5
5
5
Static
Static
Mobility Mobility
The descriptive statistics of the first experiment are given in Table 13. We can note here
that the total simulations were 250 with each simulation repeated 50 times. The successful
queries, total queries and the hop counts are also given.
- 123 -
Table 13 SACS varying with constant TTL
Descriptive Statistics
runnum
max-ttl
sacs-radius
nsucc
ntot
nhop
Valid N (listwise)
N
Minimum
Maximum
Mean
Std. Deviation
250
250
250
250
250
250
250
1
10
0
16
500
3194
250
125.50
10
20
399
500
10.00
10.00
219.51
500.00
5055
3970.77
72.313
.000
7.085
108.612
.000
559.595
For the next experiment, the descriptive statistics are given in Table 14. Here the total
set of simulation runs was 1000. The max-ttl as well as the sacs radius were both varied
periodically in the course of the experiments.
Table 14 SACS and TTL varying
Descriptive Statistics
N
Minimum
Maximum
Mean
Std. Deviation
1000
1000
1000
1000
1000
1000
1000
1
5
0
7
500
2125
1000
500.50
288.819
20
20
455
500
9885
12.50
10.00
217.40
500.00
5.593
7.075
119.092
.000
4541.87
1902.447
Runnum
max-ttl
sacs-radius
Nsucc
Ntot
Nhop
Valid N (listwise)
The next two experiments were concerned with mobility of computing devices. The de-
scriptive statistics are given in Table 15. The total simulations in this case were 250. The
max-ttl was made constant and only the sacs-radius was varied.
- 124 -
Table 15 SACS vary with constant TTL (mobility)
Descriptive Statistics
runnum
max-ttl
sacs-radius
nsucc
Ntot
Nhop
Valid N (listwise)
N
Minimum
Maximum
Mean
Std. Deviation
250
250
250
250
250
250
250
1
10
0
8
500
3369
250
125.50
10
20
349
500
10.00
10.00
194.64
500.00
5076
4113.39
72.313
.000
7.085
88.675
.000
479.169
The final set of experiments was conducted again to test mobility. This was similar to
experiment 2 since it was based on varying both the max-ttl value as well as the sacs-radius
values periodically. The descriptive statistics here are given in Table 16.
Table 16 Experiment with TTL and SACS varying (mobility)
runnum
max-ttl
sacs-radius
nsucc
Ntot
Nhop
Valid N (listwise)
Descriptive Statistics
N
Minimum
Maximum
Mean
Std. Deviation
1000
1000
1000
1000
1000
1000
1000
1
5
0
6
500
1956
1000
500.50
288.819
20
20
430
500
9883
12.50
10.00
208.20
500.00
5.593
7.075
111.068
.000
4690.41
1916.672
4.3.2.1 Discussion of results
The first set of results is the details of variation of sacs-radius and its effects on the suc-
cessful queries. Here we can note in Figure 39 that there are five values for the sacs-radius
ranging from 0 to 20. Here the value 0 implies that the SACS is not used whereas on the y-
axis, the number of successful queries are being listed. As can be noted here from the earli-
er descriptive statistics, each point represents a set of 50 repeated experiments and as such
gives a pretty good realistic example of the outcome and cannot be considered as a mere
chance. We note here that very few successful queries exist in the case of not using sacs
- 125 -
(SACS=0). However, as we make a SACS radius value of 5, the number of successful que-
ries almost increases 5 times. The increase however is not uniform with a further increase
in SACS radius. Thus, the biggest jump in successful queries is for a small SACS radius.
This is quite an interesting result because it implies that with a minimal use of SACS, there
is a significant improvement in the success rate for searchers.
Figure 39: 95% Confidence interval error plot for successful queries plotted against the increasing
sacs-radius
The next figure is based on the effects of increase of the SACS-radius on the overall
cost of execution of the queries in terms of hop counts as shown in Figure 40. The results
here are very similar to the successful queries in the previous figure. We notice that the
cost is very large in the case of not using SACS close to 5000 hops. By merely turning on
the SACS by giving sacs-radius of 5, the cost significantly decreases down to almost 3950
- 126 -
hops. There is a further decrease in cost when sacs-radius is made 10. However the jump is
not that significant. This is even further noticeable when the sacs-radius is further in-
creased. Thus it corroborates with the small-world phenomenon that just by using a small
sacs-radius of 5, queries can easily locate the content sources.
Figure 40: 95% Confidence interval plot for hop count (cost) vs an increase of the SACS-radius
In the previous figures, we note the first exploratory hypothesis about the effects of
changing the sacs-radius on both the overall cost as well as the number of successful que-
ries. The goal of the next plot as shown in Figure 41 is to test another hypothesis. This
hypothesis was based on the TTL value of the queries. So, the big question to look for here
was what if it was the TTL value of the queries that was actually resulting in the success of
- 127 -
SACS. Also one more point to examine is how this effect would be coupled with the
change in the sacs-radius. So, we can observer this effect using a scatter plot. In a scatter
plot, different values of the simulation run can be noted. Here, the y-axis is labeled in
terms of cost or hop count whereas the x-axis is giving details in terms of the TTL values
of the queries. On the other hand, the data is color coded based on the sacs radius. Here, we
can clearly see that the cost of queries without the use of SACS is always large. The differ-
ence for smaller TTL values is not as pronounced as larger TTL values. As a matter of fact,
the higher the TTL values, the more pronounced is the effect of a better cost for all sacs-
radius values. And from max-ttl value of 10 onwards, the difference in all experiments is
very clear in terms of significant reduction of hop counts/cost.
- 128 -
Figure 41: Scatter graph showing the variation of the cost in terms of hop count with a change in max-
ttl and sacs-radius
For verification of this hypothesis in terms of successful queries, we plot the next graph
in Figure 42 using mean successful queries on the y-axis color coded according to different
sacs-radius values. The x-axis depicts the max-ttl values. Now, we can note here that for
max-ttl = 5, sacs-15 appears to have the best possible successful query values. However,
the numbers of successful queries for k-random walk without SACS remains constantly
and significantly lower than any of the SACS radius simulation experiments. Furthermore,
we note that sacs-radius = 5 reflects the most benefit as it is compared with the implemen-
tation model not executing SACS. However, the differences between all sacs values are
initially very small for max-ttl of 5. Subsequently however, for an increase in sacs-radius
values, there is a corresponding increase in the successful queries. This effect is however,
- 129 -
not as pronounced as the difference between non-SACS and sacs-radius = 5 successful
queries.
Figure 42: Mean successful queries with a variation in max-ttl value and a variation in colors based on
different sacs-radius values
Up until this point, several different hypotheses have been tested and it has been discov-
ered that SACS indeed can be a better alterative as compared to k-random walk algorithm
from the domain of unstructured P2P networks. However, it is important to note here that
till now, the scenarios are for static nodes. While this is good in theoretical terms, this is
not a reflection of a realistic physical environment. In a realistic physical environment, de-
vices are not stationary. Rather, they tend to move in a Brownian random motion as noted
by Groenevelt et al. in [175]. Gonzalez et al. [176] note similar patterns in human mobility
- 130 -
models. As such, the next set of experiments is concerned with mobility modeling of the
computing device agents. The idea here is to evaluate the results of using SACS algorithm
in the case of mobility modeled as Brownian random motion.
Figure 43: 95% confidence interval plot showing the effects of increasing sacs-radius on the number of
successful queries in the case of node mobility
The first plot using mobility is for the evaluation of the effects of changing the sacs-
radius by means of measuring the number of successful queries. Here, in Figure 43, we
note that with different sacs-radius values, the results are still much better than the k-
random walk reference algorithm. However, we also note an interesting observation that
the value of successful queries increases from sacs-radius of 5 to sacs-radius of 10 but ta-
pers off at sacs-radius of 15. And furthermore sacs-radius of 20 actually starts to reduce the
number of successful queries. In other words, with the mobility problem, because of per-
- 131 -
haps the motion of the intermediate nodes over time, some of the queries tend to get lost if
there is a very large sacs-radius.
Figure 44: 95% confidence interval plot for the cost in terms of hop count with a variation of the sacs-
radius value in mobility of computing devices scenario
While we have noted the effects of the successful queries, next we want to evaluate the
effects of sacs-radius on the total cost of the search in terms of hop counts. Here, we can
again note the significant jump in cost from no SACS algorithm to sacs-radius = 5. Now,
interestingly unlike the previous plot, after sacs-radius is varied to 10, the value of cost ta-
pers off a bit and actually increases from sacs-radius of 10 to sacs-radius of 15 and
eventually sacs-radius of 20. Thus, we note here the negative effects of the mobility on the
SACS algorithm resulting in a deterioration of results as nodes move about randomly.
- 132 -
Figure 45: Scatter plot of hop count cost with a variation in max-ttl value color coded according to
sacs-radius in mobility modeling
In the next set of plots, we want to note the bigger picture so we use scatter plot and
note the cost as it varies with a change in the max-ttl value and the sacs-radius as shown in
Figure 45. In this case, we note again that the cost associated with the algorithm in case of
mobility is always better than the cases of not using SACS. However, the cost difference
shows that increasing a max-ttl value does not have much impact in terms of differences
between different sacs-radius values. One important point to note here is that initially for
sacs-radius = 5, we note the lowest cost but as max-ttl increases, the cost of sacs-radius = 5
tends to become the highest and the lowest cost being that of sacs-radius = 15 at max-ttl
value=20.
- 133 -
Figure 46: Average values of successful queries with respect to a variation in max-ttl value color coded
according to sacs-radius in the case of using mobility modeling
In the next plot, we test this hypothesis in the case of successful queries. Here for vali-
dation, instead of using a scatter plot, we use a mean value plot. In this case, we note again
the significant difference between non-SACS algorithm and different SACS runs. Now, in
terms of successful queries, initially, sacs-radius of 10 has the largest number of successful
queries. However sacs-radius of 15 wins when max-ttl value increases to 10 and keeps
winning till the value of 20. As such, we can note here that mobility models have interest-
ing effects on the SACS algorithm but SACS performs quite strongly even in this case.
- 134 -
4.3.3 Discussion of Intelligence in cas Agents
While certain types of heuristics and decision support are added inside "intelligent or ra-
tional" agents, the goal of a cas simulation is different. Intelligent engineered autonomous
agents can use game theory to make choices, execute evolutionary algorithms to find best
chromosomes with the highest fitness values and execute Monte Carlo simulations to find
the best possible outcomes based on knowledge of input distributions. The idea of cas
simulation is based on using simple but numerous agents in a system similar to life.
As discussed earlier in previous chapters, in living systems, individual cells behave in
very simple ways. The total interaction of most cells involves simple operations such as
sensing chemicals, emitting chemical molecules, using chemical molecules and cellular
motion. However, based on these few simple rules and interactions, large and complex
multi-cellular organisms are formed. In terms of learning and adaptation, unlike a typical
MAS, where a single agent learns from either supervised, unsupervised or reinforcement
learning schemes, agents in agent-based models have very simple models of internal adap-
tation. However, the real adaptation is collective. As such, while cells in the body have
very simple states, the global cas in the form of e.g. the body of a mammal adapts collec-
tively to different events in the environment. An example is that of blood clotting. While
each particular cell always performs the same actions that it was designed for, the body as
a whole adapts if there is say a wound or a bruise. Different cells are created or guided by
chemicals to a particular spot by the body and result in clot formation. A detailed descrip-
tion of the differences between these two types of agents was given earlier in Chapter 3 in
the Scientometric Analysis of Agent-based computing domain. In this chapter, we propose
the concept of exploratory agent-based modeling level of the proposed unified framework
for the modeling and simulation of cas.
- 135 -
4.4 Conclusions
In this chapter, we have given details of how a completely new cas domain applica-
tion can be modeled using exploratory agent-based modeling. We have used an application
which, to the best of our knowledge, cannot otherwise be modeled using any other simula-
tor easily because this particular application is from the "Internet of Things" domain.
Modeling Internet of things requires a flexible modeling paradigm since it uses concepts
from the domain of P2P networks, ambient intelligence, pervasive computing and wireless
sensor networks.
Our exploratory agent-based modeling of the SACS algorithm demonstrates clearly how
different cas researchers can use agent-based modeling to perform extensive simulation
experiments using parameter sweeping to come up with comprehensive hypotheses. Our
results demonstrated the effective testing and validation of various exploratory hypotheses.
We demonstrated how one hypothesis can lead to the next and how simulation experiments
can be designed to test these hypotheses. In the next chapter, our goal will be to further ex-
ploit agent-based modeling by demonstrating how ABMs can be described by means of a
pseudocode based specification coupled with a complex network model and Centrality
"footprint" to allow for a comparison of ABMs across scientific disciplines.
- 136 -
5 Proposed Framework: Descriptive Agent-based
Modeling Level
In the previous chapter we have demonstrated how agent-based modeling can be used to
develop exploratory models of cas. The key idea in exploratory modeling was to use agent-
based models to supplement research studies with simulation models with the goal of eval-
uating feasibility of further research.
An examination of the interchange of messages between researchers associated with
the mailing lists of agent-based modelers can be considered to offer a cross-sectional view
of the cognitive modeling problems faced by various cas researchers. It is pertinent to note
here that the number of NetLogo-users mailing list[177] subscribers amount close to 3, 800
subscribers and the list contains almost 13, 000 archived messages to date. Here, we can
develop a closer examination of the cognitive problems faced by researchers since it
demonstrates that while helpful, exploratory modeling approach intrinsically lacks key fea-
tures which are of importance to cas researchers. Some of these missing features result in
problems as can be noted as follows:
A. Firstly, exploratory agent-based models tend to have a narrow focus. While multi-
disciplinary cas researchers are interested in exploiting modeling constructs from
one scientific research domain for use in another, there is no easy way to transfer
knowledge related to model concepts in this manner. With models developed pri-
marily for a single specific scientific application, researchers find it difficult to
transfer knowledge from one exploratory ABM to another cas application domain
even though apparently there are inherent similarities between the models. To the
best of our knowledge, the only two means of transferring constructs is to either re-
- 137 -
use source code snippets or else replicate models based on textual descriptions
(Both approaches shall be discussed critically later in this chapter). So, a model built
for say, tumor growth might have useful constructs which could be used in a com-
pletely different scientific domain such as data aggregation in Wireless Sensor
Networks. However, even with this intuitive similarity, there is no easy translation
of model constructs from one domain to the other using a non-code but high fidelity
description of the model allowing for an almost 1-1 correspondence of model con-
structs between the specification and the ABM.
B. Another set of modeling problems can be observed in the case of model comparison
of different case studies and application domains. Again, to the best of our
knowledge, there is currently no mechanism for model comparison across cas do-
mains and applications. As an example, Social Scientists have well-developed
agent-based models and concepts for gossiping [20]. Intuitively, this could perhaps
be very similar to how a forest fire spreads in a forest. In other words, the way fire
moves from a single tree or a group of trees to its vicinity appears similar to gossip
propagation models in communities and cultures. So, it is inherently interesting to
be able to compare these models. However, once again, there is no well-known
standard or formal means of comparison of these abstract concepts gleaned from
Agent-based Models across domains. Currently, typical descriptions available for
models are textual in nature. Textual descriptions offer a limited ability to perform
comparisons. If such comparisons are performed, they are qualitative in nature and
not quantitative. As such, having the ability to compare different agent-based mod-
els quantitatively could offer deeper and objective insights into different types of
agent-based models without either requiring the model to be implemented or else
executed for collecting simulation data.
- 138 -
C. A third setback of exploratory ABMs is the difficulties in teaching model constructs.
While there is notable community interest in this domain as can be observed by the
close to 800 educators enrolled in the NetLogo-Educators mailing list[178], the only
current well-known option of teaching ABMs is via code tutorials. However, these
tutorials are often incomplete, relate to a specific application domain only or else
require learning a specific set of Language Libraries to develop simple ABMs.
As a means of exploring these problems further, we evaluated random snapshots of top-
ics and message data in a typical fortnight of "netlogo-users" mailing list. Results of the
experimentation shown in Table 17, demonstrate that almost 77% of topics and 85% of
messages are on "how to model".
On a further semantic examination of the message contents, a key problem can be clear-
ly identified that most cas researchers have difficulty extracting ideas from computer
programs or code snippets. This problem includes the difficulty in both describing models
as well as in relating "implicit mental models" to program code. Thus, while a large num-
ber of open source models are freely available online, a large number of mailing list users
still tend to ask fundamental modeling questions; questions which apparently demonstrate
the difficulty of most users in understanding ABMs using only code.
Table 17 Categorical Sampling of Netlogo-users messages for a recent two weeks period (Aug 2011)
Topic
n-Topics
%
Messages
%
Errors
How to model?
Extending/Connecting tools
Total
2
24
5
31
6.45
77.42
16.13
3
73
10
86
3.49
84.88
11.63
In this chapter, as a first step towards resolving these communication problems, we pro-
pose an extension to exploratory agent-based modeling. Our proposed DescRiptivE Agent-
- 139 -
based Modeling (DREAM) approach has been developed with the following user-centric
design goals:
A. DREAM should extend exploratory Agent-based Modeling by allowing for a quanti-
tative comparison of different models without requiring coding or execution of
simulation experiments.
B. Focus on better and more effective documentation has numerous benefits as noted
by Parnas[179] . It can also assist in reverse engineering and model replication as
noted by Li[180]. Thus, in addition to a textual description, the proposed approach
should allow for examination of an ABM visually similar to that demonstrated earli-
er in Chapter 3. As discussed earlier, visual models allow the model to be examined
and analyzed from an abstract perspective rather than requiring researchers to look
at the source code. Such an approach would thus prove useful in model and cas
comparison, description during teaching and in general, more effective knowledge
transfer and communication between multidisciplinary cas researchers.
C. Thirdly the proposed methodology should allow for a translation from these differ-
ent sub-models such as from a visual (Complex network-based) model to
pseudocode specification model to an agent-based model. Apparently this is possible
only if there is a high level of coherence between elements of the sub-models.
The proposed DREAM approach is based on merging complex network modeling ap-
proach with pseudocode-based specification model to specify Agent-based Models. It also
minimizes learning curve because Complex Networks is an existing area of research inter-
est for cas researchers as noted earlier in Chapter 2. As demonstrated earlier in Chapter 3,
Complex Networks allow for exploring data visually as well as quantitatively. Whereas
coupling a pseudocode based specification model allows for a one-to-one correspondence
of the non-code description with the ABM.
- 140 -
The structure of the rest of the chapter is as follows:
First, a generalized description of the proposed DREAM methodology is given. Next as
a validation of this modeling methodology, an experimental design of a complete case
study example is developed. This is followed by a discussion of results of the case study as
well as a critical review of the proposed methodology with existing methods of document-
ing and teaching Agent-based models such as "code templates" and "textual descriptions".
Finally we conclude the chapter.
5.1 Research Methodology
An examination of communication of cas researchers over the past several years in
mailing lists related to Agent-based modeling (such as netlogo-users[177], Repast-
interest[181] etc.) coupled with an examination of cas scientific literature as examined in
detail earlier in chapter 3 reveals that typically cas research can start in several different
ways. Some of the possible cases can be listed as follows:
1. Researchers can start with a concrete research question and need to develop a simu-
lation model to explore it further.
2. Researchers can start with a generalized cas domain but without a specific research
question in mind. In this case, they subsequently want to explore the intrinsic dy-
namics of the domain further using a simulation study.
3. Researchers can have interest in relating multidisciplinary cas studies to discover
generalized patterns related to their particular application case study/scientific do-
main. As such, they are looking to relate concepts from other models to their area of
interest either for developing models or else for comparison of these models with
their models.
- 141 -
Based on these set of ideas, next we give an overview of the proposed DREAM meth-
odology.
5.1.1 An overview of DREAM
In this section, we describe the proposed DREAM methodology in detail. Here the goal
is to provide details in the context of how cas research studies are typically conducted and
how DREAM can be applied to any agent-based modeling case study.
Development of descriptive agent-based models for inter-disciplinary
communication, comparison and learning
Figure 47: Descriptive agent-based modeling level
- 142 -
In contrast to exploratory agent-based modeling level, the proposed descriptive agent-
based modeling level of the framework entails developing a Descriptive Agent-based
Model (DREAM), which comprises of a complex network model, a quantitative centrality-
based "footprint" based on the complex network model coupled with a pseudocode-based
specification model. The goal of this framework level is to allow for inter-disciplinary
comparative studies and a free exchange of information and ideas without regards to scien-
tific disciplinary boundaries. So, if a model has been developed in say, Biological
Sciences, the model should be comparable to models developed in Social Sciences both
visually as well as quantitatively. We can observe further details of this framework level in
Figure 47.
To start with, as a means of generalization, we start by presenting a definition of
DREAMS specification framework. The presentation of this definition is inspired by the
Discrete Event System Specification (DEVS) formalism proposed by Concepcion and Zei-
gler[182].
A DREAM D is a structure
Where
N is a Complex Network model
S is a specification model
A is an Agent-based model
Where A can be further expanded as following
Where
G represents the global variables including both the input as well as the output vari-
ables
- 143 -
ASND,,EPLNPTAGGA,,,,, AG represents the agents
PT represents the patches
LN represents the links
P represents the procedures
E represents the simulation experiments
One way of understanding how DREAM can be used is by an examination of how an
ABM is structured in NetLogo as shown in Figure 48. Thus N, the complex network sub-
model can be translated into a specification model S by handling each of these sections as
shown in the figure. Furthermore, the specification model S can be translated to the ABM
A and vice versa informally.
ABM
Experiments
Globals
Input
Output
Agents
Patches
Links
Attributes
Breeds
Attributes
Attributes
Attributes
Breeds
Attributes
Procedures
Forever
Reporter
Argumented
Reporter-
Argumented
Figure 48: An overview of ABM constructs in relation with the specification model constructs
Here, we can note from the diagram that a complete ABM encompasses various
- 144 -
constructs. The first of these is the "globals" construct. As discussed earlier in the descrip-
tion of the baseline network model, this construct corresponds directly to the one in the
network structure. Thus, globals can be of two types i.e. the input globals which constitute
the model configuration parameters allowing the user to vary the behavior of the model
either manually or else programmatically and the output globals which typically represent
metrics associated with the simulation results. So, while input globals modify the behavior,
output globals represent the quantitative output metrics.
Other than the globals, the specification model can have particular attributes for the
general breed of agent (turtle in the case of NetLogo). If however, there is need to use spe-
cific types of agents, breeds can be defined which can be designed with further attributes.
Similar to agents, we have links which can themselves have attributes as well as be
classified using new breeds. If new link breeds are defined, occasionally, they themselves
are assigned new attributes. Patches, on the other hand, are general purpose constructs and
while they can also be assigned attributes, it is not possible to sub-class them into making
breeds.
While different Agent-based Modeling tools have different ways of specifying proce-
dures (such as using methods associated with a class in the case of using an Object-
Oriented programming paradigm such as Java), here we give a generic definition without
requiring procedures to be tied with agents (similar to use of procedures in Logo-based
tools). If procedures are of general purpose or utility type, they can come directly under the
umbrella of "procedures". However, if the procedures represent specifically one simulation
cycle (in other words, they need to be called repeatedly), then they would come under the
classification of "forever" procedures. In addition, if they return one or more values, then
they are classified as "reporter" procedures. If they take in arguments, they are classified
under "argumented" procedures. Finally, procedures which both take in values as argu-
- 145 -
ments and also return some values are classified as "reporter-argumented" procedures.
Eventually, in a well-structured ABM, the entire set of constructs might be useful for one
or more simulation experiments. Thus, it can be noted here that this division of constructs
has been designed so as to specifically allow the specification model to have a direct one-
to-one correspondence with the ABM constructs as noted earlier in the network model.
The proposed DREAM methodology has been designed to allow transition from the
various sub-models such as the complex network sub-model to the pseudocode-based spec-
ification model and the agent-based model by allowing for a high degree of coherence
between these sub-models. An overview of how translation of various models is possible
can be seen in Figure 49. In this figure, we can note that the modeling research ideas relat-
ed to a cas are translated firstly on the basis of an initial implicit cognitive mental model.
The first step towards developing the DREAM model is to develop a complex network
model. Subsequently for interdisciplinary comparison, the research can proceed by means
of either performing a complex network analysis or else by an informal expansion of the
network model into a pseudocode-based specification. In the case of a complex network
analysis, a result digital quantitative footprint emerges which can prove to be useful for
comparing different models. On the other hand, the network model can subsequently be
informally expanded in the form of a pseudocode-based specification model with the idea
of improving communication between researchers. Another possibility is that researchers
can use an existing agent-based model and subsequently, reverse engineer it into a pseudo-
code-based specification model. Validation to ensure the correct translation of these models
could be performed by means of informal methods such as testing, face validation and
code walkthroughs etc.
- 146 -
Implicit
Cognitive
Mental
model
Quantitative
Comparison
with other
ABMs
Complex
Network
model
Develop a network
model using base-
line network model
Compare
centrality
and other
network
measures
Complex
Network
Analysis
Explain code using
pseudocode models
Use Complex
Network tools
Expansion of
network mod-
el into
pseudocode
Agent-based
Model
Pseudocode-
based specifica-
tion
Translate pseudocode into
code
Figure 49: Pictorial representation of the DREAM Methodology
As noted by Zeigler [183] real-world complex systems need to be modeled using a mul-
tifaceted approach. As such, we can note here that the DREAM methodology can also be
used in several different ways. One way of starting modeling is when a researcher is able to
come up with a tentative implicit mental model of the proposed agent-based simulation
model, this mental model can be converted to a network sub-model using a proposed base-
line agent-based network model, which can serve as a template for developing agent-based
models6. This baseline network sub-model shown in Figure 50 can be used to develop a
complex network sub-model of new agent-based models by expanding each of the leaves
6 For the purpose of this discussion, we shall be focusing on building agent-based models using the NetLogo
tool.
- 147 -
with application case study related constructs. In the next section, we give details about the
development of this complex network model.
Figure 50: Baseline Network template Model of a complete ABM
5.1.2 Design of the Complex Network Model
Initially, to facilitate interactive study, researchers can start by developing a network
sub-model on paper. This baseline network sub-model has been designed to serve as a
template model. Leaf nodes from this model can first be expanded by connecting them to
new nodes in the domain of the application case study. Therefore, at this stage, researchers
can mind map[184] the network sub-model to describe the agent-based model at an ab-
stract level.
- 148 -
The ABM is at the center of the network and it is connected to various nodes. If we first
start with the global variables, we note that they are of two specific types. One of these is
for giving inputs to the model, typically using User Interface (UI) elements (such as slider,
switches etc.) while the other are the output global variables, which are expected to help in
retrieving the simulation metrics from the model. These variables cannot be generalized
and would vary on a case-to-case basis. So, cas researchers can add new input and output
global variables by extending the network from these two types of global variables. Using
the network approach, this step can be performed without worrying about actually going
into the details of how the model is going to be practically implemented.
After deciding the global variables, the next step could be to start looking at adding var-
ious agent types or attributes, as needed. Now, there are two possibilities of using agents;
either the generic agent breed (i.e. Turtle) can be used or else one or more new breeds can
be created. If the generic breed is used then it might still be assigned suitable attributes. In
that case, child nodes can be added to the "agent-attributes" node. If however, new agent-
breeds are to be created, they would be attached to the "agent-breeds" node. Subsequently
any attributes that the particular breed is assigned may be next attached to these agent
breeds.
After deciding the breeds, a possible next step can be to decide upon using links and
patches. Patches cannot have further breeds however new breeds for links can be defined.
Not every agent-based model will have use for specific features related to patches or links
so it would depend on the specific model requirements. Similar to the agent breeds, link
breeds can also be defined, if needed. If no new link breed is defined, then the standard
links can be given new attributes as suited to the case study.
After these object types have been defined for the model, the designer can next focus on
the procedures. If the new procedures are general purpose or helper/utility procedures,
- 149 -
nodes with their names can be attached to the "procedures" node directly. If not, then in the
case of "forever" procedures, which represent one simulation execution cycle, procedure
nodes can be connected to the "forever" node. Other options for parent nodes include the
"reporter" procedure node for procedures which return values and "argumented" procedure
node, for procedures which expect arguments for completing their processing. Finally pro-
cedures can also be defined such that they can both take in arguments as well as report
values. Such procedures can be added as children to the "reporter-argumented" parent
node. These names of different nodes are however given primarily as a template and not
meant to be strict in the sense that if in a particular case study, some specific type of proce-
dures are not required, then the related parent node can be removed from the network sub-
model. In other cases, new types of nodes may be added if it allows for a better representa-
tion of the concepts related to the model.
After deciding upon the names and types of the procedures, the last part of the network
that needs to be finalized is the different experiments that are to be executed once the simu-
lation has been completed. Each of these is developed as a single child node attached to the
"BS-Expts" parent node.
5.1.3 Analysis of the network sub-model
After the network has been designed on paper, it can be modeled on the computer. To
start with, a suitably formatted text file (e.g. using a comma separated or tab-separated
format) can be generated with columns depicting the nodes. An example table representing
the baseline model is given in Table 18.
- 150 -
Table 18 Table showing the format for creation of complex network using Cytoscape
Node from
ABM
ABM
Globals
Globals
Procedures
ABM
Agents
Agents
ABM
Procedures
Procedures
Procedures
ABM
Patches
ABM
Links
Links
Node to
BS-Expts
Globals
InputGlobals
OutputGlobals
Reporter-Argumented
Agents
Agent-Breeds
Agent-Attributes
Procedures
Forever
Reporter
Argumented
Patches
Patch-Attributes
Links
Link-Breeds
Link-Attributes
There are several ways in which networks can be extracted from a text file and im-
ported into Network tools such as the Network Workbench[162]. One possible method is to
manually create a plain text file compatible with one of the standard file formats. Subse-
quently, the file can be imported in a complex network modeling tool and visualized using
a suitable layout algorithm such as the GEM Layout in GUESS[167].
In addition to model conversion, the network sub-model can now be analyzed using
CNA as discussed in detail earlier in Chapter 2 and demonstrated in Chapter 3. Such an
analysis would thus allow for several benefits such as:
1. The different centrality measures would serve as a "digital footprint" of the
Agent-based Model.
2. The ability to discover relationships between ABM-based complex network
sub-models across domains or applications quantitatively.
After having developed and analyzed the network sub-model, specification tem-
plates can be used for developing the pseudocode-based descriptive specifications for the
- 151 -
ABM by expanding relevant nodes from the ABM complex network sub-model. These
templates are described in the next sections.
5.1.4 Goals of a pseudocode-based specification model
After the construction of the network sub-model, it can be next expanded into a pseudo-
code-based specification model. For developing this specification, the design goals of this
part of the DREAM methodology can be listed as follows:
1. Robin et al.[185] review facts about "learning curve theory" in the domain of pro-
gramming. They note the general acceptance of the fact that "novice/amateur"
programmers can take up to 10 years of programming practice and experience to
turn into "expert" programmers. As such, the proposed specification syntax should
have a short learning curve by being easy to use for multidisciplinary cas research-
ers to minimize transition difficulties without adding further to this learning curve.
2. It should be based on using a flexible set of terms in either natural language or very
basic formal mathematical terms (such as use of sets or predicates).
3. It should conform to the constructs used in an agent-based model. In other words,
the specification needs to have a strong coherence with the source code for the
agent-based model.
In other words, the key design constraint in the template syntax design is to make it
comfortable for learning and use by multidisciplinary researchers such as social scientists,
life scientists or linguists to use comfortably. The pseudocode part of this syntax can thus
be a comfortable mix and match of any syntax or form which is comprehensible and pre-
ferred by the specific cas researchers
To summarize, based on the specific user-centric design requirements of DREAM
specifications, the guidelines for writing the pseudocode are relaxed and not based on any
- 152 -
type of stringent rules as the entire goal of this modeling approach is to improve communi-
cation without causing hurdles or
requiring an extensive
learning curve
for
multidisciplinary researchers. Pseudocode used can be either "pidgin" code, a technical
term for pseudocode using formal mathematical symbols or else plain English language
pseudocode. The original idea of pidgin code is from pidgin ALGOL (short for ALGO-
rithmic Language)[186].
5.1.5 Templates for the specification model
Here in this section, generalized translation templates are presented. These templates al-
low the expansion of the complex network model to agent-based model and vice versa.
The first template is for defining a new breed. As noted below, it starts with a name and
description of the breed. Type can be either an "agent" or a "link". Finally, the breed can be
given any number of "own" or "internal variables". Writing this detailed description thus
allows the researcher to expand and refine the ideas in their implicit mental models by
writing down the specification models.
Breed BreedName: Single Sentence description of breed
Type: Agent Link
Internal Variables: <variable 1, variable 2, . . , variable n.>
variable1: Conceptual role and description of the variable 1
variable2: Conceptual role and description of the variable 2
. . .
variablen: Conceptual role and description of the variable n
After defining the breeds, a possible next step could be to define the different types of
input and output global variables which will be used in the simulation. By starting with
these variables, the designer can think about the research questions, in general or the prob-
lems to be explored, in particular. Input variables can be described in terms of the user
interface elements whereas output variables can be defined without any relation to these
elements.
- 153 -
Globals: <variable 1, variable 2, . . , variable n.>
Sliders:
variable1: Description of the variable
variable2: Description of the variable
. . .
variablen: Description of the variable n
Switch:
sw1?: Description of the switch
sw2?: Description of the switch
. . .
swn?: Description of the switch
Input:
ip1?: Description of the input
ip2?: Description of the input
. . .
ipn?: Description of the input
Output:
op1?: Description of the output metric
op2?: Description of the output metric
. . .
opn?: Description of the output metric
Next, we present the template for different types of procedures. Procedures can be of
numerous types as discussed earlier. By using this particular template, the researchers can
give details of the procedures they intend to use in the model. The template firstly de-
scribes the input (pre-conditions or the arguments taken in by the procedure). Output is the
resultant or expected outcome of the procedure. Execution can be used to note how the
procedure is called and during what phase of the simulation. Next, the context of the pro-
cedure is defined; contexts can be agent, patch or observer types. Observer context is
useful for general purpose execution or if multiple sub-contexts are going to be used in the
procedure. Finally, the procedure is defined in terms of pseudocode.
- 154 -
InputOutput
Procedure proc1: Description of procedure
Input: Description of inputs and pre-conditions expected by the procedure
Output: Resulting outcome in the simulation
Execution: Called by which procedure and during which process
Context: Observer Agent Patch
begin
1. Appropriately indented Pseudocode step
2.
3.
Appropriately indented Pseudocode step
Appropriately indented Pseudocode step
4.
. . . .
5. Appropriately indented Pseudocode step
End
Next, the specification template for experiments is presented.
Experiment exp1: Description and goal of the experiment
Inputs: < ip1 , ip2 , . . .ipn >
Setup procedures: < spr1 , spr2 , . . .sprn>
Go procedures: < gpr1 , gpr2 , . . .gprn>
Inputs:
ip1: {constant value} {[value1 , value2 , . . . valuen } {[value1
ip2: {constant value} {[value1 , value2 , . . . valuen } {[value1
. . .
ipn: {constant value} {[value1 , value2 , . . . valuen } {[value1
Stop condition: Pseudocode
Final commands: Pseudocode
incrementFinalvalue]}
incrementFinalvalue]}
incrementFinalvalue]}
In this template, we note that experiments are first described in the top of the schema.
Next, their "setup" and "go" procedures can be defined. "Setup" procedures are those pro-
cedures which will only be called at the start of a simulation experiment. "Go" procedures,
on the other hand, will be executed with each simulation step. Stop condition represents
when a particular simulation run will stop execution. Final commands are the commands
which might be executed once the simulation is terminating. Input values can be specified
either as a single value, or else a group of values or in terms of a three value list (with start,
increment and final value as its members).
- 155 -
5.2 Case study: Overview, Experimental design and Implementation
In the previous section, we have given details of the DREAM methodology by provid-
ing details of how a proposed case study in agent-based modeling may proceed. As a
validation of the proposed methodology, in this section, we give an overview of the case
study, experimental design and the implementation. This includes the development of the
DREAM model consisting of a complex network model, pseudocode-based specification
model, a quantitative centrality-based "footprint", description of the implementation as
well as the design of the simulation experiments.
5.2.1 Overview
The case study follows the complex adaptive communication networks concept in-
troduced in the previous chapter. In this case study, the goal is to explore a well-known
artificial cas concept of "flocking" and use it to evaluate how Wireless Sensor Network
nodes behave in the vicinity of complex behavior.
It is essentially based on the "Boids" model of flocking behavior introduced earlier
in chapter 2. Boids is an example of an artificial cas. In other words, it is an in-silico repre-
sentation of a set of real-world agents (e.g. birds or fish or insects etc.). The initial idea was
propagated based on an exploration of how flocks of animals and birds group together and
form complex structures based on seemingly simple and local rules. Thus, using rules such
as cognitive perception of nearby "boids", every Boid agent is able to adjust the direction
such that it neither collides with its neighbor boids nor follows any general direction. The
way this model technically works is based on a set of calculations. In a real living cas, the
calculations are deemed to be performed automatically by means of perception of motion,
however in the artificial life scenario (simulated cas), the calculations have to be actually
performed. This interesting simulation model[187], commonly available with agent-based
- 156 -
simulation tools such as NetLogo exhibits intelligent collective behavior based on local
rules followed by the components ("boids" in this particular case).
5.2.1.1 Problem statement
While the behavior of boids and flocking has been well-studied in literature[100],
to the best of our knowledge, no well-known study has been conducted previously in the
domain of examination of how physical wireless sensor nodes would react in the presence
of a group of flocking "boids". As discussed earlier, Wireless sensor networks are made up
of sensor nodes and most, if not all, simulators of WSNs are based on extensions of exist-
ing communication network simulators. As such, these simulators have difficulty in
modeling any advanced concept such as "mobility modeling" or else "environment model-
ing". This is actually a problem because WSNs are not primarily designed for
communication rather network communication is simply a means by which the WSN ex-
tracts sensing information from individual sensor nodes since this environmental sensing is
the primary goal of developing a WSN application[188]. However, it has taken considera-
ble research efforts in the domain of communication networks to come up with custom
algorithms which suit the specialized nature and limited battery problems[189] related to
the WSN nodes[190]. As such, a key focus area of these networks has previously been on
getting the prototypes and the communication of simple messages across the network
working (routing). More recent advanced algorithms include focus on data aggregation and
fusion[191]. However, previously, to the best of our knowledge, the problem of developing
a cas model for WSNs has not been studied. WSNs themselves constitute a cas because of
how they need to self-organize for performing common goals such as monitoring and sens-
ing[73]. In addition, if we couple a cas environment of flocking boids, it poses a set of
interesting questions as to how to study the interactions of two cas with each other in the
same simulation model. The goals of this simulation study can thus be noted as follows:
- 157 -
1. Evaluate the use of simple sensors in sensing complex cas behavior.
2. Evaluate the cas effects in these sensors and note the effects on sensing this
complex behavior when different sensor nodes start to die (due to battery fail-
ure). In other words, the nodes themselves are eliminated from the simulation.
So, assuming the results of the first set of experiments demonstrates if sensing
can somehow be performed, what would be the effects on sensing as WSN
nodes die down due to a battery failure at different time instances during the
simulation.
There are numerous interesting complexity problems which pose a unique set of modeling
challenges in this case study as it reflects a merger of an artificial cas along with the physi-
cal notion of a set of sensor nodes randomly deployed in the same terrain. Some of these
problems can be noted as follows:
1. Complexity due to the special nature of the environment (constituted by the boids as
well as the other sensors)
2. Complexity due to a random deployment model[192].
3. Complexity due to the gradual possibility of reduction in the total number of sensors
available to sense the proximity of the boids.
5.2.2 Agent design
In this section, we describe the design of the agents used in the simulation model.
Primarily there are two types of agents which will be used in the simulation which are de-
scribed next.
- 158 -
5.2.2.1 Boid agents
Boid agents or Boids are the mobile agents in the model. These agents exhibit complex
behaviors. The complex behaviors are collective however as in different real-world cas,
similar to the discussion in the previous chapter, these agents are able to have particular
attributes and behaviors described as follows:
1. Visibility and ability to view
Boids are able to view other Boid agents and can focus on using a certain set of in-
dividuals as their visible neighbors. In other words, this particular group of
individual boids inadvertently decides the direction chosen by the Boid agent in the
near future.
2. Mobility
Boids are able to move about the screen. This motion is in general dictated by the
neighbors. In other words, each Boid agent decides upon a certain direction to fol-
low based on its neighbors in a certain "visibility" radius.
3. Change of direction
The direction followed by each Boid is what results in the flocking behavior. Thus
this change of direction is based on certain calculations which we shall describe
next.
Boid agents change their directions based on the following three functions:
5.2.2.1.1 Separate function
This function is primarily for ensuring collision avoidance. The way it works is that each
agent first locates its nearest Boid neighbor. Next, it perceives the agent's heading and
slightly changes its own direction.
- 159 -
5.2.2.1.2 Align function
This function has a different goal. It allows the agent to change its direction towards a
certain heading. This heading is calculated by first enlisting other neighbor Boid agents,
which this agent considers as its flockmate neighbors. Next, their headings are calculated
and averaged. Based on this heading, the agent slightly turns towards this heading. The net
effect is to align in the general direction of the nearby agents.
5.2.2.1.3 Cohere function
This function is called alongside the align function. While the previous function discov-
ered the average flockmate heading, this particular function is concerned with the angles of
the other flockmate agents towards this agent. Thus, first the function requires the discov-
ery of the angles of these agents to itself but adds 180 degrees to these angles. The reason
for that is because that way, it will give the direction from this agent's perspective instead
of the other agents' perspectives. Then these angles are averaged. Finally the agent adjusts
its own heading towards this general direction slightly.
5.2.2.2 Node agents
In the same ABM as the boids model, the case study requires the placement of the sen-
sor node agents. These agents are randomly deployed. Such kind of deployment follows
the concept of "smart dust" devices[193] which can be deployed either in space[194] or on
land. In the case study, these agents have communication capabilities. Thus, these node
agents can communicate with a base station which is located within their transmission ra-
dius (single-hop). As such, they do not require multi-hop communication paradigms. The
reason for choosing a single-hop network sub-model is to allow focus on the sensing aspect
because if multi-hop networking was used, a considerable energy can be lost in developing
the routing algorithms. However, in this scenario, the goal of these sensors is to detect
- 160 -
"boids". The physical implication of detection can possibly involve the use of "infra-red"
or "image-processing" as a means of proximity detection of moving boids. In addition, bi-
nary sensors with acoustic detectors have previously been discussed by Kim et al. [195].
5.2.2.2.1
States of Node agents
Node agents have several different states as shown in Figure 51. Firstly, the agents are
active (in other words functional). If they are functional, next they can go into sensing state
if they detect a Boid in their proximity or else non-sensing state otherwise. However, if
battery power goes down, these agents would then go to inactive state.
Active
Inactive
(battery
down)
Boids
Detected
Boids not
Detected
Figure 51: State chart diagram for sensor node agents
5.2.3 Network Model
As noted earlier in the chapter, DREAM can be initiated here by developing a paper
model of network design by extending the baseline network model. The network sub-
- 161 -
model can be imported into a network modeling tool. The result can be seen as shown in
Figure 52.
Figure 52: Initial view of imported network
In its current state, the network is difficult to comprehend. As such, there is need for
network manipulation. After the network has been further manipulated using various layout
algorithms, the resulting network can be viewed in Figure 53.
This network model will be described in detail and a CNA will be performed next. The
results will be discussed in the results section of the chapter.
- 162 -
Figure 53: Network view after manipulation algorithms
5.2.4 Pseudocode-based specification
Next we develop a detailed translated specification model expanded from this network
sub-model as a means of describing the model in detail. We first start by describing the
agents and breeds.
5.2.4.1 Agents and Breeds
The model has two breeds and their definitions are given next in the form of a specifica-
tion model as below:
Breed Node: This breed represents the Wireless Sensor Nodes
Internal Variables: <boids-near, sensed?>
boids-near: All agents of breed boid which come closer to the given sensor node
sensed?: A Boolean variable which represents the state of the sensor in response to sensing
Now, the "Node" breed here is first described in the specification model. Afterwards we
note the internal variables that will be used in the simulation. There are two specific inter-
- 163 -
nal variables here. One is the "boids-near" variable. This variable is used to keep an
"agentset" of all nearby boids. The variable is key to the discovery of flocking which is
one of the basic simulation requirements for the design of the sensor nodes.
On the other hand, "sensed?" variable is a Boolean variable. It represents the state of the
sensor. In other words, when the sensor is active and is detecting boids in its vicinity, this
Boolean variable will turn true. However, if the sensor is active and not detecting any boids
in its sensing radius, the value will turn false.
The next breed is the "boid" breed, which is described below in the specification model
as follows:
Breed Boid: This breed represents the Boids (birds)
Internal Variables: <nearest-neighbor, flockmates>
nearest-neighbor: A boid agent nearest to the agent itself
flockmates: An internal model agentset of other boids which this boid considers as flockmates
The breed "Boid" represents the Boids agent type in general. Now, this particular agent
breed/type has two internal variables. The first internal variable is the "nearest-neighbor"
variable. Nearest neighbor is a variable which will be holding an agent. This single agent
will be reflecting the "internal model" of the agent regarding which other environmental
agent is physically near to this agent. This modeling concept reflects the cas concepts dis-
cussed earlier in Chapter 2. This internal model of the Boid agent keeps updating over
time. We would like to note here that this concept of internal model matches with the con-
cept of "Belief" in the well-known "Belief-Desire-Intention framework by Rao and
Georgoff[196] used to model rational agents in multiagent systems.
Next internal variable of the Boid agent reflects the flockmates "agentset". Flockmates
are again related to the belief of the current agent. In other words, the agent has this "inter-
nal model" noting that this set of agents is its flockmates. Now, since every Boid agent
might be moving in the simulation as well as continuously updating this variable, the fact
- 164 -
that one agent has other agents listed as flockmates does not guarantee that the other agents
will also keep this agent as flockmate.
After developing a specification model of the agent breeds, we next develop a model for
the various global variables. These variables are important for the configuration of the
simulation model.
5.2.4.2 Globals
Here we describe the global variables for the simulation model. Now, in this model, the
key input global variables here are four in number. These are described next in the specifi-
cation model given below.
Input Globals: <n, n-boids, visible?, max-scen?>
Sliders:
n: Used for giving input for the number of sensors in the simulation
n-boids: Used for giving input for the number of boids
Switch:
visible?: Switch used to remove boids from the screen for only observing patterns in sensors
max-scen?: Switch used to create maximum sensors with one sensor per patch
In this specification, we can note clearly now that two of the variables viz. "n" and "n-
boids" both represent inputs which are provided by the user via a "slider" UI element in the
NetLogo simulation model. Whereas "visible?" and "max-scen?" are switches or in other
words, represent Boolean input variables. "n" is used to configure the number of initial
sensors in the simulation. These sensors will be randomly deployed when the initial simu-
lation screen is setup. Having "n" as a slider variable allows for a gradual change in the
input configuration as needed in different experiments. It also assists in minimizing the ef-
fects of random sensor node deployment in the model. So, while one particular simulation
experiment might result in specific results due to certain placement of sensor nodes, repeat-
ing the simulation several times (e.g. 50 times in this particular case study) will ensure that
the specific deployment does not skew the eventual results.
- 165 -
"n-boids" slider variable is used to enter the number of boids used in the simulation.
The reason for having this variable configurable is to test the validity of hypotheses with a
variation in the number of Boid agents. So, essentially this helps answering questions such
as:
-
If a hypothesis appears to be true with a certain number of Boid agents, would the
same be true with lesser agents or else more agents?
The same can be noted about the slider described earlier i.e. depicting the initial number
of sensor nodes in the simulation.
The switch "visible?" is especially useful in case we do not want to observe the simula-
tion based on the boids. In other words, if we only want to examine the patterns being
made in the randomly deployed sensors, then we can use this switch by turning it to false.
The other switch "max-scen?" is for the maximum scenario of sensors in the sense that
in this case, sensor nodes will be located in the form of a lattice over the entire simulation
world. Thus each "patch" of the simulation will behave as a single agent. This could be
useful when instead of a randomly deployed sensor network, we want to use a set of sen-
sors which are placed at exact locations. This would again allow us to examine the validity
of the hypotheses that are to be tested in the simulations.
5.2.4.3 Procedures
After describing the breeds and the global variables, we next start to describe the vari-
ous procedures which will be part of the model. These procedures can be of different types
and we shall use the templates described earlier in the methodology section as a means of
describing and expanding each procedure in detail.
The first procedure is the setup procedure. The specification model for the setup proce-
dure is given below as follows:
- 166 -
Procedure setup: Setting up the simulation
Input: Uses global parameters from the User Interface
Output: All agents and patches are setup for simulation
Execution: Called at the start
Context: Observer
begin
1. Clear the screen.
2. Clear all nodes
3. Clear all patches
4. Create WSN nodes
5. Create Boid nodes
End
The "setup" procedure is the key procedure to every agent-based simulation model.
While the actual name of the procedure might be different, similar procedures are needed
to setup the scenarios in a simulation. Here, it can be noted that in this particular "setup",
it needs to be ensured that the remnants of the previous simulation experiment do not dis-
turb the next simulation. As such, the previous results are also cleared before the execution
of the next simulation. This would allow the simulation experiments to eventually run in an
automated fashion. Next, each of the patches is sent a message. This message is to change
their particular color to "white". The reason for doing that is to be able to better observe
the simulation visually during the animations. After this, three different functions are called
next. Each of these functions sets up part of the scenario. Dividing the setup into multiple
functions allows for code localization. Thus, if needed, the particular functions/procedures
can be modified without changing the global setup procedure. Next we describe each of
these procedures. The first of these procedures is the "make-wsn" procedure described as
follows:
- 167 -
Procedure make-wsn: Creating the nodes
Input: Uses global parameters from the User Interface
Output: All nodes are created after this function call
Execution: Called by setup procedure
Context: Observer
begin
1. If maximum sensors are needed
2. Create maximum node agents
3. Else
4. Create agents based on required numbers
End
"make-wsn" is a procedure which will setup the sensor nodes in the simulation experi-
ments. It is called without any particular context (in other words, labeled as the observer
context). Next, we describe the "init-node" procedure as follows:
Procedure init-node: To Initialize the nodes
Input: Uses internal variables of node
Output: The node is configured according to the norms of the simulation
Execution: Called by make-wsn
Context: Agent
begin
1. Initialize active sensor nodes displaying their state
End
Note here that this procedure has an "agent" context. In other words, this implies that
the procedure can be used by changing specific attributes of the agent. Here the goal is to
initialize a single Node agent for the simulation. This procedure is called by the "make-
wsn" procedure. It essentially sets the color, shape, size and the initially active state of the
sensor before the start of the simulation.
The next procedure is the "make-reg-wsn" procedure:
- 168 -
Procedure make-reg-wsn: For Creating nodes according to inputs
Input: Uses global variables from User Interface
Output: The entire wsn is configured according to the norms inputs
Execution: Called by make-wsn
Context: Observer
begin
1. Create nodes
2. Place nodes at random locations in the simulation world
End
This procedure is used to create the randomly deployed sensor network. It uses global
variables from the User interface elements and is executed by the "make-wsn" procedure.
It creates "n" nodes and then firstly initializes each node by calling "init-node" and second-
ly sets their location randomly in the simulation world. Next, it demonstrates the use of
individual intelligence of the agents by allowing "node" agents to relocate based on the
proximity of other nodes. This will allow the simulation to be more realistic since in the
real physical world, the chance of two sensors at the same location is not very probable. It
is important to note here that while this mobility is part of the procedure, it is not part of
the capability of the sensor nodes in this case study. Simply mobility of nodes is used here
to have a better and more realistic placement for simulation execution.
We next write the specification of the "make-boids" procedure given as follows:
Procedure make-boids: For creating boids according to inputs
Input: Uses global variables from User Interface
Output: Boids are created and initialized
Execution: Called by setup
Context: Observer
begin
1. Create boids
2. Initialize boids
3. Place boids at suitable random locations in the simulation environment
End
The "make-boids" procedure is a basic setup procedure. It is used to create and setup
agent attributes such as the shape, size, color and location of the Boid agents.
- 169 -
We next write the specification of the "go" procedure. The "go" procedure is responsi-
ble for executing a single step in the simulation. As such, it needs to be called repeatedly.
It takes in global parameters from the user interface. Firstly, it is expected to increment a
global environment specific tick count. While simulation runs can execute without incre-
menting the tick count, it generally makes concrete steps in the simulation, which can be
later measured (as we shall want to do in "go80" procedure). Next this procedure calls
"move-boids" and "wsn-sense" procedures. These procedures are responsible for code exe-
cution pertaining to the Boid and the Node agent respectively. Finally plotting code is
executed by calling "do-plot" which will be useful for observing the simulation evolution
temporally as shall be explained subsequently.
Procedure go: For executing one step of the simulation
Input: Uses global parameters from the User Interface
Output: Equates to a single step of each agent in the simulation
Execution: Called repeatedly for execution of the simulation
Context: Observer
begin
1. Move boids in the environment
2. Allow sensors to sense their vicinity for boids
3. Update simulation plots showing active sensing boids
End
Next, we write the specification model of the "move-boids" procedure. As discussed
earlier, this procedure is designed to be the primary procedure for Boid agents. The general
functionality here is similar to standard published "Boids" models and extends NetLogo
model library code. The basic idea in boids simulation is for boids to be close to their per-
ceived flockmates but not too close. So, here the Boid agent looks at its nearest-neighbor
and if it notes that it is too close, it calls "separate" otherwise it calls "align" and "cohere".
After adjusting the heading using these functions, the agent moves forward by a small
amount.
- 170 -
Procedure move-boids: For moving the boids for a single simulation step
Input: Uses global parameters from the User Interface
Output: Moves each simulated boid by a single simulation step
Execution: Called by the go function
Context: Observer
begin
1. Let boids maintain a list of flockmates
2. Let boids adjust their headings by means of separation, alignment and coherence
End
The first of the direction adjusting procedures that we write the specification for next is
the "separate" procedure and given as follows:
Procedure separate: For moving the boid away from the direction nearest neighbor is moving
Input: Uses no specific input parameters
Output: Moves the boid agent away from the nearest-neighbor boid
Execution: Called by move-boids procedure
Context: Agent
begin
1. Update which of the nearby flockmates are nearby
2. Find the heading of the nearest neighbor
3. Slightly change direction so as to avoid future collision
End
The logic in this procedure is based on messaging. The Boid agent requests its "nearest-
neighbor" for its heading. In a more advanced simulation, the neighbor can decide whether
or not to give the heading back. However, in this case, since the goal here is to examine the
effects of the sensing in sensor nodes, the "Boid" agent gets the heading from the "nearest-
neighbor" and then passes this as an argument to the "turn-away" procedure.
Next, we write the specification of another boid-related procedure "align" as follows:
- 171 -
Procedure align: For moving the boid toward the general direction of flockmates
Input: Uses no specific input parameters
Output: Moves the boid agent towards the average flockmate heading
Execution: Called by move-boids procedure
Context: Agent
begin
1. Update a list of current flockmates
2. Based on the flockmates' directions, find a suitable average heading
3. Change the direction to this average heading
End
We move next to describe the procedure "cohere" using the templates as follows:
Procedure cohere: For moving the boid for cohering with the flockmates
Input: Uses no specific input parameters
Output: Moves the boid agent towards an average value of heading for cohering
Execution: Called by move-boids procedure
Context: Agent
begin
1. Find the average heading towards flockmates
2. Slightly adjust direction towards this average heading
End
The next procedure "wsn-sense" is related to the sensor Node agents as well as the Boid
agents. Here in "wsn-sense" procedure, the Node agents are used to check their proximity
for Boid agents. If they are discovered then the state of the Node agents is changed both
visually as well as internally.
Procedure wsn-sense: Allows Node agents to change state based on sensing of boids
Input: No specific input values
Output: Changes the state of the sensor nodes based on the proximity of Boid agents to the Node
agent
Execution: Forever procedure called from the user interface
Context: Observer
begin
1. Make a list of boids nearby
2. Update state of sensor node based on if there is any boid nearby
End
The final procedure here is the "do-plot" procedure which performs the plotting.
- 172 -
Procedure do-plot: For plotting the current status of Node agents
Input: None
Output: Updates the plot with the current set of Node agents who are sensing Boids nearby
Execution: Called from the go procedure
Context: Observer
begin
1. Plot the number of active sensing sensor nodes
End
5.2.5 Experiments
There are two experiments that we shall be using in this simulation. The first one
demonstrates the effect of the variation of number of sensor Node agents while the second
one shows the effects of the variation in the number of Boid agents.
Experiment vary-sensors: Experiment with the effects of changing the number of Node agents
Inputs: < n-boids, n, visible?, max-scen? >
Setup procedures: < setup>
Go procedures: <go>
Repetition: 50
Inputs:
n-boids: 50
n: [100 1001000]
visible?: true
max-scen?:false
Stop condition: Ticks=1000
Final commands: None
Experiment vary-boids: Experiment for noting the effects of changing the number of Boid agents
Inputs: < n-boids, n, visible?, max-scen? >
Setup procedures: < setup>
Go procedures: <go>
Repetition: 50
Inputs:
n-boids: [100 1001000]
n: 1000
visible?: true
max-scen?:false
Stop condition: Ticks=1000
Final commands: None
- 173 -
5.3 Results and Discussion
In this section, we present the detailed results and discussion based on the complex net-
work model design as well as the simulation experiments proposed earlier in the case study
specification model.
5.3.1 Complex Network Analysis of the Network sub-model
We start by performing manipulation of the network shown earlier in Figure 52. The
first analysis is by calculation of the degree centrality measure discussed in detail earlier in
Chapter 2. Once the degree measure has been calculated for each node, it is subsequently
loaded/added to the corresponding node as an attribute. Subsequently this centrality meas-
ure is used to first resize and then colorize each node. The final visualized network is
shown in Figure 54.
- 174 -
Figure 54: Complex Network Model of Boids and WSN simulation nodes resized and colorized accord-
ing to degree centrality
Here, the first point to note is that we can clearly note peculiar characteristics of this
ABM visually without needing to look at the source code specifically. If another complex
network of a different ABM were constructed, it might have significantly different charac-
teristics. Now, firstly, we note the size of the "procedures" node, which is larger and darker
than other nodes because of the large number of procedures in this particular model. There
are also several "forever" and "Reporter" procedures which have resulted in larger node
sizes for these as well. We can also note how the network sub-model allows an examina-
tion of the agent-breeds and their respective attributes. Finally, the global variables as well
as the two behavior space experiments are visible in the network sub-model.
While the degree centrality gives an interesting overview of the network, there are other
quantitative measures which give further topological details as shall be examined next in
terms of detailed CNA. Next, we calculate various quantitative network measures of the
complex network presented in Table 19. One way of looking at this table is considering
this as a digital network footprint of this particular ABM. Thus, another ABM could possi-
bly have a completely different network footprint thus allowing for a quantitative similarity
analysis and comparison similar to the use of minutia and other features extracted from
human fingerprints for use in person identification and fingerprint indexing systems[197].
Table 19 Table of Eccentricity, Betweenness and Degree Centrality measures of the complex network
Id
ABM
Agent-Attributes
Agent-Breeds
Agents
Align
Argumented
average-flockmate-heading
average-heading-towards-flockmates
Eccentricity (%)
Betweenness (%)
0.00
0.00
8.85
6.19
0.00
5.31
0.00
0.00
15.36
0.00
3.20
3.36
0.00
3.84
0.00
0.00
- 175 -
Degree
(%)
4.44
1.11
3.33
3.33
1.11
3.33
1.11
1.11
Boid
Boid-Attributes
Boids-Near
BS-Expts
Cohere
data-of
do-plot
Flockmates
Forever
Globals
Go
go80
init-node
InputGlobals
make-boids
make-reg-wsn
make-wsn
max-scen?
move-boids
N
n-boids
nearest-neighbor
Node
Node-Attributes
OutputGlobals
Procedures
Reporter
Reporter-Argumented
sensed?
sensor-data
Separate
Setup
turn-away
turn-towards
vary-boids
vary-sensors
visible?
wsn-sense
0.00
0.00
0.00
1.77
0.00
0.00
0.00
0.00
5.31
5.31
0.00
0.00
0.00
7.08
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
7.96
7.08
0.00
2.22
3.33
1.11
3.33
1.11
1.11
1.11
1.11
3.33
3.33
1.11
1.11
1.11
5.56
1.11
1.11
1.11
1.11
1.11
1.11
1.11
1.11
2.22
3.33
1.11
37.17
17.78
7.96
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
4.44
2.22
1.11
1.11
1.11
1.11
1.11
1.11
1.11
1.11
1.11
1.11
0.00
5.76
0.00
3.84
0.00
0.00
0.00
0.00
3.84
5.76
0.00
0.00
0.00
7.68
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
2.88
3.84
0.00
20.16
5.76
14.72
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
- 176 -
Figure 55: Plot showing the Degree Centrality
The degree centrality analysis is plotted in Figure 55. The first point to note from this
analysis is that the degree centrality of the "procedures" node is the highest. The second
highest degree centrality is the "input-globals". These represent the inputs which are en-
tered from the NetLogo user interface model.
Next, the plot showing the eccentricity centrality is shown in Figure 56. Here, we note
that the highest eccentricity centrality value is again for "Procedures" node. However, the
second highest centrality node is that of "ABM" itself followed closely by "Reporter-
Argumented" which depicts such procedures which have both reporters as well as argu-
ments. This is followed by "InputGlobals".
- 177 -
Figure 56: Plot showing the eccentricity centrality
Finally, the plot showing the betweenness centrality[198] is shown in Figure 57. In the
case of betweenness centrality, following "Procedures" which has the highest centrality
value, we find "Agent-Breeds" to have the second highest value followed by "Reporter",
"Node", "InputGlobals" and "Node-Attributes".
- 178 -
Figure 57: Plot showing the betweenness centrality measure
The implications of these network topological measures are numerous. Firstly, as dis-
cussed earlier, having a network allows for a visualization-based analysis and comparison
of various agent-based models without resorting to the source code. Secondly, quantitative
measures can be used to create a network footprint database of models for a repository of
different ABMs corresponding to numerous multidisciplinary studies.
5.3.2 Simulation Measurements
While the previous section discussed the results of the CNA of the network sub-model
of the case study, starting from this sub-section, we focus on the simulation experiments
and their results. We note that the key measurement in this set of simulations is "Sensed"
(represented below by
). Sensed is the mathematical aggregation (sum) of measure-
ments of all sensors at any given time t. So, being a binary sensor, every time a sensor
turns on, it results in an increase in the Sensed value. Thus, Sensed would always be less
- 179 -
()St than or equal to the total number of available sensors in the world at that instance (
)
as follows in equation(2.6).
(2.6)
If
represents the state of a single sensor which is detecting "boids" in its proxim-
ity, from a total of i active sensors at a given time,
would be translated as the sum of
values of all currently sensing and active proximity sensors formally defined as follows in
equation (2.7):
(2.7)
5.3.3 Results from Simulation Visualization
In this section, we give a detailed overview and discussion of the simulation results
from the agent-based model constructed using the pseudocode-based specification model.
We start by giving the details of how flocking boid agents are visualized in the simulation
model. In the following Figure 58a, we can note the boids are in the simulation shown here
as "bees". Grey sensor node agents are deployed randomly in the simulation. However,
those sensor nodes which are close to the Boid agents detect them and change the color and
size to black.
- 180 -
()Nts()()sStNt()ntd()St0()()diddStnt
Figure 58:(a) shows unflocked "boids" while (b) shows flocked "boids" and the sensor nodes sensing
the boids, show up as black on the screen.
With the passage of time, we note in Figure 58b that the boids flock together considera-
bly and the number of sensors detecting the boids has greatly reduced. In other words,
from the simulation model, we can directly correlate how the Sensed measured value
should correspond and vary with flocking.
While the above figures demonstrated the concept of detection and sensing of the binary
sensors, we need to validate the simulation parameters by checking to see if the simulation
actually works. This is performed visually in Figure 59a and b. Here, the simulation execu-
tion has been performed by turning on the tracks for the mobile Boid agents and removing
the visible sensors and all other agents from the simulation screen. We can note here that at
the start of the simulation, the Boid agent tracks are more or less random and not coherent.
However, in the later case, the Boids have flocked together and while the overall move-
ment of the flocks might be considered as random, the agents have flocked in particular
ways towards directions followed by the nearest Boids giving the example of an intelligent
emergent collective behavior of flocking.
- 181 -
Figure 59: Sensed Motion tracks of "boids" showing emergence of flocking behavior (a) Unflocked (b)
Flocked
5.3.4 Quantitative results
As discussed earlier, there were two key experiments in this case study. The first was to
test hypotheses related to the sensing of Boid agents with respect to a variation in the total
number of Boids whereas the second experiment was related to the examination of meas-
urements in relation to variation to the total number of sensors. In this section, we give
detailed analysis of the simulation results.
To start with, the descriptive statistics of "vary-boids" experiment are given in Table 20.
We can note here that the total number of simulation steps were close to 0.5 million. The
total simulation runs "Runnum" were 500 where each run was repeated 50 times to ensure
that the results would minimize stochastic effects of the simulation experiments. Each ex-
periment was executed for 1000 steps. The Sensed measurement varied from a value of 118
to 999 with a mean value of around 670 and a standard deviation of around 244. Here we
can also note that the number of sensor Node agents was set to a constant value of 1000
while the number of Boid agents was varied from 100 to 1000 with increment of 100.
- 182 -
Table 20 Descriptive Statistics of vary-boids experiment
Descriptive Statistics
Runnum
n-boids
N
Step
Sensed
Valid N (listwise)
N
Minimum
Maximum
Mean
Std. Deviation
500500
500500
500500
500500
500500
500500
1
100
1000
0
118
500
1000
1000
1000
999
250.50
550.00
1000.00
500.00
670.48
144.337
287.228
.000
288.964
244.277
We can note the descriptive statistics of experiment "vary-sensors" given in Table 21. In
contrast to the first experiment set, we can observe here that firstly the variation in the
Sensed value is considerably different from the previous set of experiments varying from 1
to 267. The total number of experiments is still 1000 but the number of Boid agents is set
to a constant value of 50. In contrast, the number of sensor Node agents varies from 100 to
1000. Each simulation was executed for a duration of 1000 simulation steps (or simulation
seconds).
Table 21 Descriptive statistics of experiment vary-sensors
Descriptive Statistics
Runnum
n-boids
N
Step
Sensed
Valid N (listwise)
N
Minimum
Maximum
Mean
Std. Deviation
500500
500500
500500
500500
500500
500500
1
50
100
0
1
500
50
1000
1000
267
250.50
50.00
550.00
500.00
63.46
144.337
.000
287.228
288.964
37.980
- 183 -
5.3.5 Graphs
The previous sub-section gave details of the simulation experiments. In this section, we
use graphs as a means of evaluation of results obtained from the above two set of simula-
tion experiments.
To start with we first plot a 95% Confidence Interval plot of the Sensed value with a
variation in the number of Boid agents as shown in Figure 60. Looking at this graph, we
can see the bigger picture since we can note here that there is actually a variation in the
sensed value with the number of Boids. As the number of Boids increase, the Sensed value
also increases verifying that the model for sensing is working correctly. However, it should
also be noted that the variation in the Sensed value is not linear. This nonlinearity is essen-
tially a validation of the hypothesis discussed earlier in this chapter regarding the case
study noting there is a high level of intrinsic complexity in the mixing of two different cas
systems to form a heterogeneous cas model. The graph used is an error graph. The fact that
each point is actually having a very small set of high and low values shows the fact that
this is a consistent phenomenon from this particular case study.
In this graph, we noted how the Sensed value varied with a variation in the number of
Boid agents. Next, we want to evaluate this phenomenon further. So, we plot the mean val-
ue of Sensed vs. simulation time.
- 184 -
Figure 60: 95% Confidence Interval plot of sensed vs. number of boids
However, to ensure that we are not looking at a particular set of instances of the simula-
tion and that the hypothesis is valid for a large variation of the simulation experiments, we
plot mean values for different number of Boid agents as shown in Figure 61.
Here, the first thing we can note from this chart is that in general, with the passage of
time, as the Boids tend to flock over time, the sensed value keeps on decreasing. In other
words, the hypothesis that we developed earlier while examination of the visualization of
the simulation discussed earlier about sensing is thus proven valid here. The second obser-
vation we can note here is that the variation in the Sensed value over time is more
considerable when the number of Boid agents is smaller. With an increase in the number of
Boid agents, due to a constraint of the simulation space limitation, the variation in the
Sensed value appears to be small but noticeable. So, we can infer another hypothesis that
- 185 -
even with a large number of Boids, flocking can still be detected with a relatively high de-
gree of sensitivity.
Next, we note the interesting perturbations in the simulation experiments with the num-
ber of Boid agents visible more clearly at 500, 600 and 700 at time steps from 350 to 700.
These interesting nonlinear effects are again a demonstration of the internal nonlinearities
of this heterogeneous cas system. Thus, while emergent behavior of flocking is observable
and quantifiable at the global scale from this graph, there are significant small nonlineari-
ties which might not be observable if only that section of the graph was to be used as a test
case.
Figure 61: Mean value of sensed over time with different number of boids
This also demonstrates the effectiveness of the use of agent-based modeling in general
and the particular DREAM methodology in particular, showing how it allowed for a detec-
- 186 -
tion of complex global emergent behavior based on nonlinear interactions of a large num-
ber of components.
In the previous two graphs, we evaluated the effects of a variation of the number of
Boid agents in the simulation. Next, we focus on experiments where the number of Boids
is constant but the number of sensor Node agents varies. While the previous set of experi-
ments gave consistent results testifying to the testing and validation of several hypotheses
relating to sensing, we next want to verify and infer new hypotheses regarding the number
of sensors. In other words, we want to validate that the previous results were valid in the
case of a large variation in the number of sensor Node agents and were not a manifestation
of the particular setting of the number of sensor Node agents.
Figure 62: 95% Confidence Interval graph of Sensed with a variation in the number of sensors
- 187 -
The first graph that we evaluate here is a 95% confidence interval error graph shown in
Figure 62. This graph shows a straight line. In this case, the straight line is a very good in-
dication because it shows that the results of the Sensed value from the previous simulations
are all valid. The change in the number of sensors does not affect the results and thus the
results were all consistent and valid.
The next graph is useful as an extension of the previous graph. In this case, we plot the
mean values of Sensed against simulation time. We also plot multiple number of sensors in
the same graph so that we can see the net effects of sensor Node agents in the simulation.
The important hypothesis here to test is whether or not the hypothesis regarding sensing
nodes holds as sensor nodes are losing their battery power. Thus we note here that the vari-
ation of Sensed value still holds over time.
Figure 63: Mean Sensed value vs. simulation time with a variation in the number of sensors
- 188 -
What is really interesting here is the fact that even with a relatively small number of
sensor Node agents (100), the flocking can still be detected or "sensed" thus further vali-
dating the hypothesis that the binary sensors can actually behave intelligent collectively
and detect flocking behavior of the Boid agents. However, we can also note that if the
number of sensor Node agents increases, the variation in the initial Sensed value to the fi-
nal value is more pronounced. In other words, this helps verify the simulation further that it
is working correctly and performing the expected behavior.
5.3.6 Significance of findings
The significance of the results of these simulation experiments can be noted by the number
of possible applications as a result of developing a better understanding of this hybrid cas
system. In short, if randomly deployed sensors can be used to detect apparently distributed
mobile flocking, then some possible applications can be given as follows:
1. Sensing can be used for the identification of collective behavior of people using an
image processing or infrared approach. In other words, this technique can be used
to detect any group of people such as a group of "shoppers" flocking to a particular
set of shops in a mall even though there is a large crowd of thousands of other peo-
ple moving around.
2. This same approach can be used to detect a group of malicious attackers moving
through a crowd thus allowing for the detection of the group "flocking" without
cohering too much as that close grouping might have raised suspicions of the secu-
rity personnel.
3. Another application of this approach is in the domain of detecting a group of stealth
aircraft flocking towards a common mission. Stealth aircrafts are known to be in-
visible to radars in certain conditions. However these aircrafts are physically visible
- 189 -
to the naked eye or in the visible light. As such, while it is difficult to infer a coor-
dinated operation based on only the detection of a single aircraft, if there were a
randomly deployed group of sensors with cameras or other sensing equipment
pointing to the sky, these wirelessly connected sensors could essentially detect a
group of stealth aircraft flocking and moving towards a common goal. Using sim-
ple shape detection image processing algorithms, each sensor can simply note if a
"large" object passes over them. And this detection of "proximity" could be used to
quantify flocking at a monitoring station collecting the data from these sensors.
5.3.7 Critical Comparison of DREAM with other approaches
The previous sections were related to discussion of the simulation experiments and re-
sults of the case study. Here, we relate the DREAM approach in the bigger context of
describing agent-based models.
As discussed earlier in this chapter, DREAM approach is based on an informal method-
ology which allows for conversion between a complex network model, a digital footprint
based on centralities and a pseudocode-based specification model. This conversion is es-
sentially tied in with an informal validation process which can be followed by means of
face validation, code walkthroughs and testing.
While, to the best of our knowledge, there is no other well-known approach in literature
which describes agent-based models descriptively and quantitatively similar to our pro-
posed DREAM approach, here we perform a comparison of the approach to two somewhat
related approaches of knowledge transfer and description of agent-based models:
5.3.7.1 Template Software approach
A common approach to sharing models, as noted by Santa Fe Institute's Agent-based
modeling software Swarm's wiki page, is the template software approach[130]. This is ba-
- 190 -
sically a code-based approach for comparison of models. The way this works is that tem-
plate software for NetLogo, Swarm, Repast and MASON etc. are publicized on the
website. Interested researchers download and adapt the source code for their particular us-
age. The website defines template software as code which has been designed with the
following features in mind:
- For use in new models by means of copying and modification.
- As a means of tutorial-based teaching/learning.
- As examples of programming tasks.
-
In some cases, for comparison of various ABM tools/software.
As such, it is clear here that templates are useful but, each of them is firstly based on
code and thus specific to single platforms/languages. They also are not suitable for non-
code or quantitative comparison of models, as proposed via the DREAM approach. The
benefits of moving away from pure code to other models which do not require detailed
knowledge of programming is clearly obvious here since at times, the designer of the simu-
lation (Simulation Specialist) might be a different person than the cas researcher.
5.3.7.2 ODD approach
Grimm et al. have proposed the Overview, Design concepts, and Details (ODD) proto-
col which is a text-based descriptive approach for documenting agent-based models. This
approach has been demonstrated in ecological and social sciences [199]. The initial ODD
proposed protocol[101] was subsequently updated and is now known as the "Updated
ODD protocol". As compared with the proposed modeling approach, there are certain key
differences given as follows:
Firstly, the description in ODD is textual. In other words, it can be considered as a way
of describing agent-based models using a checklist of proposed headings. As such, it does
- 191 -
not allow a quantitative comparison such as the use of Complex Network methods as
demonstrated in the proposed modeling methodology.
1. ODD does not offer a visualization-based approach where researchers can visu-
alize an entire ABM to compare it with other ABMs.
2. ODD also does not offer a one-to-one translation of models from the actual
model to the source code. The proposed DREAM methodology allows this by
means of an informal translation process.
3. ODD also does not allow for detailed descriptions of algorithms.
Being textual and not using any pseudocode, flow charts, state charts or other algorithm
description models, it is difficult to see how a model described using ODD can be repro-
duced exactly. As Edmond and Hales [200] have previously noted that ". . .the description
of the simulation should be sufficient for others to be able to replicate (i.e. re-implement)
the simulation". Other problems with purely textual descriptions of agent-based models
have been noted by Wilensky and Rand[145] who describe the difficulties in the replication
of results based on textual description of Axelrod. It is pertinent to mention here that Axel-
rod is one of the pioneers of agent-based modeling in Social Sciences [25] with
publications in the domain dating back to the 1980's[201]. Wilensky and Rand note that
"In the replication experiment detailed herein, differences were discovered between the
Wilensky-Rand model and the Axelrod-Hammond model, and the replicated model had to
be modified to produce the original results. Moreover, the process that was required to de-
termine that the replication was successful was complicated and involved unforeseen
problems." Thus to summarize, while textual descriptions are effective in giving basic ide-
as about a simulation, it is difficult to directly implement these ideas without the use of
pseudocode as proposed by the descriptive ABM methodology in this chapter.
- 192 -
5.4 Conclusions
In this chapter, we have proposed DREAM, an extension of exploratory agent-based
modeling. DREAM allows for quantitative, visual and pseudocode based comparison and
description of agent-based models. We have described a detailed case study as a means of
demonstrating the effectiveness of using DREAM for modeling and simulation of various
cas problems and inferring hypotheses. We have also critically reviewed DREAM with
other prevalent approaches in the domain of agent-based modeling.
In the next chapter, we complete the last part of the proposed unified framework by de-
veloping a methodology of Verification and Validation of agent-based models using in-
simulation validation. The proposed Virtual Overlay Multiagent System (VOMAS) ap-
proach is demonstrated by application on three different and separate case studies
demonstrating its effectiveness in a number of multidisciplinary application case studies.
- 193 -
6 Proposed Framework: Validated Agent-based
Modeling Level
In the previous chapter, we explored Descriptive Agent-based Modeling. DREAM
allowed the developing of a non-code description of the ABM using a combination of a
complex network model, a quantitative "footprint" of the ABM as well as a pseudocode-
based specification model. However, while it solved some of the key methodological prob-
lems of ABM design, the proposed framework would still be amiss, if it did not allow for a
methodology of developing "trustable" models. Trust in any simulation model is arbitrary
because the fact of the matter is that a simulation model, however well-designed, is only as
good as its correlation with the real-world or "imagined"7 phenomenon, which it is ex-
pected to map. So, quite simply, a simulation model needs to be first verified so that it is
working as expected and then validated to ensure it models the simulated phenomena close
enough at the required level of abstraction as dictated by the particular requirements of
simulation. However, as discussed earlier in Chapter 2, this particular type of verification
and validation in the case of agent-based modeling cannot be considered in the same light
as other more traditional simulation models. Next, we give a detailed description of the
problem statement of verification and validation of agent-based modeling in the next sec-
tion.
7 Imagined such as in computer games and entertainment software, where the simulation is specifically de-
signed not to reflect the real world.
- 194 -
6.1 Problem statement
As discussed earlier in Chapter 2, while the problem of verification and validation is
applicable to any modeling and simulation approach, in the case of agent-based modeling
of complex adaptive systems, there are certain peculiarities which make their validation
different from that of traditional simulation models:
- The simulation design might not be required to develop a quantitative validation
model because the phenomena might be observable and not easily quantifiable. In
other words techniques such as Statistical validation or Empirical validation[140]
used alone, might not always be the most appropriate choice for validation [144].
- Fagiolo et al.[202] note that ABMs have characteristics such as bottom-up model-
ing, heterogeneity, bounded rationality and nonlinear networked interactions which
make them especially difficult to validate in the same way as other simulation
models. In addition, they note that two of the key problems in ABMs are "the lack
of standard technologies for constructing and analyzing agent-based models" and
"the problematic relationship between agent-based models and empirical data".
In this chapter, we propose a solution to these problems by proposing a generalized
methodology for in-simulation verification and validation of agent-based models using a
Virtual Overlay Multiagent System (VOMAS). Our approach is based on existing ideas of
verification and validation from different Agent-based Computing disciplines such as So-
cial Sciences, Software Engineering and Artificial Intelligence to develop a validation
approach which can be customized for application to multidisciplinary cas case studies. As
a means of demonstration of the generalized applicability of our proposed approach to var-
ious cas domains, we present three separate example case studies from ecological
modeling, WSNs and social sciences.
- 195 -
The rest of the chapter is structured as follows: First we give a description of the gener-
alized methodology of in-simulation validation. Next, we present experimental design for
the development of VOMAS for three example case studies. Next, in the results and dis-
cussion section, we present the simulation results followed by a detail discussion of
benefits of using the VOMAS methodology in the validation of the particular case study
model. We also critically compare the VOMAS methodology with other validation ap-
proaches of ABMs. Finally we conclude the chapter.
6.2 Proposed validate agent-based modeling methodology
In this section, we develop the generalized VOMAS methodology. To start with, we first
examine how the problem of verification and validation of computer simulations is consid-
ered by simulation experts.
A generalized view of the verification and validation process of computational simula-
tions has been presented by Sargent[203] as shown in Figure 64. This diagram
encompasses the conceptual model as well as the computerized model. As we can note
here, the validity of the conceptual model is performed by means of analysis and modeling.
Whereas from the conceptual model to the computerized simulation model, programming
and implementation can be evaluated by performing computerized model verification.
However, to ensure that the computerized model is a true representation of the actual prob-
lem, a set of thorough experimentation is needed since it allows for a confirmation of the
operational validity. Thus in the case of ABMs, the proposed methodology has to allow for
a flexible means of performing computerized model verification as well as operational and
conceptual validation.
- 196 -
Figure 64: A view of the modeling process, figure adapted from Sargent [203]
The validated agent-based modeling level of the proposed framework can be noted in
Figure 65. In this level, the development of the agent-based model is followed similar to a
Software Engineering approach. As such, the SME and the SS work together in teams to
develop requirements for the validation of the proposed model by determining invariant
contracts to be added to the simulation. These invariant contracts allow the model to be
customized for any validation paradigm and thus can be used based on the particular case
studies. Based on these meetings, the SS develops the cas ABM by adding a Virtual Over-
lay Multiagent System (VOMAS) which allow in-simulation validation of the agent-based
model using the invariant contracts.
- 197 -
Figure 65 Validated agent-based modeling framework level
6.2.1 Detailed overview of origins of VOMAS concepts
The proposed VOMAS-based methodology has basis in three different set of basic ide-
as. These ideas originate from cooperative multiagent systems, software engineering and
social sciences and are noted below as follows:
- 198 -
6.2.1.1 VOMAS basis in Software Engineering Team concepts
The temporal and interactive structure of building a VOMAS ties in closely with how
Software Engineering projects are constructed because the Software Engineering discipline
is entirely focused on developing software based on teams of various sizes. The idea stems
from the fact that simulations are also pieces of software however, to the best of our
knowledge; there is no standard methodology which uses Software Engineering practices
in the development of agent-based simulations. The VOMAS approach is thus structured
on the basis of software engineering and requires the Simulation Specialist (SS) to work
with the domain exert end user researcher or Subject Matter Expert (SME) in large scale
research simulation projects. This leads to an iterative approach where the SS develops dif-
ferent models and diagrams (e.g. using DREAM approach) and gets approval from the
client SME to proceed in each step.
A typical software development team can consist of various roles such as Manager,
Software Engineer, Analysts and Quality Assurance Engineer. While one or more team
members might have the same hat or role, or might have different roles depending upon the
stage in a particular project, all work together with the end-users to develop a piece of
software in line with the user requirements[204].
In the proposed methodology, the end users are nontechnical but multidisciplinary re-
searchers who are experts in their own domains. Thus while they are not assumed to have
advanced knowledge of modeling and simulation domains, they can intelligently work with
the SS to develop and test the simulation models. While they might not correctly identify
exactly what features are to be instilled in the software, they might be able to give con-
structive feedback on exactly how the simulation model can be proven correct in terms of
their domain-specific terminology. Thus an ecologist would be able to look at a particular
fire model and affirm that this is exactly how fires propagate in a particular area and also
- 199 -
note as to what would be needed to make a model be realistic. In addition, a Biologist re-
searcher investigating precancerous tumors might be able to observe at a tumor growth
model and affirm if the model is looking close to the implicit mental model well-known to
biologists and exactly what should the SS be looking for in an ABM tumor model for vali-
dation. Therefore, by means of using different models such as those earlier proposed in
DREAM modeling, the end user researcher SME can gradually becomes comfortable with
the software concepts helping the SS in developing a simulation model which qualifies the
user requirements.
6.2.1.2 VOMAS basis in software engineering concepts
Traon et al. [205]note that Meyer's "Design by contract" concept[206] allows for easily
including formal specification elements (referred here specifically to software-based invar-
iants, pre-conditions and post-conditions) into a software design. Using these built-in and
programmed contracts, software engineers use these software constructs as "embedded,
online oracles". The proposed methodology uses the design by contract concept. It extends
the idea to the domain of agent-based simulation models of cas. It allows for building in-
variants based on pre- and post-conditions inside agent-based models by having VOMAS
agents inside the simulation model of cas which dynamically observe the simulation and
provide visual and log-based feedback to the SS. An invariant I is technically defined as
follows:
"If a set of pre-conditions
holds in an ABM M, then M is guaranteed to undergo a
state change which will result in the post-conditions
."
Thus periodically extracted information from the simulation model can then be used to
evaluate different invariants, as and when required by the particular case study under guid-
ance from the SME. Thus, unlike the design by contract model of software design, where
- 200 -
preCpostC contract becomes an integral part of the software model, these agents need to be designed
to use minimal system resources. VOMAS agents can thus be designed to act on their own
to periodically evaluate the simulation state and results.
As a consequence of invariants, we can also define an exception E as follows:
"If pre-conditions
do not hold in an ABM M, then M is not guaranteed to undergo
any particular state change of interest to the SME"
Thus, such an exception behavior can be noted by means of execution of the simulation
model. The reduction in the number of exceptions generated from the simulation experi-
ments is thus an indicator of the improvement in the validation of the model. In addition,
exceptions can be used to validate the Invariant itself. This can be performed by means of
limiting all the conditions in a model to a specific set thus allowing for a focus on only re-
moving just one of the pre-conditions. Now, with different variations of other inputs,
exceptions can be noted. Next the pre-condition can be inserted and this would allow for
the validation of the Invariant by a complete removal of the exceptions. This is shown
graphically in the following Figure 66. We can note here the process of designing and test-
ing Invariants by means of changing pre-conditions in the proposed methodology.
- 201 -
preC Set Pre-Conditions
Observe Post-
Conditions
Remove one
Pre-Condition
Implement
Invariants
Design
Invariants
Observe
Exceptions
Figure 66: Design and testing of VOMAS Invariants by means of manipulation of the pre-conditions
6.2.1.3 VOMAS basis in multiagent system concepts
Panait and Luke [95] note that cooperative multiagent systems are a common paradigm
in multiagent models where different agents work together to a common goal. Our pro-
posed methodology has its third basis in the concept of cooperative agents. Thus, SME
can use any suitable combination of agents inside a ABM which satisfy the requirements of
validation. There are two possible extremes of this development. Either the developed
agents are part of the ABM themselves or else the agents are developed separately and
constitute agents observing the simulation from within simulation. The actual idea is to be
able to facilitate the SME in developing the validation model and would thus vary on a
case-to-case basis.
In other words, a VOMAS model built for one application domain or even one aspect of
an application domain could vary considerably from another VOMAS model built for a
different purpose. However, one primary design goal of VOMAS agents has to be so as not
to disturb the results of the simulation model since simulation results can be are very sensi-
- 202 -
tive to computation. If the computation performed by the VOMAS agents is too extensive,
then it might end up disturbing the entire simulation. Thus care needs to be taken in the
development of the in-simulation validation model so that the agents periodically report
results without affecting the actual simulation. In addition, VOMAS agents must also use a
minimal of system resources to extract information. The diagram in Figure 67 gives details
of how the overall proposed in-simulation methodology is structured.
Develop Model
Design (DREAM)
Validate using
VOMAS
Develop Validation
Questions
Meet with
SME
Execute Simula-
tions
Develop Invari-
ants
Develop in-
simulation
VOMAS
Figure 67: Methodology of building In-Simulation Validated ABMs
We can note here that the central activity of the proposed methodology is communica-
tion and collaboration with the SME. The methodology here is given from the perspective
of the SS. As such, every step of this methodology is linked with meeting and communica-
tion with the SME. How it can actually be made possible is by ensuring that the
methodology starts when the SME and SS meet and start to plan the simulation-based re-
search project. Collaboratively, the two teams of multidisciplinary researchers i.e. the
SMEs and the designers of the ABMs i.e. SS formally meet and develop the design of the
- 203 -
ABM e.g. using DREAM. Next, they discuss what would be good validation questions.
The validation questions should be succinct assisting the team to develop invariants along
with any needed pre- and post-conditions.
The next step in the methodology is to develop invariants along with the suitable pre-
and post- conditions. Finally the decision is to be made as to what are the specific agents
which will be part of the simulation. The questions at this stage depend on the particular
case study and would vary in different cases depending upon the level of computation re-
quired by the agents. So, in some cases the agents can be embedded in the actual
simulation while in other cases, the agents can be observing the simulation and validating
the invariants. Subsequently, the model is developed and gradually the invariants are added
to the ABM.
After adding each invariant, the model can be verified using the method for testing in-
variants using VOMAS presented earlier in Figure 66. After the basic testing, the
invariants can all be considered working and next the simulation experiments can be exe-
cuted. The results can subsequently be used by the team to evaluate the effectiveness of the
ABM in representation of the real-world phenomena
6.2.2 A Taxonomy of Agent-Based Validation techniques using VOMAS
In this sub-section, we examine some of the possible methods by means of which agent-
based models can be validated by means of using VOMAS as shown in Figure 68. Since
agent based models traditionally have one or more agents, what the physical implication in
the real-world is entirely up to the designer i.e. the SS. Agents act as elements or pieces in
a chess board and thus can be, at times, placed spatially in the simulation, where distance
between agents in the simulation is important either for reduction of clutter or else for a
concept of "distance" modeled by the designer. In other cases, non-spatial concepts could
- 204 -
be used in the model since spatial concepts might not make sense in that particular applica-
tion case study or discipline. In case of spatial models, it is also entirely possible that the
exact distance may not be important, but the links between agents could be important. A
detailed description of each of these follows.
In-simulation V&V of ABM
Non-Visual
Visual
Non-Spatial log-based
Spatial
Network-based
Distance-based
Placement based
Figure 68: A Taxonomy of Agent Based Validation techniques
1. Visual Validation:
Visual validation is related to the face validation technique[203] where the model
animation or results can be validated by the SME to see if the behavior is similar to that
expected in the actual system.
2. Non-visual Validation:
Non-visual validation is a generalized methodology of validation which would be
anything that is unrelated to purely face validation concepts.
a. Spatial Validation:
- 205 -
In spatial validation, the placement of agents in the simulation is im-
portant. Thus either the agents are connected based on a network or else it is
that the agents are place on a 2-D or 3-D plane and their distance or placement
is important in terms of usage in the particular case study.
b. Non-Spatial Validation:
In non-spatial validation, the validation can be performed by means of de-
veloping extensive logs. These can ensure all data is stored for future usage.
While this kind of validation allows for data storage, it also can result in some
problems. Firstly, the size of the logs might grow considerably. And with size,
the time to store the log might interfere with the simulation time. In addition,
data might be too difficult to analyze due to significant size of logs.
6.3 Case studies: Experimental Design and Implementation
In the previous section, we developed the general methodology of in-simulation verifi-
cation and validation of ABMs. In this section, we present the overview, experimental
design and implementation details for three separate case studies from different disciplines
to demonstrate the generalized applicability of the proposed validation methodology and
framework. The first discipline is in the domain of ecological modeling where we develop
a forest fire model for evaluation of the effects of forest fires on the overall forest structure
as well. The second area is in the domain of WSNs, where we develop a specialized com-
plex network model of multi-hop WSNs for observing the forest fire effects demonstrating
the applicability of the proposed methodology in different ways. The third model is from
the domain of quantitative social sciences in the domain of developing a complex network-
based model of researcher evolution over time by means.
- 206 -
6.3.1 Case Study I: Ecological modeling of forest fire spread and forest re-growth
In the first case study, we explore the use of VOMAS in a Cellular Automata (CA)
model of a forest fire simulation. To start with, we first evolve a basic and subsequently an
advanced CA model of the forest fire spread. Next, validation questions and subsequently a
VOMAS model for cross-model validation is developed using the Fire Weather Index
(FWI), a standard mathematical model for forest fire prediction. In the next section, we
describe in detail the model and experimental design.
6.3.1.1 Basic simulation Model
We start by developing a forest fire model. In the real world, a forest fire spread can be
based on different parameters such as:
1. The cause of the fire
2. The wind speed and direction
3. The amount of rain and snow in the past
4. The humidity level
5. The current rate of rain or snow
6. The particular type of Trees
7. The structural aspects of the forest
To develop the basic forest fire model, we thus need to develop several modules incre-
mentally as given in the next sub-sections.
6.3.1.1.1 Forest Creation Module
In this module, our goal is to develop the forest scenario. Our basic model extends upon
the Forest Fire model given in NetLogo model library[207] which is a simple model with
- 207 -
no advanced parameters for configuration or any means of realistic forest fire simulation.
As such, we first develop a forest made up of trees covering a random set of locations.
However, to ensure we can control the tree coverage, we give the tree coverage model a
probability (Pcov). Based on the value of Pcov, which varies from 0 to 100, we can build an
environment model with trees. With a Pcov value of 68%, we can see the forest model as
follows in Figure 69. We can also note here that the wind direction is made visible by
means of a dial at the top.
6.3.1.1.2 Basic Fire module
After the forest has been created, we can develop a baseline fire model. The way
the fire module works is that at a given random probability of fire (Pfire), a fire can start at a
random location. Now, the fire has three distinct states:
o Started
o Spreading
o Dying
In the "started" state, fire has just been initiated. After the fire has been created, it is
given the direction of spread based on the "Wind direction" parameter. After creation, the
fire next moves to the "Spreading" stage. In the spreading stage, we use a parameter "tic".
This parameter is initially zero but as fire spreads, this parameter gradually increases. Thus
depending upon the parameter "intensity", the fire keeps growing up until it reaches the
"intensity" which serves as a threshold to stopping the fire. The "Dying stage" gradually
serves as a means of stopping the fires and is reflected by a gradual change of color.
- 208 -
Figure 69: View of the forest before the fire
The effects of a fire can be seen in Figure 70. We can note here that the edges of the fire
are a different color than the center of the fire because it represents that the trees have been
destroyed in the center by now.
- 209 -
Figure 70: Forest fire spread
In addition to being able to simulate single fires, the baseline model can support more
than one fire as shown in Figure 71.
6.3.1.2 Advanced Forest Fire model version 1
While the basic or baseline forest fire model gives an interesting simulation, it does not
give any realistic fire results except the use of wind direction. We next develop an ad-
vanced forest fire model. In the advanced forest module, we want to simulate long term
fires so we want to model the effects of tree re-growth other time. In other words, while
forest fires come and destroy the forest, periodically the tree cover grows as well. So, to be
able to evaluate the advanced patterns formed due to the effects of various fires as well as
the tree re-growth, we develop tree re-growth capability.
- 210 -
Figure 71 Baseline forest fire model with two consecutive fires
Based on a re-growth percentage and a re-growth period, our simulation can be devel-
oped which can give rise to more advanced effects in forest ecological simulation. As an
example of tree re-growth based patterns with three different random forest fires, we see
the pattern as shown in Figure 72.
- 211 -
Figure 72 Advanced forest fire model version 1 with tree re-growth
6.3.1.3 Advanced Forest Fire model version 2
To make the forest fire scenario more realistic, we add several different scenarios and
parameters. The first of these is the effect of rainfall as shown in Figure 73. The effects of
rainfall are that the humidity of the location of the forest where rain is occurring is in-
creased by a random value over time.
- 212 -
Figure 73 Advanced forest fire model version 2 with rainfall in part of the forest
In addition, we allow for coupling the effects of forest fires along with snowfall as
shown in Figure 74. The physical impact of the snowfall is that the humidity as well as the
temperature of the location is affected over time e.g. with an increase in humidity level and
a decrease in temperature.
- 213 -
Figure 74 Advanced forest fire model version 2 with snowfall effects
6.3.1.4 Validation Question
Up until now, we have developed the advanced forest fire model which contains forest
creation, forest re-growth, weather effects simulation and forest fire simulation modules.
The goal of this section is to demonstrate how this model can subsequently be validated by
means of building an associated VOMAS model. Thus, while the entire model can be de-
scribed in a real case study in the form of a DREAM model described earlier in the last
chapter, here we only focus on the validation part. As described earlier in the VOMAS
Methodology, we shall start by first writing our validation question in the model. Our key
validation question can thus be stated as follows:
- 214 -
Question: How can we validate that the developed forest fire model is representative of
actual forest fires?
Now, the way this could be practically answered is by looking for either a repository of
data for forest fires or else a standard forest fire model. By performing an examination of
previous literature such as [208] by Alexandridis et al., it becomes clear that while there is
data such as either the temperatures of forests or else data in terms of the final results of
fires, it is very difficult to get any data of the exact structural aspects of forest fire spread
over time. In other words, the data availability is limited to Geographical Information Sys-
tems (GIS) based data visible via satellites. This data does not demonstrate the micro-
dynamics of the forest fires. As such, instead of this approach, which would not have guar-
anteed model validity, we have to use a different approach. One possible approach in the
absence of real data is thus to perform cross-model validation or "docking" of the models
as noted by Axtell et al. in [209]. As such, here we use this approach but in contrast to Ax-
tell's approach, which was to develop two different agent-based models, our ABM which
will be validated against a formal mathematical model of the FWI. FWI is a standard mod-
el in use by several countries including France and Canada. In France, the index has
previously been calculated and gradually updated for the last 40 years and similar practices
are prevalent in Canada as noted by Lawson et al. in [210]. Being such a stable Mathemati-
cal model thus allows FWI to be a suitable candidate for use in model validation.
The FWI is calculated using weather values input in a complex set of equations as noted
in [211] allowing prediction of forest fires[212]. In other words, FWI can be used to per-
form cross-validation of our simulation model.
- 215 -
6.3.1.5 Software Contract
The next step in our proposed methodology is to develop software contracts for the
simulation model. As such, based on the validation question that we have discussed in the
previous section, we can focus on the following contract:
Invariant Contract: If the pre-condition of a "tree is on fire" is true, then a post-
condition of "FWI value in the range of Fire Danger index" is true.
Thus, we need to be able to validate the FWI value in the simulation. Next we design
the VOMAS agents to cater for this validation.
6.3.1.6 VOMAS agent design
A naïve way of designing a VOMAS could be to have each cell calculate the FWI value
over time. However, practically speaking there is a problem in this implementation because
the FWI calculation requires calculation of almost 40 Mathematical equations (as we shall
examine next), a number of which require complex mathematical calculations such as ex-
ponential etc. So, if we were to execute a simulation that has an environment of only
and every cell of the simulation calculates the FWI in each simulation run, then
it would imply that for each simulation second, these equations would require
calculations per simulation second. In other words, such an imple-
mentation would itself result in an overhead that might invalidate the entire basic
simulation. As such, this would defeat the whole purpose of the forest fire simulation. As
an alternative to this design, we thus propose using a random number of trees as VOMAS
agents which calculate the value of the FWI and validate the value against fires. In other
words, any trees, which have a FWI value in the range of fire danger index, might have
fire. In addition, any trees having fire should have an FWI value in the danger range.
- 216 -
30030090000403600000 The actual calculation of FWI is based on items called the "Blocks". The blocks are
given as follows in Figure 75. A brief description follows of the different components is
given as follows summarized from [1], [213].
Block 1 Fine Fuel
Moisture Code
FFMC
Block 2 Duff
Moisture Code
DMC
Block 3 Drought
Code DC
Block 4Initial
Spread Index ISI
Block 5 Adjusted Duff
Moisture Code ADMC
Block 6 Fire
Weather Index
FWI
Figure 75: FWI calculation, figure adapted from Wagner[1]
6.3.1.6.1 Fire Weather Index (FWI)
FWI [1] provides the assessment of the relative fire potential based on weather
conditions. It depends on several sub-components. It is primarily calculated from the Initial
Spread Index (ISI) and Build Up Index (BUI) to provide an estimate of the intensity of fire
spread. In general, FWI can be considered to indicate the fire intensity based on a combi-
nation of an expected rate of fire spread along with the total amount of fire fuel consumed.
Next, we discuss the sub-components needed to calculate FWI.
- 217 -
6.3.1.6.2 Fine Fuel Moisture Code (FFMC)
FFMC represents the moisture contents of litter and fine fuels, 1-2cm deep, with a
typical fuel loading of about 5 tons per hectare. FFMC indicates the ease of ignition or in
other words, it formalizes the ignition probability.
6.3.1.6.3 Duff Moisture Code (DMC)
DMC is an indicator of the moisture content of the duff layer, which is essentially a
layer of compacted and organic matter in decomposing state. It predicts the probability of
fire ignition as a result of lightening. It is also representative of the rate of fuel consump-
tion.
6.3.1.6.4 Drought Code (DC)
Drought Code is an indicator of average moisture content of the deeper layers of
compacted organic matter. DC is indicative of long-term moisture conditions and deter-
mines fire resistance to extinguishing. It is also an indicator of fuel consumption.
6.3.1.6.5 Initial Spread Index (ISI)
ISI indicates the rate of fire spread immediately after ignition. In combines the
FFMC and wind speed to predict the expected rate of fire spread.
6.3.1.6.6 Build Up Index (BUI)
BUI index is a combination of the DMC and DC codes and it indicates the total
amount of fuel available for combustion. DMC has the bigger influence on the BUI.
- 218 -
6.3.1.6.7 Calculation of FWI
The details of the calculations pertinent to our simulation implementation are discussed
below however further details of these parameters are out of scope of and interest of this
thesis in general although they are publicly available in [211].
Block 1 is the Fine Fuel Moisture Code (FFMC) and is calculated based on the rainfall,
relative humidity, wind speed and the temperature as follows (F):
Block 2 is the Duff Moisture Code (DMC). It is based on the rainfall, relative humidity
and the temperature. DMC (P) is calculated as follows:
- 219 -
6.93/100/(251)100/(251)6.93/20.679(100)/100.115147.2(101)/(59.5)0.542.5()(1)42.5()(1)0.0015(150)0.50.942110.18(21.1)(1)0.ofootormoromorfofHHdwFFMCmFFrrmmrfeemrmrfeemorEHeTeE0.753(100)/100.1151.70.581.70.580.0365618100.18(21.1)(1)0.424[1(/100)]0.0694[1(/100)]*0.5810.03650.424[1(100/100)]0.0694[1(100/100)]1*0.581()*10HHodlTdodHeTeKHWHKkoeTkHWHKwkemEmE()*1059.5(250)/(147.2)KdKwwwomEEmFmm
Block 3 is the Drought code (DC) and is based on rainfall value and the temperature.
DC (D) is calculated below as follows:
Next, we note the Block 4, which represents the Initial Spread Index (ISI). ISI (R) is
calculated as follows based on the FFMC calculated earlier:
Next, we calculated the Build-up index (BUI) or the Adjusted Duff Moisture Code
(ADMC) represented as U is calculated next as follows:
- 220 -
(5.6348/43.43)6r0.921.2720100/0.50.31413ln6.2ln17.21000/(48.77)244.7243.43ln(20)1.894(1.1)(100)*10(P)100eoPoooooroeroDMCrrMebPbPbPMMrebrPMrkTHLePPorK/4000.831.278003.937400ln(800/)0.36(2.8)()0.5dDoorrDCrroQeQQordDQrVTLfDDoorDrV0.08110.13865.317()()91.9[1/(4.93*10)]0.208()()WmISIfWefFemRfWfF1.70.8/(0.4)[10.8/(0.4)][0.92(0.0114)]BUIUPDPDUPDPDP Finally these values are used to calculate the FWI as follows:
As can be seen, these calculations are necessary for the accurate calculation of the FWI
value in our case study. Thus, VOMAS agents will be performing these calculations as re-
quired. The results of the VOMAS experiments of the Forest Fire Simulation case study
will be further discussed in the results section.
6.3.2 Case Study II: Environmental Monitoring using Wireless Sensor Networks
In the previous case study, we developed a Forest Fire Simulation case study. In this
second case study, our goal is to extend the previous case study and use a set of connected
wireless sensor network nodes to monitor the forest fires. As a means of getting to this
stage, we first need to develop a formal model for a randomly deployed WSN by choosing
one of the standard models.
Technically, a connected WSN is represented as a Graph or Network
. Where
V is the set of vertices and represents the sensor nodes while E (edges) represents the con-
nectivity of the nodes. Thus for any two nodes
,
if the sensor node u is
adjacent to v. The classical connectivity model for WSNs is the Unit Disk Graph
(UDG)[214] where if any two nodes are considered adjacent if and only if their Euclidean
distance is at most 1. The UDG [215] has been considered as unrealistic connectivity mod-
els. It also does not cater for the facts such as that real world sensors might not have Omni
directional antennas and also that any small obstacles might result in disconnection[188].
An advanced version of the UDG is the General Graph (GG) model which is a general un-
- 221 -
0.8090.0230.647()0.6262()1000/(25108.64)0.1()ln2.72(0.434ln)UfFWIfDUfDebRDSB(,)GVE,uvV(,)uvE directed graph. As such, every node can be considered as an adjacent to every other node.
However, this model is again unrealistic because it can be too generous and thus it might
result in nodes which cannot be reached from a node having connections. As such, both
these models are considered as two extremes for realistic models. However, our proposed
simulation model for WSNs follows a flexible realistic version of network models which
has been considered as a reasonable model between the two extreme modeling paradigms
called the Quasi Unit Disk Graph (QUDG) proposed by Kuhn et al.[216]. Thus our WSN
model would be formally defined as follows:
Our proposed WSN model consists of nodes in a two dimensional Euclidean plane in
. All pairs of nodes with Euclidean distance
for some given
are considered
adjacent. In addition, pairs of nodes between distances
and 1 may or may not be neigh-
bors. It is a well-known fact that for
, QUDG becomes a UDG[188]. As such,
building a QUDG using the NetLogo interface, we can see the simulation output shown as
follows in Figure 76.
Next, the models for forest creation, forest fire and the WSN simulation are integrated
resulted in a single model. This sensor model is able to extract useful information such as
temperature and other parameters from the location of the sensor inside the simulation en-
vironment. With a description of the basic connected WSN model, we next move to the
validation question part of the proposed methodology.
- 222 -
2(0,1]1
Figure 76: WSN network simulation as QUDG with n=200 sensor nodes
6.3.2.1 Validation Question
The simulation in this case study is based on a set of sensor nodes connected together.
In the previous chapter, we have simulated WSNs with a unit distance from the sink with
the goal of the WSN being solved using single hop communication. In that case, the nodes
did not require connectivity between each other for communications unlike the multi-hop
scenario in the current case study. Here, each node is expected to confirm the temperature,
humidity and other parameters from the local area and then report to a remote base station.
So, in this case, our validation question can thus be defined as follows:
Question: How can we validate that the WSN is accurately monitoring the forest fire
simulation?
The solution to this question lies in following the validation of the previous ecological
case study however we shall demonstrate here as to how VOMAS development can vary
- 223 -
from one case study to another even in closely related ABMs. We can note that in the pre-
vious scenario, we used the FWI to validate if the forest fire generated was "valid" by
comparing the FWI value with the status of the fire location. In this case, we need to re-
design the VOMAS to validate the sensing capability of the sensor node.
6.3.2.2 Software Contract
In this particular scenario, we can use the following Invariant Contract:
Invariant Contract: If the pre-condition that "sensors are correctly reading the values
from their neighborhood" is true, then the outcome of the FWI calculation by the sensors
would result in a post-condition of "correct prediction of fire in the vicinity of the sensor
node".
6.3.2.3 VOMAS agent design
As noted in the generalized methodology, our next step is to decide upon the design of
the agents which will be used for performing the validation. In this case, since we want to
validate the effects of the sensor nodes, it makes sense to use some of the sensor nodes as
VOMAS agents and then use them to compare the validation criteria (FWI) with that of
their neighboring trees. Thus, if there is a fire in the neighborhood of a sensor node agent,
the sensor should predict fire based on the FWI value. The simulation results of this case
study will again be evaluated in the results section of the chapter.
6.3.3 Case Study III: Researcher Growth Simulation
While our first case study was from the domain of ecological Sciences and the second
was from telecommunications in particular, our third case study has been chosen from the
domain of Social Sciences. The goal of choosing different case studies is to be able to
show the generality of the proposed in-simulation validation methods by the use of
- 224 -
VOMAS. In this particular case study, our goal is to simulate the evolution of researchers
over time by means of an integration of ABM and Complex Network concepts.
To be specific, we start by an examination of the research process. The research process
can be considered in the light of two clear aspects. The first aspect is related to how re-
searchers perform research and get it published. The second aspect is the eventual complex
adaptive emergent effects of the popularity of specific researchers, publication Journals or
institutions over time. This particular aspect of research has previously been discussed and
evaluated empirically using case studies earlier in Chapter 3. Here we first examine the
first aspect of the publication and research-related processes.
6.3.3.1 Research Process I: Publication
A first examination of how researchers perform research can be observed in Figure 77.
Here the research workflow starts when a researcher or a group of researchers start work-
ing individually or collectively on one or more innovative ideas. The ideas are correlated
with existing literature to understand a feel of the state of the art. The researcher acquires
data, performs analysis and then subsequently attempts modeling/simulation and/or exper-
imentation in an iterative fashion striving to achieve successful results based on a research
success criteria. After apparent success, the researchers embark on the task of writing the
results in the form of a report. Once this article has been submitted for review, it is evalu-
ated by impartial and anonymous referees. The reviewers examine the report for
originality in terms of advancing the state of the art and subsequently submit their reviews
to the Editor. Eventually the referee report can result in either an acceptance for publication
or else a rejection by means of provision of a set of reviews offering guidance to the au-
thors. In the case of a rejection, the researcher is at least armed with these reviewer's
comments, so s/he can start to work again on the same project and perform more experi-
mentation, generate better results, and the process starts over again.
- 225 -
Figure 77: Research Process I: Generalized research publication Workflow
6.3.3.2 Research Process II: Citations
After a paper has been published and is available for citation by other researchers
by being listed in standard indices (such as the Google Scholar, Scopus or Thomson Reu-
- 226 -
ters), the next step in the publication life cycle is the process of citation. Citations are the
means by which the importance of a particular scientific article can be measured quantita-
tively. In other words, when other researchers read and consider a paper to be important,
they cite the paper in their own papers and the process continues. Interestingly this process
proceeds completely autonomously and thus cannot be strictly influenced by any single
entity. Eventually the citation process can, at times, result in a high prestige of a researcher.
This can be quantitatively measured by means of indices such as the h-Index. This meas-
urement was first proposed in 2005 by Hirsch [217]. In general, the impact of a scientist
can be measured by being highly cited by other authors. Egghe [218] presented a detailed
overview of the inter-relation of Hirsch with related indices. These indices have been con-
sidered as effective measures of a scientist's impact or research productivity.
Hirsch index is formally defined as:
"A scientist has index h if h of his/her Np papers have at least h citations each, and the
other (Np - h) papers have no more than h citations each."
So, as an example, suppose an author has 6 papers: 5 of them have 4 citations and the
sixth has 1 citation. Then h = 4. It is important to note here that the calculation of h-Index
is not simple. It involves several tasks which are performed behind the scenes. These in-
clude intelligent retrieval of information from some standard indexing source (such as
Google Scholar or SCI, SSCI or Scopus etc.) and then calculation of the index by measur-
ing the citations of each published paper. A well-known tool for the calculation of Hirsch
index used extensively by researchers is Ann Harzing's "Publish or Perish" [219].
Earlier in Chapter 3, we examined the process of performing Scientometric analysis of
citation data in depth by means of case studies. We noted earlier that the process of per-
forming CNA involves pruning of nodes. However, while a standard network manipulation
practice, automatic pruning can result in the loss of important and, at times, central nodes
- 227 -
from the network inadvertently. Here the goal of this case study is to propose an agent-
based model for simulation of an alternative Complex Network Modeling approach. This
approach is the Temporal Citation Networks (TCN), which involves a two step process
from bi-graph based author-paper Complex Networks such as those studied earlier in
Chapter 3. The formal methodology of developing these TCNs as an experiment of the
VOMAS methodology is as follows:
- First step is the removal of paper to paper links in the network resulting in a great
reduction in the total number of links in the network.
- The second step is to visually align the network nodes in line with the h-index of
the researcher.
After application of the two steps, the resultant network representation is capable of rep-
resenting different features which were otherwise impossible to note earlier using only
author-paper complex networks. These can be clearly noted in Figure 78. Here blue nodes
represent the papers depicting the total number of citations inside the node. The black col-
ored nodes represent the researchers, which are aligned in the 2-D Euclidean plane with a
height equal to the h-Index of the researcher, scaled to the particular vertical dimension of
the viewing plane. Thus, the goal of the case study is to evaluate the use of TCN in repre-
senting researcher evolution over time (i.e. temporally).
6.3.3.3 Agent Description
The model contains two different types of agents i.e. the researcher agents and the re-
search paper agents. The researcher agents are developed to model the researchers. As
such, they contain two variables for state maintenance. The first of these is "num-papers"
representing the total papers for the researcher. The other variable "my-papers" is repre-
sentative of all the paper agents. In other words, this variable allows for an aggregation of
these agents.
- 228 -
On the other hand, the research paper agent has three different variables. The first varia-
ble is the "tend-to-be-cited" variable. This variable ranges from 0 to 1 and is useful for
simulation of random papers. The second variable is the "num-cites" which represents the
current number of citations of the paper. The third "my-res" stores a pointer to the re-
searcher, to which this paper belongs.
Figure 78: TCN with black nodes depicting researchers placed according to their Hirsch index and
blue nodes depicting individual research papers and citations
In TCN, each paper will be replicated for each author. In other words if a paper has been
co-authored by n authors, there will be n instances of the paper in the TCN. This allows for
a reduction in clutter as well as helps in a succinct visualization to allow focusing on the
evolution of the researchers.
6.3.3.4 Validation Question
In this case study, we shall examine two different validation questions:
- 229 -
Validation Question 1: If the ABM of the TCN is able to represent any number of re-
searchers, would the number of nodes in the TCN be similar to the value predicted earlier
empirically?
Validation Question 2: Is the Hirsch index calculation of the ABM correct in terms of
the actual data?
6.3.3.5 VOMAS agent design
Here, for the first validation question, we shall design VOMAS agents who create a set
of random researcher and paper agents to validate the representational capabilities of the
ABM. However, for the second research question, we need VOMAS agents which can take
in real researcher data and simulate agents based on the evolution of this particular re-
searcher. Thus we have the following software contracts:
Invariant I: If for any TCN, the pre-condition that "any random number of researchers
are correctly represented" is true, then the post-condition that "the total number of links
should be significantly lesser than the number of links in a author-paper network with the
same data" is also true.
Invariant II: If the pre-condition that "researcher evolution data is correctly represent-
ed in a TCN" is true, then the post-condition that "the height of each node should
correspond correctly to the Hirsch Index (h) of the researcher over time" should also be
true.
6.4 Results and Discussion
In the previous section, we developed three different case studies from multiple disci-
plines. Here our goal is to present the results of the experimentation involved in the
validation of the simulations. It is pertinent to note here that while we can perform consid-
- 230 -
erably more experimentation here similar to previously done in earlier chapters in each of
these case studies, we shall however only focus on verification and validation of the effects
of the case study to demonstrate the effectiveness of the proposed generalized methodolo-
gy. Our focus here will be on the different effects of the simulation execution as well as the
resultant effects on the VOMAS validation. Here, we shall demonstrate how VOMAS
based validation builds on empirical validation methods while adding value addition in the
form of being a customizable software engineering modeling scheme.
6.4.1 Case Study I
In our first case study, our goal was to simulate the forest and forest fire dynamics
in considerable details. As such, we developed a set of advanced ABMs which allowed for
evaluation of different aspects of the case study.
To perform the simulation experiments, we notice that for validation, we have to evalu-
ate the following invariant which was described in the previous section:
"If the pre-condition of a "tree is on fire" is true, then a post-condition of "FWI value
in the range of Fire Danger index" is true."
Next we discuss first the simulation parameters followed by the results of simulation
experiments for proving the validity of this invariant.
6.4.1.1 Simulation Parameters
The experiments were conducted to empirically validate the increase of temperature as-
sociated with the forest fire spread. The simulation parameters are given below:
Number of VOMAS agents
Forest density
- 231 -
{50,100,2000}vn{60%,65%,70%}treed Fire delay in terms of simulation days
Normalized average fire growth rate (area per simulation hour):
Tree re-growth rate (new trees per simulation day in forest)
Initial average temperature of forest
Average relative humidity (%)
6.4.1.2 Validation discussion
The results of the first set of experiments can be seen in Figure 79a. Here, we note how
detected forest fires intensity increases with an increase in detected temperature. The sec-
ond result in the Figure 79b depicts the effect of the probability of fire ignition with an
increase in the FFMC.
- 232 -
{10,30,180}firet{0.001,0.002,...,0.9}spreaddfdt{0.001,0.005,0.01}newdndt(){5,10,20,30,40}oavetC{10%,20%,...,50%}aveh
Figure 79: Relationship between (a) Fire Intensity and temperature, (b) Probability of
fire ignition and fire fuel moisture code
We also note the relationship between simulated FWI vs. the FFMC value as a means of
verification of the correct working of the model as given in Figure 80. By means of a real-
istic reflection of the relation of the different values, we can thus deduce that the FWI
calculations are realistic and working.
- 233 -
Figure 80: Relation of FFMC with an increase in FWI
The table for testing the invariant is given in Table 22.
Table 22 Table of Fire Danger Index values
FWI value Fire risk
0-5
Low
5-10
Moderate
10-20
High
20-30
Very high
30+
Extreme
For validation of the invariant, the results were plotted as shown in Figure 81. As can be
noted, the randomly placed tree agents who are acting as VOMAS agents are successfully
able to detect the fire.
- 234 -
Figure 81: Tree VOMAS agents detecting fire vs. Fire area
In terms of the detected FWI in the case of fires, we can observe the detected behavior
in Figure 82. We can note here that the FWI value detected is with the fire danger as shown
earlier in the fire danger index table. Here, while the VOMAS agents are scattered
throughout the simulation environment, they are still able to detect the effects of the forest
fires of reasonable sizes.
Thus, we have noticed that by means of VOMAS, we have been able to perform valida-
tion of the case study model in addition to using empirical validation methods.
- 235 -
Figure 82: Maximum FWI detected by VOMAS tree agents plotted vs. the area of fires
6.4.2 Case Study II
In the previous case study, we validated the ecological simulation of forest fires. In the
second case study, our goal was to combine the earlier case study of fire simulation with a
multi-hop environmental monitoring WSN simulation developed as a QUDG as explained
earlier in the chapter.
We can recall here, the invariant which was designed for use in the validation of this
case study ABM. The invariant can be re-stated as follows:
If the pre-condition that "sensors are correctly reading the values from their neighbor-
hood" is true, then the outcome of the FWI calculation by the sensors would result in a
post-condition of "correct prediction of fire in the vicinity of the sensor node".
- 236 -
6.4.2.1 Simulation Parameters
Since this ABM is based on the previous ABM which was on ecological modeling of
the structural evolution of a forest, some of the parameters of the experiments are similar to
the previous case study. While others are new, being related to the sensor nodes as given
below:
Number of VOMAS agents
Forest density
Fire delay in terms of simulation days
Normalized average fire growth rate (area per simulation hour):
Tree re-growth rate (new trees per simulation day in forest)
Initial average temperature of forest
Average relative humidity (%)
Number of sensor nodes
Normalized distribution of sensors (in terms of sensors per square simulation area)
Maximum sensor communication radius
Maximum sensor sensing radius in terms of simulation measure
- 237 -
{50,100,2000}vn{60%,65%,70%}treed{10,30,180}firet{0.001,0.002,...,0.9}spreaddfdt{0.001,0.005,0.01}newdndt(){5,10,20,30,40}oavetC{10%,20%,...,50%}aveh{100,200,...,500}sensorsn{0.0013,0.026,0.013}sensorsimnAmax(){20,30,40,50}commRmax(){2,5,7,10}sensR 6.4.2.2 Validation discussion
Here, our basic simulation model can be viewed in Figure 83. As can be noted from the
simulation output, the sensors are all connected with each other however not every sensor
will be directly accessible from each other sensor using a single hop. Thus, the information
in this WSN needs to be aggregated and confirmed with local sensors before it is forward-
ed to the remote sink.
Figure 83: Basic simulation model of multi-hop WSN modeled as a QUDG monitoring forest
conditions
When the simulation is actually executed, each of the WSN nodes measure local param-
eters and display them as labels as can be noted in Figure 84. In this particular design, as
noted earlier, the number of sensors is limited as compared with the number of tree nodes
in the case of environmental modeling; as such the calculation of FWI in these nodes peri-
odically (e.g. every 1000th tick) would not be a problem for simulation execution.
- 238 -
Figure 84: WSN monitoring the local environment displaying measured parmeters
The structure of the forest as well as the weather evolves over time with fires, snowfall
and rainfall occurring at random instances. The result of a simulation where the fire has
caused an impact on a major section of the forest monitored using the WSN can be ob-
served next in Figure 85. Here, we can also note that one of the center nodes demonstrates
the fire was very recent. This phenomenon can be noted visually by means of observing the
changing of color of the appropriate WSN node to yellow.
- 239 -
Figure 85: Wireless Sensors recording forest temperature and calculating FWI
Different simulation parameters added to the simulation model can now be seen in the
configuration panel shown in Figure 86.
- 240 -
Figure 86: Configuration Options of simulation models of combining Forest Fire simulations with
monitoring simulation by a WSN
A detailed view of the simulation of several fires in a forest with observed FWI values
is given.
Figure 87: Effects of monitoring large set of fires and FWI value
- 241 -
As we can see in Figure 87, the temperature as well as the observed FWI values exhibit
similar variations. In addition, we note that the observed value of FWI greater than 20 is
demonstrating the effects of the forest fires correctly validating the ability to monitor forest
fires using the sensors.
For a formal validation of the fire detection invariant, we plot the earliest time of fire
detection vs. the average distance of forest fires as shown in Figure 88.
Figure 88 :Earliest detection of FWI spike vs. average distance of sensors from fire incident
Here we can clearly note that using the linear interpolation line, the detection fol-
lows a trend line of
with a reasonable R-squared value of
.
Here the fact that this value is greater than 0 demonstrates the possibility of prediction of
one value based on the other value (in linear regression).
- 242 -
0.9110.317yx20.572R Thus, in this case study, we have again demonstrated the effectiveness of using
VOMAS to verify and validate the simulation model using the invariant contract. In this
particular case, again, our validation was coupled with existing model validation tech-
niques of empirical validation and cross-model validation using multiple cooperative
agents in the same simulation model.
6.4.3 Case Study III
While the previous two case studies were from ecological modeling and telecommuni-
cations respectively, our third case study, which was developed in the chapter previously is
from Social Sciences and the two invariant contracts for this case study can be stated as
follows:
Invariant I: If for any TCN, the pre-condition that "any random number of researchers
are correctly represented" is true, then the post-condition that "the total number of links
should be significantly lesser than the number of links in a author-paper network with the
same data" is also true.
Invariant II: If the pre-condition that "researcher evolution data is correctly represent-
ed in a TCN" is true, then the post-condition that "the height of each node should
correspond correctly to the Hirsch Index (h) of the researcher over time" should also be
true.
6.4.3.1 Simulation parameters
In this case study, the way to understand the experimental results is that for validation of
the hypotheses, we have developed essentially two different types of validation VOMAS
mechanisms inside the same ABM. The first validation VOMAS is specifically for the first
invariant allowing for the creation of a random number of researchers using VOMAS
- 243 -
agents based on the TCN hypothesis based complex network model. The second validation
VOMAS demonstrates the validation of the second invariant by allowing for the mapping
and loading of actual Hirsch index data and demonstrating the evolution of a well-known
researcher. The simulation parameters of the two validation simulation experiments are ex-
plained below:
Number of researcher nodes
Maximum number of initial papers per researcher
Average
tendency
to be cited
is a probability which
is a uniform ran-
dom,
6.4.3.2 Validation of the representational abilities of TCN models
Researchers are each placed on the Cartesian coordinate system, based on their Hirsch
index. Each researcher, shown in black, has two labels. The first label shows the total
number of published papers and the second shows the Hirsch index. Each paper, shown in
blue, has the total citations as a label. This experiment validates the representational ability
of the TCN modeling paradigm for researchers as shown in Figure 89. Notice that this val-
idates the first invariant since we have demonstrated the capability of representation of any
number of researchers on the basis of TCN model in simulations.
Figure 89 :Simulation view of n = 60 random researchers with max-init-papers = 10
- 244 -
{10,20,...,100}resnmax(){10,20,...40}papersn(0,100)citedPrnd 6.4.3.3 Validation of Hirsch Index Calculation
For the validation of the second invariant, we shall use concepts from empirical valida-
tion by means of calculating evolution of actual Hirsch index values for a renowned
researcher in the domain of multiagent systems. The chosen researcher was Prof. Victor
Lesser because of his relatively high h-Index as noted in various citation indices. Using
"Publish or Perish" software [220], the Google scholar citation index was queried. It was
discovered firstly that the index lists a total of 649 papers. However, 546 of these papers
have valid and properly indexed years which were therefore used in the subsequent valida-
tion exercises. The first indexed paper of the researcher is indexed from the year 1968. The
data for the next 20 years for "Victor Lesser" has been plotted in Figure 90.
Figure 90: h-Index plot for twenty years for "Victor Lesser" obtained using Publish or Perish program
Next, we simulate this using our ABM as shown in Figure 91a and b which show ten
years of evolution of h-Index of Victor Lesser as depicted using TCN by means of the sim-
ulated agent-based model. The detailed results and the retrieved data (via Google scholar
index) are depicted in the experiments are given in Table 23.
- 245 -
Figure 91: a, b Validation exercise 1: Evolution of a researcher's Hirsch-index (Victor Lesser)
- 246 -
Table 23 Data for validation. Calc hI = Calculated h-Index and h-Index is the h-Index from Publish or
Perish software which obtains data via Google Scholar (Current as of July, 2011)
H-Index
Citations of papers
Calc hI
1
1
2
3
3
4
4
5
6
8
6
6
[6 3]
[20 6 3]
[20 7 6 3]
[20 7 6 6 3]
[20 7 6 6 3 2 0]
[166 20 7 6 6 3 3 2 0 0]
[166 21 20 15 7 6 6 3 3 2 0 0]
[166 122 103 71 21 20 15 8 7 6 6 3 3 2 0
0 0]
1
1
2
3
3
4
4
5
6
8
Year
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
Table 23 shows two columns for the Hirsch index. One column shows the Hirsch index
calculated using Google Scholar and the Publish or Perish program. On the other hand, the
"Calc hI" column indicates the Hirsch index calculated by the algorithm alongside the evo-
lution of the researcher. The table also shows the number of cited papers per researcher.
Thus, we can see how we have validated both the calculation of the h-Index as well as the
effectiveness of TCNs for the modeling of researcher reputation (state) co-evolution with
the changes in the topology using our agent-based model.
In terms of the number of nodes required to be displayed in traditional author-paper ci-
tation networks versus TCNs, we can observe the results as shown in Figure 92.
- 247 -
Figure 92: Validation Exercise 2: Comparison of number of nodes needed to display a author-paper
citation network vs. a TCN for Victor Lesser's papers.
Thus, by means of the proposed methodology, we have demonstrated how to verify and
validate the case study. Using two different invariants allowed us to demonstrate different
aspects from the earlier two case studies.
6.4.4 Critical comparison of VOMAS with other approaches
In the previous section, we have demonstrated the broad applicability of the proposed
in-simulation validation scheme for ABMs by application of the methods on three different
case studies from different scientific disciplines. In this section, we evaluate existing meth-
ods in contrast with our methodological approach and analyze how this validation scheme
fits in these methods prevalent in literature for the verification and validation of ABMs.
6.4.4.1 Empirical Validation
Empirical validation is concerned with using empirical data to validate the ABM. Dif-
ferent researchers such as Marks [221] and Fagiolo et al. [202] have proposed the use of
empirical validation of agent-based models. The proposed in-simulation validation meth-
odology builds on top of existing empirical validation techniques. It essentially allows for a
design of validation techniques using software engineering concepts. While it can still use
empirical validation subsequently or else perform conceptual validation depending upon
- 248 -
the particular type of simulation model or application case study (as has been shown earlier
in the case studies). Thus there is no conflict here with the original related empirical vali-
dation methodologies.
6.4.4.2 Validation using Companion Modeling Approach
In Social Sciences, there is an approach to agent-based modeling termed as the "Com-
panion Modeling Approach". In companion modeling, the researcher or modeler are the
one and the same and they work together with stakeholders from the authorities to develop
models and results. By having stakeholders examine the implications, the models are in
essence "face" validated. The process has been generalized by Bousquet and Trébuil [222]
as shown in Figure 93. The idea is to go back and forth between different stakeholders and
the researchers where the stakeholders are the authorities governing the business-
es/economic ecosystems. However, this methodology does not differentiate between the
simulation specialists and end-user researchers in any way. As such, while end-user re-
searchers might be experts in their domain of modeling, they can be separate entities from
the actual teams who can better develop the simulation models. This is one way in which
how our proposed methodology can be considered to extend upon the "companion model-
ing" concept. In addition, our methodology brings in concepts from Software Engineering
and Artificial Intelligence.
Another way in which our proposed methodology is different is the use of descriptive
agent-based models (DREAM models as discussed in previous chapter) which are able to
describe ABMs across scientific disciplines. In the proposed companion modeling ap-
proach, the goal was to develop simulations with only stakeholders from the business
ecosystem modeling and not multidisciplinary validation. Still, in this way, the proposed
VOMAS-based approach extends this companion modeling set of concepts which were
- 249 -
designed for specific application case studies to generalized multidisciplinary approaches.
Again, this implies there is no direct conflict between the two approaches and VOMAS
intelligently uses and extends existing techniques.
Figure 93: Companion modeling cycle, figure credit: Bousquet and Trébuil [222]
6.5 Conclusions
In this chapter, we have presented the proposed in-simulation verification and validation
methodology of building customized VOMAS inside different ABMs. This validated
agent-based modeling level of the framework builds upon previous framework levels such
as exploratory agent-based modeling and descriptive agent-based modeling allowing for
validated agent-based modeling. As a means of unification of all ideas and concepts, the
methods and case studies extensively involve both the use of agent-based models and
complex network models and methods in different scientific disciplines. We presented
three different case studies from different scientific disciplines including ecological model-
ing, telecommunications and social sciences demonstrating the broad applicability of the
proposed methods involving building customized validation schemes based on the particu-
lar case study.
- 250 -
7 Conclusions and Future Work
This thesis represents first steps towards the solution of a set of nontrivial problems as-
sociated with the colossal task of developing a unified framework applicable to
multidisciplinary and inter-disciplinary cas research studies. We propose a combination of
agent-based modeling and complex network approaches, two paradigms which are preva-
lent in cas modeling literature but, to the best of our knowledge, have never previously
been combined in the form of a framework enabling comprehensive cas inter-/multi-
disciplinary research studies to be carried out based on expected goals and outcomes of the
studies.
In this chapter, we first give an overview of the thesis contributions. Next, we critically
evaluate the different framework levels and case studies in relation to other state-of-the-art
techniques prevalent in cas modeling literature. This is followed by details of the limita-
tions of our work and finally we propose how the different framework levels can be
utilized in other cas research case studies in addition to how our proposed framework lev-
els can be further enhanced by other researchers in long term studies.
7.1 Overview
To demonstrate the applicability and use of each of the different proposed framework
levels, we had to work in different scientific disciplines and nomenclatures. At times, we
had to work with domain experts from a variety of backgrounds to develop the different
case studies. The application case studies ranged from consumer electronics, agent-based
computing, complex adaptive communication networks, ecological modeling, scientomet-
rics and social sciences. While the framework levels propose a systematic set of ideas for
modeling and simulation of cas depending on the data types available and the case study
objectives, by no means do they reflect a comprehensive framework for modeling all pos-
- 251 -
sible types of cas, rather, as stated in the thesis objectives, our work represent a first step
towards realizing this ambitious goal.
7.2 Research Contributions
The unified framework proposed in this thesis is structured in the form of four different
levels based on objectives of the proposed research as well as availability of suitable data.
These levels use a combination of agent-based and complex network-based modeling ap-
proaches allowing multidisciplinary researchers to better conduct research by adopting a
suitable framework level for developing models for their particular case study. Each
framework level is substantiated by one or more complete example case studies as a means
of demonstration of the generalized applicability and usage.
7.2.1 Complex Network Modeling Level for Interaction data
In chapter 3, we proposed the complex network modeling level. This level is based on
the scenario where suitable data is some available, categorized and classified. This level is
more suitable for use in the case of highly standardized interaction data repositories. To
specifically show the broad applicability of the proposed framework level, a standard data
repository, the Thomson Reuters web of knowledge, was selected. The initial data retrieved
from this source was textual in nature with a large number of tags. Since only a few set of
tags were suitable for each type of complex network, these text files were first parsed and
then analyzed for tag selection. This was followed by the extraction of suitable complex
networks. Subsequently different complex network methods were employed to extract
suitable quantitative and qualitative measures about these networks. These measures were
used to correlate the scientific implications of the research cas with the structural topologi-
cal features of each network. Using two separate case studies from two different scientific
domains of "Agent-based Computing" and "Consumer Electronics", we demonstrated the
- 252 -
broad applicability of the complex networks modeling methods. This framework level also
serves as an encompassing level for existing complex network modeling case studies.
7.2.2 Exploratory agent-based modeling level
In chapter 4, we proposed the use of Agent-based Modeling to develop exploratory cas
studies. The idea of exploratory ABM is to allow researchers to focus specifically on
demonstrating the feasibility of their ABM design by means of developing proof of con-
cept models. Similar to the complex network modeling level, exploratory agent-based
modeling level is also an encompassing level and allows different existing agent-based
modeling literature to be classified under this level.
To demonstrate the broad applicability of the proposed modeling paradigm, we demon-
strated exploratory agent-based modeling in the domain of "Internet of Things", a sub-area
of upcoming research area of Cyber-physical systems where household and commonly de-
vices can use communication networking to retrieve or expose useful information. The
case study was used as a proof-of-concept for the use of unstructured P2P search algo-
rithms in a completely different domain by demonstrating how static and mobile devices
can dynamically connect together to expose different types of content in their surround-
ings. In general, simulation results from the case study, demonstrated the effectiveness and
benefits of using the proposed exploratory agent-based modeling level for performing fea-
sibility studies and developing proof-of-concept models for aspects such as needed in e.g.
funding applications or for planning future research.
7.2.3 Descriptive agent-based modeling level
While exploratory ABM can be effective in conducting proof-of-concept feasibility
studies, occasionally the goal of modeling is to be able to explore the use of communica-
tion of models as a means of inter-disciplinary model comparison for knowledge transfer
- 253 -
and learning. Our proposed DescRiptivE Agent-based Modeling (DREAM) approach al-
lows for this type of modeling and simulation. The DREAM model consists of a
quantitative complex network model (i.e. a network model along with a quantitative "foot-
print" based on a centrality analysis of the network) coupled with a pseudocode-based
specification model. DREAM model allows for a high level of fidelity with the ABM using
a non-code modeling approach. Thus it allows for comparison of different models without
expecting end-user cas researchers to be able to program models. Researchers can develop
a DREAM model of their cas case study by extending the already-provided baseline com-
plex network model followed by performing a complex network analysis. DREAM models
also allow for a quantitative comparison of the centrality- based "footprint" of different
ABMs across application case studies in addition to a visual comparison of different mod-
els by means of the different complex networks. A case study example demonstrating the
use of DREAM in the domain of a heterogeneous ABM consisting of a well-studied
"boids" model of flocking coupled with a single-hop WSN model demonstrates the usage
of the descriptive agent-based modeling level. Simulation results obtained from the case
study experiments also indicated some interesting research outcomes in terms of how sens-
ing can be used to predict and detect group behavior in large and complex mobile agents
(such as persons in crowds or airplanes in the sky etc.).
7.2.4 Validated agent-based modeling level
The validated agent-based modeling level of the proposed framework is suitable for use
in research case studies where the validation of phenomena in the ABM is the central goal
of a particular case study. Using a combination of concepts from three different scientific
disciplines of multiagent systems, social sciences and software engineering, we propose
the use of a VOMAS approach which allows for both the segregation of roles of the Simu-
lation Specialists and the Subject Matter Experts in a team-oriented iterative methodology.
- 254 -
The goal of the VOMAS verification and validation approach is to allow cas researchers to
specify validation questions which are transformed to invariants based on pre-conditions
and post-conditions. These invariants are enforced in the ABM subsequently by custom-
ized VOMAS agents using an in-simulation validation scheme. For the validation and
demonstration of the proposed approach, three separate case studies encompassing ecolog-
ical modeling, WSN monitoring and social simulation of researchers in the form of a
Temporal Citation Network (TCN) were presented demonstrating the effectiveness and vi-
ability of the proposed validated agent-based modeling level.
7.3 Critical review of the proposed levels
The first two levels of the proposed framework can be considered as encompassing lev-
els since they allow the inclusion of existing cas literature under the proposed framework
guidelines. Thus complex network models in existing cas literature can be classified under
the complex network modeling level of the proposed framework as complex network mod-
els while existing agent-based models can be classified under the exploratory agent-based
modeling level. The rest of the two framework levels specifically propose descriptive and
validated Agent-based modeling respectively. Next, both these levels are critically com-
pared with existing methods in cas literature.
7.3.1 Critical Comparison of DREAM with other approaches
While, to the best of our knowledge, there is no other well-known approach in literature
which describes agent-based models descriptively and quantitatively similar to the pro-
posed DREAM approach as part of the descriptive agent-based modeling level, here we
review the DREAM approach to two somewhat related approaches of knowledge transfer
and description of agent-based models (discussed earlier in more details in chapter 5):
- 255 -
7.3.1.1 Template Software approach
The Santa Fe Institute proposes the use of template software approach[130] which es-
sentially entails using software code as a means of model comparison. However, while
useful, there are certain obvious limitations of this approach given as follows:
Firstly, software templates are code-based and while helpful (in line) comments can at
times augment the readability of code, the fact remains that code is always tied to a single
platforms and/or language. As such, this approach does not allow for either a non-code or a
quantitative/visual comparison of models as proposed in the descriptive agent-based mod-
eling level of the proposed framework.
Secondly, one key benefit of moving away from pure code –based approaches is the re-
duction of required programming-related learning curve for researchers to perform inter-
disciplinary model comparison without the requirement of a detailed programming
knowledge by means of visual or quantitative or pseudocode template based comparison.
7.3.1.2 ODD approach
A textual approach of describing ecological and individual-based (agent-based) models
has been proposed by Grimm et al. This approach is termed as the Overview, Design con-
cepts, and Details (ODD) protocol[199]. As compared with the proposed DREAM
modeling approach, there are certain key limitations of ODD given as follows:
Firstly, ODD is primarily textual description. A closer examination of the ODD ap-
proach reveals that it is basically a way of describing agent-based models using a proposed
checklist of headings. Secondly this approach does not allow for a quantitative comparison
such as possible by means of the use of complex network methods as described in the pro-
posed modeling methodology. Thirdly, ODD also does not offer any visualization-based
- 256 -
approach where researchers can visualize an entire ABM for comparison with another
ABM.
Another problem with the ODD approach is that it also does not offer a one-to-one
translation of models from the actual model to the source code. Finally, ODD also does not
allow for writing detailed technical descriptions of algorithms (since it is textual and does
not use pseudocode, flowcharts or UML etc. in any way to describe the algorithms). Thus
to summarize, while textual descriptions are effective in giving basic ideas about a model,
the proposed DREAM model based descriptive agent-based modeling level allows for a
number of other benefits as compared to the ODD approach.
7.3.2
Critical comparison of VOMAS with other approaches
Like the previous comparison of DREAM models, here we give the summary of a more
detailed comparison, given earlier in Chapter 6, of the proposed in-simulation validation
approach with other literature in the domain of verification and validation of ABMs.
7.3.2.1 Empirical Validation
Empirical validation is a technique used primarily in the domain of Agents in Computa-
tional Economics (ACE) and social sciences concerned with using empirical data to
validate ABM such as by Marks [221] and Fagiolo et al. [202]. While empirical validation
is an effective approach, it is not associated with a specific analysis and design approach
on how to structure the simulation study or how an agent-based model can be enhanced for
performing a systematic verification and validation. As such, the proposed in-simulation
validation methodology of the validated agent-based modeling level builds on top of these
empirical validation techniques. It essentially allows for the design of validation techniques
using software engineering concepts in addition to allowing for performing customized
conceptual validation of the case study.
- 257 -
7.3.2.2 Validation using Companion Modeling Approach
In Social Sciences, another approach to validation is termed as the "Companion Model-
ing Approach". In the companion modeling approach, researchers work together with
stakeholders from higher authorities (in the case of Business ecosystems) to develop mod-
els and get approval of the results. In other words, by involving the stakeholders, the model
is essentially "face" validated as noted by Bousquet and Trébuil [222]. The idea is to go
back and forth between different stakeholders and the researchers where the stakeholders
are the authorities governing the businesses/economic ecosystems. However, this method-
ology does not differentiate between the simulation specialists and the end-user researchers
in any way. As such, while end-user researchers might be experts in their domain of model-
ing, they can be separate entities from the actual teams who might be actually developing
the simulation models. Secondly, the companion modeling approach does not allow for a
step-by-step team-oriented approach to validation. Thirdly this approach is limited to busi-
ness stakeholder and is domain specific.
Thus, the proposed in-simulation methodology can also be considered to extend existing
concepts from the "companion modeling" approach. In addition, the proposed methodolo-
gy utilizes concepts from Software Engineering and Artificial Intelligence to form a
customizable approach to the in-simulation validation of Agent-based simulation models.
7.4 Limitation of the Methods
In the previous sections, we have given an outline of our key contributions in this thesis
followed by a critically comparison of the different framework levels with existing state-
of-the-art approaches in cas literature. Here, we give details of the limitations of the indi-
vidual framework levels as well as the limitations of case studies and the choice of the
selected ABM tool.
- 258 -
7.4.1 Limitations and proposed extensions of the framework levels
Here, we outline the limitations of the individual framework levels. Firstly, as men-
tioned earlier in this chapter, the complex network modeling and the exploratory agent-
based modeling levels are specifically designed for encompassing existing literature in cas
modeling. As such, they are purposefully designed to allow existing case studies from the
domains of complex network and agent-based modeling paradigms.
However, as noted in previous section, while ODD is a completely textual description,
none of the framework level currently tie in with the ODD model. As such, one possible
extension of the exploratory agent-based modeling level could thus be to tie in ODD with
the proposed descriptive agent-based modeling level.
For the VOMAS in-simulation modeling approach, while we have applied it to different
case studies, it has not been possible to apply and evaluate the approach to earlier business
ecosystem modeling domains, which use empirical validation and companion modeling
approaches. As such, a possible iterative improvement in this in-simulation modeling ap-
proach can be foreseen if the proposed approach were applied to case studies which have
previously been studied by researchers in the companion modeling and the empirical vali-
dation areas. This would thus validate the approach further and possibly evolve it further in
the directions of these previously well-studied approaches.
7.4.2 Selection of Case studies
Our first key limitation arises from the diversity of possible cas that exist in the real
world. While we have attempted to cover a large number of representative application case
studies, there are simply too many possible domains to be covered in a limited period of
time and are hence out of the scope of this thesis. Although we believe we have demon-
strated the methods in a number of key target sub-domains of cas such as social sciences,
- 259 -
ecological sciences as well as telecommunications, our selected case studies only reflect a
limited combination of cas. Thus, extensive further research and exploration of cas case
studies need to be performed to fully evaluate our proposed framework in other cas do-
mains.
7.4.3 Selection of Agent-based Modeling Tool
At several times during the research, other agent-based modeling tools were evaluated for
possible usage in the different case studies. These included Repast-S, Mason, Swarm,
Scratch and other toolkits, a good overview of which can be found in[141]. However, none
of these tools appeared to better suit the case studies in comparison with the NetLogo sim-
ulation tool. However, there are certain problems associated with NetLogo that limit its
usage in certain cases that can only be overcome by advanced programmers capable of de-
veloping widgets and extensions to extend NetLogo. Here, we list some of the observed
limitations:
A. NetLogo is based on primarily a single file structure. While other files can be im-
ported, it does not offer any proper reuse in the form of object oriented modules.
While NetLogo extensions can be built and compiled in Java, the Logo language it-
self requires code to be re-written or imported instead of being able to make classes
and objects for reuse.
B. While the new version (V4.1) of NetLogo allows multi-threaded execution of sever-
al experiments simultaneously, it does not directly allow for performing distributed
experiments on multiple machines. As a result, it is not very easy to execute large
scale realistic models, which might otherwise require resources offered only by a
cluster of machines.
- 260 -
C. One major problem with NetLogo is the lack of any strong debugging support. The
debugging primitives allowed by NetLogo are essentially the same as printing val-
ues, which was the norm in low-level languages such as Assembly and Machine
Languages. As such, on the one hand, while NetLogo offers excellent agent-based
modeling support, on the other, there is no way of either setting breakpoints or
watches as is the norm in most advanced Integrated Development Environments
(IDEs).
D. Finally, one other problem with NetLogo is that it does not differentiate between the
IDE and the runtime environment. As such, it is not possible to develop NetLogo
code in a separate IDE while running it in separate runtime.
7.5 Proposed Future directions
The framework and the example case studied allow for a number of possible usages in
further research studies. Some of these are given as possible future case studies as follows:
Specifically, at the complex network modeling level, the complex network approach can
be further applied to complex adaptive social and biological systems. These can include the
extraction and analysis of nucleic acid molecules secondary structures (from genomic da-
ta), analysis of social network structures and internet-scale communication networks.
In addition, the exploratory agent-based modeling level can be further applied in differ-
ent scientific disciplines such as in exploration of how complex networks can be used to
study the spread of HIV infections in sub-populations amongst other possible case studies.
In addition, the Cyber Physical Systems case study can be further enhanced by practically
implementing the study using mobile devices and performing comparative analysis of the
results obtained from the simulation study with results obtained from the physical devices.
- 261 -
The proposed descriptive agent-based modeling approach can be further applied to in-
ter-disciplinary case study applications. In addition, the results of sensing in complex
environments have numerous possible future applications. Examples of possible applica-
tions of sensing of flocking can include the following:
1. Sensing can be used for the identification of collective behavior of people using an
image processing or infrared approach. In other words, this technique can be used
to detect any group of people such as a group of "shoppers" flocking to a particular
set of shops in a mall even though there is a large crowd of thousands of other peo-
ple moving around.
2. This same approach can be used to detect a group of malicious attackers moving
through a crowd thus allowing for the detection of the group "flocking" without
cohering too much as that close grouping might have raised suspicions of the secu-
rity personnel.
3. Another application of this approach is in the domain of detecting a group of stealth
aircraft flocking towards a common mission. Stealth aircrafts are known to be in-
visible to radars in certain conditions. However these aircrafts are physically visible
to the naked eye or in the visible light. As such, while it is difficult to infer a coor-
dinated operation based on only the detection of a single aircraft, if there were a
randomly deployed group of sensors with appropriate cameras or other effective
sensing equipment pointing to the sky, these wirelessly connected sensors could po-
tentially detect a group of stealth aircraft flocking and moving towards a common
goal. Using simple shape detection image processing algorithms, each sensor can
simply note if a "large" object passes over them. And this detection of "proximity"
could be used to quantify flocking at a monitoring station collecting the data from
these sensors.
- 262 -
The validated agent-based modeling level of the framework can be used for validation
of any agent-based modeling case study. This is particularly useful in physical sciences
where validation of hypotheses can be crucial to the study. Possible future application case
studies can include the use of agent-based modeling for tumor growth analysis as well as
effects of treatment prediction, HIV spread modeling inside the patient's body, use of ABM
to study fungal adsorption of metal ions, simulation study of correlation of Helicobacter
Pylori spread in patient's gastrointestinal tract and others.
A general future research direction could take the form of development of software tools
and components for a possible unified framework. Thus researchers could implement the
framework levels in a single tool by, for example, embedding an agent-based modeling
tool such as NetLogo.
7.6 Final Words
To summarize, our framework design goals have been to develop a set of framework
levels allowing inter-disciplinary communication, collaboration and research by means of
combining complex network and agent-based modeling methods for multidisciplinary cas
researchers. While both complex network models as well as agent-based models have pre-
viously been used in different application cas studies, to the best of our knowledge, they
have not previously been combined together in the form of a single unified set of frame-
work guidelines. We have undertaken several successful case studies and reported results
across a range of selected domains, which demonstrate the effectiveness of our proposed
framework. However, our work is still only a first step towards the formulation of the en-
visaged comprehensive unified framework for the modeling and simulation of cas. It is
hoped that continuing collaborative work in this domain would result in further enhance-
ment and applications of the proposed framework.
- 263 -
Appendix 1
Here we would like to mention the keywords used for searching the ISI web of
knowledge in addition to the reasoning behind the selection. Arguably there are numerous
ways to classify as sub-domain based on keywords. In this particular case, some of the
keywords were even shared with Chemical and Biological Journals (e.g. using the word
agent for e.g. Biological agent or Chemical agent even). As such, we had to limit the search
to papers with a focus on either agent-based modeling specifically or else in the domain of
multiagent systems.
The search was thus performed on titles and the exact search from the ISI web of
knowledge was as following:
Title=(agent-based OR individual-based OR multi-agent OR multiagent OR ABM*)
AND Title=(model* OR simulat*)
Timespan=All Years. Databases=SCI-EXPANDED, SSCI, A&HCI, CPCI-S.
Date retrieved: 8th September 2010 (1064 records)
- 264 -
References
[1]C. E. Van Wagner, "Development and structure of the Canadian forest fire weather
index system," Forestry technical report, vol. 35, pp. 37, 1987.
[2]M. A. Niazi, and A. Hussain, "Social Network Analysis of trends in the consumer
electronics domain," in Consumer Electronics (ICCE), 2011 IEEE International
Conference on, 2011, pp. 219-220.
[3]M. Niazi, and A. Hussain, "Agent-based computing from multi-agent systems to agent-
based models: a visual survey," Scientometrics, pp. 1-21, 2011.
[4]M. A. Niazi, and A. Hussain, "Agent-based tools for modeling and simulation of self-
organization in peer-to-peer, ad hoc, and other complex networks,"
Communications Magazine, IEEE, vol. 47, no. 3, pp. 166-173, 2009.
[5]M. A. Niazi, and A. Hussain, "Sensing Emergence in Complex Systems," Sensors
Journal, IEEE, vol. 11, no. 10, pp. 2479-2480, 2011.
[6]M. A. Niazi, and A. Hussain, "A Novel Agent-Based Simulation Framework for Sensing
in Complex Adaptive Environments," Sensors Journal, IEEE, vol. 11, no. 2, pp.
404-412, 2011.
[7]M. Niazi, A. Hussain, A. R. Baig, and S. Bhatti, "Simulation of the research process," in
40th Conference on Winter Simulation, Miami, FL, 2008, pp. 1326-1334.
[8]M. Niazi, Q. Siddique, A. Hussain, and M. Kolberg, "Verification and Validation of an
Agent-Based Forest Fire Simulation Model," in SCS Spring Simulation
Conference, Orlando, FL, USA, 2010, pp. 142-149.
[9]M. A. Niazi, A. Hussain, and M. Kolberg, "Verification &Validation of Agent Based
Simulations using the VOMAS (Virtual Overlay Multi-agent System) approach," in
MAS&S 09 at Multi-Agent Logics, Languages, and Organisations Federated
Workshops, Torino, Italy, 2009, pp. 1-7.
[10]J. Holland, "Complex adaptive systems," Daedalus, vol. 121, no. 1, pp. 17-30, 1992.
[11]C. Gershenson, and F. Heylighen, "When Can We Call a System Self-Organizing?,"
ECAL, vol. Volume 2801/2003, pp. 606-614, 2003.
[12]J. A. Pelesko, "Self assembly: the science of things that put themselves together,"
Chapman & Hall, 2007.
[13]F. Boschetti, and R. Gray, "A Turing Test for Emergence," Advances in Applied Self-
organizing Systems, Advanced Information and Knowledge Processing M.
Prokopenko, ed., pp. 349-364: Springer London, 2008.
[14]H. Kitano, "Systems biology: a brief overview," Science, vol. 295, no. 5560, pp. 1662,
2002.
[15]F. Amadieu, C. Mariné, and C. Laimay, "The attention-guiding effect and cognitive
load in the comprehension of animations," Computers in Human Behavior, vol. 27,
no. 1, pp. 36-40, 2011.
[16]J. Epstein, "Why model?," Journal of Artificial Societies and Social Simulation, vol.
11, no. 4, pp. 12, 2008.
[17]M. Mitchell, "Complexity: a guided tour," Oxford University Press, USA, 2009.
[18]V. Grimm, and S. F. Railsback, "Individual-based Modeling and Ecology," Princeton
University Press, 2005.
[19]M. J. North, and C. M. Macal, "Managing business complexity: discovering strategic
solutions with agent-based modeling and simulation," Oxford University Press,
USA, 2007.
- 265 -
[20]N. Gilbert, and K. G. Troitzsch, "Simulation for the social Scientist," Second ed.:
McGraw Hill Education, 2005.
[21]P. L. Lollini, and F. P. Santo Motta, "Discovery of cancer vaccination protocols with a
genetic algorithm driving an agent based simulator," BMC bioinformatics, vol. 7,
no. 1, pp. 352, 2006.
[22]C. Cioffi-Revilla, "Comparing Agent-Based Computational Simulation Models in
Cross-Cultural Research," Cross-Cultural Research, vol. 45, no. 2, pp. 208-230,
May 1, 2011, 2011.
[23]N. Boccara, "Modeling complex systems," Second ed.: Springer, 2010.
[24]L. A. N. Amaral, and J. M. Ottino, "Complex networks," The European Physical
Journal B - Condensed Matter and Complex Systems, vol. 38, no. 2, pp. 147-162,
2004.
[25]R. Axelrod, "The complexity of cooperation: agent-based models of competition and
colloboration.," Princeton, NJ: Princeton University Press, 1997.
[26]M. E. J. Newman, "The Structure and Function of Complex Networks," SIAM Review,
vol. 45, no. 2, pp. 167-256, 2003.
[27]U. Nations, Second generation surveillance for HIV: The next decade, HASP, Pakistan,
Islamabad, 2010.
[28]J. A. Bondy, and U. S. R. Murty, "Graph theory with applications," MacMillan, 1976.
[29]Author ed.^eds., "Analysis of biological networks," John Wiley & Sons,, 2008, p.^pp.
Pages.
[30]W. d. Nooy, A. Mrvar, and V. Batagelj, "Exploratory Social Network Analysis With
Pajek," Cambridge, UK: Cambridge University Press, 2005.
[31]O. Balci, "Verification, validation, and accreditation," in Proceedings of the 30th
conference on Winter simulation, Washington, D.C., United States, 1998.
[32]S. Forrest, and T. Jones, "Modeling complex adaptive systems with Echo," Complex
systems: Mechanisms of adaptation, pp. 3–21, 1994.
[33]W. Wayne, "Cyber-physical Systems," IEEE Computer, vol. 42, pp. 88-89, 2009.
[34]K. Ashton, "That Internet of Things Thing," RFID Journal, 2009.
[35]J. H. Holland, "Hidden order: How adaptation builds complexity," Basic Books, 1996.
[36]M. Girvan, and M. Newman, "Community structure in social and biological
networks," Proceedings of the National Academy of Sciences of the United States of
America, vol. 99, no. 12, pp. 7821, 2002.
[37]J. H. Miller, and S. E. Page, "Complex Adaptive Systems: An Introduction to
Computational Models of Social Life," Princeton University Press, 2007.
[38]K. Kaneko, "Life: An introduction to complex systems biology," Springer Heidelberg,
Germany:, 2006.
[39]J. A. Thomson, J. Itskovitz-Eldor, S. S. Shapiro et al., "Embryonic stem cell lines
derived from human blastocysts," Science, vol. 282, no. 5391, pp. 1145, 1998.
[40]K. Nair, "Tropical forest insect pests: ecology, impact, and management," Cambridge
Univ Pr, 2007.
[41]M. D. Lowman, "Leaf growth dynamics and herbivory in five species of Australian
rain-forest canopy trees," Journal of Ecology, pp. 433-447, 1992.
[42]S. L. Lima, "Nonlethal effects in the ecology of predator-prey interactions,"
Bioscience, vol. 48, no. 1, pp. 25-34, 1998.
[43]A. M. Hein, and J. F. Gillooly, "Predators, prey, and transient states in the assembly of
spatially structured communities," Ecology, vol. 92, no. 3, pp. 549-555, 2011.
[44]H. C. Berg, "E. coli in Motion," Springer Verlag, 2004.
[45]Y. Bar Yam, "A mathematical theory of strong emergence using multiscale variety,"
Complexity, vol. 9, no. 6, pp. 15-24, 2004.
- 266 -
[46]L. Leydesdorff, "The challenge of scientometrics: The development, measurement, and
self-organization of scientific communications," Universal-Publishers, 2001.
[47]C. Chen, R. Paul, and B. O'Keefe, "Fitting the jigsaw of citation: Information
visualization in domain analysis," Journal of the American Society for Information
Science and Technology, vol. 52, no. 4, pp. 315-330, 2001.
[48]S. Card, J. Mackinlay, and B. Shneiderman, "Readings in information visualization:
using vision to think," Morgan Kaufmann, 1999.
[49]H. White, and K. McCain, "Visualizing a discipline: An author co-citation analysis of
information science, 1972-1995," Journal of the American Society for Information
Science, vol. 49, no. 4, pp. 327-355, 1998.
[50]A. Pouris, and A. Pouris, "Scientometrics of a pandemic: HIV/AIDS research in South
Africa and the World," Scientometrics, pp. 1-12, 2010.
[51]H. Hou, H. Kretschmer, and Z. Liu, "The structure of scientific collaboration networks
in Scientometrics," Scientometrics, vol. 75, no. 2, pp. 189-202, 2008.
[52]M. Sierra-Flores, M. Guzmán, A. Raga, and I. Pérez, "The productivity of Mexican
astronomers in the field of outflows from young stars," Scientometrics, vol. 81, no.
3, pp. 765-777, 2009.
[53]A. Barabási, H. Jeong, Z. Néda et al., "Evolution of the social network of scientific
collaborations," Physica A: Statistical Mechanics and its Applications, vol. 311, no.
3-4, pp. 590-614, 2002.
[54]R. Sooryamoorthy, "Scientific publications of engineers in South Africa, 1975–2005,"
Scientometrics, pp. 1-16, 2010.
[55]C. Chen, G. Panjwani, J. Proctor et al., "Visualizing the Evolution of HCI," People and
Computers XIX-The Bigger Picture, pp. 233-250, 2006.
[56]E. Sandström, and U. Sandström, "CiteSpace Visualization of Proximity Clusters in
Dentistry Research." pp. 25–28.
[57]R. Zhao, and J. Wang, "Visualizing the research on pervasive and ubiquitous
computing," Scientometrics, pp. 1-20, 2010.
[58]L. Chun-juan, C. Yue, and H. Hai-yan, "Scientometric & Visualization Analysis of
Innovation Studies International," Technology and Innovation Management, vol. 1,
2010.
[59]H. White, and B. Griffith, "Author cocitation: A literature measure of intellectual
structure," Journal of the American Society for Information Science, vol. 32, no. 3,
pp. 163-171, 1981.
[60]H. D. White, and K. W. McCain, "Visualizing a discipline: An author co-citation
analysis of information science, 1972–1995," Journal of the American Society for
Information Science, vol. 49, no. 4, pp. 327-355, 1998.
[61]H. Small, "Co citation in the scientific literature: A new measure of the relationship
between two documents," Journal of the American Society for Information Science,
vol. 24, no. 4, pp. 265-269, 1973.
[62]H. Small, and B. Griffith, "The structure of scientific literatures I: Identifying and
graphing specialties," Science studies, pp. 17-40, 1974.
[63]B. Griffith, H. Small, J. Stonehill, and S. Dey, "The structure of scientific literatures II:
Toward a macro-and microstructure for science," Science studies, vol. 4, no. 4, pp.
339-365, 1974.
[64]H. Small, "Macro-level changes in the structure of co-citation clusters: 1983–1989,"
Scientometrics, vol. 26, no. 1, pp. 5-20, 1993.
[65]E. Garfield, "Use of Journal Citation Reports and Journal Performance Indicators in
measuring short and long term journal impact," Croatian medical journal, vol. 41,
no. 4, pp. 368-374, 2000.
- 267 -
[66]M. Amin, and M. Mabe, "Impact factors: use and abuse," Medicina [Buenos Aires],
vol. 63, pp. 347-354, 2003.
[67]T. Braun, I. Dióspatonyi, S. Zsindely, and E. Zádor, "Gatekeeper index versus impact
factor of science journals," Scientometrics, vol. 71, no. 3, pp. 541-543, 2007.
[68]A. Fersht, "The most influential journals: Impact Factor and Eigenfactor," Proceedings
of the National Academy of Sciences, vol. 106, no. 17, pp. 6883, 2009.
[69]E. Garfield, "The history and meaning of the journal impact factor," Jama, vol. 295,
no. 1, pp. 90, 2006.
[70]H. Moed, T. Van Leeuwen, and J. Reedijk, "A critical analysis of the journal impact
factors of Angewandte Chemie and the journal of the American Chemical Society
inaccuracies in published impact factors based on overall citations only,"
Scientometrics, vol. 37, no. 1, pp. 105-116, 1996.
[71]M. Niazi, H. F. Ahmed, and A. Ali, "Introducing fault-tolerance and responsiveness in
web applications using SREFTIA," in
Proceedings of the International Multiconference on
Computer Science and Information Technology, Wisla, Poland, 2006, pp. 271-278.
[72]H. S. Gunawi, T. Do, J. M. Hellerstein et al., Failure as a Service (FaaS): A Cloud
Service for Large-Scale, Online Failure Drills, University of California, Berkeley,
Berkeley, 2011.
[73]F. Dressler, "Self-Organization in Sensor and Actor Networks," Wiley, 2007.
[74]B. Cohen, "Incentives build robustness in BitTorrent." pp. 68–72.
[75]C. W. Reynolds, "Flocks, herds and schools: A distributed behavioral model," in
Proceedings of the 14th annual conference on Computer graphics and interactive
techniques, 1987, pp. 25-34.
[76]E. Fischbein, M. Deri, M. S. Nello, and M. S. Marino, "The role of implicit models in
solving verbal problems in multiplication and division," Journal for Research in
Mathematics Education, vol. 16, no. 1, pp. 3-17, 1985.
[77]Z. N. O. Albert Laszlo Barabasi, "Network Biology: Understanding the cell's
functional organization," Nature Reviews, vol. 5, pp. 101-114, 2004.
[78]V. Batagelj, "Efficient algorithms for citation network analysis," Arxiv preprint
cs/0309023, 2003.
[79]K. Thulasiraman, and M. N. S. Swamy, "Graphs: theory and algorithms," Wiley
Online Library, 1992.
[80]P. Erdős, and A. Rényi, "On the evolution of random graphs," Magyar Tud. Akad. Mat.
Kutató Int. Közl., vol. 5, pp. 17–61, 1960.
[81]D. J. Watts, and S. H. Strogatz, "Collective dynamics of 'small-world'networks,"
Nature, vol. 393, no. 6684, pp. 440-442, 1998.
[82]A. L. Barabási, and R. Albert, "Emergence of scaling in random networks," Science,
vol. 286, no. 5439, pp. 509, 1999.
[83]F. M. Hulett, "The signal transduction network for Pho regulation in Bacillus subtilis,"
Molecular microbiology, vol. 19, no. 5, pp. 933-939, 1996.
[84]E. H. Davidson, "The regulatory genome: gene regulatory networks in development
and evolution," Academic Press, 2006.
[85]E. H. Davidson, "Emerging properties of animal gene regulatory networks," Nature,
vol. 468, no. 7326, pp. 911-920, 2010.
[86]Z. N. Oltvai, and A. L. Barabási, "Life's complexity pyramid," Science, vol. 298, no.
5594, pp. 763, 2002.
[87]N. Team, "Network Workbench Tool," Indiana University, Northeastern University,
and University of Michigan, 2006.
- 268 -
[88]V. Batagelj, and A. Mrvar, "Pajek-program for large network analysis," Connections,
vol. 21, no. 2, pp. 47-57, 1998.
[89]C. Chen, "CiteSpace II: Detecting and visualizing emerging trends and transient
patterns in scientific literature," Journal of the American Society for Information
Science and Technology, vol. 57, no. 3, pp. 359-377, 2006.
[90]P. Shannon, A. Markiel, O. Ozier et al., "Cytoscape: a software environment for
integrated models of biomolecular interaction networks," Genome research, vol.
13, no. 11, pp. 2498, 2003.
[91]M. Baur, M. Benkert, U. Brandes et al., "Visone Software for visual social network
analysis." pp. 554-557.
[92]M. Wooldridge, "Agent-based computing," Interoperable Communication Networks,
vol. 1, pp. 71-98, 1998.
[93]N. Jennings, "Agent-based computing: Promise and perils," in 16th Int. Joint Conf. on
Artificial Intelligence (IJCAI-99), Stockholm, Sweden, 1999, pp. 1429-1436.
[94]S. R. a. P. Norvig, "Artificial Intelligence A Modern Approach," Second ed.: Prentice
Hall, 2003.
[95]L. Panait, and S. Luke, "Cooperative multi-agent learning: The state of the art,"
Autonomous Agents and Multi-Agent Systems, vol. 11, no. 3, pp. 387-434, 2005.
[96]H. James, "Where Are All the Intelligent Agents?," vol. 22, pp. 2-3, 2007.
[97]T. Finin, R. Fritzson, D. McKay, and R. McEntire, "KQML as an agent communication
language." p. 463.
[98]J. Hendler, and D. McGuinness, "The DARPA agent markup language," IEEE
Intelligent Systems, vol. 15, no. 6, pp. 67-73, 2000.
[99]D. Connolly, F. van Harmelen, I. Horrocks et al., "Daml+ oil (march 2001) reference
description, December 2001," Internetquelle: http://www. w3. org/TR/daml+ oil-
reference, heruntergeladen am, vol. 5, 2007.
[100]V. Quera, F. S. Beltran, and R. Dolado, "Flocking Behaviour: Agent-Based
Simulation and Hierarchical Leadership," Journal of Artificial Societies and Social
Simulation, vol. 13, no. 2, pp. 8, 2010.
[101]V. Grimm, U. Berger, F. Bastiansen et al., "A standard protocol for describing
individual-based and agent-based models," Ecological Modelling, vol. 198, no. 1-2,
pp. 115-126, 2006.
[102]S. F. Railsback, and V. Grimm, "Agent-based and Individual-based Modeling: A
Practical Introduction," Princeton University Press, 2011.
[103]S. C. Bankes, "Agent-based modeling: A revolution?," 90003, National Acad
Sciences, 2002, pp. 7199-7200.
[104]E. Bonabeau, "Agent-based modeling: Methods and techniques for simulating human
systems," Proceedings of the National Academy of Sciences of the United States of
America, vol. 99, no. Suppl 3, pp. 7280, 2002.
[105]H. Devillers, J. R. Lobry, and F. Menu, "An agent-based model for predicting the
prevalence of Trypanosoma cruzi I and II in their host and vector populations,"
Journal of Theoretical Biology, vol. 255, no. 3, pp. 307-315, 2008.
[106]G. M. Dancik, D. E. Jones, and K. S. Dorman, "Parameter estimation and sensitivity
analysis in an agent-based model of Leishmania major infection," J Theor Biol, vol.
262, no. 3, pp. 398-412, Feb 7, 2010.
[107]V. A. Folcik, G. C. An, and C. G. Orosz, "The Basic Immune Simulator: an agent-
based model to study the interactions between innate and adaptive immunity,"
Theor Biol Med Model, vol. 4, pp. 39, 2007.
[108]V. Galvao, and J. G. Miranda, "A three-dimensional multi-agent-based model for the
evolution of Chagas' disease," Biosystems, vol. 100, no. 3, pp. 225-30, Jun, 2010.
- 269 -
[109]V. Galvao, J. G. Miranda, and R. Ribeiro-dos-Santos, "Development of a two-
dimensional agent-based model for chronic chagasic cardiomyopathy after stem
cell transplantation," Bioinformatics, vol. 24, no. 18, pp. 2051-6, Sep 15, 2008.
[110]Z. Guo, P. M. A. Sloot, and J. C. Tay, "A hybrid agent-based approach for modeling
microbiological systems," Journal of Theoretical Biology, vol. 255, no. 2, pp. 163-
175, 2008.
[111]C. F. Huang, J. Kaur, A. Maguitman, and L. M. Rocha, "Agent-based model of
genotype editing," Evol Comput, vol. 15, no. 3, pp. 253-89, Fall, 2007.
[112]J. Itakura, M. Kurosaki, Y. Itakura et al., "Reproducibility and usability of chronic
virus infection model using agent-based simulation; comparing with a
mathematical model," Biosystems, vol. 99, no. 1, pp. 70-8, Jan, 2010.
[113]M. Kiran, S. Coakley, N. Walkinshaw, P. McMinn, and M. Holcombe, "Validation
and discovery from computational biology models," Biosystems, vol. 93, no. 1-2,
pp. 141-150, 2008.
[114]B. J. Lao, and D. T. Kamei, "Investigation of cellular movement in the prostate
epithelium using an agent-based model," J Theor Biol, vol. 250, no. 4, pp. 642-54,
Feb 21, 2008.
[115]A. M. Bailey, M. B. Lawrence, H. Shang, A. J. Katz, and S. M. Peirce, "Agent-based
model of therapeutic adipose-derived stromal cell trafficking during ischemia
predicts ability to roll on P-selectin," PLoS Comput Biol, vol. 5, no. 2, pp.
e1000294, Feb, 2009.
[116]C. Carpenter, and L. Sattenspiel, "The design and use of an agent-based model to
simulate the 1918 influenza epidemic at Norway House, Manitoba," Am J Hum
Biol, vol. 21, no. 3, pp. 290-300, May-Jun, 2009.
[117]M. A. Rubin, J. Mayer, T. Greene et al., "An agent-based model for evaluating
surveillance methods for catheter-related bloodstream infection," AMIA Annu Symp
Proc, pp. 631-5, 2008.
[118]G. M. Odell, and V. E. Foe, "An agent-based model contrasts opposite effects of
dynamic and stable microtubules on cleavage furrow positioning," J Cell Biol, vol.
183, no. 3, pp. 471-83, Nov 3, 2008.
[119]E. J. Robinson, F. L. Ratnieks, and M. Holcombe, "An agent-based model to
investigate the roles of attractive and repellent pheromones in ant decision making
during foraging," J Theor Biol, vol. 255, no. 2, pp. 250-8, Nov 21, 2008.
[120]D. Santoni, M. Pedicini, and F. Castiglione, "Implementation of a regulatory gene
network to simulate the TH1/2 differentiation in an agent-based model of
hypersensitivity reactions," Bioinformatics, vol. 24, no. 11, pp. 1374-80, Jun 1,
2008.
[121]M. Batty, "Cities and complexity: understanding cities with cellular automata, agent-
based models, and fractals," The MIT press, 2007.
[122]X. Chen, and F. Zhan, "Agent-based modelling and simulation of urban evacuation:
relative effectiveness of simultaneous and staged evacuation strategies," Journal of
the Operational Research Society, vol. 59, no. 1, pp. 25-33, 2008.
[123]R. E. Streit, and D. Borenstein, "An agent-based simulation model for analyzing the
governance of the Brazilian Financial System," Expert Systems with Applications,
vol. 36, no. 9, pp. 11489-11501, 2009.
[124]M. H. F. Zarandi, M. Pourakbar, and I. B. Turksen, "A Fuzzy agent-based model for
reduction of bullwhip effect in supply chain systems," Expert Systems with
Applications, vol. 34, no. 3, pp. 1680-1691, 2008.
- 270 -
[125]T. Galla, "Independence and interdependence in the nest-site choice by honeybee
swarms: Agent-based models, analytical approaches and pattern formation,"
Journal of Theoretical Biology, vol. 262, no. 1, pp. 186-196, 2010.
[126]C. Gershenson, "Self-organizing traffic lights," Arxiv preprint nlin/0411066, 2004.
[127]T. Carmichael, "Complex adaptive systems and the threshold effect: Towards a
general tool for studying dynamic phenomena across diverse domains," THE
UNIVERSITY OF NORTH CAROLINA AT CHARLOTTE.
[128]U. Wilensky, "NetLogo," Center for Connected Learning Comp.-Based Modeling,
Northwestern University, vol. Evanston, IL, 1999.
[129]M. Resnick, "StarLogo: an environment for decentralized modeling and decentralized
thinking." pp. 11-12.
[130]S. W. Pages. "Code Templates for Agent-based Models," August, 2011;
http://www.swarm.org/index.php/Software_templates.
[131]J. M. Vidal, P. Buhler, and H. Goradia, "The Past and Future of Multiagent Systems,"
AAMAS Workshop on Teaching Multi-Agent Systems, 2004.
[132]J. McCarthy, "LISP 1.5 programmer's manual," 1965.
[133]J. Banks, J. S. C. II, B. L. Nelson, and D. M. Nicol, "Discrete-Event System
Simulation," Fourth ed.: Peason Education, 2005.
[134]A. M. Law, "How to build valid and credible simulation models," in Winter
Simulation Conference Miami, FL, 2008, pp. 39-47.
[135]T. H. Naylor, and J. M. Finger, "Verification of Computer Simulation Models,"
Management Science, vol. 2, pp. B92-B101, 1967.
[136]J. M. Galán, L. R. Izquierdo, S. S. Izquierdo et al., "Errors and Artefacts in Agent-
Based Modelling," Journal of Artificial Societies and Social Simulation, vol. 12,
no. 11, 2009.
[137]T. W. Lucas, S. M. Sanchez, F. Martinez, L. R. Sickinger, and J. W. Roginski,
"Defense and homeland security applications of multi-agent simulations," in
Proceedings of the 39th conference on Winter simulation: 40 years! The best is yet
to come, Washington D.C., 2007.
[138]C. Bianchi, P. Cirillo, M. Gallegati, and P. A. Vagliasindi, "Validating and Calibrating
Agent-Based Models: A Case Study," Comput. Econ., vol. 30, no. 3, pp. 245-264,
2007.
[139]J. S. Hodges, and J. A. Dewar, Is It You or Your Model Talking? A Framework for
Model Validation, RAND Corporation, Santa Monica, California.
[140]G. Fagiolo, C. Birchenhall, and P. Windrum, "Empirical Validation in Agent-based
Models: Introduction to the Special Issue " Computational Economics, vol. 30, no.
3, pp. 189-194, October, 2007, 2007.
[141]C. M. Macal, and M. J. North, "Agent-based modeling and simulation: desktop
ABMS," in Proceedings of the 39th conference on Winter simulation: 40 years!
The best is yet to come, Washington D.C., 2007.
[142]R. Axtell, "The emergence of firms in a population of agents: local increasing returns,
unstable Nash equilibria, and power law size distributions," The Brookings
Institution CSED Working Paper, vol. 3, pp. 138, 2000.
[143]R. Axtell, J. Epstein, and H. Young, "The emergence of classes in a multiagent
bargaining model," Social dynamics, pp. 191–211, 1999.
[144]S. Moss, "Alternative Approaches to the Empirical Validation of Agent-Based
Models," Journal of Artificial Societies and Social Simulation, vol. 11, no. 15,
2008.
- 271 -
[145]U. Wilensky, and W. Rand, "Making Models Match: Replicating an Agent-Based
Model," Journal of Artificial Societies and Social Simulation, vol. 10, no. 4, pp. 2,
10/31, 2007.
[146]A. Schmid, "What is the Truth of Simulation?," Journal of Artificial Societies and
Social Simulation, vol. 8, no. 4, pp. 5, 10/31, 2005.
[147]O. Barreteau, and e. al., "Our Companion Modelling Approach " Journal of Artificial
Societies and Social Simulation, vol. 6, no. 1, 2003.
[148]M. Makowsky, "An Agent-Based Model of Mortality Shocks, Intergenerational
Effects, and Urban Crime," Journal of Artificial Societies and Social Simulation,
vol. 9, No. 2, 2006.
[149]L. Cernuzzi, M. Cossentino, and F. Zambonelli, "Process models for agent-based
development," Engineering Applications of Artificial Intelligence, vol. 18, no. 2,
pp. 205-222, 2005.
[150]P. Garrido, M. Malumbres, and C. Calafate, "ns-2 vs. OPNET: a comparative study of
the IEEE 802.11 e technology on MANET environments." p. 37.
[151]A. Sobeih, W. Chen, J. Hou et al., "J-sim: A simulation environment for wireless
sensor networks." pp. 175-187.
[152]P. Levis, N. Lee, M. Welsh, and D. Culler, "TOSSIM: Accurate and scalable
simulation of entire TinyOS applications." pp. 126-137.
[153]S. Park, A. Savvides, and M. B. Srivastava, "SensorSim: a simulation framework for
sensor networks," in Proceedings of the 3rd ACM international workshop on
Modeling, analysis and simulation of wireless and mobile systems, Boston,
Massachusetts, United States, 2000, pp. 104-111.
[154]J. Polley, D. Blazakis, J. McGee et al., "Atemu: A fine-grained sensor network
simulator."
[155]I. Baumgart, B. Heep, and S. Krause, "OverSim: A flexible overlay network
simulation framework." pp. 79-84.
[156]A. Varga, "The OMNeT++ discrete event simulation system," in Proceedings of the
European Simulation Multiconference ESM'2001 (2001), Prague, 2001, pp. 319–
324.
[157]M. Jelasity, A. Montresor, G. P. Jesi, and S. Voulgaris, "The Peersim simulator," 2008.
[158]F. Mondada, G. C. Pettinaro, A. Guignard et al., "SWARM-BOT: A new distributed
robotic concept," Autonomous Robots, vol. 17, no. 2, pp. 193-221, 2004.
[159]O. Michel, "WebotsTM: Professional Mobile Robot Simulation," International
Journal of Advanced Robotic Systems, vol. 1, no. 1, pp. 39-42, 2004.
[160]E. Sahin, S. Girgin, L. Bay nd r, and A. E. Turgut, "Swarm robotics," Swarm
Intelligence: Introduction and Applications, vol. 1, pp. 87, 2008.
[161]F. Bellifemine, A. Poggi, and G. Rimassa, "Developing multi-agent systems with
JADE," Intelligent Agents VII Agent Theories Architectures and Languages, pp. 42-
47, 2001.
[162]M. Pullen, "The Network Workbench: network simulation software for academic
investigation of Internet concepts," Computer Networks, vol. 32, no. 3, pp. 365-
378, 2000.
[163]A. Barabasi, "The origin of bursts and heavy tails in human dynamics," Nature, vol.
435, no. 7039, pp. 207-211, 2005.
[164]A. Barabási, and A. Gelman, "Bursts: The Hidden Pattern Behind Everything We
Do," Physics Today, vol. 63, pp. 46, 2010.
[165]T. Reuters, "Web of science," Online factsheet Thomson Reuters, Philadelphia,
Pennsylvania (Available from:
- 272 -
www.thomsonreuters.com/content/PDF/scientific/Web_of_Science_factsheet.pdf ),
2008.
[166]R. Sedgewick, and M. Schidlowsky, "Algorithms in Java, Part 5: Graph Algorithms,"
Addison-Wesley Longman Publishing Co., Inc. Boston, MA, USA, 2003.
[167]E. Adar, "Guess: a language and interface for graph exploration." p. 800.
[168]R. W. Keyes, "The evolution of digital electronics towards VLSI," Solid-State
Circuits, IEEE Journal of, vol. 14, no. 2, pp. 193-201, 1979.
[169]S. Gamm, and R. Haeb-Umbach, "User interface design of voice controlled consumer
electronics," Philips Journal of Research, vol. 49, no. 4, pp. 439-454, 1995.
[170]J. F. Christensen, M. H. Olesen, and J. S. Kjaer, "The industrial dynamics of Open
Innovation--Evidence from the transformation of consumer electronics," Research
Policy, vol. 34, no. 10, pp. 1533-1549, 2005.
[171]R. S. Rosenbloom, and W. J. Abernathy, "The climate for innovation in industry : The
role of management attitudes and practices in consumer electronics," Research
Policy, vol. 11, no. 4, pp. 209-225, 1982.
[172]R. Hisano, and T. Mizuno, "Sales distribution of consumer electronics," Physica A:
Statistical Mechanics and its Applications, vol. 390, no. 2, pp. 309-318, 2011.
[173]S. F. Railsback, S. L. Lytinen, and S. K. Jackson, "Agent-based simulation platforms:
Review and development recommendations," SIMULATION, vol. 82, no. 9, pp.
609, 2006.
[174]N. Gershenfeld, R. Krikorian, and D. Cohen, "The Internet of Things," Scientific
American, vol. 291, no. 4, pp. 76-81, 2004.
[175]R. Groenevelt, E. Altman, and P. Nain, "Relaying in mobile ad hoc networks: the
Brownian motion mobility model," Wireless Networks, vol. 12, no. 5, pp. 561-571,
2006.
[176]M. C. Gonzalez, C. A. Hidalgo, and A. L. Barabási, "Understanding individual
human mobility patterns," Nature, vol. 453, no. 7196, pp. 779-782, 2008.
[177]U. Wilensky. "NetLogo Users Discussion Group," August, 2011;
http://groups.yahoo.com/group/netlogo-users/.
[178]U. Wilensky. "NetLogo Educators Mailing List," August, 2011;
http://groups.yahoo.com/group/netlogo-educators/.
[179]D. L. Parnas, "Precise Documentation: The Key to Better Software," The Future of
Software Engineering, S. Nanz, ed., pp. 125-148: Springer Berlin Heidelberg, 2011.
[180]Y. Li, "Reengineering a scientific software and lessons learned." pp. 41-45.
[181]C. M. Macal, and M. North. "Repast Interest users Mailing List," August, 2011;
https://lists.sourceforge.net/lists/listinfo/repast-interest.
[182]A. I. Concepcion, "DEVS Formalism: A Framework for Hierarchical Model
Development," IEEE Transactions on Software Engineering, vol. 14, pp. 228-241,
1988.
[183]B. P. Zeigler, "Multifaceted modeling methodology: Grappling with the irreducible
complexity of systems," Behavioral Science, vol. 29, no. 3, pp. 169-178, 2007.
[184]T. Buzan, and B. Buzan, "The mind map book," Pearson Education, 2006.
[185]A. Robins, J. Rountree, and N. Rountree, "Learning and teaching programming: A
review and discussion," Computer Science Education, vol. 13, no. 2, pp. 137-172,
2003.
[186]J. C. Reynolds, "The essence of Algol." pp. 67-88.
[187]U. Wilensky, "NetLogo Flocking Model," Center for Connected Learning and
Computer-Based Modeling, Northwestern University, 1998.
[188]A. Boukerche, "Algorithms and protocols for wireless sensor networks," Wiley-IEEE
press, 2009.
- 273 -
[189]G. Anastasi, M. Conti, M. Di Francesco, and A. Passarella, "Energy conservation in
wireless sensor networks: A survey," Ad Hoc Networks, vol. 7, no. 3, pp. 537-568,
2009.
[190]J. Zheng, and A. Jamalipour, "Wireless Sensor Networks: A Networking Perspective,"
Wiley-IEEE Press, 2009.
[191]J. N. Al-Karaki, R. Ul-Mustafa, and A. E. Kamal, "Data aggregation and routing in
Wireless Sensor Networks: Optimal and heuristic algorithms," Computer Networks,
vol. 53, no. 7, pp. 945-960, 2009.
[192]M. Younis, and K. Akkaya, "Strategies and techniques for node placement in wireless
sensor networks: A survey," Ad Hoc Networks, vol. 6, no. 4, pp. 621-655, 2008.
[193]J. M. Kahn, R. H. Katz, and K. S. J. Pister, "Next century challenges: mobile
networking for "Smart Dust"." pp. 271-278.
[194]C. Colombo, and C. Mcinnes, "Orbital dynamics of 'smart dust' devices with solar
radiation pressure and drag," Journal of Guidance, Control and Dynamics, 2011.
[195]W. Kim, K. Mechitov, J. Y. Choi, and S. Ham, "On target tracking with binary
proximity sensors." pp. 301-308.
[196]A. S. Rao, and M. P. Georgeff, "BDI agents: From theory to practice," in First
International Conference on Multiagent Systems, San Francisco, 1995, pp. 312–
319.
[197]R. Cappelli, M. Ferrara, and D. Maltoni, "Fingerprint Indexing Based on Minutia
Cylinder-Code," Pattern Analysis and Machine Intelligence, IEEE Transactions on,
no. 99, pp. 1-1, 2011.
[198]L. C. Freeman, "A set of measures of centrality based on betweenness," Sociometry,
pp. 35-41, 1977.
[199]V. Grimm, U. Berger, D. L. DeAngelis et al., "The ODD protocol: A review and first
update," Ecological Modelling, vol. 221, no. 23, pp. 2760-2768, 2010.
[200]B. Edmonds, and D. Hales, "Replication, replication and replication: Some hard
lessons from model alignment," Journal of Artificial Societies and Social
Simulation, vol. 6, no. 4, 2003.
[201]R. Axelrod, and W. D. Hamilton, "The evolution of cooperation," Science, vol. 211,
no. 4489, pp. 1390, 1981.
[202]G. Fagiolo, A. Moneta, and P. Windrum, "A Critical Guide to Empirical Validation of
Agent-Based Models in Economics: Methodologies, Procedures, and Open
Problems," Computational Economics, vol. 30, no. 3, pp. 195-226, 2007.
[203]R. G. Sargent, "Verification and validation of simulation models," in Proceedings of
the 40th Conference on Winter Simulation, Miami, Florida, 2008, pp. 157-169.
[204]F. P. Brooks, "The mythical man-month," Addison-Welsley, 1995.
[205]Y. Le Traon, B. Baudry, and J. M. Jezequel, "Design by Contract to Improve Software
Vigilance," Software Engineering, IEEE Transactions on, vol. 32, no. 8, pp. 571-
586, 2006.
[206]J. M. Jazequel, and B. Meyer, "Design by contract: the lessons of Ariane," Computer,
vol. 30, no. 1, pp. 129-130, 1997.
[207]U. Wilensky, "NetLogo Fire model," Center for Connected Learning and Computer-
Based Modeling, Northwestern University 1997.
[208]A. Alexandridis, D. Vakalis, C. I. Siettos, and G. V. Bafas, "A cellular automata model
for forest fire spread prediction: The case of the wildfire that swept through Spetses
Island," Applied Mathematics and Computation, vol. 204, no. 1, pp. 191-201, 2008.
[209]R. Axtell, R. Axelrod, J. M. Epstein, and M. D. Cohen, "Aligning simulation models:
A case study and results," Computational & Mathematical Organization Theory,
vol. 1, no. 2, pp. 123-141, 1996.
- 274 -
[210]B. Lawson, P. F. R. Centre, and J. Turner, "Weather in the Canadian Forest Fire
Danger Rating System: a user guide to national standards and practices,"
Canadian Forestry Service, 1978.
[211]C. E. Van Wagner, "Structure of the Canadian forest fire weather index," Environment
Canada, Forestry Service, 1974.
[212]B. J. Stocks, T. J. Lynham, B. D. Lawson et al., "The Canadian Forest Fire Danger
Rating System: An Overview," The Forestry Chronicle, vol. 65, no. 6, pp. 450-457,
1989.
[213]W. J. de Groot, "Interpreting the Canadian Forest fire Weather Index (FWI) System,"
In proc of the fourth Central Region Fire Weather Index Committee Scientific and
technical Seminar, Edmonton Canada, 1998.
[214]B. N. Clark, C. J. Colbourn, and D. S. Johnson, "Unit disk graphs," Discrete
mathematics, vol. 86, no. 1-3, pp. 165-177, 1990.
[215]Y. Wang, "Topology control for wireless sensor networks," Wireless Sensor Networks
and Applications, pp. 113-147, 2008.
[216]F. Kuhn, R. Wattenhofer, and A. Zollinger, "Ad hoc networks beyond unit disk
graphs," Wireless Networks, vol. 14, no. 5, pp. 715-729, 2008.
[217]J. Hirsch, "An index to quantify an individual's scientific research output,"
Proceedings of the National Academy of Sciences, vol. 102, no. 46, pp. 16569,
2005.
[218]L. Egghe, "The Hirsch-index and related impact measures," Annual Review of
Information Science and Technology, vol. 44, pp. 65-114, 2010.
[219]A. Harzing, and R. van der Wai, "Google Scholar as a new source for citation
analysis," Ethics in Science and Environmental Politics(ESEP), vol. 8, no. 1, pp.
61-73, 2008.
[220]A. Harzing, and R. van der Wal, "A Google Scholar h-index for journals: An
alternative metric to measure journal impact in economics and business," Journal
of the American Society for Information Science and Technology, vol. 60, no. 1, pp.
41-46, 2009.
[221]R. Marks, "Validating Simulation Models: A General Framework and Four Applied
Examples," Computational Economics, vol. 30, no. 3, pp. 265-290, 2007.
[222]F. Bousquet, and G. Trébuil, "Introduction to companion modeling and multi-agent
systems for integrated natural resource management in Asia," Companion Modeling
and Multi-Agent Systems for Integrated Natural Resource Management in Asia, pp.
1-20, 2005.
- 275 -
|
1301.1848 | 2 | 1301 | 2013-01-15T05:35:56 | The forest consensus theorem | [
"cs.MA",
"cs.DM",
"eess.SY",
"math.CO",
"math.OC"
] | We show that the limiting state vector in the differential model of consensus seeking with an arbitrary communication digraph is obtained by multiplying the eigenprojection of the Laplacian matrix of the model by the vector of initial states. Furthermore, the eigenprojection coincides with the stochastic matrix of maximum out-forests of the weighted communication digraph. These statements make the forests consensus theorem. A similar result for DeGroot's iterative pooling model requires the Cesaro (time-average) limit in the general case. The forests consensus theorem is useful for the analysis of consensus protocols. | cs.MA | cs |
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, 2013
1
The forest consensus theorem
Pavel Chebotarev, Rafig Agaev
Abstract
We show that the limiting state vector in the differential model of consensus seeking with an arbitrary
communication digraph is obtained by multiplying the eigenprojection of the Laplacian matrix of the
model by the vector of initial states. Furthermore, the eigenprojection coincides with the stochastic
matrix of maximum out-forests of the weighted communication digraph. These statements make the
forest consensus theorem. A similar result for DeGroot's iterative pooling model requires the Ces`aro
(time-average) limit in the general case. The forest consensus theorem is useful for the analysis of
consensus protocols.
Consensus, Eigenprojection, Matrix exponent, DeGroot's iterative pooling.
Index Terms
The continuous-time model of consensus seeking in a multiagent system has the form [22],
[10]
I. INTRODUCTION
xi(t) = ui(t),
n
ui(t) = −
X
j=1
aij (xi(t) − xj(t)) ,
i = 1, . . . , n,
(1)
(2)
where xi(t) is the state of the i'th agent and aij ≥ 0 is the weight with which agent i takes into
account the discrepancy with agent j. The matrix form of the model (1) -- (2) is:
where x(t) = (x1(t), . . . , xn(t))T , L is the Laplacian matrix of the model (1) -- (2):
x(t) = −L x(t),
L = diag(A 1) − A,
(3)
(4)
A = (aij)n×n, and 1 = (1, . . . , 1)T .
The nonsymmetric Laplacian matrices of this kind were studied in [1], [9], [5].
In this paper, we present the forest consensus theorem stating that for an arbitrary non-negative
matrix A and any trajectory x(t) satisfying (1) -- (2),
x(t) = J x(0)
lim
t→∞
P. Chebotarev and R. Agaev are with the Institute of Control Sciences of the Russian Academy of Sciences, 65 Profsoyuznaya
Street, Moscow 117997, Russia.
August 4, 2018
DRAFT
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, 2013
2
holds, where J is the eigenprojection of L, which coincides with the matrix ¯J of maximum
out-forests of the communication digraph corresponding to A.
A similar result, which involves the Ces`aro limit, holds for the discretization of the model
(1) -- (2).
The paper is organized as follows. After introducing the necessary notation and summarizing
the preliminary results, in Section III we prove the forest consensus theorem. Section IV is
devoted to the properties of the limiting state of the model. Section V contains a numerical
example; in Section VI, we show that the classical results on a communication digraph having
a spanning diverging tree immediately follow from the forest consensus theorem. Finally,
Section VII presents a counterpart of the forest consensus theorem for the discretization of
the model (1) -- (2).
II. BASIC CONCEPTS AND PRELIMINARY RESULTS
A. Eigenprojections and functions of matrices
Let A ∈ Cn×n be an arbitrary square matrix. Let ν = ind A be the index of A, i.e., the
smallest k ∈ {0, 1, . . .} such that rank Ak+1 = rank Ak, where A0 is identified with the identity
matrix I. A is nonsingular iff ν = 0. The index of a singular matrix is the index of its eigenvalue
0, i.e., the multiplicity of 0 as the root of its minimal polynomial, or, equivalently, the size of
the largest Jordan block with zero diagonal in its Jordan form. If ν = 1 then the algebraic and
geometric multiplicities of 0 coincide (in this case, the eigenvalue 0 of A is called semisimple).
Let R(A) and N (A) be the range and the null space of A, respectively. The eigenprojection
[24] of A at eigenvalue 0 is1 a projection (i.e. an idempotent matrix) Z such that R(Z) = N (Aν)
and N (Z) = R(Aν). In other words, Z is a projection to N (Aν) along R(Aν). In the
case of a singular matrix A, following [25], we call Z the eigenprojection of A (without
mentioning eigenvalue 0). The eigenprojection is unique because an idempotent matrix is
uniquely determined by its range and null space (see, e.g., [7, Sections 2.4 and 2.6]).
Eigenprojections underlie the definition of the components of a matrix which, in turn, are
used to define ϕ(A) for differentiable functions ϕ : C → C (see either of [14, Chapter 5], [7,
Section 2.5], [17], [16]), in the theory of generalized inverse matrices, as well as in the numerous
applications of matrix analysis.
Let λ1, . . . , λs be all distinct eigenvalues of A; let νk be the index of λk defined as the index
of A − λkI. According to the theory of matrix components [14, Chapter 5], for any function
ϕ : C → C having finite derivatives ϕ(j)(λk) of the first νk − 1 orders at λ1, . . . , λs, ϕ(A) is
defined as follows:
s
νk−1
ϕ(A) :=
X
k=1
X
j=0
ϕ(j)(λk) Zkj,
(5)
where the derivative ϕ(0) of order 0 is the value of ϕ and Zkj are the components of A defined
by
Zkj = (j!)−1(A − λkI)j Zk0.
1Such an eigenprojection is also called the principal idempotent [15].
August 4, 2018
(6)
DRAFT
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, 2013
3
Here, the component Zk0 is the eigenprojection of A − λkI (k = 1, . . . , s) also called the
eigenprojection of A at λk.
For more details on eigenprojections, see, e.g., [3].
B. The stochastic matrix of maximum out-forests
A matrix A = (aij) of the model (1) -- (2) determines a weighted communication digraph Γ
with vertex set V (Γ) = {1, . . . , n}: Γ has the (j, i) arc with weight wji = aij whenever aij > 0
(i.e., when agent j influences agent i). Thus, arcs of Γ are oriented in the direction of influence;
the weight of an arc is the degree of influence.
A diverging tree is a weakly connected (i.e., its corresponding undirected graph is connected)
digraph in which one vertex, called the root, has indegree zero and the other vertices have
indegree one. A diverging tree is said to diverge from its root. Spanning diverging trees are also
called out-arborescences or out-branchings [20]. A diverging forest is a digraph all of whose
weak components (i.e., maximal weakly connected subdigraphs) are diverging trees. The roots
of these trees form the set of roots of the diverging forest.
Definition 1. Any spanning diverging forest of a digraph Γ is called an out-forest of Γ. An
out-forest F of Γ is a maximum out-forest of Γ if Γ has no out-forest with a greater number of
arcs than in F . The out-forest dimension of Γ is the number of components in any maximum
out-forest of Γ.
The weight of a weighted digraph is the product of its arc weights. The matrix ¯J = ( ¯J ij) of
maximum out-forests of a weighted digraph Γ is defined as follows:
¯J ij =
fij
f
,
i, j = 1, . . . , n,
(7)
where f is the total weight of all maximum out-forests of Γ and fij is the total weight of those of
them that have i belonging to a tree diverging from j. In Proposition 1, we list some properties
of L and ¯J (cf. [10], [8], [4]) which are useful for the analysis of consensus protocols.
Proposition 1. Let L be the Laplacian matrix of the model (1) -- (2). Let ¯J be the matrix of
maximum out-forests of the corresponding communication digraph Γ whose out-forest dimension
is d. Then:
1) L is singular (since L1 = 0);
2) If λ 6= 0 is an eigenvalue of L, then Re(λ) > 0 [2, Proposition 9];
3) ind L = 1 [9, Proposition 12];
4) rank L = n − d; rank ¯J = tr ¯J = d [1, Proposition 11];
5) ¯J is row stochastic as by definition, Pn
6) ¯J is the eigenprojection of L [9, Proposition 12], which implies that ¯J 2 = ¯J;
7) L ¯J = ¯JL = 0 [1, Theorem 5]; N ( ¯J) = R(L), R( ¯J) = N (L) (by items 3 and 6);
8) ¯J = lim
j=1 fij = f for any i ∈ {1, . . . , n};
(I + αL)−1 [1, Theorem 6];
α→∞
August 4, 2018
DRAFT
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, 2013
4
9) ¯J = C(0)(cid:14)h(0), where2 C(λ) is the quotient of the matrix polynomial λh(λ)I and the
binomial λI − L, λh(λ) being the minimal polynomial of L (this follows from [14, Eq. (22)
in Chapter 5]);
tr(L ¯J k−1) , k = 1, . . . , n − d,
10) ¯J = ¯J n−d, where ¯J n−d is defined recursively: ¯J k = I − k L ¯J k−1
¯J 0 = I, and L ¯J n−d = 0 [2, Section 4] or [9, Section 5].
An elementwise characterization of ¯J is given in [1, Theorem 2′].
Theorem 1. Let x(t) be a solution of (3). Then
III. THE FOREST CONSENSUS THEOREM
x(t) = J x(0),
lim
t→∞
(8)
where J is the eigenprojection of L. Moreover, J coincides with the matrix ¯J of maximum
out-forests of the communication digraph corresponding to L.
Proof. All solutions of (3) satisfy the identity [14, Eq. (46) in Chapter 5]
x(t) = e−Ltx(0).
According to (5) e−Lt is representable in the form (cf. Eq. (12) of Chapter 4 in [13])
s
νk−1
j=0
where λ1, . . . , λs are all distinct eigenvalues of L.
k=1
X
X
e−Lt =
Zkj tje−λkt,
(9)
(10)
Since L is singular, we can set λ1 = 0. Then Z1j are the components of L corresponding to the
characteristic root 0. By item 3 of Proposition 1, ν1 = ind L = 1 and by item 6 of Proposition 1,
Z10, the eigenprojection of L denoted by J, coincides with the matrix ¯J of maximum out-forests
of the communication digraph Γ.
Since the components Zkj of L are independent of t while, by item 2 of Proposition 1,
Re(λk) > 0 (k ≥ 2), we have
lim
t→∞
s
νk−1
X
k=2
X
j=0
Zkj tje−λkt = 0.
Finally, (9) -- (11) and ν1 = 1 yield
lim
t→∞
x(t) = lim
t→∞
e−Ltx(0) = Z10 x(0) = J x(0).
2In some cases, the expression [3, Theorem 1] can be more convenient for calculations.
August 4, 2018
(11)
⊓⊔
DRAFT
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, 2013
5
IV. THE PROPERTIES OF THE ASYMPTOTIC STATE
Now we need some additional notation. A basic bicomponent of a digraph Γ is any maximal
(by inclusion) strongly connected weighted subdigraph of Γ such that there are no arcs coming
into it from outside. By [1, Proposition 6], the number of basic bicomponents in Γ is equal to
the out-forest dimension d of Γ.
Let x(∞) be the limiting state vector of the model (1) -- (2): x(∞) = limt→∞ x(t).
Corollary 1 (of Theorem 1). Let K be a basic bicomponents of Γ; let j be a vertex of K. It
holds that:
1) If i is a vertex of K or i is reachable (by a directed path) from K and not reachable from
the other basic bicomponents of Γ, then xi(∞) = xj(∞) and xi(∞) is equal to the value
of consensus for the communication digraph K alone;
2) If vertex i is reachable from several basic bicomponents of Γ, then xi(∞) is between the
minimum and maximum elements of x(∞) that correspond to these basic bicomponents
(and is strictly between them if the minimum and maximum differ);
3) If vertex i is not in a basic bicomponent of Γ, then x(∞) is independent of xi(0).
Corollary 1 is easily proved using the row stochasticity of ¯J and two simple facts which follow
from [1, Theorem 2′]. The facts are: (1) ¯J ij 6= 0 if and only if j belongs to a basic bicomponent
of Γ and i is reachable from j; (2) If i and j belong to the same basic bicomponents in Γ, then
the i-row and j-row of ¯J are equal, while the i-column and j-column are proportional.
Using time shift and item 7 of Proposition 1 we have3
Corollary 2 (of Theorem 1). Let x(t) be a solution of (3). Then for any t ∈ R, ¯J x(t) = x(∞).
Consequently, for any t1, t2 ∈ R, ¯J(x(t1) − x(t2)) = 0, i.e., (x(t1) − x(t2)) ∈ N ( ¯J) = R(L).
Consider the weighted communication digraph Γ shown in Fig. 1. It has two basic bicompo-
nents whose vertex sets are {1, 2} and {3, 4, 5}.
V. EXAMPLE
1
1
2
2
3
6
3
4
7
2
1
2
4
1
2
2
5
Figure 1. A communication digraph Γ.
3Cf. [6], equation between (16) and (17).
August 4, 2018
DRAFT
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, 2013
6
The Laplacian matrix (4) of the model (1) -- (2) corresponding4 to Γ is:
L =
2 −2
−1
1
0
0
0
1
0
0
0
0 −2
0
0 −3 −2
0 −4
0
0
0
0
0
0 −1
4 −2
0 −2
0
2
0
0
0
0
0
0
5
0
0
0
0
0
0
.
0
0 −1 5
The spectrum of L is real: (0, 0, 2, 3, 5, 5, 5), which is not generally the case for a Laplacian
matrix of a weighted digraph. On the other hand, L is not diagonalizable as the geometric
multiplicity of the triple eigenvalue 5 is 1. The minimal polynomial of L is
To find Z10 = J, one can use item 9 of Proposition 1:
λh(λ) = λ(λ − 2)(λ − 3)(λ − 5)3.
Z10 =
C(0)
h(0)
,
(12)
where h(0) = (−2)(−3)(−5)3 = −750, while C(0) can be determined using (55) and (56) in
[14, Chapter 4]: Ψ(0, µ) = (µ − 2)(µ − 3)(µ − 5)3 and
C(0) = Ψ(0 · I, L) = (L − 2I)(L − 3I)(L − 5I)3.
Substituting C(0) and h(0) into (12) yields
Z10 = J = ¯J =
1
750
250 500
250 500
0
0
0
0
0 0 0
0 0 0
0 300 150 300 0 0
0 300 150 300 0 0
0 300 150 300 0 0
60 120 0 0
60 264 132 264 0 0
150 300 120
30
0
0
0
(13)
(14)
.
The matrix ¯J can also be found using item 10 of Proposition 1: as well as (13), this involves
four matrix multiplications, but it does not require the knowledge of the nonzero eigenvalues
or the minimal polynomial of L. Direct enumeration of forests has no practical value for the
computation of ¯J . However, to give the reader a little taste of this "forestry", in Fig. 2 we present
all 32 maximum out-forests of Γ and their weights. The total weight of them is f = 750; by the
definition (7), it is the common denominator of the entries of ¯J = ( ¯J ij). The numerator of ¯J ij
is the total weight fij of the maximum out-forests in which i belongs to a tree diverging from j.
For example, for ¯J 65, these are the forests #10, 12, 14, 16, 26, 28, 30, and 32 whose weights
are 16, 4, 16, 4, 32, 8, 32, and 8, respectively, so that f65 = 120 and ¯J 65 = 120
750 , in concordance
with (14).
4On this correspondence, see Section II-B.
August 4, 2018
DRAFT
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, 2013
2
4
2
5
1
1
1
1
2
3
6
3
7
1
2
2
6
3
4
7
2
4
2
5
1
1
2
4
2
5
1
2
3
6
3
4
7
Forest #1; weight = 12
Forest #2; weight = 32
Forest #3; weight = 48
1
1
2
3
6
1
3
7
2
5
1
1
4
1
2
2
6
3
4
7
2
5
1
1
4
1
2
3
6
3
4
7
2
5
1
4
1
Forest #5; weight = 6
Forest #6; weight = 16
Forest #7; weight = 24
1
1
2
3
6
3
7
1
5
1
1
4
1
2
2
2
6
3
4
7
5
1
1
4
1
2
2
3
6
3
4
7
5
1
4
1
2
1
1
1
7
2
4
2
5
2
6
2
1
3
7
Forest #4; weight = 8
2
5
4
1
2
6
2
1
3
7
Forest #8; weight = 4
4
1
2
2
6
2
1
3
7
5
5
Forest #9; weight = 6
Forest #10; weight = 16
Forest #11; weight = 24
Forest #12; weight = 4
1
1
2
4
1
2
3
6
3
7
1
5
1
1
2
4
1
2
2
6
3
4
7
5
1
1
2
3
6
2
4
1
3
4
7
5
1
1
2
4
1
2
6
2
1
3
7
Forest #13; weight = 6
Forest #14; weight = 16
Forest #15; weight = 24
Forest #16; weight = 4
1
2
2
3
6
1
3
7
2
4
2
5
1
2
2
4
2
5
1
2
2
6
3
4
7
2
2
3
6
3
4
7
2
4
2
5
1
2
2
4
2
5
2
6
2
1
3
7
Forest #17; weight = 24
Forest #18; weight = 64
Forest #19; weight = 96
Forest #20; weight = 16
1
2
2
3
6
3
7
1
2
4
1
5
1
2
2
2
6
3
4
7
2
4
1
5
1
2
2
3
6
3
4
7
2
4
1
5
1
2
2
5
4
1
2
6
2
1
3
7
Forest #21; weight = 12
Forest #22; weight = 32
Forest #23; weight = 48
Forest #24; weight = 8
1
2
2
3
6
3
7
1
4
1
2
5
1
2
2
2
6
3
4
7
4
1
2
5
1
2
2
3
6
3
4
7
4
1
2
5
1
2
4
1
2
2
6
2
1
3
7
Forest #25; weight = 12
Forest #26; weight = 32
Forest #27; weight = 48
Forest #28; weight = 8
1
2
2
3
6
1
3
7
2
4
1
5
1
2
2
4
1
2
2
6
3
4
7
5
1
2
2
3
6
2
4
1
3
4
7
5
1
2
2
4
1
2
6
2
1
3
7
Forest #29; weight = 12
Forest #30; weight = 32
Forest #31; weight = 48
Forest #32; weight = 8
Figure 2. The maximum out-forests of the communication digraph in the Example.
5
5
August 4, 2018
DRAFT
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, 2013
8
Using [1, Proposition 9] the set of all maximum out-forests of Γ can be described as follows.
1. Choose an arbitrary spanning diverging tree in each basic bicomponent of Γ. 2. Choose any
maximum out-forest in the digraph obtained from Γ by removing all arcs belonging to the
basic bicomponents. Combining the chosen trees and forest gives a maximum out-forest of Γ;
every desirable forest can be obtained in this way. A more detailed algorithm for constructing
maximum out-forests can be found in [1, Section 5].
Let x(0) = (1, 10, 5, 7, 9,∗,∗)T (the last two components are "free": they correspond to
"nonbasic" vertices which, by Corollary 1, do not affect the limiting state vector). By Theorem 1,
x(t) = J x(0) = (7, 7, 7, 7, 7, 7, 7)T ,
lim
t→∞
i.e., asymptotic consensus is achieved. On the other hand, if x(0) = (0, 6, 3, 9, 10,∗,∗)T, then
x(t) = J x(0) = (4, 4, 7, 7, 7, 5.2, 6.64)T,
lim
t→∞
and asymptotic consensus is achieved only within the basic bicomponents having vertex sets
{1, 2} and {3, 4, 5}, but not for the whole set of agents.
Thus, a system satisfying (1) -- (2) has its domain of convergence to consensus, that is, the set
of initial states x(0) such that the product J x(0) gives a vector with all equal components. In
[4], this domain (obviously, it is a subspace of Rn) is characterized and it is shown that when
x(0) does not belong to the domain, then there is still some reasonable "quasi-consensus". It can
by obtained by first, projecting x(0) onto the domain of convergence and second, applying the
coordination protocol (1) -- (2). Say, for the initial states of the form x(0) = (0, 6, 3, 9, 10,∗,∗)T
which were considered above, the value of such a "quasi-consensus" is 5.82.
VI. ON COMMUNICATION DIGRAPHS OF OUT-FOREST DIMENSION 1
Suppose that the communication digraph Γ has a spanning diverging tree or, equivalently, the
out-forest dimension of Γ is one (d = 1). In this (and only this) case, by item 4 of Proposition 1,
rank ¯J = 1 holds, so by item 5,
¯J = 1vT
l ,
(15)
is any row of ¯J and vT
l
where vT
l and 1 span the left and right null
l
spaces of L, respectively. Thus, Theorem 1 yields the following familiar necessary and sufficient
condition of achieving consensus.
1 = 1. By items 7 and 4, vT
Corollary 3 (of Theorem 1). If the communication digraph Γ of the model (1) -- (2) has a spanning
diverging tree, then for any initial state x(0), x(t) converges to the consensus
lim
t→∞
x(t) = (vT
l x(0)) 1,
(16)
where vl is the unique left eigenvector of L associated with 0 and satisfying vT
1 = 1. Conversely,
l
if for each initial state x(0), x(t) tends to a consensus, then Γ has a spanning diverging tree.
For the more restricted case of a strongly connected digraph Γ, a representation similar to (16)
l , where
was obtained in [22, Theorem 3]. In this case, it was shown that limt→∞ e−Lt = vrvT
August 4, 2018
DRAFT
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, 2013
9
vl and vr are, respectively, the left and right eigenvectors of L associated with 0 and satisfying
l vr = 1. Before Theorem 1, the authors of [22] mention that 1 can be substituted for vr.
vT
Corollary 3 coincides with [20, Theorem 3.12] (see also [20, Proposition 3.11] and Lemma 1.3
in [23]). The case of a communication digraph Γ having a spanning diverging tree was recently
considered in [6] where Lemma 3 presents an analog of (16). However, the multiplier 1/√n in
[6, Eq. (18)] is not correct due to an invalid step in the proof.
Finally, observe that Theorem III.8 in [26] can also be derived from Theorem 1.
VII. A DISCRETE COUNTERPART OF THE FOREST CONSENSUS THEOREM
Consider the discretization of the model (3):
x(t + τ ) − x(t)
τ
= −Lx(t)
(17)
with a small fixed τ ∈ R. Let y(k) := x(kτ ), k = 0, 1, . . . , be the state vector with the discrete-
time dynamics determined by (17). Then
y(k) = (I − τ L) y(k − 1),
k = 1, 2, . . . .
Setting
and observing [1, Section 8] that P is row stochastic whenever
P := I − τ L
0 < τ ≤(cid:16)max
we obtain DeGroot's iterative pooling model [12]:
i X
j6=i
aij(cid:17)−1
y(k) = P y(k − 1),
k = 1, 2, . . . .
(18)
(19)
(20)
(21)
Matrix (19) is sometimes called the Perron matrix with parameter τ of the weighted digraph Γ.
Obviously, P is the linear part of the series expansion of e−τ L.
Let us compare the asymptotic properties of the model (3) and its discrete analog (21). From
(21) one has
(22)
A necessary and sufficient condition of the convergence of {P k} under (20) is the aperiodicity
of P. On the other hand, the Ces`aro (time-average) limit
y(k) = P ky(0),
k = 0, 1, . . . .
P ∞ := lim
k→∞
1
k
k
X
i=1
P i
(23)
exists for any stochastic P and coincides with limk→∞ P k whenever the latter exists. Otherwise,
if P is periodic with period s, then P ∞ = s−1(cid:0)P (1) + . . . + P (s)(cid:1) , where P (1), . . . , P (s) are the
limits of the converging subsequences of {P k}: P (i) = lim
The discrete-time counterpart of Theorem 1 is an immediate consequence of well-known
results. Yet, for ease of comparison with Theorem 1, we represent it in the form of a theorem.
P js+i, i = 1, . . . , s.
j→∞
August 4, 2018
DRAFT
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, 2013
10
Theorem 2. Let sequence y(k) satisfy (21), where P is defined by (19) -- (20). Then
lim
k→∞
1
k
k
X
i=1
y(i) = J y(0),
where J is the eigenprojection of L, which coincides with the matrix ¯J of maximum out-forests
of the communication digraph Γ corresponding to L. Moreover, if (20) is satisfied strictly, then
y(k) = J y(0).
lim
k→∞
(24)
Proof. Meyer [21] and Rothblum [24] have shown that P ∞ is the eigenprojection of I − P .
Hence, by (19) and the definition of eigenprojection, P ∞ = J . Now applying the Ces`aro limit
to (22) and using (23) and item 6 of Proposition 1 one obtains the first assertion of Theorem 2.
Alternatively, the identity P ∞ = ¯J coincides with the Markov chain tree theorem first proved
by Wentzell and Freidlin [27] and rediscovered in [19], [18]. This identity provides the first
assertion of Theorem 2 along the same lines.
Finally, if (20) is satisfied strictly, then P has a strictly positive diagonal. In this case, by
Gersgorin's theorem, P has no eigenvalues of modulus 1 except for 1. Hence, P is not periodic
and {P k} converges, which yields (24).
⊓⊔
Obviously, the only essential difference between Theorem 2 and Theorem 1 is the use of the
Ces`aro limit in the case of a periodic matrix P. With a similar "Ces`aro" addendum one can
easily formulate a discrete-time counterpart of the Corollary 3 of Section VI.
To compute the matrix J = ¯J, one can use items 8 -- 10 of Proposition 1, constructive
characterizations (h), (j) or (l) in [3, Section 2], or [11, Proposition 2].
REFERENCES
[1] R. P. Agaev and P. Y. Chebotarev, "The matrix of maximum out forests of a digraph and its applications," Automation and
Remote Control, vol. 61, no. 9, pp. 1424 -- 1450, 2000.
[2] -- -- , "Spanning forests of a digraph and their applications," Automation and Remote Control, vol. 62, no. 3, pp. 443 -- 466,
2001.
[3] -- -- , "On determining the eigenprojection and components of a matrix," Automation and Remote Control, vol. 63, no. 10,
pp. 1537 -- 1545, 2002.
[4] -- -- , "The projection method for reaching consensus and the regularized power limit of a stochastic matrix," Automation
and Remote Control, vol. 72, no. 12, pp. 2458 -- 2476, 2011.
[5] R. Agaev and P. Chebotarev, "On the spectra of nonsymmetric Laplacian matrices," Linear Algebra and its Applications,
vol. 399, pp. 157 -- 168, 2005.
[6] K. Amelin, N. Amelina, O. Granichin, and O. Granichina, "Multi-agent stochastic systems with switched topology
and noise," in 13th ACIS International Conference on Software Engineering, Artificial Intelligence, Networking and
Parallel/Distributed Computing (SNPD). Kyoto, Japan: IEEE, 2012, pp. 438 -- 443.
[7] A. Ben-Israel and T. N. E. Greville, Generalized Inverses: Theory and Applications, 2nd ed. New York: Springer, 2003.
[8] P. Chebotarev, "Comments on "Consensus and cooperation in networked multi-agent systems"," Proceedings of the IEEE,
vol. 98, no. 7, pp. 1353 -- 1354, 2010.
[9] P. Chebotarev and R. Agaev, "Forest matrices around the Laplacian matrix," Linear Algebra and its Applications, vol. 356,
pp. 253 -- 274, 2002.
[10] P. Y. Chebotarev and R. P. Agaev, "Coordination in multiagent systems and Laplacian spectra of digraphs," Automation
and Remote Control, vol. 70, no. 3, pp. 469 -- 483, 2009.
[11] -- -- , "Addendum to the paper "On determining the eigenprojection and components of a matrix"," Automation and Remote
Control, vol. 72, no. 3, p. 626, 2011.
August 4, 2018
DRAFT
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, 2013
11
[12] M. H. DeGroot, "Reaching a consensus," Journal of the American Statistical Association, vol. 69, no. 345, pp. 118 -- 121,
1974.
[13] F. R. Gantmacher, Applications of the Theory of Matrices. New York: Interscience, 1959.
[14] -- -- , The Theory of Matrices. New York: Chelsea, 1959.
[15] R. E. Hartwig, "More on the Souriau-Frame algorithm and the Drazin inverse," SIAM Journal on Applied Mathematics,
vol. 31, pp. 42 -- 46, 1976.
[16] N. J. Higham, Functions of Matrices: Theory and Computation, 2nd ed. Philadelphia: SIAM, 2008.
[17] P. Lancaster and M. Tismenetsky, The Theory of Matrices, 2nd ed. New York: Academic Press, 1985.
[18] T. Leighton and R. L. Rivest, "Estimating a probability using finite memory," in Foundations of Computation Theory.
Berlin-Heidelberg: Springer, 1983, vol. 158/1983, pp. 255 -- 269.
[19] -- -- , "The Markov chain tree theorem," Laboratory of Computer Science, MIT, Cambridge, Mass., Computer Science
Technical Report MIT/LCS/TM-249, 1983.
[20] M. Mesbahi and M. Egerstedt, Graph Theoretic Methods in Multiagent Networks.
Princeton, NJ: Princeton University
Press, 2010.
[21] C. D. Meyer, Jr., "The role of the group generalized inverse in the theory of finite Markov chains," SIAM Review, vol. 17,
no. 3, pp. 443 -- 464, 1975.
[22] R. Olfati-Saber and R. M. Murray, "Consensus problems in networks of agents with switching topology and time-delays,"
IEEE Transactions on Automatic Control, vol. 49, no. 9, pp. 1520 -- 1533, 2004.
[23] W. Ren and Y. Cao, Distributed Coordination of Multi-agent Networks: Emergent Problems, Models, and Issues. London:
Springer, 2011.
[24] U. G. Rothblum, "Computation of the eigenprojection of a nonnegative matrix at its spectral radius," in Stochastic Systems:
Modeling, Identification and Optimization II, ser. Mathematical Programming Study, R. J.-B. Wets, Ed. Amsterdam:
North-Holland, 1976, vol. 6, pp. 188 -- 201.
[25] -- -- , "A representation of the Drazin inverse and characterizations of the index," SIAM Journal on Applied Mathematics,
vol. 31, no. 4, pp. 646 -- 648, 1976.
[26] C. N. Taylor, R. W. Beard, and J. Humpherys, "Dynamic input consensus using integrators," in Proceedings of the 2011
American Control Conference (ACC 2011), San Francisco, CA, 2011, pp. 3357 -- 3362.
[27] A. D. Wentzell and M. I. Freidlin, "On small random perturbations of dynamical systems," Russian Mathematical Surveys,
vol. 25, no. 1, pp. 1 -- 55, 1970.
August 4, 2018
DRAFT
|
0911.0231 | 3 | 0911 | 2011-01-12T02:58:06 | Synchronized Task Decomposition for Cooperative Multi-agent Systems | [
"cs.MA",
"cs.DC",
"eess.SY"
] | It is an amazing fact that remarkably complex behaviors could emerge from a large collection of very rudimentary dynamical agents through very simple local interactions. However, it still remains elusive on how to design these local interactions among agents so as to achieve certain desired collective behaviors. This paper aims to tackle this challenge and proposes a divide-and-conquer approach to guarantee specified global behaviors through local coordination and control design for multi-agent systems. The basic idea is to decompose the requested global specification into subtasks for each individual agent. It should be noted that the decomposition is not arbitrary. The global specification should be decomposed in such a way that the fulfilment of these subtasks by each individual agent will imply the satisfaction of the global specification as a team. First, it is shown by a counterexample that not all specifications can be decomposed in this sense. Then, a natural follow-up question is what the necessary and sufficient condition should be for the proposed decomposability of a global specification. The main part of the paper is set to answer this question. The case of two cooperative agents is investigated first, and a necessary and sufficient condition is presented and proven. Later on, the result is generalized to the case of arbitrary finite number of agents, and a hierarchical algorithm is proposed, which is shown to be a sufficient condition. Finally, a cooperative control scenario for a team of three robots is developed to illustrate the task decomposition procedure. | cs.MA | cs |
Technical Report: NUS-ACT-10-001-Ver.3:
Synchronized Task Decomposition for
Cooperative Multi-agent Systems
Preprint, Accepted by Automatica (entitled as
Guaranteed Global Performance Through Local
Coordinations)
Mohammad Karimadini, and Hai Lin,
Abstract
It is an amazing fact that remarkably complex behaviors could emerge from a large collection of very
rudimentary dynamical agents through very simple local interactions. However, it still remains elusive on
how to design these local interactions among agents so as to achieve certain desired collective behaviors.
This paper aims to tackle this challenge and proposes a divide-and-conquer approach to guarantee
specified global behaviors through local coordination and control design for multi-agent systems. The
basic idea is to decompose the requested global specification into subtasks for each individual agent. It
should be noted that the decomposition is not arbitrary. The global specification should be decomposed
in such a way that the fulfillment of these subtasks by each individual agent will imply the satisfaction
of the global specification as a team. Formally, a given global specification can be represented as an
automaton A, while a multi-agent system can be captured as a parallel distributed system. The first
question needs to be answered is whether it is always possible to decompose a given task automaton
A into a finite number of sub-automata Ai, where the parallel composition of these sub-automata Ai
is bisimilar to the automaton A. First, it is shown by a counterexample that not all specifications can
M. Karimadini and H. Lin are both from the Department of Electrical and Computer Engineering, National University of
Singapore, Singapore. Corresponding author, H. Lin [email protected]
be decomposed in this sense. Then, a natural follow-up question is what the necessary and sufficient
conditions should be for the proposed decomposability of a global specification. The main part of
the paper is set to answer this question. The case of two cooperative agents is investigated first, and
necessary and sufficient conditions are presented and proven. Later on, the result is generalized to the
case of arbitrary finite number of agents, and a hierarchical algorithm is proposed, which is shown to
be a sufficient condition. Finally, a cooperative control scenario for a team of three robots is developed
to illustrate the task decomposition procedure.
I. INTRODUCTION
Strongly driven by its great potential in both civilian and industrial applications [1], [2],
multi-agent systems has rapidly emerged as a hot research area at the intersection of control,
communication and computation [3], [4], [5], [6]. The key issue in multi-agent system design is
how to explicitly set the local interaction rules such that certain desirable global behaviors can be
achieved by the team of cooperative agents[7]. Although it is known that sophisticated collective
behaviors could emerge from a large collection of very elementary agents through simple local
interactions, we still lack knowledge on how to change these rules to achieve or avoid certain
global behaviors. As a result, it still remains elusive on how to design these local interaction rules
so as to make sure that they, as a group, can achieve the specified requirements. In addition,
we would like to point out that the desired collective behavior for a group of mobile agents
could be very complicated, and might involve the coordination of many distributed independent
and dependent modules in a parallel and environmental awareness manner. Hence, it poses new
challenges that go beyond the traditional path planning, output regulation, or formation control[8],
[9], [10], [5].
This paper aims to propose a decomposition approach applicable in divide-and-conquer design
for cooperative multi-agent systems so as to guarantee the desired global behaviors. The core
idea is to decompose a global specification into sub-specifications for individual agents, and
then design local controllers for each agent to satisfy these local specifications, respectively. The
decomposition should be done in such a way that the global behavior is achieved provided that
all these sub-specifications are held true by individual agents. Hence, the global specification
is guaranteed by design. In order to perform this idea, several questions are required to be
answered, such as how to describe the global specification and subtasks in a succinct and formal
way; how to decompose the global specification; whether it is always possible to decompose,
and if not, what are the necessary and sufficient conditions for decomposability.
To formally describe the specification, a deterministic finite automaton is chosen here to
represent the global specifications for multi-agent systems, due to its expressibility for a large
class of tasks[11], [12], its similarity to our human logical commands, and its connection to
the temporal logic specifications[13], [14]. Accordingly, we will focus on the logical behavior
of a multi-agent system and model its collective behavior through parallel composition [15]. It
is assumed that each agent is equipped with a local event set, containing both private events
and common events (shared with other agents). Furthermore, it is assumed that the global task
automaton is defined over the union of all agents' events. Then, the decomposition problem can
be stated as follows. Given the global desired behavior represented as a deterministic automaton,
how to obtain the local task automata from it such that the composition of the obtained local
task automata is equivalent to the original global task automaton?
Since the global task automaton is defined over the union of all agents' event sets, a reasonable
way to obtain the local task automata is through natural projections with respect to each agent's
event set. Namely, the agent will ignore the transitions marked by the events that are not in its
own event set, i.e., blinds to these moves. The obtained automaton will be a sub-automaton of
the global task automaton by deleting all the moves triggered by blind events of the agent. Given
a task automaton and sets of local events, it is always feasible to do the projection operation,
but the question is whether the obtained sub-automata preserve the required specifications in the
sense that the fulfillment of each agent with its corresponding sub-task automaton will imply the
satisfaction of the global specification as a group. Unfortunately, by a simple counterexample,
it can be shown that the answer is not always. Then, a natural follow-up question is what
the necessary and sufficient condition should be for the proposed decomposability of a global
specification. The main part of the paper is set to answer this question.
Similar automaton decomposition problem has been studied in the computer science literature.
Roughly speaking, two different classes of problems have been studied, so far. The first problem
is to design the event distribution so as to make the automaton decomposable, which is typically
studied in the context of concurrent systems. For example, [16] characterized the conditions for
decomposition of asynchronous automata in the sense of isomorphism based on the maximal
cliques of the dependency graph. The isomorphism equivalence used in [16] is however a strong
condition, in the sense that two isomorphic automata are bisimilar but not vise versa [11]. In
many applications bisimulation relation suffices to capture the equivalence relationship. On the
other hand, the second class of problems assumes that the distribution of the global event set is
given and the objective is to find conditions on the automaton such that it is decomposable. This is
usually called synthesis modulo problem [15] that can be investigated under three types of equiv-
alence: isomorphism, bisimulation and language equivalence. Bisimulation synthesis modulo for
a global automaton was addressed in [17], by introducing necessary and sufficient conditions
for automaton decomposition based on language product of the automaton and determinism of
its bisimulation quotient. Obtaining the bisimulation quotient, however, is generally a difficult
task, and the condition on language product relies on language separability[18], which is another
form of decomposability. These problems motivate us to develop new necessary and sufficient
conditions that can characterize the decomposability based on the investigation of events and
strings in the given automaton.
In this paper, we identify conditions on the global specification automaton in terms of its
private and common events for the proposed decomposability, which are shown to be necessary
and sufficient for the case of two agents. Later on, the result is generalized to the case of
arbitrary finite number of agents, and a hierarchical algorithm is proposed, which is shown to
be a sufficient condition. Furthermore, it is shown that if the global task is decomposable, then
designing the local controller for each agent to satisfy its corresponding sub-task will lead the
entire multi-agent system to achieve the global specification. To illustrate the decomposition
approach, a coordination and control scenario has been developed and implemented on a team
of three robots.
The rest of the paper is organized as follows. Preliminary lemmas, notations, definitions and
problem formulation are represented in Section II. Section III introduces the necessary and
sufficient conditions for decomposition of an automaton with respect to parallel composition and
two local event sets. The algorithm for the hierarchical extension of the automaton decomposition
is given in Section IV. To illustrate the task decomposition, an implementation result is given
on a cooperative multi-robot system example in Section V. Finally, the paper concludes with
remarks and discussions in Section VI. The proofs of lemmas are given in the Appendix.
II. PROBLEM FORMULATION
We first recall the definition of an automaton [12].
Definition 1: (Automaton) An automaton is a tuple A = (Q, q0, E, δ) consisting of
• a set of states Q;
• the initial state q0 ∈ Q;
• a set of events E that causes transitions between the states, and
• a transition relation δ ⊆ Q × E × Q such that (q, e, q′) ∈ δ if and only if δ(q, e) = q′ (or
q e→ q′ ).
In general, automaton has also an argument Qm which is the set of final or marked states, the
states are marked when it is desired to attach a special meaning to them, such as accomplishing.
This argument is dropped when the notion of marked state is not of interest or it is clear from
the content.
As δ is a partial map from Q × E into Q, in general, not all states are reachable from the initial
state. The accessible portion of the automaton is defined as
Definition 2: (Accessible Operator; [11]) Consider an automaton A = (Q, q0, E, δ). The oper-
ator Ac(.) excludes the states and their attached transitions that are not reachable from the initial
state, and is defined as Ac(A) = (Qac, q0, E, δac) with Qac = {q ∈ Q∃s ∈ E∗, δ(q0, s) = q} and
δac = δQac × E → Qac, restricting δ to the smaller domain of Qac. As AC(.) has no effect on
the behavior of the automaton, from now on we assume A = Ac(A).
The transition relation can be extended to a finite string of events, s ∈ E∗, where E∗ stands for
Kleene − Closure of E (the set of all finite strings over elements of E), as follows δ(q, ε) = q,
and δ(q, se) = δ(δ(q, s), e) for s ∈ E∗ and e ∈ E. We focus on deterministic task automata
that are simpler to be characterized, and cover a wide class of specifications. The qualitative
behavior of a deterministic discrete event system (DES) is described by the set of all possible
sequences of events starting from initial state. Each such a sequence is called a string, and the
collection of strings represents the language generated by the automaton, denoted by L(A). Given
a language L, L ⊆ E∗ is the prefix-closure of L defined as L = {s ∈ E∗∃t ∈ E∗, st ∈ L},
consisting of all prefixes of all the strings in L. The existence of a transition over string s ∈ E∗
from a state q ∈ Q, is denoted by δ(q, s)!, and considering a language L, by δ(q, L)! we mean
∀ω ∈ L : δ(q, ω)!.
To describe the decomposability condition in the main result and during the proofs, we define
successive event pair and adjacent event pair as follows.
Definition 3: (Successive event pair) Two events e1 and e2 are called successive events if
∃q ∈ Q : δ(q, e1)! ∧ δ(δ(q, e1), e2)! or δ(q, e2)! ∧ δ(δ(q, e2), e1)!.
Definition 4: (Adjacent event pair) Two events e1 and e2 are called adjacent events if ∃q ∈
Q : δ(q, e1)! ∧ δ(q, e2)!.
To compare the task automaton and its decomposed automata, we use the simulation and
bisimulation relations [11].
Definition 5: (Simulation and Bisimulation) Consider two automata Ai = (Qi, q0
1, 2. A relation R ⊆ Q1 × Q2 is said to be a simulation relation from A1 to A2 if
i , E, δi), i =
1, q0
2) ∈ R
1) (q0
2) ∀ (q1, q2) ∈ R, δ1(q1, e) = q′
2, (q′
1, q′
1, then ∃q′
2 ∈ Q2 such that δ2(q2, e) = q′
2) ∈ R.
If R is defined for all states and all events in A1, then A1 is said to be similar to A2 (or A2
simulates A1), denoted by A1 ≺ A2 [11].
If A1 ≺ A2, A2 ≺ A1 and R is symmetric then A1 and A2 are said to be bisimilar (bisimulate
each other), denoted by A1
∼= A2 [19].
In general, bisimilarity implies languages equivalence but the converse does not necessarily
hold true [20].
Definition 6: (Isomorphism, [21]) Isomorphism is one of the strongest equivalences relations
between automata. Two automata Ai = (Qi, q0
i , E, δi), i = 1, 2, are said to be isomorphic, if
there exists an isomorphism θ from A1 to A2 defined as a bijective function θ : Q1 → Q2 such
that θ(q0
2, and θ(δ1(q, e)) = δ2(θ(q), e), ∀q ∈ Q1, e ∈ E.
1) = q0
By this definition, two isomorphic automata are bisimilar, but bisimilar automata are not
necessarily isomorphic (see Example 3).
In this paper, we assume that the task automaton AS and the sets of local events Ei are all
given. It is further assumed that AS is deterministic while its event set E is obtained by the
union of local event sets, i.e., E = ∪iEi. The problem is to check whether the task automaton
AS can be decomposed into sub-automata ASi on the local event sets Ei, respectively, such that
the collection of these sub-automata ASi is somehow equivalent to AS. The equivalence is in the
sense of bisimilarity as defined above, while the composition process for these sub-automata ASi
could be in the usual sense of parallel composition as defined below. Parallel composition is used
to model the interactions between automata and represent the logical behavior of multi-agent
systems. Parallel composition is formally defined as
Definition 7: (Parallel Composition [12]) Let Ai = (Qi, q0
i , Ei, δi), i = 1, 2, be automata.
The parallel composition (synchronous composition) of A1 and A2 is the automaton A1A2 =
(Q = Q1 × Q2, q0 = (q0
1, q0
2), E = E1 ∪ E2, δ), with δ defined as
(δ1(q1, e), δ2(q2, e)) ,
(δ1(q1, e), q2) ,
(q1, δ2(q2, e)) ,
undefined,
if
δ1(q1, e)!, δ2(q2, e)!,
e ∈ E1 ∩ E2
if δ1(q1, e)!, e ∈ E1\E2;
if δ2(q2, e)!, e ∈ E2\E1;
otherwise
;
.
∀(q1, q2) ∈ Q, e ∈ E : δ((q1, q2), e) =
The parallel composition of Ai, i = 1, 2, ..., n is called parallel distributed system, and is
defined based on the associativity property of parallel composition [11] as
An = A1 k (A2 k (· · · k (An−1 k An))).
n
k
i=1
Ai = A1 k ... k
A reasonable guess for task automaton decomposition is to use natural projections with respect
i , takes
to agents' event set. Natural projection over strings is denoted by pEi = pi : E∗ → E∗
a string from the event set E and eliminates events in it that do not belong to the event set
Ei ⊆ E. The natural projection is formally defined on the strings as
Definition 8: (Natural Projection on String, [11]) Consider a global event set E and its local
event sets Ei, i = 1, 2, ..., n, with E =
inductively defined as
n
∪
i=1
Ei. Then, the natural projection pi : E∗ → E∗
i
is
pi(ε) = ε;
∀s ∈ E∗, e ∈ E : pi(se) =
pi(s)e if e ∈ Ei;
otherwise.
pi(s)
The natural projection is also defined on automata as Pi(AS) : AS → ASi, where, ASi are
obtained from AS by replacing its events that belong to E\Ei by τ-moves (representing silent
or unobservable transitions), and then, merging the τ-related states. The τ-related states form
equivalent classes defined as follows.
Definition 9: (Equivalent class of states, [16]) Consider an automaton AS = (Q, q0, E, δ) and
Ei. Then, the relation ∼ Ei (or ∼i) is the least
local event sets Ei, i = 1, 2, ..., n, with E =
equivalence relation on the set Q of states such that δ(q, e) = q′ ∧ e /∈ Ei ⇒ q ∼Ei q′, and
[q]Ei = [q]i denotes the equivalence class of q defined on ∼ Ei. In this case, q and q′ are said
n
∪
i=1
to be τ-related.
The natural projection is then formally defined on an automaton as follows.
Definition 10: (Natural Projection on Automaton) Consider an automaton AS = (Q, q0, E, δ)
and local event sets Ei, i = 1, 2, ..., n, with E =
with δi([q]Ei, e) = [q′]Ei
1.
δ(q1, e) = q′
n
∪
i=1
Ei. Then, Pi(AS) = (Qi = Q/∼Ei, [q0]Ei, Ei, δi),
if there are states q1 and q′
1 such that q1 ∼Ei q, q′
1 ∼Ei q′, and
The following example elaborates the concept of natural projection on a given automaton.
Example 1: Consider an automaton AS:
/ •
a
•
with the event set E = E1 ∪ E2
e2
e4
•
/ •
e1
b
/ •
and local event sets E1 = {a, b, e1}, E2 = {a, b, e2, e4}. The natural projections of AS into E1
is obtained as P1(AS): •
by replacing {e2, e4} ∈ E\E1 with τ and merging the τ-
•
e1
b
•
related states. Similarly, the projection P2(AS) is obtained as P2(AS):
a
e2
/ •
•
a
e4
/ •
b
/ • .
To investigate the interactions of transitions in two automata, particularly in P1(AS) and
P2(AS), the interleaving of strings is defined, based on the path automaton as follows.
Definition 11: (Path Automaton) A sequence q1
e2→ ... en→ qn is called a path automaton,
characterized by its initial state q1 and string s = e1e2...en, denoted by P A(q1, s), and defined
as P A(q1, s) = ({q1, ..., qn}, {q1}, {e1, ..., en},δP A), with δP A(qi, ei) = qi+1, i = 1, ..., n − 1.
P A(q′
1, s′) is defined, similarly.
e1→ q2
Definition 12: (Interleaving) Given two sequences q1
m, the interleaving of their corresponding strings, s = e1e2...en and s′ = e′
q′
by ss′, and defined as ss′ = L{P A(q1, s)P A(q′
1, s′)}.
e1→ q2
e2→ ... en→ qn and q′
1e′
2...e′
1
e′
2→ ...
e′
e′
1→ q′
m→
2
m, is denoted
Example 2: Consider three strings s1 = e1a, s2 = ae2 and s3 = ae1. Then the interleaving
of s1 and s2 is s1s2 = e1ae2 while the interleaving of two strings s2 and s3 becomes s2s3 =
{ae1e2, ae2e1}.
Based on these definitions, we are now ready to formally define the decomposability of an
automaton with respect to parallel composition and natural projections as follows.
Definition 13: (Automaton decomposability, or bisimulation synthesis modulo [15]) A task
automaton AS with the event set E and local event sets Ei, i = 1, ..., n, E =
Ei, is said to be
decomposable with respect to parallel composition and natural projections Pi : AS → Pi (AS),
n
∪
i=1
/
/
/
/
/
5
5
o
o
o
o
/
/
/
4
4
/
/
i = 1, · · · , n, if
Pi (AS) ∼= AS.
n
k
i=1
Now, let us see a motivating example to illustrate the decomposition procedure.
Example 3: Consider the task automaton AS and its local event sets in Example 1. This au-
∼=
tomaton is decomposable with respect to parallel composition and natural projections, since AS
P1(AS)P2(AS), leading to L(AS) = L(P1(AS)P2(AS)) = {ε, a, ae1, ae1b, e2, e2e4, e2e4b}.
Two automata
/ ?>=<
89:;q0
e1
6nnnnnn
(PPPPPP
e2
?>=<
89:;q1
?>=<
89:;q4
e2
(PPPPPP
6nnnnnn
e1
and
?>=<
89:;q2
a
/ ?>=<
89:;q3
/GFED
@ABCq′
0
e1
8qqqqqqq
&MMMMMMM
e2
GFED
@ABCq′
1
GFED
@ABCq′
4
e2
/GFED
@ABCq′
2
/GFED
@ABCq′
5
e1
a
a
/GFED
@ABCq′
3
/GFED
@ABCq′
6
Note that, these two automata are bisimilar (with the bisimulation relation R = {(q0, q′
with E = E1 ∪ E2, E1 = {a, e1}, E2 = {a, e2} are other examples of decomposable automata.
0), (q1, q′
1),
6)}), but not isomorphic. The first one is concrete (decom-
(q2, q′
posable in the sense of isomorphism; satisfying F D (forward diamond) and ID (independent
3), (q4, q′
4), (q2, q′
2), (q3, q′
5), (q3, q′
diamond)) as well as decomposable (in the sense of bisimulation). The latter automaton, on the
other hand, is decomposable, but not concrete.
Examples 3 shows a decomposable task automaton; however, to deal with the top - down
design we need to understand whether any task automaton is decomposable or not.
Problem 1: Given a deterministic task automaton AS and local event sets Ei, i = 1, · · · , n, is
it always possible to decompose AS with respect to parallel composition and natural projections
Pi, i = 1, · · · , n?
To answer this question, we examine the following example.
Example 4: Consider an automaton AS:
/ •
e1
/ •
e2
/ • , with local event sets E1 = {e1} and E2 = {e2}. The parallel composition
of P1(AS) :
e1
/ •
/ • and P2(AS) :
e2
/ •
/ • is P1(AS)P2(AS):
/ •
e1
(PPPPPP
e2
•
•
.
/ •
e2
6nnnnnn
e1
One can observe that AS ≺ P1(AS)P2(AS) but P1(AS)P2(AS) ⊀ AS leading to L(AS) =
{ε, e1, e1e2} ⊆ L(P1(AS)P2(AS)) = {ε, e1, e1e2, e2, e2e1}, but L(P1(AS)P2(AS)) * L(AS).
Therefore, AS is not decomposable with respect to parallel composition and natural projections
Pi, i = 1, 2.
(
/
6
(
/
6
/
/
/
8
&
/
/
/
/
/
/
/
/
/
/
/
/
(
/
6
Therefore, not all automata are decomposable with respect to parallel composition and natural
projections. Then, a natural follow-up question is what makes an automaton decomposable. It
can be formally stated as follows.
Problem 2: Given a deterministic task automaton AS and local event sets Ei, i = 1, · · · , n,
what are the necessary and sufficient conditions that AS is decomposable with respect to parallel
Pi (AS) ∼=
composition and natural projections Pi : AS → Pi (AS), i = 1, · · · , n, such that
AS?
n
k
i=1
This problem will be addressed in the following two sections. In the next section, we will deal
with the case of two agents and present necessary and sufficient conditions for decomposability.
Then, the result is generalized in Section IV to a finite number of agents, as a sufficient condition.
III. TASK DECOMPOSITION FOR TWO AGENTS
In order for AS
∼= P1(AS)P2(AS), from the definition of bisimulation, it is required to
have AS ≺ P1(AS)P2(AS); P1(AS)P2(AS) ≺ AS, and the simulation relations are symmetric.
These requirements are provided by the following three lemmas. Firstly, the following simulation
relationship always holds true.
Lemma 1: Consider any deterministic automaton AS with event set E = E1 ∪ E2, local event
sets Ei, and natural projections Pi, i = 1, 2. Then AS ≺ P1(AS)P2(AS).
Proof: See the Appendix for proof.
This lemma shows that, in general, P1(AS)P2(AS) simulates AS. The similarity of P1(AS)P2(AS)
to AS, however, is not always true (see Example 4), and needs some conditions as stated in the
following lemma.
Lemma 2: Consider a deterministic automaton AS = (Q, q0, E = E1 ∪ E2, δ) and natural
projections Pi : AS → Pi(AS), i = 1, 2. Then, P1(AS)P2(AS) ≺ AS if and only if AS satisfies
the following conditions: ∀e1 ∈ E1\E2, e2 ∈ E2\E1, q ∈ Q, s ∈ E∗:
• DC1 : [δ(q, e1)! ∧ δ(q, e2)!] ⇒ [δ(q, e1e2)! ∧ δ(q, e2e1)!];
• DC2 : δ(q, e1e2s)! ⇔ δ(q, e2e1s)!, and
• DC3 : ∀s, s′ ∈ E∗, sharing the same first appearing common event a ∈ E1 ∩ E2, s 6= s′,
q ∈ Q: δ(q, s)! ∧ δ(q, s′)! ⇒ δ(q, p1(s)p2(s′))! ∧ δ(q, p1(s′)p2(s))!.
Proof: See the Appendix for proof.
Next, we need to show that the two simulation relations R1 (for AS ≺ P1(AS)P2(AS)) and
R2 (for P1(AS)P2(AS) ≺ AS), defined by the above two lemmas, are symmetric.
Lemma 3: Consider an automaton AS = (Q, q0, E = E1 ∪ E2, δ) with natural projections
Pi : AS → Pi(AS), i = 1, 2. If AS is deterministic, AS ≺ P1(AS)P2(AS) with the simulation
relation R1 and P1(AS)P2(AS) ≺ AS with the simulation relation R2, then R−1
1 = R2 (i.e.,
∀q ∈ Q, z ∈ Z: (z, q) ∈ R2 ⇔ (q, z) ∈ R1) if and only if DC4: ∀i ∈ {1, 2}, x, x1, x2 ∈ Qi,
x1 6= x2, e ∈ Ei, t ∈ E∗
i , δi(x, e) = x1, δi(x, e) = x2: δi(x1, t)! ⇔ δi(x2, t)!.
Proof: See the proof in the Appendix.
Based on these lemmas, the main result on task automaton decomposition is given as follows.
Theorem 1: A deterministic automaton AS = (Q, q0, E = E1 ∪ E2, δ) is decomposable with
respect to parallel composition and natural projections Pi : AS → Pi(AS), i = 1, 2, such that
∼= P1(AS)P2(AS) if and only if AS satisfies the following decomposability conditions (DC):
AS
∀e1 ∈ E1\E2, e2 ∈ E2\E1, q ∈ Q, s ∈ E∗,
• DC1: [δ(q, e1)! ∧ δ(q, e2)!] ⇒ [δ(q, e1e2)! ∧ δ(q, e2e1)!];
• DC2: δ(q, e1e2s)! ⇔ δ(q, e2e1s)!, and
• DC3: ∀s, s′ ∈ E∗, sharing the same first appearing common event a ∈ E1 ∩ E2, s 6= s′,
q ∈ Q: δ(q, s)! ∧ δ(q, s′)! ⇒ δ(q, p1(s)p2(s′))! ∧ δ(q, p1(s′)p2(s))!;
• DC4: ∀i ∈ {1, 2}, x, x1, x2 ∈ Qi, x1 6= x2, e ∈ Ei, t ∈ E∗
i , δi(x, e) = x1, δi(x, e) = x2:
δi(x1, t)! ⇔ δi(x2, t)!.
Proof: According to Definition 5, AS
∼= P1(AS)P2(AS) if and only if AS ≺ P1(AS)P2(AS)
(that is always true due to Lemma 1), P1(AS)P2(AS) ≺ AS (that it is true if and only if DC1,
DC2 and DC3 are true, according to Lemma 2) and the simulation relations are symmetric, i.e.,
R−1
1 = R2(that for a deterministic automaton AS, when AS ≺ P1(AS)P2(AS) with simulation
relation R1 and P1(AS)P2(AS) ≺ AS with simulation relation R2, due to Lemma 3, R−1
1 = R2
∼= P1(AS)P2(AS) if and only if DC1,
holds true if and only if DC4 is satisfied). Therefore, AS
DC2, DC3 and DC4 are satisfied.
Remark 1: Intuitively, the decomposability condition DC1 means that for any successive or
adjacent pair of private events (e1, e2) ∈ {(E1\E2, E2\E1), (E2\E1, E1\E2)} (from different
private event sets), both orders e1e2 and e2e1 should be legal from the same state, unless they
are mediated by a common string.
Furthermore, e1e2 and e2e1 are not required to meet at the same state (unlike F D and ID in
[16]); but due to DC2, any string s ∈ E∗ after them should be the same, or in other words, if e1
and e2 are necessary conditions for occurrence of a string s, then any order of these two events
would be legal for such occurrence (see Example 3). Note that, as a special case, s could be ε.
The condition DC3 means that if two strings s and s′ share the same first appearing common
event, then any interleaving of these two strings should be legal in AS. This requirement is due
to synchronization of projections of these strings in P1(AS) and P2(AS).
The last condition, DC4, ensures the symmetry of mutual simulation relations between AS
and P1(AS)P2(AS). Given the determinism of AS, this symmetry is guaranteed when each local
task automaton bisimulates a deterministic automaton, leading to the existence of a deterministic
automaton that is bisimilar to P1(AS)P2(AS). If the simulation relations are not symmetric,
then some of the sequences that are allowed in AS will be disabled in P1(AS)P2(AS).
The notion of language decomposability [22] is comparable with DC2 and means that any
order of any successive events in any string of the language specification should be legal, or at
least one of its projections (from the viewpoint of the corresponding local observer) should
be capable of distinguishing this order. It also embodies a notion similar to DC3, stating
that the global languages specification should contain all possible interleaving languages of
all local languages. This notion, however, is not capable of capturing the other two conditions
on the decision on the switch between adjacent transitions (DC1), and existence of deterministic
bisimilar automata to P1(AS) and P2(AS) (DC4). The automaton decomposability conditions in
this result, in terms of bisimulation, besides checking the capability of local plants on decision
making on the orders of event (DC2), they should also be capable of decision making on the
switches (DC1), and moreover, the synchronization of local task automata should not lead to an
illegal behavior(DC3), and also ensures that like AS, P1(AS)P2(AS) also has a deterministic
behavior (DC4).
The decomposability conditions can be then paraphrased as follows: Any decision on order
or switch between two transitions that cannot be made locally (by at least one local controller)
should not be critical globally (any result of the decision should be allowed); and interpretation
of the global task by the team of local plants should neither allow an illegal behavior (a string
that is not in global task automaton), nor disallow a legal behavior (a string that appears in the
global task automaton).
The following four examples illustrate the decomposability conditions for decomposable and
undecomposable automata.
Example 5: This example illustrates the concept of decision making on switching between the
events, mentioned in Remark 1. Furthermore, it shows an automaton that satisfies DC2, DC3 and
DC4, but not DC1, leading to undecomposability. The automaton AS:
with
e1
/ •
(PPPPPP
e2
•
•
local event sets E1 = {e1} and E2 = {e2}, is not decomposable as the parallel composition of
e2
e1
e2
e1
P1(AS) :
/ •
/ • and P2(AS) :
/ •
/ • is P1(AS)P2(AS) :
6mmmmmm
e1
which does not bisimulate AS. Here, AS is not decomposable with respect to parallel composition
and natural projections Pi, i = 1, 2, since two events e1 ∈ E1\E2 and e2 ∈ E2\E1 do not respect
DC1, as none of the local plant takes in charge of decision making on the switching between
these two events. One can observe that, if in this example e1 ∈ E1\E2 and e2 ∈ E2\E1 were
with local
separated by a common event a ∈ E1 ∩ E2, then
(PPPPPP
•
e2
e2
a
/ •
/ •
•
/ •.
/ •
(PPPPPP
e1
•
•
event sets E1 = {e1, a} and E2 = {e2, a}, was decomposable, since the decision on the switch
between e1 and a could be made in E1 and then E2 could be responsible for the decision on
the order of a and e2.
Example 6: The automaton AS in Example 3 shows an automaton that respects DC1, DC3
and DC4, but is undecomposable due to violation of DC2. Here, AS is not decomposable
since none of the local plants take in charge of decision making on the order of two events
e1 ∈ E1\E2 and e2 ∈ E2\E1. If e1 ∈ E1\E2 and e2 ∈ E2\E1 were separated by a com-
/ • with local event sets
mon event a ∈ E1 ∩ E2, then the automaton
E1 = {e1, a} and E2 = {e2, a}, was decomposable, since the decision on the orders of
e1 and a and then a and e2 could be made in E1 and then E2, subsequently. As another
with E1 = {a, e1}, E2 =
example, consider an automaton AS:
/ •
/ •
/ •
e1
e2
e2
a
a
•
/ •
/ •
/ •
e1
6nnnnnn
(PPPPPP
e2
•
/ •
e1
/
/
/
(
/
/
/
/
/
/
/
(
/
6
/
/
/
(
/
/
/
/
/
/
/
/
6
(
/
{a, e2}, leading to P1(AS)P2(AS):
. The transition δ(z0, e2e1a)!
e2
e2
•
•
•
e1
e1
e1
e2
e2
•
`•
•
e1
e1
e1
/ •
•
e2
e2
/ •
a
/ •
in P1(AS)P2(AS), but ¬δ(q0, e2e1a)! in AS. Therefore, AS is not decomposable. If the lower
branch was continued with a transition on a after e2e1, then the automaton was decomposable
(See Example 3).
Example 7: This example illustrates an automaton that satisfies DC1, DC2 and DC4, but it
is undecomposable as it does not fulfil DC3, since new strings appear in P1(AS)P2(AS) from
the interleaving of two strings in P1(AS) and P2(AS), but they are not legal in AS. Consider the
with E1 = {a, e1}, E2 = {a, e2}, leading
task automaton AS:
/ •
a
•
e1
6nnnnnn
(PPPPPP
a
to P1(AS) ∼=
/ •
e1
6nnnnnn
(PPPPPP
e2
•
•
e2
(PPPPPP
6nnnnnn
e1
•
a
/ •
/ •
e2
a
/ •
, P2(AS) ∼=
•
•
/ •
e2
6nnnnnn
(PPPPPP
a
•
•
a
/ •
and
/ •
e2
P1(AS)P2(AS): •
•
a
a
e2
e1
e2
•
•
a
a
e1
•
•
e2
/ •
•
that is not bisimilar to AS since two strings e2a
/ •
e2
/ •
and e1ae2 are newly generated, while they do not appear in AS, although both P1(AS) and
P2(AS) are deterministic.
Example 8: This example illustrates an automaton that satisfies DC1 and DC2, and DC3,
but is undecomposable as it does not fulfil DC4. Consider the task automaton
AS:
with E1 = {a, b, e1}, E2 = {a, b}, leading to
a
b
/ ?>=<
89:;q2
/ ?>=<
89:;q3
/ ?>=<
89:;q0
e1
6nnnnnn
(PPPPPP
a
?>=<
89:;q1
?>=<
89:;q4
P1(AS):
/ ?>=<
89:;x0
e1
6mmmmmm
(QQQQQQ
a
?>=<
89:;x1
a
/ ?>=<
89:;x2
b
/ ?>=<
89:;x3
, P2(AS):
?>=<
89:;x4
/ ?>=<
89:;y0
a
6nnnnnn
(PPPPPP
a
?>=<
89:;y1
b
, and
/ ?>=<
89:;y2
?>=<
89:;y3
o
o
/
O
O
o
o
/
/
O
O
O
O
o
o
/
/
(
/
6
(
/
6
/
/
/
6
(
/
/
6
(
/
o
o
o
o
/
/
/
o
o
o
o
/
/
/
/
/
6
(
/
/
/
6
(
/
/
6
(
P1(AS)P2(AS):
/ ?>=<
89:;z0
e1
6nnnnnn
(QQQQQQ
a
?>=<
89:;z4
?>=<
89:;z1
a
(QQQQQQ
a
which is not bisimilar to AS. This task
b
/ ?>=<
89:;z3
?>=<
89:;z2
?>=<
89:;z5
automaton AS satisfies DC1 and DC2 as contains no successive/adjacent transitions defined on
different local event sets. But, it does not fulfil DC4, although any string in T = {p1(s)p2(s′),
p1(s′)p2(s)} (s and s′ are the top and bottom strings in AS and share the first appearing common
event a ∈ E1∩E2), appears in AS. The reason is that there exists a transition on string e1a from z0
to z5 that stops in P1(AS)P2(AS), whereas, although e1a transits from q0 in AS, it does not stop
afterwards. This illustrate dissymmetry in simulation relations between AS and P1(AS)P2(AS).
Note that AS ≺ P1(AS)P2(AS) with the simulation relation R1 over all events in E, from all
states in Q into some states in Z, as R1 = {(q0, z0), (q1, z1), (q2, z2), (q3, z3), (q4, z4)}. Moreover,
P1(AS)P2(AS) ≺ AS with the simulation relation R2 over all events in E, from all states in
Z into some states in Q, as R2 = {(z0, q0), (z1, q1), (z2, q2), (z3, q3), (z4, q4), (z5, q2)}. Therefore,
although AS ≺ P1(AS)P2(AS) and P1(AS)P2(AS) ≺ AS, P1(AS)P2(AS) ≇ AS, since
∃(z5, q2) ∈ R2, whereas (q2, z5) /∈ R1. If for stoping of string e1a in P1(AS)P2(AS), there was
a state in Q reachable from q0 by e1a and stopping there, then we would have ∀q ∈ Q, z ∈ Z :
(q, z) ∈ R1 ⇔ (z, q) ∈ R2 and P1(AS)P2(AS) ∼= AS.
It should be noted that the condition DC4 not only applies for nondeterminism on common
events, but also it requires any nondeterminism on private event also to have a bisimilar determin-
istic counterpart. For example, consider the task automaton AS:
E1 = {e1, a}, E2 = {e2, a}. The parallel composition of P1(AS):
/ •
e1
(PPPPPP
e2
/ •
e1
(PPPPPP
e1
•
•
•
•
a
6nnnnnn
with
•
•
e2
e1
a
/ •
(with
/ •
nondeterministci transition on private event e1) and P2(AS):
is P1(AS)P2(AS):
/ •
a
(PPPPPP
e2
•
•
which is not bisimilar to AS.
•
e2
C
8888888
a
e1
•
e1
•
/ •
•
e2
/ •
e1
C
8888888
e2
e1
•
/
/
(
/
/
6
(
/
/
/
(
/
/
6
/
/
/
/
(
/
/
/
/
(
/
/
/
/
C
C
/
/
o
o
The automaton AS:
/ •
e1
6nnnnnn
(PPPPPP
a
•
•
a
/ •
with E = E1 ∪ E2, E1 = {a, e1}, E2 =
/ •
e2
{a, e2}, P1(AS):
/ •
e1
6nnnnnn
(PPPPPP
a
•
•
a
and P2(AS):
/ •
/ •
ple of an undecomposable automaton that violates both DC3 and DC4. It violates DC3 since
δ(z0, e1ae2)! in P1(AS)P2(AS), but ¬δ(q0, e1ae2)! in AS, and it does not satisfy DC4 since
P2(AS) is nondeterministic and is not bisimilar to a deterministic automaton, leading to a string
in P1(AS)P2(AS) that e2 is disallowed after a while there in no such restriction in AS. If AS
was AS:
/ • , and
, then P2(AS) ∼=
/ •
/ •
e2
a
/ •
/ •
•
a
e2
•
e2
/ •
is an exam-
a
6nnnnnn
(PPPPPP
a
•
/ •
e1
6nnnnnn
(PPPPPP
a
•
/ •
e2
AS was decomposable.
Remark 2: Example 8 also shows that the determinism of AS does not reduce the bisim-
ulation synthesis problem to language equivalence synthesis problem. Note that here, AS and
P1(AS)P2(AS) are language equivalent, but not bisimilar. The reason is that although AS is
deterministic, and AS ≺ P1(AS)P2(AS), P1(AS)P2(AS) ≺ AS, the simulation relations are
not symmetric due to existence of nondeterministic strings in P1(AS)P2(AS) that can not be
replaced by a deterministic one. The nondeterminism in P1(AS)P2(AS) is inherited from a
,
nondeterminism in P2(AS). If AS was in the form of AS:
a
b
/ •
/ •
•
/ •
e1
6nnnnnn
(PPPPPP
a
•
b
/ •
then P1(AS) ∼= AS, P2(AS) ∼=
/ •
a
/ •
b
/ • and P1(AS)P2(AS) ∼= AS.
IV. HIERARCHICAL DECOMPOSITION
The previous section showed the decomposition of an automaton with respect to the parallel
composition and two local event sets. However, in practice, multi-agent systems are typically
comprised of many individual agents that work as a team. The proposed procedure of decomposi-
tion can be generalized for more than two agents. However, the problem becomes rapidly complex
as the number of agents increases. It is then advantageous to have a hierarchical decomposition
/
/
6
(
/
/
/
6
(
/
/
6
(
/
/
/
6
(
/
/
/
/
/
/
/
6
(
/
/
/
/
method to have only two individual event sets at a time for decomposition. Consider a task
automaton AS to be decomposed with respect to parallel composition and individual event sets
Ei, i = 1, 2, ..., n, so that E =
Ei. We propose the following algorithm as a sufficient condition
for hierarchical decomposition of the given task automaton.
n
∪
i=1
Algorithm 1: (Hierarchical Decomposition Algorithm)
n
∪
i=1
Ei, Σ = {E1, ..., En}, K = {1, ..., n}.
1) E =
2) i = 1, find k ∈ K such that Σi = Ek ∈ Σ, ¯Σi = ∪
decomposability conditions DC1-DC4 in Theorem 1, i.e., AS
j∈K\k
Ej, so that AS satisfies
∼= PΣi(AS)P ¯Σi(AS).
3) K = K\k, Σ = {Ej}j∈K, AS = P ¯Σi(AS), i = i + 1, go to Step 2.
4) Continue until i = n − 1 or no more decomposition is possible in i = m − 1, m ≤ n. Then
Σm = ¯Σm−1, and hence, AS is decomposable with respect to parallel composition and
natural projections into {Σ1, · · · , Σm} ⊆ Σ, if the algorithm proceeds up to i = m − 1.
Remark 3: The algorithm will terminate due to finite number of states and local event sets.
If the algorithm successfully proceeds to step n − 1, the automaton AS is decomposable and we
obtain a complete decomposition of the global specifications into subtasks for each individual
agent. However, it is unclear whether the algorithm can successfully terminate for any decom-
posable task automaton (necessity). The computational complexity of the algorithm in the worst
case is of order O(n2(E2 × Q × κ + Σ
a∈E1∩E2
pa(L(AS))2)), where κ = max
t∈L(AS )
t, assuming
replaced by
the number of appearing events as the length of loops. In practice, during the iterations, E is
Ej which is decreasing with respect to iteration. Moreover, the second term in
the complexity expression shows that the less number of common events and the less appearance
K
∪
j=1
of common events in AS, the less complexity.
Once the subtasks are obtained, the next step is the design of controllers for each agent
to achieve these subtasks respectively. The following result shows that the fulfillment of the
decomposed subtasks will imply the global specifications for the multi-agent systems. Before
stating the theorem, following two lemmas are presented to be used for the proof.
Lemma 4:
PΣ1(AS) k PΣ2(AS) k · · · k PΣm−1(AS) k PΣm(AS) ∼=
PΣ1(AS) k (cid:0)PΣ2(AS) k (cid:0)· · · k (cid:0)PΣm−1(AS) k PΣm(AS)(cid:1)(cid:1)(cid:1)
Proof: Algorithm 1 is a direct extension of Theorem 1 combined with Lemma 4. Then,
∼= Pi(AS), i = 1, 2, · · · , n, due to Lemma 5.2
choosing local controllers ACi, so that ACi k APi
leads to
(ACi k APi) ∼=
Pi(AS) ∼= AS.
n
k
i=1
n
k
i=1
Proof: See the proof in the Appendix.
Lemma 5: If two automata A2 and A4 (bi)simulate, respectively, A1 and A3, then A2 k A4
(bi)simulates A1 k A3, i.e.,
1) (A1 ≺ A2) ∧ (A3 ≺ A4) ⇒ (A1 k A3 ≺ A2 k A4);
∼= A2 k A4);
2) (A1
∼= A4) ⇒ (A1 k A3
∼= A2) ∧ (A3
Proof: See the proof in the Appendix.
Theorem 2: Consider a plant, represented by a deterministic parallel distributed system A∆ =
n
k
i=1
APi, with given local event sets Ei, i = 1, ..., n, and given specification represented by a
Ei. If the Algorithm
deterministic decomposable automaton AS
∼= Pi(AS),
1 continues up to i = n − 1, then designing local controllers ACi, so that ACi k APi
i = 1, 2, · · · , n, derives the global closed loop system to satisfy the global specification AS, in
the sense of bisimilarity, i.e.,
Pi(AS), with E =
(ACi k APi) ∼= AS.
n
k
i=1
n
∪
i=1
∼=
n
k
i=1
Now, if DC1-DC4 is reduced to DC1-DC3 (Theorem 1 is reduced into Lemma 2), then
Pi(AS) ∼= AS is reduced into
Pi(AS) ≺ AS, and hence, choosing local controllers ACi,
n
k
i=1
n
k
i=1
so that ACi k APi ≺ Pi(AS), i = 1, 2, · · · , n, due to Lemma 5.1 leads to
Pi(AS) ≺ AS. Therefore,
(ACi k APi) ≺
n
k
i=1
n
k
i=1
Corollary 1: Considering the plant and global task as stated in Theorem 2, if DC1-DC4 is
reduced to DC1-DC3 in Algorithm 1 and it continues up to i = n − 1, then designing local
controllers ACi, so that ACi k APi ≺ Pi(AS), i = 1, 2, · · · , n, derives the global closed loop
(ACi k APi) ≺ AS.
system to satisfy the global specification AS, in the sense of similarity, i.e.,
Remark 4: It should be noted that in this approach, the parallel composition requires a fixed
n
k
i=1
communication pattern among the "local" automata to synchronize on their common events. This
framework is therefore suitable for static distributed systems. For moving agent systems, the
agents are required to provide large enough communication range to ensure that the connectivity
is preserved during the movement, to ensure the correct synchronization on the common events.
V. EXAMPLE
Consider a cooperative multi-robot system (MRS) configured in Figure 1. The MRS consists of
three robots R1, R2 and R3. All robots have the same communication and positioning capabilities.
Furthermore, the robot R2 has the rescue and fire-fighting capabilities, while R1 and R3 are
normal robots with the pushing capability. Initially, all of them are positioned in Room 1. Rooms
2 and 3 are accessible from Room 1 by one-way door D2 and two-way doors D1 and D3, as
shown in Figure 1. All doors are equipped with spring to be closed automatically, when there
is no force to keep them open.
Fig. 1. The environment of MRS coordination example.
Assume that Room 2 requests help for fire extinguishing. After the help announcement, the
Robot R2 is required to go to Room 2, urgently from D2 and accomplish its task there and come
back immediately to Room 1. However, D2 is a one-way door, and, D1 is a heavy door and needs
cooperation of two robots R1 and R3 to be opened. To save time, as soon as the robots hear the
help request from Room 2, R2 and R3 go to Rooms 2 and 3, from D2 and D3, respectively, and
then R1 and R3 position on D1, synchronously open D1 and wait for accomplishment of task
of R2 in Room 2 and returning to Room 1 (R2 is fast enough). Afterwards, R1 and R3 move
backward to close D1 and then R3 returns back to Room 1 from D3. All robots then stay at
Room 1 for the next task.
These requirements can be translated into a task automaton for the robot team as it is illustrated
in Figure 2, defined over local event sets E1 = {h1, R1toD1, R1onD1, F W D, D1opened, R2in1,
BW D, D1closed, r}, E2 = {h2, R2to2, R2in2, D1opened, R2to1, R2in1, r}, and E3 = {h3, R3to3,
R3in3, R3toD1, R3onD1, F W D, D1opened, R2in1, BW D, D1closed, R3to1, R3in1, r}, with hi:=
Ri received help request, i = 1, 2, 3; RjtoD1:=command for Rj to position on D1, j = 1, 3;
RjonD1:= Rj has positioned on D1, j = 1, 3; F W D:= command for moving forward (to open
D1); BW D:= command for moving backward (to close D1) ; D1opened:= D1 has been opened;
D1closed:= D1 has been closed; r:= command to go to initial state for the next implementation;
Ritok:= command for Ri to go to Room k, and Riink:= Ri has gone to Room k, i = 1, 2, 3,
k = 1, 2, 3.
AS:
h1
R1toD1
R1onD1
R3to3
h3
•
•
•
/ •
#FFFFFFF
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
cGGGGGGGGGGGGGGGGG
#GGGGGG
R2in1
•
•
•
•
•
•
•
•
•
•
•
R3in3
#GGGGGG
#GGGGGG
#GGGGGG
•
•
/ •
R3toD1
#GGGGGG
#GGGGGG
•
/ •
R3onD1
#GGGGGG
/ •
/ •
/ •
/ •
F W D
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
•
•
•
•
h2
R2to2
R2in2
D1opened
/ •
/ •
/ •
R2to1
/ •
/ •
/ •
•
R3in1
R3to1
•
•
•
r
/ •
h2
•
•
R2to2
#FFFFFFF
#GGGGGG
#GGGGGG
cGGGGGGGGGGGGGGGGG
#GGGGGG
R2in2
•
D1Closed
•
•
•
•
•
•
•
•
•
/ •
#FFFFFFF
#FFFFFFF
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
cGGGGGGGGGGGGGGGGG
#GGGGGG
BW D
•
•
•
•
•
•
•
•
•
•
Fig. 2. Task automaton AS for robot team.
The states show the labels of transitions after occurrence, and hence are not labeled, here.
Meaning that δ(x, e) = y is interpreted as follows: Once e occurs in x, it transits into y and this
state y is labeled by e(occurred). For example, when "help request" event happens, the state of
each robot transits from the initial state to the state "the robot received help request".
/
/
/
#
/
/
#
/
/
#
#
O
O
#
#
/
/
/
#
/
/
#
#
#
#
/
/
#
/
/
#
/
/
#
#
O
O
#
#
#
#
/
/
/
#
/
/
#
#
#
#
#
/
/
#
/
/
#
/
/
#
#
c
#
#
#
/
/
/
#
#
#
/
/
/
#
c
#
/
/
/
#
/
/
/
#
c
o
o
o
o
We check decomposability condition for this global task automaton with respect to Σ1 = E2
and ¯Σ1 = E1 ∪E3 for the first stage in Algorithm 1. Firstly, {h2, R2to2, R2in2} = Σ1\ ¯Σ1 can oc-
cur in any order with respect to {h1, R1toD1, R1onD1, h3, R3to3, R3in3, R3toD1, R3onD1, F W D}
= ¯Σ1\Σ1, as it is shown in the global automaton in Figure 2, satisfying DC1 and DC2. Moreover,
D1opened, R2in1 and r are common events, provided by R1, R2 and R3, respectively, and
informed to the other two robots upon occurrence.
Since {D1opened, F W D} ⊆ ¯Σ1, {R2in2, D1opened} ⊆ Σ1, {D1opened, R2to1} ⊆ Σ1,
{R2to1, R2in1} ⊆ Σ1, {R2in1, BW D} ⊆ ¯Σ1, {BW D, D1closed} ⊆ ¯Σ1, {D1closed, R3to1} ⊆
¯Σ1, {R3to1, R3in1} ⊆ ¯Σ1, {R3in1, r} ⊆ ¯Σ1, {r, h1} ⊆ ¯Σ1, {r, h2} ⊆ Σ1, {r, h3} ⊆ ¯Σ1, and
hence all these successive transitions satisfy DC1 and DC2. Furthermore, all common events
{D1opened, R2in1, r} ⊆ Σ1∩ ¯Σ1 appear in only one branch, and hence, DC3 is satisfied. Finally,
PΣ1(AS) and P ¯Σ1(AS) are both deterministic, and hence, DC4 is satisfied.
Therefore, due to Theorem 1, AS can be decomposed into P2(AS) = PΣ1(AS) = A2 and P1,3 =
P ¯Σ1(AS). The second stage of hierarchical decomposition, decomposes P1,3(AS) into A1 =
P1(AS) and A3 = P3(AS). The private transitions defined over E1\E3 = {h1, R1toD1, R1onD1}
can occur in any order with respect to the transitions defined over the private local event set
E3\E1 = {h3, R3to3, R3in3, R3toD1, R3onD1}. Since {R3onD1, F W D} ⊆ E3, {R1onD1, F W D}
⊆ E1, {F W D, D1opened} ⊆ E1 ∩ E3, {D1opened, R2in1} ⊆ E1 ∩ E3, {R2in1, BW D} ⊆
E1 ∩ E3, {BW D, D1closed} ⊆ E1 ∩ E3, {D1closed, R3to1} ⊆ E3, {R3to1, R3in1} ⊆ E3,
{R3in1, r} ⊆ E3, {r, h3} ⊆ E3 and {r, h1} ⊆ E1, then DC1 and DC2 are satisfied. Furthermore,
since all common events {F W D, D1opened, R2in1, BW D, D1closed, r} ⊆ E1 ∩ E3 appear in
only one branch in P1,3(AS), therefore, there are no pairs of strings violating DC3, and hence,
DC3 is also satisfied. Moreover, P1(AS) and P2(AS) are both deterministic, and consequently,
P1,3(AS) satisfies DC4. The results of two decomposition stages are shown in Figures 3 and 4,
such that P1(AS) k P2(AS) k P3(AS) ∼= AS.
It can be seen that the design of supervisor to satisfy these individual task automata is easier
than the design of a global supervisor to satisfy the global specification. Furthermore, since the
specification is determined for each agent, the global task can be achieved in a decentralized
fashion.
Discussions on the design of supervisory control (for local local plants and local task automata)
and also refining of the low level continuous controllers can be found in [11], [12], [23], [24]
h1
R1toD1
R1onD1
P1,3(AS):
/ •
r
•
#GGGGGG
cGGGGGG
R3in1
•
•
•
#GGGGGG
#GGGGGG
cGGGGGG
•
/ •
•
•
R3to1
D1closed
#GGGGGG
#GGGGGG
#GGGGGG
cGGGGGG
•
•
•
•
•
h3
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
cGGGGGG
/ •
•
•
•
•
BW D
R3to3
#GGGGGG
#GGGGGG
#GGGGGG
#GGGGGG
cGGGGGG
•
•
•
•
•
R2in1
R3in3
#GGGGGG
#GGGGGG
#GGGGGG
•
•
/ •
R3toD1
#GGGGGG
#GGGGGG
•
/ •
R3onD1
#GGGGGG
/ •
F W D
#GGGGGG
•
D1opened
P2(AS):
h2
/ •
/ •
R2to2
R2in2
/ •
/ •
D1opened
R2to1
/ •
r
/ •
@A
R2in1
/ •
BCD
Fig. 3. P2(AS) for R2 and P{E1∪E3}(AS) for the team {R1, R3}.
P1(AS):
h1
/ •
R1onD1
/ •
/ •
/ •
R1toD1
F W D
D1opened
/ •
/ •
/ •
R2in1
BW D
P2(AS):
P3(AS):
r
•
/ •
@A
/ •
@A
D1closed
h2
/ •
/ •
R2to2
R2in2
/ •
/ •
D1opened
R2to1
/ •
r
•
R2in1
/ •
BCD
h3
/ •
/ •
R3to3
R3in3
R3onD1
/ •
/ •
/ •
R3toD1
D1opened
/ •
/ •
F W D
/ •
R2in1
r
BW D
/ •
/ •
D1closed
R3to1
/ •
R3in1
/ •
BCD
Fig. 4. P1(AS) for R1; P2(AS) for R2 and P3(AS) for R3.
and [9].
This scenario has been successfully implemented on a team of three ground robots, shown in
Figure 1.
/
/
/
#
/
/
#
/
/
#
#
O
O
#
/
/
/
#
/
/
#
#
c
/
/
#
/
/
#
/
/
#
#
c
#
/
/
/
#
/
/
#
#
c
/
/
#
/
/
#
/
/
#
#
c
/
/
/
#
c
o
o
/
/
/
/
/
/
/
O
O
/
/
/
/
/
/
/
O
O
o
o
/
/
/
/
/
/
/
O
O
/
/
/
/
/
/
/
/
/
/
/
/
/
O
O
VI. CONCLUSION
The paper proposed a formal method for automaton decomposition that facilitates the top-
down distributed cooperative control of multi-agent systems. Given a set of agents whose logical
behaviors can be modeled as a parallel distributed system, and a global task automaton, the paper
has the following contributions: firstly, we have provided necessary and sufficient conditions for
decomposability of an automaton with respect to parallel composition and natural projections
into two local event sets; secondly, the approach has been extended into a sufficient condition
for an arbitrary finite number of agents, using a hierarchical algorithm, and finally, we have
shown that if a global task automaton is decomposed into the local tasks for the individual
agents, designing the local supervisors for each agent, satisfying the local tasks, guarantees that
the closed loop system of the team of agents satisfies the global specification.
The implementation of this approach is decentralized, in the sense that there is no central
unit to coordinate the agents; however, they need communication to synchronize on common
events. This approach differs from the classical decentralized supervisory control that refers to
the configuration of a monolithic plant, controlled by several supervisors that are distributed
into different nodes [11]. The proposed approach is more suitable for those applications with
distributed configurations for both plant and supervisor, such as multi-robot coordination systems
and sensor/actuator networks. In addition, due to associativity property of parallel composition,
the proposed approach can be modular such that when a new task automaton is introduced to the
systems, one can decompose the new global task automaton, and then compose the new local
task automata with the corresponding old local task automata.
A. Proof for Lemma 1
VII. APPENDIX
We prove AS ≺ P1(AS)P2(AS)) by showing that R = {(q, z) ∈ Q × Z∃s ∈ E∗, δ(q0, s) =
q, z = ([q]1, [q]2)} is a simulation relation, defined on all events in E and all reachable states in
AS. Consider AS = (Q, q0, E = E1∪E2, δ), Pi(AS) = (Qi, q0
i , Ei, δi), i = 1, 2, P1(AS)P2(AS) =
(Z, z0, E, δ). Then, ∀q, q′ ∈ Q, e ∈ E, δ(q, e) = q′, according to definition of natural projection
(Definition 10) [q′]i =
δi([q]i, e)
[q]i
if e ∈ Ei;
if e /∈ Ei
, i = 1, 2, and due to definition of parallel compo-
([q]1, δ2([q]2, e)) ,
(δ1([q]1, e), [q]2) ,
(δ1([q]1, e), δ2([q]2, e)) ,
sition (Definition 7) δ(([q]1, [q]2), e) = ([q′]1, [q′]2) =
if e ∈ E1 ∩ E2;
if e ∈ E1\E2;
if e ∈ E2\E1.
This is true for any q ∈ Q, particularly for q0. This reasoning can be repeated for any reachable
state in Q. Therefore, starting from q0 and taking (q0, Z0 = ([q0]1, [q0]2)) ∈ R, from the above
construction, it follows that for any reachable state in Q (∃s ∈ E∗, δ(q0, s) = q) and q′ ∈ Q,
e ∈ E, δ(q, e) = q′, there exists z = ([q]1, [q]2), z′ = ([q′]1, [q′]2) such that δ(z, e) = z′, and we
can take (q, z) ∈ R and (q′, z′) ∈ R. Therefore, R = {(q, z) ∈ Q × Z∃s ∈ E∗, δ(q0, s)!, z =
([q]1, [q]2)} is a simulation relation, defined over all e ∈ E, and all reachable states in AS, and
hence, AS ≺ P1(AS)P2(AS).
B. Proof for Lemma 2
We use two following lemmas during the proof.
Lemma 6: Consider a deterministic automaton AS = (Q, q0, E = E1 ∪ E2, δ). Then DC1 ∧
DC2 ⇒ [∀s ∈ E∗, δ(q0, s)! ⇒ δ(q0, p1(s)p2(s))! in AS.
This lemma means that for any transition defined on a string in AS, all path automata defined
on the interleaving of p1(s) and p2(s) in P1(AS)P2(AS) are simulated by AS, provided DC1
and DC2.
0...αk
mk, αk
Proof: Consider a deterministic automaton AS = (Q, q0, E = E1 ∪ E2, δ), a string s ∈ E∗,
δ(q0, s) = q, and its projections p1(s), p2(s) with δ1(x0, p1(s)) = x, δ2(y0, p2(s)) = y and
(x, y) ∈ δ((x0, y0), p1(s)p2(s)), in P1(AS), P2(AS) and P1(AS)P2(AS), respectively. Any
string s can be written as s = ω1γ1... ωKγK, with ωk ∈ [E\(E1 ∩ E2)]∗, γk ∈ (E1 ∩ E2)∗,
0 = ε, p1(γk) = p2(γk) =
p1(ωk) = αk = αk
0 = ε. The case mk = 0, nk = 0, rk = 0, K = 0, results in p1(ωk) = ε,
γk = γk
p2(ωk) = ε, γk = ε and s = ε. Based on this setting and definition of parallel composition, for
k = 0, ..., K, i = 0, ..., mk, j = 0, ..., nk and r = 0, ..., rk, the interleaving p1(s)p2(s) is evolved
in P1(AS)P2(AS) as follows: ∀(xk+i, yk+j) ∈ Q1×Q2: δ((xk+i, yk+j), αk
i ), yk+j),
i )=(δ1(xk+i, αk
j )),
j )=(xk+i, δ2(yk+j, βk
δ((δ1(xk+i, αk
δ((xk+i, δ2(yk+j, βk
j )), δ((xk+i, yk+j), βk
j )), δ((xk+mk+r, yk+nk+r), γk
j )=((δ1(xk+i, αk
i )=(δ1(xk+i, αk
i ), δ2(yk+j, βk
i ), δ2(yk+j, βk
0 = ε, p2(ωk) = βk = βk
i ), yk+j), βk
j )), αk
0 ...βk
nk, βk
0 ...γk
rk, γk
r )) =
(δ1(xk+mk+r, γk
r ), δ2(yk+nk+r, γk
r )) with δ1(xk+i, αk
i ) =
xk+i
xk+i+1
if αk
if αk
i = ε
i 6= ε
, δ2(yk+j, βk
j ) =
yk+j
ifγk
if γk
.
i , βk
if βk
if βk
r )) =
yk+j+1
j = ε
j 6= ε
r = ε
r 6= ε
(xk+mk+r, yk+nk+r)
r ), δ2(yk+nk+r, γk
and (δ1(xk+mk+r, γk
(xk+mk+r+1, yk+nk+r+1)
Moreover, DC1 and DC2 collectively imply that ∀e1 ∈ E1\E2, e2 ∈ E2\E1, q ∈ Q, [δ(q, e1)!∧
δ(q, e2)!] ∨ δ(q, e1e2)! ∨ δ(q, e2e1)! ⇒ δ(q, e1e2)! ∧ δ(q, e2e1)! which particularly means that
∀k ∈ {0, ...K}, ∀αk
j , qi,j ∈ Q, i = 0, ..., mk, j = 0, ..., nk, r = 0, ..., rk: δ(qk+i,k+j, αk
i )! with a simulation relation R(ωk) = {((xk+i, yk+j), qk+i,k+j), (δ1(xk+i, αk
j )), ((δ1(xk+i, αk
j αk
j αk
i )), ((xk+i, δ2(yk+j, βk
i ), δ2(yk+j, βk
δ(qk+i,k+j, βk
δ(qk+i,k+j, αk
i ))} from transitions defined on p1(ωk)p2(ωk)
αk
i βk
into AS. For the transitions on the common events, the evolutions are δ(qk+mk+r,k+nk+r, γk
r )! in
AS for r = 0, ..., rk, leading to simulation relation R(γk) = {((xk+mk+r, yk+nk+r),qk+mk+r,k+nk+r),
r ))} from transitions on p1(γk)p2(γk)
((δ1(xk+mk+r, γk
into AS. Therefore, for i = 0, ..., mk, j = 0, ..., nk, r = 0, ..., rk, R =
R(ωk) ∪ R(γk) defines
a simulation relation from P A(z0, p1(s)p2(s)) (in P1(AS)P2(AS)) into AS.
j )), δ(qk+i,k+j, βk
j )), δ(qk+i,k+j, βk
r )), δ(qk+mk+r,k+nk+r, γk
j )), ((δ1(xk+i, αk
r ),δ2(yk+nk+r, γk
i βk
i ), yk+j),
i ), δ2(yk+j, βk
j )), δ(qk+i,k+j,
K
∪
k=0
j )! ∧
Lemma 7: If DC1 and DC2 hold true, then ∀s, s′ ∈ E∗, δ(q0, s)!, δ(q0, s′)!, s 6= s′, pE1∩E2(s),
pE1∩E2(s′) do not start with the same a ∈ E1 ∩ E2, then δ(q0, p1(s)p2(s′))! ∧ δ(q0, p1(s′)p2(s))!
in AS.
Proof: The antecedent of Lemma 7 addresses three following cases: (Case 1): s = ω1,
1; (Case 2): s = ω1aω2, s′ = ω′
1, and (Case 3): s = ω1aω2, s′ = ω′
1bω′
2, where, ω1, ω′
1 ∈
s′ = ω′
[E\(E1 ∩ E2)]∗, ω2, ω′
2 ∈ (E1 ∪ E2)∗, a, b ∈ E1 ∩ E2.
For Case 1, setting K = 1, rk = 0, taking p1(ω1) = α = α0...αm ∈ (E1\E2)∗, p2(ω′
1) =
m′ ∈ (E1\E2)∗ and p2(ω1) = β = β0...βn ∈
0...β′
n′ ∈ (E2\E1)∗, p1(ω′
β′ = β′
(E2\E1)∗, similar to the first part of Lemma 6, it follows that δ(q0, p1(ω1)p2(ω′
δ(q0, p1(ω′
1) = α′ = α′
1)p2(ω1))! in AS.
0...α′
1))! and
For Case 2, from Case 1 and Lemma 6 it follows that δ(q0, p1(s)p2(s′))! = δ(q0, [p1(ω1)p2(ω′
1)]
ap1(ω2))! and δ(q0, p1(s′)p2(s))! = δ(q0, [p1(ω′
1)p2(ω1)]ap2(ω2))! in AS.
For the third case, from definition of parallel composition combined with the first two cases and
also Lemma 6, δ(q0, p1(s)p2(s′))! leads to δ(q0, [p1(ω1)p2(ω′
p2(ω′
a[p1(ω2)p2(ω2)])! and δ(q0, [p1(ω′
1)p2(ω1)]b[p1(ω′
1)]b[p1(ω′
2)p2(ω′
2)p2(ω′
2))])! in AS.
2)])! in AS and similarly, δ(q0, p1(s′)p2(s))! results in δ(q0, [p1(ω′
1)p2(ω1)]
1)]a[p1(ω2)p2(ω2)])! and δ(q0, [p1(ω1)
Therefore, for all three cases, δ(q0, p1(s)p2(s′))! in AS and δ(q0, p1(s′)p2(s))! in AS.
Now, Lemma 2 is proven as follows.
p1(s)p2(s′)
p1(s)p2(s′)
−→
p1(s)
−→ x in P1(AS) and a transition y0
Sufficiency: The set of transitions in P1(AS)P2(AS), is defined as T = {(x0, y0)
−→ (x, y) in P1(AS)P2(AS) is the interleaving
(x, y) ∈ Q1 × Q2}, where, (x0, y0)
p2(s′)
−→ y in P2(AS) (projections of
of transitions x0
s′
transitions q0
−→ q′, respectively, in AS). T can be divided into three sets
of transitions corresponding to a division {Γ1, Γ2, Γ3} on the set of interleaving strings Γ =
s′
−→ q′, q, q′ ∈ Q, s, s′ ∈ E∗}, where, Γ1 = {p1(s)p2(s′) ∈ Γs = s′},
{p1(s)p2(s′)q0
Γ2 = {p1(s)p2(s′) ∈ Γs 6= s′, pE1∩E2(s) and pE1∩E2(s′) do not start with the same event}, and
Γ3 = {p1(s)p2(s′) ∈ Γs 6= s′, pE1∩E2(s) and pE1∩E2(s′) start with the same event }.
s−→ q and q0
s−→ q, q0
Now, for any s, s′, δ(q0, s)!, δ(q0, s′)!, in AS, both δ(q0, p1(s)p2(s′))! and δ(q0, p1(s′)p2(s))!
are guaranteed, for Γ1, due to Lemma 6; for Γ2, due to Lemma 7, and for Γ3, due to combination
of DC3 and Lemma 6 (For simplification, in DC3, s and s′ can be started from any q, instead
of q0, and the strings between q0 and q are checked by Lemma 6).
Necessity: The necessity is proven by contradiction. Suppose that AS simulates P1(AS)P2(AS),
but ∃e1 ∈ E1\E2, e2 ∈ E2\E1, q ∈ Q, s ∈ E∗ s.t. (1): [δ(q, e1)! ∧ δ(q, e2)!] ∧ ¬[δ(q, e1e2)! ∧
δ(q, e2e1)!]; (2): ¬[δ(q, e1e2s)! ⇔ δ(q, e2e1s)!], or (3): ∃s, s′ ∈ E∗, sharing the same first
appearing common event a ∈ E1 ∩ E2, s 6= s′, q ∈ Q: δ(q, s)! ∧ δ(q, s′)! ∧ ¬[δ(q, p1(s)p2(s′))! ∧
δ(q, p1(s′)p2(s))!].
In the first case, due to definition of parallel composition, the expression [δ(q, e1)! ∧ δ(q, e2)!]
leads to δ(z, e1e2)! = δ(z, e2e1)!, where z ∈ Q1 × Q2 in P1(AS)P2(AS) corresponds to q ∈ Q
in AS. Therefore, δ(z, e1e2)! ∧ δ(z, e2e1)!, but, ¬[δ(q, e1e2)! ∧ δ(q, e2e1)!]. This means that
P1(AS)P2(AS) ⊀ AS which is a contradiction. The second case means ∃e1 ∈ E1\E2, e2 ∈
E2\E1, q ∈ Q, s ∈ E∗ s.t. [δ(q, e1e2s)! ∨ δ(q, e2e1s)!] ∧ ¬[δ(q, e1e2s)! ∧ δ(q, e2e1s)!]. From defini-
tion of parallel composition, then δ(q, e1e2s)!∨δ(q, e2e1s)! implies that δ(z, e1e2)! = δ(z, e2e1)!,
for some z ∈ Q1 × Q2 corresponding to q ∈ Q. Consequently, from definition of transition
relation and AS ≺ P1(AS)P2(AS) it turns to δ(z, e1e2s)! = δ(z, e2e1s)!, meaning that
δ(z, e1e2s)! ∧ δ(z, e2e1s)!, but, ¬[δ(q, e1e2s)! ∧ δ(q, e2e1s)!]. This in turn contradicts with
the similarity assumption of P1(AS)P2(AS) ≺ AS. The third case also leads to the con-
tradiction as causes the violation of the simulation relation from P1(AS)P2(AS) into AS as
δ(q, s)! ∧ δ(q, s′)! leads to δ(z, p1(s)p2(s′))! ∧ δ(z, p2(s)p1(s′))! in P1(AS)P2(AS), whereas
¬[δ(q, p1(s)p2(s′))! ∧ δ(q, p2(s)p1(s′))!].
C. Proof for Lemma 3
To prove Lemma 3, we use following two lemmas together with Lemma 5. Firstly, the
following lemma is introduced to characterize the symmetric property of simulation relations.
Lemma 8: Consider two automata A1 and A2, and let A1 be deterministic, A1 ≺ A2 with the
1 = R2 if and
simulation relation R1 and A2 ≺ A1 with the simulation relation R2. Then, R−1
only if there exists a deterministic automaton A′
1 such that A′
1
∼= A2.
Proof:
Sufficiency: A1 ≺ A2, A2 ≺ A1 and A′
1
that due to determinism of A1 and A′
of bisimulation, A1
1 lead to A1
∼= A2, and consequently, R−1
1 = R2.
∼= A′
1. Finally, since A′
1
∼= A2, collectively, result in A1 ≺ A′
1 and A′
1 ≺ A1,
∼= A2, from transitivity
Necessity: The necessity is proven by contradiction as follows. Consider two automata A1 =
(X, x0, E, δ1), A2 = (Y, y0, E, δ2), let A1 be deterministic, A1 ≺ A2 with the simulation relation
∼= A2),
R1, A2 ≺ A1 with the simulation relation R2 and suppose that R−1
∼= A2. This means that
but there does not exist a deterministic automaton A′
∃s ∈ E∗, σ ∈ E, y1, y2 ∈ Y , δ2(y0, s) = y1, δ2(y0, s) = y2, δ2(y1, σ)!, but ¬δ2(y2, σ)!. From
A2 ≺ A1, δ2(y0, s) = y1∧δ2(y1, σ)! implies that ∃x1 ∈ X, δ1(x0, s) = x1∧δ1(x1, σ)!. On the other
hand, A1 is deterministic, and hence, ∀x2 ∈ X, δ1(x0, s) = x2 ⇒ x2 = x1. Therefore, A2 ≺ A1
necessarily leads to (y2, x1) ∈ R2. But, ∃σ ∈ E such that δ1(x1, σ)! ∧ ¬δ2(y2, σ)!, meaning that
(y2, x1) ∈ R2 ∧ ¬(x1, y2) /∈ R1, i.e., R−1
6= R2, that contradicts with the hypothesis, and the
1
necessity is followed.
1 = R2 (and hence A1
1 such that A′
1
Next, let A1 and A2 to be substituted by AS and P1(AS)P2(AS), respectively, in Lemma 8.
Then, the existence of A′
S in Lemma 8 is characterized by the following lemma.
1 = A′
Lemma 9: Consider a deterministic automaton AS and its natural projections Pi(AS), i = 1, 2.
∼= P1(AS)P2(AS) if and only if
Then, there exists a deterministic automaton A′
S such that A′
S
there exist deterministic automata P ′
i (AS) such that P ′
i (AS) ∼= Pi(AS), i = 1, 2.
Proof: Let AS = (Q, q0, E = E1∪E2, δ), Pi(AS) = (Qi, qi
i = 1, 2, P1(AS)P2(AS) = (Z, z0, E, δ), P ′
Lemma 9 is then presented as follows.
1(AS)P ′
0, Ei, δi), P ′
2(AS) = (Z ′, z′
i, q′
i (AS) = (Q′
0, E, δ′
i),
0,i, Ei, δ′
). The proof of
Lemma 5, P ′
1, 2 implies that δ′
(Definition 7), δ′
i (AS) such that P ′
Sufficiency: The existence of deterministic automata P ′
1 and δ′
is a function, and hence P ′
i (AS) ∼= Pi(AS), i = 1, 2 leads to P ′
S = P ′
i (AS) ∼= Pi(AS), i =
2 are functions, and consequently from definition of parallel composition
2(AS) is deterministic. Moreover, from
2(AS) ∼= P1(AS)P2(AS), meaning that
∼= P1(AS)P2(AS).
2(AS) such that A′
S
Necessity: The necessity is proven by contraposition, namely, by showing that if there does
i (AS) ∼= Pi(AS), for i = 1 or i = 2, then
1(AS)P ′
1(AS)P ′
1(AS)P ′
i (AS) such that P ′
there exists a deterministic automaton A′
not exist deterministic automata P ′
there does not exist a deterministic automaton A′
S such that A′
S
∼= P1(AS)P2(AS).
Without loss of generality, assume that there does not exist a deterministic automaton P ′
1(AS)
1(AS) ∼= P1(AS). This means that ∃q, q1, q2 ∈ Q, e ∈ E1, t2 ∈ (E2\E1)∗, t ∈ E∗,
such that P ′
δ(q, t2e) = q1, δ(q, e) = q2, ¬(δ(q1, t) = q2 ⇔ δ(q2, t)!), meaning that δ(q1, t)! ∧ ¬δ(q2, t)! or
¬δ(q1, t)! ∧ δ(q2, t)!. Again without loss of generality we consider the first case and show that
it leads to a contradiction. From the first case, δ(q1, t)! ∧ ¬δ(q2, t)!, and definition of natural
projection, it follows that δ1([q]1, e) = [q1]1, δ1([q1]1, p1(t))!, δ1([q]1, e) = [q2]1, ¬δ1([q2]1, p1(t))!,
δ2([q]2, p2(e)) = [q2]2, ¬δ2([q2]2, p2(t))!, and hence, δ(([q]1, [q]2), e) = ([q1]1, [q1]2), δ(([q1]1, [q1]2),
p1(t))!, whereas δ(([q]1, [q]2), e) = ([q1]1, [q2]2), ¬δ(([q1]1, [q2]2), p1(t))! in P1(AS)P2(AS),
∼= P1(AS)P2(AS),
implying that there does not exist a deterministic automaton A′
S such that A′
S
and the necessity is proven.
Now, Lemma 3 is proven as follows.
Sufficiency: DC4 implies that there exist deterministic automata P ′
i (AS) such that P ′
1(AS)P ′
i (AS) ∼=
2(AS) ∼=
2(AS) such
Pi(AS), i = 1, 2. Then, from Lemmas 5 and 9, it follows, respectively, that P ′
P1(AS)P2(AS), and that there exists a deterministic automaton A′
that A′
S
∼= P1(AS)P2(AS) that due to Lemma 8, it results in R−1
1 = R2.
S = P ′
1(AS)P ′
Necessity: Let AS be deterministic, AS ≺ P1(AS)P2(AS) with the simulation relation R1
and P1(AS)P2(AS) ≺ AS with the simulation relation R2, and assume by contradiction that
R−1
1 = R2, but DC4 is not satisfied. Violation of DC4 implies that for i = 1 or i = 2, there
i (AS) ∼= Pi(AS). Therefore, due to
∼= P1(AS)P2(AS),
S such that A′
S
i (AS) such that P ′
Lemma 9, there does not exist a deterministic automaton A′
and hence, according to Lemma 8, it leads to R−1
1
does not exists a deterministic automaton P ′
6= R2 which is a contradiction.
D. Proof for Lemma 4
Lemma 4 comes from the associativity property of parallel decomposition [11] as for automata
Ai, i = 1, ..., m: A1 k A2 k · · · k Am−1 k AEm = A1 k (A2 k (· · · k (Am−1 k Am))).
E. Proof for Lemma 5
Lemma 5.1 is proven by showing that the relation R = {((q1, q3), (q2, q4))(q1, q2) ∈ R1 and
(q3, q4) ∈ R2} is a simulation relation, where, R1 and R2 are the respective simulations from
A1 to A2 and from A3 to A4.
Consider Ai = (Qi, q0
i , Ei, δi), i = 1, ..., 4, A1A3 = (Q1,3, (q0
3), E = E1 ∪ E3, δ1,3),
4), E = E2 ∪ E4, δ2,4), E1 = E3 and E2 = E4. Then, ∀(q1, q3), (q1, q3)′ ∈
A2A4 = (Q2,4, (q0
Q1,3, e ∈ E, q2 ∈ Q2, q4 ∈ Q4 such that δ1,3((q1, q3), e) = (q1, q3)′, (q1, q2) ∈ R1 and (q3, q4) ∈
R2, according to definition of parallel composition (Definition 7), we have (q1, q3)′ = (q′
1, q0
2, q0
1, q′
3) =
(δ1(q1, e), δ3(q3, e)) ,
(δ1(q1, e), q3) ,
(q1, δ3(q3, e)) ,
if δ1(q1, e)!, δ3(q3, e)!, e ∈ E1 ∩ E3;
if δ1(q1, e)!, e ∈ E1\E3;
if δ3(q3, e)!, e ∈ E3\E1;
and due to definition of simulation (Definition 5), A1 ≺ A2 and A3 ≺ A4, it follows that
i, i = 2, 4
∃q′
∃q′
i ∈ Qi, δi(qi, e) = q′
2 ∈ Q2, δ2(q2, e) = q′
2
4 ∈ Q4, δ4(q4, e) = q′
4
if e ∈ E2 ∩ E4
if e ∈ E2\E4
if e ∈ E4\E2
∃q′
This, in turn, due to definition of parallel composition implies that ∃(q′
2, q′
4) ∈ Q2,4 such that
δ2,4((q2, q4), e) = (q2, q4)′ =
(δ2(q2, e), δ4(q4, e)) ,
(δ2(q2, e), q4) ,
(q2, δ4(q4, e)) ,
if e ∈ E2 ∩ E4;
if e ∈ E2\E4;
if e ∈ E4\E2.
Therefore, ∀(q1, q3), (q1, q3)′ ∈ Q1,3, (q2, q4) ∈ Q2,4, e ∈ E, such that δ1,3((q1, q3), e) = (q1, q3)′
and ((q1, q3), (q2, q4)) ∈ R, then ∃(q2, q4)′ ∈ Q2,4, δ2,4((q2, q4), e) = (q2, q4)′, ((q1, q3)′, (q2, q4)′) ∈
R. This together with ((q0
4)) ∈ R, by construction, leads to A1A3 ≺ A2A4.
3), (q0
1, q0
2, q0
Now, to prove Lemma 5.2, we define the relation ¯R = {((q2, q4), (q1, q3))(q2, q1) ∈ ¯R1 and
(q4, q3) ∈ ¯R2}, where, ¯R1 and ¯R2 are the respective simulation relations from A2 to A1 and from
A4 to A3, and then similar to the proof of the first part, we show that ¯R is a simulation relation.
∼= A2A4 it remains to show that ∀(q1, q3) ∈ Q1,3, (q2, q4) ∈ Q2,4:
Now, to show that A1A3
((q1, q3), (q2, q4)) ∈ R ⇔ ((q2, q4), (q1, q3)) ∈ ¯R. This is proven by contradiction. Suppose
that ∃(q1, q3) ∈ Q1,3, (q2, q4) ∈ Q2,4 such that ((q1, q3), (q2, q4)) ∈ R ∧ ((q2, q4), (q1, q3)) /∈ ¯R,
or ((q2, q4), (q1, q3)) ∈ ¯R ∧ ((q1, q3), (q2, q4)) /∈ R. We prove that the first hypothesis leads
to contradiction, and the contradiction of the second hypothesis is followed, similarly. The
expression ((q1, q3), (q2, q4)) ∈ R ∧ ((q2, q4), (q1, q3)) /∈ ¯R means that ∃s ∈ E∗, δ1,3((q0
(q1, q3), δ2,4((q0
δ2,4((q2, q4), σ)! ∧ ¬δ1,3((q1, q3), σ)!. From Definition 7, δ2,4((q2, q4), σ)! means that
3), s) =
4), s) = (q2, q4), ∀e ∈ E, δ1,3((q1, q3), e)!: δ2,4((q2, q4), e)!; but, ∃σ ∈ E,
2, q0
1, q0
δ2(q2, σ)!, δ4(q4, σ)!
δ2(q2, σ)!
δ4(q4, σ)!
if e ∈ E2 ∩ E4;
if e ∈ E2\E4;
if e ∈ E4\E2.
Consequently, from (q2, q1) ∈ ¯R1 and E1 = E2 (due to A1
to A3
∼= A4), and Definition 5, it follows that
δ1(q1, σ)!, δ3(q3, σ)!
∼= A2), (q4, q3) ∈ ¯R2 and E3 = E4 (due
if e ∈ E2 ∩ E4 = E1 ∩ E3;
if e ∈ E2\E4 = E1\E3;
if e ∈ E4\E2 = E3\E1,
δ1(q1, σ)!
δ3(q3, σ)!
that from Definition 7 leads to δ1,3((q1, q3), σ)! which contradicts with the hypothesis and the
proof is followed.
REFERENCES
[1] V. R. Lesser, "Cooperative multiagent systems: A personal view of the state of the art," IEEE Transactions on Knowledge
and Data Engineering, vol. 11, pp. 133 -- 142, 1999.
[2] P. U. Lima and L. M. Custdio, Multi-Robot Systems, Book Series Studies in Computational Intelligence, Book Innovations
in Robot, Mobility and Control. Berlin: Springer Berlin / Heidelberg, 2005, vol. 8.
[3] J. Choi, S. Oh, and R. Horowitz, "Distributed learning and cooperative control for multi-agent systems," Automatica,
vol. 45, no. 12, pp. 2802 -- 2814, 2009.
[4] E. Semsar-Kazerooni and K. Khorasani, "Multi-agent team cooperation: A game theory approach," Automatica, vol. 45,
no. 10, pp. 2205 -- 2213, 2009.
[5] G. E. Fainekos, A. Girard, H. Kress-Gazit, and G. J. Pappas, "Temporal logic motion planning for dynamic robots,"
Automatica, vol. 45, no. 2, pp. 343 -- 352, 2009.
[6] Z. Ji, Z. Wang, H. Lin, and Z. Wang, "Brief paper: Interconnection topologies for multi-agent coordination under leader-
follower framework," Automatica, vol. 45, no. 12, pp. 2857 -- 2863, 2009.
[7] V. Crespi, A. Galstyan, and K. Lerman, "Top-down vs bottom-up methodologies in multi-agent system design," Auton.
Robots, vol. 24, no. 3, pp. 303 -- 313, 2008.
[8] P. Tabuada and G. Pappas, "Linear time logic control of discrete-time linear systems," Automatic Control, IEEE Transactions
on, vol. 51, no. 12, pp. 1862 -- 1877, Dec. 2006.
[9] C. Belta, A. Bicchi, M. Egerstedt, E. Frazzoli, E. Klavins, and G. Pappas, "Symbolic planning and control of robot motion
[grand challenges of robotics]," Robotics and Automation Magazine, IEEE, vol. 14, no. 1, pp. 61 -- 70, March 2007.
[10] M. Kloetzer and C. Belta, "Automatic deployment of distributed teams of robots from temporal logic motion specifications,"
Robotics, IEEE Transactions on, vol. 26, no. 1, pp. 48 -- 61, feb. 2010.
[11] C. G. Cassandras and S. Lafortune, Introduction to discrete event systems. USA: Springer, 2008.
[12] R. Kumar and V. K. Garg, Modeling and Control of Logical Discrete Event Systems. Norwell, MA, USA: Kluwer
Academic Publishers, 1999.
[13] R. Gotzhein, "Temporal logic and applications: a tutorial," Comput. Netw. ISDN Syst., vol. 24, no. 3, pp. 203 -- 218, 1992.
[14] P. Wolper, "Constructing automata from temporal logic formulas: a tutorial," pp. 261 -- 277, 2002.
[15] M. Mukund, From global specifications to distributed implementations, in B. Caillaud, P. Darondeau, L. Lavagno (Eds.),
Synthesis and Control of Discrete Event Systems, Kluwer. Berlin: Springer Berlin / Heidelberg, 2002.
[16] R. Morin, "Decompositions of asynchronous systems," in CONCUR '98: Proceedings of the 9th International Conference
on Concurrency Theory. London, UK: Springer-Verlag, 1998, pp. 549 -- 564.
[17]
I. Castellani, M. Mukund, and P. S. Thiagarajan, "Synthesizing distributed transition systems from global specification," in
Proceedings of the 19th Conference on Foundations of Software Technology and Theoretical Computer Science. London,
UK: Springer-Verlag, 1999, pp. 219 -- 231.
[18] Y. Willner and M. Heymann, "Supervisory control of concurrent discrete-event systems," International Journal of Control,
vol. 54, pp. 1143 -- 1169, 1991.
[19] C. Zhou, R. Kumar, and S. Jiang, "Control of nondeterministic discrete-event systems for bisimulation equivalence,"
Automatic Control, IEEE Transactions on, vol. 51, no. 5, pp. 754 -- 765, may 2006.
[20] R. Alur, T. Henzinger, G. Lafferriere, and G. Pappas, "Discrete abstractions of hybrid systems," Proceedings of the IEEE,
vol. 88, no. 7, pp. 971 -- 984, jul 2000.
[21] M. V. Lawson, Finite Automata, Book Series in Control Engineering, Book Handbook of Networked and Embedded Control
System. Springer Boston / Birkhuseg, 2007, vol. I.
[22] K. Rudie and W. Wonham, "Think globally, act locally: decentralized supervisory control," Automatic Control, IEEE
Transactions on, vol. 37, no. 11, pp. 1692 -- 1708, nov 1992.
[23] P. Tabuada and G. Pappas, "From discrete specifications to hybrid control," in Decision and Control, 2003. Proceedings.
42nd IEEE Conference on, vol. 4, Dec. 2003, pp. 3366 -- 3371 vol.4.
[24] C. Belta, V. Isler, and G. Pappas, "Discrete abstractions for robot motion planning and control in polygonal environments,"
Robotics, IEEE Transactions on, vol. 21, no. 5, pp. 864 -- 874, oct. 2005.
|
1910.13196 | 1 | 1910 | 2019-10-29T11:06:56 | Deep Decentralized Reinforcement Learning for Cooperative Control | [
"cs.MA",
"cs.LG",
"eess.SY",
"eess.SY"
] | In order to collaborate efficiently with unknown partners in cooperative control settings, adaptation of the partners based on online experience is required. The rather general and widely applicable control setting, where each cooperation partner might strive for individual goals while the control laws and objectives of the partners are unknown, entails various challenges such as the non-stationarity of the environment, the multi-agent credit assignment problem, the alter-exploration problem and the coordination problem. We propose new, modular deep decentralized Multi-Agent Reinforcement Learning mechanisms to account for these challenges. Therefore, our method uses a time-dependent prioritization of samples, incorporates a model of the system dynamics and utilizes variable, accountability-driven learning rates and simulated, artificial experiences in order to guide the learning process. The effectiveness of our method is demonstrated by means of a simulated, nonlinear cooperative control task. | cs.MA | cs |
Deep Decentralized Reinforcement
Learning for Cooperative Control (cid:63)(cid:63)
Florian Kopf, (cid:63) Samuel Tesfazgi, (cid:63)
Michael Flad and Soren Hohmann
Institute of Control Systems, Karlsruhe Institute of Technology (KIT),
76131 Karlsruhe, Germany (e-mail: [email protected])
Abstract: In order to collaborate efficiently with unknown partners in cooperative control
settings, adaptation of the partners based on online experience is required. The rather general
and widely applicable control setting, where each cooperation partner might strive for individual
goals while the control laws and objectives of the partners are unknown, entails various challenges
such as the non-stationarity of the environment, the multi-agent credit assignment problem,
the alter-exploration problem and the coordination problem. We propose new, modular deep
decentralized Multi-Agent Reinforcement Learning mechanisms to account for these challenges.
Therefore, our method uses a time-dependent prioritization of samples, incorporates a model of
the system dynamics and utilizes variable, accountability-driven learning rates and simulated,
artificial experiences in order to guide the learning process. The effectiveness of our method is
demonstrated by means of a simulated, nonlinear cooperative control task.
Keywords: Reinforcement Learning, Deep Learning, Learning Control, Shared Control,
Decentralized Control, Machine Learning, Non-stationary Systems, Nonlinear Control.
1. INTRODUCTION
In numerous control problems including highly-automated
driving, robotics and manufacturing plants, several entities
(e.g. machines and/or humans) are required to collaborate
in order to achieve complex control objectives. Although
the cooperating partners' goals usually do not completely
contradict each other, the partners might have individual
preferences. Suitable partners need to be flexible enough
to account for the preferences of each other while repre-
senting their interests. We refer to this kind of setting
as Cooperative Control (Kopf et al., 2018) 1 in order to
emphasize that partners need to cooperate with each other
and make compromises when conflicts occur. However, this
does not necessarily imply that they are facing a so-called
fully cooperative setting with a single global goal. Instead
individual goals for the agents are allowed. Bearing the
vision of future human-machine collaboration and plug-
and-play machine-machine cooperation in mind, we focus
on the case where the partners do not know the others'
control laws or objective functions and no explicit com-
munication is used. This decentralized setting requires the
partners to constantly adapt to each other based on online
experience.
Due to its generalization capabilities and major successes
in the single-agent Reinforcement Learning (RL) setting,
Multi-Agent Reinforcement Learning (MARL) has re-
cently become the focus of increasing attention in order to
solve Cooperative Control problems. Compared to single-
(cid:63) These authors contributed equally to this work.
(cid:63)(cid:63)This work has been submitted to IFAC for possible publication.
1 Alternatively termed Mixed Cooperative-Competitive Control
(Lowe et al., 2017).
agent RL, the multi-agent case is inherently more complex
as agents directly or indirectly interact with each other and
their common environment.
A major challenge occurring here is the non-stationarity
of the dynamics from the local perspective of each agent
which violates the Markov property that is commonly
assumed in RL. Besides the severe challenge of non-
stationarity, it is in general difficult to deduce to what ex-
tent an agent contributed to state transitions and thus the
rewards received as each agent is capable of manipulating
the environment. This is known as the multi-agent credit
assignment problem (Chang et al. (2004)). Additionally,
the exploration-exploitation trade-off common to RL even
worsens in the cooperative case. This is due to other learn-
ing agents which might concurrently explore. Matignon
et al. (2012) refer to this problem as alter-exploration. Fur-
thermore, the coordination problem states that successful
cooperation requires the agents to coordinate their con-
trols in order to avoid e.g. shadowed equilibria (Matignon
et al. (2012)). Finally, in order to cope with the majority of
control problems, we require compatibility with continuous
state and control spaces and nonlinear systems and do
not assume restrictions concerning the structure of the
agents' objectives. However, as a system model is usually
available in control engineering as a result of model design
or an identification process, we desire to incorporate this
beneficial knowledge into our method. Due to causality,
we assume that the joint control signals of other agents
are not instantaneously measurable at run time but are
retrospectively measurable or deducible.
1.1 Related Work
In the following, a short overview regarding related work
concerning cooperative control will be given and analyzed
w.r.t. our problem. One possible approach as proposed by
Kopf et al. (2019) is to identify and constantly update the
aggregated control law of all other agents. Relying on a
model of the system dynamics, this allows a simulation-
based optimization of the cooperative control problem.
The concept of opponent or partner modeling is also
discussed by Lowe et al. (2017) (Section 4.2 therein). When
facing a dynamic game setting, another approach to cope
with unknown partners in cooperative scenarios is given by
the identification of associated cost functionals as done by
Kopf et al. (2017) and Inga et al. (2018) and a subsequent
optimization. In the human-machine context, this setting
is motivated by the assumption that human motion can
be modeled by means of optimal control (Scott, 2004).
In contrast to these methods, the following approaches
avoid the need to identify the partners' cost functionals
or control
laws. Among these methods, Adaptive Dy-
namic Programming in the Cooperative Control setting
(Vamvoudakis and Lewis (2011); Kopf et al. (2018)) fo-
cuses on efficient adaptation from a control-oriented per-
spective but has more restricting assumptions regarding
reward structures and system dynamics compared to deep
RL methods. Thus, the following methods either rely on
extensions to Deterministic Policy Gradient (DPG) meth-
ods (Silver et al. (2014)) or extensions to Deep Q-Networks
(DQN) (Mnih et al. (2015)).
Among the DPG methods, either all agents need to know
the policy parameters of all others (Gupta et al., 2017),
explicit opponent modeling is required when facing our
problem (cf. (Lowe et al., 2017, Section 4.2)), or all agents
share the same critic and a global reward function (Foer-
ster et al., 2018), i.e. the agents are not decentralized and
fully cooperative. Furthermore, the DPG based methods
suffer from increasing variance in multi-agent domains (cf.
Lowe et al. (2017) and Foerster et al. (2018)), which desta-
bilizes the training process particularly with independently
learning agents. Concerning the DQN-based methods, they
either work in the fully cooperative setting with finite state
and control spaces (Foerster et al. (2017); Matignon et al.
(2007)), are limited to finite control spaces (Omidshafiei
et al., 2017) or finite state and control spaces (Palmer
et al., 2018).
1.2 Contributions of This Paper
As none of the deep MARL methods in literature ful-
fills our control-oriented requirements, we propose a new
approach for cooperative control in continuous state and
control spaces. Our method does not depend on the ex-
plicit identification of the other agents' behavior. Instead,
an adapting automation is explored, which is not reliant
on the premise of other agents behaving optimally and
is expected to facilitate a high degree of generalizability
across domains and partners. Compared to recent deep RL
methods in the multi-agent domain, we face the challenge
of decentralized agents with no knowledge of the partners'
control strategies or objectives and no explicit means of
communication and present three new, modular mecha-
nisms which explicitly address the associated challenges.
Although the deep MARL methods in Section 1.1 cannot
applied be directly to our problem setting, they reveal
reoccurring mechanisms which we rely on: First, exten-
sions to the experience replay memory (ERM) in order
to counteract the difficulty of applying experience replay
in non-stationary environments. Second, variable learning
rates in order to induce coordination and facilitate the
use of a temporal dimension in the sampling process. We
propose Temporal Experience Replay (TER) to account
for the non-stationarity of the environment each agent
faces. The main idea behind TER is a time-dependent
prioritization of samples in the experience replay memory.
Furthermore, we introduce the idea of Imagined Experi-
ence Replay (IER), which benefits from a model of the
system dynamics and grounds the training process by
means of fictional experiences. IER can be understood as
an adaptation of the idea of imagination rollouts (cf. Gu
et al. (2016)) to cope with the challenges encountered in
multi-agent settings. In addition, in order to address the
multi-agent credit assignment problem, we propose a new
mechanism of variable learning rates. Our accountability-
driven approach termed impact Q-learning (IQL) ties the
learning rate to the agent's contribution towards the joint
control. We further combine IQL and IER to simulate
targeted cooperation scenarios in order to exhaust poten-
tial coordination between agents. Finally, the mechanisms
are made dependent on an exploration rate such that the
influence of each distinct concept is varied according to
its current utility. This increases their effectiveness and
reduces issues connected to alter-exploration.
2. FORMAL PROBLEM DEFINITION AND
PREREQUISITES
We now formalize our problem definition and introduce
prerequisites on which our proposed mechanisms rely on.
2.1 Formal Problem Definition
Consider a discrete-time system f : X × U → X that is
controlled by N agents given by
xk+1 = f (xk, u1,k, . . . , uN,k),
(1)
where xk ∈ X ⊆ Rn denotes the state at time step k,
ui,k ∈ Ui ⊆ R the control of agent i ∈ N = {1, . . . , N}
and U = U1 ×···× UN the joint control space. Depending
on the current state xk and controls ui,k, each agent i ∈ N
experiences a reward ri that results from a reward function
gi : X × U → R, i.e.
ri,k = gi(xk, u1,k, . . . , uN,k).
(2)
The goal of each agent is to adapt his control law πi : X →
Ui in order to maximize his value
∞(cid:88)
∞(cid:88)
V π
i (xk) =
γk
i ri,k =
γk
i gi(xk, π1(xk), . . . , πN (xk)),
k=0
k=0
(3)
i.e. the long-term discounted reward under the tuple
laws π = (π1, . . . , πN ), where γi ∈ [0, 1)
of control
denotes a discount factor. Thus, our deterministic game
setting (in contrast to the stochastic game definition in
Bu¸soniu et al. (2010)) is defined by the tuple G =
(X, U1, . . . , UN , f, g1, . . . , gN , γ1, . . . , γN ). Our problem is
then formalized as follows.
Problem 1. Given the game G, each agent i ∈ N knows
the system dynamics f and his own reward function gi.
Furthermore, each agent i ∈ N receives his current reward
ri,k at time step k and is able to deduce the previous
controls uj,k−1, ∀j ∈ N \ {i} of other agents but has no
access to the current controls uj,k, other agents' control
laws πj, their reward functions gj or actual rewards. In
this setting, each agent i ∈ N aims at adapting his control
law πi in order to maximize V π
i as defined in (3).
2.2 Prerequisites Concerning Deep Q-Networks
Our algorithm is based on DQN. Thus, the fundamental
concepts of Q-learning and DQN are introduced in the
following. Q-learning (Watkins (1989)) is an iterative
algorithm which intends to learn an optimal state-action-
value function Q∗. Here,
Q∗(xk, uk) = max
π
Qπ(xk, uk)
(4)
holds, where Qπ(xk, uk) represents the discounted long-
term cost, if an agent is in state xk and applies the control,
i.e. action, uk at time step k and follows the control law
π thereafter. The relevance of Q∗ becomes clear as the
optimal control law maximizing the long-term discounted
reward (cf. (3) for N = 1) is given by
π∗(xk) = arg max
Q∗(xk, uk).
uk
The update rule in order to estimate Q∗ is given by
Q(xk, uk) ← Q(xk, uk)
+ αk
rk + γ max
u
Q(xk+1, u) − Q(xk, uk)
,
(5)
(cid:105)
(cid:125)
(cid:104)
(cid:124)
(cid:123)(cid:122)
δk
(6)
where δk denotes the temporal difference (TD) error and
αk ∈ (0, 1] a learning rate. The TD error δk thus measures
the difference between the current Q-function estimate
Q(xk, uk) and the TD target rk + γ maxu Q(xk+1, u). The
tuple χk = (xk, uk, rk, xk+1) is taken from interaction with
the environment.
In order to extend Q-learning to continuous state spaces,
function approximators such as deep neural networks
which parametrize the Q-function have been introduced.
Here, the work of Mnih et al. (2015) marked a break-
through, as the introduction of Experience Replay (ER)
significantly improved training. The idea is to randomize
training samples in order to remove correlation between
observed state-transition sequences. Therefore, experience
tuples χk are stored in an ER memory (ERM) M at each
time step k. A Q-learning update is then performed by
sampling (e.g. uniformly at random) from the ERM and
minimizing the associated squared TD error δk. In order
to account for continuous control spaces, Gu et al. (2016)
introduced the concept of Normalized Advantage Func-
tions (NAF), allowing to deduce an analytical expression
in order to solve (5).
3. DECENTRALIZED COOPERATIVE CONTROL
METHOD
In order to gain control of the challenges associated
with Problem 1, we propose a time-dependent mechanism
termed Temporal Experience Replay (TER) to account
for the non-stationary environment, include known system
dynamics by means of Imagined Experience Replay (IER)
and use variable learning rates with the proposed Impact
Q-Learning (IQL) in order to induce coordination. As
these mechanisms can be applied in a modular fashion,
they are separately introduced and then combined in Sec-
tion 3.4.
3.1 Temporal Experience Replay (TER)
(7)
(8)
The proposed method of Temporal Experience Replay at-
tempts to unify the idea of favoring more recent expe-
riences with the concept of more probable sampling of
experiences according to a prioritization factor. Analogue
to Prioritized Experience Replay (Schaul et al., 2016), we
suggest to bias the sampling process. However, instead of
utilizing the TD error for the prioritization, we propose to
focus towards recent experiences by introducing a temporal
prioritization τ , which is proportional to the time that has
passed since collection kc of the state-transition:
τkc (k) = exp(cid:0) − k − kc(cid:1) + ξtemp,
with the sampling probability Pkc (k) given by
τkc (k)(cid:80)
l τl(k)
.
Pkc(k) =
In (7) the optional offset ξtemp can be used to ensure that
experiences are sampled with non-zero probability and the
term k denotes the current time step. Hence, to compute
(7) and (8) at runtime, the experience tuple has to be
extended by the respective current time step, producing
the new tuple:
χ∗
k = (xk, uk, rk, xk+1, k).
(9)
The underlying idea is that agents are more capable of
adjusting to ever changing policies of other agents by
experiencing recent state-transitions tuples more often
than old ones. However, the TER as described by (7)
and (8) is impractical, as it leads to two major issues:
Firstly, similar to approaches that restrict the memory size
itself, the proposed temporal prioritization suffers from
biasing the ERM too much towards recent experiences.
This can lead to over-fitting of an agent's policy. Secondly,
the temporal prioritization increases the computational
complexity of the sampling process to a degree that is
not feasible in practice. This is due to the computation
of the temporal prioritization τkc(k) itself, as it has to be
updated for each experience tuple at every time step.
To overcome both of these issues, a two step sampling
process is proposed. Initially, a macro-batch B of size
B is sampled uniformly at random from the complete
experience replay buffer M. Subsequently, a smaller mini-
batch T of size t, with t < B, is sampled from B utilizing
the temporally prioritized probabilities given in (8). By
dividing the sampling process into two manageable parts,
both of the above mentioned problems are solved. The
macro-batch B is only of size B, thus, the computational
complexity of calculating the temporal priorities τkc (k) is
equally reduced to B. Additionally, the initial macro-batch
is sampled uniformly at random, which reduces the risk of
overemphasizing experiences related to recent episodes.
In order to account for the varying exploration rate εk
of agents at different stages of the training process, we
propose an additional exploration rate dependency of
B yielding a time-dependent macro-batch size Bk. TER
attempts to induce adaptation to other agents policy
changes. Therefore, it is most effective when the partners'
policies start to converge and are less influenced by ex-
ploration noise. Thus, experiences should be sampled uni-
formly at random during early training (i.e. when εk ≈ 1),
which can be achieved by choosing Bk close to the mini-
batch size t, whereas during later training stages, i.e. once
εk → 0, Bk should approach the final macro-batch size B.
Consequently, we choose
Bk = (B − t)(1 − εk) + t.
(10)
3.2 Imagined Experience Replay (IER)
The above mentioned augmentations to ER attempt to
either stabilize the training process in order to make
agents less susceptible to changing environment dynamics
or bias learning towards recent experiences to enable
agents to adapt to changes in the dynamics. In any
case it is acknowledged that the other agents behavior is
indissociable from the dynamics of the environment, which
is generally a reasonable presumption given independent
and decentralized agents. However, due to the assumption
that a system model is available, it becomes possible to
ground the training process through simulated experiences
in which the partners' controls are marginalized leading to
stationary environment dynamics. This is the fundamental
idea of our second proposed modification to the ER, which
is termed Imagined Experience Replay (IER).
The concept of IER was inspired by the imagination
roll-outs developed by Gu et al. (2016), who proposed
the idea of accelerating the training process by utilizing
a learned model to simulate artificial experiences that
were then added to the replay buffer. Differently, IER is
used here to simulate experiences, which would not occur
under normal circumstances. Specifically, all other agents'
controls u−i = {u1, . . . , ui−1, ui+1, . . . , uN} are set to zero
while retaining the agent's control ui unchanged. Given a
regular experience
χi,k = (xk, ui,k, ri,k, xk+1)
(11)
for agent i which occurred at time k, the successor state
xk+1 and received reward ri,k can be substituted by
utilizing the underlying system dynamics f and reward
function gi: 2
χi,k =(cid:0)xk, ui,k, gi(xk, ui,k, u−i,k), p(xk, ui,k, u−i,k)(cid:1).
χi,k =(cid:0)xk, ui,k, gi(xk, ui,k, 0)
(cid:125)
(12)
Subsequently, an imagined experience χ can be simulated
by replacing the other agents' controls u−i by 0 yielding
the imagined successor state xk+1 and reward ri,k:
, p(xk, ui,k, 0)
(cid:123)(cid:122)
(cid:123)(cid:122)
(cid:1).
(13)
(cid:124)
(cid:124)
(cid:125)
ri,k
xk+1
In contrast to the imagination roll-out of Gu et al. (2016),
the imagined experiences in (13) are not stored to the
actual ERM and sampled from there. Instead, an explo-
ration rate dependent probability P is utilized to deter-
mine, whether an imagined experience is computed in
convenience
2 For
and
p(xk, ui,k, u−i,k) evaluate gi(·) and p(·) at the state xk while
agent i applies the control ui,k and all other agents controls are
denoted by the tuple u−i,k.
gi(xk, ui,k, u−i,k)
notation,
of
addition to the sampled, observed experience. Once used
for training, the imagined experience is discarded in order
to reduce the risk of overemphasizing artificial experiences
in which partners are non-existent. During the initial
training phase, agents predominantly explore random con-
trols. Hence, it is not possible to infer other agents policies
from observations, and the application of IER, in order to
stabilize the training process, is most useful, as experiences
are simulated in which solely the agent i interacts with the
environment. These imagined experiences, at this stage of
training, are essentially observations for which the explo-
ration noise of other agents is not present. Consequently,
the probability P of simulating imagined experiences χi,k
at time step k is proposed to be proportional to the current
exploration rate εk, i.e. P (k) ∼ εk.
However, during the later stages of training, when policies
start to converge and are less influenced by exploratory
controls, the coordination between agents and the adap-
tation to the partners' policies becomes more important.
Upon closer examination it can be seen that by gener-
ating artificial experiences in which cooperation between
agents is simulated, the IER can potentially be utilized
to induce coordination between agents. Opposite to (13),
these imagined coordination experiences are more useful
during later stages of training. Therefore, the respective
sampling probability is proposed as Pcoord(k) ∼ (1 − εk).
When generating experiences with the purpose of inducing
coordination, it has to be considered that the final algo-
rithm is required to entail mixed cooperative-competitive
task types. Thus, it cannot generally be presumed that
the agents respective goals are compatible. Consequently,
it is proposed that the IER is utilized to simulate three
additional scenarios:
(1) In order to induce coordination, the control ui of
agent i is discarded, causing it to be idle:
idle =(cid:0)x, 0, gi(x, 0, u−i), p(x, 0, u−i)(cid:1).
χ(i)
(14)
Therefore, by utilizing χ(i)
idle, agent i can observe how
other agents behave by themselves and whether the
resulting environment transitions are beneficial.
(2) Here, the first of two cooperation scenarios χ(i)
coop1 is
simulated. For this purpose, the agent's controls ui
are set to be equal to the average of all other agents'
joint control u−i:
resulting in:
∀j∈N\{i}
coop1 =(cid:0)x, u−i, gi(x, u−i, u−i), p(x, u−i, u−i)(cid:1).
χ(i)
(15)
(3) In the second cooperation scenario χ(i)
coop2, the inverse
is generated. Each element of the other agents' joint
control u−i is modified to be equal to ui:
coop2 =(cid:0)x, ui, gi(x, ui, u−i), p(x, ui, u−i)(cid:1),
χ(i)
(16)
with uj := ui,∀j ∈ {1, . . . , N}.
By imagining these three scenarios, potential coordination
possibilities are exhausted. The first one specifically en-
ables agents to evaluate whether being idle leads to an
acceptable reward, which in competitive environments is
(cid:88)
uj,
u−i =
1
N − 1
generally discouraged and will be evaluated correspond-
ingly, whereas the second and third evaluate the effect if
agent i imitates the average control of the others or if all
agents stick to the control of agent i.
3.3 Impact Q-Learning (IQL)
In the previous subsections, different additions to the ERM
were proposed, which focused mainly on providing stabil-
ity to counteract the problem of a non-stationary environ-
ment. An additional, often utilized mechanism in MARL
are variable learning rates. The hysteretic (Matignon et al.
(2007)) and lenient (Palmer et al. (2018)) method are two
representatives of optimistic learners, which are generally
well suited to induce coordination. However, the core prin-
ciple of optimistic agents, which reduce their learning rate
given negative experiences, is diametrically opposed to
the challenge of multi-agent credit assignment. In order
to credit agents correctly with respect to the observed
outcome, it is necessary for them to not only learn notably
from positive experiences but also from negative ones. Fur-
thermore, the concept of tying the learning rate to rewards
(Matignon et al. (2007)) is itself flawed to combat credit
assignment, as the attention an agent should pay to certain
experiences ideally does not depend on the quality of the
outcome, but on the contribution of an agent towards the
observed outcome.
Therefore, we propose a novel approach for solving the
multi-agent credit assignment problem using variable
learning rates. We attempt to tie the variable learning rate
to the actual contribution of an agent towards the observed
state transitions. This is facilitated by the retrospective
observation of all agents' controls, as it enables each agent
to compare its control ui,k at time step k to the separately
remaining joint control u−i,k of all other agents. To this
end we introduce a novel quantity, called impact factor
λi,k =
,
(17)
N(cid:80)
ui,k
uj,k
j=1
which describes the agent's relative contribution to the
joint control, and thus, to experienced state transitions. In
order to enable the computation of a meaningful impact
factor λ in (17) it is presupposed for IQL that agents share
the same control space U1 = . . . = UN and that all agents'
controls manipulate the system equally. Subsequently, the
update rule (6) for an agent i can be modified to apply
different learning rates depending on the agent's impact
factor:
δk ← rk + γ max
Qi(xk+1, u) − Qi(xk, uk),
u
Qi(xk, uk) + αδk
Qi(xk, uk) + σδk
Qi(xk, uk) + βδk
if 1.0 ≥ λi,k > λhigh
if λhigh ≥ λi,k ≥ λlow
if λlow > λi,k ≥ 0,
Qi(xk, uk) ←
(18)
with 0 < β < σ < α < 1.
In (18), the Q-learning
update rule is partitioned into three distinct impact ranges
with which it is possible to differentiate whether an agent
had a high, medium, or low influence towards a state
transition. Hence, the amount an agent learns from an
experience is proportional to its respective contribution
or impact. When a positive experience is observed, the
Q-value estimate is only increased heavily, if the agent
can be credited for the event. On the other hand, when
a punishment occurs, the agent is mainly discouraged
from the corresponding state-action pair,
if the agent
is at least in part responsible. Particularly this kind
of accountability-driven learning behavior is required for
agents to overcome the credit assignment challenge.
3.4 Algorithm
Upon closer examination of IER,
it can be seen that
the concept of simulated experiences specifically for the
coordination scenarios in (14), (15), and (16) may reduce
the coordination problems severity, but also produces
additional computational effort. Thus, it is advisable to
limit these calculations to state-action pairs with high
potential for coordination. This can be done by utilizing
the computed impact factors. In (18), three intervals with
different degrees of an agent's impact were distinguished.
When analyzing the ones corresponding to learning rates
α and β, it can be seen that the potential for coordination
is limited here, because the agent either predominantly
contributes towards the state-transition or only has a
minor impact. However, for the case of medium learning
rates σ, the impact of agents, particularly in the case of
only few agents, is distributed more evenly, which in turn
increases the need for coordination. In this instance, the
simulation of different IER scenarios is most powerful and
the trade-off between computational effort and induced
coordination most beneficial. Further, two distinct kinds
of medium-impact experiences are distinguished. Either
the agent's control ui,k and the average of all remaining
controls u−i,k work in the same direction, or against each
other. This can be determined by sampling an experience
χk and computing a coordination coefficient ψk as such:
ψi,k = sgn(u−i,k · ui,k).
(19)
If ψi,k equals 1, agent i and the others work in the same
direction and it is not necessary to simulate the coordi-
nation experiences χidle, χcoop1 and χcoop2. Instead, the
learning rate σ, which is normally used for λhigh ≥ λi,k ≥
λlow in (18), is substituted by the larger learning rate α.
Therefore, agents are induced to emphasize cooperative
experiences during the learning process. In the case that
agents act in opposing directions, ψi,k equals -1. Besides
the sampled experience χk, the artificial experiences χidle,
χcoop1 and χcoop2 are simulated, and subsequently, the
agent is trained on all of them. Here the lowest learn-
ing rate β is applied, because the trained on experiences
have not actually occurred and are only imagined for
coordination purposes. Thus, the instances for which the
computational strenuous task of simulating multiple co-
ordination experiences is required, can be reduced greatly
and focused to occasions connected to the highest expected
learning progress. The resulting algorithm after finally
assembling the above described mechanisms is described in
Algorithm 1, where U (a, b) denotes a uniform distribution
in the intervall [a, b].
4. RESULTS
In this section, the previously described algorithm is
trained on a control task. Subsequently, the method's
effectiveness is evaluated.
Algorithm 1 Deep impact Q-learning with TER and IER
are defined by the nonlinear differential equations
weights θ and θ, ER buffer M ← ∅ with
size M , exploration rate ε = 1
learning rates α, σ, and β, ERM size M ,
target update frequency m, decay rate ,
minimum exploration rate εmin, maximum
number of episodes Emax and maximum
time steps per episode Kmax
1: Input: macro-batch size B, mini-batch size t,
2:
3:
4:
5:
6:
7: Initialize: Q(x, u; θ) and Q(x, u; θ) with random
8:
9:
10: for episode e = 1, . . . , Emax do
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
k = 0
while episode not terminated and k ≤ Kmax do
With probability ε select random control uk
Otherwise select uk = arg maxu Q(xk, u; θ)
Execute uk and observe rk, xk+1
Store tuple (xk, uk, rk, xk+1, k) in M
Compute Bk = (B − t)(1 − εk) + t
Sample uniformly at random B of size Bk
Compute τkc(k) for transition in B
for each χ ∈ T do
Sample T of size t ∼ Pkc(k) = τkc(k)/(cid:80)
(cid:1)
Compute χc =(cid:0)xc, uc, rc, xc+1
Compute λc = ui,c/(cid:80)
Extract time of collection c from χ
Draw random variable w ∼ U (0, 1)
if w < εc (exploration rate at time c) then
Set yc = rc + γ maxu Q(xc+1, u; θ )
Update θ with learning rate β for yc
(7)
l τl (8)
26:
27:
28:
29:
end if
(10)
(17)
juj,c
Set yc = rc + γ maxu Q(xc+1, u; θ )
if λc > λhigh then
Update θ with learning rate α for yc
else if λhigh ≥ λc ≤ λlow then
if sgn (u−i,c · ui,c) ≥ 0 then
Update θ with learning rate α for yc
else
Update θ with learning rate σ for yc
end if
if sgn(u−i,c · ui,c) < 0 and εc < w then
Compute χidle,c, χcoop1,c, χcoop2,c
Set target yidle,c, ycoop1,c, and ycoop2,c
Update θ with learning rate β for
yidle,c, ycoop1,c, and ycoop2,c
end if
else
Update θ with learning rate β for yc
end if
end for
Every m steps, update target network: θ ← θ
k = k + 1
end while
Decay exploration rate: ε ← max[ · ε; εmin]
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
49:
50:
51:
52:
53: end for
θk =
g sin(θk) − cos(θk)
(cid:34)
Fk,res + mpole l(cid:2) θ2
4
3
l
mpole + mcart
(cid:34)−Fk,res − mpole l θ2
(cid:35)
k sin(θk) − θk cos(θk)(cid:3)
− mpole cos2(θk)
mpole + mcart
k sin(θk)
(cid:35)
,
(20)
,
mpole + mcart
sk =
(21)
where g = −9.8 m/s2, mpole = 0.1 kg, mcart = 1.0 kg,
l = 0.5 m (half-pole length) and Fk,res ∈ [−10 N, 10 N]
(clipped sum of forces). In (20) and (21), θk denotes the
angular displacement of the pole from 0 rad, which is
defined by the pole standing perfectly upright. The cart's
position is defined by sk with the center being at 0 m and
. The
successor state xk+1 according to (1) is calculated using
the semi-implicit Euler method with a discrete time step
of 0.02 s.
the system state is given by xk = (cid:2)sk sk θk
(cid:3)(cid:124)
θk
We assume that one agent focuses on balancing the pole
upright, while the other agent is rewarded depending on
the position of the cart. Thus, the first agent receives a
reward r1,k of 1 for each time step k in which the pole angle
θk ∈ (−0.21 rad, 0.21 rad). If the episode is terminated,
a reward of -1 is observed. On the contrary, the second
agent's reward r2,k solely depends on the current cart
position sk. Specifically, a step-wise reward function is
defined as such:
+5,
+1,
0,
−1,
if sk − s∗ < 0.1 m
if 0.1 m ≤ sk − s∗ < 0.5 m
if 0.5 m ≤ sk − s∗ < 2.4 m
if episode is terminated,
r2,k =
(22)
with the target position denoted by s∗. We choose
s∗ = 0 m. Agent 2 receives the highest reward in a small
range around the target position, while the received reward
is reduced step-wise once a certain boundary distance is
exceeded. At the beginning of each of the Emax = 2000
training episodes, the cart is initiated uniformly at random
with the initial position s0 ∼ U (−2.3 m, 2.3 m) and the
initial pole angle θ0∼ U (−0.085 rad, 0.085 rad). An episode
is terminated once one of the intervals sk ∈ [−2.4, 2.4] m
or θk ∈ [−0.21, 0.21] rad is exceeded or Kmax = 3000 time
steps have passed. We set γ = 0.999.
In our work, a dueling network architecture with NAFs as
introduced by Gu et al. (2016) was used. Additionally, mul-
tiple fully connected layers and dropout layers are stacked
in front of the dueling network architecture to process ob-
servations. A description of the parameters corresponding
to the network architecture and the hyperparameters used
for training is given in the Appendix A.
4.2 Simulations
4.1 Example System and Network Architecture
For the simulated environment, we use a customized two-
player-variation of the OpenAI gym (Brockman et al.
(2016)) cart-pole problem. Here, two agents balance a
pole, which is hinged to a movable cart, by concurrently
applying forces to the cart's base. The system dynamics
Fig. 1 shows the resulting state-value estimates V (x) of
both agents at different training episodes, where each
data point is averaged over sk and θk for reasons of
presentability. After 1000 episodes of training, agent 1
expects the highest return along the pole angle of 0 rad.
The lowest state-values are estimated for θ close to the
(a) agent 1, 1000 episodes
(b) agent 1, 2000 episodes
(c) agent 2, 1000 episodes
(d) agent 2, 2000 episodes
Fig. 1. State-value estimates of the agents at different training stages. Each data point is averaged over sk and θk.
controlled yet. However,
in later training stages (2000
episodes), the agents develop understanding concerning
coordination possibilities (guided by IER) and their rel-
ative contribution (thanks to IQL) allowing the agents
to move the cart to a desired state without terminating
the episode yielding much higher rewards. The decreased
steepness of the state-value gradients when comparing
V (x) after 1000 and 2000 episodes is a result of the agent's
increased control capabilities allowing to transition from a
state with low rewards to a state associated with higher
rewards.
It is noticeable that without the mechanisms proposed in
Section 3, the agents were not able to learn to stabilize the
cart-pole at all with the given parametrization. Thanks
to IER and IQL and an appropriate focus on recent
experiences due to TER, the agents were successful at
adapting to each other. The agents also learned to perform
complex trajectories, which included deflecting the pole
close to the terminal positions and angles (cf. Fig. 2). Thus,
it is possible for them to flexibly control the cart-pole even
in difficult situations.
5. CONCLUSION
In this paper, new mechanisms have been proposed in
order to account for challenges arising in deep Multi-
Agent Reinforcement Learning problems with restricted
information. Two novel extensions to experience replay
were presented. First, TER allows the sampling process to
properly reflect the fact that recent experiences carry more
information regarding the current control laws of cooper-
ation partners and are thus better suited to counteract
the non-stationarity compared to outdated experiences.
Second, artificial experiences denoted as IER complement
the experience replay memory. In the early training stage,
alter-exploration problems are reduced due to simulated
transitions in which the agents interact separately with
the environment. Later, coordination is induced to ex-
haust the cooperation potential between agents as adap-
tation becomes feasible. Finally, these experience replay
enhancements are supplemented by a mechanism termed
IQL. Here, the relative contribution of the agent towards
the observed outcome is accounted for by means of an
impact factor which adapts an agent's learning rate. Our
algorithm was evaluated on a simulated cart-pole-problem,
where two agents successfully learned to cooperate.
(a)
(b)
Fig. 2. Cart-pole position and pole angle for an example
initialization using the learned controllers.
terminal pole angles θ = ±0.21 rad. Analogously, after
1000 episodes, agent 2 evaluates states close to the desired
target position s∗ = 0 m as most beneficial. This maximum
state-value drops abruptly with slight deviation from s∗.
Fig. 1b and 1d show the agents' estimated state-values
after 2000 episodes of training. For both agents, the
expected returns have generally increased compared to
the previous training stage. Because they successfully
learned how to jointly balance the pole without moving
the cart outside the boundaries, the available time steps to
accumulate rewards is increased. Additionally, both agents
learned to appropriately reduce V (x) close to all terminal
states.
Example trajectories of the cart position sk and pole angle
θk resulting from the trained control law are depicted in
Fig. 2. In Fig. 2a, the boundaries for the highest 5 point
reward range and the lower 1 point reward range for the
seconds agent's position control are depicted in red and
orange and the red lines in Fig. 2b mark the terminal
conditions. It can be seen that the agents are capable of
moving the cart from the initial position to the desired
target while holding the pendulum upright.
4.3 Discussion
After 1000 training episodes, the state-value estimations
are predominantly dependent on the state dimension as-
sociated with the individual preferences as this yields the
highest rewards while the cart-pole cannot be successfully
REFERENCES
Brockman, G., Cheung, V., Pettersson, L., Schneider, J.,
Schulman, J., Tang, J., and Zaremba, W. (2016). Openai
gym. Arxiv.org, 1606.01540.
Bu¸soniu, L., Babuska, R., and de Schutter, B. (2010).
Multi-agent reinforcement learning: An overview.
In
Innovations in Multi-Agent Systems and Applications,
183 -- 221. Springer, Berlin, Heidelberg.
Chang, Y.h., Ho, T., and Kaelbling, L.P. (2004). All
learning is local: Multi-agent learning in global reward
games.
In Advances in neural information processing
systems, 807 -- 814.
Foerster, J., Nardelli, N., Farquhar, G., Afouras, T., Torr,
P.H.S., Kohli, P., and Whiteson, S. (2017). Stabilising
experience replay for deep multi-agent reinforcement
learning. In Proceedings of the 34th ICML, volume 70
of PMLR, 1146 -- 1155.
Foerster, J.N., Farquhar, G., Afouras, T., Nardelli, N.,
and Whiteson, S. (2018). Counterfactual multi-agent
policy gradients. In 32nd AAAI Conference on Artificial
Intelligence.
Gu, S., Lillicrap, T., Sutskever, I., and Levine, S. (2016).
Continuous deep q-learning with model-based accelera-
tion. In ICML, 2829 -- 2838.
Gupta, J.K., Egorov, M., and Kochenderfer, M. (2017).
Cooperative multi-agent control using deep reinforce-
ment learning. In Autonomous Agents and Multiagent
Systems, volume 10642 of Lecture Notes in Computer
Science, 66 -- 83. Springer Int. Publishing, Cham.
Inga, J., Eitel, M., Flad, M., and Hohmann, S. (2018).
Evaluating human behavior in manual and shared con-
trol via inverse optimization.
In 2018 IEEE Interna-
tional Conference on Systems, Man, and Cybernetics,
2699 -- 2704.
Kopf, F., Ebbert, S., Flad, M., and Hohmann, S. (2018).
Adaptive dynamic programming for cooperative control
with incomplete information.
In 2018 IEEE Interna-
tional Conference on Systems, Man and Cybernetics.
Kopf, F., Inga, J., Rothfuss, S., Flad, M., and Hohmann,
Inverse reinforcement learning for iden-
IFAC-
S. (2017).
tification in linear-quadratic dynamic games.
PapersOnLine, 50(1), 14902 -- 14908.
Kopf, F., Nitsch, A., Flad, M., and Hohmann, S. (2019).
Partner approximating learners
(pal): Simulation-
accelerated learning with explicit partner modeling in
multi-agent domains. arXiv e-prints, arXiv:1909.03868.
Lowe, R., Wu, Y., Tamar, A., Harb, J., Abbeel, O.P., and
Mordatch, I. (2017). Multi-agent actor-critic for mixed
cooperative-competitive environments. In Advances in
Neural Information Processing Systems, 6379 -- 6390.
Matignon, L., Laurent, G.J., and Le Fort-Piat, N. (2007).
Hysteretic q-learning: An algorithm for decentralized re-
inforcement learning in cooperative multi-agent teams.
In 2007 IEEE/RSJ International Conference on Intelli-
gent Robots and Systems, 64 -- 69.
Matignon, L., Laurent, G.J., and Le Fort-Piat, N.
(2012). Independent reinforcement learners in cooper-
ative markov games: A survey regarding coordination
problems. The Knowledge Engineering Review, 27(1),
1 -- 31.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Ve-
ness, J., Bellemare, M.G., Graves, A., Riedmiller, M.,
Fidjeland, A.K., Ostrovski, G., Petersen, S., Beattie, C.,
Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wier-
stra, D., Legg, S., and Hassabis, D. (2015). Human-level
control through deep reinforcement learning. Nature,
518(7540), 529 -- 533.
Omidshafiei, S., Pazis, J., Amato, C., How, J.P., and Vian,
J. (2017). Deep decentralized multi-task multi-agent
reinforcement learning under partial observability.
In
Proceedings of the 34th ICML, volume 70 of PMLR,
2681 -- 2690.
Palmer, G., Tuyls, K., Bloembergen, D., and Savani,
R. (2018).
Lenient multi-agent deep reinforcement
learning. In Proceedings of the 17th Int.Conference on
Autonomous Agents and MultiAgent Systems, 443 -- 451.
Schaul, T., Quan, J., Antonoglou, I., and Silver, D. (2016).
In Proceedings of the
Prioritized experience replay.
International Conference on Learning Representations.
Scott, S.H. (2004). Optimal feedback control and the
neural basis of volitional motor control. Nature reviews.
Neuroscience, 5(7), 532 -- 546.
Silver, D., Lever, G., Heess, N., Degris, T., Wierstra, D.,
and Riedmiller, M. (2014). Deterministic policy gradient
algorithms. 31st ICML.
Vamvoudakis, K.G. and Lewis, F.L. (2011). Multi-player
non-zero-sum games: Online adaptive learning solution
of coupled hamilton -- jacobi equations. Automatica,
47(8), 1556 -- 1569.
Watkins, C.J.C.H. (1989). Learning from Delayed Re-
wards. Ph.D. thesis, King's College, Cambridge, UK.
Appendix A. NETWORK ARCHITECTURE AND
TRAINING PARAMETER
Table A.1. Hyperparameters of the network architecture
hyperparameter
number of hidden layers
neurons per hidden layer
dropout probability
activation f. hidden layer
initialization of hidden layer
activation f. output layer A/C
initial weights all layers A/C
optimizer
value
3
64
0.2
LeakyReLU,
α = 0.01
Xavier uniform,
∼ U (−0.5, 0.5)
linear
∼ U (−1, 1)
Adam, β1 = 0.9,
β2 = 0.999, no gradient
clipping, decay, fuzz
factor or AMSGrad
Huber loss
error metric
target network update frequency m 4000
Table A.2. Hyperparameters of the algorithm
value
0.999
0
1 × 105
hyperparameter
discount factor γ
ξtemp
ERM size M
macro-batch size B 256
mini-batch size t
α learning rate
σ learning rate
β learning rate
λhigh
λlow
exploration
80
5 × 10−4
2 × 10−4
5 × 10−5
0.8
0.2
εmin = 0.01, decay rate = 0.999
|
1901.10421 | 1 | 1901 | 2019-01-01T17:22:43 | A structured approach for the implementation of distributed manufacturing simulation | [
"cs.MA"
] | Manufacturing has been changing from a mainly inhouse effort to a distributed style in order to meet new challenges owing to globalization of markets and worldwide competition. Distributed simulation provides an attractive solution to construct cross enterprise simulations to evaluate the viability of the proposed distributed manufacturing enterprises. However, due to its complexity and high cost distributed simulation failed to gain a wide acceptance from industrial users. The main objective of this paper is to address these issues and present a new structured approach to implement distributed simulation with cost effective and easy to implementable tools. A simplified approach for model partitioning for distributed simulation is also included in the proposed approach. The implementation of distributed manufacturing simulation is illustrated with Arena, Microsoft Message Queue (MSMQ) and Visual Basic for Applications (VBA). | cs.MA | cs | A STRUCTURED APPROACH FOR THE IMPLEMENTATION OF
DISTRIBUTED MANUFACTURING SIMULATION
Sameh M. Saad, Terence Perera and Ruwan Wickramarachchi
School of Engineering
Sheffield Hallam University
Sheffield, S1 1WB, United Kingdom
KEYWORDS
Distributed manufacturing
simulation software, IDEF0, Middleware
ABSTRACT
simulation, commercial
solution
Manufacturing has been changing from a mainly in-
house effort to a distributed style in order to meet new
challenges owing to globalization of markets and world-
wide competition. Distributed simulation provides an
attractive
to construct cross enterprise
simulations to evaluate the viability of the proposed
distributed manufacturing enterprises. However, due to
its complexity and high cost distributed simulation
failed to gain a wide acceptance from industrial users.
The main objective of this paper is to address these
issues and present a new structured approach to
implement distributed simulation with cost effective and
easy to implementable tools. A simplified approach for
model partitioning for distributed simulation is also
included in the proposed approach. The implementation
of distributed manufacturing simulation is illustrated
with Arena, Microsoft Message Queue (MSMQ) and
Visual Basic for Applications (VBA).
INTRODUCTION
In today's highly competitive world, manufacturing
enterprises are confronted with growing competition,
the evolution of new markets, more and more
sophisticated consumer demand, and
increasingly
complex global political and economic scenarios. In
order to lower the costs, increase profits, reduce product
development times, enhance products, and react to
environmental changes more positively manufacturing
enterprises are moving towards open architectures for
integrating their activities with those of their suppliers,
customers and partners. In manufacturing, companies
may form strategic partnerships for outsourcing some of
their operational activities, share resources or joint
development of products and services etc., leading to
formation of virtual manufacturing enterprises which
operate in distributed manufacturing environment. To
facilitate
the creation of virtual manufacturing
enterprises, potential partners must be quickly able to
evaluate whether it will be profitable for them to
participate
the proposed enterprise. Simulation
provides a capability to conduct experiments rapidly to
in
predict and evaluate the results of manufacturing
decisions (McLean and Leong, 2001).
As Law and McComas (1998) noted manufacturing
is one of the largest application areas of simulation, with
the first uses dating back to at least early 1960s.
However, traditional sequential simulation alone may
not sufficient to simulate highly complex Distributed
Manufacturing Enterprises (DME). In such situations,
distributed simulation provides a promising alternative
to construct cross enterprise simulations. The use of
distributed simulation allows each partner to hide any
proprietary information in the implementation of the
individual simulation, simulate multiple manufacturing
systems at different degrees of abstraction levels, link
simulation models built using different simulation
software, to take advantage of additional computing
power, simultaneous access to executing simulation
models for users in different locations, reuse of existing
simulation modes with
little modifications etc.
(Venkateswaran et al., 2001; McLean and Riddick,
2000; Gan et al., 2000; Taylor et al., 2001). However,
Peng and Chen (1996) noted that as a technique, parallel
and distributed simulation is not very successful in
manufacturing. Most of the distributed manufacturing
simulations developed were implemented with either
simulation languages or general purpose programming
languages such as C++ and Java. This calls expertise not
only in distributed simulation but also in programming
too. Moreover, general business community is not very
receptive towards distributed simulation due to its
complexity, long development times, involvement of
step learning curves, high costs etc.
Another important issue needs to be addressed
when designing a distributed simulation is partitioning
of the simulation model into sub-models or logical
processes (LPs). Efficiency and effectiveness of a
distributed simulation system depends on partitioning of
the system. Some of the existing approaches require
executing of the whole model sequentially in order to
collect data before partitioning and mapping carried out
based on data collected. However, a simulation is
executed in distributed manner because its inability to
run sequentially due to size, complexity, requirements
for more computing resources, or need
to run
geographically distributed manner etc. This creates a
dilemma for users especially in business organizations,
to design parallel and distributed
who
intend
simulations.
approach
implement
It
(DMS).
The objective of this paper is to present a new
distributed
to
simplified
manufacturing simulation
includes a
simplified approach to model partitioning and mapping,
and simulation model development processes for DMS.
Instead of implementing the distributed simulation with
programming languages, we are proposing to develop
the system using commercial simulation software.
BACKGROUND
and
simulation
Distributed
technologies with
effective platform
combines
distributed
computing
traditional sequential
simulation techniques. In recent years, popularity of
distributed computing applications increased due to
proliferation of inexpensive and powerful workstations,
improvements in networking technologies, low cost
equipment and incremental scalability. Hence, the use
of network of workstations has been evolving into a
popular
for distributed
simulation. However,
low communication speeds,
shortage of network bandwidth and the ever increasing
demand for network resources may result slowing down
the execution speed of the distributed simulation model.
Although the networked workstations are slower than
dedicated machines, they may be fast enough and may
require much less specialist expertise to put them to use
with a fraction of a cost of the price needed for a
dedicated parallel processing computer (Cassel and
Pidd, 2001).
the
century,
Throughout
the world
of
manufacturing has changed from a mainly in-house
effort to a distributed style of manufacturing. As the
term distributed manufacturing implies, DMEs which
also known as virtual manufacturing enterprises are
ephemeral organizations in which several companies
collaborate to produce a single product or product line
(Venkateswaran et al., 2001). Participating in this type
of collaboration allow partner organizations to use their
knowledge, resources and in particular manufacturing
expertise
take advantage of new business
opportunities and/or gain a competitive advantage that
are on a larger scale than an individual partner could
handle alone.
PROPOSED APPROACH
Modeling and Model Partitioning for Distributed
Manufacturing Simulation
to
and
requirements
and perhaps
problems,
alternative
solutions for many systems especially complex and
large ones, it is desirable to build a conceptual model
before transforming it into a computer simulation
model. Conceptual model is a simulation developer's
way of translating modeling requirements (ie. What to
be represented by simulation) into detailed design
framework (ie. How it is to be done), from which the
software that will make up the simulation can be built
(Pace, 1999). Furthermore conceptual model is the
ultimate expression of the system functionality and
should be the basis for testing and verification and
validation procedures (Haddix, 2001).
analyzing
A conceptual model developed with an appropriate
modeling approach and modeling
tool facilitates
partitioning of the DME model into LPs. Modeling
approaches specify the way models are to be developed
while Modeling tools provide a standard means of
describing
systems. Hierarchical
modeling approach was selected since it provides a way
of managing
systems by
considering them as a collection of sub-systems (Kiran,
1998). In a distributed simulation system these sub
systems are represented by simulation models that are
independently created, modified and saved. Pidd and
Castro (1998) also noted that many large systems are
inherently hierarchical.
scale complex
IDEF0 was chosen as the modeling technique for
the proposed approach, and it has been widely used in
industry due to its user-friendliness, computer support,
rigor and conciseness, and well documented rules and
procedures (Pandya, 1995; Kateel et al., 1996). Number
of authors including Cheng-Leong et al. (1999), Cheng-
Leong (1999), Whiteman et al. (1997), Rensburg and
Zwemstra (1995) have highlighted the usefulness of
IDEF0 as a model
in
simulation. Another benefit of using IDFE0 with
commercial simulation software is that IDEF0 structure
of the model can easily be transformed into simulation
model. Figure 1 shows a part of simulation model
developed by Arena for an IDEF0 model.
representation
technique
large
Process 1
A21
Process 2
A22
Process 1
Assign 1
Process 2
0
0
Station 1
Route 1
The degree to which the simulation results are able
to characterize the system under study is directly related
to the degree the simulation model characterizes the
system (Luna, 1993). In order to understand the
Figure 1: IDEF0 diagram and Arena simulation model
With hierarchical modeling approach and IDEF0
technique, LPs that can function independently could be
identified based on
interactions between different
sections. In the IDEF0 model these interactions are
represented by number of lines between boxes that
represent different sections of the enterprise.
Once the LPs are identified, they could be validated
to make sure that LPs represent individual entities of the
enterprise, and the entire enterprise when considered
together. The validated LPs could be assigned (mapped)
into workstations
in a computer network before
converting them into computer simulation models and
execute as a distributed simulation. Figure 2 shows the
proposed approach for modeling, model partitioning and
mapping for the distributed manufacturing simulation.
Identify an appropriate modeling technique
Transform LPs into simulation models
Figure 2: Proposed approach for modeling, model
partitioning and mapping
in
order
algorithms
The main difference between existing approaches
and the proposed approach is the stage of partitioning
carried out in the simulation methodology. According to
most of the current approaches partitioning is done only
after the system is converted into a computer program
the
using
to minimize
the
communication overheads and optimize
load
balance. To
the distributed
simulation
development process it is proposed to partition the
conceptual model into LPs and assign them into
workstations before transforming LPs into computer
simulation models. It is also assumed that networked
workstations are freely available to assign LPs and only
one LP is mapped into a workstation.
simplify
Develop the Conceptual model
Validate the conceptual mode
Partition the conceptual model into LPs
Validate LPs
Mapping
Development of the Distributed Manufacturing
Simulation
the
pose
Distributed
Simulations
unique
synchronization constraints due to their underlying
sense of time. When the simulation state can be
simultaneously changed by different processes, actions
by one process can affect actions of another (Nicol,
1993). In order to make sure that each LP processes
arriving messages in their timestamped order and not in
their real time arriving order, individual simulation
models needed to be synchronized. This requirement is
referred to as local causality constraint (Fujimoto,
1990). Optimistic protocols are implemented by saving
simulation state at different points of time and rolling
back to a previous time point if local causality
constraint is violated. If a programming language is
used to develop the simulation, then state saving
mechanism can be integrated into the distributed
simulation engine itself. Since commercial simulation
packages generally do not allow saving simulation state
at different time points and rolling back to previous time
points, it was decided to employ conservative protocol
to synchronize the distributed manufacturing simulation.
Conservative approaches strictly
local
causality constraint and guarantee that each model will
only process events in non-decreasing timestamp order.
Determining a value for lookahead is one of the most
important and difficult aspect of conservative protocol.
However, it was assumed that minimum-processing
times (which can be used as lookahead values) for LPs
can be calculated. An approximate synchronization
mechanism,
distributed
manufacturing applications has been proposed by Saad
et al. (2003).
especially
suitable
impose
In order to synchronize and pass parameters,
simulation models need to communicate with each
other. Communication methods provided by operating
systems often require complex programming. In a
distributed simulation, middleware provides simple and
reliable solution for this problem. Middleware is a class
of software designed to help manage the complexity and
heterogeneity inherent in distributed system. It contains
a set of enabling services which allow multiple
processes running on one or more computers to interact
across a network. Analysis of past literature reveals
number
distributed
manufacturing systems and supply chains using tools
such as HLA, CORBA and GRIDS (see Venkateswaran
et al., 2001, Taylor et al., 2001; Gan et al., 2000;
McLean and Riddick, 2000). For
the proposed
approach, Microsoft Message Queue (MSMQ), a
Message Oriented Middleware was selected to link
simulation models.
attempts
simulate
As MSMQ is integrated into newer versions of
Windows operating systems and available as an
additional component for Windows NT, 98 and 98, it
of
to
for
in
provides a cost effective solution for message passing.
MSMQ interacts with simulation model through an
application program interface (API). Arena simulation
software was used
this study as commercial
simulation software to demonstrate the implementation.
However, other commercial simulation software such as
Automod, Promodel, Witness etc. can also be used for
this purpose. Both Arena and MSMQ support Visual
Basic for Applications (VBA) and C++. Since,
programming of Arena with VBA
is more
straightforward than with C++, it was decided to use
VBA
the API. VBA also offers a
programming environment similar to popular Visual
Basic programming language.
to develop
API developed for MSMQ could send messages
containing parameters obtained from simulation model
to a queue in the same computer or directly to another
remote computer. API that resides in the remote
computer extracts these messages from the queue and
passes the parameters to the simulation model (Figure
3).
Arena model
Arena model
API
API
pq
sq
pq
sq
Network
Figure 3: MSMQ, API and Arena
ILLUSTRATION
In order to illustrate the development of distributed
simulation with Arena, MSMQ and VBA, following
brief case study is presented.
A, B and C firms are proposing to form a
distributed manufacturing enterprise (Figure 4)
to
produce a new product called XYZ. Firm A is to
produce and process parts X and Z. Part Z is sent to
Firm C and part X is sent to Firm B which posses an
expensive processing unit for further processing. Firm B
is also to produce part Y and assemble Parts X and Y
together to form component XY which is then sent to
Firm C for further processing and final assembly. At
Firm C, component XY and part Z are to be further
processed and assembled together to produce product
XYZ. In addition to processing of parts X, Y, Z,
component XY and product XYZ, three firms also
produce their own products independently. Parts are to
be passed in batches of 1000s and transfer time from
one firm to another firm was assumed as 10 hours.
Before committing on the DME, firms want to evaluate
the feasibility in terms of capacity utilization and how
the proposed venture affects their existing operations.
As firms are reluctant to pass information of their
processes to other firms, it was agreed to develop 3
models separately and run them in a distributed
simulation environment.
Firm C
Firm B
Firm A
Figure 4: A model for distributed manufacturing
Three Arena simulation models were independently
developed, verified and validated for firms A, B and C.
Entities created within models were used instead of
inputs from other models for B and C. Then B and C
were modified by replacing the 'Create module' with a
'Create block' and adding a 'VBA block' just before the
'Dispose module' (Figures 5 and 6). 'Create block' can
be used to release entities into the model which created
by API of the model to represent output from A and/ or
B. Two MSMQ queues were created
in each
workstation, one to accept inputs (pq) and the other to
synchronize (sq) the distributed simulation. Once batch
of 1000 units were processed at A or B, the batch goes
through a VBA block and API written in VBA sends a
message to destination model (Figure 6). When a
message is reached its destination queue, it is processed
automatically with built-in 'qevent' event. Once a
message comes to 'pq', 'qevent' creates an entity and
schedules to release it after 'transfer time' at 'Create
block'. The 'Separate module' adds additional 999 units
to make a batch of 1000 which was passed as output
from previous model (Figure 5). At model C, output
from A and B can be identified by message label.
Figures 7 and 8 show sample code written for VBA
block and 'qevent' respectively.
Create
Separate 1
0
Original
0
Duplicate
Figure 5: Adding output received from other models as
input
Batch 1
VBA
1
Dispose 1
0
0
Figure 6: Passing output to destination model VBA
block
Private Sub VBA_Block_1_Fire()
Dim qinfo As MSMQQueueInfo
Set qinfo = New MSMQQueueInfo
qinfo.FormatName = "DIRECT = OS:ENG-4l30-10
\private$\mbq"
Dim qQueue As MSMQQueue
Set qQueue = qinfo.Open(MQ_SEND_ACCESS,
MQ_DENY_NONE)
Dim qMsg As MSMQMessage
Set qMsg = New MSMQMessage
qMsg.Label = "A"
qMsg.Body = "1000"
qMsg.Send qQueue
qQueue.Close
End Sub
Figure 7: Sample code of VBA block
Sub qEvent_Arrived(ByVal Queue As Object, ByVal
cursor As Long)
Dim vEntityIndex As Long
Dim vPictureIndex As Long
Dim aQueue As MSMQQueue
Set aQueue = Queue
Dim qMsg As MSMQMessage
Set qMsg = aQueue.Receive(, , , 0)
vPictureIndex =
ThisDocument.Model.SIMAN.SymbolNumber
("Picture.Package")
vEntityIndex = ThisDocument.Model.SIMAN
.EntityCreate
Call ThisDocument.Model.SIMAN.EntitySetPicture
(vEntityIndex, vPictureIndex)
If qMsg.Label = "A" Then
Call ThisDocument.Model.SIMAN
.EntitySendToBlockLabel(vEntityIndex
, 10, "CreateBlockA")
Else
Call ThisDocument.Model.SIMAN
.EntitySendToBlockLabel(vEntityIndex,
10, "CreateBlockB")
End If
aQueue.EnableNotification qEvent
End Sub
Figure 8: Sample code to create and schedule an entity
to represent input (of model C)
DISCUSSION
This paper presented a simplified approach to
implement a distributed manufacturing simulation.
Although a distributed manufacturing application was
used to illustrate the implementation, the proposed
approach may be able to use in other application areas
too. The main benefit of this research is the simplified
approach employed when developing the distributed
simulation models using commercial
simulation
software and, connecting and
in
them
running
distributed environment with MSMQ and VBA. Not
only Arena, but also other simulation software packages
such as Promodel, Automod and Witness can be used
to implement the simulation. Furthermore, simulation
models developed with different simulation software
can be connected together and run as a distributed
simulation as long as they support either VBA or C++.
Cost involved can be kept low as no additional costs
involved with middleware and application program
interface (API), provided workstations are running on a
windows operating system. The proposed approach also
encourages reusability of existing simulation models.
Existing simulation models developed for traditional
sequential simulation require only minor modifications
to adopt for a distributed simulation.
It is expected that this simplified approach may
address criticisms made against distributed simulation
because of its complexity to develop and implement,
higher costs involved, need for more expertise etc.
Working of the simulation model can be animated easily
as commercial simulation software
for
implementation. Animation may play a very effective
role in convincing the benefits of simulation to non
simulation users such as managers, workers etc.
is used
highly
systems
complex
in performance of
Main shortcoming of the proposed approach is the
sacrifice made
the distributed
simulation mainly in terms of speedup. It may also not
be feasible to employ the proposed approach to
implement
as
telecommunications systems, computer networks, logic
circuits etc. However, applications which are not
executed in distributed manner only to gain speedups,
and applications that specifically require executing in
distributed simulation environment are ideally fit for the
approach we presented in this paper.
CONCLUSIONS
such
The proposed approach addressed some of the
criticisms leveled against distributed simulation with
cost effective and simplified implementation approach
for distributed manufacturing simulation.
Performance comparisons between distributed
simulation implemented using proposed approach and
conventional approaches may provide an opportunity to
fully understand the benefits and the shortcoming of our
work. Unlike sequential simulation, output from a
distributed simulation can be obtained for individual
models as well as for
the whole system under
investigation. Since outcome of the simulation effort
depends on the output of the simulation, it may worth
individual models and for
investigating strategies to identify and generate output
locally at
the entire
distributed simulation system.
REFERENCES
Cassel, R.A. and M. Pidd. 2001. "Distributed discrete event
simulation using the three-phase approach and Java."
Simulation: Practice and Theory 8, No.8, 491-507.
Cheng-Leong, A., K.L. Phengm and G.R.K. Leng. 1999.
"IDEF*: a comprehensive modeling methodology for the
development of manufacturing enterprise systems."
International Journal of Production Research 37, No.17,
3839-3859.
Cheng-Leong, A. 1999. "Enactment of IDEF models."
International Journal of Production Research 37, No.15,
3383-3397.
Fujimoto, R.M. 1990. "Parallel discrete event simulation."
Communications of the ACM 33, No.10, 30-53.
Gan, B.P., L. Liu, S. Jain, S.J. Turner, W. Cai, and W. Hsu.
2000. "Distributed supply chain simulation across
enterprise boundaries." Proceedings of the 2000 Winter
Simulation Conference 1245-1251.
revisited: A
for modeling &
the 2001 Simulation
Haddix, F. 2001. "Conceptual modeling
developmental model
simulation." Proceedings of
Interoperability Workshops
approach
Kateel, G., M. Kamath and D. Pratt. 1996. "An overview of
CIM enterprise modeling methodologies." Proceedings of
the 1996 Winter Simulation Conference 1000-1007.
Kiran, A.S. 1998. "Hierarchical modeling: A simulation based
application." Proceedings of the 1998 IEEE International
Conference on Systems, Man and Cybernetics 3079-3089.
Law, A.M. and M.G. McComas. 1998. "Simulation of
manufacturing systems." Proceedings of the 1998 Winter
Simulation Conference 49-52.
Luna, J.J. 1993. "Hierarchical relations in simulation models."
Proceedings of the 1993 Winter Simulation Conference
132-137.
McLean, C. and F. Riddick. 2000. "The IMS mission
architecture for distributed manufacturing simulation."
Proceedings of the 2000 Winter Simulation Conference,
1539-1548.
McLean, C. and S. Leong. 2001. "The role of simulation in
International Working
strategic manufacturing."
Conference on Strategic Manufacturing 239-250.
Nicol, D.M. 1993. "The cost of conservative synchronization
in parallel discrete event simulations." Journal of the
Association for Computing Machinery 40, No.2, 304-333.
Pace, D.K. 1999. "Development and documentation of a
simulation conceptual model." Proceedings of the 1999
Simulation Interoperability Workshops.
Pandya, K.V. 1995. "Review of modeling techniques and tools
for decision making in manufacturing management." IEE
Proceedings of Science, Measures and Technology.
Peng, C. and F.F. Chen. 1996. "Parallel discrete event
simulation of manufacturing systems: A
technology
survey." Computers and Industrial Engineering 31,
No.1/2, 327-330.
Pidd, M. and B.R. Castro. 1998. "Hierarchical modular
modeling in discrete simulation." Proceedings of the1998
Winter Simulation Conference, 383-390.
Resenburg, A.V. and N. Zwemstra. 2002. "Implementing
IDEF techniques as simulation modeling specifications."
Computers and Industrial Engineering 29, No.1-4, 467-
471.
Saad, S.M., T. Perera, and R. Wickramarachchi. 2003.
"Simulation of distributed manufacturing enterprises: A
new approach." To be published in Proceedings of the
2003 Winter Simulation Conference.
Taylor, S.J.E., R. Sudra, T. Janahan, G. Tan and J Ladbrook.
2001. "Towards COTS distributed simulation using
GRIDS." Proceedings of the 2001 Winter Simulation
Conference 1372-1379.
Venkateswaran, J., M.Y.K. Jafferali and Y Son. 2001.
"Distributed simulation: An enabling technology for the
evaluation of virtual enterprises." Proceedings of the 2001
Winter Simulation Conference 856-862.
Whitman, L., B. Huff and A Presley. 1997. "Structured
models and dynamic systems analysis: The integration of
the IDEF0/IDEF3 modeling methods and discrete event
simulation." Proceedings of the 1997 Winter Simulation
Conference.
control,
systems
integration,
AUTHOR BIOGRAPHIES
SAMEH M. SAAD (BSc, MSc, PhD, CEng, MIEE,
ILTM) is a Reader in Advanced Manufacturing Systems
and Enterprise modeling and management and
Postgraduate Course Leader at
the Systems and
Enterprise Engineering Division, one of the three
divisions in the School of Engineering, Sheffield
Hallam University, United Kingdom. Dr Saad's
research interests revolve around aspects of design and
analysis of manufacturing systems, production planning
and
reconfigurable
manufacturing systems, manufacturing responsiveness,
enterprise modeling and management and next
generation of manufacturing systems. He has published
over 55 articles in various national and international
academic journals and conferences. His contact email
address is <[email protected]>
TERRENCE PERERA (BSc, PhD) is Professor and
Head of Enterprise and Systems Engineering at School
of Engineering, Sheffield Hallam University. He also
leads the Systems Modeling and Integration Research
Group. His current research interests include the
implementation, integration and practice of virtual
modeling tools within all industrial sectors. His email
address is <[email protected]>
RUWAN WICKRAMARACHCHI is a PhD student
at Sheffield Hallam University, United kingdom. He
received his MPil and BSc degrees from University of
Cambridge, United Kingdom and University of
Kelaniya, Sri Lanka respectively. His main research
interest focused on distributed enterprise simulation
distributed manufacturing
with
applications.
by
<[email protected]>
contacted
emphasis
can
He
on
be
|
1604.03747 | 1 | 1604 | 2016-04-13T12:56:27 | Behavior of Self-Motivated Agents in Complex Networks | [
"cs.MA",
"cs.GT",
"physics.soc-ph"
] | Traditional evolutionary game theory describes how certain strategy spreads throughout the system where individual player imitates the most successful strategy among its neighborhood. Accordingly, player doesn't have own authority to change their state. However in the human society, peoples do not just follow strategies of other people, they choose their own strategy. In order to see the decision of each agent in timely basis and differentiate between network structures, we conducted multi-agent based modeling and simulation. In this paper, agent can decide its own strategy by payoff comparison and we name this agent as "Self-motivated agent". To explain the behavior of self-motivated agent, prisoner's dilemma game with cooperator, defector, loner and punisher are considered as an illustrative example. We performed simulation by differentiating participation rate, mutation rate and the degree of network, and found the special coexisting conditions. | cs.MA | cs | BEHAVIOR OF SELF-MOTIVATED AGENTS IN COMPLEX NETWORKS
Sundong Kim
Dept. of Industrial and Systems Engineering
Jin-Jae Lee
Dept. of Physics
KAIST, Daejeon, South Korea
KAIST, Daejeon, South Korea
ABSTRACT
Traditional evolutionary game theory describes how certain strategy spreads throughout the system where
individual player imitates the most successful strategy among its neighborhood. Accordingly, player
doesn't have own authority to change their state. However in the human society, peoples do not just fol-
low strategies of other people, they choose their own strategy. In order to see the decision of each agent in
timely basis and differentiate between network structures, we conducted multi-agent based modeling and
simulation. In this paper, agent can decide its own strategy by payoff comparison and we name this agent
as "Self-motivated agent". To explain the behavior of self-motivated agent, prisoner's dilemma game
with cooperator, defector, loner and punisher are considered as an illustrative example. We performed
simulation by differentiating participation rate, mutation rate and the degree of network, and found the
special coexisting conditions.
1
INTRODUCTION
1.1 Overview
Traditional evolutionary game describes how successful strategy spreads in population(Jörgen M.
Weibull., 1995). Depending on the payoff structure, agent follows the strategy of most successful neigh-
bor. In that way the whole system adopt a certain strategy. But human society is more complicated that it
is not explained by only adopting strategy. Still, there are many peoples trying to follow how other people
live, but not everyone makes a decision by mimicking their role model. For example, by only using
adopting strategy, we can't explain how great politician grew up from bed neighborhood. Of course there
is more probability that he/she will be up to he/she's neck in crime. But he/she might have dreamed of
changing the world, or had more advantage to get a good score while he/she is in bed neighborhood. Pre-
vious researches hadn't focused on this characteristic of agent so that it is not a sufficient model to ex-
plain any autonomous society.
In this paper, we suggest the decision process of "Self-motivated agent". Self-motivated agent can
guess which types of strategies are available regardless of their neighborhood, and it can voluntarily
choose its strategy by comparing payoff and change its state by itself. Participants play prisoner's dilem-
ma game with their neighborhood and find the maximum payoff that they can get, and finally change to
relevant state.
In our model, there are four types of player - cooperator, defector, loner and punisher. To explain so-
ciety more accurately, we added loner and punisher into traditional prisoner's dilemma game. Loner, an
agent with indifferent mind, gets fixed payoff(Dawes, 1980), and punisher is an agent introduced to pe-
nalize defector. (Fehr & Gächter, 2002; Rockenbach & Milinski, 2006). And we also considered other
three factors, participation rate and mutation rate, and network structure.
Since every agent makes their decision at different time, we can say that only some of the players join
the game on each time. To model this factor, we suggest participation rate to set how many players will
participate in a time period. Surprisingly, this probability not only controls the speed of dynamics, but al-
so becomes the important factor of deciding the overall behavior of agents. And mutation is also adopted
Kim and Lee
in our model (Glenn W. Rowe, 1985). It corresponds to mutation in the DNA, or in cultural evolution it
describes individuals experimenting with new behavior. Sometimes human-being also makes decision by
throwing dice, which means they leave it to chance. Mutation is happened regardless of whether the agent
is a participant or not.
Complex network structure is also considered (S.Boccaletti et al., 2005). Six different networks are
used in our paper, and we found different coexisting condition by changing network degree and structure.
1.2
Previous Research
As a famous example of game theory, prisoner's dilemma is a kind of non-zero-sum game where two
people participate. If both players cooperate, they will get the largest benefit together. However, if both
players choose selfish strategy to maximize personal gain, then they can't get any profit and hence it turns
out to be the dilemma. Even there are a lot of people, not two people with two-strategy of cooperators and
defectors, we know that defectors prevail through replicator dynamics (Hofbauer & Sigmund, 1998). This
result depicts our sociology (Dawes, 1980).
From the point of view of evolutionary theory however, there is a problem that defectors prevail.
Robert Axelord and William. D. Hamilton for the first time to solve these problems by using this game
theory (Axelrod & Hamilton, 1981). This is just a matter of evolution theory and involve a lot of social
issues(Fehr & Fischbacher, 2003; Wedekind & Milinski, 2000). In this way, the effort what you looking
for that cooperator can emerge and persists through game theory is said Evolutionary Game Theory.
In 1992, Martin A. Nowak and Robert M. May introduced a spatial structure (Nowak & May, 1992).
Individuals are located in the normal grid, and collect the aggregate rewards of interaction with neighbors
to mimic the strategy of the most successful neighbor. Studies on attributes about spatial structure in gen-
eral two-strategy game was carried (Ch Hauert, 2002). Additionally, in snow drift game, the effect of spa-
tial structure gave birth to other result (Christoph Hauert & Doebeli, 2004). Instead of these two-strategy
games, they have also been studying games that some elements have been added.
In biology, the phenomenon that three-strategy coexist have been founded (Zamudio & Sinervo,
2000). And these systems have been simulated in rock-paper-scissors game (Kerr, Riley, Feldman, &
Bohannan, 2002). From the point of view of evolutionary game theory, individual can follow a strategy of
cooperators, defectors, and loners (Dawes, 1980). Regardless of neighbors, loners have the fixed payoff,
hence they get more payoff than defectors. On the other hand, they don't get benefit by participating in
the game, so cooperators prevail around loners. Therefore, strategy of these three types is cyclic domi-
nant, so they coexist and overcome the dilemma (Hauert et al., 2002). Additionally, the concept of pun-
isher is suggested. Punisher has the opportunity to punish and imposes to co defecting co-players but
these actions are costly (Fehr & Gächter, 2002; Rockenbach & Milinski, 2006).
In 2009, the concept of mutation in the game of four strategies, cooperators, defectors, loners, and
punishers was introduced (Traulsen et al., 2009). A mutation occurs with probability mutation rate and a
mutant switches to a different random strategy. The dynamics has changed as the mutation rate increases.
Change of dynamic describes the cultural evolution differs from the genetic reproduction which has rare
mutations.
Previous researches have been carried out on lattice structures. More recently, interest in which lattice
structure is extended by graphs and social networks has increased and it also affects evolutionary game
theory. In the beginning, there is a research about evolutionary dynamics on graphs (Lieberman et al.,
2005). Each vertex of graph represents an individual and weighted edges of graph represent reproductive
rate. The result shows that amplifying random drift or selection in evolutionary dynamics is determined
by forms of graphs. In addition, it said how to access evolutionary games on graphs.
Along this study, the critical point to prevail cooperators was founded by analytical and numerical
simulations in two-strategy game of cooperators and defectors on several graphs and networks (Ohtsuki,
Hauert, Lieberman, & Nowak, 2006). This system followed 'death-birth' process: in each time step, a
random individual is chosen to die, and the neighbors compete for the empty site proportional to their fit-
ness. Also, previous research demonstrated that agent's adaptive expectation plays an important role in
cooperation emergence on complex networks (Ohtsuki et al., 2006).
Kim and Lee
2 MODEL
2.1
Strategies of Agent
In our game, four different players exist. Cooperator is an agent that gives benefit b to all players except
loner, but to be a cooperator it has to pay cost c for neighboring players except loner. Cooperator also gets
benefit by neighboring cooperator. Defector also gets benefit by neighboring cooperator without paying
any cost, but it gets certain amount of penalty from neighboring punisher. In the simple game with these
two types of agent, defector always has advantages over cooperator (Hofbauer & Sigmund, 1998). Loner
is an agent who is not interested in this game, so it gets fixed payoff regardless of having any types of
neighbor. Becoming loner is the best strategy if there is no cooperator around, but if there exists any co-
operator neighbor agent will not choose to be a loner directly. At least it has to calculate different payoff
functions if we assume total payoff of loner is bigger than benefit. Loner's relevant payoff between its
opponent is set to be . Punisher is an agent introduced to penalize defector like police officer.
In our paper, punisher is not a type of cooperator, their function is just giving punishment to neighboring
defectors. It is also benefited by neighboring cooperator but it has to bear small penalty ) to penal-
ize neighboring defector. Comparing cost of altruism, penalty of defect and small penalty to penalize, c
is established. Because each cooperator has to pay maximum cost by considering network
degree of d, and it has to be smaller than and . Otherwise agent will not choose to do altruistic behav-
ior. We brought the concept of punisher and loner from previous research (Traulsen et al., 2009), but we
put minor change of their role and payoff functions – in our model punisher doesn't act as a cooperator,
and cooperator doesn't pay costs for neighboring loner .
2.2
4-Person prisoner's dilemma game
Let's consider people doing teamwork. We can divide people's behavior into four different types. People
who trust other people and shows altruism – cooperator, free-rider who only gets benefit from teammate –
defector, person who is not interested in or not joining that work – loner, person who criticize bad team-
mate, but gets bed reputation because he always lays out a sermon – punisher (Greenberg J., & Baron,
R.A, 1997). Agent in our model acts exactly as these four different types of people and decide their role
by playing 4-person prisoner's dilemma game with their neighbors. And the result will show how differ-
ent behavior agent shows according to change of parameter and network structure.
The model consists of N agents playing 4-person prisoner's dilemma game. Agents are located on the
vertices of network. In every simulation step, agents play game with other agents within their own neigh-
borhood. The agent can choose four strategies: cooperation, defection, indifferent, or punish. For example
if player A is defector and player B is punisher, punisher give penalty to defector, and it also get minus
payoff while penalizing, so their payoff for this game is ( , ). The payoff matrix is as follows :
Cooperator
Defector
Loner
Punisher
Cooperator
( b – c , b – c )
( b , – c )
( , – c )
( b , – c )
Defector
( , b )
( 0 , 0 )
( , 0 )
( , )
Loner
( , )
( 0 , )
( , )
( 0 , )
Punisher
( , b )
( , )
( , 0 )
( 0 , 0 )
Table 1 : 4-person prisoner's dilemma game used in our model
By adding loner and punisher from the traditional prisoner's dilemma game, dominant strategy is to
choose loner for both players. So assuming that the agent plays prisoner's dilemma game one by one with
Kim and Lee
its neighbor, we can estimate that all players eventually become loner. But if we consider the society of
degree more than 1, agent has to deal with multiple neighbors. Present paper, we assume that each agent
knows the strategy of their neighbors. Since each agent is facing with their neighborhoods in the game, it
gathers all information of their neighborhoods states. And calculate how much total payoff it can get if it
changes into certain state. Detailed explanation is in the next section.
2.3
Update Mechanism
To see the dynamics of transition and find different behavior, we added participation probability on our
model. If participation rate is 1, every player in the game changes their state each time. In each step, par-
ticipants are randomly selected depending on participation rate, and every participating agent calculates
their expected payoff using the information of their neighbor. And player finally changes its state that
guarantees maximum payoff. If multiple maximum payoff value exists, agent picks one state randomly.
In our simulation, small random number ɛ is added to the payoff function so that agent can pick the max-
imum payoff between four.
We only considered agent's first-step neighbor to get payoff function. For example, there is an agent
calculating their options and decide to be a defector. If there is a punisher on its neighbor and whether
punisher has another defector on their neighborhood or not, penalty value for that defector doesn't change.
The number of corresponding neighborhood – cooperator, defector, loner, punisher is counted as
For example, the payoff function of cooperator can be evaluated as (
) because it get benefits from neighboring cooperators and pay costs c to ( ) neigh-
bors for altruistic acts.
The payoff and final strategy can be described as follows :
( )
( )
2.4
Spatial Structure
Six types of complex networks are used in this paper. Lattice networks of Moore-neighborhood (Network
degree d = 8) are used as a baseline to see the coexisting behavior depending on participation rate and
mutation rate. And lattice network of von-Neumann neighborhood(d=4) and network with d=16 are used
to compare how degree change affects behavior of agents.
And Cellular(Rives, A. W., & Galitski, T. 2003), Core-Periphery(Borgatti, S. P., & Everett, M. G.
2000), Erdős-Rényi,(Erdős, P., & Rényi, A. 1960), Scale-Free(Barabási , A. –L., & Albert, R. 1999),
Small-World(Watts, D. J., & Strogatz S. H. 1998) network are used. These networks are complex net-
works on the ground of having heterogeneous network degree for each node.
Previous research has proven that there are different critical points for cooperators to prevail over the sev-
eral types of complex networks. (Ohtsuki et al., 2006). And cooperation frequency fluctuations in a Bara-
bási and Albert (BA) network and Watts and Strogatz (WS) small world networks are compared (Bo.X,
2012). In this paper, we found different coexisting points and interesting phenomena depending on net-
work structure and degree.
2.5
Participation rate and Mutation rate
Kim and Lee
Participation rate p and mutation rate plays the important role in our simulation. In our setting, partici-
pation rate decides how many agents decide to change their state on each time and mutation rate describes
individuals experimenting with new behaviors. In our model, randomly selected participants play game
first, and mutation occurs with probability in each update step. Previous study suggests that adjusting
mutation rate can result significant change in the societies (Traulsen et al., 2009), and we found different
stationary states depending on participation rate and mutation rate.
3
RESULT & DISCUSSION
3.1
Simulation Setting
Initially, 2,500 agents are located on the grid space, and four types of agents are randomly distributed
throughout the space. The degree of the network is 8 by querying Moore neighborhood. For visualizing,
we used grid space with world height 50 and width 50, and the bias is removed by using torus space. Par-
ticipation rate p and mutation rate is varied in the interval [0, 1] and benefit b, cost c, penalty , small
penalty is set to be 100, 5, 150, 50 respectively, And loner's total payoff for this game is set to be b for
Moore neighborhood. (
. Since initial place of agent located can affect the final coexisting proba-
bility, we simulated 30 times by changing random seeds. In our simulation, we simulated by controlling
participation rate and mutation rate together by batch run. We picked 4 participation rate p = 0.0001, 0.01,
0.1, 1 by changing mutation rate = 0, 0.0001, 0.001, 0.01, 0.01. Our null hypothesis is when p = 1, = 0.
Table 2 and 3 shows the simulation parameters and virtual experiment design.
Figure Value
Number Of Agent
2500
Space Height
Space Width
50
50
Benefit (b)
100
Cost (c)
5
Penalty (β)
Small Penalty (γ)
50
15
Loner's payoff (σ)
12.5
Table 2 : Simulation parameters
Experiment Variable name
Experiment Design
Implication
Participation rate (p)
Mutation rate (
Simulation ending time
0.001, 0.01, 0.1, 1(4 cases) Participation rate of each agent
0, 0.0001, 0.001, 0.01, 0.1 (5 cases) Randomly mutated rate
10000 ticks Stable point for simulation
Each experiment is replicated 30 times by changing random seeds.
Table 3 : Virtual experiment design of scenario of interests
3.2
Participation rate & Mutation rate
Kim and Lee
Participation and Mutation rate are important factors that decide coexisting condition. We simulated 4-
agent prisoner's dilemma game by controlling Participation rate and Mutation rate. Figure 1 describes
simple example of how coexisting probability and behavior changes depending on the participation rate
and mutation rate. Red, blue, green, black dots are cooperator, defector, loner, and punisher agents. Left is
a stable situation with p = 1, = 0. Middle is loner-dominant society after changing mutation rate to =
0.01. Right you can see coexistence again by changing p = 0.1. We can say that stable structure is col-
lapsed by adopting mutation, because cooperators who give benefit have a chance to change into other
agents and their neighborhoods decide to change themselves in the next step. But by put participation rate
into our simulation, we can slow down the second part of the action and give more chance to survive the
coexisting structure.
(Fixed parameters : 2,500 agents in grid space with Moore neighborhood, b=100, c=5, =150, =50)
Figure 1 : Simulation results of changing participation rate and mutation rate
(Left : p = 1, = 0, Middle : p = 1, = 0.01, Right : p = 0.1, = 0.01)
Interesting fact is that, when mutation rate is small cooperator is dominant when participation rate is
We define coexistence value to measure how equally four types of agents are spread throughout the so-
ciety, which is equal to √
. By using this value, we can easily conclude that condition,
p = 1, = 0, results in highest coexistence value.
the smallest. For example, = 0.0001, 0.001 number of cooperators is the largest when p = 0.001.
But becomes 0.01, cooperator is dominant when p = 0.01. When , number of cooperator is the
largest when p = 0.1. On the other hand, when p = 0.001, number of cooperator is the largest when
As participation rate becomes higher, number of cooperator become larger when higher value
of mutation rate is obtained.
Table 3 summarized our results by calculating average number of each agents and result of t-test. We
can find that coexisting condition changes depending on participation rate and mutation rate. And Figure
2 shows that becomes larger when more agents mutate with small participation rate. To compare with
other networks, number of agent is set as 1,000. Null hypothesis is p = 1, = 0 for the case of = 0. And
for each participation rate, = 0 is set as null hypothesis. (+: P < 0.05, *: P < 0.01).
Participation
Mutation
Cooperator
Defector
Loner
rate
0.001
0.001
0.001
0.001
0.01
rate
0.0001
0.001
0.01
0.1
0.0001
Mean
277.0
377.9
268.4
251.4
162.4
t-value
-6.6*
18.4*
-16.7*
-24.3*
-15.1*
Mean
211.1
217.7
234.1
247.0
134.8
t-value
-4.2*
-2.9*
1.6
7.6*
-11.0*
Mean
310.5
195.1
258.7
248.9
567.2
t-value
8.4*
-9.5*
13.4*
8.3*
13.9*
Punisher
Mean
t-value
Coexistence
Value
201.5
209.3
238.7
252.3
135.7
-4.6* 245.9
-4.5* 240.7
4.7* 249.6
10.4* 249.9
-12.3* 202.6
0.01
0.01
0.01
0.1
0.1
0.1
0.1
1
1
1
1
0.001
0.01
0.1
1
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0
0
0
0
171.6
380.3
271.1
142.0
3.7
140.0
375.6
100.0
1.4
4.7
212.1
315.5
271.5
268.2
269.1
-16.6*
32.6*
-0.1
-23.9*
-377.8*
-19.2*
28.3*
-30.3*
-408.7*
-350.0*
-7.7*
7.8*
0.4
-0.2
0.0
Kim and Lee
145.9
214.6
236.7
118.0
10.0
132.9
205.1
84.6
2.3
12.0
174.3
230.0
200.1
204.1
201.2
-12.1*
5.4*
16.5*
-21.7*
-221.7*
-15.2*
0.4
-23.8*
-287.5*
-163.1*
-5.4*
3.9*
-0.2
0.6
0.0
529.6
196.5
252.7
622.0
977.7
590.6
202.6
731.0
994.7
973.4
445.6
226.1
330.7
336.0
335.2
16.1*
-45.8*
-27.4*
21.8*
302.6*
18.6*
-41.4*
28.4*
383.1*
285.9*
7.5*
-12.0*
-0.3
0.1
0.0
153.0
208.6
238.9
118.0
8.7
136.5
216.3
84.2
1.6
10.2
168.3
228.5
197.8
191.9
194.4
-9.3* 212.2
3.6* 240.5
16.5* 249.5
-15.3* 187.3
-202.8* 23.6
-12.7* 196.8
9.7* 241.1
-25.3* 151.1
-405.0* 8.5
-197.0* 27.3
-5.6* 229.5
6.0* 247.4
0.6 244.1
-0.6 243.7
0.0 243.7
Table 3 : Different coexisting condition by changing participation and mutation rate (Moore, 1000)
Figure 2 : Coexistence value (left) and Number of cooperator (right) of each cases
3.3
Network degree
We conducted same simulation by changing network into von-Neumann neighborhood, which has net-
work degree of 4. Interestingly, we found that number of loner increase than the previous simulation with
Moore neighborhood. The main reason loner becomes dominant is because degree of neighborhood is de-
creased to 4, which means that chance of having cooperator neighborhood is decreased. As agent doesn't
have any cooperator neighborhood, they can't choose to be cooperator, defector, and loner because their
payoff function becomes negative or 0. But in our payoff function, loner get non-zero payoff for any cir-
cumstances so agent with no cooperator neighbor will become loner eventually. On the other hand, by
simulating network with degree = 16, larger number of cooperator and coexistence value is observed.
Null hypothesis here is the same case of Moore neighborhood. (+: P < 0.05, *: P < 0.01).
Participation
Mutation
Cooperator
Defector
Loner
Punisher
Coexistence
Value
rate
0.001
0.001
0.001
0.001
0.01
rate
0.0001
0.001
0.01
0.1
0.0001
Mean
51.1
206.2
243.4
248.6
7.5
t-value Mean
-116.7*
94.8
-66.8*
234.7
-9.0*
245.0
249.9
-1.1
-211.7*
13.5
t-value Mean
756.7
327.9
261.3
248.6
965.4
-47.0*
5.2*
4.0*
1.3
-125.8*
t-value Mean
97.5
231.3
250.1
252.6
13.6
72.4*
42.3*
1.0
-0.1
153.3*
t-value
-39.5* 137.5
6.6* 246.1
4.9* 249.8
0.1 249.9
-113.7* 34.0
Kim and Lee
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0
0
0
0
34.9
205.0
249.5
2.8
2.9
34.3
215.5
1.6
0.3
3.6
41.7
106.8
67.6
63.0
72.1
-111.0*
-59.6*
-9.2*
-244.4*
-2.1+
-84.7*
-52.3*
-257.6*
-10.6*
-2.1+
-112.1*
-69.5*
-109.7*
-92.6*
-92.2*
73.8
233.5
245.8
4.4
8.9
75.9
230.3
2.1
1.0
8.8
89.0
169.0
90.3
84.2
94.3
-33.6*
6.3*
4.0*
-143.2*
-1.6
-29.4*
8.1*
-172.2*
-6.0*
-3.3*
-42.0*
-13.0*
-48.9*
-45.4*
-39.2*
816.5
328.9
255.5
988.3
979.7
812.8
317.0
993.6
997.6
978.4
785.3
558.4
753.4
768.8
737.3
65.0*
36.4*
1.0
170.8*
1.5
52.3*
30.6*
182.1*
5.3*
2.5*
75.0*
37.8*
69.4*
57.7*
54.9*
74.9
231.9
248.9
4.4
8.7
77.9
237.1
2.6
1.0
9.2
85.2
165.8
88.8
84.1
95.5
-39.4* 112.0
8.5* 245.8
3.8* 249.9
-139.3* 15.3
0.0 21.7
-26.9* 113.3
9.5* 247.2
-139.1* 9.7
-2.0+ 4.0
-1.2 23.1
-36.7* 125.6
-17.2* 202.2
-49.3* 142.1
-37.8* 136.1
-35.9* 147.9
Table 4 : Simulation result with von-Neumann neighborhood lattice
0.01
0.01
0.01
0.1
0.1
0.1
0.1
1
1
1
1
0.001
0.01
0.1
1
Participation
Mutation
Cooperator
Defector
Loner
Punisher
Coexistence
rate
0.001
0.001
0.001
0.001
0.01
0.01
0.01
0.01
0.1
0.1
0.1
0.1
1
1
1
1
0.001
0.01
0.1
1
rate
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0
0
0
0
Mean
263.6
422.4
292.5
256.4
172.4
262.2
419.1
290.6
153.5
155.6
264.1
414.6
138.7
127.1
142.7
267.3
283.3
197.2
198.5
192.1
t-value Mean
225.3
219.1
238.7
245.2
145.0
223.8
226.7
241.3
126.1
131.0
227.6
218.9
116.3
106.7
120.9
218.4
270.3
161.5
164.7
159.8
-3.4*
14.5*
9.6*
2.4*
2.0
21.3*
16.9*
8.2*
2.0
34.7*
25.6*
14.3*
10.6*
28.0*
25.1*
13.8*
-8.4*
-18.7*
-14.5*
-24.9*
t-value Mean
273.6
131.7
234.4
247.5
538.6
276.9
131.7
227.7
596.8
578.8
277.6
140.6
630.2
660.6
611.4
287.2
166.7
476.0
472.5
486.5
3.3*
0.4
2.1+
-0.7
2.2+
14.4*
3.7*
1.7+
1.8+
34.1*
20.1*
4.5*
9.7*
31.4*
31.3*
9.3*
5.5*
-12.0*
-11.0*
-16.5*
t-value Mean
237.4
226.9
234.9
251.3
144.0
237.0
221.8
240.6
123.6
134.7
229.8
227.1
114.8
105.6
124.6
226.5
279.7
165.2
164.3
161.6
-7.2*
-25.5*
-9.6*
-0.6*
-2.3+
-30.1*
-33.4*
-13.0*
-1.7+
-36.3*
-40.1*
-27.5*
-10.7*
-31.4*
-29.8*
-20.1*
-12.8*
15.0*
12.3*
22.0*
t-value
Value
8.5* 249.2
5.3* 229.3
-1.5 249.0
-0.5 250.1
2.2+ 209.9
15.6* 249.1
4.0* 229.5
0.9 249.0
1.2 194.4
33.1* 199.6
18.2* 248.8
2.9* 232.0
10.0* 184.8
33.5* 175.4
29.2* 190.4
12.1* 248.2
7.9* 244.5
-10.2* 223.7
-9.0* 224.5
-14.1* 221.6
Table 5 : Simulation result with lattice, degree=16
3.4
Network Structure
We conducted same simulation by applying complex network structures. Cellular, Core-periphery, Erdős-
Rényi, Small-world, Scale-free network with same number of node and edge are used. Since network de-
gree of each node diverse in complex network, random seed which decides agent's initial distribution
over the network becomes critical here. For example, if node with degree 100 is chosen as a cooperator,
its 100 neighbors will get benefit by it, so overall agent dynamics will reflect that positive effect. In that
case, much more cooperators will be survived to the end. Figure 3 shows how different the stable states of
agent will change, depending on the initial condition by changing random seed (Network information:
Erdős-Rényi random network with , ).
100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
Kim and Lee
Punisher
Loner
Defactor
Cooperator
Random Seed (1-30)
Figure 3 : Different agent behavior by changing random seed
We used meta-network assessment and analysis tool ORA(Carley, Kathleen M et al., 2012) to generate
six different stylized networks. To restrict our focus into network structure, we fixed the number of nodes
to 1,000 and made number of edge as close as 8,000. We didn't remove any surplus edges not to destruct
network characteristic. Table 6 shows detail explanation of each network.
Network
Number of
Number of
Node
Edge
Other information
Grid(Moore)
Grid(von-
Neumann)
Lattice
Cellular
Core-Periphery
Erdős-Rényi,
1,000, 2,500
1,000, 2,500
Used Grid space provided by Repast
Used Grid space provided by Repast
1,000
1,000
1,000
1,000
8,000 Directed Degree of 8 - Number of neighbor : 16
7,995 Number of Cells : 52, Inner Density : 0.40, Outer Density : 0.10
8,003 Proportion of core nodes : 0.13, Density of core nodes : 0.50
8,000 Total density : 0.008 (symmetric)
Scale-Free
1,000
8,004
Small-World
1,000
8,005
Probability of node connecting to core : 0.008
Initial node count(core) : 40, Initial density(core) : 0.01
Number of node connected by edge : 891 (109 nodes has degree 0)
Number of neighbors : 8, Probability of removing neighbor : 0.05
Probability of adding for neighbor : 0.055, Power law exponent :
0.055
Network made by ORA, which is directed graph, but we used it as un-directed.
Table 6 : Specific Network Information
We found the result of networks having heterogeneous degree is more fluctuated than homogeneous net-
work by observing standard deviation. Figure 4, 5 shows the mean and standard deviation of each agent
type. In the condition of p=0.1, μ=0.001, Erdős-Rényi and Scale-free network has the biggest standard
deviation because of their diversity of node degree. Cooperator is not appeared in lattice network since
stable structure is broken because of mutation rate (Section 3.2). However by changing graph structure,
we can find coexisting behavior in three types of networks – Erdős-Rényi, Scale-free, Small-world.
Simulation results by varying network structures are in APPENDIX.
Kim and Lee
Figure 4 : Average number of agent by changing network structure (p=0.1, μ=0.001)
Figure 5 : Standard deviation of agent by changing network structure(p=0.1, μ=0.001)
3.5
Limitation of our simulation model
In our model, we used pre-defined participation rate and mutation rate value, to verify that those parame-
ters affect our model stochastically. But in the real world agent can decide themselves whether joining for
certain game on each time, it is not decided by force. And agents can choose different strategy by their
own utility function, not just combining the payoff depending on their neighborhood. Making agents pay-
off heterogeneous depends on each agent's background will be needed to show organizational behavior
Kim and Lee
better. In order to use self-motivated agent model in the real world, calibration depending on the situation
is needed. Diverse type of real world dataset will be needed to explain the society.
Second, we assume that each agent knows current state of its neighbors. But in real life, information
imbalance makes not everyone can decide their decision in optimal way. Also more accurate initial condi-
tion will be considered.
Last, stochastic difference can be occurred, if sequence of applying participation rate and mutation
rate is changed. Although we reflect participation and mutation is independent, our result shows that
there is a correlation between two variables.
4
CONCLUSIONS
In this paper, we studied the behavior of 4 different types of self-motivated agents playing N-person pris-
oner's dilemma game. By using agent based modeling, we can reflect the agent's mutation and participa-
tion in stochastic way. And we could observe different behavior by changing network structure.
We found the impact of participation rate and mutation rate on agent society. Degree of network is al-
so very critical factor changing population of each agent type. And we found that structure of network is
important in that each agent affect society different according to heterogeneity. Besides some cases espe-
cially in Erdős-Rényi, network, we found the phenomenon that initial location of agent affect their coex-
isting condition critically.
For further work, we will compare how self-motivated agent behaves different with agents in tradi-
tional evolutionary games. By comparing different coexisting condition, we can analyze which model ex-
plains our human nature and society better. Also, we will put different utility function while calculating
payoff for emphasizing heterogeneity of agent. Last, we will prove how participation rate affects with
mutation rate in analytical way.
A APPENDIX
Table 5 to 8 is the simulation result of agents' behavior in different complex networks. We can find that
participation – mutation relation is also followed in complex networks. Null hypothesis is p = 1, = 0 for
the case of = 0. And for each participation rate, = 0 is set as null hypothesis. (+: P < 0.05, *: P < 0.01).
Participation
Mutation
Cooperator
Defector
Loner
Punisher
t-value
Coexistence
Value
rate
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0
0
0
0
rate
0.001
0.001
0.001
0.001
0.01
0.01
0.01
0.01
0.1
0.1
0.1
0.1
1
1
1
1
0.001
0.01
0.1
1
t-value Mean
101.8
153.8
230.4
249.2
20.4
139.7
156.2
237.2
3.1
21.2
123.3
163.8
1.1
5.8
21.6
131.4
138.3
3.7
4.8
2.8
Mean
115.1
417.2
299.1
254.2
9.8
145.4
413.0
293.5
2.8
13.1
135.9
416.5
1.5
7.4
14.1
134.0
216.9
5.8
8.5
4.9
Table 7 : Simulation result with Cellular Network
t-value Mean
527.2
140.2
237.3
246.7
949.9
469.9
140.3
229.2
990.6
942.0
495.0
144.9
996.9
982.7
941.3
513.3
165.9
988.1
983.1
990.3
t-value Mean
255.9
288.6
233.3
249.7
19.8
245.1
291.1
240.4
3.5
23.4
245.2
276.4
0.5
4.1
23.3
219.2
478.9
2.5
3.5
2.0
38.3*
-14.0*
27.8*
41.2*
-9.2*
-43.5*
-424.3*
-328.1*
3.7*
-8.6*
-46.4*
-336.0*
4.8*
-2.2+
-10.0*
-44.2*
-119.7*
-0.8
-1.8+
0.0
-24.5*
72.4*
34.2*
15.2*
3.0*
23.1*
119.7*
110.2*
-6.9*
2.3*
28.5*
133.6*
-4.9*
1.5
4.7*
26.1*
33.3*
0.6
1.7+
0.0
-7.7*
8.6*
45.4*
47.4*
8.3*
21.6*
55.8*
84.6*
-2.5*
6.2*
27.5*
60.0*
-3.4*
2.2+
7.5*
20.8*
20.7*
1.0
1.8+
0.0
-39.2* 199.4
-63.4* 225.8
-92.6* 248.5
-95.4* 250.0
7.1* 44.1
34.1* 219.9
61.1* 226.6
110.9* 248.9
0.0 13.2
10.9* 49.7
33.4* 212.3
79.6* 228.7
-7.1* 5.3
3.0* 20.3
9.7* 50.8
36.5* 211.0
58.4* 221.0
0.9 15.1
1.9+ 19.4
0.0 12.7
Kim and Lee
Participation
Mutation
Cooperator
Defector
Loner
Punisher
Coexistence
Value
rate
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0
0
0
0
rate
0.001
0.001
0.001
0.001
0.01
0.01
0.01
0.01
0.1
0.1
0.1
0.1
1
1
1
1
0.001
0.01
0.1
1
Mean
26.4
180.3
240.9
251.2
2.9
26.3
178.4
238.8
0.2
3.5
26.0
186.3
0.0
0.4
3.3
28.3
26.6
0.0
0.0
0.0
t-value Mean
100.3
214.6
234.7
252.7
15.1
101.3
216.5
240.8
0.6
14.8
104.7
214.9
0.0
0.3
19.1
120.1
200.5
0.0
0.0
0.0
Table 8 : Simulation result with Core-Periphery Network
t-value Mean
834.3
399.2
288.8
249.7
955.2
833.8
394.4
280.6
996.9
960.6
835.8
387.9
1000.0
999.1
958.0
809.4
668.4
1000.0
1000.0
1000.0
t-value Mean
38.9
206.1
236.4
246.7
26.8
38.6
211.3
240.0
2.3
20.7
33.9
211.9
0.0
0.4
20.5
43.0
104.5
0.0
0.0
0.0
14.2*
-60.6*
-163.0*
-189.3*
-5.5*
-17.0*
-137.1*
-270.0*
-1.5
-4.7*
-24.2*
-206.2*
-1.0
-6.1*
-4.6*
-25.5*
-37.7*
0.0
0.0
0.0
-0.1
64.3*
101.4*
104.6*
8.2*
17.3*
79.2*
105.5*
2.5*
5.2*
21.5*
97.3*
0.0
3.3*
7.8*
18.2*
18.2*
0.0
0.0
0.0
-10.4*
39.4*
59.7*
64.2*
3.7*
6.6*
85.4*
93.1*
1.2
3.5*
18.8*
86.8*
0.0
3.6*
2.9*
12.3*
22.9*
0.0
0.0
0.0
rate
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0
0
0
0
rate
0.001
0.001
0.001
0.001
0.01
0.01
0.01
0.01
0.1
0.1
0.1
0.1
1
1
1
1
0.001
0.01
0.1
1
t-value Mean
55.2
193.0
239.1
247.1
80.8
33.7
203.5
241.9
1.3
107.6
36.7
204.0
0.1
0.8
450.2
41.2
194.9
73.4
37.1
26.6
t-value Mean
Mean
541.1
222.0
248.8
422.2
235.5
286.9
252.6
252.7
261.2
127.5
501.6
131.6
241.8
416.6
244.1
280.5
1.5
0.3
249.3
120.6
504.1
135.0
240.7
412.3
0.2
0.0
1.1
0.2
45.9
132.1
512.6
154.6
318.4
454.6
116.5
93.8
42.6
56.2
30.0
22.0
Table 9 : Simulation result with Erdős-Rényi, Network
t-value Mean
181.7
136.1
239.2
247.2
530.8
332.9
138.7
233.4
996.7
522.1
325.1
145.3
999.7
997.8
372.1
288.9
32.2
716.3
864.0
921.4
-27.7*
-0.7
20.2*
20.4*
0.4
-35.9*
41.4*
67.8*
-158.5*
3.1*
-0.4
51.7*
-253.4*
-145.5*
34.6*
13.7*
18.3*
2.5*
0.6
0.0
-14.6*
-11.9*
-76.2*
-86.2*
1.3
8.6*
136.2*
75.9*
-572.7*
2.8*
21.7*
146.8*
-899.0*
-379.0*
7.7*
25.1*
117.4*
2.2+
1.0
0.0
8.7*
46.6*
91.3*
94.8*
-3.2*
-44.9*
-320.3*
-215.2*
233.0*
-6.6*
-66.3*
-354.8*
373.2*
189.9*
-23.9*
-61.5*
-817.2*
-3.2*
-0.9
0.0
t-value
-20.1* 96.3
5.4* 237.5
12.1* 249.3
24.5* 250.1
5.3* 32.5
18.7* 96.2
79.1* 238.2
94.4* 249.4
2.1+ 4.2
3.3* 32.0
18.8* 93.7
94.3* 239.5
1.0 0.0
3.2* 2.4
4.0* 33.4
27.6* 104.3
55.0* 138.9
0.0 0.0
0.0 0.0
0.0 0.0
24.9* 186.3
-20.5* 229.2
-36.2* 249.3
-33.7* 249.9
4.8* 194.4
69.1* 164.9
30.0* 230.9
48.2* 249.3
-128.2* 4.9
6.4* 202.8
90.7* 168.8
54.5* 232.9
-184.6* 0.9
-88.3* 3.7
4.6* 178.5
67.8* 175.2
26.0* 173.6
4.2* 154.8
1.1 93.6
0.0 63.4
Participation
Mutation
Cooperator
Defector
Loner
Punisher
t-value
Coexistence
Value
Participation
Mutation
Cooperator
Defector
Loner
Punisher
rate
0.001
0.001
0.001
0.001
0.01
0.01
0.01
0.01
0.1
rate
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
Mean
93.0
338.5
252.5
222.3
100.2
84.6
339.2
248.8
50.8
t-value Mean
31.6
176.3
213.3
217.4
96.0
28.6
176.5
211.4
100.5
-63.5*
2.8*
-24.7*
-43.0*
3.8*
26.7*
142.8*
92.6*
2.6*
t-value Mean
-100.1*
341.3
6.4*
136.1
22.4*
213.0
27.2*
225.0
3.1*
476.3
24.5*
383.2
48.7*
139.3
75.4*
215.9
3.0*
713.1
t-value Mean
425.0
240.0
212.3
224.6
218.4
394.7
236.2
213.6
26.4
28.2*
43.5*
63.2*
64.1*
-6.7*
-52.7*
-452.3*
-214.7*
-3.0*
9.0* 143.7
-29.8* 210.1
-66.5* 222.1
-51.7* 222.3
7.8* 177.9
55.4* 138.3
44.2* 210.7
84.6* 221.9
2.5* 99.0
t-value
Coexistence
Value
Kim and Lee
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0
0
0
0
0.1
0.1
0.1
1
1
1
1
0.001
0.01
0.1
1
65.8
84.7
337.1
82.9
113.5
107.9
94.9
331.3
17.8
0.0
0.0
3.7*
23.1*
121.8*
3.4*
4.6*
6.0*
24.8*
46.7*
1.4
0.0
0.0
141.7
394.9
228.4
39.8
45.7
152.6
410.4
356.3
33.2
0.0
0.0
Table 10 : Simulation result with Scale-Free Network
-8.1*
-61.0*
-435.5*
-3.7*
-5.5*
-11.0*
-53.1*
-354.4*
-1.6
0.0
0.0
543.4
381.8
138.9
631.9
504.8
380.9
355.0
44.1
830.5
891.0
891.0
140.2
29.3
184.9
136.4
226.9
247.0
32.2
159.2
9.6
0.0
0.0
4.2*
26.8*
52.8*
3.7*
5.6*
6.2*
22.0*
19.3*
1.4
0.0
0.0
rate
0.001
0.001
0.001
0.001
0.01
0.01
0.01
0.01
0.1
0.1
0.1
0.1
1
1
1
1
0.001
0.01
0.1
1
rate
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0.0001
0.001
0.01
0.1
0
0
0
0
t-value Mean
169.3
189.0
238.1
249.2
212.9
155.4
195.8
241.7
212.3
210.2
168.3
196.8
210.9
214.5
216.8
180.0
222.5
207.9
202.3
200.1
t-value Mean
Mean
327.0
286.1
258.2
416.7
235.9
291.6
250.6
254.5
214.5
248.4
328.7
275.5
247.6
422.9
245.0
284.1
205.9
242.7
223.9
249.2
319.2
280.4
244.0
419.1
196.7
233.1
203.9
241.9
219.5
248.9
301.4
296.9
346.6
327.1
198.3
223.7
197.9
221.3
219.5
196.3
Table 11 : Simulation result with Small-World Network
16.6*
17.0*
55.0*
81.8*
-3.5*
-16.1*
-112.7*
-63.4*
-3.3*
-6.7*
-19.6*
-107.4*
-1.8+
-3.7*
-5.3*
-22.1*
-67.7*
-1.1
0.0
0.0
t-value Mean
217.6
136.1
234.6
244.3
324.0
240.1
134.2
229.8
338.8
316.8
232.5
141.5
359.3
339.8
314.8
221.1
103.8
370.1
378.4
384.0
-11.8*
-12.7*
7.0*
12.5*
1.1
-12.6*
-3.8*
13.4*
2.0+
1.9+
-5.9*
-2.4*
2.3+
3.4*
3.4*
-3.8*
3.0*
1.3
0.0
0.0
-9.3*
25.3*
-12.7*
-31.9*
3.9*
9.5*
60.3*
27.3*
3.8*
5.8*
14.7*
65.6*
2.2+
4.2*
4.7*
15.0*
27.4*
0.8
0.0
0.0
9.6* 163.2
73.4* 139.1
48.9* 210.9
2.9* 129.9
4.2* 156.1
5.1* 198.4
57.8* 145.2
27.5* 169.7
1.6 46.6
0.0 0.0
0.0 0.0
-4.9* 242.3
-24.8* 229.4
-42.0* 249.0
-41.0* 249.6
4.1* 246.2
27.9* 241.1
11.9* 229.0
20.0* 249.3
2.6* 244.8
7.8* 246.9
24.1* 243.3
12.3* 231.0
0.1 242.8
2.0+ 244.9
4.8* 247.1
19.3* 244.3
19.5* 226.2
0.6 241.7
0.0 240.6
0.0 239.9
Participation
Mutation
Cooperator
Defector
Loner
Punisher
t-value
Coexistence
Value
AUTHOR BIOGRAPHIES
SUNDONG KIM is a Master candidate at Department of Industrial and Systems Engineering, KAIST.
His email address is [email protected].
JIN-JAE LEE is a Ph.D. candidate at Department of Physics, KAIST.
His email address is [email protected]
REFERENCES
Axelrod, R., & Hamilton, W. D. (1981). The evolution of cooperation. Science, 211(4489), 1390-1396.
Barabási , A. –L., & Albert, R. (1999). Emergence of Scaling in Random Networks. Science, 286(5439),
509-512.
Kim and Lee
Bo, X.(2012). Prisoner's dilemma game on complex networks with agent's adaptive expectations. Journal
of Artificial Societies and Social Simulation, 15(3) 3.
Borgatti, S. P & Everett, M. G. (2000). Models on core/periphery structures. Social Networks, Volume 21,
Issue 4, 375-395.
Carley, Kathleen M., Pfeffer, J., Reminga, J., & Storrick, J., Columbus, D. (2012) ORA User's guide 2013.
Defense Technical Information Center document.
Dawes, R. M. (1980). Social dilemmas. Annual review of psychology, 31(1), 169-193.
Erdős, P., & Rényi, A. (1960). On the evolution of random graphs. Publications of the Mathematical
Institute of the Hungarian Academy of Science. 5, 17.
Fehr, E., & Fischbacher, U. (2003). The nature of human altruism. Nature, 425(6960), 785-791.
Fehr, E., & Gächter, S. (2002). Altruistic punishment in humans. Nature, 415(6868), 137-140.
Glenn, W. Rowe. (1985). Mutations, mixed strategies and game theory. Journal of Theoretical Biology,
Volume 117, Issue 2, 291-302.
Greenberg J., & Baron, R.A. (1997). Behavior in organizations, 6th ed. Englewood Cliffs, NJ: Prentice
Hall.
Hauert, C. (2002). Effects of space in 2× 2 games. International Journal of Bifurcation and Chaos,
12(07), 1531-1548.
Hauert, C., De Monte, S., Hofbauer, J., & Sigmund, K. (2002). Volunteering as red queen mechanism for
cooperation in public goods games. Science, 296(5570), 1129-1132.
Hauert, C., & Doebeli, M. (2004). Spatial structure often inhibits the evolution of cooperation in the
snowdrift game. Nature, 428(6983), 643-646.
Hofbauer, J., & Sigmund, K. (1998). Evolutionary games and population dynamics: Cambridge
University Press.
Kerr, B., Riley, M. A., Feldman, M. W., & Bohannan, B. J. (2002). Local dispersal promotes biodiversity
in a real-life game of rock–paper–scissors. Nature, 418(6894), 171-174.
Jörgen W. Weibull. (1995). Evolutionary game theory. Cambridge, MA : The M.I.T.Press. Cloth: ISBN
0-262-23181-6; Paper: ISBN 0-262-73121-5
Lieberman, E., Hauert, C., & Nowak, M. A. (2005). Evolutionary dynamics on graphs. Nature,
433(7023), 312-316.
Nowak, M. A., & May, R. M. (1992). Evolutionary games and spatial chaos. Nature, 359(6398), 826-829.
Ohtsuki, H., Hauert, C., Lieberman, E., & Nowak, M. A. (2006). A simple rule for the evolution of
cooperation on graphs and social networks. Nature, 441(7092), 502-505.
Rives, A. W., & Galitski, T. (2003). Modular organization of cellular networks. Proceedings of the
national Academy of Sciences, 100(3), 1128-1133.
Rockenbach, B., & Milinski, M. (2006). The efficient interaction of indirect reciprocity and costly
punishment. Nature, 444(7120), 718-723.
S.Boccaletti, V.Latora, Y.Moreno, M.chavez, D.-U. Hwang. (2005). Complex networks: Structure and
dynamics. Physics Reports, Volume 424, Issues 4-5, 175-308.
Traulsen, A., Hauert, C., De Silva, H., Nowak, M. A., & Sigmund, K. (2009). Exploration dynamics in
evolutionary games. Proceedings of the National Academy of Sciences, 106(3), 709-712.
Watts, D. J., & Strogatz, S. H. (1998). Collective dynamics of 'small-world' networks. Nature, 393, 440-
442.
Wedekind, C., & Milinski, M. (2000). Cooperation through image scoring in humans. Science,
288(5467), 850-852.
Zamudio, K. R., & Sinervo, B. (2000). Polygyny, mate-guarding, and posthumous fertilization as
alternative male mating strategies. Proceedings of the National Academy of Sciences, 97(26),
14427-14432.
|
0709.2410 | 2 | 0709 | 2007-09-19T15:29:37 | Distributed Decision Through Self-Synchronizing Sensor Networks in the Presence of Propagation Delays and Asymmetric Channels | [
"cs.MA",
"cs.DC"
] | In this paper we propose and analyze a distributed algorithm for achieving globally optimal decisions, either estimation or detection, through a self-synchronization mechanism among linearly coupled integrators initialized with local measurements. We model the interaction among the nodes as a directed graph with weights (possibly) dependent on the radio channels and we pose special attention to the effect of the propagation delay occurring in the exchange of data among sensors, as a function of the network geometry. We derive necessary and sufficient conditions for the proposed system to reach a consensus on globally optimal decision statistics. One of the major results proved in this work is that a consensus is reached with exponential convergence speed for any bounded delay condition if and only if the directed graph is quasi-strongly connected. We provide a closed form expression for the global consensus, showing that the effect of delays is, in general, the introduction of a bias in the final decision. Finally, we exploit our closed form expression to devise a double-step consensus mechanism able to provide an unbiased estimate with minimum extra complexity, without the need to know or estimate the channel parameters. | cs.MA | cs |
Distributed Decision Through Self-Synchronizing
Sensor Networks in the Presence of Propagation
Delays and Asymmetric Channels
Gesualdo Scutari, Sergio Barbarossa and Loreto Pescosolido
Dpt. INFOCOM, Univ. of Rome "La Sapienza ", Via Eudossiana 18, 00184 Rome, Italy
E-mail: {scutari, sergio, loreto}@infocom.uniroma1.it.∗
Manuscript submitted on January 10, 2007. Finally accepted on September 15, 2007.
Abstract
In this paper we propose and analyze a distributed algorithm for achieving globally optimal
decisions, either estimation or detection, through a self-synchronization mechanism among linearly
coupled integrators initialized with local measurements. We model the interaction among the nodes
as a directed graph with weights (possibly) dependent on the radio channels and we pose special
attention to the effect of the propagation delay occurring in the exchange of data among sensors, as
a function of the network geometry. We derive necessary and sufficient conditions for the proposed
system to reach a consensus on globally optimal decision statistics. One of the major results proved
in this work is that a consensus is reached with exponential convergence speed for any bounded delay
condition if and only if the directed graph is quasi-strongly connected. We provide a closed form
expression for the global consensus, showing that the effect of delays is, in general, the introduction
of a bias in the final decision. Finally, we exploit our closed form expression to devise a double-
step consensus mechanism able to provide an unbiased estimate with minimum extra complexity,
without the need to know or estimate the channel parameters.
1 Introduction and Motivations
Endowing a sensor network with self-organizing capabilities is undoubtedly a useful goal to increase the
resilience of the network against node failures (or simply switches to sleep mode) and avoid potentially
dangerous congestion conditions around the sink nodes. Decentralizing decisions decreases also the
vulnerability of the network against damages to the sink or control nodes. Distributed computation
over a network and its application to statistical consensus theory has a long history (see, e.g., [1, 2]),
including the pioneering work of Tsitsiklis, Bertsekas and Athans on asynchronous agreement problem
for discrete-time distributed decision-making systems [3] and parallel computing [4, 5]. A sensor
network may be seen indeed as a sort of distributed computer that has to evaluate a function of the
∗This work has been partially funded by the WINSOC project, a Specific Targeted Research Project (Contract Number
0033914) co-funded by the INFSO DG of the European Commission within the RTD activities of the Thematic Priority
Information Society Technologies, and by ARL/ERO Contract N62558-05-P-0458.
Part of this work was presented in the IEEE Workshop on Signal Processing Advances in Wireless Communications,
(SPAWC-2006), July 2006, for undirected graphs.
1
measurements gathered by each sensor, possibly without the need of a fusion center. This problem may
be addressed by taking into account the vast literature on distributed consensus/agreement algorithms.
These techniques have received great attention in the recent years within the literature on cooperative
control and multiagent systems [6]−[22]. In particular, the conditions for achieving a consensus over
a common specified value, like a linear combination of the observations, was solved for networked
continuous-time dynamic systems by Olfati-Saber and Murray, under a variety of network topologies,
also allowing for topology variations during the time necessary to achieve consensus [7, 8]. The discrete-
time case was addressed by Tsitsiklis in [4] (see also [5]). Many recent works focused on the distributed
computation of more general functions than the average of the initial measurements. These include
average-max-min consensus [12], geometric mean consensus [13], and power mean consensus [14]. A
study on the class of smooth functions that can be computed by distributed consensus algorithms was
recently addressed in [15]. Application of consensus algorithms to data fusion problem and distributed
filtering was proposed in [16, 17]. The study of the so called alignment problem, where all the agents
eventually reach an agreement, but without specifying how the final value is related to the initial
measurements, was carried out in [18]-[22]. Two recent excellent tutorials on distributed consensus
and agreement techniques are [23] and [24].
Consensus may be also seen as a form of self-synchronization among coupled dynamical systems.
In [25, 26], it was shown how to use the self-synchronization capabilities of a set of nonlinearly coupled
dynamical systems to reach the globally optimal Best Linear Unbiased Estimator (BLUE), assuming
symmetric communication links. As well known, the BLUE estimator coincides with the Maximum
Likelihood (ML) estimate for the linear observation model with additive Gaussian noise. In particular,
it was shown in [26] that reaching a consensus on the state derivative, rather than on the state itself (as
in [7]−[24]), allows for better resilience against coupling noise. In [27], the authors provided a discrete-
time decentralized consensus algorithm to compute the BLUE, based on linear coupling among the
nodes, as a result of a distributed optimization incorporating the coupling noise.
The consensus protocols proposed in [7], [9]−[21] and [24]−[27] assume that the interactions among
the nodes occur instantaneously, i.e., without any propagation delay. However, this assumption is not
valid for large scale networks, where the distances among the nodes are large enough to introduce a
nonnegligible communication delay. There are only a few recent works that study the consensus or
the agreement problem, in the presence of propagation delays, namely [8], [28]−[32], [47] focusing on
continuous-time systems, and [3, 4], [5, Ch. 7.3], [22], dealing with discrete-time protocols. More
specifically, in [8, 28] the authors provide sufficient conditions for the convergence of a linear average
consensus protocol, in the case of time-invariant homogeneous delays (i.e., equal delay for all the links)
and assuming symmetric communication links. The most appealing feature of the dynamical system
in [8] is the convergence of all the state variables to a known, delay-independent, function (equal to
the average) of the observations. However, this desired property is paid in terms of convergence, since,
in the presence of homogeneous delays, the system of [8] is able to reach a consensus if and only if
the (common) delay is smaller than a given, topology-dependent, value. Moreover, the assumption
of homogeneous delays and symmetric links is not appropriate for modeling the propagation delay in
a typical wireless network, where the delays are proportional to the distance among the nodes and
communication channels are typically asymmetric.
2
The protocol of [8] was generalized in [29] to time-invariant inhomogeneous delays (but symmetric
channels) and in [30] to asymmetric channels. The dynamical systems studied in [29, 30] are guaranteed
to reach an agreement, for any given set of finite propagation delays, provided that the network
is strongly connected. Similar results, under weaker (sufficient) conditions on the (possibly time-
varying) network topology, were obtained in [3, 4], [5, Ch.7.3] and [22] for the convergence of discrete-
time asynchronous agreement algorithms. However, since the final agreement value is not a known
function of the local measurements, agreement algorithms proposed in the cited papers are mostly
appropriate for the alignment of mobile agents, but they cannot be immediately used to distributively
compute prescribed functions of the sensors' measurements, like decision tests or global parameter
estimates. In [47], the authors studied the convergence properties of the agreement algorithms proposed
in [29]−[31] and provided a closed form expression of the achievable consensus in the case of strongly
connected balanced digraphs, under the assumption that the initial conditions are nonnegative and the
state trajectories remain in the nonnegative orthant of the state space. However, the final agreement
value depends on delays, network topology and initial conditions of each node, so that the bias is
practically unavoidable.
In summary, in the presence of propagation delays, classical distributed
protocols reaching consensus/agreement on the state [6]−[24], [27]−[31], [47] cannot be used to achieve
prescribed functions of the sensors' measurements that are not biased by the channel parameters.
Ideally, we would like to have a totally decentralized system able to reach a global consensus on
a final value which is a known, delay-independent, function of the local measurements (as in [8]), for
any given set of inhomogeneous propagation delays (as in [29]) and asymmetric channels (as in [30]).
In this paper, we propose a distributed dynamical system having all the above desired features. More
specifically, we consider a set of linearly coupled first-order dynamical systems and we fully characterize
its convergence properties, for a network with arbitrary topology (not necessarily strongly connected,
as opposed to [29]−[31]) and (possibly) asymmetric communication channels. In particular, we consider
an interaction model among the sensors that is directly related to the physical channel parameters.
The network is modeled as a weighted directed graph, whose weights are directly related to the (flat-
fading) channel coefficients between the nodes and to the transmit power. Furthermore, the geometry-
dependent propagation delays between each pair of nodes, as well as possible time offsets among the
nodes, are properly taken into account. The most appealing feature of the proposed system is that a
consensus on a globally optimal decision statistic is achieved, for any (bounded) set of inhomogeneous
delays and for any set of (asymmetric) communication channels, with the only requirement that the
network be quasi-strongly connected and the channel coefficients be nonnegative.1
In particular, our main contributions are the following: i) We provide the necessary and sufficient
conditions ensuring local or global convergence, for any set of finite propagation delays and network
topology; ii) We prove that the convergence is exponential, with convergence rate depending, in general,
on the channel parameters and propagation delays; iii) We derive a closed form expression for the final
consensus, as a function of the attenuation coefficients and propagation delays of each link; iv) We
show how to get a final, unbiased function of the sensor's measurements, which coincides with the
globally optimal decision statistics that it would have been computed by a fusion center having error-
1This last requirement, if not immediately satisfied, requires some form of phase compensation at the receiver.
3
free access to all the nodes. The paper is organized as follows. Section 2 describes the proposed
first-order linearly coupled dynamical system and shows how to design the system's parameters and
the local processing so that the state derivative of each node converges, asymptotically, to the globally
optimal decision statistics. Section 3 contains the main results of the paper, namely the necessary and
sufficient conditions ensuring the global or local convergence of the proposed dynamical system, in the
presence of inhomogeneous propagation delays and asymmetric channels. Finally, Section 4 contains
numerical results validating our theoretical findings and draws some conclusions.
2 Reaching Globally Optimal Decisions Through Self-Synchronization
It was recently shown in [33] that, in many applications, an efficient sensor network design should
incorporate some sort of in-network processing. In this paper, we show first a class of functions that
can be computed with a totally distributed approach. Then, we illustrate the distributed mechanism
able to achieve the globally optimal decision tests.
2.1 Forms of Consensus Achievable With a Decentralized Approach
If we denote by yi, i = 1, . . . , N , the (scalar) measurement taken from node i, in a network composed
of N nodes, we have shown in [25, 26] that it is possible to compute any function of the collected data
expressible in the form
f (y1, y2, . . . , yN ) = hXN
XN
i=1
ci
i=1
cigi(yi)
,
where {ci} are positive coefficients and gi(·), i = 1, . . . , N, and h(·) are arbitrary (possibly nonlinear)
real functions on R, i.e., gi, h : R 7→ R, in a totally decentralized way, i.e., without the need of a sink
node. In the vector observation case, the function may be generalized to the vector form
(1)
(2)
f (y1, y2, . . . , yN ) = h NXi=1
Ci!−1 NXi=1
Cigi(yi)! ,
k=1 is the vector containing the observations {yi,k}L
where yi = {yi,k}L
k=1 taken be sensor i, and gi(·)
and h(·) are arbitrary (possibly nonlinear) real functions on RL, i.e., gi, h : RL 7→ R, and {Ci} are
arbitrary square positive definite matrices.
Even though the class of functions expressible in the form (1) or (2) is not the most general one,
nevertheless it includes many cases of practical interest, as shown in the following examples.
Example 1: ML or BLUE estimate. Let us consider the case where each sensor observes a vector
in the form
yi = Aiξ + wi,
i = 1, . . . , N,
(3)
where yi is the M × 1 observation vector, ξ is the L × 1 unknown common parameter vector, Ai is
the M × L mixing matrix of sensor i, and wi is the observation noise vector, modeled as a circularly
symmetric Gaussian vector with zero mean and covariance matrix Ri. We assume that the noise
vectors affecting different sensors are statistically independent of each other, and that each matrix
4
Ai is full column rank, which implies M ≥ L. As well known, in this case the globally optimal ML
estimate of ξ is [34]:
ξM L = f (y1, y2, . . . , yN ) = NXi=1
AT
i R−1
i Ai!−1 NXi=1
AT
i R−1
i yi! .
(4)
i Ai)−1AT
This expression is a special case of (2), with Ci = AT
If the noise pdf is unknown, (4) still represents a meaningful estimator, as it is the BLUE [34].
i Ai and gi(yi) = (AT
i R−1
i R−1
i R−1
i yi.
Example 2: Detection of a Gaussian process with unknown variance embedded in Gaus-
sian noise with known variance. Let us consider now a detection problem. Let yi[k] denote the
signal observed by sensor i, at time k. The detection problem can be cast as a binary hypothesis test,
where the two hypotheses are
H0 :
H1 :
yi[k] = wi[k],
yi[k] = si[k] + wi[k],
i = 1, . . . , N,
k = 1, . . . , K,
(5)
where si[k] is the useful signal and wi[k] is the additive noise. Let us consider the case where the
random sequences si[k] and wi[k] are spatially uncorrelated and modeled as zero mean independent
Gaussian random processes. A meaningful model consists in assuming that the noise variance is known,
let us say equal to σ2
w, whereas the useful signal variance is not. Under these assumptions, the optimal
detector consists in computing the generalized likelihood ratio test (GLRT) [34, 35]
T (y) = f (y1, y2, . . . , yN ) =
1
K
NXi=1
KXk=1
i [k] 1
σ2
w
y2
−
w! −
w ! ,
log bPi + σ2
σ2
w
NXi=1
(6)
and comparing it with a threshold that depends on the desired false alarm rate. In (6), the term bPi
denotes the ML estimate of the signal power at node i, given by
1
bPi + σ2
bPi = 1
K
,
y2
i [k] − σ2
w!+
KXk=1
w (cid:17) [35].
w(cid:17) − log(cid:16) bPi+σ2
σ2
where (x)+ , max (0, x) . Also in this case, it is easy to check that (6) is a special case of (1), with
ci = 1 and gi(yi) = 1
− 1
w
k=1 y2
bPi+σ2
These are only two examples, but the expressions (1) or (2) can be used to compute or approximate
KPK
i [k](cid:16) 1
σ2
w
more general functions of the collected data, like, maximum and minimum [12], geometric mean [13],
power mean [14] and so on, through appropriate choice of the parameters involved.
2.2 How to Achieve the Consensus in a Decentralized Way
The next question is how to achieve the aforementioned optimal decision statistics with a network
having no fusion center. In [25, 26] we proposed an approach to solve this problem using a nonlinear
interaction model among the nodes, based on an undirected graph and with no propagation delays in
the exchange of information among the sensors. In this paper, we consider a linear interaction model,
but we generalize the approach to a network where the propagation delays are taken into account and
the network is described by a weighted directed graph (or digraph, for short), which is a fairly general
5
model to capture the non reciprocity of the communication links governing the interaction among the
nodes.
The proposed sensor network is composed of N nodes, each equipped with four basic components:
i) a transducer that senses the physical parameter of interest (e.g., temperature, concentration of
contaminants, radiation, etc.); ii) a local processing unit that processes the measurement taken by the
node; iii) a dynamical system, whose state is initialized with the local measurements and it evolves
interactively with the states of nearby sensors; iv) a radio interface that transmits the state of the
dynamical system and receives the state transmitted by the other nodes, thus ensuring the interaction
among the nodes2.
Scalar observations. In the scalar observation case, the dynamical system present in node i evolves
according to the following linear functional differential equation
K
ci Xj∈Ni
t ∈ [−τ , 0],
aij (xj(t − τ ij; y) − xi(t; y)) ,
t > 0,
i = 1, . . . , N,
(7)
xi(t; y) = gi(yi) +
xi(t; y) =eφi(t),
where xi(t; y) is the state function associated to the i-th sensor that depends on the set of measurements
y = {yi}N
i=1; gi(yi) is a function of the local observation yi, whose form depends on the specific decision
test; K is a positive coefficient measuring the global coupling strength; ci is a positive coefficient that
may be adjusted to achieve the desired consensus; τ ij = Tij + dij/c is a delay incorporating the
propagation delay due to traveling the internode distance dij, at the speed of light c, plus a possible
time offset Tij between nodes i and j. The sensors are assumed to be fixed so that all the delays
are constant. We also assume, realistically, that the maximum delay is bounded, with maximum
signal received from node i and transmitted from node j. In such a case, we have aij =qPjhij2/dη
value τ = maxij τ ij. The coefficients aij are assumed to be nonnegative and, in general, dependent
on transmit powers and channel parameters. For example, aij may represent the amplitude of the
ij,
where Pj is the power of the signal transmitted from node j; hij is a fading coefficient describing the
channel between nodes j and i; η is the path loss exponent. The nonnegativity of aij requires some
form of channel compensation at the receiver side, like in the maximal ratio receiver, for example, if
the channel coefficients {hij}ij are complex variables. Furthermore, we assume, realistically, that node
i "hears" node j only if the power received from i exceeds a given threshold. In such a case, aij 6= 0,
otherwise aij = 0. The set of nodes that sensor i hears is denoted by Ni = {j = 1, . . . , N : aij 6= 0}.
Observe that, in general, aij 6= aji, i.e., the channels are asymmetric. It is worth noticing that the
state function of, let us say, node i depends, directly, only on the measurement yi taken by the node
itself and only indirectly on the measurements gathered by the other nodes.
In other words, even
though the state xi(t; y) gets to depend, eventually, on all the measurements, through the interaction
with the other nodes, each node needs to know only its own measurement.
Because of the delays, the state evolution (7) for, let us say, t > 0, is uniquely defined provided
2The state value is transmitted by modulating an appropriate carrier. Because of space limitations, it is not possible
to go into the details of this aspect in this paper. In parallel works, we have shown that a pulse position modulation of
ultrawideband signals may be a valid candidate for implementing the interaction mechanism. Some preliminary remarks
on the implementation of the proposed protocol can be found in [38].
6
that the initial state variables xi(t; y) are specified in the interval from −τ to 0. The initial conditions
of (7) are assumed to be taken in the set of continuously differentiable and bounded functions eφi(t)
mapping the interval [−τ , 0] to R (see Appendix B for more details).
Vector observations. If each sensor measures a set of, let us say, L physical parameters , the coupling
mechanism (7) generalizes into the following expression3
aij (xj(t − τ ij; y) − xi(t; y)) ,
t > 0,
i = 1, . . . , N,
(8)
xi(t; y) = gi(yi) + K Q−1
i Xj∈Ni
t ∈ [−τ , 0],
xi(t; y) =eφi(t),
where xi(t) is the L-size vector state of the i-th node; gi(yi) is a vector function of the local observation
k=1, i.e. gi : RL 7→ RL; and Qi is an L × L non-singular matrix that is a free parameter to
yi = {yi,k}L
be chosen according to the specific purpose of the sensor network.
As in (7), the initial conditions of (8) are assumed to be arbitrarily taken in the set of continuously
differentiable and bounded (vectorial) functions eφi(t) mapping the interval [−τ , 0] to RL, w.l.o.g..
2.3 Self-Synchronization
Differently from most papers dealing with average consensus problems [6]−[24] and [27]−[31], where
the global consensus was intended to be the situation where all dynamical systems reach the same
state value, we adopt here the alternative definition already introduced in our previous work [26]. We
define the consensus (through network synchronization) with respect to the state derivative, rather
than to the state, as follows.
Definition 1 Given the dynamical system in (7) (or (8)), a solution {x⋆
to be a synchronized state of the system, if
i (t; y)} of (7) (or (8)) is said
x⋆
i (t; y) = ω⋆(y),
∀i = 1, 2, . . . , N.
(9)
The system (7) (or (8)) is said to globally synchronize if there exists a synchronized state as in (9),
and all the state derivatives asymptotically converge to this common value, for any given set of initial
conditions {eφi}, i.e.,
k xi(t; y) − ω⋆(y)k = 0,
lim
t7→∞
∀eφi, i = 1, 2, . . . , N,
(10)
where k·k denotes some vector norm and {xi(t; y)} is a solution of (7) (or (8)). The synchronized state
is said to be globally asymptotically stable if the system globally synchronizes, in the sense specified in
(10). The system (7) (or (8)) is said to locally synchronize if there exist disjoint subsets of the nodes,
called clusters, where the nodes in each cluster have state derivatives converging, asymptotically, to the
same value, for any given set of initial conditions {eφi}.
Observe that, according to Definition 1, if there exists a globally asymptotically stable synchronized
state, then it must necessarily be unique (in the derivative). In the case of local synchronization instead,
the system may have multiple synchronized clusters, each of them with a different synchronized state.
3We assume that the coupling coefficients aij are the same for all estimated parameters. This assumption is justified
by the fact that all the parameters are sent through the same physical channel.
7
As it will be shown in the next section, one of the reasons to introduce a novel definition of consensus,
different from the classical one [6]−[24], [27]−[31], is that the convergence on the state derivative, rather
than on the state, is not affected by the presence of propagation delays and there is a way to make the
final consensus value to coincide with the globally optimal decision statistic. In the ensuing sections,
we will provide necessary and sufficient conditions for the system in (7) (or (8)) to locally/globally
synchronize according to Definition 1, along with the closed form expression of the synchronized state.
3 Necessary and Sufficient Conditions for Self-Synchronization
The problem we address now is to check if, in the presence of propagation delays and asymmetric
communication links, the systems (7) and (8) can still be used to achieve globally optimal decision
statistics in the form (1) and (2), in a totally distributed way. To derive our main results, we rely on
some basic notions of digraph theory. To make the paper self-contained, in Appendix A.1 we recall
the basic definitions of weak, quasi-strong and strong connectivity (WC, QSC, and SC, for short) of
a digraph. In Appendix A.2, we recall the algebraic properties of the Laplacian matrix associated to
a digraph and we derive the properties of its left eigenvectors, as they will play a fundamental role
in computing the achievable forms of consensus. In Appendix B, we provide sufficient conditions for
the marginal stability of linear delayed differential equations, as they will be instrumental to prove the
main results of this paper.
The next theorem provides necessary and sufficient conditions for the proposed decentralized ap-
proach to achieve the desired consensus in the presence of propagation delays and asymmetric com-
munication links.
Theorem 1 Let G ={V , E } be the digraph associated to the network in (7), with Laplacian matrix L.
Let γ = [γ1, . . . , γN ]T be a left eigenvector of L corresponding to the zero eigenvalue, i.e., γT L = 0T
N .
Given the system in (7), assume that the following conditions are satisfied:
a1 The coupling gain K and the coefficients {ci} are positive; the coefficients {aij} are non-negative;
a2 The propagation delays {τ ij} are time-invariant and finite, i.e., τ ij ≤ τ = maxi6=j τ ij < +∞,
∀i 6= j;
a3 The initial conditions are taken in the set of continuously differentiable and bounded functions
mapping the interval [−τ , 0] to RN .
Then, system (7) globally synchronizes for any given set of propagation delays, if and only if the
digraph G is Quasi-Strongly Connected (QSC). The synchronized state is given by
ω⋆(y) =
,
γiaijτ ij
(11)
XN
i=1
γicigi(yi)
XN
i=1
γici + KXN
i=1Xj∈Ni
where γi > 0 if and only if node i can reach all the other nodes of the digraph through a strong path,
or γi = 0, otherwise.
The convergence is exponential, with rate arbitrarily close to r , − mini{Re{si}
: p(si) =
0 and si 6= 0}, where p(s) is the characteristic function associated to system (7) (cf. Appendix B).
8
Proof. See Appendix C.
Theorem 1 has a very broad applicability, as it does not make any particular reference to the
network topology. If, conversely, the topology has a specific structure, then we may have the following
forms of consensus.4
Corollary 1 Given system (7), assume that conditions a1)-a3) of Theorem 1 are satisfied. Then,
1. The system globally synchronizes and the synchronized state is given by
ω⋆(y) = gr(yr),
r ∈ {1, 2, . . . , N },
(12)
if and only if the digraph G contains only one spanning directed tree, with root node given by
node r.
2. The system globally synchronizes and the synchronized state is given by (11) with all γi's positive
if and only if the digraph G is Strongly Connected (SC). The synchronized state becomes
ω⋆(y) =
i=1
XN
ci + KXN
cigi(yi)
i=1Xj∈Ni
XN
i=1
if and only if, in addition, the digraph G is balanced.
,
aijτ ij
(13)
3. The system locally synchronizes in K disjoint clusters C1, . . . , CK ⊆ {1, . . . , N },5 with synchro-
nized state derivatives for each cluster
x⋆
q(t; yk) =
γicigi(yi)
Xi∈Ck
γici + KXi∈CkXj∈Ni
Xi∈Ck
,
∀q ∈ Ck,
k = 1, . . . , K,
(14)
γiaijτ ij
with yk = {yi}i∈Ck , if and only if the digraph G is weakly connected (WC) and contains a spanning
directed forest with K Root Strongly Connected Components (RSCC).6
The proof of this corollary is a particular case of Theorem 1, except that we exploit the structure
of the left eigenvector corresponding to the zero eigenvalue of the Laplacian matrix L, as derived
in Appendix A.2. The previous results can be extended to the vector case in (8), according to the
following.7
Theorem 2 Given the system (8), assume that conditions a2)-a3) of Theorem 1 are satisfied and that
the matrices {Qi} are positive definite. Then, the system synchronizes for any given set of propagation
delays, if and only if the digraph G is QSC. The synchronized state is given by
ω⋆(y) ,
NXi=1
γiQi + IL ⊗K
NXi=1
NXj∈Ni
γiaijτ ij
−1 NXi=1
γiQigi(yi)! ,
(15)
4We focus, w.l.o.g., only on WC digraphs.
In the case of non WC digraphs, Corollary 1 applies to each disjoint
component of the digraph.
5In general, the clusters C1, . . . , CK are not a partition of the set of nodes {1, · · · , N }.
6Please, see Appendix A.1 for the formal definition of (root) strongly connected component.
7The proof is omitted because of space limitations, but it follows the same guidelines as the proof for the scalar case,
since the interaction among the nodes is the same.
9
where ⊗ denotes the Kronecker product and γi > 0 if and only if node i can reach all the other
nodes of the digraph by a strong path, or γi = 0, otherwise. The convergence is exponential, with
rate arbitrarily close to r , − mini{Re{si} : p(si) = 0 and si 6= 0}, where p(s) is the characteristic
function associated to system (8).
3.1
Impact of propagation delays on convergence
The impact of delays in consensus-achieving algorithms has been analyzed in a series of works [5,
Ch.7.3], [8], [22], [28]−[31]. Among these works, it is useful to distinguish between consensus al-
gorithms, [8, 23], where the states of all the sensors converge to a prescribed function (typically the
average) of the sensors' initial values, and agreement algorithms, [5, Ch.7.3], [22], [28]−[31], [47], where
the goal is to make all the states to converge to a common value, but without specifying how this value
has to be related to the initial values. In our application, we can only rely on consensus algorithms,
where the final consensus has to coincide with the globally optimal decision statistic.
The consensus algorithms analyzed in [8, 23] assume the same delay value for all the links, i.e.,
τ ij = τ , and symmetric channels, i.e., aij = aji, for all i 6= j. Under these assumptions, the average
consensus in [8, 23] is reached if and only if the common delay τ is upper bounded by [8]
τ <
π
2
1
λN
,
(16)
where λN denotes the maximum eigenvalue of the Laplacian associated to the undirected graph of the
network (cf. Appendix A). Since, for any graph with N nodes, we have [39]
N
N − 1
dmax ≤ λN ≤ 2 dmax,
(17)
with dmax denoting the maximum graph degree, increasing dmax imposes a strong constraint on the
maximum (common) tolerable delay. This implies, for example, that networks with hubs (i.e., nodes
with very large degrees) that are commonly encountered in scale-free networks [36], are fragile against
propagation delays, when using the consensus algorithms of [8, 23].
In our application, we were motivated to extend the approach of [8, 23] to the general case of
inhomogeneous delays and asymmetric channel links. Nevertheless, in spite of the less restrictive
assumptions, Theorem 1 shows that our proposed algorithm is more robust against propagation delays,
since its convergence capability is not affected by the delays. Moreover, our approach is valid in the
more general case of asymmetric communications links and the final value is not simply the average
of the measurements, but a weighted average of functions of the measurements that can be made to
coincide with the desired globally optimal decision statistic in the form (1) or (2), through a proper
choice of the coefficients ci.
An intuitive reason for explaining the main advantage of our approach is related to the use of an
alternative definition of global consensus: As opposed to conventional methods requiring consensus on
the state value, i.e.
[6]−[24], [28]−[31], we require the convergence over the state derivative, i.e., we
only require that the state trajectories converge towards parallel straight lines. The slope must be the
same for all the trajectories and it has to coincide with the desired decision statistic. But the constant
terms of each line may differ from sensor to sensor. This provides additional degrees of freedom that,
eventually, make our approach more robust against propagation delays or link coefficients.
10
3.2 Effect of network topology on final consensus value
Theorem 1 generalizes all the previous (only sufficient) conditions known in the literature [8], [28]−
[31] for the convergence of linear agreement/consensus protocols in the presence of propagation delays,
since it provides a complete characterization of the synchronization capability of the system for any
possible degree of connectivity in the network (not only for SC digraphs), as detailed next.
In general, the digraph modeling the interaction among the nodes may have one of the following
structures: i) the digraph contains only one spanning directed tree, with a single root node, i.e., there
exists only one node that can reach all the other nodes in the network through a strong directed path;
ii) the digraph contains more than one spanning directed tree, i.e., there exist multiple nodes (possibly
all the nodes), strongly connected to each other, that can reach all the other nodes by a strong directed
path; iii) the digraph is weakly connected and contains a spanning forest, i.e., there exists no node
that can reach every other node through a strong directed path; iv) the digraph is not even weakly
connected. The last case is the least interesting, as it corresponds to a set of isolated subnetworks,
that can be analyzed independently of each other, as one of the previous cases. In the first two cases,
according to Theorem 1, system (7) achieves a global consensus, whereas in the third case the system
forms clusters of consensus with, in general, different consensus values in each cluster, i.e., the system
synchronizes only locally.
In other words, a global consensus is possible if and only if there exists at least one node (the root
node of the spanning directed tree of the digraph) that can send its information, directly or indirectly,
to all the other nodes. If no such a node exists, a global consensus cannot be reached. However, a local
consensus is still achievable among all the nodes that are able to influence each other.
Interestingly, the closed form expressions of the synchronized state given by (11) confirms the above
statements: The observation yi of, let us say, sensor i affects the final consensus value if and only if
such an information can reach all the other nodes by a strong directed path. As a by-product of this
result, we have the following special cases (Corollary 1): If there is only one node that can reach all the
others, then the final consensus depends only on the observation taken from that node; on the other
extreme, the final consensus contains contributions from all the nodes if and only if the digraph is SC.
Moreover, if a node contributes to the final value, it does that through a weight that depends on
its in/out-degree. The set of weights {γi} in (11) can be interpreted as a measure of the "symme-
try/asymmetry" of the communication links in the network: Some of these weights are equal to each
other if and only if the subdigraph associated to the corresponding nodes is balanced (or undirected).
The synchronized state, as given in (11), suggests also an interesting interpretation of the consensus
formation mechanism of system (7), based on the so called condensation digraph.8 From (11) in fact,
one infers that all the nodes that are SC to each other (usually referred to as nodes of a strongly
connected component (SCC)) produce the same effect on the final consensus as an equivalent single
node that represents the consensus within that SCC. In fact, one may easily check if system in (7)
locally or globally synchronizes and which nodes contribute on the consensus, simply reducing the
original digraph to its equivalent condensation digraph and looking for the existence of a spanning
8Please, refer to Appendix A.1 for the definition of condensation digraph and the procedure for reducing a digraph
into its condensation digraph.
11
directed tree in the condensation digraph. The only SCCs of the original digraph that will provide
a contribution on the final consensus are the SCCs associated to the root nodes of the condensation
digraph.
As an additional remark, the possibility to form clusters of consensus, rather than a global con-
sensus, depends on the (channel) coefficients aij.
given in Section 2.2, they may be altered by changing the transmit powers Pj. As a consequence, the
nodes with the highest transmit power will be the most influential ones. If, for example, we want to
If, e.g., these coefficients have the expression as
set a certain parameter on each node, like, e.g., a decision threshold, we can use the same consensus
mechanism used in this paper by assigning, for example, the desired value to, let us say, node i, and
select the transmit powers so that node i is the only node that can reach every other node.
3.3 How to get unbiased estimates
The closed form expression of the synchronized state, as given in (11) (or in (15)), is valid for any given
digraph (not only for undirected graphs as in [8]). Expression (11) shows a dependence of the final
consensus on the network topology and propagation parameters, through the coefficients {aij}, {γi}
and the delays {τ ij}. This means that, even if the propagation delays do not affect the convergence
of the proposed system, they introduce a bias on the final value, whose amount depends on both the
delays and coefficients {aij}. The effect of propagation parameters and network topology on the final
synchronized state is also contained in the eigenvector γ of the Laplacian L. This implies that the
final consensus resulting from (11) cannot be made to coincide with the desired decision statistics as
given by (1), except that in the trivial case where all the delays are equal to zero and the digraph is
balanced (and thus strongly connected). However, expression (11) suggests a method to get rid of any
bias, as detailed next.
The bias due to the propagation delays can be removed using the following two-step algorithm.9
We let the system in (7) to evolve twice: The first time, the system evolves according to (7) and we
denote by ω⋆(y) the synchronized state; the second time, we set gi(yi) = 1 in (7), for all i, and the
system is let to evolve again, denoting the final synchronized state by ω⋆(1). From (11), if we take the
ratio ω⋆(y)/ω⋆(1), we get
which coincides with the ideal value achievable in the absence of delays. Thus, this simple double-step
algorithm allows us to remove the bias term depending on the delays and on the channel coefficients,
without requiring the knowledge or estimate of neither set of parameters.
If the network is strongly connected and balanced, γi = 1, ∀i and then the compensated consen-
sus coincides with the desired value (1).
If the network is unbalanced, the compensated consensus
ω⋆(y)/ω⋆(1) does not depend on the (channel) coefficients {aij}, but it is still biased, with a bias
dependent on γ, i.e., on the network topology. Nevertheless, this residual bias can be eliminated in
a decentralized way according to the following iterative algorithm. Let us denote by Nr the number
of nodes in the RSCC of the digraph. At the beginning, every node sets gi(yi) = 1 and ci = 1,
9We focus only on the scalar system in (7), because of space limitation.
12
ω⋆(y)
ω⋆(1)
i=1
= XN
XN
γicigi(yi)
γici
i=1
,
(18)
i = 1, . . . , N and the network is let to evolve. The final consensus value is denoted by ω⋆(1). Then, the
network is let to evolve Nr times, according to the following protocol. At step i, with i = 1, . . . , Nr,
the nodes within the i-th SCC, set gi(yi) = 1, while all the other nodes set gk(yk) = 0 for all k 6= i. Let
us denote by ω⋆(ei) the final consensus value, where ei is the canonical vector having all zeros, except
the i-th component, equal to one. Repeating this procedure for all the SCC's, at the end of the Nr
steps, each node is now able to compute the ratio ω⋆(ei)/ω⋆(1), which coincides with γi := γi/Pk γk.
Thus, after Nr + 1 steps, every node knows its own (normalized) γi. This value is subsequently used to
compensate for the network unbalance as follows. The compensation is achieved by simply setting, at
each node ci = ci/γi. In fact, with this setting, the final ratio ω⋆(y)/ω⋆(1) coincides with the desired
unbiased expression given by (1), for all i such that γi 6= 0. If the digraph G is SC, this procedure
corresponds indeed to make the Laplacian matrix L(G ) balanced, so that the left eigenvector of L
associated to the zero eigenvalue be proportional to the vector 1N . This procedure only needs some
kind of coordination among the nodes to make them working according to a described scheduling. It
is important to remark that, since the eigenvector γ does not depend on the observations {yi}, the
proposed algorithm is required to be performed at the start-up phase of the network and repeated
only if the network topology or the channels change through time. In summary, we can eliminate the
effect of both delays and channel parameters on the final consensus value, thus achieving the optimal
decision statistics as given in (1), with a totally decentralized algorithm, at the price of a slight increase
of complexity and the need for some coordination among the nodes.10
3.4 Asymptotic convergence rate
Theorem 1 generalizes previous results on the convergence speed of classical linear consensus protocols
(see, e.g.
[8, 23]) to the case in which there are propagation delays.
In spite of the presence of
delays, the proposed system still converges to the consensus with exponential convergence rate, i.e.,
k x(t; y) − ω⋆(y)k ≤ O(cid:0)ert(cid:1), where the convergence factor r < 0 is defined in Theorem 1. Thus,
the convergence speed is dominated by the slowest "mode" of the system. Moreover, as expected,
the presence of delays affects the convergence speed, as the roots of characteristic equation p(s) =
0 associated to system (7) depend, in general, on both network topology and propagation delays.
Unfortunately, a closed form expression of this dependence is not easy to derive, and the roots of
p(s) = 0 need to be computed numerically.
In the special case of negligible delays instead, we can provide a bound of the convergence factor as a
function of the channel parameters through the eigenvalues of the Laplacian matrix L. Setting τ ij = 0
in (7), the characteristic equation associated to (7) becomes p(s) = sI + L = 0, whose solutions are
just the eigenvalues of −L. Under the assumptions of Theorem 1, it follows that
r = − min
{Re{λi} : λi ∈ σ{L}, and λi 6= 0},
i
(19)
where σ{L} denotes the spectrum of Laplacian matrix L. The value of r is negative if and only if
the digraph G is QSC [45, Lemma 2]. Using (19), we can obtain bounds on the convergence rate as a
10In [47], the authors provided a closed form expression of the agreement obtainable with an SC balanced network
achieving consensus on the state. However, that expression depends on the channel parameters and on the initial
conditions in a way that the bias cannot be eliminated through a distributed procedure.
13
function of the network topology, as shown next. When the digraph is strongly connected we have11
r ≤ κ , −λ2(cid:18) 1
2
(DγL + LT Dγ)(cid:19) < 0,
(20)
where Dγ , diag(γ1, · · · , γN ), γ is the left eigenvector of L associated to the zero eigenvalue, normal-
ized so that kγk∞ = 1, and λ2(A) denotes the second smallest eigenvalue of the symmetric matrix A.
It follows from (20) that system (7) reaches a consensus with rate at least κ. As special cases of (20)
we have the following: i) In the case of balanced digraphs, Dγ = I and thus κ = −λ2(cid:0) 1
If the digraph is undirected (and connected) Dγ = I, L = LT and thus κ = −λ2 (L), where λ2 (L) is
also known as the algebraic connectivity of the digraph [39]. For the case in which the digraph is QSC,
2 (L + LT )(cid:1); ii)
some bounds of r can be found in [44, 46] using the generalization of the classical definition of alge-
braic connectivity. Moreover, interestingly, the convergence rate of the system, under the assumption
of Theorem 1, can be related to the convergence rates of the SCCs of condensation digraph associated
to the system. Because of lack of space, we suggest the interested reader to check [45] for more details.
In conclusion, according to the above results, we infer that the convergence rate of the proposed
consensus algorithm, at least in the absence of propagation delays, is the same as that of the classical
linear protocols achieving consensus on the state.
4 Numerical Results
In this section, we illustrate first some examples of consensus, for different network topologies. Then, we
show an application of the proposed technique to an estimation problem, in the presence of random link
coefficients. In both examples, the analog system (7) is implemented in discrete time, with sampling
step size Ts = 10−3.
Example 1: Different forms of consensus for different topologies
In Figure 1, we consider three topologies (top row), namely: (a) a SC digraph, (b) a QSC digraph
with three SCCs, and (c) a WC (not QSC) digraph with a spanning forest composed by two trees. For
each digraph, we also sketch its decomposition into SCCs (each one enclosed in a circle), corresponding
to the nodes of the associated condensation digraph (whose root SCC is denoted by RSCC). In the
bottom row of Figure 1, we plot the dynamical evolutions of the state derivatives of system (7) versus
time, for the three network topologies, together with the theoretical asymptotic values predicted by
(11) (dashed line with arrows). As proved by Theorem 1, the dynamical system in Figure 1a) achieves
a global consensus, since the underlying digraph is SC. The network of Figure 1b), instead, is not
SC, but the system is still able to globally synchronize, since there is a set of nodes, in the RSCC
component, able to reach all the other nodes. The final consensus, in such a case, contains only
the contributions of the nodes in the RSCC, since no other node belongs to the root of a spanning
directed tree of the condensation digraph. Finally, the system in Figure 1c) cannot reach a global
consensus since there is no node that can reach all the others, but it does admit two disjoint clusters,
corresponding to the two RSCCs, namely RSCC1 and RSCC2. The middle lines of Figure 1c) refer to
the nodes of the SCC component, not belonging to either RSCC1 or RSCC2, that are affected by the
11The inequality in (20) follows from the fact that Re{λi} ≥ κ for all nonzero eigenvalues {λi} of L [45].
14
consensus achieved in the two RSCC components, but that cannot affect them. Observe that, in all
the cases, the state derivatives of the (global or local) clusters converge to the values predicted by the
closed form expression given in (11), (12) or (14), depending on the network topology.
2(cid:13)
1(cid:13)
13(cid:13)
12(cid:13)
11(cid:13)
3(cid:13)
5(cid:13)
4(cid:13)
14(cid:13)
10(cid:13)
6(cid:13)
7(cid:13)
8(cid:13)
9(cid:13)
RSCC(cid:13)
3(cid:13)
5(cid:13)
9(cid:13)
2(cid:13)
1(cid:13)
11(cid:13)
4(cid:13)
14(cid:13)
10(cid:13)
SCC(cid:13) 2(cid:13)
13(cid:13)
12(cid:13)
SCC(cid:13) 1(cid:13)
6(cid:13)
7(cid:13)
8(cid:13)
RSCC(cid:13) 1(cid:13)
3(cid:13)
RSCC(cid:13) 2(cid:13)
6(cid:13)
5(cid:13)
9(cid:13)
7(cid:13)
8(cid:13)
4(cid:13)
14(cid:13)
10(cid:13)
2(cid:13)
1(cid:13)
11(cid:13)
SCC(cid:13)
13(cid:13)
12(cid:13)
110(cid:13)
100(cid:13)
.(cid:13)
_(cid:13)x(cid:13) i(cid:13) ((cid:13)t(cid:13))(cid:13)
90(cid:13)
80(cid:13)
70(cid:13)
60(cid:13)
50(cid:13)
0(cid:13)
200(cid:13)
400(cid:13)
i(cid:13)t(cid:13)e(cid:13)r(cid:13)a(cid:13)t(cid:13)i(cid:13)o(cid:13)n(cid:13) i(cid:13)n(cid:13)d(cid:13)e(cid:13)x(cid:13)
(a)(cid:13)
600(cid:13)
800(cid:13)
1000(cid:13)
.(cid:13)
i(cid:13) ((cid:13)t(cid:13))(cid:13)
_(cid:13)x(cid:13)
110(cid:13)
100(cid:13)
90(cid:13)
80(cid:13)
70(cid:13)
60(cid:13)
50(cid:13)
0(cid:13)
.(cid:13)
_(cid:13)x(cid:13)
i(cid:13) ((cid:13)t(cid:13))(cid:13)
110(cid:13)
100(cid:13)
90(cid:13)
80(cid:13)
70(cid:13)
60(cid:13)
RSCC(cid:13) 1(cid:13)
RSCC(cid:13) 2(cid:13)
600(cid:13)
800(cid:13)
1000(cid:13)
50(cid:13)
0(cid:13)
200(cid:13)
200(cid:13)
400(cid:13)
i(cid:13)t(cid:13)e(cid:13)r(cid:13)a(cid:13)t(cid:13)i(cid:13)o(cid:13)n(cid:13)i(cid:13)n(cid:13)d(cid:13)e(cid:13)x(cid:13)
(b)(cid:13)
600(cid:13)
800(cid:13)
1000(cid:13)
400(cid:13)
i(cid:13)t(cid:13)e(cid:13)r(cid:13)a(cid:13)t(cid:13)i(cid:13)o(cid:13)n(cid:13) i(cid:13)n(cid:13)d(cid:13)e(cid:13)x(cid:13)
((cid:13)c(cid:13))(cid:13)
Figure 1: Consensus for three different network topologies: a) SC digraph; b) QSC digraph with three SCCs;
c) WC digraph with a two trees forest; Ts = 10−3, τ = 50Ts, K = 30, N=14.
Example 2: Distributed optimal decisions through self-synchronization. The behaviors
shown in the previous example refer to a given realization of the topology, with given link coefficients,
and of the observations.
In this example, we report a global parameter representing the variance
obtained in the estimate of a scalar variable. Each sensor observes a variable yi = Aiξ + wi, where
wi is additive zero mean Gaussian noise, with variance σ2
i . The goal is to estimate ξ. The estimate is
performed through the interaction system (7), with functions gi(yi) = yi/Ai and coefficients ci = A2
i /σ2
i ,
chosen in order to achieve the globally optimal ML estimate. The network is composed of 40 nodes,
randomly spaced over a square of size D. The size of the square occupied by the network is chosen in
order to have a maximum delay τ = 100Ts. We set the threshold on the amplitude of the minimum
useful signal to zero, so that, at least in principle, each node hears each other node. The corresponding
digraph is then SC. To simulate a practical scenario, the channel coefficients aij are generated as
i.i.d. Rayleigh random variables, to accommodate for channel fading. Each variable aij has a variance
depending on the distance dij between nodes i and j, equal to12 σ2
ij = Pj/(1 + d2
ij).
In Figure 2, we plot the estimated average state derivative (plus and minus the estimation standard
deviation), as a function of the iteration index. Figure 2a) refers to the case in which there is only
12We use the attenuation factor 1/(1 + d2
ij ) instead of 1/d2
ij to avoid the undesired event that, for dij < 1 the received
power might be greater than the transmitted power.
15
200
180
160
140
120
100
80
60
40
20
θ
σ
±
θ
m
SN Rm = 5dB
200
180
160
140
120
100
80
60
40
20
θ
σ
±
θ
m
SN Rm = 5dB
SN RC = 20dB
0
0
50
100
150
250
200
300
iteration index
350
400
450
500
0
0
50
100
150
250
200
300
iteration index
350
400
450
500
Figure 2: Estimated parameter vs. convergence time in the absence [subplot a)] and in the presence [subplot
(a)
(b)
b)] of receiver noise: a) ML estimate by a centralized system (dotted lines); b) estimate by the proposed method,
with no propagation delays, (dashed and dotted lines plus "×"for the average value); c) estimate by the proposed
method, with propagation delays (dashed lines plus "⊳" for the average value); d) estimate by the two-step
estimation method leading to (18) (solid lines plus "◦" for the average value).
observation noise, but there is no noise in the exchange of information among the nodes. Conversely,
Figure 2b) refers to the case where the node interaction is noisy, so that the state evolution of each
sensor is described by the state equation zi(t) = xi(t) + vi(t), with xi(t) given by (7), where vi(t) is
white Gaussian noise; the SNR is ξ2/σ2
w = 20 dB. The averages are taken across all the nodes, for 100
independent realizations of the network, where, in each realization we generated a new topology and
a new set of channel coefficients and noise terms. The results refer to the following cases of interest:
a) ML estimate achieved with a centralized system, with no communication errors between nodes and
fusion center (dotted lines); b) estimate achieved with the proposed method, with no propagation
delays, as a benchmark term (dashed and dotted lines plus × marks for the average value); c) estimate
achieved with the proposed method, in the presence of propagation delays (dashed lines plus triangles
for the average value); d) estimate achieved with the two-step estimation method leading to (18) (solid
lines plus circles for the average value). From Figure 2 we can see that, in the absence of delays, the
(decentralized) iterative algorithm based on (7) behaves, asymptotically, as the (centralized) globally
optimal ML estimator. In the presence of delays, we observe a clear bias (dashed lines), due to the
large delay values, but with a final estimation variance still close to the ML estimator's. Interestingly,
the two-step procedure leading to (18) provides results very close to the optimal ML estimator, with
no apparent bias, in spite of the large delays and the random channel fading coefficients. The only
price paid with the two-step procedure, besides time, is a slight increase of the variance due to taking
the ratio of two noisy consensus values, as evidenced in Fig. 2 b).
5 Concluding remarks
In conclusion, in this paper we have proposed a totally decentralized sensor network scheme capable
to reach globally optimal decision tests through local exchange of information among the nodes, in
16
the presence of asymmetric communication channels and inhomogeneous time-invariant propagation
delays. The method is particularly useful for applications where the goal of the network is to take
decisions about a common event. Differently from the average consensus protocols available in the
literature, our system globally synchronizes for any set of (finite) propagation delays if and only if
the underlying digraph is QSC, with a final synchronized state that is a known function of the sensor
measurements. In general, the synchronized state depends on the propagation parameters, such as
delays and communication channels. Nevertheless, exploiting our closed form expression for the final
consensus values, we have shown how to recover an unbiased estimate, for any set of delays and channel
coefficients, without the need to knowing or estimating these coefficients. This desirable result is a
distinctive property of the consensus achievable on the state derivative and cannot be obtained using
classical consensus/agreement algorithms that reach a consensus on the state variables. If we couple
the nice properties mentioned above with the properties reported in [26], where we showed that, in
the absence of delays, the consensus protocol proposed in this paper and in [26] is also robust against
coupling noise, we have, overall, a good candidate for a distributed sensor network.
As in many engineering problems, the advantages of our scheme come with their own prices. Three
issues that deserve further investigations are the following: i) the states grow linearly with time; ii) the
coefficients aij are nonnegative; and iii) a change of topology affects the convergence properties of the
proposed scheme. The first issue has an impact on the choice of the radio interface responsible for the
exchange of information between the nodes. To avoid the need for transmitting with a high dynamic
range, the nodes must transmit a nonlinear, bounded function of the state value. One possibility, as
proposed in [38], is to associate the state value to the phase of sinusoidal carrier or to the time shift of
a pulse oscillator. The second point requires that the receiver be able to compensate for possible sign
inversions. As far as the switching topology is concerned, it would be useful to devise methods aimed
at increasing the resilience of our method against topology changes. But it is worth keeping in mind
that the previous aspects are only the reverse of the medal of a method capable to achieve a globally
optimal decision for any set of delays and for asymmetric channels.
6 Appendix
A Directed Graphs
The interaction among the sensors is properly described by a directed graph. For the reader's con-
venience, in this section, we briefly review the notation and basic results of graph theory that will
be used throughout this paper. For the reader interested in a more in-depth study of this field, we
recommend, for example, [39]−[41].
A.1 Basic Definitions
To take explicitly into account the possibility of unidirectional links among the network nodes, we
represent the information topology among the nodes by their (weighted) directed graph.
Directed graph. Given N nodes, a (weighted) directed graph (or digraph) G is defined as G ={V , E },
where V , {v1, . . . , vN } is the set of nodes (or vertices) and E ⊆ V × V is a set of edges (i.e., ordered
17
pairs of the nodes), with the convention that eij , (vi, vj) ∈ E (i.e., vi and vj are the head and the
tail of the edge eij, respectively) means that the information flows from vj to vi. A digraph is weighted
if a positive weight is associate to each edge, according to a proper map W : E → R+, such that if
eij , (vi, vj) ∈ E , then W (eij ) = aij > 0, otherwise aij = 0. We focus in the following on weighted
digraphs where the weights of loops (vi, vi) are zero, i.e., aii = 0 for all i. If (vi, vj) ∈ E ⇔(vj, vi) ∈ E
(and aij = aji, ∀i 6= j), then the graph is said to be (weighted) undirected. For any node vi ∈ V , we
define the information neighbor of vi as
Ni , {j = 1, . . . , N : eij = (vi, vj) ∈ E } .
The set Ni represents the set of indices of the nodes sending data to node i.
The in-degree and out-degree of node vi ∈ V are, respectively, defined as:
degin(vi) ,
NXj=1
aij,
and
degout(vi) ,
aji.
NXj=1
(21)
(22)
Observe that for undirected graphs, degin(vi) = degout(vi).
We may have the following class of digraphs.
Balanced digraph The node vi of a digraph G ={V , E } is said to be balanced if and only if its
in-degree and out-degree coincide, i.e., degin(vi) = degout(vi). A digraph G ={V , E } is called balanced
if and only if all its nodes are balanced, i.e.,
aij =
NXj=1
NXj=1
aji,
∀i = 1, . . . , N.
(23)
Path/cycle A strong path (or directed chain) in a digraph G ={V , E } is a sequence of distinct nodes
v0, . . . , vq ∈ V such that (vi, vi−1) ∈ E , ∀i = 1, . . . , q. If v0 ≡ vq, the path is said to be closed. A weak
path is a sequence of distinct nodes v0, . . . , vq ∈ V such that either (vi−1, vi) ∈ E or (vi, vi−1) ∈ E ,
∀i = 1, . . . , q. A (strong) cycle (or circuit) is a closed (strong) path.
Directed tree/forest A digraph with N nodes is a (rooted) directed tree if it has N − 1 edges
and there exists a distinguished node, called the root node, which can reach all the other nodes by a
(unique) strong path. Thus a directed tree cannot have cycles and every node, except the root, has
one and only one incoming edge.13 A digraph is a (directed) forest if it consists of one or more directed
s} of a digraph G ={V , E }, with Vs ⊆ V and Es ⊆ E ∩ (Vs × Vs) , is a
trees. A subgraph Gs={Vs, E
spanning directed tree (or a spanning directed forest), if it is a directed tree (or a directed forest) and
it has the same node set as G , i.e., Vs ≡ V . We say that a digraph G ={V , E } contains a spanning tree
(or a spanning forest) if there exists a subgraph of G that is a spanning directed tree (or a spanning
directed forest).
Connectivity. In a digraph there are many degrees of connectedness. In this paper we focus on the
following. A digraph is Strongly Connected (SC) if, for every pair of nodes vi and vj, there exists a
strong path from vi to vj and viceversa. A digraph is Quasi -Strongly Connected (QSC) if, for every
13Observe that some literature (e.g., [20, 21]) defines this concept using the opposite convention for the orientation of
the edges.
18
pair of nodes vi and vj, there exists a node r that can reach both vi and vj by a strong path. A digraph
is weakly connected (WC) if any pair of distinct nodes can be joined by a weak path. A digraph is
disconnected if it is not weakly connected.
According to the above definitions, it is straightforward to see that strong connectivity implies quasi
strong connectivity and that quasi strong connectivity implies weak connectivity, but the converse, in
general, does not hold. For undirected graphs instead, the above notions of connectivity are equivalent:
An undirected graph is connected if any two distinct nodes can be joined by a path. Moreover, it is
easy to check that the quasi strong connectivity of a digraph is equivalent to the existence of at least
one spanning directed tree in the graph (see, e.g., [42, p. 133 ]).
Condensation Digraph When a digraph G is WC, it may still contain strongly connected subgraphs.
A maximal subgraph of G which is also SC is called Strongly Connected Component (SCC) of G 14 [40].
Using this concept, any digraph G ={V , E } can be partitioned into SCCs, let us say G1,{V1, E
1},
. . . , GK ,{VK, E
K }, where Vj ⊆ V and Ej ⊆ E denote the set of nodes and edges lying in the j-
th SCC, respectively. Using this structure, one can reduce the original digraph G to the so called
condensation digraph G ⋆={V ⋆, E ⋆}, by substituting each node set Vi of each SCC Gi of G with a
distinct node v⋆
j if and only if there exists
some edges from the i-th SCC Gi and the j-th SCC Gj [40, Ch. 3.2]. An SCC that is reduced to the
root node of a directed tree of the condensation digraph is called Root SCC (RSCC). Observe that, by
i ∈ V ⋆of G ⋆, and introducing an edge in G ⋆ from v⋆
i to v⋆
definition, the condensation digraph has no cycles [40, Lemma 3.2.3]. Building on this property, we
have the following.
Lemma 1 Let G ⋆={V ⋆, E ⋆} be the condensation digraph of G , composed by K nodes. Then, the nodes
K ∈ V ⋆, so that all the existing edges in G ⋆ are in the form
of G ⋆ can always be ordered as v⋆
1, . . . , v⋆
(v⋆
i , v⋆
j ) ∈ E ⋆,
with
1 ≤ j < i ≤ K,
(24)
where v⋆
1 has zero in-degree.
1, . . . , v⋆
The ordering v⋆
1, remove v⋆
K satisfying (24) can be obtained by the following iterative procedure. Start-
1 and all its out-coming edges from G ⋆. Since the reduced digraph with K − 1
ing from v⋆
nodes has no (strong) cycles by construction, there must exist a node with zero in-degree in it. Let us
denote such a node by v⋆
j ), with j > 2, can exist in the reduced
digraph (and thus in G ⋆). This justifies (24) for i = 2 and j = 1, 2. The rest of (24), for j > 2, is
obtained by repeating the same procedure on the remaining nodes.
2. Then, no edges in the form (v⋆
2, v⋆
The connectivity properties of a digraph are related to the structure of its condensation digraph,
as given in the following Lemma (we omit the proof because of space limitations).
Lemma 2 Let G ⋆ be the condensation digraph of G . Then: i) G is SC if and only if G ⋆is composed
by a single node; ii) G is QSC if and only if G ⋆contains a spanning directed tree; iii) if G is WC, then
G ⋆ contains either a spanning directed tree or a (weakly) connected directed forest.
The concept of condensation digraph is useful to understand the network synchronization behavior,
as shown in Section 3. A similar idea was already used in [10, 21] to study leadership problems in
coordinated multi-agent systems.
14Maximal means that there is no larger SC subgraph containing the nodes of the considered component.
19
A.2 Spectral properties of a Digraph
We recall now some basic relationships between the connectivity properties of the digraph and the
spectral properties of the matrices associated to the digraph, since they play a fundamental role in the
stability analysis of the system proposed in this paper. In the following, we denote by 1N and 0N the
N -length column vector of all ones and zeros, respectively.
Given a digraph G ={V , E }, we introduce the following matrices associated to G :
• The N × N adjacency matrix A is composed of entries [A]ij
, aij, i, j = 1, . . . , N , equal to the
weight associated with the edge eij, if eij ∈ E , or equal to zero, otherwise;
• The degree matrix ∆ is a diagonal matrix with diagonal entries [∆]ii
, degin(vi);
• The (weighted) Laplacian L is defined as
XN
−aij
k6=i=1
aik,
if j = i,
if j 6= i.
(25)
[L]ij
,
Using the adjacency matrix A and the degree matrix ∆, the Laplacian can be rewritten in
compact form as L , ∆ − A.15
By definition, the Laplacian matrix L in (25) has the following properties:
i) it is a diagonally
dominant matrix [43]; ii) it has zero row sums; and iii) it has nonnegative diagonal elements. From
i)-iii), invoking Gersgorin's disk Theorem [43], we have that zero is an eigenvalue of L corresponding
to a right eigenvector in the Null{L} ⊇ span{1N }, i.e.,
L 1N = 0N ,
(26)
while all the other eigenvalues have positive real part. This also means that rank(L) ≤ N − 1.
Moreover, from (23) and (26), it turns out that balanced digraphs can be equivalently characterized
in terms of the Laplacian matrix L: A digraph is balanced if and only if 1N is also a left eigenvector
of L associated with the zero eigenvalue, i.e.,
N L = 0T
1T
N ,
(27)
or equivalently 1
2 (L + LT ) is positive semidefinite.
The relationship between the connectivity properties of a digraph and the spectral properties of its
Laplacian matrix are given by the following.
Lemma 3 Let G ={V , E } be a digraph with Laplacian matrix L. The multiplicity of the zero eigenvalue
of L is equal to the minimum number of directed trees contained in a spanning directed forest of G .
Corollary 2 The zero eigenvalue of L is simple if and only if G contains a spanning directed tree (or,
equivalently, it is QSC).
15Observe that the definition of Laplacian matrix as given in (25) coincides with that used in the classical graph theory
literature, except for the convention we adopted in the orientation of the edges. This leads to L expressed in terms of
the in-degrees matrix, rather than of the out-degrees matrix. Our choice is motivated by the physical interpretation we
gave to the edges' weights, as detailed in Section 2.2.
20
Lemma 3 comes directly form Theorem 9 and Theorem 10 of [44]. Corollary 2 was independently
proved in many papers, such as [9, Corollary 2], [20, Lemma 2]. Observe that, since the strong
connectivity of the digraph implies QSC, the results provided in [8, 21] for SC digraphs, can be
obtained as special case of Corollary 2. Specifically, we have the following.
Corollary 3 Let G ={V , E } be a digraph with Laplacian matrix L. If G is SC, then L has a simple
zero eigenvalue and a positive left-eigenvector associated to the zero eigenvalue.
According to Corollary 2, because of (26), the Laplacian of a QSC digraph has an isolated eigenvalue
equal to zero, corresponding to a right eigenvector in the span{1N }. Observe that, for undirected
graphs, Corollary 3 can be stated as follows: rank(L) = N − 1 if and only if G is connected [39]. For
directed graphs, instead, the "only if" part does not hold.
We derive now the structure of the left-eigenvector γ of the Laplacian matrix L associated to the
zero eigenvalue, as a function of the network topology. This result is instrumental to prove the main
theorem of this paper. We have the following.
Lemma 4 Let G ={V , E } be a digraph with N nodes and Laplacian matrix L. Assume that G is QSC
with K SCC's G1,{V1, E
bered w.l.o.g. so that G1 coincides with the RSCC of G . Then, the left-eigenvector γ = [γ1, . . . , γ N ]T
of L associated to the zero eigenvalue has the following structure
K}, with Vi ⊆ V , Ei ⊆ E , Vi = ri andPi ri = N, num-
1}, . . . , GK ,{VK, E
γi =( > 0,
= 0,
iff vi ∈ V1,
otherwise.
(28)
If G1 is also balanced, then γr1
, [γ1, . . . , γr1]T ∈ span{1r1}, where r1, V1 .
Proof. Since the digraph G is QSC, the set V1 contains either all N or 0 < r1 < N nodes of G .
In the former case, V1 ≡ V and thus the digraph is SC, by definition. Hence, according to Corollary
3 (cf. Appendix A.2), the Laplacian matrix L(G ) has a simple zero eigenvalue with left-eigenvector
γ > 0. If, in addition, G is balanced (and thus also SC), then γ ∈ span {1N } .
We consider now the latter case, i.e., 0 < r1 < N. According to Lemma 2, the condensation digraph
K ∈ V ⋆ (associated
1 (associated to the SCC G1 of G ). From Lemma 1, we
K ∈ V ⋆ are ordered according to (24), so that the Laplacian
G ⋆={V ⋆, E ⋆} of G contains a spanning directed tree composed by K nodes v⋆
to the K SCCs G1, . . . , GK ), with root node v⋆
1, . . . , v⋆
assume, w.l.o.g., that the nodes v⋆
matrix of G ⋆ be a lower triangular matrix.
1, . . . , v⋆
Using the relationship between the original digraph G and its condensation digraph G ⋆ (cf. Ap-
pendix A.1), the Laplacian matrix L(G ) of G can be written as a lower block triangular matrix, i.e.
with L1 = L(G1) ∈ Rr1×r1 denoting the r1 × r1 Laplacian matrix associated to the root SCC G1, and
Bk = Lk + Dk ∈ Rrk×rk , where Lk = L(Gk) is the rk × rk Laplacian matrix of the k-th SCC Gk
L(G ) =
,
(29)
L1
0
∗ B2
∗
∗
∗
. . .
. . .
0
...
. . .
. . .
0
∗ BK
21
of G and Dk is a nonnegative diagonal matrix whose i-th entry is equal to the sum of the weights
associated to the edges outgoing from the nodes in G1, . . . , Gk−1 and incoming in the i-th node in Gk.16
Observe that, since G is QSC, each Dk has at least one positive entry; otherwise the SCC Gk would
be decoupled from the other SCCs.
Since each Gk is SC by definition, we have rank(Lk) = rk − 1 (cf. Corollary 3) and Null(Lk) =
span(1rk ) (see (26)). Using these properties and the fact that Dk1rk 6= 0rk , we have that the rows
(columns) of Lk + Dk are linearly independent, or equivalently,
rank(Bk) = rk,
∀k = 2, . . . , K.
(30)
Using (29) and (30), we derive now the structure of the left-eigenvector γ of L(G ) in (29). Parti-
tioning γ = [γ T
rK ]T according to (29), with γ rk ∈ Rrk , we have
r1, γT
r2, . . . , γT
γ T
r1L1 = 0T
r1
and γT
rk Bk = 0T
rk ,
k = 2, . . . , K,
which, using Corollary 3 and (30), provides
γr1 > 0r1
and γ rk = 0rk ,
k = 2, . . . , K,
respectively. If, in addition, G1 is balanced, then γr1 ∈ span {1r1} .
(31)
(32)
B Preliminary Results on Linear Functional Differential Equations
In this section, we introduce some basic definitions on linear functional differential equations and prove
some intermediate results that will be extensively used in the proof of Theorem 1, as given in Appendix
C.
To formally introduce the concept of functional differential equations, we need the following nota-
tions: let RN be an N -dimensional linear vector space over the real numbers with vector norm k·k ; de-
noting by τ the maximum time delay of the system, let C , C(cid:0)[−τ , 0], RN(cid:1) [or C1 , C1(cid:0)[−τ , 0], RN(cid:1)]
be the Banach space of continuous (or continuously differentiable) functions mapping the interval
[−τ , 0] to RN with the topology of uniform convergence, i.e., for any φ ∈ C (or φ ∈ C1) the norm of
φ is defined as φs = sup−τ ≤ϑ≤0 kφ(ϑ)k.
Since in this paper we considered only linear coupling among the differential equations as given in
(7), we focus on Linear homogeneous Delayed Differential Equations (LDDEs) with a finite number
of heterogeneous non-commensurate delays17 [48]-[51]:
·
xi(t) = ki Xj∈Ni
aij (xj(t − τ ij) − xi(t)) ,
i = 1, . . . , N,
t > t0,
(33)
where ki is any arbitrary positive constant. Equation (33) means that the derivative of the state
variables x at time t depends on t and x(ϑ), for ϑ ∈ [t − τ , t]. Hence, to uniquely specify the evolution
16We assume, w.l.o.g., that in each SCC Gk the nodes are numbered from 1 to rk.
17Observe that the LDDE (33) falls in the more general class of LDDE's studied in [48]-[50]. In fact, it is straightforward
to check that LDDE (33) can be rewritten in the canonical form of [48, Ch. 6, Eq. (6.3.2)], [50, Ch. 3, Eq. (3.1)].
22
of the state beyond time t0, it is required to specify the initial state variables x(t) in a time interval of
length τ , from t0 − τ to t0, i.e.,
xi(t0 + ϑ) = φi(ϑ),
ϑ ∈ [−τ , 0],
i = 1, . . . , N .
(34)
Here after, the initial conditions φ = {φi}i are assumed to be taken in the set C1
differentiable) functions that are bounded in the norm18 φs = sup−τ≤ϑ≤0 kφ(ϑ)k∞, i.e.,
β of (continuously
,(φ ∈ C1 : φs = sup
−τ ≤ϑ≤0
kφ(ϑ)k∞ ≤ β < +∞) .
C1
β
(35)
Given φ ∈ C1
β and t0, let x[t0, φ](t) denote the function x at time t, with initial value φ at time
t0, i.e., x(t0 + ϑ) = φ(ϑ), with ϑ ∈ [−τ , 0]. For the sake of notation, we define x[φ](t) , x[0, φ](t).
A function x[t0, φ](t) is said to be a solution to equation (33) on [t0 − τ , t0 + t1) with initial value φ
at time t0, if there exist t1 ≥ 0 such that: i) x ∈ C ([t0 − τ , t0 + t1], Rn) ; ii) x[t0, φ](t) satisfies (33),
∀t ∈ [t0, t0 + t1]; and iii) x(t0 + ϑ) = φ(ϑ), with ϑ ∈ [−τ , 0]. It follows from [50, Theorem 1.2] that
such a solution to equation (33) exists and is unique. We focus now on two concepts related to the
trajectories of (33), namely boundedness and stability.
Definition 2 Given system (33), a solution x[t0, φ](t) is bounded if there exists a β = β(t0, φ) such
that kx[t0, φ](t)k < β(t0, φ) for t ≥ t0 − τ . The solutions are uniformly bounded if, for any α > 0,
there exists a β = β(α) > 0 such that for all t0 ∈ R, φ ∈ C and φs < α we have kx[t0, φ](t)k < β for
all t ≥ t0.
As far as the stability notion is concerned, it is not at all different from its counterpart for unretarded
systems, except for the different assumptions on the initial conditions. The interested reader may refer,
e.g., to [50, 51] for an in-depth treatment of this topic. We focus here on effective methods on proving
the stability of LDDEs. Since system (33) is linear, the stability analysis can be carried out either
in the time-domain [50, 51] or in the frequency-domain [48, 49], as for classical ordinary differential
equations (see, e.g., [52]). In this paper we focus on the latter approach. The same conclusions can
also be obtained using the time-domain analysis, based on Lyapunov-Krasovskii functional [53]. Before
stating the major result, we need first the following intermediate definitions.
Let C+ = {s ∈ C : Re{s} > 0}, C− = {s ∈ C : Re{s} < 0}, and C+ be the closure of C+, i.e.,
C+ = {s ∈ C : Re{s} ≥ 0}. Denoting by Hn×m the set of n × m matrices whose entries are analytic19
and bounded functions in C+, let us introduce the N × N diagonal degree matrix ∆ ≥ 0N ×N and the
complex matrix H(s) ∈ CN ×N , defined respectively as
∆ , diag (k1 degin(v1), ..., kN degin(vN ))
and
[H(s)]ij
,( 0,
kiaije−sτ ij ,
if i = j,
if i 6= j,
(36)
with degin(vi) given in (22) (see Appendix A). Observe that H(s) ∈ HN ×N . We can now provide the
main result of this section, stated in the following lemma.
18We used, without loss of generality, as vector norm k·k in RN the infinity norm k·k∞ , defined as kxk∞
, maxi xi.
19A complex function is said to be analytic (or holomorphic) on a region D ⊆ C if it is complex differentiable at every
point in D, i.e., for any z0 ∈ D the function satisfies the Cauchy-Riemann equations and has continuous first partial
derivatives in the neighborhood of z0 (see, e.g., [54, Theorem 11.2]).
23
Lemma 5 Given system (33), assume that the following conditions are satisfied:
b1. The initial value functions φ ∈ C1
β, and the solutions x[φ](t) are bounded;
b2. The characteristic equation associated to (33)
p(s) , det (sI + ∆ − H(s)) = 0,
(37)
with ∆ and H(s) defined in (36), has all roots {sr}r ∈ C−, with at most one simple root at
s = 0.20
Then, system (33) is marginally stable, i.e., ∀φ ∈ C1
β and Re{s1} < c < 0, there exist t1 and α,
with t0 < t1 < +∞ and 0 < α < +∞, independent of φ, and a vector x∞, with kx∞k < +∞, such that
kx[φ](t) − x∞k ≤ α φs ect,
∀t > t1.
(38)
Proof. Under assumption b1), according to [48, Theorem 6.5], the stability properties of system (33)
are fully determined by the roots of the characteristic equation associated to (33), as detailed next.21
Denoting by M = {Re{sr} : p(sr) = 0} the set of real parts of the characteristic roots {sr}r, it follows
from [48, Theorem 6.7] that, ∀φ ∈ C1
β and c /∈ M, there exist t1 and α, with 0 < t1 < +∞ and
0 < α < +∞ independent on φ, such that
≤ α φs ect,
∀t > t1,
(39)
x[φ](t) − lim
l→+∞ Xsr∈Cl: Re{sr}>c
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
pr(t)esrt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
where each pr is a (vectorial) polynomial of degree less than the multiplicity of sr, Cl denotes a contour
in the complex plane of radius increasing with l and centered around s = 0 (see [48, Sec. 4.1] for more
details on how such contours Cl need to be chosen), and the sum in (39) is taken over all characteristic
roots srwithin the contour Cl and to the right of the line Re{sr} = c. Observe that, since the number
of such roots is finite22 [50, Theorem 1.5] (see also [48, Ch. 6.8 and Ch. 12]), the limit in (39) is always
well-defined.
Using (39), we prove now that, under b2), system (33) is marginally stable. Invoking the property
that, for any given γ ∈ R, the number of characteristic roots sr with Re{sr} > γ is finite, one can
always choose the constant c /∈ M in (39) so that Re{s1} < c < Re{s0} = 0, which leads to
kx[φ](t) − p0k ≤ α φs ect,
∀t > t1,
(40)
where we have explicitly used the assumption that the (possible) root s0 = 0 is simple and that there
are no roots with positive real part.
It follows from [48, Corollary 6.2] that this condition is also
sufficient to guarantee that p0 is bounded, i.e., kp0k < +∞. Setting in (40) x∞ = p0, we obtain (38),
which completes the proof.
20We assume, w.l.o.g., that the roots {sr} are arranged in nonincreasing order with respect to the real part, i.e.,
0 = Re{s0} > Re{s1} ≥ Re{s2} ≥ ....
21Observe that assumption b1) is only sufficient for the existence of the Laplace transform of the solutions to (33).
22The reader who is familiar with Linear Retarded Functional Differential Equations (LRFDE) may observe that this
result comes directly from the fact that the characteristic equation (37) does not have neutral roots [48, Ch. 12].
24
Remark 1.
It follows from Lemma 5 that, under assumptions b1)-b2), all the solutions to (33)
asymptotically convergence to the constant vector x∞. Moreover, equation (38) provides an estimate
of the convergence rate of the system: as for systems without delays, the convergence speed of (33) is
exponential23 with rate arbitrarily close to Re{s1} that, in general, depends on network topology and
delays.
Remark 2. Lemma 5 generalizes results of [49, 50], where the authors provided alternative conditions
for the asymptotic stability of a Linear Retarded Differential Equation (LRDE). Interestingly, Lemma
5 contains some of the conditions of [49, 50] as a special case: System (33) is asymptotically stable if
all the characteristic roots of (37) have negative real part [see (38)]. This conclusion is the same as
that for linear unretarded systems (see, e.g., [52]).
Moreover, Lemma 5 can be easily generalized to include the cases in which one is interested in
"oscillatory" behaviors of system (33). It is straightforward to see that bounded oscillations arise if
assumption b2) is replaced by the following condition: All the roots of characteristic equation (37)
have negative real part and the roots with zero real part are simple.
C Proof of Theorem 1
In the following, for the sake of notation simplicity, we drop the dependence of the state function from
the observation, as this dependence does not play any role in our proof.
C.1 Sufficiency
We prove that, under a1)-a3), the quasi strong connectivity of digraph G associated to the network in
(7) is a sufficient condition for the system (7) to synchronize and that the synchronized state is given
by (11). To this end, we organize the proof according to the following two steps.
We first show that, under a1)-a2) and the quasi-strong connectivity of G , the set of LRFDEs (7)
admits a solution in the form
x⋆
i (t) = αt + x⋆
i,0,
i = 1, . . . , N,
(41)
if α = ω⋆, where ω⋆ is defined in (11) and {x⋆
i,0} are constants that depend in general on the system
parameters and the initial conditions. This guarantees the existence of the desired synchronized state
(cf. Definition 1). Then, invoking results of Appendix B, we prove that, under a1)-a3) and the quasi-
strong connectivity of G , such a synchronized states is also globally asymptotically stable (according
to Definition 1).
C.1.1 Existence of a synchronized state
Let us assume that conditions a1)-a2) are satisfied and that G ={V , E } associated to (7) is QSC.
The synchronized state in the form (41) is a solution to (7) if and only if it satisfies equations (7)
(cf. Appendix B), i.e., if and only if there exist α and {x⋆
i,0} such that the following system of linear
equations is feasible:
∀i = 1, . . . , N,
(42)
ci∆ωi(α)
K
+ Xj∈Ni
aij(cid:0)x⋆
j,0 − x⋆
i,0(cid:1) = 0,
23We say that x(t) converges exponentially toward x∞ with rate r < 0 if kx(t) − x∞k ≤ O `ert´ .
25
where
∆ωi(α) , gi(yi) − α1 +
K
ci Xj∈Ni
aij τ ij .
(43)
Introducing the weighted Laplacian L = L(G ) associated to digraph G (cf. Section A), the system in
(42) can be equivalently rewritten as
Lx⋆
0 =
1
K
Dc∆ ω(α),
(44)
, [x⋆
1,0, . . . , x⋆
N,0]T , Dc , diag(c1, . . . , cN ), and ∆ ω(α) , [∆ ω1(α), . . . , ∆ ωN (α)]T , with
where x⋆
0
∆ ωi(α) defined in (43). Observe that, under a1)-a2) and the quasi-strong connectivity of G , the
graph Laplacian L has the following properties (cf. Corollary 2):
rank(L) = N − 1, N (L) = span {1N } ,
and N (LT ) = span {γ} ,
(45)
where N (L) denotes the (right) null-space of L, and γ is a left eigenvector of L corresponding to the
(simple) zero eigenvalue of L, i.e., γT L = 0T .
Assume now that α is fixed.
It follows from (45) that, for any given α, system (44) admits a
solution if and only if Dc∆ ω(α) ∈ span {L} . Because of (45), we have
Dc∆ ω(α) ∈ span {L} ⇔ γT Dc∆ ω(α) = 0.
(46)
It is easy to check that the value of α that satisfies (46) is α = ω⋆, with ω⋆ defined in (11). Hence, if
α = ω⋆, the synchronized state in the form (41) is a solution to (7), for any given set of {τ ij}, {gi},
{ci}, {aij} and K 6= 0. The structure of the left eigenvector γ associated to the zero eigenvalue of L
as given in (11) comes directly from Lemma 4.
Setting α = ω⋆, system (44) admits ∞1 solutions, given by
x⋆
0 =
1
K
L♯Dc∆ ω(ω⋆) + span{1N } , x0 + span{1N },
where
x0 ,
1
K
L♯Dc∆ ω(ω⋆),
(47)
(48)
∆ωi(ω⋆) is obtained by (43) setting α = ω⋆ and L♯
A is the generalized inverse of the Laplacian L [56].
C.1.2 Global Asymptotic Stability of the Synchronized State
To prove the global asymptotic stability of the synchronized state of system (7), whose existence has
been proved in Appendix C.1.1, we use the following intermediate result (see Appendix B for the
definitions used in the lemma).
Lemma 6 ([55, Theorem 2.2]) Let H(s) ∈ HN ×N and ρ (H(s)) denote the spectral radius of H(s).
Then, ρ (H(s)) is a subharmonic24 bounded (above) function on C+.
24See, e.g., [54, Ch. 12], [55], for the definition of subharmonic function.
26
We first rewrite system (7) in a more convenient form, as detailed next. Consider the following
change of variables
Ψi(t) , xi(t) − (ω⋆t + xi,0) ,
i = 1, . . . , N,
(49)
where ω⋆ and {xi,0} are defined in (11) and (48), respectively, so that the original system (7) can be
equivalently rewritten in terms of the new variables {Ψi(t)}i as
Ψi(t) = ∆ωi(ω⋆) +
K
ci Xj∈Ni
aij (Ψj(t − τ ij) − Ψi(t) + xj,0 − xi,0) ,
i = 1, . . . , N,
t ≥ 0,
(50)
Ψi(ϑ) = φi(ϑ) ,eφi(ϑ) − ω⋆ϑ − xi,0, ϑ ∈ [−τ , 0],
where eφ are the initial value functions of the original system (7). Using (48) [see also (42), with
α = ω⋆], system (50) becomes
aij (Ψj(t − τ ij) − Ψi(t)) ,
i = 1, . . . , N,
t ≥ 0,
(51)
Ψi(t) = ki Xj∈Ni
Ψi(ϑ) = φi(ϑ), ϑ ∈ [−τ , 0],
where, for the sake of convenience, we defined ki , K/ci > 0, for i = 1, . . . , N.
It follows from (51) that the synchronized state of system (7), as given in (41), is globally asymp-
totically stable (according to Definition 1) if system in (51) is marginally stable (cf. Appendix B). To
prove the marginal stability of system (51), it is sufficient to show that system (51) satisfies Lemma
5. To this end, we organize the rest of the proof in the following two steps.
Step 1. We show that, under a1), a3), all the solutions Ψ[φ](t) to system (51) are uniformly bounded,
as required by assumption b1) of Lemma 5;
Step 2. We prove that, under a1)-a3) and the quasi-strong connectivity of the digraph, the charac-
teristic equation associated to system (51) has all the roots in C− and a simple root in s = 0,
which satisfies assumption b2) of Lemma 5.
Step 1. Given any arbitrary β < +∞, assumptions a1), a3) are sufficient to guarantee that all the
trajectories of (51) are uniformly bounded (cf. Definition 2), as shown next. Since φ ∈ C1
β, we have
(see (35))
Ψi (ϑ) ≤ β,
∀i = 1, . . . , N, ϑ ∈ [−τ , 0].
(52)
Condition (52) is sufficient for {Ψi[φ](t)} to be uniformly bounded for all t > 0. In fact, assume that
{Ψi[φ](t)} are not bounded. Then, according to Definition 2 of Appendix B, there must exist some
t > 0 and a set J ⊆ {1, . . . , N } such that25
kΨ(t)k∞ ≤ β,
∀t < t,
and
Ψj(t) =( < 0,
This would imply that, for some t > t, Ψj(cid:0)t(cid:1) > β.
Ψj(cid:0)t(cid:1) = β and
> 0,
·
Given (53)-(54) and j ∈ J , we have two possibilities, namely:
25For the sake of notation, we omit in the following the dependence of Ψ[φ](t) on φ.
27
if Ψj(cid:0)t(cid:1) = −β,
if Ψj(cid:0)t(cid:1) = β,
j ∈ J .
(53)
(54)
i) Ψj(cid:0)t(cid:1) = β and
we have
·
Ψj(t) > 0. Since Ψi(t − τ ji) < β, ∀i = 1, . . . , N (see (53)) and kj > 0, from (51)
·
Ψj(t) = kj Xi∈Nj
aji(cid:0)Ψi(t − τ ji) − Ψj(cid:0)t(cid:1)(cid:1) = kj Xi∈Nj
aji(cid:0)Ψi(t − τ ji) − β(cid:1) ≤ 0,
where in the last inequality, we used (52). Expression (55) contradicts the assumption
(55)
·
Ψj(t) > 0.
ii) Ψj(cid:0)t(cid:1) = −β and
·
Ψj(t) < 0. However, since
·
Ψj(t) = kj Xi∈Nj
aji(cid:0)Ψi(t − τ ji) − Ψj(t)(cid:1) = kj Xi∈Nj
aji(cid:0)Ψi(t − τ ji) + β(cid:1) ≥ 0,
(56)
a contradiction with
·
Ψj(t) < 0 results.
Thus, any solution Ψ[φ](t) to (51) is bounded and remains in C1
β for any t > 0.
Step 2. We study now the characteristic equation (37) associated to system (51), assuming that
a1)-a3) are satisfied and that digraph G is QSC. First of all, observe that, since ∆ − H(0) = KDcL,
we have
p(0) = det (∆ − H(0)) = K det (Dc) det (L) = 0,
(57)
where ∆ and H(s) are defined in (36), and the last equality in (57) is due to the properties of
Laplacian matrix L = L(G ) of digraph G (see (26)). It follows from (57) that p(s) has a root in s = 0,
corresponding to the zero eigenvalue of the Laplacian L (recall that K det (Dc) 6= 0). Since the digraph
is assumed to be QSC, according to Corollary 2, such a root is simple.
Thus, to complete the proof, we need to show that p(s) does not have any solution in C+\{0}, i.e.,
det (sI+∆ − H(s)) 6= 0,
∀s ∈ C+\{0}.
(58)
Since sI+∆ is nonsingular in C+\{0} [recall that, under a1), ∆ ≥ 0, with at least one positive
diagonal entry], (58) is equivalent to
which leads to the following sufficient condition for (58):
det(cid:16)I− (sI+∆)−1 H(s)(cid:17) 6= 0,
ρ (s) , ρ(cid:16)(sI+∆)−1 H(s)(cid:17) < 1,
∀s ∈ C+\{0},
∀s ∈ C+\{0}.
(59)
(60)
Since (sI+∆)−1 ∈ HN ×N and H(s) ∈ HN ×N (cf. Appendix B), it follows from Lemma 6 that the
spectral radius ρ (s) in (60) is a subharmonic function on C+. As a direct consequence, we have, among
26 and satisfies the maximum modulus principle
all, that ρ (s) is a continuous bounded function on C+
(see, e.g., [54, Ch. 12]): ρ (s) achieves its global maximum only on the boundary of C+.27 Since ρ (s)
is strictly proper in C+, i.e., ρ (s) → 0 as s → +∞ while keeping s ∈ C+, it follows that
sup
s∈C+
ρ (s) < sup
s∈C+
ρ (s) ≤ sup
ω∈R
ρ (jω) .
(61)
26Observe that ρ (s) is well-defined in s = 0, and ρ (0) = 1.
27According to the maximum modulus theorem [54, Theorem 10.24], the only possibility for ρ (s) to reach its global
maximum also on the interior of C+ is that ρ (s) be constant over all C+, which is not the case.
28
Using (61), we infer that condition (60) is satisfied if
ρ (jω) = ρ(cid:16)(jωI+∆)−1 H(jω)(cid:17) < 1,
∀ω ∈ R\{0}.
(62)
For any matrix norm k·k, since ρ (A) ≤ kAk ∀A ∈ CN ×M [43, Theorem 5.6.9], using the maximum
row sum matrix norm k·k∞ defined as [43, Definition 5.6.5]
kAk∞
we have
, max
r=1,...,N
ρ (jω) ≤ (cid:13)(cid:13)(cid:13)(jωI+∆)−1 H(jω)(cid:13)(cid:13)(cid:13)∞
= max
r=1,...,N(cid:12)(cid:12)(cid:12)(cid:12)
kr degin(vr)
jω + kr degin(vr)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 1,
MXq=1(cid:12)(cid:12)(cid:12)[A]rq(cid:12)(cid:12)(cid:12) ,
r=1,...,NXq6=r(cid:12)(cid:12)(cid:12)(cid:12)
= max
kr arq
jω + kr degin(vr)
(63)
(64)
e−jωτ rq(cid:12)(cid:12)(cid:12)(cid:12)
where the first r diagonal blocks are the Laplacian matrices of the r RSCCs of G , and the {Bk}, with
k > r, are the nonsingular matrices associated to the remaining SCCs. Each of these matrices can be
where in the last inequality the equality is reached if and only if ω = 0. It follows from (64) that
ρ (jω) < 1 for all ω 6= 0, which guarantees that condition (60) is satisfied.
This proves that assumption b2) of Lemma 5 holds true. Hence, all the trajectories Ψ[φ](t) → Ψ∞
as t → +∞, with exponential rate arbitrarily close to r , {mini Re{si} : p(si) = 0 and si 6= 0}, where
p(s) is defined in (37) and Ψ∞ satisfies the linear system of equations LΨ∞ = 0, whose solution is
Ψ∞ ∈ span{1N } (Corollary 2). In other words, system (51) exponentially reaches the consensus on
the state.
C.2 Necessity
We prove the necessity of the condition by showing that, if the digraph G of (7) is not QSC, different
clusters of nodes synchronize on different values. This local synchronization is in contrast with the
definition of (global) synchronization, as given in Definition 1. Hence, if the overall system has to
synchronize, the digraph associated to the system must be QSC.
Assume that the digraph G associated to (7) is not QSC, but WC with K SCCs and, let us say,
r ≤ K (distinct) RSCCs . Then, according to Lemma 2, the condensation digraph G ⋆={V ⋆, E ⋆}
contains a spanning directed forest with r distinct roots (associated to the r RSCCs of the K SCCs of
G ). Ordering the nodes v⋆
K ∈ V ⋆ according to Lemma 1, and exploring the relationship between
G ⋆and G (cf. Appendix C.1.1), one can write the Laplacian matrix L = L(G ) as an r-reducible matrix
[43], i.e.,
1, . . . , v⋆
L(G ) =
,
(65)
L1
0
0
∗
∗
∗
0
. . .
· · ·
. . .
0 Lr
· · ·
· · ·
. . .
∗
∗
∗
∗ Br+1
∗
∗
∗
∗
· · ·
· · ·
0
...
...
· · ·
...
. . .
. . .
0
∗ BK
29
written as the linear combination of the Laplacian matrix of the corresponding SCC and a nonnegative
diagonal matrix with at least one positive diagonal entry. The structure of L given in (65) shows that
the RSCCs associated to the first r diagonal blocks are totally decoupled from each other. Hence, (at
least) the state derivatives of the nodes in each of these r RSCCs reach a common value (since the
corresponding subdigraphs are SC by construction) that, in general, is different for any of the SCCs.
This is sufficient for the overall system not to reach a global synchronization.
References
[1] E. Eisenberg, D. Gale, "Consensus of Subjective Probabilities: The Pari-mutuel Method," The Annals of
Mathematical Statistics, pp. 165 -- 168, Mar. 1959.
[2] M. H. DeGroot, "Reaching a Consensus," Journal of the American Statistical Association, pp. 118 -- 121,
Mar. 1974.
[3] J. N. Tsitsiklis, D. P. Bertsekas, M. Athans, "Distributed Asynchronous Deterministic and Stochastic
Gradient Optimization Algorithms," IEEE Trans. on Automatic Control, pp. 803 -- 812, Sep. 1986.
[4] J. N. Tsitsiklis, "Problems in Decentralized Decisions Making and Computation," Ph.D. Dissertation, MIT,
Cambridge, MA, Nov. 1982. Available at http://web.mit.edu/jnt/www/Papers/PhD-84-jnt.pdf.
[5] D. P Bertsekas and J.N. Tsitsiklis, Parallel and Distributed Computation: Numerical Methods, Athena
Scientific, 1989.
[6] V. Borkar, P. Varaiya, "Asymptotic Agreement in Distributed Estimation," IEEE Trans. on Automatic
Control, pp. 650 -- 655, 1982.
[7] R. Olfati-Saber and R. M. Murray, "Consensus Protocols for Networks of Dynamic Agents," in Proc. of the
2003 American Control Conference, June 2003.
[8] R. Olfati-Saber, and R.M. Murray, "Consensus Problems in Networks of Agents with Switching Topology
and Time-Delays," IEEE Trans. on Automatic Control, vol. 49, pp. 1520-1533, Sep., 2004.
[9] W. Ren, R. W. Beard, and T. W. McLain, "Coordination Variables and Consensus Building in Multiple
Vehicle Systems,"in Proc. of the Block Island Workshop on Cooperative Control, Springer-Verlag Series,
vol. 309, p. 171-188, 2005.
[10] C. W. Wu, "Agreement and Consensus Problems in Groups of Autonomous Agents with Linear Dynamics,"
in Proc. of IEEE International Symposium on Circuits and Systems (ISCAS 2005), pp. 292 - 295, 23-26
May 2005.
[11] L. Xiao, S. Boyd, and S. Lall, "A scheme for Robust Distributed Sensor Fusion based on Average Consen-
sus," in Proc. of the Fourth International Symposium on Information Processing in Sensor Networks, IPSN
2005, 15 April 2005, pp. 63 -- 70.
[12] J. Cort´es, "Finite-time Convergent Gradient Flows with Applications to Network Consensus,"Automatica,
vol. 42, no. 11, pp. 1993 -- 2000, November 2006.
[13] R. Olfati-Saber, E. Franco, E. Frazzoli, and J. S. Shamma, "Belief Consensus and Distributed Hypothesis
Testing in Sensor Networks," Lecture Notes in Control and Information Sciences, vol 331, pp. 169 - 182,
July 2006.
[14] D. Bauso, L. Giarre, and R. Pesenti, "Nonlinear Protocols for Optimal Distributed Consensus in Networks
of Dynamic Agents," Systems & Control Letters, vol 55, no. 11, pp. 918-928, July 2006.
[15] J. Cort´es, "Distributed Algorithms for Reaching Consensus on General Functions,"submitted to Automatica,
October 2006. Available at http://www.soe.ucsc.edu/∼jcortes/research/data/2006 Co-auto.pdf.
[16] R. Olfati-Saber and J. Shamma, "Consensus Filters for Sensor Networks and Distributed Sensor Fusion,"in
Proc. of the Joint 44th IEEE Conference on Decision and Control and European Control Conference (CDC-
ECC'05), Seville, Spain, December 2005.
30
[17] R. Olfati-Saber, "Distributed Kalman Filter with Embedded Consensus Filters,"in Proc. of the Joint 44th
IEEE Conference on Decision and Control and European Control Conference (CDC-ECC'05), Seville,
Spain, December 2005.
[18] A. Jadbabaie, J. Lin, A. S. Morse, "Coordination of Groups of Mobile Autonomous Agents Using Nearest
Neighbor Rules," IEEE Trans. on Automatic Control, pp. 988 -- 1001, June 2003.
[19] L. Moreau, "Stability of Multiagent Systems with Time-dependent Communication Links," IEEE Trans-
actions on Automatic Control, vol. 50, no. 2, pp. 169-182, February 2005.
[20] Z. Lin, B. Francis, and M. Maggiore, "Necessary and Sufficient Graphical Conditions for Formation Control
of Unicycles,"in IEEE Trans. on Automatic Control, pp. 121-127, 2005.
[21] Z. Jin and R. M. Murray, "Consensus Controllability for Coordinated Multiple Vehicle Control, in Proc. of
Sixth Int. Conf. on Cooperative Control and Optimization, Feb. 1 -- 3, 2006, Gainesville, Florida.
[22] V. D. Blondel, J. M. Hendrickx, A. Olshevsky, and J. N. Tsitsiklis, "Convergence in Multiagent Coordina-
tion, Consensus, and Flocking," in Proc. of the Joint 44th IEEE Conference on Decision and Control and
European Control Conference (CDC-ECC'05), Seville, Spain, December 2005.
[23] R. Olfati-Saber, J. A. Fax, R. M. Murray, "Consensus and Cooperation in Networked Multi-agent Systems,"
in Proc. of the IEEE, vol. 95, no. 1, pp. 215-233, Jan. 2007.
[24] W. Ren, R. W. Beard, and E. M. Atkins, "Information Consensus in Multivehicle Cooperative Control:
Collective Group Behavior Through Local Interaction," IEEE Control Systems Magazine, vol. 27, no. 2,
pp. 71-82, April 2007.
[25] S. Barbarossa, "Self-organizing Sensor Networks with Information Propagation based on Mutual Coupling
of Dynamic Systems," in Proc. of IWWAN 2005, London, UK, 2005.
[26] S. Barbarossa, and G. Scutari, "Decentralized Maximum Likelihood Estimation for Sensor Networks Com-
posed of Self-synchronizing Locally Coupled Oscillators,"to appear on IEEE Transactions on Signal Pro-
cessing, July 2007. Available at http://arxiv.org/abs/cs.DC/0612042.
[27] I. Schizas, A. Ribeiro, and G. B. Giannakis "Distributed Estimation with Ad Hoc Wireless Sensor Net-
works," in Proc. of European Signal. Process. Conf. (EUSIPCO 06), Florence, Italy, September 4-8, 2006.
[28] M.G. Earl, and S.H. Strogatz, "Synchronization in Oscillator Networks with Delayed Coupling: A Stability
Criterion,"Physical Rev. E, Vol. 67, pp. 1-4, 2003.
[29] A. Papachristodoulou and A. Jadbabaie, "Synchronization in Oscillator Networks: Switching Topologies
and Presence of Nonhomogeneous delays,"in Proc of the 43th IEEE Conference on Decision and Control
and European Control Conference (ECC-CDC 2005), Seville, Spain, December 2005.
[30] D.J. Lee and M.W. Spong, "Agreement With Non-uniform Information Delays," in Proc. of the American
Control Conference (ACC 2006), Minneapolis, MN, June 14-16, 2006.
[31] L. Wang and F. Xiao, "A New Approach to Consensus Problem for Discrete-Time Multiagent Systems with
Time-Delays," in Proc. of the American Control Conference (ACC 2006), Minneapolis, MN, June 14-16,
2006.
[32] G. Scutari, S. Barbarossa, and L. Pescosolido, "Optimal Decentralized Estimation Through Self-
synchronizing Networks in the Presence of Propagation Delays,"in Proc. of IEEE Workshop on Signal
Processing Advances in Wireless Communications, (SPAWC-2006), Cannes, France, July 2006.
[33] A. Giridhar, and P.R. Kumar, "Computing and Communications Functions," IEEE Trans. on Information
Theory, pp. 388 -- 404, March 2005.
[34] S. M. Kay, Fundamentals of Statistical Signal Processing, Prentice Hall Signal Proc. Series, Upper Saddle
River, NJ, 1998.
[35] L. Pescosolido, S. Barbarossa and G. Scutari, "Decentralized Detection and Localization Through Sensor
Networks Designed As a Population of Self-Synchronizing Oscillators,"in Proc. of the IEEE International
Conference on Acoustics, Speech and Signal Processing, (ICASSP 2006), Vol. 4, pp. 981-984, Toulouse,
May 15-19, 2006.
[36] A.-L. Barabasi and R. Albert, "Emergence of Scaling in Random Networks," Science, vol. 286, pp. 509-512,
1999.
31
[37] S. Barbarossa, G. Scutari, A. Swami, "Achieving Consensus in Self-organizing Wireless Sensor Networks:
The impact of Network Topology on Energy Consumption," in Proc. of the IEEE International Conference
on Acoustics, Speech and Signal Processing, (ICASSP 2007), Honolulu, Hawai, April 15-20, 2007.
[38] S. Barbarossa and G. Scutari, "Bio-inspired Sensor Network Design: Distributed Decision Through Self-
synchronization," IEEE Signal Processing Magazine, vol. 24, no. 3, May pp. 2635, 2007.
[39] M. Fiedler, "Algebraic Connectivity of Graphs", Journal of Czechoslovak Math., vol. 23, pp. 298 -- 305, 1973.
[40] R. A. Brualdi and H. J. Ryser, Combinatorial Matrix Theory, Cambridge University Press, 1991.
[41] C. Godsil, and G. Royle, Algebraic Graph Theory, Springer, New York, 2001.
[42] C. Berge and A. Ghouila -- Houri, Programming, Games and Transportation Networks, John Wiley & Sons,
New York, 1965.
[43] R. Horn, and C. R. Johnson, Matrix Analysis, Cambridge University Press, 1985.
[44] C. W. Wu, "On Bounds of Extremal Eigenvalues of Irreducible and m-reducible Matrices,"Linear Algebra
and its Applications, vol. 402, pp. 29-45, June 2005.
[45] C. W. Wu, "On Rayleigh -- Ritz Ratios of a Generalized Laplacian Matrix of Directed Graphs,"Linear Algebra
and its Applications, vol. 402, pp. 207227, February 2005.
[46] C. W. Wu, "Synchronization in an Array of Chaotic Systems Coupled via a Directed Graph,"in Proc. of
the IEEE Int. Symposium on Circuits and System (ISCASS 2005), Kobe, Japan, May 23-26, 2005.
[47] V. Chellaboina, W. M. Haddad, Q. Hui, and J. Ramakrishnan, "On System State Equipartitioning and
Semistability in Network Dynamical Systems with Arbitrary Time-Delays,"in Proc of the 44th IEEE Con-
ference on Decision and Control (CDC 2006), San Diego, CA, USA, December 13-15, 2006.
[48] R. Bellman and K.L. Cooke, Differential-Difference Equations, New York Academic Press, 1963.
[49] G. Stepan, Retarded Dynamical Systems: Stability and Characteristic Functions, Pitman Research Notes
in Mathematics Series, Longman Scientific & Technical, 1989.
[50] K. Gu, V.L. Kharitonov, J. Chen, Stability of Tyme-Delay Systems, Control Engineering Series, Birkhauser,
2002.
[51] J. K. Hale, S.M.V. Lunel, Introduction to Functional Differential Equations, Springer-Verlag, 1991.
[52] T. Kailath, Linear Systems, Prentice Hall, 1979.
[53] J. R. Haddock and J. Terjeki, "Liapunov-Razumikhin Functions and an Invariance Principle for Functional
Differential Equations,"Jour. of Differential Equations, vol. 48, pp. 95-122, 1983.
[54] W. Rudin, Real and Complex Analysis, McGraw-Hill, International Student Ed., 1970.
[55] S. Boyd and S. A. Desoer, "Subharmonic Functions and Performance Bounds on Linear Time-Invariant
Feedback Systems,"in IMA Jour. of Mathematical Control & Information, Vol. 2, pp. 153-170, 1985.
[56] S.L. Campbell and C.D. Meyer, Generalized Inverses of Linear Transformations, Dover Publications, 1991.
32
|
0911.2900 | 1 | 0911 | 2009-11-15T18:09:08 | Computation Speed of the F.A.S.T. Model | [
"cs.MA",
"physics.soc-ph"
] | The F.A.S.T. model for microscopic simulation of pedestrians was formulated with the idea of parallelizability and small computation times in general in mind, but so far it was never demonstrated, if it can in fact be implemented efficiently for execution on a multi-core or multi-CPU system. In this contribution results are given on computation times for the F.A.S.T. model on an eight-core PC. | cs.MA | cs |
Computation Speed of the F.A.S.T. Model
Tobias Kretz
PTV Planung Transport Verkehr AG
Stumpfstrasse 1
D-76131 Karlsruhe
[email protected]
July 19, 2021
Abstract
The F.A.S.T. model for microscopic simulation of pedestrians was
formulated with the idea of parallelizability and small computation
times in general in mind, but so far it was never demonstrated, if it
can in fact be implemented efficiently for execution on a multi-core or
multi-CPU system. In this contribution results are given on computa-
tion times for the F.A.S.T. model on an eight-core PC.
1
Introduction
Pedestrians and vehicles alike are extended objects in space. For simulation
models [1 -- 6] this means that one has to guarantee a mutual exclusion vol-
ume, i.e. that they do not overlapp. This exclusion volume includes more
than the mere body or vehicle, but as well a headway whose size increases
monotonically with speed [7, 8]. This can be achieved in at least two ways:
either all agents compute their next movement step in parallel and poten-
tially emerging conflicts about exclusion volumes are solved afterwards [9],
or the movement is done sequentially: a strategy with which it's easy to
prevent conflicts generally. However, the kind of update procedure has a
strong influence on the dynamics of the system [9 -- 11]. Obviously parallel
update fits better for parallel computing attempts and as a lucky coinci-
dence parallel update has proven to usually yield more realistic results when
physical systems are simulated than sequential update [12, 13].
The F.A.S.T. model [14 -- 19] tries to make use of both strategies, as
the planning process for the next movement step is done in parallel and
the actual movement sequentially to avoid a computationally costly conflict
resolution lateron. Additionally computationally costly calculations (like
1
exponential or trigonometric functions) are only made use of in the planning
process, while actual movement only consists of very simple commands and
calculations.
So, each time step consists of a computationally rather expensive plan-
ning phase, where data common to all agents is only read (easy to parallelize)
and a computationally cheaper actual motion phase, difficult to parallelize,
as it writes to data structures common for all agents. In this contribution re-
sults of measurements of the computation time of a first parallelized version
of the algorithm are given.
2 Technical Details
The parameters of the F.A.S.T. model were chosen to be kS = 1.2 and
kother = 0. In a simulation all agents had the same maximum speed. All
calculations were done for all maximum speeds vm = 1 to vm = 5, but as
with vm = 4 the fundamental diagram of Weidmann [20, 21] is reproduced
quite well [17], the focus in the following results section is on vm = 4.
The computation times given in the next section refer to the simulation
of 396 simulation time steps equivalating to 396 simulated seconds.
The simulations were carried out on a PC with two Xeon E5320 quadcore
processors and 20 GB RAM. The parallelization was done using OpenMP
[22], and the source code was compiled using the Visual C++ 8 (Visual
Studio 2005) compiler.
Parallel computing was made use of only in the process of choosing a
desired cell for each agent:
b l o c k s i z e=max( min ( number of agents / cores , 3 2 7 6 7 ) , 1 ) ;
#pragma omp p a r a l l e l num threads ( c o r e s )
{
#pragma omp for s c h e d u l e ( dynamic , b l o c k s i z e )
for ( int i = 0 ;
i < number of agents ;
i ++)
c h o o s e d e s i r e d c e l l
( i ) ;
}
3 Results and Conclusions
One of the main results from figures 1 to 7 is that with a parameter config-
uration that reproduces Weidmann's fundamental diagram fairly well, and
using all eight cores, a real time computation speed could be achieved when
about 182,000 agents were in the simulation simultaneously. However, sim-
ulating more complex situations like counterflow or non-trivial route choice
would require additional calculations and reduce simulation speed.
2
Figure 1: Number of agents that can be simulated in real time in dependence
of number of cores.
Figure 2: Computation time in dependence of number of agents at fixed
maximum speed vm = 4.
Figure 3: Computation speed factor in dependence of number of cores at
fixed maximum speed vm = 4.
3
Figure 4: Computation speed factor in dependence of number of agents
when eight cores are used (interpolation with splines).
Figure 5: Computation time in dependence of maximum speed for 40,000
agents.
Figure 6: Computation time in dependence of number of agents for 8 cores.
4
Figure 7: Computation speed factor in dependence of number of cores for
40,000 agents.
The computation speed factors with eight cores compared to using only
one core of the same PC were found to be in the range 4.3 to 5.0. This
confirms that the initial intention is met to have a model suited for parallel
computation.
The wide range of factors in figure 4 may be a hint that there might be
more efficient partitions of the parallel calculation parts.
Apart from the model efficiently making use of a high number of cores,
Intel has released a CPU (Xeon 5482) with a -- depending on the kind of
computation -- 20% to 50% higher performance. If it is possible to make
use of this performance increase, a real-time simulation could be achieved
with well beyond 200,000 agents; a stadium size (40,000 agents) evacuation
simulation in 5% to 15% of real time, as evacuation always implies that
the average number of active agents during the course of the simulation is
roughly half the initial number.
References
[1] R. Wiedemann, "Simulation des Strassenverkehrsflusses",
Schriftenreihe des IfV 8 (1974) . Institut fur Verkehrswesen.
Universitat Karlsruhe.
[2] K. Nagel, "Particle Hopping Models and Traffic Flow Theory", Phys.
Rev. E 53 (1996) 4655 -- 4672, cond-mat/9509075v1.
[3] D. Chowdhury, L. Santen, and A. Schadschneider, "Statistical Physics
of Vehicular Traffic and Some Related Systems", Phys. Rep. 329
(2000) 199 -- 329, cond-mat/0007053v1.
[4] D. Helbing, "Traffic and related self-driven many-particle systems",
Rev. Mod. Phys. 73 (2001) no. 4, 1067 -- 1141, cond-mat/0012229v2.
5
[5] A. Schadschneider, W.W.F. Klingsch, H. Klupfel, T. Kretz,
C. Rogsch, and A. Seyfried, "Evacuation Dynamics: Empirical
Results, Modeling and Applications", in Meyers [29], p. 3142.
arXiv:0802.1620 [physics.soc-ph]. ISBN:978-0-387-75888-6.
[6] A. Schadschneider, H. Klupfel, T. Kretz, C. Rogsch, and A. Seyfried,
"Fundamentals of Pedestrian and Evacuation Dynamics", in Bazzan
and Klugl [28], ch. VI, pp. 124 -- 154. ISBN:978-1-60566-226-8.
[7] A. Seyfried, B. Steffen, W.W.F. Klingsch, and M. Boltes, "The
fundamental diagram of pedestrian movement revisited", J. Stat.
Mech 10002 (2005) , arXiv:physics/0506170 [physics.soc-ph].
[8] U. Chattaraj, A. Seyfried, and P. Chakroborty, "Comparison of
Pedestrian Fundamental Diagram Across Cultures", Advances in
Complex Systems 12 (2009) 393 -- 405, arXiv:0903.0149
[physics.soc-ph].
[9] A. Kirchner, H. Klupfel, K. Nishinari, A. Schadschneider, and
M. Schreckenberg, "Discretization Effects and the Influence of
Walking Speed in Cellular Automata Models for Pedestrian
Dynamics", J. Stat. Mech. P10011 (2004) ,
arXiv:cond-mat/0410706 [cond-mat.stat-mech].
[10] A. Schadschneider and M. Schreckenberg, "Garden of Eden States in
Traffic Model", J. of Phys. A 31 (1998) L225 -- L231,
arXiv:cond-mat/9801061 [cond-mat.stat-mech].
[11] N. Rajewsky, L. Santen, A. Schadschneider, and M. Schreckenberg,
"The asymmetric exclusion process: comparison of update
procedures", Journal of statistical physics 92 (1998) no. 1, 151 -- 194,
arXiv:cond-mat/9710316 [cond-mat.stat-mech].
[12] C. Rogsch, A. Schadschneider, and A. Seyfried, "Simulation of human
movement by cellular automata models using different update
schemes", in Human Behaviour in Fire 2009, pp. 543 -- 548.
Interscience, Robinson College, Cambridge, UK, 2009.
ISBN:978-0-9556548-3-1.
[13] C. Rogsch, A. Schadschneider, and A. Seyfried, "How to Select the
Right One - Update Schemes for Pedestrian Movement in Simulation
and Reality", in Traffic and Granular Flow 09, S. Dai et al., ed., p. 0.
Springer Berlin Heidelberg New York, Shanghai, 2010. to be
published in TGF'09 proceedings.
[14] T. Kretz and M. Schreckenberg, "Moore and more and symmetry", in
Waldau et. al. [23], pp. 297 -- 308. arXiv:0804.0318 [cs.MA].
ISBN:978-3-540-47062-5.
6
[15] T. Kretz and M. Schreckenberg, "F.A.S.T. -- Floor field- and
Agent-based Simulation Tool", in Chung and Dumont [26], ch. 8,
pp. 125 -- 135. arXiv:physics/0609097 [physics.comp-ph].
ISBN:978-1420095098.
[16] T. Kretz and M. Schreckenberg, "The F.A.S.T.-Model", in El Yacoubi
et. al. [24], pp. 712 -- 715. arXiv:0804.1893 [cs.MA].
ISBN:3-540-40929-7.
[17] T. Kretz, Pedestrian Traffic -- Simulation and Experiments. PhD
thesis, Universitat Duisburg-Essen, February, 2007.
http://deposit.d-nb.de/cgi-bin/dokserv?idn=983420513.
urn:nbn:de:hbz:464-20070302-120944-7.
[18] T. Kretz, M. Kaufman, and M. Schreckenberg, "Counterflow
Extension for the F.A.S.T.-Model", in Umeo et. al. [25], pp. 555 -- 558.
arXiv:0804.4336 [cs.MA]. ISBN:978-3-540-79991-7.
[19] T. Kretz, "Pedestrian Traffic: on the Quickest Path", Journal of
Statistical Mechanics: Theory and Experiment P03012 (2009) ,
arXiv:0901.0170 [physics.soc-ph].
[20] U. Weidmann, "Transporttechnik der Fussganger -
Transporttechnische Eigenschaften des Fussgangerverkehrs",
Schriftenreihe des IVT 90, ETH Zurich, 3, 1993.
http://www.ped-net.org/uploads/media/weidmann-1993_01.pdf.
Zweite, erganzte Auflage, in German.
[21] S. Buchmuller and U. Weidmann, "Parameters of pedestrians,
pedestrian traffic and walking facilities", Schriftenreihe des IVT 132,
ETH Zurich, 2007. Note: This is an English summary and update of
[20].
[22] OpenMP Architecture Review Board, "Openmp." Digital, 2009.
http://openmp.org/wp/about-openmp/.
[23] N. Waldau, P. Gattermann, H. Knoflacher, and M. Schreckenberg,
eds., Pedestrian and Evacuation Dynamics 2005. Springer Berlin
Heidelberg New York, Vienna, December, 2006.
ISBN:978-3-540-47062-5.
[24] S. El Yacoubi, B. Chopard, and S. Bandini, eds., Cellular Automata -
7th International Conference on Cellular Automata for Research and
Industry, ACRI 2006. Springer-Verlag Berlin Heidelberg, Perpignan,
France, September, 2006. ISBN:3-540-40929-7.
7
[25] H. Umeo, S. Morishita, K. Nishinari, T. Komatsuzaki, and S. Bandini,
eds., Cellular Automata -- 8th International Conference on Cellular
Automata for Research and Industry, ACRI 2008, Proceedings.
Springer-Verlag Berlin Heidelberg, Yokohama, Japan, September,
2008. ISBN:978-3-540-79991-7.
[26] E. Chung and A.-G. Dumont, eds., Transport simulation: Beyond
traditional approaches. EPFL press, Lausanne, CH, 2009.
ISBN:978-1420095098.
[27] C. Rogsch, A. Schadschneider, W.W.F. Klingsch, and
M. Schreckenberg, eds., Pedestrian and Evacuation Dynamics 2008.
Springer Berlin Heidelberg New York, Wuppertal, 2009. ISBN:
978-3-642-04503-5.
[28] A. Bazzan and F. Klugl, eds., Multi-Agent Systems for Traffic and
Transportation Engineering. Information Science Reference, Hershey,
PA, USA, April, 2009. ISBN:978-1-60566-226-8.
[29] R.A. Meyers, ed., Encyclopedia of Complexity and Systems Science.
Springer, Berlin Heidelberg New York, 2009. ISBN:978-0-387-75888-6.
8
|
1303.4695 | 1 | 1303 | 2013-03-15T17:01:47 | NetLogo Implementation of an Evacuation Scenario | [
"cs.MA"
] | The problem of evacuating crowded closed spaces, such as discotheques, public exhibition pavilions or concert houses, has become increasingly important and gained attention both from practitioners and from public authorities. A simulation implementation using NetLogo, an agent-based simulation framework that permits the quickly creation of prototypes, is presented. Our aim is to prove that this model developed using NetLogo, albeit simple can be expanded and adapted for fire safety experts test various scenarios and validate the outcome of their design. Some preliminary experiments are carried out, whose results are presented, validated and discussed so as to illustrate their efficiency. Finally, we draw some conclusions and point out ways in which this work can be further extended. | cs.MA | cs | NetLogo Implementation of an Evacuation Scenario
João Emílio Almeida, Zafeiris Kokkinogenis
Artificial Intelligence and Computer Science
Laboratory (LIACC)
Faculty of Engineering, University of Porto (FEUP)
Porto, Portugal
{ joao.emilio.almeida, pro08017}@fe.up.pt
Abstract - The problem of evacuating crowded closed spaces,
such as discotheques, public exhibition pavilions or concert
houses, has become
increasingly
important and gained
attention both from practitioners and from public authorities.
A simulation implementation using NetLogo, an agent-based
simulation framework that permits the quickly creation of
prototypes, is presented. Our aim is to prove that this model
developed using NetLogo, albeit simple can be expanded and
adapted for fire safety experts test various scenarios and
validate the outcome of their design. Some preliminary
experiments are carried out, whose results are presented,
validated and discussed so as to illustrate their efficiency.
Finally, we draw some conclusions and point out ways in which
this work can be further extended.
Keywords: Agent-Based Simulation, NetLogo, Evacuation
Simulation, Emergency Planning, Building evaluation, Egress.
I.
INTRODUCTION
Evacuating large crowds is a challenge under any
circumstances. The evacuation from large facilities during an
emergency or disaster is a much more complex task because
chaos and panic add to highly density populations even more
uncertainty and stress. Studying crowd behaviour
in
emergency situations is difficult since it often requires
exposing real people to the actual, possibly dangerous,
environment. Fire drills (figure 1) are possible approaches to
study this phenomenon. However, as they hardly recreate the
truly panic conditions, people tend not to take it seriously. A
simulation tool could serve as a viable alternative.
Figure 1. Fire drills
Crowds and pedestrians have been empirically studied for
the past decades [1, 2]. The evaluation methods applied
were based on direct observation, photographs, and time -
lapse films. Apart from behavioural investigations, the main
Rosaldo J. F. Rossetti
Artificial Intelligence and Computer Science
Laboratory (LIACC)
Faculty of Engineering, University of Porto (FEUP)
Porto, Portugal
[email protected]
goal of these studies was to develop computer animated
realistic applications, for
the game
industry, design
elements of pedestrian facilities, or planning guidelines for
architectural building and urban design.
In their common environment pedestrians tend to show
some basic attributes. For example people always try to find
the shortest and easiest way to reach their destination. If
possible they avoid detours, even if the shortest way is
crowded. The basic principle is the "least effort principle",
which means everyone tries to reach their goal as fast as
possible spending the least amount of energy and time.
Observations made for crowds in emergency situations
feature typically the same patterns. As people try to leave the
building as fast as possible, the desired velocity increases
which
formations. As
to some characteristic
leads
nervousness increases there is less concern about comfort
zone and finding the most convenient and shortest way. It is
observable, for example, that if people have to leave a
building in an emergency situation and they don't know the
structure of the building well enough, they would run for the
exit they used as an entrance, even if other exits might be
easier to reach or even safer [1].
In this paper we focus our attention on the evaluation of
to the architectural layout of a closed public space as it can
be a museum, cinema, discotheque or concert hall. Thus are
not taken into consideration typical behavioral theories we
find in crowd simulations. For this work we used the
NetLogo [3] agent-based modelling and simulation platform
to
rapidly prototype simple, yet
realistic, “what-if”
evacuation scenarios.
The rest of the paper is organized as it follows: in the
section two some crowd and evacuation simulation related
works are presented. In the section 3 we present the model
implementation, in section 4 and 5 the model validation and
the experimental result discussion are shown respectively.
Finally in section 6 we give space to some final remarks and
future works.
II. RELATED WORKS
There are three main reasons for developing computer
simulations for crowd behaviours: firstly, to test scientific
theories and hypotheses; secondly, to test design strategies;
and finally, to recreate the phenomena about which to
theorize
[4]. Computer models
for emergency and
evacuation situations have been developed and much
research on panics has been of empirical nature and carried
out by researchers from the social sciences [1, 2, 5, 6, and 7].
Santos et al. [8] offer a critical review of selected
simulation models of evacuation. Also, authors have
identified social sciences approaches that could improve
contemporary simulation models. They argue that social
sciences could provide
important new directions
to
simulation models of emergency evacuations. Helbing et al.
[1, 2] developed a continuous pedestrian model based on
plausible interactions, and pointed out that pedestrian
dynamics shows various collective phenomena, such as
arching, clogging and herding [5]. According to their
findings, every simulation model should reproduce such
behaviours in order to be rather realistic. They believe that
the above models can serve as an example linking collective
behaviour as a phenomenon of mass psychology (from the
socio-psychological perspective) to the view of an emergent
collective pattern of motion (from the perspective of
physics). Their simulations suggest
the optimal
that
behaviour in escape situations is a suitable mixture of
individualistic and herding behaviour. Kaup et al. [9] present
a simulation model for emergency planning and crowd
management purposes. Other simulation studies, combined
with optimization algorithms, aim at
improving
the
evacuation efficiency. This can be achieved in terms of
evacuation times, assessment and analysis of evacuation
plans, as well as routes/path optimizations. Coelho [10] has
deeply studied the mathematical and analytical models,
comparing and analyzing them on a systematic basis. In
Klüpfel et al. [11], the analysis of evacuation processes on-
board is taken into consideration. Here authors apply cellular
automata to reproduce crowd motion for the detection of
possible bottleneck situations during the evacuation process.
Filippoupolitis et al. [12] address the building evacuation
optimization problem. Su et al. [13] developed a discrete-
event computer simulation model for assessing evacuation
programs and provide a comprehensive idea of evacuation
plans for hospital buildings in the event of a possible bomb
threat.
III. MODEL IMPLEMENTATION
In this paper we describe the use of NetLogo as a rapid
prototyping tool for an agent-based evacuation simulation
study of a closed space with the purpose to assess the
building’s architectural layout in terms of evacuation times
and occupants exit rate, only. In this section, we present how
our problem was modelled and which abstractions were used
to achieve it. Furthermore, complexities and constraints
inherent to this problem were identified. From that, a
simplified model of an abstraction of the application domain
was created without losing key aspects, such as certain
degrees of realism. Our purpose is to simulate the evacuation
of a closed space evaluating the layout of its egresses
considering a simple behaviour of the occupants only. The
goal is to evaluate and compare scenarios’ performance,
changing the location and width of the exits using as metric
the total evacuation time.
As agents, we defined the occupants of a closed
environment randomly distributed. All agents have the same
characteristics of adult population. Each agent will head
always toward the nearest exit. If other agents occupy the
space around him, he will wait for his turn to move on. No
overlapping is allowed. This scenario is quite simple but we
can still observe some of the typical arching and clogging
phenomena. The devised scenario consists on a square
configuration of 55 x 55 patches. Considering that each
patch has 1 m width, they will have an area of 1 m2.
Therefore, the space has 55 m x 55 m = 3025 m2. Occupants
are represented as a circle, and only one person per patch is
allowed. Thus, the maximum occupancy density is 1 person
per meter-square. An instantiation of the model can be seen
in figure 2.
Figure 2. NetLogo model implementation
Patches are in one of three states: inside, outside or wall.
Outside patches are black, inside white and the walls are
grey. The space limits, or walls, are special patches that
cannot be used by occupants. Exits are controlled using
“sliders”, and their width might vary from 0 to 1, 2 or 3,
depending on the situations. To make this space more
realistic, the exits were located following the Portuguese law
(detailed in the next section), requiring a certain number and
distance between them. Some exits were placed having their
width controlled using the correspondent slider. Each
position corresponds to a meter. So, an exit with width of 2
units means that the door has 2 meter width. If zero means
that the exit is closed and nobody can go outside through it.
The number of occupants is also variable. For the sake of
simplification, the closed space is represented with no
obstacles inside. This happens sometimes in the real life,
when rock concerts happen in gymnasiums.
When
the simulation setup starts, occupants are
distributed inside the area of interest. After pressing the “go”
bottom, meaning that a complete evacuation order was
given, for instance due to a fire or bomb threat, all occupants
start moving towards the nearest exit. The simulation ends
when all occupants exit the space. The movement of each
occupant is considered to be 1 m/s. If each iteration (or tick)
their respective width will depend on the expected number of
occupants. Passage Unit (PU) concept is basically defined
through empirical studies and determines the minimum
width through which a certain number of egresses can leave
the space safely. It is considered that for each group of 100
occupants one PU is required. In Portuguese Fire Code, the
first PU is 0.9m, second 1.4m and onwards is n×0.6m where
n is the number of PU considered. For the present study, it
was assumed that each PU corresponds to 1m width.
Another constraint regards the distribution of the exits that
must be independent. In other words, exits considered
independent if they are located from each other a distance far
enough to form an angle of no less than 45º to the user
observing both exits.
For the given example, having 1,000 occupants, the
Portuguese Fire Code requires a minimum of 5 independent
exits, with 2 PU each, resulting in a total of 10 PU (10 x
0.6m = 6m width). In our example this width is 10m.
V. EXPERIMENTAL RESULTS AND DISCUSSION
After setting up the model, some experiments were
carried out. Taking advantage of the random location of
occupants, at least three runs were performed for each
scenario. The number of occupants was fixed to 1,000. The
exits varied from 1 to 8, as well as varied their width PU.
The results are presented in Table I.
Table I. Results of experiments using Netlogo
For each of the eight possible exits the parameter (figure
3) value varies from 0 (meaning the door is closed) to 2, the
PU value.
lasts a second, then the occupants shall move from one patch
to another in each tick.
IV. VALIDATION METHODOLOGY
To the validation of the model, mathematical and
empirical knowledge were used. A comparison using
common analytical models (such as Van Bogaert method or
the Predetchenky & Milinskii formulae) was carried out to
validate the data obtained through NetLogo, as explained
below.
A. The Van Bogaert analytical method
The Van Bogaert [10] approach assumes that horizontal
movement speed is 1.5 m/s. Total evacuation time depends
on two main factors: total number of occupants and the flow
of the exits. The maximum time to evacuate the space is
given by:
Tmax = S × I × Fd × H × R × 300 [s] (1)
found
(1) are coefficients
The components of
empirically, representing the various factors that influence
the final result. S depends on the area (in this case S=3), I
reflects the building compartment (in this case is none, thus
I=0.75), Fd is the density factor (Fd = 0.36 for maximum
density), H is related with the building height (H=1 since it
is a one storey building), and R is risk type (R=1 for ordinary
risks). Replacing the coefficients in (1), we have the
following expression:
Tmax = 3 × 0.75 × 0.36 × 1 × 1 × 300 = 243 [s]
So, according this method, the total evacuation time should
be 4 minutes (or to be more precise, 243 seconds).
B. The Predetchenky & Milinskii method
Predetchenky & Milinskii [10, 14] developed analytical
methods to calculate the evacuation speed of crowd. They
were the ground basis to yield the following formulae:
2 − 217 Da + 5 7)/60 (2)
3 + 434 Da
4 − 380 Da
V = (112 Da
VVE = VV × (1,49 − 0,36 Da) (3)
Where V (2) is the walking velocity [m/s] in normal
situation and VVE (2) is the horizontal velocity [m/s] in
emergency situation. Da corresponds to the dimensionless
density given by:
Da = (N×Ph)/A (4)
In (4), N is the total number of persons, Ph the area
occupied by a person (typically is 0.125 m2) and A is the
area of the space in m2 (3025 m2 in the example). Replacing
the values in the equations, first in (4) Da = 0.041322, then in
(2) V = 0.81246 m/s and VVE = 1.19848 m/s.
C. The Portuguese fire code
The Portuguese Fire Code considers, regarding the exits:
a) the maximum distance from any given point to the nearest
exit; b) the number of exits; c) the sum of the exits width.
For this model we assume that the minimum distance
requirement to the exits was valid. The number of exits and
Figure 3. Model parameters
exits widthtotal width#1#2#312345678remarks11112611273125801000000only 1 exit with 1PU212670666656010000102 opposit exits 1 PU each224364366368020000202 opposit exits 2 PU each414358340356010110104 opposit exits 1 PU each428202190197020220204 exits with 2UP each5210187179183202220205 exits with 2PU each6212156170158202222026 exits with 2UP each8216136143142222222228 exits with 2UP eachruns (seconds to exit)PUexits parameters These results show that increasing the number of exits
reduces total evacuation time. There is a relation between the
exits and width, even when the total width is the same, more
exits result in a shorter evacuation time. That is something
someone could guess by intuition. Nonetheless, using
simulations such as the one presented in this work, such an
assumption can be demonstrated and confirmed.
Considering the scenario that better represents the
Portuguese Fire Codes, with 5 exits having 2 PU of width
each, and running simulations three times, the results
obtained were 187, 179, 183, seconds respectively (figure 4).
Comparing the results with the validation reference methods,
described previously, the simulation speed was 1 m/s,
whereas the other methods use 1.5 m/s (i.e. in Van Bogaert)
and 1.19848 m/s (i.e. in Predetchenkii & Milinskii). The
total evacuation time, according
to the Van Bogaert
analytical method should be
than 243 seconds.
less
Therefore, considering the results for the selected scenario
are always smaller than that reference value, even if
occupants speed is slower than the other methods suggest.
So the Portuguese Fire Code requirements give better results
than the other methods.
Figure 4. Simulation results
VI. CONCLUSIONS AND FUTURE WORK
this paper we presented an evacuation model
In
implemented in the NetLogo platform. Although the model
may appear
its conceptualisation
simplistic,
to be
encompasses some aspects of the observed system in the real
world. As such, we are able to observe typical emerging
behaviour patterns during evacuation situations, namely
arching, clogging and other herding phenomena. The
proposed model presents a satisfactory degree of validity
within a plausible range of situations. In order for the model
to achieve higher degree of validity, more realistic features
should be incorporated, both in the agent behaviour model as
well as in the representation of the environment. Regarding
the environment, different layouts should be taken into
account (including rooms, hallways and obstacles). It will
be very desirable
the
implementation of a realistic
fire/smoke diffusion model as well, where heat, visibility
and smoke can be accounted for. This ultimately would
influence occupant’s decisions as their cognitive abilities
would be also affected. Another important issue is the
development of a social force and psychological model to
represent the agent internal state providing a more realistic
simulation of the human decision-making mechanisms, as
was suggested in [1,2]. These features will give a very
reliable and true perspective of an emergency case in a
closed space offering to the designers the possibility to better
understand the dynamics that take place in such situations.
From what we have experienced, our approach is feasible
and may support more complex representations of this kind
of applications, proving to be an important tool for designing
and validation of evacuation plans and strategies.
ACKNOWLEDGMENT
This project has been partially supported by FCT
(Fundação para a Ciência e a Tecnologia), the Portuguese
Agency for R&D, under grants SFRH/BD/72946/2010,
SFRH/BD/67202/2009
REFERENCES
[1] D. Helbing, I. Farkas, P. Molnar and T. Vicsek, “Simulating of
pedestrian crowds
in normal and evacuation situations”,
in
M.Schreckenberg, S.D. Sharma(ed.) Pedestrian and Evacuation
Dynamics. Springer Verlag Berlin and Heidelberg, pp. 21-58, (2001)
[2] D. Helbing, P. Molnar, I. Farkas and K. Bolay: Self organizing
pedestrian movement, in Environment and Planning B: Planning &
Design (2001)
[3] U. Wilensky, NetLogo. http://ccl.northwestern.edu/netlogo/. Center
for Connected Learning
and Computer-Based Modeling,
Northwestern University. Evanston, IL (1999)
[4] X. Pan, C.S. Han, K. Dauber and K.H. Law, A multi-agent based
framework for the simulation of human and social behaviors during
emergency evacuations. AI & Society (2007)
[5] C. W. Reynolds, Flocks, Herds, and Schools: A Distributed
Behavioral Model. Proceedings of SIGGRAPH ’87, Computer
Graphics, 21(4), pages 25-34, July (1987)
[6] A.L Coelho, Modelação de Evacuação de Edifícios Sujeitos à Acção
de um Incêndio (in Portuguese). Ph.D. Dissertation, LNEC, Lisboa
(1997)
[7] E. Cordeiro, A Influência do Comportamento das Pessoas e suas
Limitações na Evacuação dos Edifícios (in Portuguese), LNEC,
Lisboa (2009)
[8] G. Santos and B.E. Aguirre, A Critical Review of Emergency
Evacuation Simulation Models. NIST Workshop on Building
Occupant Movement during Fire Emergencies June 9-10. National
Institute of Standards and Technology, U.S. Department of
Commerce (2004)
[9] D.J. Kaup, T.L. Clarke, L. Malone and R. Olosn, Crowd Dynamics
Simulation Research. Proceedings of
the Summer Computer
Simulation Conference. Calgary, Canada, pp. 365-370. (2006)
[10] A.L Coelho, Incêndios em Edifícios (in Portuguese). Edições Orion,
Lisboa (2010)
[11] H. Klüpfel, T. Meyer-Koning and J. Wahle, Microscopic Simulation
of Evacuation Processes on Passenger Ships
[12] A. Filippoupolitis, E. Gelenbe, D. Gianni, L. Hey, G. Loukas and S.
Timotheou, Distributed Agent-based Building Evacuation Simulator.
Proceedings of the Summer Computer Simulation
Conference.
Edinburgh, Scotland, pp. 46- 53. (2008)
[13] S. Su, S.T. Tsai, C.L.Shih, R.J.Kuo, H.C.Wang and J.C.Chen, Use of
Computer Simulation in the Evacuation System for Hospitals.
Proceedings of the Summer Computer Simulation
Conference.
Edinburgh, Scotland, pp. 205-212. (2008)
[14] V.M. Predtechenskii and A.I. Milinskii. Planning for Foot Traffic in
Buildings. Amerind Publishing Co. Pvt. Ltd., New Delhi (1978)
|
cs/0210023 | 1 | 0210 | 2002-10-25T16:32:59 | Geometric Aspects of Multiagent Systems | [
"cs.MA",
"cs.AI"
] | Recent advances in Multiagent Systems (MAS) and Epistemic Logic within Distributed Systems Theory, have used various combinatorial structures that model both the geometry of the systems and the Kripke model structure of models for the logic. Examining one of the simpler versions of these models, interpreted systems, and the related Kripke semantics of the logic $S5_n$ (an epistemic logic with $n$-agents), the similarities with the geometric / homotopy theoretic structure of groupoid atlases is striking. These latter objects arise in problems within algebraic K-theory, an area of algebra linked to the study of decomposition and normal form theorems in linear algebra. They have a natural well structured notion of path and constructions of path objects, etc., that yield a rich homotopy theory. | cs.MA | cs | Geometric Aspects of Multiagent Systems
Timothy Porter
School of Informatics,
University of Wales Bangor,
Bangor,
Gwynedd, LL57 1UT,
Wales, U.K.
Email: [email protected]
July 27, 2011
Abstract
Recent advances in Multiagent Systems (MAS) and Epistemic Logic within Distributed Systems Theory,
have used various combinatorial structures that model both the geometry of the systems and the Kripke
model structure of models for the logic. Examining one of the simpler versions of these models, interpreted
systems, and the related Kripke semantics of the logic S 5n (an epistemic logic with n-agents), the similarities
with the geometric / homotopy theoretic structure of groupoid atlases is striking. These latter ob jects arise
in problems within algebraic K-theory, an area of algebra linked to the study of decomposition and normal
form theorems in linear algebra. They have a natural well structured notion of path and constructions of
path ob jects, etc., that yield a rich homotopy theory.
In this paper, we examine what an geometric analysis of the model may tell us of the MAS. Also the
analogous notion of path will be analysed for interpreted systems and S 5n -Kripke models, and is compared
to the notion of ‘run’ as used with MASs. Further progress may need adaptions to handle S 4n rather than
S 5n and to use directed homotopy rather than standard ‘reversible’ homotopy.
2
0
0
2
t
c
O
5
2
]
A
M
.
s
c
[
1
v
3
2
0
0
1
2
0
/
s
c
:
v
i
X
r
a
1
Geometric Aspects of Multiagent Systems
Timothy Porter
1
Introduction
In many studies of distributed systems, a multiagent model is used. An agent is a processor, sensor or finite
state machine, interconnected by a communication network with other ‘agents’. Typically each agent has a local
state that is a function of its initial state, the messages received from other agents, observations of the external
environment and possible internal actions. It has become customary when using formal models of distributed
systems to use modal epistemic logics as one of the tools for studying the knowledge of such systems. The
basic such logic for handling a system with n-agents is one known as S 5n . Unless the system is very simple the
actual logic will be an extension of that basic one, that is, it may have more axioms. For instance, the way the
various agents are connected influences the logic in subtle ways. Suppose that agent 1 sends all its information
immediately to agents 2 and 3, then if we denote by Kiφ, the statement that agent i ‘knows’ proposition φ, we
clearly expect within the logic of that system that K1φ ⇒ K2φ ∧ K3φ.
The logic S 5n is obtained from ordinary propositional logic by adding ‘knowledge operators’, Ki as above.
(In the literature the notation Kiφ is often replaced by ✷iφ.) It models a community of ideal knowledge agents
who have the properties of veridical knowledge (everything they know is true), positive introspection (they know
what they know) and negative introspection (they know what they do not know). These properties are reflected
in the axiom system for the logic. The axioms include all propositional tautologies, plus the schemata of axioms:
Ki (φ ⇒ ψ) ⇒ (Kiφ ⇒ Kiψ), Kiφ ⇒ φ, Kiφ ⇒ KiKiφ, and ¬Ki¬φ ⇒ Ki¬Ki¬φ, where i ∈ A, the set of
‘agents’. (We will see an alternative presentation of the logic later on.) Two comments worth making are (i)
several of these axioms and in particular the last one - negative introspection - are considered computationally
infeasible and (ii) ideas such as common knowledge (represented by an additional operator, C ) can be introduced
to give a richer extended language. Here however we will be restrict attention largely to models for S 5n and
extensions that may reflect the geometry of the system being modelled. How is this ‘epistemic analysis used in
practice? We mention three examples. One is given in [19] (§1.9) as due to Halpern and Zuck. It shows the
way in which epistemic operators give compact and exact specifications of protocols that are verifiably ‘safe’.
Another worth mentioning is the analysis of AI data / knowledge searches, such as the Muddy Children problem
(cf. Lomuscio and Ryan, [17]). Finally the study of knowledge based programming, [12], in which languages
one may require statements such as : if i knows φ, set x = 0, by formalising what ‘knows’ means in this context
requires analyses of this type. The book, [19], and several of the papers cited here contains numerous further
examples.
The classical models for multimodal logics, and for S 5n and its extensions in particular, are combinatorial
models known as Kripke frames and, for S 5n , Kripke equivalence frames. These consist of a set W , called the
set of possible worlds, and n-equivalence relations ∼i , one for each agent. The interpretation of ∼i is that if w1 ,
w2 are two possible worlds and w1 ∼i w2 , then agent i cannot tell these two worlds apart. In a series of papers
and books (see in particular [6]) Fagin, Halpern, Moses and Vardi, in various combinations, have put forward a
simpler combinatorial model known as an interpreted system. These have the same formal expressive power as
Kripke frames, but are nearer the intuition of interacting agents than is the more abstract Kripke model.
In each case the underlying frame / set of global states, has a very similar combinatorial structure to that
underlying a structure, global actions (or groupoid atlases ), introduced by A. Bak, [1, 2]. These arose from
an analysis of algebraic problems related to the solution of systems of linear equations over arbitrary rings.
(The mathematical area is algebraic K-theory and lies at the interface between algebraic topology and algebra
/ algebraic geometry). Any action of a group on a set leads to a set of orbits. These are the equivalence classes
for an equivalence by a ‘reachability’ or ‘accessibility’ relation generated by the group action. (Translating and
weakening to a monoid action, one has a variant of the reachability of states in a finite automaton.) In a global
action, the set X is divided up into a family of patches, each of which has a group attached, which acts on
that patch (see below for the more detailed definition). If the patches all coincide the resulting ‘single domain
global action’ is essentially a set with a collection of (possibly independent/ possibly interacting) group actions.
As group actions yield groupoids by a well known construction, and the resulting equivalence relations are also
groupoids, a useful generalisation of global actions is that of groupoid atlas introduced by Bak, Brown, Minian
2
and the author, [3]. These therefore present a context in which both the algebraic ideas and the logical models
of S 5n can fit. Moreover both global actions and groupoid atlases have a rich homotopy theory. This homotopy
theory is based on a notion of path that, suitably modified, bears an uncanny resemblance to that of the ‘runs’
considered in multiagent systems, but seems to be better structured and, in fact, more computationally realistic.
The point of this paper is to examine these models in some more detail and to start the analysis of the
necessary modifications to the global action/groupoid atlas homotopy theory that will allow its application to
the problem of the geometric analysis of multiagent systems: how does the geometry of a multiagent system
influence its inherent logic and thus its computational ability? The author’s hope is that such an analysis
will aid in three specific problem areas: firstly, any analysis of systems such as these hits the combinatorial
explosion problem, the effective state space is too large for efficient search to be implemented. By reducing
the search space via homotopical methods, it is expected that some progress in this can be achieved. Next,
some distributed systems can be modularised thereby aiding verification that their description and behaviour
matches their specification. This modularisable attribute should be identifiable by a combination of algebraic
and geometric tools. A related question here is as to whether or not it is better to group a set of agents together
as one ‘super-agent’ and under what condition can this be done without changing the behaviour of the system
for the worse.1 This, of course, presupposes a mechanism for comparison of MASs with different numbers of
agents, a point to which we will return. The final hope is that a geometric homotopical overview may aid in
the description and handling of know ledge evolution without MAS.
Acknowledgements
The research in this paper would not have been possible without conversations with Tony Bak, George
Janelidze, and various participants at the 2000 category theory conference at the Villa Olmo, Como, and
therefore without the assistance of ARC grant Number 859 and INTAS grants (INTAS 93-436 and 97-31961).
It has also benefitted from conversations and discussions with Eric Goubault, Jean Goubault-Larrecq and other
members of the Geometric Methods in Theoretical Computer Science group.
2 Preliminaries.
(As references for basic modal logic, try Kracht, [13], Meyer and van der Hoek, [19] and Blackburn, de Ryke
and Venema, [4].) In the following, at least to start with, there will be n-agents and A = {1, 2, . . . , n} will
denote the set of such ‘agents’.
S 5 and S 5n .
To introduce these logics fairly formally we suppose given a set of variables and form a language Lω (n) given
by
φ ::= pλ ⊥ ¬φ φ1 ∨ φ2 Miφ
where the pλ are the propositional variable ordered by the finite ordinals λ, and Mi is a modality for each agent
i = 1, . . . , n.
In contrast to some treatments, we are using operators, Mi , corresponding to “possibility” , rather than
“knowledge” operators, i.e. we interpret Miφ as “agent i considers φ is possible”. The relation with Kiφ (“agent
i knows φ”) is Mi = ¬Ki¬φ, “agent i does not know that φ is false”. For computational purposes these may
be expected to yield different methods, since ¬ is not well behaved computationally, however for this paper we
will not be considering computational/ implementational problems, so the M v. K debate need not concern us
greatly.
A logic in Lω (n) is any set Λ of Lω (n)-formulae such that
• Λ includes all Lω (n)-formulae that are instances of tautologies,
and
• Λ is closed under the inference rule
i.e. detachment or modus ponens
if φ, φ → ψ ∈ Λ then ψ ∈ Λ
1Our models of agents tend to work as if they are given ‘atomic’ entities, however if they interact or if they themselves consist
of ‘subagents’, (processors), a different grouping into ‘full agents’ may be beneficial to analysis, optimisation and verification.
3
The logic is uniform if it is closed under the rule of uniform substitution of Lω (n)-formulae for propositional
variables and is normal if it contains the schemata
(K ) Mi (ψ ∨ χ) → Mi (ψ) ∨ Mi (χ)
(N ) ¬Mi (⊥)
and monotonicity (for each i):
if ψ → χ ∈ Λ then Miψ → Miχ ∈ Λ.
As is well known, S 5n is defined to be the smallest normal logic in Lω (n) containing
(T )i φ → Miφ
(4)i MiMiφ → Miφ
and
(B )i φ → KiMiφ,
(so is K 4.BT or S 4.B in the notation used in Kracht, [13], p.72).
Of course φ → ψ is shorthand for ¬φ ∨ ψ
As we are in the classical rather than the intuitionistic case, it is easy to rewrite this in terms of Ki instead
of Mi . The logic S 51 is usually called S 5.
The related logic S 4n , mentioned earlier, does not require the schemata (B )i .
The usual semantics of S 5n is given by Kripke equivalence frames and models .
Kripke equivalence frames
An equivalence frame (or simply frame ) F = (W, ∼1 , . . . , ∼n ) consists of a set W with, for each i ∈ A, an
equivalence relation ∼i on W . Elements of W are called worlds and are denoted w, w′ , etc. We will write [w]i
for the equivalence class of the element w ∈ W for the ith equivalence relation, ∼i . A Kripke frame is very like
a labelled transition system, but it has equivalence relations rather than partial orders as its basic relational
structure. The logic gives a semi-static view of the system. To get a dynamic aspect one needs to look at
knowledge evolution, cf. for example, Lomuscio and Ryan, [17].
An equivalence (Kripke) model (or simply model ) M = (F, π) is a frame F together with a relation
where P = {pλ : λ ∈ N}. This relation yields an interpretation
πR : W → P (P ),
Rπ ⊆ P × W,
which interprets as : πR (w) is the set of basic propositions “true” at w, or dually a valuation
Lπ : P → P (W )
giving : Lπ(p) is the set of worlds at which p is “true”. Of course πR and Lπ contain the same information and
will be merged in notation to π when no confusion will result.
1 , . . . , ∼′
n) is a function f : W → W ′ such
A weak map or weak morphism of frames f : F → F ′ = (W ′ , ∼′
that for each i,
The map f will give a map of models f : (F, π) → (F ′ , π ′ ) if
if w ∼i w′ , then f (w) ∼′
i f (w′ ).
W
W ′
f
"EEEEEEEE
yyyyyyyy
π
π ′
P (P )
commutes.
Weak maps are too weak to react well with the logic so a stronger notion of bounded morphism (or p-
morphism) is also used.
A weak morphism f : F → F ′ of frames is bounded if for each i, 1 ≤ i ≤ n and u ∈ W , v ′ ∈ W ′ ,
i v ′ if and only if there is a v ∈ W with f (v) = v ′ and u ∼i v .
f (u) ∼′
4
/
/
"
Remark:
A discussion of some of the properties of the resulting categories of frames and weak maps (or of frames and
bounded maps) can be found in [21]. In this paper we will not be considering bounded morphisms nor models
in any great detail due to restrictions on space.
Global states for interpreted systems
Interpreted systems were first proposed by Fagin, Halpern, Moses and Vardi, [6] to model distributed systems.
They give simple combinatorial models for some of the formal properties of multiagent systems. As before one
has a set, A = {1, 2, . . . , n}, of agents, and now one assumes each agent i can be in any state of a set Li of local
states. In addition one assumes given a set Le of possible states of the ‘environment’. More formally:
A set of global states (SGS) for an interpreted system is a subset S of the product Le × L1 × . . . × Ln with
each Le , Li non-empty. If S = Le × L1 × . . . × Ln , then the SGS is called a hypercube, cf. [14].
The idea behind allowing the possibility of considering a subset and not just the whole product is that
some points in Q L = Le × Qn
i=1 Li may not be ‘feasible’, because of explicit or implicit constraints present
in the multiagent system (MAS). The explicit way these constraints might arise is usually not considered
central for the general considerations of the multimodal logic approach to MASs, yet it seems clear that it
represents the interconnection of the network of agents and, if the local states are the states of a finite state
automaton, questions of reachability may also arise. This will be where the ‘topology’ of the MAS is most
clearly influencing the combinatorial topology of the model. As a simple example, suppose we have agent 1
acts solely as a sensor for agent 2, so anything agent 1 knows, agent 2 automatically knows, K1φ ⇒ K2φ. The
effect of this can be illustrated where L1 has two local states, x1 and x2 . In x1 , p is true; in x2 , ¬p is true.
Suppose L2 has 5 local states, y1 , . . . , y5 , and p is true only in y1 and y2 , ¬p being true in the remainder. Then
S = {(x1 , y1 ), (x1 , y2 ), (x2 , y3 ), (x2 , y4 ), (x2 , y5 )} is as large a SGS (or more precisely interpreted system as the
valuation plays a role) as one can get within this setting. The situation mentioned earlier, K1φ ⇒ K2φ ∧ K3φ,
will lead to a similar 3-dimensional example. The link between the structure of the SGS, the logic inherent
in the interrelations between agents and the computational power of the system is subtle, see [16] for a set of
examples. Other restrictions may also play a role. Agents may share resources, e.g.
in a context where they
need to access a distributed database and one agent may block another from an otherwise feasible transition.2
Any SGS yields a Kripke frame. If we write L = (Le , L1 , . . . , Ln) and (S, L) for an SGS based on L, then
set F (S, L) to be the frame with S as its set of possible worlds with ∼i defined by:
ℓ ∼i ℓ′ ⇔ ℓi = ℓ′
i ,
i.e. ℓ and ℓ′ correspond to the same local state for agent i. For simplicity we will assume that Le is a singleton
set.
There are notions of weak map and bounded map of SGSs and an adjoint equivalence between the categories
of frames and those of SGSs modulo a notion of essential equivalence. If F = (W, ∼1 , . . . , ∼n ) is an equivalence
frame, then for each agent i, let Wi = W/ ∼i , be the set of equivalence classes of elements of W for the relation,
∼i , and set W = (W1 , . . . , Wn ). There is a ‘diagonal’ function
given by
∆ : W → Y W ,
∆(w) = ([w]1 , . . . , [w]n )
and (∆W, W ) is an SGS. Setting G (F ) = (∆W, W ) gives the functor, left adjoint to F , that is used in [21] to
prove the equivalence mentioned above.
Mathematical Interlude: Global Actions and Groupoid Atlases.
A very similar structure to a Kripke equivalence frame is that of a global action, see [1, 2]. Their generalisation
in [3] to groupoid atlases gives a context where both Kripke equivalence frames and global actions coexist and
it is a situation with a well defined and quite well behaved homotopy theory, therefore it yields a potential tool
for the geometric analysis of MASs.
2For simplicity, it is assumed that each local agent is a reversible transition system. Thus if a transition s → s′ can occur in
Li , then s′
∗ s as well, i.e. we can get back from s′ to s by some sequence of transitions, even if this requires reinitialising Li ,
→
but any given transition in Li may not be feasible at some state of S , being blocked by the actions of other agents, whence the
complication of the system.
5
The prime example of a global action is a set X with a family of groups acting on it. In particular if G is
a group (in the usual mathematical sense) then given a family of subgroups {Hi : i ∈ I } of G, we can consider
the actions of each Hi on the set of elements of G by left multiplication. The important point to note is that
the different subgroups Hi may be related, e.g. we may have Hi ⊆ Hj , which implies structural relationships
between the equivalence relations generated by the actions of Hi and Hj . In more detail, we have for each i ∈ I ,
an equivalence relation ∼i on the set of elements of G defined by
x ∼i y if and only if there is some hi ∈ Hi with x = hiy .
If Hi ⊆ Hj , then x ∼i y implies x ∼j y , which is exactly the sort of relationship that results from ‘knowledge
passing’ within a MAS, (cf. [15]). In a global action or groupoid atlas, this relationship is explicitly specified
from the start. The example above is a single domain global action as there is one set on which all the groups
act. The general form assumes only that the groups act on subsets of the ‘domain’. This adds additional
flexibility and adaptability to the concept. (In addition to the notes, [3], the original definition and discussion
of global actions can be found in [1, 2].)
A global action A consists of a set XA , together with a family {(GA )α y (XA )α α ∈ ΦA} of group actions
on subsets (XA )α ⊆ XA . The various local groups (GA )α and the corresponding subsets (local patches ), (XA )α ,
are indexed by the index set, ΦA , called the coordinate system of A. This set ΦA is equipped with a reflexive
relation, written ≤, and it is required that
- if α ≤ β in ΦA , then (GA )α leaves (XA )α ∩ (XA )β invariant (so (XA )α ∩ (XA )β is a union of equivalence
classes for the (GA )α -action), and
- there is given for each pair α ≤ β , a group homomorphism
(GA )α≤β : (GA )α → (GA )β
such that if σ ∈ (GA )α and x ∈ (XA )α ∩ (XA )β , then
σx = (GA )α≤β (σ)x.
This second axiom says that if α and β are explicitly related and their domains intersect then the two actions
are related on that intersection. Again this is the sort of structural compatibility that arises in MASs, except
that, as so far considered, interpreted systems, etc. do not allow for the ‘multi-patch’ setting.
Any global action yields on each (XA )α an equivalence relation due to the (GA )α -action. The equivalence
classes (local orbits or local components) for these ‘local equivalence relations’ form a structure that is sometimes
useful, regardless of what group actions are used, i.e. we need the local equivalence relations rather than the
local groups that were used to derive them. As both equivalence relations and group actions yield groupoids
(small categories in which all the morphisms are isomorphisms), it is convenient to adapt the notion of global
actions to give a generalisation which handles the local equivalence relations as well. This generalisation is
called a groupoid atlas in [3]. Fuller details of that transition are given in that source.
A groupoid atlas A on a set XA consists of a family of groupoids (GA )α defined with ob ject sets (XA )α
which are subsets of XA . These local groupoids are indexed by an index set ΦA , called the coordinate system of
A, which is equipped with a reflexive relation, written ≤. This data is required to satisfy:
(i) if α ≤ β in ΦA , then (XA )α ∩ (XA )β is a union of components of (GA )α , i.e.
if x ∈ (XA )α ∩ (XA )β , and
g ∈ (GA )α , g : x → y , then y ∈ (XA )α ∩ (XA )β ;
and
(ii) if α ≤ β in ΦA , then there is given a groupoid morphism,
→ (GA )β (cid:12)(cid:12)(cid:12)(XA )α∩(XA )β
(GA )α (cid:12)(cid:12)(cid:12)(XA )α∩(XA )β
defined between the restrictions of the local groupoids to the intersections, and which is the identity on ob jects.
,
Example 1. (From Kripke frames to Groupoid Atlases.)
Let X be a set and ∼i , i = 1, 2, . . . , n, n equivalence relations on X . Then F = (X, ∼1 , . . . , ∼n) is a Kripke
frame, but also, if we specify the local groupoids
Gi := Ob jects X, arrows x1 →i x2 if and only if x1 ∼i x2 ,
6
and Φ to be discrete, i.e. “ ≤ ” = “ = ”, we have a simple groupoid atlas, A(F ), (cf. example 2 §2 of [3]). In
fact later we will introduce a second method for turning a frame into a groupoid atlas.
Example 2. The Line
The simplest non-trivial groupoid is I . This is the groupoid corresponding to the Kripke frame W = {0, 1},
∼ = the indiscrete / universal equivalence relation so 1 ∼ 0. (If the number of equivalence relations / agents
is needed to be kept constant, then set ∼i = ∼ for i = 1, . . . , n. This is sometimes useful, but should not
concern us too much for the moment.
The line, L, is obtained by placing infinitely many copies of I end to end, so
L
:= the set, Z, of integers
Φ := Z ∪ {−∞}, where − ∞ ≤ −∞, − ∞ < n for all n ∈ Z and n ≤ n,
but that gives all related pairs.
What about models?
The above construction (Example 1 and later on its variant) gives us a way to think of Kripke frames and
SGSs as groupoid atlases, but they do not directly consider the interpretations / valuations that are needed
if Kripke models and interpreted systems are to be studied via that combinatorial gadgetry. Given a frame
F = (W, ∼1 , . . . , ∼n ) and an interpretation
we can get a bounded morphism of frames
π : W → P (P ),
π∗ : F → S Λ
P
for Λ = S 5n or an extension. This frame S Λ
P is the canonical frame for the logic Λ with the given set of variables
P . Its ‘possible worlds’ are the Λ-maximal sets of Lω (n)-formulae, and π∗ assigns to a world w the set of φ
such that (F, w) =π φ, i.e. the set of φ valid at the world w given π as interpretation, (cf. [13] p.63). Thus, if
the homotopy theory of frames / SGSs informs us about their ‘geometry’, the homotopy theory of frames over
S Λ
P should inform us of the corresponding ‘geometry’ of the Λ-models in each context. Because of this and our
relative ignorance of ‘homotopy over’ in this context, we will put models aside for this paper and concentrate
on frames and SGSs.
3 Morphisms, Runs, Curves and Paths.
In the previous section, we have seen that groupoid atlases form a class of structures that encompasses Kripke
equivalence frames as well as more general ob jects such as global actions. This by itself need not be useful. The
general notion of a categorical model of a situation demands that serious attention be paid to the morphisms.
We have seen that Kripke frames and interpreted systems have both weak morphisms and bounded morphisms
available for use, the latter preserving more of the internal logic. We therefore need to consider morphisms
of groupoid atlases. The payoff will be if the known function space structure on certain classes of morphisms
between groupoid atlases (cf. [1, 2]) can allow a similar structure to be made available for models of MASs.
A function f : A → B between the underlying ob ject sets of two groupoid atlases is said to support the
structure of a weak morphism if it preserves local frames, (the term comes from the original work on global
actions and is not connected with the model theoretic meaning). Here a local frame in A is a set {x0 , . . . , xp}
of ob jects in some connected component of some (GA )α , i.e. α ∈ Φ and there are arrows gi : x0 → xi in
(GA )α for i = 1, 2, . . . , p. The function f preserves local frames if for {x0 , . . . , xp } is a local frame in A then
{f (x0 ), . . . , f (xp )} is a local frame in B.
Any weak morphism of Kripke frames will give a weak morphism of the corresponding groupoid atlases,
but not conversely since if A = (W, ∼1 , . . . , ∼n ) and B = (W ′ , ∼′
1 , . . . , ∼′
n), the notion of weak morphism of
groupoid atlases allows f : W → W ′ to ignore which agents are involved, i.e. {w0 , . . . , wp } is a local frame in
A if there is some agent i such that w0 ∼i wk , k = 1, . . . , p, so agent i considers these worlds equivalent; if
f preserves this local frame, then there is some agent j such that f (w0 ), . . . , f (wp ) are considered equivalent
by that agent. Note however that agent j need not be the same agent as agent i, nor necessarily have the
same position in the lists of agents if the sets of agents in the two cases are represented by disjoint lists. In
7
fact the number of agents in the context of the two Kripke frames did not actually need to be the same for
a weak morphism to exist between them. This added flexibility would seem to be essential when discussing
modularisation, as mentioned above, but also for interacting MASs and the resulting interaction between the
corresponding epistemic logics, however note these are weak morphisms so the link with the logic here is fairly
weak. There is a notion of strong morphism of groupoid atlases, but this will not be strong enough either for
the logic. In fact Bak’s notion of stong morphisms in this particular context reduces to that of weak morphisms.
The difference is that in a weak morphism the ‘reasons’, i.e. the elements gi above, that a set of ob jects forms a
local frame is not considered a part of the data of the morphism, with strong morphisms this data is recorded.
In the groupoid atlases derived from Kripke frames, there is only one ‘reason’ possible. If it exists, it is unique!
Thus the difference between the two types of morphism can be safely ignored for the moment. We hope to
return to morphisms of groupoid atlases that correspond to bounded morphisms of Kripke frames and SGSs in
a future paper.
Runs (cf. [19], p. 59).
A run in a Kripke model M = (F, π) associated with a distributed system / MAS is simply a finite or infinite
sequence of states s(1) , s(2) , . . . , with s(i) ∈ S , the set of possible worlds of F , here being thought of as being an
SGS for an interpreted (S, L).
A weakness in this definition is that no apparent restriction is put on adjacent states in a run. This thus
ignores essential structure in the SGS, and any link between runs and morphisms is not immediately clear.
Because of this, we will take the view that as formulated, this notion of ‘run’ is not quite adequate for the
analysis of these systems. It needs refining, bringing it nearer to the mathematical notion, not just for aesthetic
reasons but also because it does not do the job for which it was ‘designed’ ! It does work well in some situations
however. If we, for the moment, write x → x′ to mean x ∼i x′ for some i and then extend to the corresponding
category (reflexive, transitive closure) to give x →∗ y , then (cf. again [19], p. 60), for hypercubes with more
than one agent, any two states are related via →∗ .
Lemma 1
If M is a hypercube SGS associated to a distributed system with more than one agent, then given any states s, t
in S , s →∗ t.
Proof
If s = (s1 , . . . , sn ) and t = (t1 , . . . , tn ) then
(s1 , . . . , sn ) → (t1 , s2 , . . . , sn ) → (t1 , . . . , tn ).
The first arrow comes from ∼i , and i 6= 1, the second from ∼1 .
Of course, this argument may fail if (S, L) is not a hypercube as simple examples show.
Proposition 1
If M is a Kripke frame or SGS, considered, as above, as a groupoid atlas, then any weak morphism
for which f (n) = f (0) for al l n < 0, determines a run s(i) = f (i) in S
f : L → M
(cid:4)
(cid:4)
Often the form of the set of global states is not specified that precisely. Sometimes local transition functions
are used so that the Li are transformed into “local transition systems” with the actions involved being coupled
with each other (cf. for instance, the V SK systems of Wooldridge and Lomuscio [18]). The feasible runs would
seem in any case to be those for which s(i+1) is reachable from s(i) so there is some set of transitions in the
various agents that leads from s(i) to s(i+1) , or precisely:
a run (s(k) ) is feasible if for each k = 1, 2, . . . , s(k) →∗ s(k+1) .
Of course, hidden within this notion is a certain potential for concurrency. We do not specify in (s(k) ) how
s(k) becomes s(k+1) except that by some set of local transitions within the state spaces of the different agents,
components of s(k) have changed to become those of s(k+1) and at all times the resulting intermediate list of
local states is a valid one, i.e. is a list of global states within S .
Looking at finer granularity, assume that s(k) and s(k+1) are not linked directly, i.e. s(k) →∗ s(k+1) but it is
not the case that s(k) → s(k+1) .
8
If s, t are two states in (S, L), we will write
H C (s, t) = {x ∈ Y L :
for each i, 1 ≤ i ≤ n, xi = si or ti }
and say this is the hypercube interval between s and t .
Proposition 2
Suppose (s(k) ) is a run in M = (S, L). If H C (s(k) , s(k+1) ) ⊂ S for each k , then there is a morphism
of groupoid atlases satisfying f (n) = s(1) , n ≤ 2
f : L → M
f (2k) = s(k) ,
k = 1, 2, . . .
(cid:4)
In other words, if at each stage, the hypercube interval between adjacent states of a run is contained in S ,
we can replace (s(k) ) by a curve. Within each hypercube interval, there are many possible concurrent paths
between adjacent states of the run. We therefore have not only that a ‘curve’ can be given to represent the
run, but the different representing curves are in some sense ‘homotopic’, i.e. equivalent via deformations (or
interleaving equivalence). Of course, the condition is far from being necessary. If each s(k) →∗ s(k+1) , we could
find a curve, but would not be able to specify it as closely. The intermediate ‘odd’ points of the curve can be
given, up to interleaving equivalence, as in the earlier lemma.
The precise definition of a curve in a groupoid atlas is given as follows:
If A is a groupoid atlas, a (strong) curve in A is a (strong) morphism of groupoid atlases
f : L → A,
so for each n, one gets a β ∈ ΦA and f (im ) : f (n) → f (n + 1) in (GA )β , where we have written f (im ) for fG (im ),
where im : m → m + 1 in (XL )m . N.B. the β and gβ are part of the specification of the strong curve. The
corresponding weak notion of curve only asks for the existence of β and gβ , but does not specify them
A (free) path in A will be a curve that stabilises to a constant value on both its left and right ends, i.e. it is
an f : L → A such that there are integers N − ≤ N + with the property that
for all n ≤ N − , f (n) = f (N −),
for all n ≥ N + , f (n) = f (N +),
Of more use for modelling runs is the notion of a based path (i.e. when a basepoint / initial state is specified,
but no final point is mentioned). Given a basepoint a0 ∈ A, a based path in (A, a0 ) is a ‘free’ path that stabilises
to a0 on the left, i.e. f (N − ) = a0 . We can similarly define a based curve by requiring merely left stabilisation
at a0 . Runs correspond to such curves in which N − = 1.
4 Ob jects of Curves, and Paths
Within the interpreted systems approach to MASs, a set R of runs is often considered as a model (see, for
example, [11] or [12]). The equivalence frame structure given to R may involve the local history of each
processor / agent or merely the various ‘points’ visited at the same time; see the discussion in [19] p.39. The
groupoid atlas viewpoint provides a local frame structure on R that is canonical, but, of course, that will need
evaluating for its relevance to the problems of MASs.
Let A be a groupoid atlas with coordinate system ΦA , underlying set XA , etc, as before. We will write
Curves(A) for the set of curves in A.
If f : L → A is a curve in A, a function β : L → ΦA frames f if β is a function such that
(i) for m ∈ L, f (m) ∈ (XA )β (m) ;
(ii) for m ∈ L, there is a b in ΦA with b ≥ β (m), b ≥ β (m + 1) and a f (im ) : f (m) → f (m + 1) in (GA )b .
Remarks:
(a) The intuition is that the local set containing f in Curves(A) will consist of curves passing through the same
9
local sets (XA )α in the same sequence. The idea of a framing of f is that β picks out the local sets (XA )β (m)
that receive f (m). Condition (ii) then ensures that these choices are compatible with the requirement that f
be a curve.
(b) We have used several times the groupoid atlas associated to a Kripke frame or SGS. The set Φ in that case
was just the set of agents with the discrete order. This use of the discrete order is too simplistic in general as it
hides the relationships between the agents. Mathematically this simple model breaks down first on considering
framings, since the condition (ii) implies b ≥ β (m) and b ≥ β (m + 1) so β (m) = β (m + 1) if the order is discrete,
but then if β is to frame f , f must never have left a single equivalence class of the Kripke frame which was not
the intention! To avoid this silly restriction, we can replace the set of agents by the finite non-empty subsets of
that set.
Kripke frames to Groupoid Atlases revisited.
Suppose F = (X, ∼1 , . . . , ∼n ) is a Kripke frame. Define a new groupoid atlas A′(F ) by :
XA′ (F ) = X , the underlying set of F ;
ΦA′ (F ) = the set of non-empty subsets of A ordered by ⊇, i.e. α ≤ β if α ⊇ β ;
XA′ (F ) α = X , for all α ∈ ΦA′ (F )
and
∼α= T{∼i : i ∈ α}, i.e. the equivalence relation
x ∼α y ⇔ ^
i∈α
(x ∼i y )
Remark.
We can think of A′ (F ) as a ‘subdivision’ of A(F ), rather like the barycentric subdivision of a simplicial
complex, a construction to which it is very closely related. To any global action or groupoid atlas, one can
assign two simplicial complexes; see Appendix. These encode valuable geometric information about the system
and relate to the interaction of the different equivalence classes.
(Fuller details can be found in [3].) Our
subdivision above makes no significant change to the homotopy information encoded in the corresponding
complexes.
This subdivision is just what is needed to encode runs in ‘framings’. Logically, it seems to correspond to the
enrichment of our language with ‘group common knowledge’ operators Kα , α ⊆ A, or dually ‘group possibility’
operators Mα , where
Kαφ = ^
i∈α
Kiφ, etc.
Here it should be possible to adapt the ‘subdivision’ to reflect more closely the geometry of the distributed
system. For instance, not all finite sets of agents might be included as there might be no direct link between
certain of them. The clique complexes used in analyses of scheduling problems in distributed systems and in
the theory of traces may be relevant here.
Now let A be a general groupoid atlas and let AL be the following data for a groupoid atlas:
XAL = Curves(A);
ΦAL = {β : L → ΦA β frames some curve f in A}
For β ∈ ΦAL ,
(XAL )β = {f ∈ Curves(A) β frames f }
(GAL )β = {(σm ) source(σm ) ∈ (XAL )β , σm ∈ (GAL )β (m)}
Note that it is easy to see that target(σm ) is also in (XAL )β in this situation (see lemma in section 4 of [3]).
Finally define
β ≤ β ′ ⇔ β (m) ≤ β ′ (m) for all m ∈ L.
Proposition 3
With the above notation, AL is a groupoid atlas. If A is a global action, then so is AL .
(cid:4)
It is natural to ask if A = A′ (F ) for F a Kripke frame, is AL associated to some Kripke frame. In general the
answer would seem to be no as there will be more than one local ‘patch’, (XAL )β , in this case and the index
set is that of al l framings. In fact this structure is not obviously in the MAS literature. Each framing β of
a curve f in A′(F ) defines a sequence (β (m)) of finite non-empty subsets of the set of agents, satisfying the
10
Kripke frame version of condition (ii) namely that if m ∈ L, there is a b with b ≥ β (m), b ≥ β (m + 1) and
f (m) ∼b f (m + 1). In other words b ⊆ β (m) ∩ β (m + 1) and f (m) ∼b f (m + 1). For a given set of runs, the
framings may reflect a possiblility of some modularisation as they indicate which agents are idle during the run.
This raises an interesting problem of using the framings to optimise use of resources.
Each of the local groupoids in A′ (F )L is an equivalence relation on that local patch. Given f , f ′ ∈ (XA′ (F ) )β ,
so β frames both f and f ′ , they will be equivalent if
f (m) ∼β (m) f ′ (m)
for each m. These linked pairs together with the fact that f (m) ∼b f (m + 1) and f ′ (m) ∼b′ f ′ (m + 1), for some
b, b′ ⊆ β (m) ∩ β (m + 1) give a pattern rather like a ladder of linked ‘squares’. This is more or less a ‘homotopy’
between f and f ′ .
Remark
Perhaps a framing can best be thought of as the sequence of those subsets of A involved in a computation
at each instant. Thus a particular agent may be idle throughout a run if no framing of that run/curve involves
that agent. Sometimes more than one agent is involved in a transition at a particular time step, so if, at time m,
the corresponding set of agents is β (m), this interprets as saying that the two global states f (m) and f (m + 1)
are β (m)-equivalent, i.e. f (m) ∼i f (m + 1) for al l the agents i in β (m). (In a SGS, which we can imagine as
4 , s′
a hypercube for simplicity, then, for example, if n = 5, f (m) = (s1 , s2 , s3 , s4 , s5 ), and f (m) = (s1 , s2 , s3 , s′
5 )
are {1, 2, 3}-equivalent.)
The notion of homotopy between based curves should correspond to that of a path in AL , where we do need
‘path’ not curve so that it stabilises to the given two curves at the two ‘ends’ of that path. We will not explore
this here due to lack of space. The basics of a general treatment of homotopy for groupoid atlases can be found
in [3] and in more detail in [20] (from the point of view of a cylinder based theory as against a cocylinder theory
as would be natural from the viewpoint we have explored here). Another extremely useful source for this type
of theory is [10].
Cartesian closedness?
Clearly the ob ject of paths as defined corresponds to a mapping ob ject with domain L. This raises the
important but difficult question of the cartesian closedness of the category of groupoid atlases and more impor-
tantly of the part of it corresponding to the Kripke frames. Bak has shown [1, 2] that global actions do allow
a function space construction that is well behaved on a large class of examples. A closely related construc-
tion occurs with equilogical spaces as defined by Scott, [23]. These are T0 -spaces together with an equivalence
relation. Kripke equivalence frames for a ‘single agent system’ give equilogical spaces and equilogical spaces
form a cartesian closed category. No analogues of equilogical spaces for systems of n-agents seem to have been
developed. Similarly no analogues are known where different models have different numbers of ‘agents’. Yet
from a logical point of view and for an an adequate logical language to handle multiagent systems, some setting
in which a cartesian closed category structure is available is clearly desirable.
5 Conclusions, Critique and Future Directions
In this paper, I have tried to examine some of the methodological links between the theory of global actions
/ groupoid atlases and the general context of combinatorial models for studying multiagent systems. Within
the space available, no firm conclusions can be reached as to the potential usefulness of these links, but the
possibility of a better structured ob ject of runs in a distributed system has been shown that extends the Kripke
frames of runs considered in the MAS literature.
What has not been done? It is clear that a more detailed examination of homotopy is required, especially
with respect to its interpretation in terms of computation. The problem of cartesian closedness has been noted,
but, deliberately, set aside due to lack of space and firm knowledge. The whole question of the relationship
between these constructions and bounded morphisms (and thus with the logic) has also been set aside. It is
conjectured that bounded morphisms will form a (subclass of the) class of ‘fibrations’ in the homotopy theory
of this context, since, for a bounded morphism f : A → B of frames, for each i, the condition corresponds to
being a fibration of groupoids.
Finally, but crucially, the computational infeasibility of S 5n suggests that a separate study using S 4n ,
probably in an intuitionistic form, cf.
[8, 9], will be worth doing. This will presumably need a version of
11
directed homotopy, but which variant of the many available, cf. [7], will best suit is not yet clear. (The ideas
in [10] are also very relevant here.) Perhaps then, some deeper evaluation of how the geometry of a distributed
system influences its inherent logic and thus its computational ability will become possible.
References
[1] A. Bak, Global Actions: The algebraic counterpart of a topological space , Uspeki Mat. Nauk., English
Translation: Russian Math. Surveys, 525, (1997), 955 –996.
[2] A. Bak, Topological Methods in Algebra , in S. Caenepeel and A. Verschoren, eds., Rings, Hopf Algebras
and Brauer Groups , number 197 in Lect. Notes in Pure and Applied Math, 1998, M. Dekker, New York.
[3] A. Bak, R. Brown, G. Minian, and T. Porter, Global Actions, Groupoid Atlases and Related Topics, Draft
lecture notes, Bangor, 2002.
[4] P. Blackburn, M. de Rijke, and Y. Venema, Modal Logic, Cambridge Tracts in Theoretical Computer
Science, Cambridge U.P., 2001.
[5] C. H. Dowker, Homology Groups of Relations, Annals of Maths. 56 (1952) 84 - 95.
[6] R. Fagin, J. Y. Halpern, Y. Moses and M.Y. Vardi, Reasoning about Know ledge, MIT Press, 1995.
[7] E. Goubault,
Some Geometric Perspectives
in Concurrency Theory,
http://www.di.ens.fr/∼goubault/GOUBAULTpapers.html).
Submitted
(2001)
(Available
from
[8] J. Goubault-Larrecq and E. Goubault, Order theoretic, geometric and combinatorial models fof intuition-
istic S4 proofs, in IMLA’99, 1999.
[9] J. Goubault-Larrecq and E. Goubault, On the geometry of Intuitionistic S4 proofs, submitted to Homotopy,
Homology and its Applications.
[10] M. Grandis, An intrinsic homotopy theory for simplicial complexes, with applications to image analysis,
Appl. Categ. Structures 10 (2002), 99-155.
[11] J. Y. Halpern and Y. Moses, Knowledge and Common knowledge in a Distributed Environment, J. A.C.M.,
37(3) (1990) 549-587.
[12] J. Y. Halpern and Y. Moses, Using counterfactuals in Knowledge-based programming, in: Theoretical
Aspects of Rationality and Knowledge, 1999, ed. Y. Gilboa.
[13] M. Kracht, Tools and Techniques in Modal Logic, Studies in Logic No. 142, Elsevier, 1999.
[14] A. Lomuscio and M. Ryan, On the relation between Interpreted Systems and Kripke Models, Proceedings
of the AI97 Workshop on Theoretical and Practical Foundation of Intel ligent Agents and Agent-Oriented
Systems, Perth (WA), December. Springer lecture notes in AI vol. 1441.
[15] A. Lomuscio and M. Ryan, Ideal Agents Sharing (some!) knowledge, in Proc. ECAI-98. 13th European
Conference on Artificial Intel ligence, John Wiley & Sons, 5pp, 1998.
[16] A. Lomuscio and M. Ryan, A spectrum of modes of knowledge sharing between agents, Proceedings of the
sixth international workshop on Agent Theories, Architectures, and Languages (ATAL-99), Springer LNAI
c(cid:13) volume 1757.
[17] A. Lomuscio and M. D. Ryan, An algorithmic approach to knowledge evolution, Artificial Intel ligence for
Engineering Design, Analysis and Manufacturing (AIEDAM), Vol. 13, No. 2 (Special issue on Temporal
Logic in Engineering), 23pp, 1999.
[18] A. Lomuscio and M. Wooldridge, Multi-Agent VSK Logic, Proceedings of JELIA2000, 7th European Work-
shop on Logic in Artificial Intelligence. Malaga. October. Springer LNAI c(cid:13) volume 1919.
12
[19] J.-J. Meyer and W. van der Hoek, Epistemic Logic for AI and Computer Science, Number 41 in Cambridge
Tracts in Theoretical Computer Science, Cambridge University Press, 1995.
[20] E. G. Minian, Generalized cofibration categories and global actions, K-Theory 20 (2000) 37 - 95.
[21] T. Porter, Interpreted systems and Kripke models for multiagents systems from a categorical perspective,
preprint, UWB, School of Informatics, 2002.
[22] T. Porter, Categorical perspectives on multiagent systems, Research Report, UWB, School of Informatics,
in preparation.
[23] D. S. Scott, A new category? Domains, Spaces and Equivalence Relations, Unpublished Manuscript,
available http://www.cs.cmu.edu/Groups/LTC/.
13
Appendix : Simplicial complexes from Kripke frames
The two constructions mentioned in the main text are classical, dating in their initial forms to the embryonic
algebraic topology of the 1920s and 30s. The local equivalence classes of a Kripke frame give a covering of the
underlying set, X of the frame. It thus gives a relation from X to the set Y of equivalence classes. Abstracting,
let R ⊂ X × Y be a relation. (In our case xRy is exactly x ∈ y , where y is an equivalence class for any of
the equivalence relations). Using a formulation due to Dowker, [5], any such relation determines two simplicial
complexes
(i) K = KR :
- the set of vertices is the set, Y ;
- p-simplex of K is a set {y0 , · · · , yp} ⊆ Y such that there is some x ∈ X with xRyj for j = 0, 1,
· · · , p.
(ii) L = LR :
- the set of vertices is the set X ;
- a p-simplex of K is a set {x0 , · · · , xp } ⊆ X such that there is some y ∈ Y with xiRy for i = 0, 1, · · · , p.
These are, in some sense, dual constructions. In the topological context, KR is often called the Nerve of the
covering and LR the Vietoris complex.
As a simple example, let X = {1, 2, . . . , 6},
a ∼1 b if a − b is a multiple of 2;
a ∼2 b if a − b is a multiple of 3.
This corresponds to a hypercube, L1 × L2 , with L1 having 3 elements, L2 having 2. Y has 5 elements. X
has 6.
KR is a bipartite graph:
{3, 6}
{2, 5}
{1, 4}
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
ffffffffffffffffffffffffffffff
IIIIIIIII
IIIIIIIII
uuuuuuuuu
uuuuuuuuu
{1, 3, 5}
{2, 4, 6}
LR is a prism with two filled triangular faces:
1
4
5
2
3
6
They have both the homotopy type of a figure 8. (For instance in the prism, shrink the triangles to points
and then shrink one vertical edge.)
The main result of Dowker’s paper was that for an arbitrary relation R, the two complexes have the same
homotopy type. The question of the influence of the homotopy type of these complexes on the complexity of
searches in the state space of the original MAS seems to be a very interesting one.
14
|
1612.08048 | 2 | 1612 | 2017-01-19T16:03:42 | Liquid Democracy: An Analysis in Binary Aggregation and Diffusion | [
"cs.MA",
"cs.AI",
"cs.SI"
] | The paper proposes an analysis of liquid democracy (or, delegable proxy voting) from the perspective of binary aggregation and of binary diffusion models. We show how liquid democracy on binary issues can be embedded into the framework of binary aggregation with abstentions, enabling the transfer of known results about the latter---such as impossibility theorems---to the former. This embedding also sheds light on the relation between delegation cycles in liquid democracy and the probability of collective abstentions, as well as the issue of individual rationality in a delegable proxy voting setting. We then show how liquid democracy on binary issues can be modeled and analyzed also as a specific process of dynamics of binary opinions on networks. These processes---called Boolean DeGroot processes---are a special case of the DeGroot stochastic model of opinion diffusion. We establish the convergence conditions of such processes and show they provide some novel insights on how the effects of delegation cycles and individual rationality could be mitigated within liquid democracy.
The study is a first attempt to provide theoretical foundations to the delgable proxy features of the liquid democracy voting system. Our analysis suggests recommendations on how the system may be modified to make it more resilient with respect to the handling of delegation cycles and of inconsistent majorities. | cs.MA | cs |
An Analysis in Binary Aggregation and Diffusion∗
Liquid Democracy:
Zo´e Christoff(cid:63) and Davide Grossi(cid:63)(cid:63)
(cid:63) Department of Philosophy, University of Bayreuth
(cid:63)(cid:63) Department of Computer Science, University of Liverpool
[email protected]
[email protected]
November 5, 2018
Abstract
The paper proposes an analysis of liquid democracy (or, delegable
proxy voting) from the perspective of binary aggregation and of binary
diffusion models. We show how liquid democracy on binary issues can
be embedded into the framework of binary aggregation with abstentions,
enabling the transfer of known results about the latter -- such as impossi-
bility theorems -- to the former. This embedding also sheds light on the
relation between delegation cycles in liquid democracy and the probability
of collective abstentions, as well as the issue of individual rationality in a
delegable proxy voting setting. We then show how liquid democracy on
binary issues can be modeled and analyzed also as a specific process of dy-
namics of binary opinions on networks. These processes -- called Boolean
DeGroot processes -- are a special case of the DeGroot stochastic model of
opinion diffusion. We establish the convergence conditions of such pro-
cesses and show they provide some novel insights on how the effects of
delegation cycles and individual rationality could be mitigated within liq-
uid democracy.
The study is a first attempt to provide theoretical foundations to the
delgable proxy features of the liquid democracy voting system. Our anal-
ysis suggests recommendations on how the system may be modified to
make it more resilient with respect to the handling of delegation cycles
and of inconsistent majorities.
∗Working paper: The paper collects work presented at: Dynamics in Logic IV, TU
Delft, November 2016; seminars at the Computer Science Departments of the University of
Leicester and the University of Oxford, December 2016; Dutch Social Choice Colloquium,
December 2016. The authors wish to thank the participants of the above workshops and
seminars for many helpful suggestions. The authors wish also to thank Umberto Grandi for
many insightful comments on an earlier version of this paper. Both authors acknowledge
support for this research by EPSRC under grant EP/M015815/1.
1
Contents
1 Introduction
2 Binary Aggregation with Abstention
2.1 Opinions and Opinion Profiles . . . . . . . . . . . . . . . . . . . . . . .
2.2 Aggregators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Agenda conditions
2.4 Properties of aggregators
. . . . . . . . . . . . . . . . . . . . . . . . .
2.5 Characterizing quota rules . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . .
2.6
Impossibility in Binary Aggregation with Abstentions
3 Liquid Democracy as Binary Aggregation
3.1 Binary Aggregation via Delegable Proxy . . . . . . . . . . . . . . . . .
3.1.1 Proxy Opinions and Profiles . . . . . . . . . . . . . . . . . . . .
3.1.2 Delegation Graphs . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.3 Proxy Aggregators . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Two Issues of Delegable Proxy . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
Individual & Collective Rationality . . . . . . . . . . . . . . . .
. . . . . . . . . .
3.3.1 One man -- One vote . . . . . . . . . . . . . . . . . . . . . . . .
3.3.2
Impossibility . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Section Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1 Cycles and Abstentions
3.2.2
3.3 Embedding in Binary Aggregation with Abstentions
4 Liquid Democracy as Binary Opinion Diffusion
4.2 Convergence of BDPs
4.1 Binary aggregation and binary influence . . . . . . . . . . . . . . . . .
4.1.1 DeGroot Processes and Opinion Diffusion . . . . . . . . . . . .
4.1.2 Boolean DeGroot processes . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.2 Context: convergence in DeGroot processes . . . . . . . . . . .
4.2.3 Two results . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.4 Liquid Democracy as a BDP . . . . . . . . . . . . . . . . . . .
4.2.5 Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Excursus: unanimity and 2-colorability . . . . . . . . . . . . . . . . . .
4.4 BDPs on logically interdependent issues . . . . . . . . . . . . . . . . .
4.5 Section Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5 Fixpoint Logics for BDPs
Influence graphs as Kripke models
5.1
. . . . . . . . . . . . . . . . . . . .
5.2 Modal µ-calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 On the logic of convergence in BDPs . . . . . . . . . . . . . . . . . . .
6 Conclusions
A Binary aggregation (without abstention)
B Relevant terminology from graph theory
2
3
4
4
5
6
7
8
10
10
10
11
11
12
12
12
14
15
15
16
17
17
18
18
18
19
20
20
21
22
23
23
26
27
28
28
28
29
31
34
35
1
Introduction
Liquid democracy [2] is a form of democratic decision-making considered to
stand between direct and representative democracy.
It has been used, advo-
cated and popularized by local and even national parties (e.g., Demoex1 in
Sweden, and Piratenpartei2 in Germany) to coordinate the behavior of party
representatives in assemblies, as well as campaigns (e.g., Make Your Laws3 in
the US). At its heart is voting via a delegable proxy, also called sometimes tran-
sitive proxy. For each issue submitted to vote, each agent can either cast its
own vote, or it can delegate its vote to another agent -- a proxy -- and that agent
can delegate in turn to yet another agent and so on. This differentiates liquid
democracy from standard proxy voting [24, 27], where proxies cannot delegate
their vote further. Finally, the agents that decided not to delegate their votes
cast their ballots (e.g., under majority rule, or adaptations thereof), but their
votes now carry a weight consisting of the number of all agents that, directly or
indirectly, entrusted them with their vote.
Scientific context and contribution Analyses of standard (non-delegable)
proxy voting from a social choice-theoretic perspective -- specifically through the
theory of spatial voting -- have been put forth in [1] and [17]. Delegable proxy
has not, to the best of our knowledge, been object of study so far, with the
notable exception of [6] which focuses specifically on algorithmic aspects of a
variant of liquid democracy (which the authors refer to as viscous democracy)
with applications to recommender systems.
The objective of the paper is to provide a first analysis, via formal methods,
of the liquid democracy voting system based on delegable proxy. This, we
hope, should point to a number of future lines of research and stimulate further
investigations into this and related systems.
Outline The paper starts in Section 2 by introducing some preliminaries on
the theory of binary aggregation, which is the framework of reference for this
study. It is then structured in two parts. This preliminary section presents also
novel results on binary aggregation with abstentions. The first part (Section
3) studies voting in liquid democracy from the point of view of the delegation
of voting power: we study delegable proxy aggregators using the machinery of
binary and judgment aggregation. This allows us to shed novel light on some is-
sues involved in the liquid democracy system, in particular: the issue of circular
delegation, and the issue of individual irrationality when voting on logically in-
terdependent issues. The second part (Sections 4 and 5) studies voting in liquid
democracy as a very specific type of opinion diffusion on networks, whereby del-
egation is rather interpreted as the willingness to copy the vote of a trustee. We
show that this perspective provides some interesting insights on how to address
1demoex.se/en/
2www.piratenpartei.de
3www.makeyourlaws.org
3
the above mentioned issues of circular delegations and individual irrationality.
Section 6 concludes the paper and outlines some on-going lines of research.
2 Binary Aggregation with Abstention
The formalism of choice for this paper is binary aggregation [14] with absten-
tion.4 This preliminary section is devoted to its introduction.
2.1 Opinions and Opinion Profiles
A binary aggregation structure (BA structure) is a tuple A = (cid:104)N, P, γ(cid:105) where:
• N = {1, . . . , n} is a non-empty finite set individuals s.t. N = n ∈ N;
• P = {p1, . . . , pm} is a non-empty finite set of issues (P = m ∈ N), each
represented by a propositional atom;
• γ ∈ L is an (integrity) constraint, where L is the propositional language
constructed by closing P under a functionally complete set of Boolean
connectives (e.g., {¬,∧}).
write ±p to denote one element from {p,¬p}, and ±P to denote(cid:83)
An opinion function O is an assignment acceptance/rejection values (or,
truth values) to the set of issues P. Thus, O(p) = 0 (respectively, O(p) = 1) in-
dicates that opinion O rejects (respectively, accepts) the issue p. Syntactically,
the two opinions correspond to the truth of the literals p or ¬p. For p ∈ P we
p∈P {p,¬p},
which we will refer to as the agenda of A. Allowing abstention in the frame-
work of binary aggregation amounts to considering incomplete opinions: an
incomplete opinion is a partial function from P to {0, 1}. We will study it as a
function O : P → {0, 1,∗} thereby explicitly denoting the undetermined value
corresponding to abstention.
We say that the incomplete opinion of an agent i is consistent if the set of
formulas {p Oi(p) = 1} ∪ {¬p Oi(p) = 0} ∪ {γ} can be extended to a model
of γ (in other words, if the set is satisfiable).
Intuitively, the consistency of
an incomplete opinion means that the integrity constraint is consistent with
i's opinion on the issues she does not abstain about. We also say that an
incomplete opinion is closed whenever the following is the case:
if the set of
propositional formulas {p Oi(p) = 1}∪{¬p Oi(p) = 0}∪{γ} logically implies
p (respectively, ¬p), then Oi(p) = 1 (respectively, Oi(p) = 0). That is, individual
opinions are closed under logical consequence or, in other words, agents cannot
abstain on issues whose acceptance or rejection is dictated by their expressed
opinions on other issues. The set of incomplete opinions is denoted O∗ and
the set of consistent and closed incomplete opinions O∗
c . As the latter are the
opinions we are interested in, we will often refer to them simply as individual
opinions.
4The standard framework of binary aggregation without abstention is sketched in the
appendix for ease of reference.
4
An opinion profile O = (O1, . . . , On) records the opinion, on the given set
of issues, of every individual in N . Given a profile O the ith projection O is
denoted Oi (i.e., the opinion of agent i in profile O). Let us introduce some
more notation. We also denote by O(p) = {i ∈ N Oi(p) = 1} the set of agents
accepting issue p in profile O, by O(¬p) = {i ∈ N Oi(p) = 0} and by O(±p) =
O(p) ∪ O(¬p) the set of non-abstaining agents. We write O =−i O(cid:48) to denote
that the two profiles O and O(cid:48) are identical, except for possibly the opinion of
voter i.
2.2 Aggregators
c )N → O∗,
Given a BA structure A, an aggregator (for A) is a function F : (O∗
mapping every profile of individual opinions to one collective (possibly incom-
plete) opinion.5 F (O)(p) denotes the outcome of the aggregation on issue p.
The benchmark aggregator is the issue-by-issue strict majority rule (maj), which
accepts an issue if and only if the majority of the non-abstaining voters accept
that issue:
1 if O(p) > O(¬p)
0 if O(¬p) > O(p)
∗
otherwise
maj(O)(p) =
(1)
We will refer to this rule simply as 'majority'.
Majority can be thought of as a quota rule.
In general, quota rules in
binary aggregation with abstention are of the form: accept when the proportion
of non-abstaining individuals accepting is above the acceptance-quota, reject
when the proportion of non-abstaining individuals is above the rejection-quota,
and abstain otherwise:6
Definition 1 (Quota rules). Let A be an aggregation structure. A quota rule
(for A) is defined as follows, for any issue p ∈ P, and any opinion profile
O ∈ O∗:7
1
0
∗
F (O)(p) =
if O(p) ≥ (cid:100)q1(p) · O(±p)(cid:101)
if O(¬p) ≥ (cid:100)q0(p) · O(±p)(cid:101)
otherwise
(2)
where for x ∈ {0, 1}, qx is a function qx : P → (0, 1] ⊂ Q assigning a positive
rational number smaller or equal to 1 to each issue, and such that, for each
5It is therefore worth stressing that, in this paper, we study aggregators that are resolute
(that is, output exactly one value), even though they allow for collective abstention.
6There are several ways to think of quota rules in the presence of abstentions. Instead of
a quota being a proportion of non-abstaining agents, one could for instance define rules with
absolute quotas instead: accept when at least n agents accept, independently of how many
agents do not abstain. In practice, voting rules with abstention are often a combination of
those two ideas: accept an issue if a big enough proportion of the population does not abstain,
and if a big enough proportion of those accept it.
7The definition uses the ceiling function (cid:100)x(cid:101) denoting the smallest integer larger than x.
5
p ∈ P:
(3)
A quota rule is called: uniform if, for all pi, pj ∈ P, qx(pi) = qx(pj); it is called
symmetric if, for all p ∈ P, q1(p) = q0(p).
qx(p) > 1 − q(1−x)(p),
Notice that the definition excludes trivial quota.8 It should also be clear
c )N → O∗ as desired.9
that, by (3) the above defines an aggregator of type (O∗
Notice also that if the rule is symmetric, then (3) forces qx > 1
2 .
(cid:109)
(cid:108)O(±p)+1
2 < q1, q0 ≤ 1
2
Example 1. The majority rule (1) is a uniform and symmetric quota rule where
q1 and q0 are set to meet the equation (cid:100)q1(p) · O(±p)(cid:101) = (cid:100)q0(p) · O(±p)(cid:101) =
, for any issue p and profile O. This is achieved by setting the quota
as 1
maj as a class of quota rules yielding the same collective opinions.
N+1
2N . More precisely one should therefore consider
2 + 1N =
Example 2. The uniform and symmetric unanimity rule is defined by setting
q1 = q0 = 1. A natural uniform but asymmetric variant of unanimity can be
obtained by setting q1 = 1 and q0 = 1N .
Let us finally note an important difference between quota rules in binary
aggregation with abstentions vs. without abstentions. In a framework without
abstentions quota rules are normally defined by a unique acceptance quota q1,
the rejection quota being uniquely determined as q0 = 1−q1. As a consequence,
the majority rule, when N is odd, is the only unbiased quota rule in the
standard framework. This is no longer the case when abstentions are considered.
A novel characterization of the majority rule will be given in Section 2.5.
2.3 Agenda conditions
Definition 2 (simple/evenly negatable agenda). An agenda ±P is said to be
simple if there exists no set X ⊆ ±P such that: X ≥ 3, and X is minimally
γ-inconsistent, that is:
• X is inconsistent with γ
• For all Y ⊂ X, Y is consistent with γ (or, γ-consistent).
An agenda is said to be evenly negatable if there exists a minimal γ-inconsistent
set X ⊆ ±P such that for a set Y ⊆ X of even size, X\Y ∪ {¬p p ∈ Y } is
γ-consistent. It is said to be path-connected if there exists p1, . . . , pn ∈ ±P
such that p1 =c p2, . . . , pn−1 =x pn where pi =c pi+1 (conditional entailment)
denotes that there exists X ⊆ ±P, which is γ-consistent with both pi and ¬pi+1,
and such that {p} ∪ X ∪ {γ} logically implies pi+1.
8Those are quotas with value 0 (always met) or > 1 (never met). Restricting to non-trivial
quota is not essential but simplifies our exposition.
9What needs to be avoided here is that both the acceptance and rejection quota are set so
low as to make the rule output both the acceptance and the rejection of a given issue
6
We refer the reader to [18, Ch. 2] for a detailed exposition of the above
conditions. We provide just a simple illustrative example.
Example 3. Let P = {p, q, r} and let γ = (p ∧ q) → r. ±P is not simple. The
set {p, q,¬r} ⊆ ±P is inconsistent with γ, but none of its subsets is.
2.4 Properties of aggregators
We start by recalling some well-known properties of aggregators from the judg-
ment aggregation literature, adapted to the setting with abstention:
Definition 3. Let A be an aggregation structure. An aggregator F : (O∗
O∗ is said to be:
unanimous iff for all p ∈ P, for all profiles O and all x ∈ {0, 1,∗}: if for all
i ∈ N, Oi(p) = x, then F (O)(p) = x. I.e., if everybody agrees on a value,
that value is the collective value.
c )N →
anonymous iff for any bijection µ : N → N , F (O) = F (Oµ), where Oµ =
(cid:11). I.e., permuting opinions among individuals does not
(cid:10)Oµ(1), . . . , Oµ(n)
affect the output of the aggregator.
p-dictatorial iff there exists i ∈ N (the p-dictator) s.t. for any profile O, and
all x ∈ {0, 1}, Oi(p) = x iff F (O)(p) = x. I.e., there exists an agent
whose definite opinion determines the group's definite opinion on p. If
F is p-dictatorial, with the same dictator on all issues p ∈ P, then it is
called dictatorial.
p-oligarchic iff there exists C ⊆ N (the p-oligarchs) s.t. C (cid:54)= ∅ and for any
profile O, and any value x ∈ {0, 1}, F (O)(p) = x iff Oi(p) = x for all
i ∈ C. I.e., there exists a group of agents whose definite opinions always
determine the group's definite opinion on p. If F is p-oligarchic, with the
same oligarchs on all issues p ∈ P, then it is called oligarchic.
monotonic iff, for all p ∈ P and all i ∈ N : for any profiles O, O(cid:48), if O =−i
if F (O)(p) = 1, then
i(p) ∈ {0,∗}, then: if F (O)(p) =
I.e., increasing support for a definite collective
O(cid:48): (i) if Oi(p) (cid:54)= 1 and O(cid:48)
F (O(cid:48))(p) = 1; and (ii) if Oi(p) (cid:54)= 0 and O(cid:48)
0, then F (O(cid:48))(p) = 0.
opinion does not change that collective opinion.
i(p) ∈ {1,∗}, then:
independent iff, for all p ∈ P, for any profiles O, O(cid:48): if for all i ∈ N, Oi(p) =
O(cid:48)
i(p), then F (O)(p) = F (O(cid:48))(p). I.e., the collective opinion on each issue
is determined only by the individual opinions on that issue.
neutral iff, for all p, q ∈ P, for any profile O: if for all i ∈ N , Oi(p) = Oi(q),
I.e., all issues are aggregated in the same
then F (O)(p) = F (O)(q).
manner.
systematic iff it is neutral and independent.
I.e., the collective opinion on
issue p depends only on the individual opinions on this issue.
7
responsive iff for all p ∈ P, there exist profiles O, O(cid:48) such that F (O)(p) = 1
and F (O(cid:48))(p) = 0. I.e., the rule allows for an issue to be accepted for
some profile, and rejected for some other.
unbiased iff for all p ∈ P, for any profiles O, O(cid:48) : if for all i ∈ N , Oi(p) = 1 iff
i(p) = 0 (we say that O(cid:48) is the "reversed" profile of O), then F (O)(p) =
O(cid:48)
1 iff F (O(cid:48))(p) = 0. I.e., reversing all and only individual opinions on
an issue p (from acceptance to rejection and from rejection to acceptance)
results in reversing the collective opinion on p.
rational iff for any profile O, F (O) is consistent and closed. I.e., the aggre-
gator preserves the constraints on individual opinions.
Example 4. It is well-known that majority is not rational in general. The
standard example is provided by the so-called discursive dilemma, represented
by the BA structure (cid:104){1, 2, 3} ,{p, q, r} , r ↔ (p ∧ q)(cid:105). The profile consisting of
O1 = p ∧ q ∧ r, O2 = p ∧ ¬q ∧ ¬r, O3 = ¬p ∧ q ∧ ¬r, returns an inconsistent
majority opinion maj(O) = p ∧ q ∧ ¬r (cf. [18]).
Finally, let us defined also the following property. The undecisiveness of
an aggregator F on issue p for a given aggregation structure is defined as the
number of profiles which result in collective abstention on p:
c F (O)(p) = ∗}.
u(F )(p) = {O ∈ O∗
(4)
2.5 Characterizing quota rules
As a typical example, consider the aggregator maj: it is unanimous, anonymous,
monotonic, systematic, responsive and unbiased, but, as mentioned above, it
is not rational in general. However, it can be shown (cf.
[18, 3.1.1]) that
aggregation by majority is collectively rational under specific assumptions on
the constraint:
Fact 1. Let A be a BA structure with a simple agenda. Then maj is rational.
Proof. If the agenda ±P is simple, then all minimally inconsistent sets have
cardinality 2, that is, are of the form {ϕ,¬ϕ} such that ϕ = ¬ϕ for ϕ, ψ ∈ P.
W.l.o.g. assume ϕ = pi and ψ = pj. Suppose towards a contradiction that
there exists a profile O such that maj(O) is inconsistent, that is, maj(O)(pi) =
maj(O)(pj) = 1, and ϕ = ¬ψ. By the definition of maj (1) it follows that
O(pi) > O(¬pi) and O(pj) > O(¬pj). Since pi = ¬pj by assumption,
and since individual opinions are consistent and closed, O(¬pj) ≥ O(pi) and
O(¬pi) ≥ O(pj). From the fact that O(pi) > O(¬pi) we can thus conclude
that O(¬pj) > O(pj). Contradiction.
May's theorem [23] famously shows that for preference aggregation, the ma-
jority rule is in fact the only aggregator satisfying a specific bundle of desirable
properties. A corresponding characterization of the majority rule is given in
8
judgment aggregation without abstention: when the agenda is simple, the ma-
jority rule is the only aggregator which is rational, anonymous, monotonic and
unbiased [18, Th. 3.2]. We give below a novel characterization theorem, which
takes into account the possibility of abstention (both at the individual and at
the collective level). To the best of our knowledge, this is the first result of this
kind in the literature on judgment and binary aggregation with abstention.
We first prove the following lemma:
Lemma 1. Let F be a uniform and symmetric quota rule for a given A. The
following holds: 1
if and only if F = arg minG u(G)(p), for
all p ∈ P.
2 < q1 = q0 ≤ N+1
2N
Proof. The claim is proven by the following series of equivalent statements.
(a) A uniform and symmetric quota rule F has quota such that 1
2 < q1 =
q0 ≤ N+1
2N . (b) A uniform and symmetric quota rule F has quota such that
(cid:100)q1(p)O(±p)(cid:101) = (cid:100)q0(p)O(±p)(cid:101) =
for any profile O and issue p.
(c) For any O ∈ O∗
c and p ∈ P, u(O)(p) = ∗ if and only if O(p) = O(¬p),
that is, an even number of voters vote and the group is split in half. (d) F =
arg minG u(G)(p), for all p ∈ P.
(cid:108)O(±p)+1
(cid:109)
2
That is, the quota rule(s) corresponding to the majority rule (Example 1) is
precisely the rule that minimizes undecisiveness.
We can now state and prove the characterization result:
Theorem 1. Let F : (O∗
holds:
c )N → O∗ be an aggregator for a given A. The following
1. F is a quota rule if and only if it is anonymous, independent, monotonic,
and responsive;
2. F is a uniform quota rule if and only if it is a neutral quota rule;
3. F is a symmetric quota rule if and only if it is an unbiased quota rule;
4. F is the majority rule maj if and only if it is a uniform symmetric quota
rule which minimizes undecisiveness.
Proof. Claim 1 Left-to-right: Easily checked. Right-to-left: Let F be an
anonymous, independent, monotonic, and responsive aggregator. By anonymity
and independence, for any p ∈ P, and any O ∈ O∗
c , the only information deter-
mining the value of F (O)(p) are the integers O(p) and O(¬p).
By responsiveness, there exists a non-empty set of profiles S1 = {O ∈
O∗F (O)(p) = 1}. Pick O to be any profile in S1 with a minimal value of
O(p)
O(±p) and call this value q1. Now let O(cid:48) be any profile such that O(cid:48) =−i O and
O(cid:48)(p)
O(cid:48)(±p) > q1. This implies that Oi(p) = 0 and O(cid:48)
i(p) = 1. By monotonicity, it
follows that F (O(cid:48))(p) = 1. By iterating this argument a finite number of times
O(p)
O(±p) ≥ q1, we have that F (O)(p) = 1. Given
we conclude that whenever
9
that q1 was defined as a minimal value, we conclude also that if F (O)(p) = 1,
then O(p)
O(p±) ≥ q1. The argument for q0 is identical.
Claims 2 & 3 follow straightforwardly from the definitions of uniform quota
rule (Definition 1) and of neutrality (Definition 3) and, respectively, from the
definitions of symmetric quota rules (Definition 1) and of unbiasedness (Defini-
tion 3) .
1
Claim 4 Left-to-right. Recall that maj is defined by quota 1
2 < q1 = q0 ≤
2 + 1N (Example 1). It is clear that maj is uniform and symmetric. The claim
then follows by Lemma 1. Right-to-left. By Lemma 1 if an aggregator minimizes
2 + 1N . These quota
undecisiveness then its quota are set as 1
define maj (Example 1).
2 < q1 = q0 ≤ 1
By the above theorem and Fact 1, it follows that, on simple agendas, majority
is the only rational aggregator which is also responsive, anonymous, systematic
and monotonic.
2.6
Impossibility in Binary Aggregation with Abstentions
The following is a well-know impossibility result concerning binary aggregation
with abstentions:
Theorem 2 ([11, 10]). Let A be a BA structure whose agenda is path connected
c )N → O∗ is independent,
and evenly negatable. Then if an aggregator F : (O∗
unanimous and collectively rational, then it is oligarchic.
We will use this result to illustrate how impossibility results from binary
aggregation with abstentions apply to delegable proxy voting on binary issues.
3 Liquid Democracy as Binary Aggregation
In this section we provide an analysis of liquid democracy by embedding it in the
theory of binary aggregation with abstentions presented in the previous section.
To the best of our knowledge, this is the first attempt at providing an analysis
of delegable proxy voting using social-choice theoretic tools, with the possible
exception of [16].
In what follows we will often refer to delegable proxy voting/aggregation
simply as proxy voting/aggregation.
3.1 Binary Aggregation via Delegable Proxy
In binary aggregation with proxy, agents either express an acceptance/rejection
opinion or delegate such opinion to a different agent.
10
3.1.1 Proxy Opinions and Profiles
Let a BA structure A be given and assume for now that γ = (cid:62), that is, all issues
are logically independent. An opinion O : P → {0, 1} ∪ N is an assignment of
either a truth value or another agent to each issue in P, such that Oi(p) (cid:54)= i
(that is, self-delegation is not an expressible opinion). We will later also require
proxy opinion to be individually rational, in a precise sense (Section 3.2.2). For
simplicity we are assuming that abstention is not a feasible opinion in proxy
voting, but that is an assumption that can be easily lifted in what follows.
We call functions of the above kind proxy opinions to distinguish them from
standard (binary) opinions, and we denote by P the set of all proxy opinions, Pc
the set of all consistent proxy opinions, P N being the set of all proxy profiles.
3.1.2 Delegation Graphs
Each profile O of proxy opinions (proxy profile in short) induces a delegation
graph GO = (cid:104)N,{Rp}p∈P(cid:105) where for i, j ∈ N :
(cid:26) Oi(p) = j
Oi(p) ∈ {0, 1}
iRpj ⇐⇒
if i (cid:54)= j
otherwise
(5)
The weight of an agent i w.r.t. p in a delegation graph GO is given by its
p (i.e., the reflexive and transitive closure of Rp):10
The expression iRpj stands for "i delegates her vote to j on issue p". Each Rp
is a so-called functional relation. It corresponds to the graph of an endomap on
N . So we will sometimes refer to the endomap rp : N → N of which Rp is the
graph. Relations Rp have a very specific structure and can be thought of as a
set of trees whose roots all belong to cycles (possibly loops).
indegree with respect to R∗
wO
vidual carries the same weight, independently of the structure of the delegation
graph. Alternative definitions of weight are of course possible and we will come
back to this issue later.11
For all p ∈ P, we also define the function gp : N → ℘(N ) such that gp(i) =
pi(cid:9). This definition of weight makes sure that each indi-
pi and (cid:64)k : jRpk(cid:9). The function associates to each agent i (for a
i (p) = (cid:8)j ∈ N jR∗
(cid:8)j ∈ N jR∗
given issue p), the (singleton consisting of the) last agent reachable from i via a
path of delegation on issue p, when it exists (and ∅ otherwise). Slightly abusing
notation we will use gp(i) to denote an agent, that is, the guru of i over p when
gp(i) (cid:54)= ∅.
If gp(i) = {i} we call i a guru for p. Notice that gp(i) = {i} iff
rp(i) = i, that is, i is a guru of p iff it is a fixpoint of the endomap rp.
If the delegation graph GO of a proxy profile O is such that, for some Rp,
there exists no i ∈ N such that i is a guru of p, we say that graph GO (and
profile O) is void on p. Intuitively, a void profile on p is a profile where no voter
expresses an opinion on p, because every voter delegates her vote to somebody
else.
10 We recall that the reflexive transitive closure R∗ of a binary relation R ⊆ N 2 is the
smallest reflexive and transitive relation that contains R.
11See also footnote 15 below.
11
Given a BA structure A, a proxy aggregation rule (or proxy aggregator) for
A is a function pv : P N → O∗ that maps every proxy profile to one collective
incomplete opinion. As above, pv(O)(p) denotes the outcome of the aggregation
on issue p.
3.1.3 Proxy Aggregators
The most natural form of voting via delegable proxy is a proxy version of the
majority rule we discussed in Section 2:12
1
0
∗
if (cid:80)
if (cid:80)
i (p) >(cid:80)
i (p) >(cid:80)
i∈O(p) wO
i∈O(¬p) wO
otherwise
pvmaj(O)(p) =
i∈O(¬p) wO
i∈O(p) wO
i (p)
i (p)
(6)
(cid:80)
That is, an issue is accepted by proxy majority in profile O if the sum of
the weights of the agents who accept p in O exceeds the majority quota, it
is rejected if the sum of the weights of the agents who reject p in O exceeds
the majority quota, and it is undecided otherwise.
It should be clear that
i (p) = {i ∈ NOgi (p) = 1} (and similarly for ¬p), that is, the sum
of the weights of the gurus accepting (rejecting) p is precisely the cardinality of
the set of agents whose gurus accept (reject) p.
c → O∗ a proxy
In general, it should be clear that for any quota rule F : O∗
variant pvF of F can be defined via an obvious adaptation of (6).
i∈O(p) wO
To fix intuitions further about proxy voting it is worth discussing another
example of aggregator, proxy dictatorship. It is defined as follows, for a given
d ∈ N (the dictator) any proxy profile O and issue p:
pvd(O)(p) =
Ogp(d)
∗
if gp(d) (cid:54)= ∅
otherwise
(7)
(cid:40)
That is, in a proxy dictatorship, the collective opinion is the opinion of the guru
of the dictator, when it exists, and it is undefined otherwise.
3.2 Two Issues of Delegable Proxy
3.2.1 Cycles and Abstentions
It should be clear from the definition of proxy aggregators like pvmaj, that such
aggregators rely on the existence of gurus in the underlying delegation graphs.
If the delegation graph Rp on issue p contains no guru, then the aggregator has
access to no information in terms of who accepts and who rejects issue p. To
avoid bias in favor of acceptance or rejection, such situations should therefore
result in an undecided collective opinion. That is for instance the case of pvmaj.
However, such situations may well be considered problematic, and the natural
question arises therefore of how likely they are, at least in principle.
12On the importance of majority decisions in the current implementation of liquid democ-
racy by liquid feedback cf. [2, p.106].
12
Fact 2. Let A be a BA structure where γ = (cid:62) (i.e., issues are independent).
If each proxy profile is equally probable (impartial culture assumption), then the
probability that, for each issue p, the delegation graph Rp has no gurus tends to
1
e2 as n tends to infinity.
Proof. The claim amounts to computing the probability that a random proxy
profile O induces a delegation graph Rp that does not contain gurus (or equiv-
alently, whose endomap rp : N → N has no fixpoints) as n tends to infinity.
Now, for each agent i, the number of possible opinions on a given issue p (that
is, functions O : {p} → {0, 1}∪ N ) is (N\{i})∪{0, 1} = n + 1 (recall i cannot
express "i" as an opinion). The number of opinions in which i is delegating
her vote is n − 1. So, the probability that a random opinion of i about p is an
opinion delegating i's vote is n−1
n+1 . Hence the probability that a random profile
consists only of delegated votes (no gurus) is ( n−1
n+1 )n. The claimed value is then
established through this series of equations:
(cid:18) n − 1
(cid:19)n
lim
n→∞
n
(cid:19)n
(cid:19)n
(cid:19)n
n + 2
(cid:18) n
(cid:18) 1
(cid:18) 1
(cid:18)
n
n+2
(cid:19)
1 + 2
n
1
(1 + 2
1
n )n
= lim
n→∞
= lim
n→∞
= lim
n→∞
= lim
n→∞
=
=
n )n
limn→∞(1 + 2
1
e2
This completes the proof.
Now contrast the above simple fact with the probability that all agents
abstain on an issue when each voter either expresses a 1 or 0 opinion or abstains
(that is, the binary aggregation with abstentions setting studied earlier). In that
case the probability that everybody abstains tends to 0 as n tends to infinity.
Fact 2 should obviously not be taken as a realistic estimate of the effect of
cycles on collective abstention, as the impartial culture assumption is a highly
idealized assumption. Election data should ideally be used to assess whether del-
egation cycles ever lead large parts of the electorate to 'lose their vote', possibly
together with refinements of the above argument that take into consideration
realistic distributions on proxy profiles, and therefore realistic delegation struc-
tures. Nonetheless, Fact 2 does flag a potential problem of cyclical delegations
as sources of abstention which has, to the best of our knowledge, never been
discussed. The mainstream position on cyclical delegations [2, Section 2.4.1]
is:13
13Cf. also [3].
13
"The by far most discussed issue is the so-called circular delegation
problem. What happens if the transitive delegations lead to a cycle,
e.g. Alice delegates to Bob, Bob delegates to Chris, and Chris del-
egates to Alice? Would this lead to an infinite voting weight? Do
we need to take special measures to prohibit such a situation? In
fact, this is a nonexistent problem: A cycle only exists as long as
there is no activity in the cycle in which case the cycle has no effect.
As already explained [. . . ], as soon as somebody casts a vote, their
(outgoing) delegation will be suspended. Therefore, the cycle natu-
rally disappears before it is used. In our example: If Alice and Chris
decide to vote, then Alice will no longer delegate to Bob, and Chris
will no longer delegate to Alice [. . . ]. If only Alice decides to vote,
then only Alice's delegation to Bob is suspended and Alice would
use a voting weight of 3. In either case the cycle is automatically
resolved and the total voting weight used is 3."
We will discuss later (Section 4) a possible approach to mitigate this issue
by suggesting a different interpretation of liquid democracy in terms of influence
rather than delegation.
3.2.2 Individual & Collective Rationality
In our discussion so far we have glossed over the issue of logically interdepen-
dent issues and collective rationality. The reason is that under the delegative
interpretation of liquid democracy developed in this section individual ratio-
nality itself appears to be a more debatable requirement than it normally is in
classical aggregation.
A proxy opinion Oi is individually rational if the set of formulas
{γ} ∪(cid:8)p ∈ P Ogp(i)(p) = 1(cid:9) ∪(cid:8)¬p ∈ P Ogp(i)(p) = 0(cid:9)
(8)
is satisfiable (consistency), and if whenever (8) entails ±p, then ±p belongs to it
(closedness).14 That is, the integrity constraint γ is consistent with i's opinion
on the issues she does not delegate on, and the opinions of her gurus (if they
exist), and those opinions, taken together, are closed under logical consequence
(w.r.t. the available issues).
The constraint in (8) captures, one might say, an idealized way of how del-
egation works: voters are assumed to be able to check or monitor how their
gurus are voting, and always modify their delegations if an inconsistency arises.
The constraint remains, however, rather counterintuitive under a delegative in-
terpretation of proxy voting. Aggregation via delegable proxy has at least the
potential to represent individual opinions as irrational (inconsistent and/or not
logically closed).
Like in the case of delegation cycles we will claim that the interpretation of
liquid democracy in terms of influence to be developed in Section 4, rather than
14Cf. the definition of individual opinions in Section 2.
14
in terms of delegation, makes individual rationality at least as defensible as in
the classical case.
3.3 Embedding in Binary Aggregation with Abstentions
3.3.1 One man -- One vote
Aggregation in liquid democracy as conceived in [2] should satisfy the principle
that the opinion of every voter, whether expressed directly or through proxy,
should be given the same weight:
"[. . . ] in fact every eligible voter has still exactly one vote [. . . ] unre-
stricted transitive delegations are an integral part of Liquid Democ-
racy. [. . . ] Unrestricted transitive delegations are treating delegat-
ing voters and direct voters equally, which is most democratic and
empowers those who could not organize themselves otherwise" [2,
p.34-36]
In other words, this principle suggests that aggregation via delegable proxy
should actually be 'blind' for the specific type of delegation graph. Making
this more formal, we can think of the above principle as suggesting that the
only relevant content of a proxy profile is its translation into a standard opinion
profile (with abstentions) via a function t : P → O∗ defined as follows:
for
any i ∈ N and p ∈ P: t(Oi(p)) = Ogp(i) if gp(i) (cid:54)= ∅ (i.e., if i has a guru
for p), and t(Oi(p)) = ∗ otherwise. Clearly, if we assume proxy profiles to be
individually rational, the translation will map proxy opinions into individually
rational (consistent and closed) incomplete opinions. By extension, we will
denote by t(O) the incomplete opinion profile resulting from translating the
individual opinions of a proxy profile O.
The above discussion suggests the definition of the following property of
proxy aggregators: a proxy aggregator pv has the one man -- one vote prop-
erty (or is a one man -- one vote aggregator) if and only if pv = t ◦ F for
some aggregator F : O∗
c → O∗ (assuming the individual rationality of proxy
profiles).15
The class of one man -- one vote aggregators can therefore be studied simply
as the concatenation t ◦ F where F is an aggregator for binary voting with
abstentions, as depicted in the following diagram:
a delegation graph. The weight w(i) of i is (cid:80)
15It should be clear that not every proxy aggregator satisfies this property. By means of
example, consider an aggregator that uses the following notion of weight accrued by gurus in
(cid:96)(i,j) where (cid:96)(i, j) denotes the length
of the delegation path linking j to i. This definition of weight is such that the contribution
of voters decreases as their distance from the guru increases. Aggregators of this type are
studied in [6].
j∈R∗(i)
1
15
O
t
pvF
t(O)
F
F (t(O))
which gives us a handle to study a large class of proxy voting rules
Example 5. Proxy majority pvmaj (6) is clearly a one man -- one vote rule ag-
gregator. It is easy to check that, for any proxy profile O: pvmaj(O) = maj(t(O)).
The same holds for proxy dictatorship (7). It is easy to see that proxy dictator-
ship pvd is such that for any proxy profile O: pvd(O) = d(t(O)), where d is the
standard dictatorship (of d ∈ N ).
It follows that for every proxy aggregator pvF = t ◦ F the axiomatic ma-
chinery developed for standard aggregators can be directly tapped into. Char-
acterization results then extend effortlessly to proxy voting, again providing a
strong rationale for the use of majority in proxy aggregation:
Fact 3 (Characterization of proxy majority). A one man -- one vote proxy ag-
gregator pv = t ◦ F for a given A is proxy majority pvmaj iff F is anonymous,
independent, monotonic, responsive, neutral and minimizes undecisiveness.
Proof. This follows from the definition of t and Theorem 1.
It follows that on simple agendas and assuming the individual rationality of
proxy profiles, proxy majority is the only rational aggregator which is anony-
mous, independent, monotonic, responsive, neutral and minimizes undecisive-
ness.
3.3.2
Impossibility
Similarly, there are many ways in which pursue the opposite embedding, from
standard aggregation into proxy voting. For example, we can define a func-
tion s : Oc → Pc from opinion profiles to individually rational proxy profiles
as follows. For a given opinion profile O, and issue p ∈ P consider the set
{i ∈ N Oi(p) = ∗} of individuals that abstain in O and take an enumeration
1, . . . , m of its elements, where m = {i ∈ N Oi(p) = ∗}. The function is de-
fined as follows, for any i ∈ N and p ∈ P: s(Oi(p)) = Oi(p) if Oi(p) ∈ {0, 1},
s(Oi(p)) = i + 1 mod m, otherwise.16 A translation of this type allows to think
c → O∗ as the concatenation s ◦ pv, for some
of standard aggregators F : O∗
proxy aggregator pv:
16Notice that since self-delegation (that is, Oi(p) = i) is not feasible in proxy opinions, this
definition of s works for profiles where, on each issue, either nobody abstains or at least two
individuals abstain. A dummy voter can be introduced for that purpose.
16
O
s
s(O)
Fpv
pv
pv(s(O))
Fact 4. Let A be such that its agenda is path connected and evenly negatable.
For any proxy aggregator pv, if s◦ pv is independent, unanimous and collectively
rational, then it is oligarchic.
Proof. It follows directly from the definition of s and Theorem 2.
3.4 Section Summary
The section has provided a very simple model of delegable proxy voting within
the framework of binary aggregation. This has allowed us to put liquid democ-
racy in perspective with an established body of results in the social choice
theory tradition, and highlight two of its problematic aspects, which have so far
gone unnoticed: the effect of cycles on collective indecisiveness, and the issue of
preservation of individual rationality under delegable proxies.
An independent, purpose-built axiomatic analysis for liquid democracy fo-
cused on its more characteristic features (like the one man -- one vote property)
is a natural line of research, which we do not pursue here.
4 Liquid Democracy as Binary Opinion Diffu-
sion
Proxy voting can also be studied from a different perspective. Imagine a group
where, for each issue p, each agent copies the 0, 1 opinion of a unique personal
"guru". Imagine that this group does so repeatedly until all agents (possibly)
reach a stable opinion. These new stable opinions can then be aggregated as
the 'true' opinions of the individuals in the group. The collective opinion of a
group of agents who either express a 0, 1 opinion or delegate to another agent
is (for one man -- one vote proxy aggregators) the same as the output obtained
from a vote where each individual has to express a 0, 1 opinion but gets there
by copying the opinion of some unique "guru" (possibly themselves). In this
perspective, a proxy voting aggregation can be assimilated to a (converging)
process of opinion formation.
The above interpretation of liquid democracy is explicitly put forth in [2]:
"While one way to describe delegations is the transfer of voting
weight to another person, you can alternatively think of delegations
as automated copying of the ballot of a trustee. While at assemblies
with voting by a show of hands it is naturally possible to copy the
17
vote of other people, in Liquid Democracy this becomes an intended
principle" [2, p. 22].
The current section develops an analysis of this interpretation, and highlights
some of its advantages over the delegation-based one studied earlier.
4.1 Binary aggregation and binary influence
The section develops a very simple model of binary influence based on the stan-
dard framework of binary aggregation (see Appendix B for a concise presenta-
tion). For simplicity, in this section we assume agents are therefore not allowed
to abstain, although this is not a crucial assumption for the development of our
analysis.
4.1.1 DeGroot Processes and Opinion Diffusion
In [9], DeGroot proposes a simple model of step-by-step opinion change under
social influence. The model combines two types of matrices. Assuming a group
of n agents, a first n× n matrix represents the weighted influence network (who
influences whom and how much), and a second n × m matrix represents the
probability assigned by each agent to each of the m different alternatives. Both
the agents' opinion and the influence weights are taken within [0, 1] and are
(row) stochastic (each row sums up to 1). Given an opinion and an influence
matrix, the opinion of each agent in the next time step is obtained through
linear averaging.
Here we focus on a specific class of opinion diffusion processes in which
opinions are binary, and agents are influenced by exactly one influencer, possibly
themselves, of which they copy the opinion. The model captures a class of
processes which lies at the interface of two classes of diffusion models that have
remained so far unrelated: the stochastic opinion diffusion model known as
DeGroot's [9], and the more recent propositional opinion diffusion model due to
[15]. The diffusion processes underpinning liquid democracy -- which we call here
Boolean DeGroot processes (BDPs) -- are the {0, 1} special case of the DeGroot
stochastic processes and, at the same time, the special case of propositional
opinion diffusion processes where each agent has access to the opinion of exactly
one neighbor (cf. Figure 1).
4.1.2 Boolean DeGroot processes
Here we focus on the Boolean special case of a DeGroot process showing its
relevance for the analysis of liquid democracy. Opinions are defined over a BA
structure, and hence are taken to be binary. Similarly, we take influence to
be modeled by the binary case of an influence matrix. Influence is of an "all-
or-nothing" type and each agent is therefore taken to be influenced by exactly
one agent, possibly itself. The graph induced by such a binary influence matrix
(called influence graph) is therefore a structure G = (cid:104)N, R(cid:105) where R ⊆ N 2 is
a binary relation where iRj is taken here to denote that "i is influenced by j".
18
Figure 1: BDPs lie in the intersection of DeGroot processes and propositional
opinion diffusion processes.
Such relation is serial (∀i ∈ N,∃j ∈ N : iRj) and functional (∀i, j, k ∈ N if iRj
and iRk then j = k). So each agent i has exactly one successor (the influencer),
possibly itself, which we denote R(i). It should be clear that influence graphs
are the same sort of structures we studied earlier in Section 3 under the label
'delegation graph'.
An influence profile G = (G1, . . . , Gm) records how each agent is influenced
by each other agent, with respect to each issue p ∈ P. Given a profile G the ith
projection Gi denotes the influence graph for issue pi, also written Gp.
So let us define the type of opinion dynamics driving BDPs:
Definition 4 (BDP). Now fix an opinion profile O and an influence profile G.
Consider the stream O0, O1, . . . , On, . . . of opinion profiles recursively defined
as follows:
• Base: O0 := O
• Step: for all i ∈ N , p ∈ P, On+1
i
(p) := On
Rp(i)(p).
where Gp = (cid:104)N, Rp(cid:105). We call processes defined by the above dynamics Boolean
DeGroot processes (BDPs).
It should be clear that the above dynamics is the extreme case of linear
averaging applied on binary opinions and binary influence.
As noted above, BDPs are also the special case of processes that have recently
been proposed in the multi-agent systems literature as propositional opinion
diffusion processes [15], i.e., cases where 1) the aggregation rule is the unanimity
rule (an agent adopts an opinion if and only if all her influencers agree on it),
and 2) each agent has exactly one influencer. We will come back to the link
with propositional opinion diffusion in some more detail later in Section 4.3.
4.2 Convergence of BDPs
When do the opinions of a group of individuals influencing each other stabi-
lize? Conditions have been given, in the literature, for the general paradigms
19
BDPsDeGroot ProcessesPropositional Opinion Diffusion of which BDPs are limit cases. This section introduces the necessary graph-
theoretic notions and briefly recalls those results before giving a characterization
of convergence for BDPs.
4.2.1 Preliminaries
We start with some terminology. We say that the stream of opinion profiles
O0, O1, . . . , On, . . . converges if there exists n ∈ N such that for all m ∈ N, if
m ≥ n, then Om = On.
We will also say that a stream of opinion profiles converges for issue p if there
exists n ∈ N such that, for all m ∈ N, if m ≥ n, then Om(p) = On(p). Given
a stream of opinion profiles starting at O we say that agent i ∈ N stabilizes
in that stream for issue p if there exists n ∈ N such that On
i (p) for
any m > n. So a BDP on influence graph G starting with the opinion profile
O is said to converge if the stream O0, O1, . . . , On, . . . generated according to
Definition 4 where O = O0 converges. Similarly, A BDP is said to converge
for issue p if its stream converges for p, and an agent i in the BDP is said to
stabilize for p if it stabilizes for p in the stream generated by the BDP.
i (p) = Om
Notice first of all that influence graphs have a special shape:17
Fact 5. Let G be an influence graph and C be a connected component of G.
Then C contains exactly one cycle, and the set of nodes in the cycle is closed.
Proof. Assume that C does not contain any cycle. Since N is finite and since no
path can repeat any node, any path in C is finite too. Let i be the last element
of (one of) the longest path(s) in C. Then i does not have any successor, which
contradicts seriality. So C contains at least one cycle. Let S be the set of nodes
of a cycle in C. Assume that S is not closed: for some i ∈ S and j /∈ S, iRj.
Since S is a cycle, there is also some k ∈ S, such that iRk, which contradicts
functionality. Therefore, the nodes of any cycle in C forms a closed set. Now
assume that C contains more than one cycle. Since the nodes of each cycle forms
a closed set, there is no path connecting any node inside a cycle to any node in
any other cycle, which contradicts connectedness. So C contains a unique cycle,
whose nodes form a closed set.
Intuitively, influence graphs of BDPs then look like sets of confluent chains
aiming together towards common cycles.
4.2.2 Context: convergence in DeGroot processes
For the general case of DeGroot processes, an influence structure guarantees that
any distribution of opinions will converge if and only if "every set of nodes that
is strongly connected and closed is aperiodic" [20, p.233]. In the propositional
opinion diffusion setting, sufficient conditions for stabilization have been given
by [15, Th. 2]: on influence structures containing cycles of size at most one
17Please consult Appendix B for the relevant terminology from graph theory.
20
(i.e, only self-loops), for agents using an aggregation function satisfying (ballot-
)monotonicity and unanimity18, opinions will always converge in at most at
most k + 1 steps, where k is the diameter of the graph.19 The results below
show how BDPs are an interesting limit case of both DeGroot and propositional
opinion diffusion processes.
4.2.3 Two results
It must be intuitively clear that non-convergence in a BDP is linked to the
existence of cycles in the influence graphs. However, from the above observation
(Fact 5), we know that nodes in a cycle cannot have any influencers outside this
cycle, and hence that cycles (including self-loops) can only occur at the "tail" of
the influence graph. Hence, if the opinions in the (unique) cycle do not converge,
which can only happen in a cycle of length ≥ 2, the opinions of the whole
population in the same connected component will not converge. The above
implies that for any influence graphs with a cycle of length ≥ 2, there exists a
distribution of opinions which loops. This brings us back to convergence result
for general (not necessarily Boolean) DeGroot processes. Indeed, for functional
and serial influence graphs, a closed connected component is aperiodic if and
only if its cycle is of length 1.
Fact 6. Let G be an influence profile. Then the following are equivalent:
1. The BDP converges for any opinion profile O on G.
2. For all p ∈ P, Gp contains no cycle of length ≥ 2.
3. For all p ∈ P, all closed connected components of Gp are aperiodic.
Proof. 2) ⇒ 1) Let p ∈ P and assume that Gp contains no cycle of length
≥ 2 and has diameter k. Let Cp be a connected component of Gp. By Fact 5,
Cp contains a unique cycle, which, by assumption, is of length 1. Hence, Cp
is aperiodic. Let i be the node in the cycle. The opinion of i will spread to
all nodes in Cp after at most k steps. Therefore, all BDPs on G will converge
after at most l steps, where l is the maximum within the set of diameters of
Gp for all p ∈ P.
1) ⇒ 3) We proceed by contraposition. Assume that for
some p ∈ P, a connected component Cp of Gp contains a cycle of length k ≥ 2.
By 5, this cycle is unique, and therefore the greatest common divisor of the
cycles lengths of Cp is k, so Cp is not aperiodic. Let S be the set of nodes in
the cycle. Let O be such that for some i, j ∈ S with distance d from i to j,
Oi(p) (cid:54)= Oj(p). Then Oi(p) will not converge, but enter a loop of size k: for
18Notice that the rule underpinning BDP, that is the 'guru-copying' rule on serial and
functional graphs, trivially satisfies those constraints.
19A second sufficient condition for convergence is given by [15]: when agents use the una-
nimity aggregation rule, on irreflexive graphs with only vertex-disjoint cycles, such that for
each cycle there exists an agent who has at least two influencers, opinions converge after at
most N steps. Note that no BDP satisfies this second condition.
21
all x ∈ N, Ox×k
Trivial.
i
(p) (cid:54)= O(x×k)+d
i
(p). Hence, O does not converge.
3) ⇒ 2)
It is worth noticing that one direction (namely from 3 to 1) of the above
result is actually a corollary of both the convergence result for DeGroot processes
stated at the beginning of this section (cf.
[20]), and of a known convergence
result for propositional opinion diffusion [15, Th. 2], also stated earlier.
The above gives a characterization of the class of influence profiles on which
all opinion streams converge. But we can aim at a more general result, charac-
terizing the class of pairs of opinion and influence profiles which lead to conver-
gence:
Theorem 3. Let G be an influence profile and O be an opinion profile. Then
the following statements are equivalent:
1. The BDP converges for O on G.
2. For all p ∈ P, there is no set of agents S ⊆ N such that: S is a cycle in
Gp and there are two agents i, j ∈ S such that Oi(p) (cid:54)= Oj(p).
Proof. 1) ⇒ 2) We proceed by contraposition. Let p ∈ P, S ⊆ N be a cycle
in Gp, i, j ∈ S, and Oi(p) (cid:54)= Oj(p). Let k be the length of the cycle and d be
the distance from i to j. Then Oi(p) will enter a loop of size k: for all x ∈ N,
i (p) (cid:54)= Oxk+d
2) ⇒ 1) Assume S ⊆ N be such that S is a cycle in
Oxk
Gp, and for all i, j ∈ S, Oi(p) = Oj(p). Then, for all j ∈ S, and all x ∈ N,
j (p) = Oi(p) and for all f ∈ N /∈ S with distance d from f to i, for all x ∈ N,
Ox
such that x ≥ d, Ox
(p).
i
f (p) = Oi(p).
This trivially implies that the class of opinion profiles which guarantees con-
vergence for any influence profile, is the one where everybody agrees on every-
thing already. Note that the only stable distributions of opinions are the ones
where, in each connected component in G, all members have the same opinion,
i.e, on BDPs, converging and reaching a consensus (within each connected com-
ponent) are equivalent, unlike in the stochastic case. Moreover, for an influence
profile where influence graphs have at most diameter d and the smallest cycle in
components with diameter d is of length c, it is easy to see that if a consensus
is reached, it will be reached in at most d − c steps, which is at most n − 1.
Finally observe that Theorem 3 subsumes Fact 6. If Gp contains only cycles
of length 1 (second statement in Fact 6) then, trivially, no two agents in a cycle
can disagree (second statement in Theorem 3).
4.2.4 Liquid Democracy as a BDP
We have seen (Section 3) that each proxy profile O induces what we called a
delegation graph GO = (cid:104)N, Rp(cid:105) for each issue p. Delegation graphs are the same
sort of structures we referred to in the current section as influence graphs. So
22
each proxy profile O can be associated to a BDP by simply assigning random
0 or 1 opinions to each voter delegating her vote in O. It is then easy to show
that for each connected component C of GO, if C has a guru with opinion x,
then that component stabilizes in the BDP on opinion x for each assignment
of opinions to the delegating agents in O. Vice versa, if C stabilizes on value
x in the BDP for each assignment of opinions to the delegating agents in O,
then C has a guru whose opinion is x. This establishes a direct correspondence
between voting with delegable proxy and Boolean deGroot processes. However,
BDPs offer an interesting and novel angle on the issue of cyclical delegations,
to which we turn now.
4.2.5 Cycles
As discussed earlier (Section 3.2.1), cycles are a much discussed issue in liquid
democracy.
Its proponents tend to dismiss delegation cycles as a non-issue:
since the agents forming a cycle delegate their votes, none of them is casting a
ballot and the cycles get resolved essentially by not counting the opinions of the
agents involved in the cycle [2]. We stressed this solution as problematic in the
'vote-delegation' interpretation of liquid democracy as it has the potential to
discard large numbers of opinions. The elimination of cycles not only hides to
aggregation the opinions of the agents involved in cycles, but also the opinions
of agents that may be linked to any of those agents by a delegation path. In
other words information about entire connected components in the delegation
graph may be lost.
We argue that the 'vote-copying' interpretation of the system -- formalized
through BDPs -- offers novel insights into possible approaches to cycles in del-
egable proxy. Theorems 3 and 6 offer an alternative solution by showing that
not all cycles are necessarily bad news for convergence: cycles in which all
agents agree still support convergence of opinions, and therefore a feasible ag-
gregation of opinions by proxy. This suggests that alternative proxy voting
mechanisms could be designed based on opinion convergence behavior rather
than on weighted voting.
4.3 Excursus: unanimity and 2-colorability
In the above, we have worked at the intersection of two models of opinion
diffusion, the DeGroot model, and the propositional opinion diffusion model.
However, there is more to say about how the two frameworks relate.
Let us take a brief detour towards a generalisation of BDPs corresponding to
the case of propositional opinion diffusion with the unanimity rule, where agents
can have several influencers and change their opinions only if all their influencers
disagree with them. This means that we relax the functionality constraint on
influence graphs. We will show how the two frameworks meet again: some non-
stabilizing opinion cases under the unanimity rule correspond to a special class
among the 'semi-Boolean' cases of DeGroot processes where opinions are still
binary but influence does not need to be.
23
We define the dynamics of opinions under the unanimity rule in the obvious
way:
Definition 5 (UP). Fix an opinion profile O and a (serial but non-necessarily
functional) influence profile G. Consider the stream O0, O1, . . . , On, . . . of opin-
ion profiles recursively defined as follows:
• Base: O0 := O
• Step: for all i ∈ N and all p ∈ P:
(cid:26) On
On+1
i
(p) =
if for some j, k ∈ Rp(i),On
i (p)
j (p) otherwise, where j ∈ Rp(i)
On
j (p) (cid:54)= On
k (p)
(9)
where Gp = (cid:104)N, Rp(cid:105). We call processes defined by the above dynamics Unanim-
ity Processes (UPs).
We give a sufficient condition for non-convergence of UPs:
Lemma 2. Let G be a (serial and non-necessarily functional) influence profile
and O be an opinion profile, such that, for some p ∈ P, for all i, j ∈ C, where
C is a connected component of Gp: if i ∈ Rp(j), then Oi(p) (cid:54)= Oj(p). Then O
does not converge in UP.
Proof. Let G be a (serial and non-necessarily functional) influence profile, and
O be an opinion profile, such that, for some p ∈ P, for all i, j ∈ C with C
if i ∈ Rp(j), then Oi(p) (cid:54)= Oj(p). Then, by
a connected component of Gp:
definition of UPs, for all i ∈ C, O1
i (p) (cid:54)= Oi(p), and by repeating the same
(p) (cid:54)= On
argument, for all n ∈ N, On+1
i (p).
i
Intuitively, the above condition for non-convergence corresponds to a situ-
ation of global maximal disagreement: all agents (of a connected component)
disagree with all their influencers. Recall that a graph is properly k-colored if
each node is assigned exactly one among k colors and no node has a successor of
the same color, and consider the two possible opinions on issue p as colors. The
above result can be reformulated in terms of proper 2 colorings, as follows: if for
some p ∈ P, O properly colors Gp, then O does not converge. In such a case, all
agents will change their opinion on p at every step, entering an oscillation of size
2. So the maximal state of disagreement is the maximally unstable case of the
dynamics. Note that this limit case of opinion distribution is yet another special
case of DeGroot processes, another example within the intersection between the
two frameworks of propositional opinion diffusion and DeGroot.
The possibility of such a distribution of opinions on p relies on the influence
graph Gp being 2-colorable, which is again a requirement about the lengths of
its cycles:
it is 2-colorable if and only if it contains no cycle of odd length.
However, non 2-colorability is not a sufficient condition for convergence of UPs
in general: a simple cycle of three agents, for instance, is not 2-colorable but
does not guarantee convergence either (as illustrated above with the convergence
24
conditions for BDPs). Nevertheless, there is a class of influence profiles for
which being 2-colorable is a necessary condition of non-convergence of UPs, the
symmetric ones:
Lemma 3. Let G be a symmetric (serial and non-necessarily functional) influ-
ence profile and O be an opinion profile. The following statements are equiva-
lent:
1. O converges in UP on G;
2. For all p ∈ P, for all connected component C of Gp, there are i, j ∈ C,
such that i ∈ Rp(j), and Oi(p) = Oj(p), where Gp = (cid:104)N, Rp(cid:105).
Proof. 2) ⇒ 1) Assume that for any p ∈ P, for any connected component C
of Gp, there exist i, j ∈ C, such that Rp(j) and Oi(p) = Oj(p). By definition of
UP, this implies that Oi(p) is stable, and that all agents with distance ≤ k will
be stable after at most k steps. 1) ⇒ 2) This follows from Lemma 2.
This means that opinions on a given p will converge if and only if two agents
influencing each other on p already agree on it. We can therefore, as we did
for BDPs, characterize the class of influence profiles for which all (symmetric)
opinion profiles converge in UPs:
Theorem 4. Let G be a symmetric (serial and non-necessarily functional)
influence profile. The following statements are equivalent:
1. All opinion profiles O, converge in UPs on G.
2. For all p ∈ P, and all connected components of C ⊆ Gp, C is not 2-
colorable (contains cycle(s) of odd length), where Gp = (cid:104)N, Rp(cid:105).
Proof. 2) ⇒ 1) Let p ∈ P and C be connected component of Gp with diameter
k. Let C contain a cycle of length c, with c odd. Let O be an arbitrary opinion
profile. Since c is odd, there exist i, j ∈ S such that j ∈ Rp(i) and Oi(p) = Oj(p).
By definition of UP, this implies that Oi(p) is stable, and that all agents with
distance ≤ k will be stable after at most k steps. Hence, O converges. 1) ⇒ 2)
This follows from Lemma 3.
Note that, while the basic modal language cannot capture graph 2-colorability,
it can capture non 2-colorability, and therefore capture the class of symmetric
(serial and non-necessarily functional) influence profiles which guarantee con-
vergence of UPs. We leave the detail out for space reason.
We have shown that, for UPs in general, convergence (in a connected com-
ponent) is not guaranteed if it contains no odd cycles, and that symmetric UPs
guarantee convergence as soon as they contain some odd cycle. However, con-
taining an odd cycle is a very "easy" requirement for a real-life influence network
to meet (it corresponds to a non-zero clustering coefficient). By contrast, re-
call that BDPs guarantee convergence (on any opinion profile) only when they
contain only cycles of size 1, which is a rather implausible requirement to be
satisfied on real influence networks.
25
4.4 BDPs on logically interdependent issues
So far we have assumed the aggregation to happen on a set of issues without
constraint (or rather with γ = (cid:62)). In this subsection we study what happens
in the presence of a constraint γ (cid:54)= (cid:62). BDPs on aggregation structures with
constraints may lead individuals to update with logically inconsistent opinions.
But the diffusion perspective whereby agents copy the opinion of trustees rather
than delegating their voting right better lends itself to an assumption of indi-
vidual rationality.
The following processes are simple adaptations of BDPs where agents update
their opinions only if the opinions of their influencers, on the respective issues,
are consistent with the constraint.20
Definition 6. Fix an opinion profile O, an influence profile G, and a constraint
γ. Consider the stream O0, O1, . . . , On, . . . of opinion profiles recursively defined
as follows:
• Base: O0 := O
• Step: for all i ∈ N , p ∈ P,
On+1
i
(p) :=
(cid:26) On
Rp(i)(p)
i (p)
On
if (cid:86)
p∈P On
Rp(i)(p) ∧ γ is consistent
otherwise
where Gp = (cid:104)N, Rp(cid:105). We call processes defined by the above dynamics individ-
ually rational BDPs.
Individually rational BDPs converge in some cases in which BDPs do not.
There are cases in which there is disagreement in the cycles but the process still
converges, because of the safeguard towards individual rationality built into the
dynamics.
Example 6. Consider the following example. Let N = {1, 2}, P = {p, q} and
γ = {p ↔ ¬q}. Let then G =(cid:10)N,{Ri}i∈P
(cid:11) be as follows: 1Rq1, 2Rq2, 1Rp2
and 2Rp1. Finally let O be such that O1(p) = O2(q) = 1, O2(p) = O1(q) = 0.
Voters 1 and 2 form a non-unanimous cycle, but O is a stable opinion profile.
The example shows that direction 1) ⇒ 2) of Theorem 3 does not hold for
individually rational BDPs: some individually rational BDPs may stabilize even
in the presence of disagreement within a cycle. Intuitively, the reason why this
happens is that individually rational BDPs that stabilize even when disagree-
ments occur within cycles do so because their cycles are not "synchronized".
In the above example, given the constraint p ↔ ¬q, the only way to get stabi-
lization starting from a situation respecting the constraint is to have a cycle of
influence for q which goes 'in the opposite direction' from the one from p, all
other cases would amount to violate the constraint.
20Other update policies are of course possible. A recent systematic investigation of opinion
diffusion on interconnected issues is [7].
26
Beyond this simple example, we want to find out what happens with more
complex constraints and what are the conditions for individually rational BDPs
to converge. Let us first show that direction 2) ⇒ 1) of Theorem 3 still holds,
that is, individually rational BDPs without disagreement in their cycles always
converge:
Theorem 5. Let G be an influence profile, O be an opinion profile, and γ a con-
straint. Then the following holds: if for all p ∈ P, for all S ⊆ N such that S is a
cycle in Gp, and all i, j ∈ S: Oi(p) = Oj(p), then the individually rational BDP
for O, G and γ converges in at most k steps, where k ≤ max{diam(Gp)p ∈ P}.
Proof. Assume that for all p ∈ P, for all S ⊆ N such that S is a cycle in
Gp, for all i, j ∈ S: Oi(p) = Oj(p). Consider an arbitrary i ∈ N . Let ki(p)
be the distance from i to the closest agent in a cycle of Gp, and let ki denote
max{ki(p)p ∈ P}. We show that for any ki ∈ N, Oki
is stable.
i
• If ki = 0: i is its only infuencer, therefore O0
• If ki = n + 1: Assume that for all agents j such that kj = n, Okj
is stable.
This implies that all influencers of i are stable. We need to consider the
following cases:
i is stable.
j
1. If(cid:86)
2. If(cid:86)
stable.
p∈P Om
p∈P Om
Rp(i)(p) ∧ γ is not consistent, then it will never be: On
Rp(i)(p) ∧ γ is consistent, then On+1
is stable.
i
is
i
This completes the proof.
4.5 Section Summary
In this section we studied a very simple class of opinion diffusion processes on
networks (Boolean DeGroot processes, BDPs), which precisely capture the vote-
copying behavior suggested by a standard interpretation of the liquid democracy
system. Interestingly these processes lie at the interface of two so far uncon-
nected network diffusion models: the well-known DeGroot processes -- of which
BDPs constitute the binary special case -- and of propositional opinion diffusion
processes -- of which BDPs constitute the special case where the set of neighbors
is a singleton. We established necessary and sufficient conditions for conver-
gence, which can be captured in modal fixpoint logics as we will show in the
next section. We argued that these results provide a novel angle on the issue of
delegation cycles in liquid democracy.
There are a number of further questions concerning, especially, individually
rational BDPs that we leave for future investigations: What are the necessary
conditions for their stabilization? What opinions are reachable? And, in par-
ticular, when is a consensus reached? Finally, one could consider other types of
influence policies than the one used in individually rational BDPs. For instance,
agents may be allowed to 'pass through' an inconsistent state at some point, in
27
which case one can wonder under which conditions the process can still converge
to a consistent state. Indeterministic policies would also make sense, where an
agent confronted with inconsistent opinions from her influencers keeps one of
the closest consistent opinions set, rather than not being influenced at all (cf.
[7]).
5 Fixpoint Logics for BDPs
In this section we show how a well-established logic for formal verification can be
readily used to specify and reason about properties of BDPs, and in particular
their convergence. The logic is the so-called µ-calculus. This points to a so-far
unexplored interface between fixpoint logics and models of opinion dynamics --
like the DeGroot model and propositional opinion diffusion. The section moves
some first steps in that direction along the lines of another recent work [4],
where the mu-calculus, and extensions thereof, have been applied to the study
of dynamical systems.
5.1
Influence graphs as Kripke models
We treat influence graphs as Kripke (multi-relational) models [22, 8].
Definition 7. We call an influence model a tuple M = (cid:104)N, G, O(cid:105) where G =
(Gp1, . . . , Gpm ) is an influence profile, and O : P −→ 2N is an opinion profile
over P, that is, a valuation function.
One can therefore easily interpret a modal language over influence models,
where modalities are interpreted on the accessibility relations in G. That is,
to each graph Gp we associate modalities [p] and (cid:104)p(cid:105). We will give the details
below, but let us immediately note that the class of (possibly infinite) influence
graphs would then be characterized by the following properties, for any p ∈ P:
[p] ϕ → (cid:104)p(cid:105) ϕ
(cid:104)p(cid:105) ϕ → [p] ϕ
(seriality)
(10)
(11)
More precisely, for any influence profile G = (Gp1, . . . , Gpm ), formula [pi] ϕ →
(cid:104)pi(cid:105) ϕ (respectively, (cid:104)pi(cid:105) ϕ → [pi] ϕ) is valid in such graph -- that is, true in any
pointed influence model built on such graph -- if and only if each Gpi consists
of a serial (respectively, functional) relation.21 Put otherwise, on serial and
functional graphs the modal box and diamond are equivalent.
(functionality)
5.2 Modal µ-calculus
Lµ : ϕ ::= p ⊥ ¬ϕ ϕ ∧ ϕ (cid:104)p(cid:105) ϕ µp.ϕ(p)
The language of the µ-calculus expands the basic modal language with a
least fixpoint operator µ. Here is the BNF of the language:
21These are known results from modal correspondence theory (cf. [5]).
28
where p ranges over P and ϕ(p) indicates that p occurs free in ϕ (i.e., it is
not bounded by fixpoint operators) and under an even number of negations.22
In general, the notation ϕ(ψ) stands for ψ occurs in ϕ. The usual definitions for
Boolean and modal operators apply. Intuitively, µp.ϕ(p) denotes the smallest
formula p such that p ↔ ϕ(p). The greatest fixpoint operator ν can be defined
from µ as follows: νp.ϕ(p) := ¬µp.¬ϕ(¬p).
We interpret Lµ on influence models as follows:
Definition 8. Let ϕ ∈ Lµ. The satisfaction of ϕ by a pointed influence model
(M, i) is inductively defined as follows:
M, i (cid:54)= ⊥
M, i = p ⇐⇒ i ∈ O(p), for p ∈ P
M, i = ¬ϕ ⇐⇒ i (cid:54)∈ (cid:107)ϕ(cid:107)M
M, i = ϕ1 ∧ ϕ2 ⇐⇒ i ∈ (cid:107)ϕ1(cid:107)M ∩ (cid:107)ϕ2(cid:107)M
M, i = (cid:104)p(cid:105) ϕ ⇐⇒ i ∈ {j ∃k : jGpk & k ∈ (cid:107)ϕ(cid:107)M}
M, i = µp.ϕ(p) ⇐⇒ i ∈(cid:92){X ∈ 2N (cid:107)ϕ(cid:107)M[p:=X] ⊆ X}
where (cid:107)ϕ(cid:107)M[p:=X] denotes the truth-set of ϕ once O(p) is set to be X. As usual,
we say that: ϕ is valid in a model M iff it is satisfied in all points of M, i.e.,
M = ϕ; ϕ is valid in a class of models iff it is valid in all the models in the
class.
We list some relevant known results about Kµ. The logic has a sound and
(weakly) complete axiom system [28]. The satisfiability problem of Kµ is decid-
able [26]. The complexity of the model-checking problem for Kµ is known to be
in NP ∩ co-NP [19]. It is known that the model-checking problem for a formula
of size m and alternation depth d on a system of size n can be solved by the nat-
ural fixpoint-approximation algorithm with (time) complexity of O((m · n)d+1)
[12], where the alternation depth of a formula of Lµ is the maximum number
of µ/ν alternations in a chain of nested fixpoint subformulas.23 Finally, the µ-
calculus is known to be invariant for bisimulation (cf. [5]). It is actually known
to correspond to the bisimulation-invariant fragment of monadic second-order
logic [21].
5.3 On the logic of convergence in BDPs
Each stream of opinion profiles O0, O1, . . . , On, . . . corresponds to a stream of
influence models M0,M1, . . . ,Mn, . . ..
From the point of view of an influence model M = (cid:104)N, G, O(cid:105) the BDP
dynamics of Definition 4 can therefore be recast in terms of updates of the
valuation function O as follows:
22This syntactic restriction guarantees that every formula ϕ(p) defines a set transformation
which preserves ⊆, which in turn guarantees the existence of least and greatest fixpoints by
the Knaster-Tarski fixpoint theorem (cf. [25]).
23The reader is referred to, e.g. [13], for the precise definition.
29
• Base: O0 := O
• Step: On+1(p) := (cid:107)[p] p(cid:107)Mn .
That is, the interpretation of p at step n + 1 is the interpretation of [p] p at step
n. Equivalently, the interpretation of ¬p at step n + 1 is the interpretation of
[p]¬p at step n.
Lemma 4. Let M = (cid:104)N, G, O(cid:105) be an influence model. The two following
statements are equivalent:
1. i ∈ N is stable for p;
2. The pointed model (M, i) satisfies:24
stb(p) := νx. ± p ∧ [p] x
(12)
Proof. First of all observe that, by the semantics of the µ-calculus, formula
stb(p) denotes the largest fixpoint of function ±p∧[p] (·), that is, formula [p∗]±p
where [p∗] is the modal box interpreted over the reflexive and transitive closure of
Gp. 1) ⇒ 2) Assume that i is stable for p and suppose towards a contradiction
that M, i (cid:54)= stb(p). By what said above, it follows that there exists a j such that
Oi(p) (cid:54)= Oj(p) which is connected by a finite Gp path to i. By the functionality
of influence models and the dynamics of Definition 4 then at some stage n in
the stream of opinion profiles it should hold that On
i (p) = Oj(p), against the
assumption that i be stable for p. 2) ⇒ 1) Assume M, i = stb(p). By what
said above, this implies that there exists no j such that Oi(p) (cid:54)= Oj(p) which is
connected by a finite Gp path to i. It follows that in the stream generated by
the BDP dynamics i cannot change its opinion, and hence it is stable.
Theorem 6. Let M = (cid:104)N, G, O(cid:105) be an influence model. The two following
statements are equivalent:
1. i ∈ N stabilizes for issue p ∈ P;
2. The pointed model (M, i) satisfies:
µx.stb(p) ∨ [p] x
such fixpoint as (cid:83)
(13)
Proof. First of all observe that, by the semantics of the µ-calculus µx.stb(p) ∨
[p] x denotes the smallest fixpoint of equation x ↔ stb(p)∨[p] x. By the Knaster-
Tarski theorem and the fact that influence models are finite, we can compute
0≤n<ω(cid:107)stb(p)n(cid:107) where (cid:107)stb(p)0(cid:107) = (cid:107)stb(p) ∨ [p]⊥(cid:107) (notice
that [p]⊥ ↔ ⊥ on influence models) and (cid:107)stb(p)n+1(cid:107) = (cid:107)stb(p) ∨ [p] stb(p)n(cid:107).
So, by Lemma 4 i belongs to (cid:107)µx.stb(p)∨[p] x(cid:107) either i is stable for issue p or has
24Notice that ±p is used as a variable ranging over {p, ¬p}. Technically the above formula
is to be read as a scheme for νx.p ∧ [p] x and νx. ± ¬p ∧ [p] x.
30
access in a finite number of steps to a an agent who is stable for p. 1) ⇒ 2)
Assume that i stabilizes for issue p ∈ P. So there exists a stage n in the
stream of profiles generated through Definition 4 at which On
i (p) for
all m > n. By Lemma 4, (cid:104)N, G, On(cid:105) , i = stb(p). It follows that i is connected
through a finite Gp-path to an agent j such that M, j = stb(p). By what
established above we thus have that M, i = µx.stb(p)∨ [p] x. 2) ⇒ 1) Assume
M, i = µx.stb(p) ∨ [p] x. It follows that i is connected through a finite Gp-path
to an agent j such that M, j = stb(p). By Lemma 4 j is therefore stable and
therefore i will stabilize for p.
i (p) = Om
So the formula that expresses the stabilization of the agents' opinions on
one issue is µx. (νy. ± p ∧ [p] y) ∨ [p] x. Informally, the theorem states that in a
BDP an agent reaches a stable opinion if and only if it has an indirect influencer
(linked by an influence path) whose all direct and indirect influencer have the
same opinion. Notice that such formula has alternation depth 0. So an off-the-
shelf model-checking algorithm for the µ-calculus can check stabilization in time
O(m · n) with n being the size of the model and m the size of the formula.
Now confront this with the earlier Theorem 6. Since the convergence of the
BDP is equivalent to the stabilization of all agents on all issues p (either on p
or ¬p), we have the following corollary:
Corollary 1. The BDP for an opinion profile O based on influence graph G
converges if and only if
(cid:104)N, G, O(cid:105) , i = U
µx.stb(p) ∨ [p] x
(14)
(cid:94)
p∈P
for any agent i ∈ N , where U denotes the universal modality (cf. [5]).
So the above formula characterizes the property of convergence for a BDP.
Since the process of voting in a liquid democracy system can be modeled by a
BDP, the formula also characterizes precisely when voting by delegable proxy
results in a 1 or 0 opinion on a given issue.
6 Conclusions
The paper has moved the first steps towards the development of theoretical
foundations for the voting system of liquid democracy based on delegable proxy.
We have pursued two lines of research linked to two interpretations com-
monly associated to the proxy character of liquid democracy: the delegation of
voting right to trustees, vs. the copying of the votes of influencers. The first
interpretation has led us to develop a simple model of liquid democracy based
on the theory of binary and judgment aggregation. This has allowed us to study
liquid democracy as a form of binary aggregation with abstentions. The second
interpretation has led us to study liquid democracy through extremely simple
31
models of opinion diffusion corresponding to the Boolean special case of the
stochastic processes of opinion diffusion known as DeGroot processes. We have
argued that studying aggregation in liquid democracy through this lens offers
important advantages with respect to the handling of delegation cycles and the
preservation of individual rationality. Through this second perspective we have
also shown how off-the-shelf logical techniques can be used to analyze properties
(such as convergence) of the diffusion process underpinning liquid democracy.
References
[1] Dan Alger. Voting by proxy. Public Choice, 126(1-2):1 -- 26, jan 2006.
[2] J. Behrens, A. Kistner, A. Nitsche, and B. Swierczek. Principles of Liquid
Feedback. 2014.
[3] Jan Behrens. Cyclical delegations: Myths or disaster. Liquid Democracy
Journal, 3, 2015.
[4] Johan van Benthem. Oscillations, Logic, and Dynamical Systems. In Su-
jata Ghosh and Jakub Szymanik, editors, The Facts Matter. Essays on
Logic and Cognition in Honour of Rineke Verbrugge, pages 9 -- 22. College
Publications, 2015.
[5] Patrick Blackburn, Maarten de Rijke, and Yde Venema. Modal Logic.
Cambridge University Press (CUP), 2001.
[6] Paolo Boldi, Francesco Bonchi, Carlos Castillo, and Sebastiano Vigna.
Viscous democracy for social networks. Communications of the ACM,
54(6):129, jun 2011.
[7] Sirin Botan. Propositional opinion diffusion with constraints. ILLC Master
of Logic Thesis, 2016.
[8] Zo´e Christoff and Jens Ulrik Hansen. A logic for diffusion in social networks.
Journal of Applied Logic, 13(1):48 -- 77, mar 2015.
[9] Morris H. Degroot. Reaching a Consensus. Journal of the American Sta-
tistical Association, 69(345):118 -- 121, mar 1974.
[10] Franz Dietrich and Christian List. Judgment aggregation without full ra-
tionality. Social Choice and Welfare, 31(1):15 -- 39, jul 2007.
[11] Elad Dokow and Ron Holzman. Aggregation of binary evaluations with
abstentions. Journal of Economic Theory, 145(2):544 -- 561, mar 2010.
[12] E. Allen Emerson. Model Checking and the Mu-calculus. In Descriptive
Complexity and Finite Models, Proceedings of a DIMACS, 1996.
32
[13] E. Allen Emerson, Charanjit S. Jutla, and A. Prasad Sistla. On model
checking for the mu-calculus and its fragments. Theoretical Computer Sci-
ence, 258(1-2):491 -- 522, may 2001.
[14] Umberto Grandi and Ulle Endriss. Lifting integrity constraints in binary
aggregation. Artif. Intell., 199:45 -- 66, 2013.
[15] Umberto Grandi, Emiliano Lorini, and Laurent Perrussel. Propositional
Opinion Diffusion. In Proceedings of the 2015 International Conference on
Autonomous Agents and Multiagent Systems, AAMAS '15, pages 989 -- 997,
Richland, SC, 2015. International Foundation for Autonomous Agents and
Multiagent Systems.
[16] James Green-Armytage. Direct democracy by delegable proxy. 2005.
[17] James Green-Armytage. Direct voting and proxy voting. Const Polit Econ,
26(2):190 -- 220, dec 2014.
[18] Davide Grossi and Gabriella Pigozzi. Judgment Aggregation: A Primer.
Synthesis Lectures on Artificial Intelligence and Machine Learning, 8(2):1 --
151, mar 2014.
[19] Erich Gradel and Martin Otto. On logics with two variables. Theoretical
Computer Science, 224(1-2):73 -- 113, aug 1999.
[20] Matthew O. Jackson. Social and Economic Networks. Princeton University
Press, Princeton, NJ, USA, 2008.
[21] David Janin and Igor Walukiewicz. On the expressive completeness of the
propositional mu-calculus with respect to monadic second order logic. In
CONCUR '96: Concurrency Theory, pages 263 -- 277. Springer, 1996.
[22] Fenrong Liu, Jeremy Seligman, and Patrick Girard. Logical dynamics of
belief change in the community. Synthese, 191(11):2403 -- 2431, 2014.
[23] Kenneth O. May. A Set of Independent Necessary and Sufficient Conditions
for Simple Majority Decision. Econometrica, 20(4):680, oct 1952.
[24] James C. Miller. A program for direct and proxy voting in the legislative
process. Public Choice, 7-7(1):107 -- 113, sep 1969.
[25] Colin Stirling. Modal Mu-Calculus. In Texts in Computer Science, pages
103 -- 132. Springer, 2001.
[26] Robert S. Streett and E. Allen Emerson. The propositional mu-calculus
is elementary. In Automata Languages and Programming, pages 465 -- 472.
Springer, 1984.
[27] Gordon Tullock. Computerizing politics. Mathematical and Computer Mod-
elling, 16(8-9):59 -- 65, aug 1992.
33
[28] Igor Walukiewicz. Completeness of Kozen's Axiomatisation of the Propo-
sitional mu-Calculus. Information and Computation, 157(1-2):142 -- 182, feb
2000.
A Binary aggregation (without abstention)
The formalism of choice for this paper is binary aggregation [14]. A binary
aggregation structure (BA structure) is a tuple A = (cid:104)N, P, γ(cid:105) where:
• N = {1, . . . , n} is a finite set individuals s.t. N = n ∈ N;
• P = {p1, . . . , pm} is a finite set of issues (P = m ∈ N), each represented
by a propositional atom;
• γ ∈ L is an (integrity) constraint, where L is the propositional language
constructed by closing P under a functionally complete set of Boolean
connectives (e.g., {¬,∧})
An opinion O : P → {0, 1} is an assignment of truth values to the set
of issues P, and the set of all opinions is denoted by O. The opinion of an
agent i is said to be "consistent" whenever Oi = γ, that is, i's opinion sat-
isfies the integrity constraint. The set of all consistent opinions is denoted
Oc = {O ∈ O O = γ}. Thus, O(p) = 0 (respectively, O(p) = 1) indicates that
opinion O rejects (respectively, accepts) the issue p. Syntactically, the two opin-
ions correspond to the truth of the literals p or ¬p. For p ∈ P we write ±p to
denote one element from {p,¬p}. An opinion profile O = (O1, . . . , On) records
the opinion, on the given set of issues, of every individual in N . Given a profile
O the ith projection O is denoted Oi (i.e., the opinion of agent i in profile O).
We also denote by O(p) = {i ∈ N Oi(p) = 1} the set of agents accepting issue
p in profile O and by O(p−) = {i ∈ N Oi(p) = 0}.
Given a BA structure A, an aggregation rule (or aggregator) for A is a
function F : (Oc)N → O, mapping every profile of consistent opinions to one
collective opinion in O. F (O)(p) denotes the outcome of the aggregation on
issue p. A benchmark aggregator is issue-by-issue strict majority rule (maj),
which accepts an issue if and only if the majority of the population accepts it:
maj(O)(p) = 1 ⇐⇒ O(p) ≥ N + 1
2
.
(15)
It is well-known that aggregation by majority does not preserve consistency.
The standard example is provided by the discursive dilemma, represented by
the BA structure (cid:104){1, 2, 3} ,{p, q, r} , r ↔ p ∧ q(cid:105). The profile consisting of O1 =
p ∧ q ∧ r, O2 = p ∧ ¬q ∧ ¬r, O3 = ¬p ∧ q ∧ ¬r, returns an inconsistent majority
opinion maj(O) = p ∧ q ∧ ¬r.
34
B Relevant terminology from graph theory
Let G = (cid:104)N, R(cid:105) be a graph and R∗ be the transitive and symmetric closure of
R. A path is a sequence of nodes (cid:104)i1, . . . , ik(cid:105), such that, for all l ∈ {1, . . . , k},
ilRil+1. The distance between two nodes i, j is the length of the shortest path
(cid:104)i, . . . , j(cid:105) between them. The diameter of a graph is the maximal distance
between any two nodes related by a path. A cycle is a path of length k such
that i1 = ik. A set of nodes S ⊆ N is said to be:
a cycle in G if all elements in S are in one cycle of length S,
connected if for any i, j ∈ S: iR∗j,
strongly connected if for any i, j ∈ S: there is a path (cid:104)i, . . . , j(cid:105),
closed if for any i ∈ S, j /∈ S, it is not the case that iRj,
a connected component if for any i, j ∈ N : iR∗j if and only if i, j ∈ S,
aperiodic if the greatest common divisor of the lengths of its cycles is 1.
35
|
1711.00728 | 1 | 1711 | 2017-10-31T11:18:06 | Crossing Behaviour of Social Groups: Insights from Observations at Non-signalized Intersection | [
"cs.MA",
"physics.soc-ph"
] | Environmental, demographical and psychological factors have a demonstrated impact on risky crossing behaviour. In this work we focus on the potential influence of social factors on the considered phenomenon (i.e. group crossing decision). We present the results of a video-recorded observation about the crossing behaviour of singles and dyads at non-signalized intersections. Results showed that crossing behaviour is characterized by three distinct phases: (i) approaching, (ii) appraising (decision making) and (iii) crossing. Dyads walk slower than single pedestrians in all phases. The crossing behaviour of dyads is characterized by the emergence of a leader who takes the decision to cross first, followed by the companion. However, there is no difference between the accepted safety gap of singles and dyads. Understanding factors influencing the crossing decision of social groups represents an important result supporting the development of agent-based simulations of pedestrian-vehicle interactions. | cs.MA | cs |
Crossing Behaviour of Social Groups:
Insights from Observations at Non-signalized
Intersection
Andrea Gorrini, Luca Crociani, Giuseppe Vizzari, Stefania Bandini
Abstract Environmental, demographical and psychological factors have a demon-
strated impact on risky crossing behaviour. In this work we focus on the poten-
tial influence of social factors on the considered phenomenon (i.e. group crossing
decision). We present the results of a video-recorded observation about the cross-
ing behaviour of singles and dyads at non-signalized intersections. Results showed
that crossing behaviour is characterized by three distinct phases: (i) approaching,
(ii) appraising (decision making) and (iii) crossing. Dyads walk slower than single
pedestrians in all phases. The crossing behaviour of dyads is characterized by the
emergence of a leader who takes the decision to cross first, followed by the compan-
ion. However, there is no difference between the accepted safety gap of singles and
dyads. Understanding factors influencing the crossing decision of social groups rep-
resents an important result supporting the development of agent-based simulations
of pedestrian-vehicle interactions.
1 Introduction
As highlighted by the WHO [16], road accidents represent the eighth leading cause
of death in the world population: 1.2 million people are killed on roads every year.
Despite recent efforts, the measures currently in place to reduce the phenomenon
are mainly aimed at protecting car occupants. However, pedestrians are some of the
most vulnerable road users, with a percentage of fatalities corresponding to 22% of
the overall victims (26% in EU, 14% in USA).
Andrea Gorrini ( ) · Luca Crociani · Giuseppe Vizzari · Stefania Bandini
CSAI research center, University of Milano-Bicocca, Viale Sarca 336 - U14, 20126 Milan (Italy).
e-mail: {name.surname}@disco.unimib.it
Stefania Bandini
RCAST, The University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo 153-8904 (Japan).
1
2
Gorrini, A., Crociani, L., Vizzari, G., Bandini, S.
To effectively contrast the social costs of pedestrian-car accidents it is necessary
to identify the determining factors of risky crossing decisions by means of a multi-
disciplinary approach (e.g., traffic psychology, transportation engineering, safety
science, computer science). This is aimed at supporting public institutions in the
design of effective and safe infrastructures and traffic management solutions.
In this context, the results presented in [15, 12] highlighted for instance the
impact of the physical features of the infrastructure on crossing behaviour (e.g.,
road width, traffic volumes, type of intersection). Other researches focused on the
complex coordination of the motor and cognitive abilities involved in crossing be-
haviour, such as: locomotion [15]; perception and attention [8]; attitude towards
hazardous situations [3]. Other studies [2, 7] showed the impact of ageing on cross-
ing behaviour, due to the progressive decline of these functions.
These works showed the relevant impact of environmental, demographical and
psychological factors on risky crossing behaviour. However, there is still a lack of
knowledge about the impact of social factors on the phenomenon, due to limited or
controversial results. Urban cross-walks are characterized indeed by the presence
of multiple pedestrians, and people often cross the street in the presence of other
familiar or unfamiliar pedestrians.
The contributions about this topic can be classified according two different ap-
proaches. On one hand, the results presented in [9, 13] showed that the presence
of other pedestrians is associated with a reduction of cautiousness in crossing de-
cision at signalized cross-walks. Moreover, the results presented in [4] showed that
pedestrians imitate the behaviour of others to judge the gap from oncoming vehicles,
with potentially dangerous crossing decisions due to an overestimation of crossing
possibilities. This phenomenon is based on the power of group to cause a diffusion
of responsibility [6], as each member feels that the responsibility for violating the
norms and for taking risky crossing decisions is shared with the rest of the group.
On the other hand, the results collected by [14, 1] showed that people are more
lax in risky crossing decisions standing by their own than when standing in groups.
These findings can be explained considering the theory of social control [11], which
describes the mechanism behind obedient behaviour as the motivation to be re-
warded just for being conformist to the group norms.
Starting from these assumptions, the present work was focused on studying the
crossing behaviour of social groups (groups composed of familiar pedestrians, such
as relatives and friends) through a video-recorded observation at a non-signalized in-
tersection. The research was aimed at comparing data about the crossing behaviour
of single pedestrians and two-members groups (i.e. dyads): the simplest and most
frequently observed type of group.
The methodology which sets the current work is presented in Section 2, with
reference to data collection and data analysis. The results of the observation (e.g.,
traffic volumes, Level of Service, crossing phases, speeds, safety gap) are presented
in Section 3. The paper concludes with remarks about the achieved results and their
future use for the further development of a computer-based simulator of pedes-
trian/vehicles interaction in urban contexts [5].
Crossing Behaviour of Social Groups
3
Fig. 1 A screen shot from the tracker tool used for data analysis.
Locomotion behaviour
Verbal behaviour
Non Verbal behaviour
- high spatial cohesion and coordination while walking
- waiting dynamics to regroup in case of separation
- leader/follower dynamics in sudden changes of direction
- talking while walking
- physical contact
- body and gaze orientation to the each other
- gesticulation while talking and/or indicating
Table 1 The check-list used by the coders for identifying social groups.
2 Data Collection
The video-recorded observation [7] was performed in 2015 (from 10:45 am, 73 min-
utes in total) at a non-signalized intersection in Milan (Italy), which was selected by
means of a preliminary analysis related to the localisation of road traffic accidents.
Results showed that the area was characterised by a high number of pedestrian/car
accidents in the past years.
A first phase of manual counting activity allowed to estimate and characterise the
observed traffic volumes. An ad hoc designed check-list (see Tab. 1), comprising a
set of locomotion and communication indicators, was used to support the annotators
in the identification of social groups from the video-images. Then, a time stamping
activity allowed to measure the Level of Service in the observed intersection.
A second phase of data analysis was based on the use of the open source software
Tracker Video Analysis and Modelling Tool (see Fig. 1). The tracker allowed to
correct the distortion of the video images (due to due to the wide lens and the nearly
4
Gorrini, A., Crociani, L., Vizzari, G., Bandini, S.
zenith perspective of the camera), and to track a sample of pedestrians considering
one frame every ten (every 0.4 sec). The data set (including the X, Y coordinates and
the associated frames t) was exported for data analysis.
3 Results
3.1 Traffic Volumes
The observed vehicular traffic volumes (1379 veh, 18.89 veh/min) were constituted
for the majority by cars (67%). The direction of movements of vehicles was equally
distributed. The majority of pedestrians (585 ped, 8.01 ped/min) were singles (65%).
Dyads were a significant portion of the total counted pedestrians: 26% dyads; 8%
triples; 1% other. A Cohen's Kappa analysis showed a moderate agreement between
the two independent coders in the identification of social groups (K = 0.47).
3.2 Level of Service
The Level of Service (LOS) [10] describe the degree of efficiency of an intersection,
by measuring the additional travel time experienced by drivers and pedestrians as
they travel/walk through a road segment. At non-signalized intersections LOS E is
the minimum acceptable design standard.
The LOS have been estimated by time stamping the delay of drivers due to ve-
hicular traffic and crossing pedestrian flows (time for deceleration, queue, stopped
delay, acceleration), and the delay of pedestrians due to drivers' non compliance to
their right of way on zebra (waiting, start-up delay). Results showed that the average
delay of drivers (3.20 s/veh ± 2.73 sd) and pedestrians (1.29 s/ped ± .21 sd) corre-
sponded to LOS A: nearly all drivers found freedom of operation; low risk-taking
in crossing behaviour.
3.3 Speeds and Crossing Phases
A sample of 49 adult pedestrians (27 singles and 11 dyads, from about 18 y.o. until
65 y.o.) was selected for video-tracking analysis. The sample was selected avoiding
situations such as: platooning of vehicles on the roadway, the joining of pedestrians
already crossing, and in general situations influencing the direct interaction between
pedestrians and drivers. Part of the selected crossing episodes was characterised by
the multiple interaction between the crossing pedestrian and two cars oncoming
Crossing Behaviour of Social Groups
5
Fig. 2 An exemplification of the trend analysis performed on the time series of dyads' speeds.
from the near and the far lane. Pedestrians' gender was balanced, but not considered
for data analysis.
The speeds pedestrians have been analysed among the time series of video
frames. The trend of speeds was analysed by calculating the difference between the
moving average (MA, time period length: 0.8 s) and the cumulative average (CA) of
the entire frames series. This allowed to smooth out short-term fluctuations of data
and to highlight longer-term deceleration/acceleration trends. According to results,
crossing behaviour is defined as composed of three distinctive phases:
1. approaching phase: the pedestrian travels on the side-walk with a stable speed
(Speed MA - CA (cid:39) 0);
2. appraising phase: the pedestrian approaching the cross-walk decelerates to eval-
uate the distance and speed of oncoming vehicles (safety gap). We decided to
consider that this phase starts with the first value of a long-term deceleration
trend (Speed MA - CA < 0);
3. crossing phase: the pedestrian decides to cross and speed up. The crossing phase
starts from the frame after the one with the lowest value of speed before a long-
term acceleration trend (Speed Min).
A two-way ANOVA1 showed a significant difference among the speeds of pedes-
trians while approaching, appraising and crossing [F(2,146) = 66.981, p < 0.001],
and a significant difference between the speeds of singles and dyads [F(1,146) =
40.652, p < 0.001] (see Tab. 2 and Fig. 3). A series of post hoc Tukey test showed a
non significant difference between the speeds of all pedestrians while approaching
and crossing (p > 0.05). The difference between the speeds of singles and dyads
was significant among all the crossing phases (p < 0.001).
Results showed that the crossing behaviour of dyads (see Figure 2) is charac-
terised by the emergence of a leader who completes the appraising first and takes
the decision to cross, followed by the companion. An independent samples t-test
analysis showed that the difference between the speeds of group leaders and follow-
ers was not significant among all the three crossing phases (p > 0.05).
1 All statistics have been executed by using IBM SPSS Statistics 24, at the p < 0.01 level.
6
Gorrini, A., Crociani, L., Vizzari, G., Bandini, S.
(a) Singles
(b) Dyads
Fig. 3 The speed of single pedestrians and dyads among the crossing phases.
Table 2 The speed of singles and dyads among the crossing phases.
Crossing Phases
Approaching speed
Appraising speed
Crossing speed
Speed of Singles
1.277 m/s ± 0.182 sd
0.943 m/s ± 0.214 sd
1.354 m/s ± 0.181 sd
Speed of Dyads
1.123 m/s ± 0.155 sd
0.702 m/s ± 0.280 sd
1.127 m/s ± 0.131 sd
Results demonstrated that the crossing behaviour of both singles and dyads is
based on a significant deceleration in proximity of the curb to evaluate the dis-
tance and speed of oncoming vehicles to safely cross. Dyads walked in average
17% slower than singles among the three crossing phases, decelerating the 11%
more than singles while appraising.
3.4 Accepted Safety Gap
The term safety gap denotes the ratio between the pedestrians's evaluation of the
distance of an approaching vehicle and its speed (not taking into account accelera-
tion/deceleration trends) to decide to safely cross avoiding collision.
A one-way ANOVA showed that the safety gap accepted by singles (3.982 s ±
2.549 sd), group leaders (4.355 s ± 2.491 sd) and followers (4.242 s ± 2.585 sd) was
not significantly different (p > 0.05) (see Table 3). Moreover, a one-way analysis of
variance showed that the time duration of the appraising phase of singles (2.785 s ±
1.474 sd), group leaders (3.527 s ± 2.297 sd) and followers (3.164 s ± 1.682) were
not significantly different (p > 0.05).
Crossing Behaviour of Social Groups
7
Table 3 The safety gap accepted by singles, group leaders and followers.
Singles
3.982 s ± 2.549 sd
Safety Gap
Vehicle Distance 16.399 m ± 7.753 sd
Vehicle Speed
Appraising Time 2.785 s ± 1.474 sd
Group Leaders
4.355 s ± 2.491 sd
19.152 m ± 9.967 sd
Followers
4.242 s ± 2.585 sd
17.325 m ± 9.583 sd
16.534 km/h ± 6.737 sd 16.525 km/h ± 5.662 sd 15.695 km/h ± 4.479 sd
3.527 s ± 2.297 sd
3.164 s ± 1.682
Results showed also that, compared to leaders, followers started appraising with
a delay of 1.055 s (± 1.032 sd) and they decide to cross with a delay of 0.691 s (±
0.671 sd). Further analysis about the relative positions of group members showed
that the large majority of group leaders preceded followers while appraising, and
that all leaders preceded followers when they decided to cross. Results highlighted
that the leaders were able to better evaluate the distance and speed of oncoming ve-
hicles according to the position, and so they take the decision to cross first followed
by the companion.
4 Final Remarks
A video-recorded observation was performed at a non-signalized intersection to
study the crossing behaviour of social groups, comparing data among single pedes-
trians and dyads. Trend analyses on the speeds of the tracked pedestrians showed
that crossing behaviour is characterised by three distinct phase: approaching, ap-
praising and crossing. In particular, the appraising phase consists on a significant
deceleration in proximity of the curb to evaluate the distance and speed of oncom-
ing vehicle to safely cross (i.e. decision making).
Results showed that dyads walked in average 17% slower than singles among
the three crossing phases, decelerating the 11% more than singles while appraising.
Group crossing behaviour is based on the emergence of a leader who completes the
appraising first and decide to cross, followed by the companion. However, followers
did not merely imitate the crossing behaviour of the leaders in judging the safety
gap from oncoming vehicles, since no significant difference was found in the time
duration of the appraising phase between them. Finally, no significant difference
was found in the safety gap accepted by singles and dyads.
Future works will be aimed at enlarging the sample of tracked pedestrians for
data analysis, including also groups composed by unfamiliar pedestrians. More-
over, we consider the possibility to test the impact of social imitation among unfa-
miliar pedestrians on risky crossing behaviour, such as non-compliant jaywalking
behaviour out of zebra-striped crossing.
While no significant difference has emerged in comparing the crossing decision
of singles and dyads, the presented results could be of notable interest for those in-
volved in modelling and simulation of urban interactions. In particular, the results of
8
Gorrini, A., Crociani, L., Vizzari, G., Bandini, S.
the observation will be used to support the further development of a computer-based
simulator of pedestrian/vehicles interactions at non-signalised intersections [5], fo-
cusing on modelling heterogeneous crossing speed among pedestrians.
Acknowledgements The Italian policy was complied in order to exceed the ethical issues about
the privacy of the people recorded without their consent. The authors thank Fatima Zahra Anouar
for her fruitful contribution in data analysis. The recorded video tape and the annotated data set are
available for research purposes.
References
1. Dang, N.T., Cavallo, V., Thomson, J., Gorrini, A., Piwowarczyk, A., Vienne, F., Bandini, S.,
Pierre, G.S.: Crossing the street in dyads: Social influence on pedestrian behavior in an im-
mersive virtual environment. In: Proceedings of the International Conference on Road Safety
and Simulation, Delft, Netherlands (in press, 2017)
2. Domm`es, A., Le Lay, T., Vienne, F., Dang, N.T., Beaudoin, A.P., Do, M.C.: Towards an expla-
nation of age-related difficulties in crossing a two-way street. Accident Analysis & Prevention
85, 229 -- 238 (2015)
3. Evans, D., Norman, P.: Understanding pedestrians' road crossing decisions: an application of
the theory of planned behaviour. Health Education Research 13(4), 481 -- 489 (1998)
4. Faria, J.J., Krause, S., Krause, J.: Collective behavior in road crossing pedestrians: the role of
social information. Behavioral Ecology 21(6), 1236 -- 1242 (2010)
5. Feliciani, C., Crociani, L., Gorrini, A., Vizzari, G., Bandini, S., Nishinari, K.: A simulation
model for non-signalized pedestrian crosswalks based on evidence from on field observation.
Intelligenza Artificiale (in press, 2017)
6. Festinger, L., Pepitone, A., Newcomb, T.: Some consequences of de-individuation in a group.
The Journal of Abnormal and Social Psychology 47(2S), 382 (1952)
7. Gorrini, A., Vizzari, G., Bandini, S.: Towards modelling pedestrian-vehicle interactions: Em-
pirical study on urban unsignalized intersection. In: Proceedings of the 8th International Con-
ference on Pedestrian and Evacuation Dynamics, Hefei, China (in press, 2016)
8. Hamed, M.M.: Analysis of pedestrians' behavior at pedestrian crossings. Safety Science
38(1), 63 -- 82 (2001)
9. Harrell, W.A.: Factors influencing pedestrian cautiousness in crossing streets. The Journal of
Social Psychology 131(3), 367 -- 372 (1991)
10. Highway Capacity Manual: Highway Capacity Manual. Transportation Research Board, Na-
tional Research Council, Washington, DC (2010)
11. Hirschi, T., Gottfredson, M.R.: The generality of deviance. New Jersey (1994)
12. Perumal, V., et al.: Study on pedestrian crossing behavior at signalized intersections. Journal
of traffic and transportation engineering (English edition) 1(2), 103 -- 110 (2014)
13. Ren, G., Zhou, Z., Wang, W., Zhang, Y., Wang, W.: Crossing behaviors of pedestrians at
signalized intersections: observational study and survey in china. Transportation Research
Record: Journal of the Transportation Research Board (2264), 65 -- 73 (2011)
14. Rosenbloom, T.: Crossing at a red light: Behaviour of individuals and groups. Transportation
research part F: traffic psychology and behaviour 12(5), 389 -- 394 (2009)
15. Sisiopiku, V.P., Akin, D.: Pedestrian behaviors at and perceptions towards various pedestrian
facilities: an examination based on observation and survey data. Transportation Research Part
F: Traffic Psychology and Behaviour 6(4), 249 -- 274 (2003)
16. WHO: Global status report on road safety 2015. World Health Organization (2015)
|
1508.00023 | 2 | 1508 | 2016-01-15T23:11:59 | Work Capacity of Freelance Markets: Fundamental Limits and Decentralized Schemes | [
"cs.MA",
"cs.CY"
] | Crowdsourcing of jobs to online freelance markets is rapidly gaining popularity. Most crowdsourcing platforms are uncontrolled and offer freedom to customers and freelancers to choose each other. This works well for unskilled jobs (e.g., image classification) with no specific quality requirement since freelancers are functionally identical. For skilled jobs (e.g., software development) with specific quality requirements, however, this does not ensure that the maximum number of job requests is satisfied. In this work we determine the capacity of freelance markets, in terms of maximum satisfied job requests, and propose centralized schemes that achieve capacity. To ensure decentralized operation and freedom of choice for customers and freelancers, we propose simple schemes compatible with the operation of current crowdsourcing platforms that approximately achieve capacity. Further, for settings where the number of job requests exceeds capacity, we propose a scheme that is agnostic of that information, but is optimal and fair in declining jobs without wait. | cs.MA | cs | Work Capacity of Freelance Markets: Fundamental
Limits and Decentralized Schemes
Avhishek Chatterjee, Lav R. Varshney, and Sriram Vishwanath
1
6
1
0
2
n
a
J
5
1
]
A
M
.
s
c
[
2
v
3
2
0
0
0
.
8
0
5
1
:
v
i
X
r
a
Abstract
Crowdsourcing of jobs to online freelance markets is rapidly gaining popularity. Most crowdsourcing platforms
are uncontrolled and offer freedom to customers and freelancers to choose each other. This works well for unskilled
jobs (e.g., image classification) with no specific quality requirement since freelancers are functionally identical. For
skilled jobs (e.g., software development) with specific quality requirements, however, this does not ensure that the
maximum number of job requests is satisfied. In this work we determine the capacity of freelance markets, in terms
of maximum satisfied job requests, and propose centralized schemes that achieve capacity. To ensure decentralized
operation and freedom of choice for customers and freelancers, we propose simple schemes compatible with the
operation of current crowdsourcing platforms that approximately achieve capacity. Further, for settings where the
number of job requests exceeds capacity, we propose a scheme that is agnostic of that information, but is optimal
and fair in declining jobs without wait.
freelance markets, capacity, queuing theory, decentralized algorithms
Index Terms
I. INTRODUCTION
Methods and structures for information processing have been changing. Enabled by the proliferation of modern
communication technologies, globalization and specialization of workforces has led to the emergence of new
decentralized models of informational work. Moreover, the millennial generation now entering the workforce often
favors project-based or job-based work, as in crowdsourcing and social production [2], [3], rather than long-term
commitments [4]. Indeed over the last decade, more than 100 ‘human clouds’ have launched with a variety of
structures. These platforms serve clients by harnessing external crowds, and global enterprises similarly harness
their internal crowds [5]–[7], making use of human cognitive surplus for information processing [8].
Platforms follow different collective intelligence models [9], [10], and require different strategies for allocating
informational work to workers. In crowdsourcing contest platforms like InnoCentive and TopCoder, there is self-
selection: work is issued as an open call and anyone can participate in any job; the best submission wins the reward
[11]–[14]. In microtask crowdsourcing platforms like Amazon Mechanical Turk, any worker is assumed able to
do any job and so first-come-first-serve strategies are often used; level of reliability may be considered in optimal
allocation [15]. In freelance markets like oDesk and Elance, however, specialized jobs must be performed by skilled
workers: allocation requires careful selection from the large pool of variedly-skilled freelancers.
Freelance market platforms serve as spot markets for labor by matching skills to tasks, often performing on-
demand matching at unprecedented scales. For example, oDesk had 2.5 million workers and nearly 0.5 million clients
in 2013 [10]. Herein we study allocation and scheduling of informational work within these kinds of platforms, via
a queuing framework. We aim to establish fundamental limits through a notion of work capacity, and also develop
decentralized algorithms, which are easily-computed, that nearly achieve these performance limits.
Freelancers may have one or more skills (that are known, cf. [16]) and jobs may have multiple parts, called tasks,
that require separate skills. Due to job skill requirement variety and limited freelancer ability, it is often not possible
Part of the material in this paper was presented at IEEE INFOCOM, Hong Kong, April-May 2015 [1].
A. Chatterjee is with the Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX 78712 USA.
(e-mail: [email protected]).
L. R. Varshney is with the Department of Electrical and Computer Engineering and the Coordinated Science Laboratory, University of
Illinois at Urbana-Champaign, Urbana, IL 61801 USA. (e-mail: [email protected]).
S. Vishwanath is with the Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX 78712
USA. (e-mail: [email protected]).
2
to find a freelancer that meets all requirements for a job: a job may have to be divided among freelancers. Moreover,
a task in a job may require so much time that even the task may have to be divided among multiple freelancers.
There are reputation systems within freelance market platforms, so freelancers have a reputation level as well as
minimum acceptable hourly rate and skills, which allow worker categorization. Some freelancers are adaptable in
terms of hours available to spend on a particular type of task, whereas others pre-specify hours available for each
kind of task. Here we consider the non-adaptable setting where, for example, a freelancer may be available for 20
hours (per week) of any C++ or Java programming, or may be available for 10 hours (per week) of C++ and 10 hours
of Java. Studying limits for adaptable freelancers and designing centralized schemes (and their approximations) are
similar, but the distributed schemes require a different approach.
The objective of the platform is to find a good allocation of jobs (and tasks) to freelancers. Since working on a
task requires synchronization among freelancers, work can only start when the whole task has been allocated. On
the other hand, for some jobs there are interdependencies between different tasks [17] and hence, for these jobs all
tasks must be allocated before the job starts. Moreover, some jobs may require all parts to be done by freelancers
with the same level of expertise for uniform quality and money spent. These considerations lead to concepts of
decomposability and flexibility that are central to our development.
In the ethos of self-selection, it is desirable for crowd systems to not be centrally controlled, but rather for jobs
and freelancers to choose each other. Currently, this may happen randomly or greedily. This is clearly not optimal,
as the following example illustrates. Consider two types of jobs (single task) and two categories of freelancers. A
type 1 job can be served by either of the worker categories (example, lower reputation requirement) whereas type 2
jobs can only be served by category 2 workers. If freelancers and jobs are allocated arbitrarily then it may happen
that type 1 uses many category 2 freelancers and many type 2 jobs remain unserved.
Optimal centralized allocation of informational tasks under the constraints of crowd systems is related to hard
combinatorial problems such as the knapsack problem. Compared to scheduling problems in computer science [18],
communication networks [19], [20], and operations research [21], crowd systems face challenges of freedom of
self-selection, need for decentralized operation, and uncertainty in resource availability.
Prior works in the information theory, networking, and queueing literatures are similar to our work in terms of
theoretical framework, performance metrics, and the nature of performance guarantees, but are not directly related.
The notion of capacity of a resource-shared system where jobs are queued until they are served and the notion of a
capacity-achieving resource allocation scheme for this kind of system came to prominence with the work of Tassiulas
and Ephremides [22], [23]. The capacity concept and capacity-achieving schemes were subsequently developed for
applications in communication networks [19], [20], [24], [25], cloud computing [26], online advertising [27], [28],
and power grids [29], among others. With the advent of cloud services, large-scale systems have attracted significant
research interest: resource allocation schemes and their performance (queueing delays, backlogs, etc.) in the large-
scale regimes have been studied [30]–[33].
In this paper, our goal is to understand the fundamental limits (capacity) of freelance markets and ways to
achieve this ultimate capacity. We first develop a centralized scheme for achieving these maximum allocations
where a central controller makes all job allocation decisions. Given the potential large scale of platforms, we also
discuss low-complexity approximations of the centralized scheme that almost achieve the limit. Finally, with an eye
towards giving flexibility to customers (job requesters) in choosing freelancers, we propose simple decentralized
schemes with minimal central computation that have provable performance guarantees. Further, since job arrival
and freelancer availability processes are random (and sometimes non-stationary), we also address ways to adapt
when the system is operating outside its capacity limits.
II. SYSTEM MODEL
We first provide formal definitions of the nature of informational work and workers, and establish notation.
Freelancers (or agents) are of L categories. In each category l ∈ [L], there are M l types of agents depending
on their skill sets and available hours. There are S skills among agents of all categories and types. An agent of
category l and type i has a skill-hour vector hl
Jobs posted on the platform are of N types. Each type of job j ∈ [N ] needs a skill-hour service rj, i.e. rj,s
hours of skill s. A part of a job of type j involving skill s is called a (j, s)-task if rj,s > 0, which is the size of
this task.
i,s available hours for work involving skill s ∈ [S].
i, i.e. hl
3
A job of type j can only be served by agents of categories l ∈ N (j) ⊂ [L]. This restriction is captured by a
bipartite graph G = ([N ], [L], E), where a tuple (j, l) /∈ E ⊂ [N ] × [L] implies that category l agents cannot serve
jobs of type j.
On the platform, jobs are allocated at regular time intervals to available agents, these epochs are denoted by
t ∈ {1, 2, . . .}. Jobs that arrive after epoch t has started are considered for allocation in epoch t + 1, based on
agents available at that epoch. Unallocated jobs (due to insufficient number of skilled agents) are considered again
in the next epoch.
Jobs arrive according to a ZN
number of jobs of type j that arrive in scheduling epoch t.
The stochastic process of available agents at epoch t is U(t) = (U1(t), U2(t), . . . , UL(t)). For each agent
+ -valued stochastic process A(t) = (A1(t), A2(t), . . . , AN (t)), where Aj(t) is the
M l(t)(cid:1) denotes the number of available agents of different types at epoch
2(t), . . . , U l
category l, Ul(t) =(cid:0)U l
1(t), U l
t.
We assume processes A(t) and U(t) are independent of each other and that each of these processes is independent
and identically distributed for each t.1 We also assume that each of these processes has a bounded (Frobenius norm)
covariance matrix. Let Γ(·) be the distribution of U(t), and let λ = E[A(t)] and µl = E[Ul(t)] for l ∈ [L] be the
means of the processes.
At any epoch t, only an integral allocation of a task (say (j, s)) is possible. A set of tasks t1, t2, . . . , tn of size
r1, r2, . . . , rn of skill s can be allocated to agents 1, 2, . . . , m only if available skill-hours for skill s of these agents
h1, h2, . . . , hm satisfy
n(cid:88)
m(cid:88)
vip ≤ hi,
vqj ≥ rj, j ∈ [n], i ∈ [m]
for some {vpq ≥ 0}.
Whether different tasks of a job can be allocated at different epochs and across different categories of agents
p=1
q=1
depend on the type of the job.
Definition 1. A type of job j is called non-decomposable (decomposable) if different tasks comprising it are (are
not) constrained to be allocated at the same epoch.
Definition 2. A type of job j is called inflexible (flexible) if different tasks as well parts of tasks comprising it are
(are not) constrained to be allocated to the same category of agents.
In a system with only decomposable jobs, given a set of {ul = (ul
M l) : l ∈ [L]} agents (that is, ul
i
agents of category l and of type i within that category), a number aj,s of (j, s)-tasks can be allocated only if there
exist non-negative {zl
j,s : (l, j, s) ∈ [L] × [N ] × [S]} satisfying
2, . . . , ul
1, ul
(cid:88)
(cid:88)
l
j∈[N ]
zl
j,s = aj,s, zl
j,srj,s ≤ (cid:88)
zl
i∈[M l]
j,s = 0 if (j, l) /∈ E, for all j ∈ [N ], s ∈ [S],
i,s, for all l ∈ [L], s ∈ [S].
ihl
ul
(1)
On the other hand, given a set of {ul = (ul
decomposable jobs, aj jobs of type j (for each j) can be allocated only if there exist non-negative {zl
[L] × [N ]} satisfying
M l) : l ∈ [L]} agents in a system with only non-
j,s : (l, j) ∈
1, ul
2, . . . , ul
(1) and aj,s = aj for all j, s.
(2)
Intuitively, the conditions imply that required skill-hours for the set of jobs is less than the available skill-hours
of agents. The {zl
j,s} capture a possible way of dividing tasks across multiple category of agents, as they can be
interpreted as the number (possibly fraction) of (j, s)-tasks allocated to l-category agents. Note that conditions (1)
and (2) are necessary for allocations of decomposable and non-decomposable jobs respectively. These conditions
only imply that there exist possible ways of splitting jobs and tasks across different categories of agents to ensure
integral number of tasks (jobs) are allocated in case of decomposable (non-decomposable) jobs.
1Most of our results can be extended to stationary ergodic processes.
4
For a system with only flexible jobs, different parts of a task can be allocated to different categories and a category
+ . Thus flexible and decomposable
can be allocated parts of tasks. Hence, {zl
(non-decomposable) jobs need to satisfy condition (1) (condition (2)) which we refer to as FD (FND).
j,s} can possibly take any value in RLN S
For inflexible jobs, a necessary condition the allocation must satisfy is that each category gets the same integral
number of (j, s)-tasks for all s, j, i.e.,
j,s ∈ Z+ s.t. zl
zl
j,s = zl
j,s(cid:48) for all s, s(cid:48), j.
(3)
An allocation of inflexible and decomposable (non-decomposable) jobs needs to satisfy conditions (1) (condition
(2)) and (3), which we refer to as ID (IND).
For simplicity, in this work we focus on systems with only a single one of these four classes of jobs.2 For brevity
we use the same abbreviations to refer to class of job, as we use for the necessary conditions. Thus, we have FD,
FND, ID, and IND systems.
agents is as follows: λ(N ) =(cid:80)N
a job generally requires more diversity than a single agent possesses, implying M =(cid:80)
seen in practice, we assume λj(N ) = ω(1),∀j ∈ [N ] and (cid:80)
In crowd systems, the scaling of number of job and agent types, rate of job arrivals, and number of available
j=1 λj scales faster than N, i.e. λ(N ) = ω(N ) or limN→∞ N/λ(N ) = 0 and the
number of skills S scale slower than N, i.e. S = o(N ). In practice, a job requires at most a constant number of
skills d, implying there are Ω(Sd) possible job types. On the other hand, the number of skills of an agent d(cid:48) < d as
). L = O(1),
as it relates to variation in reputation levels and hourly rates, and so M = o(N ). Beyond these system scalings
j:rj,s>0 λj(N ) = Ω (N c) for all s ∈ [S], for some
l M l = O(Sd(cid:48)
c > 0. In the sequel, we assume these scaling patterns and refer to them as crowd-scaling.
III. CAPACITY, OUTER REGION, AND CENTRALIZED ALLOCATION
In this section we study the limits of a freelance market with centralized allocation and present a centralized
algorithm that achieves the limit. We also discuss a simpler upper bound for the capacity region in terms of first-order
statistics of the system. These results on ultimate system limits and ways to achieve them are not only important in
their own right, but also serve as benchmarks for later discussion of decentralized schemes that provably achieve
nearly the same limits.
To formally characterize the maximal supportable arrival rate of jobs we introduce some more notation. For each
j ∈ [N ], let Qj(t) be the number of unallocated jobs that are in the crowd system just after allocation epoch t− 1.
As defined above, Aj(t) is the number of jobs of type j that arrive between starts of epochs t− 1 and t. Let Dj(t)
be the number of jobs of type j that have been allocated to agents at epoch t; we call a job allocated only when
all parts have been allocated. Thus the evolution of the process Qj(t) can be written as:
Qj(t + 1) = Qj(t) + Aj(t) − Dj(t).
(4)
Note that at any epoch t, at most Qj(t) + Aj(t) type j jobs can be allocated, as this is the total number of type j
jobs at that time and hence Dj(t) ≤ Qj(t) + Aj(t), implying Qj(t) ≥ 0.
Notation and Convention. We denote the interior and the closure of a set C by C and ¯C, respectively. When
we say λ = (λ1, λ2, . . . , λN ) ∈ Λ ⊂ RN S
+ we mean λS = ((λ1, λ1, . . . , S times) , (λ2, λ2, . . . , S times) , . . .) ∈ Λ.
Also, whenever we say Λ ⊆ (⊇)Λ(cid:48) for Λ(cid:48) ⊂ RN
Definition 3. An arrival rate λ is stabilizable if there is a job allocation policy P under which Q(t) = (Qj(t), j ∈ [N ])
has a finite expectation, i.e., lim supt→∞ E[Qj(t)] < ∞, for all j. The crowd system is called stable under this
policy.
Definition 4. CΓ, a closed subset of RN
agent-availability process if any λ ∈ CΓ is stabilizable and any λ /∈ CΓ is not stabilizable.
+ is the capacity region of a crowd system for a given distribution Γ of the
+ , we mean for any λ ∈ RN
+ , λ ∈ Λ(cid:48) ⇐ (⇒)λS ∈ Λ.
2Extension to combinations of multiple classes is not much different but requires more notation.
A. Capacity Region and Outer Region
agents u =(cid:0)ul
Let us characterize the capacity regions of different classes of crowd systems. For any given set of available
i : 1 ≤ i ≤ M l, 1 ≤ l ≤ L(cid:1), we define the set of different types of tasks (aj,s) that can be allocated
in a crowd system. Note that the necessary conditions to be satisfied for tasks to be allocated are specific to the
class of crowd system.
5
Using the explicit conditions (1), (2), and (3) for tasks (jobs) to be allocated, we define CFD(u), CFND(u),
CID(u), and CIND(u) as the set of tasks that can be allocated in FD, FND, ID, and IND systems respectively for
given availability u. We denote these sets generically by C(u) and refer to conditions FD, IFD, FND, and IND
generically as crowd allocation constraint or CAC.
(cid:110)
(aj,s ∈ Z+) : ∃(cid:16)
(cid:17)
zl
j,s
satisfying CAC
,
and C(u) is the convex hull of C(u).
C(u) :=
The following theorem generically characterizes capacity regions of different crowd systems.
Theorem 5. Given a distribution Γ of agent-availability, i.e., Γ(u) = Pr (U(t) = u), for a λ /∈ ¯C(Γ) there exists
no policy under which the crowd system is stable, where
(cid:111)
.
λ =
(cid:88)
u∈ZM
+
(cid:88)
l
C(Γ) =
Γ(u)λ(u) : M =
M l, λ(u) ∈ C(u)
For FD, FND, and IND systems, for any λ ∈ C(Γ) there exists a policy such that the crowd system is stable.
Proof: See Appendix II.A.
This implies that for FD, FND, and IND systems capacity region CΓ = ¯C(Γ) and for ID systems CΓ ⊆ ¯C(Γ)
(possibly strict). Note that the conditions FD, FND, ID, and IND (generically CAC) are necessary conditions for a
valid allocation. The above theorem implies these conditions are also sufficient, except for ID systems. In Sec. V,
we present an alternate characterization of the capacity region for inflexible systems.
Note C(Γ) depends on the distribution of agent availability Γ, but it is hard to obtain this distribution for large
and quickly-evolving systems in practice. Hence, a characterization in terms of simpler system statistics is of use.
Below is a characterization of a region beyond which no arrival rate can be stabilized. Borrowing terminology from
multiterminal Shannon theory, we call this the outer region.
For any set J ⊂ [N ], define N (J) = {l ∈ [L] : ∃j ∈ J s.t. (j, l) ∈ E} and the closed subset of RN
+ ,
λ : ∀J ⊂ [N ],∀s,
(cid:88)
j∈J
λjrj,s ≤ (cid:88)
(cid:88)
l∈N (J)
i∈M l
Cout
µ =
.
ihl
µl
i,s
Theorem 6. For any distribution Γ with mean µ, CΓ ⊆ Cout
µ .
Proof: See Appendix II.B.
µ
µ
because Cout
In general, CΓ is a strict subset of Cout
only captures the balance of skill-hours in the crowd-
system, i.e. average skill-hours requirement is no more than average availability, but partial allocation of a task is
not acceptable in a crowd system. Moreover, for non-decomposable jobs all tasks of a job have to be allocated
simultaneously. Hence, meeting an average skill-hour balance criterion may be far from being sufficient for stability.
For inflexible systems the requirements are even stricter, which is likely to increase the gap between the outer region
and the true capacity region. In Sec. V we present a tighter outer region for inflexible systems.
µ may be non-empty when CΓ is empty. For example, consider a simple non-decomposable
crowd system with N = L = 1 and M 1 = S = 2. Let each job require 1 hour of both skills, type i agents have
only 1 hour available for skill i and none for other skills, U1(t) be uniformly distributed on {(0, 10), (10, 0)}, and
λ = (4, 4). Then clearly λ ∈ Cout
µ , but note that at any time there is only one type of skill available, hence no job
can be allocated. This implies CΓ = ∅.
In certain scenarios Cout
6
B. Centralized Allocation
Though there exists a policy for each λ ∈ CΓ that stabilizes the system, these policies may differ based on
λ and may depend on the job-arrival and agent availability statistics. Changing policies based on arrival rate and
statistics is not desirable in practical crowd systems due to the significant overhead. Below we describe a centralized
statistics-agnostic allocation policy which stabilizes any λ ∈ CΓ. Later we discuss computational cost of this policy
for different classes of crowdsourcing system and present simpler distributed (or almost distributed) schemes with
provable performance guarantees under some mild assumptions.
To describe the scheme we introduce some more notation. Let Qj,s(t) be the number of s-tasks (skill s) of type
j jobs just after the allocation epoch t− 1 and let Dj,s(t) be the number of s-tasks (skill s) of type j jobs allocated
at epoch t. Then
Qj,s(t + 1) = Qj,s(t) + Aj(t) 1 (rj,s > 0) − Dj,s(t).
Note that for all t, due to the CAC condition on allocation, Dj,s(t) ∈ C(U(t)). Moreover, there is an additional
restriction that Dj,s(t) ≤ Qj,s(t) + Aj(t), as there are Qj,s(t) + Aj(t) part s of job type j in the system at that
time, which in turn implies Qj,s(t) ∈ Z+ for all t. Note that as Dj,s(t) ∈ C(U(t)), for non-decomposable systems
Qj,s = Qj,s(cid:48) for all j, s, s(cid:48), whereas for decomposable systems they may differ.
Algorithm 1 MaxWeight Task Allocation (MWTA)
Input: {Qj,s(t) : j ∈ [N ], s ∈ [S]}, A(t) and U(t) at t
Output: Allocation of jobs to agents
MaxWeight
(cid:16)
(cid:17)
zl
j,s(t) : l, j, s
(cid:16)
(cid:17)
= arg
max
j,s∈Z+:l,j,s)
(zl
zl
j,s
satisfy CAC with aj,s = ∆j,s(t)∀j, s.
(cid:88)
j,s
Qj,s(t)∆j,s
s.t.
Task Allocation
for j = 1 : N do
Order j-type jobs arbitrarily, Oj
for s = 1 : S do
Use order Oj among non-zero (j, s)-tasks
l = 1
while l ≤ L and(cid:80)l−1
Allocate [(cid:80)l−1
j,s : (cid:80)l
k=1 zk
j,s < Qj,s(t) + Aj(t) do
k=1 zl
k=1 zk
k] (j, s) tasks to category l. Here tasks [x : x + y] are task set I =
{(cid:100)x(cid:101),···(cid:98)x + y(cid:99)} (in the ordering Oj), ((cid:100)x(cid:101) − x) fraction of task (cid:100)x(cid:101) and 1 + x + y − (cid:100)x + y(cid:101) fraction
of task (cid:100)x + y(cid:101).
l ← l + 1
end while
Order agents of category l arbitrarily
for s = 1 : S do
Agents pick maximum (as per availability constraint) tasks (or part) in order from (cid:80)
j min(zl
j,s, Qj,s(t) +
end for
end for
for l = 1 : L do
Aj(t))rj,s hours
end for
end for
We propose the MaxWeight Task Allocation (MWTA) policy, Alg. 1, to allocate tasks to agents at epoch t
based only on the knowledge of Q(t), A(t), and U(t), and therefore statistics-agnostic. It is based on MaxWeight
matching [19], [20].
j,s} that satisfies CAC. The following theorem
implies that MWTA allocates tasks optimally. The proof of the theorem is based on adapting the proof of optimality
It is apparent that the MaxWeight part of the algorithm finds a {zl
7
of the MaxWeight algorithm under the constraints and assumptions of crowd systems. It implicitly relies on the
following result.
Proposition 7. For any u and Q, and {zl
feasible allocation for FD, FND, and IND systems.
j,s} satisfying CAC, the Task Allocation part of MWTA (Alg. 1) gives a
Proof: See Appendix II.C.
Theorem 8. MWTA (Alg. 1) stabilizes FD, FND, or IND crowd systems for any arrival rate λ ∈ CΓ (for respective
CΓ).
Proof: See Appendix II.D.
IV. SINGLE-CATEGORY SYSTEMS AND DECENTRALIZED ALLOCATIONS
There is effectively a single category of agents in many platforms with a large population of new freelancers,
whose reputations are based on evaluation tests for skills and who are paid at a fixed rate. Hence designing efficient
allocation schemes for single-category systems are of particular interest, as this population of agents are significant
in ever-evolving crowd systems. Insights drawn from single-category systems are also useful in controlling multi-
category systems, Sec. V.
j,s = aj,s ∈ Z+ and hence the feasibility condition (1) reduces
uihi,s for all s ∈ [S], aj,s ∈ Z+,
For a single category system (L = 1), note that z1
aj,srj,s ≤(cid:88)
(cid:88)
to:
j
i
(cid:80)
with condition (2) additionally requiring aj,s = aj,s(cid:48) for all j, s, s(cid:48). Thus, C(u) is the set of {aj,s} satisfying the
above conditions for respective classes of jobs and C(Γ) is the weighted (by Γ(u)) sum of convex hulls of C(u)s,
here CΓ = ¯C(Γ).
(cid:80)
µ has a simple characterization as well. As for any j ∈ [N ], (j, 1) ∈ E, and N (J) = 1 for all J ⊂ [N ],
Cout
(cid:110)
λ :(cid:80)
i∈[M ] µihi,s. Thus it is sufficient to satisfy the inequality for J = [N ], and hence,
l∈N (J)
Cout
i∈[M ] µihi
µ =
The MaxWeight computation in MWTA for single-category systems turns out to be the following integer linear
i,s = (cid:80)
j∈[N ] λjrj ≤(cid:80)
i∈[M l] µl
(cid:111)
ihl
.
program (ILP), which is related to knapsack problems.
(5)
(cid:88)
∆j,srj,s ≤(cid:88)
j,s
arg max
{∆j,s:j,s}
s.t. (cid:88)
Qj,s∆j,s
uihi,s∀s ∈ [S],
j
i
∆j,s = ∆j,s(cid:48),∀s, s(cid:48), j (only for ND)
For decomposable and non-decomposable systems, this is a single knapsack and multi-dimensional knapsack
problem [34], respectively, and hence NP-hard. There exist fully polynomial time approximations (FPTAS) for
single knapsack, whereas for multi-dimensional knapsack only polynomial time approximations (PTAS) are possible
1− ∈ CΓ}. Also, note
[34]. With this approximation, say 1 − , the MWTA policy stabilizes (1 − )CΓ = {λ : λ
that for large crowd systems each λi is large and hence stabilizing any λ with λ + 1 ∈ CΓ is almost optimal. The
above ILP can be relaxed to obtain a linear program, an allocation based on which achieves this approximation
(see Appendix I.A for details).
A. Decentralized Allocations
Now we show that due to the structure of the crowd allocation problem and the fact that crowd systems are large,
simple allocation schemes with minimal centralized control achieve good performance under mild assumptions
on arrival and availability processes. Interestingly, though the centralized optimal allocation requires solving a
knapsack problem at each epoch and greedy schemes are known to be sub-optimal for knapsack problems [34], we
propose two simple greedy schemes that are almost optimal with good performance guarantees. One of them, called
8
GreedyAgent allocation provably performs well for decomposable systems and offers the freedom of selection to
freelancers. Another, called GreedyJob allocation has provable performance guarantees for both decomposable and
non-decomposable systems while allowing customers (job requesters) the freedom of selection. Thus, in some sense,
this shows that though greedy algorithms can be suboptimal for an arbitrary allocation instant (at each epoch), for
a dynamical system over long time, its performance is good.
Algorithm 2 GreedyAgent Allocation
Input: A(t)
Output: Job to agent allocations
A: set of agents, T : set of tasks
while A and T non-empty do
Agents in A contend (pick random numbers) and a wins
for each skill with non-zero skill hour do
a picks as many integral tasks as it can pick
if a has remaining available hour then
a Picks from remaining parts of the partially allocated task
if a has remaining available hour then
a picks part of any unallocated task
end if
Remove fully allocated tasks from T
end if
end for
A = A\{a}
end while
Tasks with partial allocations are not allocated
In GreedyAgent allocation (Alg. 2), agents themselves figure out the allocation via contention, without any central
control. Agents need no knowledge about the agent population, but do need information on the available pool of
jobs and have to agree on certain norms. In most freelance market platforms, this information is readily available,
and so an algorithm like this is natural. As expected, this scheme may not be able to stabilize any arrival rate in
CΓ for any ergodic job-arrival and agent-availability processes, but it has good theoretical guarantees under some
mild assumptions on the job arrival and agent availability processes.
Definition 9. A random variable X is Gaussian-dominated if E[X 2] ≤ E[X]2 + E[X] and for all θ ∈ R,
E[eθ(X−E[X])] ≤ exp
Definition 10. A random variable X is Poisson-dominated if for all θ ∈ R, E[eθ(X−E[X])] ≤ e
(cid:110) ((E[X 2]−E[X]2)θ2
E[X](eθ−θ−1).
(cid:111)
2
Note that these domination definitions imply that the variation of the random variable around its mean is dominated
in a moment generating function sense by that of a Gaussian or Poisson random variable.3 Such a property is
satisfied by many distributions including Poisson and binomial that are used to model arrival processes for many
systems, e.g., telephone networks, internet, call centers, and some freelance markets [5], [6]. It is not hard to
show that sub-Gaussian distributions (standard in machine learning [36]) that are symmetric around their mean, are
Gaussian-dominated.
The following theorem gives a guarantee on the performance of GreedyAgent, under mild restrictions on the job-
arrival and agent-availability processes. Independence assumptions are not too restrictive for large crowd systems,
where jobs and agents may come from different well-separated geographies or organizational structures.
Theorem 11. If the arrival processes {Aj(t)} and the agent availability processes {Ui(t)} are i.i.d. across time
and independent across types (jobs and agents) and all these processes are Gaussian-dominated (and/or Poisson-
dominated), then for any given α ∈ (0, 1], there exists an Nα such that GreedyAgent allocation stabilizes any
3Domination in this sense is used in bandit problems [35].
(cid:111)
unallocated tasks is O(S log N ) with probability 1 − o(cid:0) 1
arrival rate λ ∈ (1 − α)Cout
N ≥ Nα. Moreover, for any arrival rate in (1 − α)Cout
1−α λ ∈ Cout
µ :=
(cid:110)
λ :
µ
1
(cid:1).
N 2
for any single-category decomposable crowd system with
µ , at steady state, after an allocation epoch, the number of
9
Proof: See Appendix II.E.
As CΓ ⊆ Cout
µ , this implies that the greedy scheme stabilizes an arbitrarily large fraction of the capacity region,
under the assumptions on the arrival and availability processes. As S = o(N ), more specifically O(N c) for c < 1,
the above bound on number of jobs imply that there are o(N ) unallocated tasks at any time. This in turn implies
that unallocated tasks per type (average across types) is o(1), i.e., vanishingly small number of tasks per type are
unallocated as the system scales.
In GreedyAgent, there is no coordination among agents while picking tasks within jobs. Hence in a non-
decomposable system, many tasks may be picked by agents but only few complete jobs are allocated. As more and
more jobs accumulate, the chance of this happening increases, resulting in more accumulation. This can result in
the number of accumulated jobs growing without bound, as formalized below.
Proposition 12. There exists a class of non-decomposable crowd systems with Poisson-dominated (as well as
Gaussian-dominated) distributions of arrival and availability, such that the system is not stable under GreedyAgent
allocation.
Proof: See Appendix II.F.
Hence, we propose another simple greedy scheme that works for both decomposable and non-decomposable
systems. The GreedyJob allocation scheme (Alg. 3) is completely distributed and hence a good fit for crowd
systems. GreedyJob has similar performance guarantees for both decomposable and non-decomposable systems as
GreedyAgent has for decomposable systems only.
Algorithm 3 GreedyJob Allocation
Input: U(t)
Output: Job to agent allocations
J : set of all jobs
while Available skill-hours of agents and J (cid:54)= ∅ do
Jobs in J contend (pick random numbers) and J wins
if J finds agents to allocate all tasks then
Allocate to those agents
else
J does not allocate anything
end if
J = J \{J}
end while
i µihi,s −(cid:80)
Theorem 13. If the arrival processes {Aj(t)} and the agent availability processes {Ui(t)} are i.i.d. across time and
independent across types (jobs and agents), all these processes are Gaussian-dominated (and/or Poisson-dominated)
i µihi,s(cid:48) is O(subpoly(N )), then for any given α ∈ (0, 1], ∃Nα such that GreedyJob
for any single-category crowd-system
µ , at steady state, after an allocation epoch, total number
and ∀s, s(cid:48),(cid:80)
(cid:111)
of unallocated jobs (adding all types) is O(log N ) with probability 1 − o(cid:0) 1
allocation stabilizes any arrival rate λ ∈ (1− α)Cout
with N ≥ Nα. Moreover, for any arrival rate in (1−α)Cout
1−α λ ∈ Cout
µ :=
(cid:1).
(cid:110)
λ :
µ
1
N 2
Proof: See Appendix II.G.
In Sec. V we propose a decentralized scheme for multi-category systems that uses the two single-category
decentralized schemes as building blocks. At the end of Sec. V we briefly discuss the suitability of these decentralized
schemes for crowd systems in terms of implementability on crowd platforms.
10
V. MULTI-CATEGORY SYSTEMS
Sec. III characterized the capacity region and developed an optimal centralized scheme for crowd systems in,
whereas Sec. IV discussed simple decentralized schemes for single-category systems. Here we return to multi-
category systems, briefly discussing computational aspects of MWTA, followed by an alternate approach to the
capacity and outer region of inflexible systems that yields a simple optimal scheme. We also present a decentralized
scheme based on insights from the optimal scheme and the decentralized allocations in Sec. IV.
The MWTA scheme, which is throughput optimal for FD, FND, and IND systems, involves solving an NP-hard
problem for multi-category systems. For multi-category systems this is from the general class of packing integer
programs, for which constant factor approximation algorithms exist under different assumptions on the problem
parameters [37]. These assumptions do not generally hold for MaxWeight allocation under the CAC constraint.
Rather, we follow the same steps of LP relaxation and obtain a scheme that stabilizes any λ for λ + 1 ∈ CΓ, since
for large systems this is better than any arbitrarily close approximation scheme (as λ → ∞ as N → ∞).
A. Inflexible System
Below we present a characterization of the capacity region of inflexible systems in terms of the bipartite graph
G = (V, E), which captures the restriction of job-agent allocations.
Theorem 14. Any λ can be stabilized if λ ∈ CI, where
λ =
(cid:88)
,
CI =
is the capacity region of a single category system with an agent availability distribution Γl = Γ(cid:0)U l
where C(l)
Moreover, no λ /∈ ¯CI can be stabilized, i.e., for inflexible systems the capacity region CΓ = ¯CI.
l∈[L]
Γl
Γl
λ(l) : λ(l) ∈ C(l)
j = 0 for all (j, l) /∈ E
, λl
i : i ∈ [M l](cid:1).
Proof: See Appendix II.H.
This theorem has the following simple consequence. Consider separate pools of agents for each different category,
cf. [7], which has agent-availability distributions {Γl : l ∈ [L]}. Each such pool (category) of agents l can stabilize
job-arrival rates in C(l)
. Thus if the job arrival process of each job type j can be split in such a way that pool
j : j} ∈ C(l)
(category) l of agents sees an arrival rate λl
,
the system would be stable.
j > 0 only if (j, l) ∈ E, while ensuring that {λl
j, where λl
Γl
Γl
In a server farm where jobs can be placed on any of the server queues, the join-shortest-queue (JSQ) policy
stabilizes any stabilizable rate [20]. JSQ gives an arriving job to the server with the shortest queue and each server
serves jobs in FIFO order. For multi-category crowd systems, we can draw a parallel between servers and agent
pools. In addition we have constraints on job placement given by G and also have to do allocations of jobs among
the agents in the pool optimally (unlike JSQ we do not have FIFO/LIFO specified). Thus we have to adapt JSQ
appropriately based on our insights about optimal operation of crowd systems.
We propose a statistics-agnostic scheme, JLTT-MWTA (Alg. 4) that has two parts: JLTT (join least total task)
directs arrivals to appropriate pools of agents and MWTA allocates jobs in each pool separately. Letting Ql
j,s(t) be
the number of unallocated (j, s)-tasks in lth pool just after epoch t − 1, JLTT uses these quantities to direct jobs
to appropriate pools whereas MWTA uses them to allocate tasks within each pool.
The JLTT part is computationally light. The central controller only needs to know Q(t) and has to pick the
j,s) pools of agents to direct jobs (type j). To perform MWTA in each pool, a PTAS,
(cid:80)
minimally loaded (minl
FPTAS, or LP relaxation scheme can be used.
s Ql
Unlike JSQ, where service discipline in each server is fixed and the goal is to place the jobs optimally, we have
jobs with multi-dimensional service requirements from time-varying stochastic servers (agent-availability) and have
to place jobs as well as discipline the service in each random and time-varying virtual pool. Thus optimality of
JSQ cannot be claimed in our case. But as stated below, JLTT division followed by MWTA allocation is indeed
optimal.
Theorem 15. JLTT-MWTA stabilizes any λ ∈ CI.
Proof: See Appendix II.I.
11
Algorithm 4 JLTT-MWTA: Divide and Allocate
Input: A(t), U(t), Q(t)
Output: Job division and allocation
Create pool l with category l agents (∀l ∈ [L])
JLTT: Join Least Total Task
for each (j, s) do
Count number of unallocated (j, s)-tasks in pool l: Ql
Divide Aj(t)1(rj,s > 0) tasks equally among pools arg minl:(j,l)∈E
j,s(t)
end for
In each pool l run MWTA for single-category system
(cid:80)
s Ql
j,s(t)
An important aspect of JLTT-MWTA is that job allocations within each pool can happen independently of each
other. The central controller only has to make a decision on how to split the jobs based on the current system
state information. This allows a more distributed allocation along the lines of following hierarchical organizational
structure [38]. First, the central controller divides jobs for different agent-pools based on {Ql
j,s}. Then in each agent
pool, allocations are according to GreedyJob allocation, which works for both decomposable and non-decomposable
single-category systems. The distributed scheme that we propose here is an improvisation of the above JLTT scheme
followed by GreedyJob allocation in each pool. We call it Improvised JLTT and GreedyJob Allocation (Alg. 5).
Algorithm 5 Improvised JLTT and GreedyJob Allocation
Input: A(t), U(t), Q(t)
Output: Job division and allocation
Improvised JSQ for each job-type j and each skill s:
j(t)
n = 0
N l
j = Ql
While n < Aj(t)
Send 1 task to the category l∗ with lowest
index among arg minl:(j,l)∈E N l
j
Increase N l∗
End While
j and n each by 1
Allocations within each pool l:
Run GreedyJob allocation
j values.
j,s = Ql
First note that unlike JLTT-MWTA, here we only maintain number of unallocated jobs and do not maintain
number of unallocated tasks for each skill s. This is because as GreedyJob allocation is used as allocation scheme
in each pool, Ql
j,s(cid:48) for all s, s(cid:48).
This algorithm is also simple to implement. The central controller only needs to track the number of unallocated
j. For any arriving job of type j, the central controller sends
j are reset
j. This continues until the next epoch, when the N l
j(t)) from the previous epoch and set N l
j = Ql
j and updates N l
jobs (Ql
the job to the pool with minimum N l
to new Ql
Recall that Sec. I gave a simple example of a fully distributed scheme where jobs pick agents greedily (from
the set of feasible agents as per G) and showed it was not a good scheme. Improvisation of JLTT is proposed for
a better performance guarantee, while GreedyJob in each pool is proposed for implementability and freedom of
selection for customers. It is not hard to prove Improvised JLTT followed by MWTA is optimal for any arrival and
availability process satisfying the assumptions of Sec. II. Below we present performance guarantee for Improvised
JLTT and GreedyJob allocation.
To present performance guarantees of the distributed scheme we give an outer region CO for the system, along
the lines of the alternative characterization CI of the capacity region CΓ for inflexible systems.
Theorem 16. Inflexible crowd systems cannot be stabilized for λ /∈ CO, where CO =
and Cout
is the outer region for the single category system comprising the lth category (pool) of agents with
(cid:110)
λ : λ =(cid:80)
µl
l∈[L] λ(l) where λ(l) ∈ Cout
µl
(cid:111)
,
12
µl = E(cid:2)Ul(cid:3).
Proof: See Appendix II.J.
For job allocation in server farms, extant performance guarantees are mostly for symmetric load, i.e., symmetric
(almost) service and job arrival rates and regular graphs, cf. [30], [31]. Unlike server farms, symmetric load (in
terms of skill-hours) is not guaranteed in crowd systems by symmetric arrival rates and graphs. This is because
different types of jobs have different skill and hour requirements. The following guarantee for crowd systems is
for bounded asymmetry (sub-polynomial variation) in agent availability, complete graph, asymmetric job arrival
rates, and asymmetric job requirements (extendable to regular graphs with additional assumptions on symmetry of
job arrival rates and requirements). Note that because of the inflexibility constraint, a multi-category system with
a complete graph is not equivalent to a single-category system.
Theorem 17. Without loss of generality assume the same ordering of agent types in each category, i.e., M l =
i (t)}
M l(cid:48)
are i.i.d. across time and independent across types (jobs and agents), all these processes are Gaussian-dominated
i,s(cid:48)) are O (subpoly(N )) and G is
complete bipartite, then for any given α ∈ (0, 1], ∃Nα such that Improvised JLTT and Greedy-job stabilizes any
λ ∈ (1− α)CO :=
and the maximum number of unallocated jobs (across all types) is O(log N )
i for all l, l(cid:48), i. If the arrival processes {Aj(t)} and the agent availability processes {U l
i and maxl,s,s(cid:48) (cid:80)
i maxl,l(cid:48) µl
i,s − hl
i − µl(cid:48)
i = hl(cid:48)
= M/L and hl
(and/or Poisson-dominated),(cid:80)
1−α λ ∈ CO(cid:111)
(cid:1).
with probability 1 − o(cid:0) 1
(cid:110)
λ :
i µl
i(hl
1
N 2
Proof: See Appendix II.K.
The proofs of Thm. 11, 13, and 17 are all based on constructing queue-processes (different for the algorithms)
that stochastically dominate the number of unallocated jobs, and bounding the steady state distributions of these
processes using Loynes’ construction and moment generating function techniques.
B. Implementation of Decentralized Schemes on Crowd Platforms
We have described the allocation schemes at the level of system abstraction and discussed their performance.
These schemes can be easily implemented on crowd platforms as well.
GreedyAgent allocation is completely decentralized, only requiring agents to abide by a norm for picking partial
tasks, which can be enforced by randomized vigilance and penalizing norm violators in reputation. If the payments
are the same, as it generally is in single-category systems where all jobs require same quality, there is no incentive
for agents to deviate from the norm.
Any arbitrary contention method among agents will work for the algorithm, and hence the crowdsourcing platform
only needs to ensure that no two allocations are done simultaneously (as practiced in airlines booking). Multiple
allocations can also be allowed by the platform if they do not conflict. Here the platform has to ensure that an
agent can place requests only for an amount of tasks it can actually perform, given the constraints on available
hours. Also, only one agent can request for a task or a certain part of it. Once the agent has been declined, it can
place request(s) for task(s) of the same or lesser hours. This can either be enforced by appropriate modification
of the portals by keeping tracks of total hours of requests placed or by vigilance. GreedyJob can also be easily
implemented on a crowd platform. The platform has to ensure that jobs request agents and not the other way
around. One way to implement this is to allow jobs to place requests for agents while ensuring they do not request
more than the required service. Also, the platform has to ensure that skill-hour requests of no two jobs collide.
This again can be ensured by serializing the requests as above. Agents are expected to accept a requested task, as
there is no difference between tasks involving same skill since payments are the same. This can also be ensured
by linking agent rating to rate of task-request acceptance.
It is apparent that GreedyJob offers choice to customers and GreedyAgent offers choice to agents. By allowing
a customer (or an agent) to decline an approaching agent (or a customer) request and to explore more options,
only one option at a time, the platform can provide freedom of choice to agents and customers under both schemes
while operating at capacity.
In case of multi-category systems, the platform only needs to direct arriving jobs to the appropriate pool of
agents, based on current backlog; the rest of the allocation happens as per GreedyJob. Directing a job to a category
of agents can be implemented in a crowd platform by making the job visible only to freelancers of that category and
13
vice versa (similar to filtering done by search engines and online social networks) or through explicit hierarchical
organization into pools [38].
VI. BEYOND THE CAPACITY REGION
We have now characterized the capacity (and outer) regions of different classes of crowd systems, shown the
existence of computationally feasible centralized schemes that achieve these regions, and presented simple distributed
schemes with minimal centralized intervention and good performance guarantees for any arrival rate within the
capacity region. In crowd systems, however, arrival rates may not be within the capacity region, since the platform
may have little or no control on resource (freelancer) planning, unlike traditional communication networks or cloud
computing systems. Hence, an important aspect of crowd systems is to turn down job requests. Indeed, deciding to
decline a job must be done as soon as the job arrives because dropping a job after first being accepted adversely
affects the reputation of the crowd platform.
Here, we propose a centralized scheme for a crowd system to decline jobs on arrival in a way that is fair
across all job types. Our scheme is statistics-agnostic and works even for independent but non-stationary arrival
and availability processes.
We solve the following problem: given an arrival rate λ, design a statistics-agnostic policy to accept (1 − β)λ
jobs on average and allocate them appropriately such that β is the minimum for which the crowd system is stable.
Note that if λ ∈ CΓ, the minimum β is 0, else, it is strictly positive. We want to design a statistics-agnostic policy
without the knowledge of λ and CΓ. As a benchmark, we consider the following problem for > 0, when λ and
CΓ are known.
min β ∈ [0, 1]
s.t. (1 − β)λ + 1 ∈ CΓ.
(6)
Given β∗, optimum of (6), (1 − β∗)λ is within of the optimal rate of accepted jobs for which the system is
stabilizable.
As we want a scheme that is agnostic of λ and CΓ, we propose the following simple scheme, for ν > 0 and
j,s(t) is the number of unallocated accepted (j, s)-tasks ((j, s)-tasks directed to category l) in the system.
Ql
β(t) = arg min
β∈[0,1]
(cid:40)
β(cid:80)
j Aj(t) − νβ(cid:80)
β(cid:80)
j Aj(t) − νβ(cid:80)
(cid:40)
For accepted jobs run
j,s:rj,s>0
j,s:rj,s>0 minl Ql
Qj,s(t)Aj(t)
(I)
j,s(t)Aj(t) (II)
Job is accepted w.p. 1 − β(t), accepted jobs A(t) (I & II)
MWTA
JLtT-MWTA (II)
(I)
(7)
Steps marked by I (II) are applicable for FD, FND and IND (ID and IND) systems.
Theorem 18. Crowd system with jobs accepted and allocated according to (7) is stable and (cid:80)
(cid:80)T
Aj(t)] can be made arbitrarily small for an appropriately chosen ν, for all sufficiently large T .
E[(cid:80)
j λj(1 − β∗) −
1
T
j
t=1
Proof: See Appendix II.L.
(cid:80)T
This theorem demonstrates that by following the job acceptance and allocation method (7), the crowd system
can be stabilized while ensuring the average number of accepted jobs per allocation epoch is arbitrarily close to
the optimal number of accepted jobs per allocation period. Note that as all jobs (across all types) are accepted with
the same probability, the above result also implies that (1 − β∗)λj − 1
E[ Aj(t)] is small. It can be shown
that the above scheme works for time-varying systems (E[A(t)] = λ(t) and U ∼ Γt()) as well guaranteeing small
, where β∗(t) is the solution of (6) for λ = λ(t) and CΓ = CΓt.
limT→∞ 1
T
j
sectionConclusion
Human information processing, structured through freelance markets, is an emerging structure for performing
informational work by capturing the cognitive energy of the crowd. It is important to understand the fundamental
limits and optimal designs for such systems.
λj(t)(1 − β∗(t)) − E[ Aj(t)]
(cid:80)T
(cid:80)
(cid:16)
(cid:17)
t=1
t=1
T
14
In this work we provide a characterization of the work capacity of crowd systems and present two statistic-agnostic
job allocation schemes MWTA (flexible jobs) and JLTT-MWTA (inflexible jobs) to achieve limits. To ensure low
computational load on the crowd platform provider and freedom of choice for job requesters, we present simple
decentralized schemes, GreedyAgent, GreedyJob, and Improvised JLTT-GreedyJob that (almost) achieve capacity
with certain performance guarantees. These decentralized schemes are easy to implement on crowd platforms,
require minimal centralized control, and offer freedom of self-selection to customers: all desirable qualities for
any crowd platform. Due to quick evolution and unpredictability of freelancer resources, crowd systems may often
operate outside capacity, which inevitably results in huge backlogs. Backlogs hurt the reputation of the platform,
and so we also propose a scheme that judiciously accepts or rejects jobs based on the system load. This scheme
is fair in accepting jobs across all types and accepts the maximum number of jobs under which the system can be
stable.
REFERENCES
[1] A. Chatterjee, L. R. Varshney, and S. Vishwananth, “Work capacity of freelance markets: Fundamental limits and decentralized schemes,”
in Proc. 2015 IEEE INFOCOM, Apr. 2015, pp. 1769–1777.
[2] D. Tapscott and A. D. Williams, Wikinomics: How Mass Collaboration Changes Everything, expanded ed. New York: Portfolio
Penguin, 2006.
Press, 2006.
Institute, 2011.
[3] Y. Benkler, The Wealth of Networks: How Social Production Transforms Markets and Freedom. New Haven, CT: Yale University
[4] D. Bollier, The Future of Work: What It Means for Individuals, Businesses, Markets and Governments. Washington, DC: The Aspen
[5] D. F. Bacon, E. Bokelberg, Y. Chen, I. A. Kash, D. C. Parkes, M. Rao, and M. Sridharan, “Software economies,” in Proc. FSE/SDP
Workshop Future Softw. Eng. Research (FoSER 2010), Nov. 2010, pp. 7–12.
[6] M. Vukovic and O. Stewart, “Collective intelligence applications in IT services business,” in Proc. IEEE 9th Int. Conf. Services Comput.
(SCC), Jun. 2012, pp. 486–493.
[7] L. R. Varshney, S. Agarwal, Y.-M. Chee, R. R. Sindhgatta, D. V. Oppenheim, J. Lee, and K. Ratakonda, “Cognitive coordination of
global service delivery,” arXiv:1406.0215v1 [cs.OH]., Jun. 2014.
[8] C. Shirky, Cognitive Surplus: Creativity and Generosity in a Connected Age. Penguin, 2010.
[9] T. W. Malone, R. Laubacher, and C. Dellarocas, “The collective intelligence genome,” MIT Sloan Manage. Rev., vol. 51, no. 3, pp.
[10] K. J. Boudreau and K. R. Lakhani, “Using the crowd as an innovation partner,” Harvard Bus. Rev., vol. 91, no. 4, pp. 60–69, Apr.
[11] D. DiPalantino and M. Vojnovi´c, “Crowdsourcing and all-pay auctions,” in Proc. 10th ACM Conf. Electron. Commer. (EC’09), Jul.
[12] N. Archak and A. Sundararajan, “Optimal design of crowdsourcing contests,” in Proc. Int. Conf. Inf. Syst. (ICIS), 2009, p. 200.
[13] K. J. Boudreau, N. Lacetera, and K. R. Lakhani, “Incentives and problem uncertainty in innovation contests: An empirical analysis,”
Manage. Sci., vol. 57, no. 5, pp. 843–863, May 2011.
[14] G. V. Ranade and L. R. Varshney, “To crowdsource or not to crowdsource?” in Proc. AAAI Workshop Human Comput. (HCOMP’12),
[15] D. R. Karger, S. Oh, and D. Shah, “Budget-optimal task allocation for reliable crowdsourcing systems,” Oper. Res., vol. 62, no. 1, pp.
[16] C.-J. Ho and J. W. Vaughan, “Online task assignment in crowdsourcing markets,” in Proc. 26th AAAI Conf. Artif. Intell., Jul. 2012,
[17] D. V. Oppenheim, L. R. Varshney, and Y.-M. Chee, “Work as a service,” in Advanced Web Services, A. Bouguettaya, Q. Z. Sheng, and
F. Daniel, Eds. Springer, 2014, pp. 409–430.
[18] J. Kleinberg and ´E. Tardos, Algorithm Design. Addison-Wesley, 2005.
[19] M. J. Neely, Stochastic Network Optimization with Application to Communication and Queueing Systems. Morgan & Claypool
[20] R. Srikant and L. Ying, Communication Networks: An Optimization, Control and Stochastic Networks Perspective.
Cambridge
21–31, Spring 2010.
2013.
2009, pp. 119–128.
Jul. 2012, pp. 150–156.
1–24, Jan.-Feb. 2014.
pp. 45–51.
Publishers, 2010.
University Press, 2014.
[21] M. L. Pinedo, Scheduling: Theory, Algorithms, and Systems. Springer, 2012.
[22] L. Tassiulas and A. Ephremides, “Stability properties of constrained queueing systems and scheduling policies for maximum throughput
in multihop radio networks,” IEEE Trans. Autom. Control, vol. 37, no. 12, pp. 1936–1948, Dec. 1992.
[23] ——, “Dynamic server allocation to parallel queues with randomly varying connectivity,” IEEE Trans. Inf. Theory, vol. 39, no. 2, pp.
[24] A. Eryilmaz, R. Srikant, and J. R. Perkins, “Stable scheduling policies for fading wireless channels,” IEEE/ACM Trans. Netw., vol. 13,
[25] M. J. Neely and E. Modiano, “Capacity and delay tradeoffs for ad hoc mobile networks,” IEEE Trans. Inf. Theory, vol. 51, no. 6, pp.
[26] S. T. Maguluri, R. Srikant, and L. Ying, “Stochastic models of load balancing and scheduling in cloud computing clusters,” in Proc.
2012 IEEE INFOCOM, Mar. 2012, pp. 702–710.
466–478, Mar. 1993.
no. 2, pp. 411–424, Apr. 2005.
1917–1937, Jun. 2005.
15
[27] I. Menache, A. Ozdaglar, R. Srikant, and D. Acemoglu, “Dynamic online-advertising auctions as stochastic scheduling,” in Proc.
Workshop Econ. Netw. Syst. Comput. (NetEcon ’09), Jul. 2009.
[28] B. Tan and R. Srikant, “Online advertisement, optimization and stochastic networks,” IEEE Trans. Autom. Control, vol. 57, no. 11, pp.
2854–2868, Nov. 2012.
[29] S. Chen, P. Sinha, and N. B. Shroff, “Scheduling heterogeneous delay tolerant tasks in smart grid with renewable energy,” in Proc.
51st IEEE Conf. Decision Control, Dec. 2012, pp. 1130–1135.
[30] J. N. Tsitsiklis and K. Xu, “Queueing system topologies with limited flexibility,” in Proc. ACM SIGMETRICS Int. Conf. Meas. Model.
Comput. Syst., Dec. 2013, pp. 167–178.
[31] M. Bramson, Y. Lu, and B. Prabhakar, “Randomized load balancing with general service time distributions,” in Proc. ACM SIGMETRICS
Int. Conf. Meas. Model. Comput. Syst., Jun. 2010, pp. 275–286.
[32] V. Shah and G. de Veciana, “Performance evaluation and asymptotics for content delivery networks,” in Proc. 2014 IEEE INFOCOM,
[33] D. Wang, “Computing with unreliable resources: Design, analysis and algorithms,” Ph.D. thesis, Massachusetts Institute of Technology,
Apr.-May 2014, pp. 2607–2615.
Cambridge, MA, Jun. 2014.
[34] H. Kellerer, U. Pferschy, and D. Pisinger, Knapsack Problems. Springer, 2004.
[35] S. Bubeck and N. Cesa-Bianchi, “Regret analysis of stochastic and nonstochastic multi-armed bandit problems,” Found. Trends Mach.
[36] R. Keshavan, A. Montanari, and S. Oh, “Matrix completion from noisy entries,” in Advances in Neural Information Processing Systems
Learn., vol. 5, no. 1, pp. 1–122, Dec. 2012.
22. Cambridge, MA: MIT Press, 2009, pp. 952–960.
[37] C. Chekuri and S. Khanna, “On multidimensional packing problems,” SIAM J. Comput., vol. 33, no. 4, pp. 837–851, 2004.
[38] S. Agarwal, S. Kenkre, V. Pandit, and B. Sengupta, “Studying the evolution of skill profiles in distributed, specialization driven service
delivery systems through work orchestration,” in Proc. SRII Global Conf. 2011, Apr. 2011, pp. 201–213.
APPENDIX I
A. Computation for Centralized Allocation
For a single category system (L = 1), note that z1
(cid:88)
aj,srj,s ≤(cid:88)
j
i
j,s = aj,s ∈ Z+ and hence the feasibility condition (1) becomes
uihi,s for all s ∈ [S], aj,s ∈ Z+,
with condition (2) additionally requiring aj,s = aj,s(cid:48) for all j, s, s(cid:48). Thus, C(u) is the set of {aj,s} satisfying the
above conditions for respective classes of jobs (as well as systems) and C is the weighted (by Γ(u)) sum of convex
hulls of C(u)s.
µ has a simple characterization as well. As for any j ∈ [N ], (j, 1) ∈ E, and N (J) = 1 for all J ⊂ [N ],
Cout
l∈N (J)
i∈[M ] µihi,s. Thus it is sufficient to satisfy the inequality for J = [N ], and hence,
s =(cid:80)
i∈[M l] µl
(cid:80)
(cid:80)
ihl,i
λjrj ≤ (cid:88)
µihi
i∈[M ]
.
Coutµ =
j∈[N ]
(cid:88)
λ :
(cid:88)
aj,srj,s ≤(cid:88)
j,s
Qj,s∆j,s
arg max
∆j,s:j,s
s.t. (cid:88)
The MaxWeight computation in MWTA for single-category decomposable systems turns out to be the following
integer linear program (ILP), which is related to knapsack problems.
uihi,s∀s ∈ [S],
(8)
j
i
poly(cid:0) 1
(cid:1) computations. Moreover, there exists faster greedy algorithms that achieve 1
This problem is an integer program, hence it is not clear whether this problem can be solved efficiently at all
instants. In fact, it is a so-called unbounded knapsack problem for a given u and Q. This problem is known to
be NP-hard [34]. There is a pseudo-polynomial algorithm based on dynamic programming which solves it exactly,
but the runtime may depend on Qj,s. This dynamic programming-based algorithm can be converted into a fully
polynomial time approximation schemes (FPTAS) which achieves any (1 − ) approximation of the problem in
2 approximation and can be
converted into a polynomial time approximation scheme (PTAS) that achieves (1 − ) approximation in poly(n
)
computations. Thus we can conclude that though the MWTA algorithm is computationally hard for single-category
decomposable system, there exist efficient approximation schemes. It is not hard to show (Prop. 19 below) that an
algorithm that gives (1 − ) approximation of the optimization problem in MWTA can stabilize any λ for which
(1−) ∈ CD
Γ .
λ
1
16
given by,
For single-category non-decomposable systems, feasibility condition (2) of an allocation of aj jobs of type j is
(cid:88)
ajrj,s ≤(cid:88)
uihi,s∀s ∈ [S], aj ∈ Z+.
uihi,s and aj,s = aj∀s
.
i
Hence the stabilizable region changes accordingly to
j
CΓ =
Γ(u)λ(u) : λ(u) ∈ CN D(u)
(cid:40)(cid:88)
u
CN D(u) = conv
(cid:88)
j
,
(cid:41)
ajrj,s ≤(cid:88)
(cid:33)
i
aj,s ∈ Z+ :
(cid:32)(cid:88)
(cid:88)
∆j,srj,s ≤(cid:88)
s
j
Qj,s∆j,s
arg max
∆j,s:j
s.t. (cid:88)
where conv{·} is the convex hull.
Hence, in this case, the MWTA allocation needs to solve
uihi,s, ∆j,s = ∆j,s(cid:48), for all s, s(cid:48) ∈ [S],
(9)
j
i
and then divide the allocated jobs arbitrarily among agents while meeting their per skill time-availability, as there
is only one category of agents.
This problem is also a knapsack-like integer program, with an additional constraint that the number of (j, s)-items
has to be the same as the number of (j, s(cid:48))-items for all j, s and s(cid:48). Such a problem is called a multi-dimensional
knapsack problem. This problem is also NP-hard. Moreover, provably there cannot exist a fully polynomial time
approximation scheme for this problem [34].
For a multi-dimensional knapsack problem, there exists an approximation scheme that achieves an approximation
factor equal to the dimension d [34]. This approximation scheme can be converted into a PTAS with complexity
O (N κ) and an approximation factor of 1 − , where κ is strictly increasing with dimension and 1
. In the case of
our setting, the number of skills S (dimension) is large and may scale with N, hence this scheme is not suitable.
Though the total number of skills S can scale with N, in most cases the number of skill-parts that a type j
job has is a constant, i.e., rj has most coordinates as 0. Thus it is apparent from the objective function that the
optimal choice of ∆j,s is 0 for the corresponding coordinates s. This allows us to rewrite the optimization problem
as another multi-dimensional knapsack problem with constant dimensions given by maxj {s : rj,s > 0}.
For this problem we can use the PTAS to obtain arbitrarily close approximation and hence can stabilize a
rate-region arbitrarily close to CN D. But the complexity of this algorithm is very high as complexity scales
super-exponentially (N k) with the approximation factor (unlike decomposable systems where we have an FPTAS,
polynomial in 1
).
Note that our goal is not to solve (9) optimally, but to have a fast allocation scheme that can stabilize a large
fraction of CN D. In this regard we can take a different approach that exploits basic characteristics of a crowd
system. Since N and λj(N ) for j ∈ [N ] are large in most crowd systems, if an allocation scheme stabilizes any
rate λ for λ + c1 ∈ C, then it stabilizes
such a scheme would stabilize almost all of C. Motivated by this, we propose the following allocation scheme,
which is a modification of (9).
(cid:17)C. Note that as λi(N ) scales with N, this implies that
(cid:32)(cid:88)
(cid:88)
xjrj,s ≤(cid:88)
{xj, j} = arg max
xj∈R:j
j
uihi,s, for all s,
s.t. (cid:88)
1 − maxi
(cid:33)
(cid:16)
(10)
Qj,s
xj
1
λi(N )
s
and allocate ∆j = (cid:98)xj(cid:99) jobs of type j to the agents, splitting arbitrarily while meeting time-availability constraints.
j
i
s Qj,s) xj ≥ (cid:80)
j ((cid:80)
Note that since in (10), the variables are relaxed to R from Z, (cid:80)
j ((cid:80)
j ((cid:80)
s Qj,s) ∆j ≥(cid:80)
∆j = (cid:98)xj(cid:99) ≥ xj − 1, hence(cid:80)
j ((cid:80)
Qj,s(t) ∆j,s(t) ≤(cid:88)
s Qj,s) zj. Again,
The following proposition guarantees that a proposed LP-relaxation scheme stabilizes any λ with λ + 1 ∈ CN D.
Proposition 19. Let P be an allocation scheme that at epoch t does an allocation {∆(t)} instead of { ∆(t)} of
the MWTA allocation scheme, which satisfies
s Qj,s) (zj − 1).
Qj,s(t)∆j,s(t) +
(cid:88)
(cid:88)
Qj,s(t)δ,
17
or,
j,s
(cid:88)
j,s
j,s
Qj,s(t) ∆j,s(t) ≤(cid:88)
j,s
(1 − )
j,s
Qj,s(t)∆j,s(t),
stabilizes any rate λ ∈ C if λ + δ1 ∈ C or
1
1− λ ∈ C respectively.
Proof: This proof follows the same steps as the proof of Thm. 8. We first prove the result for an allocation
Qj,s(t)∆j,s(t) +
Qj,s(t)δ.
(cid:88)
j,s
with
(cid:88)
j,s
Qj,s(t) ∆j,s(t) ≤(cid:88)
(cid:88)
− E
Qj,s(t)
j,s
as follows:
E [L (Q(t + 1)) − L (Q(t))Q(t)]
(cid:88)
≤ E
Qj,s(t)Aj(t)Q(t)
j,s
j,s
We can bound the last term as above, because ( ∆j,s − δ)2 ≤ ∆2
(cid:16) ∆j,s(t) − δ
(cid:88)
A2
(cid:88)
+ E
j (t) + ( ∆j,s − δ)2(cid:17)Q(t)
(cid:16)
(cid:17)Q(t)
(cid:88)
− E
Qj,s(t) ∆j,s(t)Q(t)
j,s + δ2. Thus,
j,s
(cid:88)
k γkdk = ν(u) for γk > 0,(cid:80)
j,s Qj,s(t)λj ≤ E(cid:104)(cid:80)
j,s
E [L (Q(t + 1)) − L (Q(t))Q(t)] ≤ B + E
Qj,s(t)(Aj(t) + δ)Q(t)
Now note that if λ is such that λ + δ1 ∈ C, then we can write it in terms of convex combinations of d ∈ C(u)
j,s
j,s
and follow the same steps as in the proof of Thm. 8.
Similarly for the other case of constant factor approximation we have:
− E
Qj,s(t) ∆j,s(t)(1 − )Q(t)
(cid:88)
j,s
E [L (Q(t + 1)) − L (Q(t))Q(t)] ≤ B + E
Qj,s(t)Aj(t)Q(t)
j,s ≤ ∆2
j,s.
as (1 − )2 ∆2
(1− )ν(u)Γ(u) and(cid:80)
j,s, (cid:80)
(cid:80)
If λ ∈ (1−)C, then by definition of C and (1−)C, there exist ν(u) ∈ C(u) and {dk(u) ∈ C(u)} such that, λ ≤
j,s Qj,s(t) ∆j,s(t) ≤
. Then following the proof of Thm. 8, the result
(cid:105)
j,s Qj,s(t) ∆j,s(t)Q(t)
k γk = 1. As for any dk for a given u,(cid:80)
j,s Qj,s(t)dk
follows.
In this appendix, we present proofs of the main results in Secs. III–V. As mentioned earlier, most of these results
extend to systems with stationary and ergodic arrival and availability processes, but here we only present results
for i.i.d. processes.
APPENDIX II
18
A. Proof of Theorem 5
(cid:0)λ(cid:48)(u) ∈ C(u) : u ∈ ZM
+
Here we only prove the converse part, i.e., λ /∈ C cannot be stabilized by any policy. For the direct part, it is
sufficient to prove there exists a scheme that stabilizes any λ ∈ C, and so the proof of Thm. 8 below is sufficient.
First we prove that C is a convex subset of RN
+ . If λ, λ(cid:48) ∈ C, then there exist (cid:0)λ(u) ∈ C(u) : u ∈ ZM
(cid:1) and
+
(cid:1) such that(cid:88)
Γ(u)λ(u) = λ, (cid:88)
Γ(u)λ(cid:48)(u) = λ(cid:48).
Thus for any γ ∈ [0, 1],
u
γλ + (1 − γ)λ(cid:48) =
(cid:88)
u
Γ(u)(γλ(u) + (1 − γ)λ(cid:48)(u).
Note that C(u) is convex since it is the convex hull of C(u); hence γλ(u) + (1 − γ)λ(cid:48)(u) ∈ C(u), which in turn
implies γλ + (1 − γ)λ(cid:48) ∈ C. This proves convexity of C.
Thus ¯C is a closed convex set. Hence for any λO /∈ ¯C, there exists a hyperplane hT x = c that separates ¯C and
λO, i.e., there exists an > 0 for any λ ∈ ¯C such that hT λO ≥ hT λ + .
u
Hence under any policy:
E(cid:2)hT Q(t + 1)(cid:3) = E(cid:2)hT (Q(t) + A(t) − D(t))(cid:3)
= E(cid:2)hT Q(t) + A(t) − ∆(t)+(cid:3)
where ∆(t) is the number of possible departure under the scheme if there were infinite number of jobs of each
type, and ·+ is shorthand for max(·, 0). As x+ is a convex function of x, hT Q(t) + A(t) − ∆(t)+ is a convex
function of Q(t), A(t), and ∆(t). Thus by Jensen’s inequality:
E(cid:2)hT Q(t) + A(t) − ∆(t)+(cid:3) ≥ hT E [Q(t)] + E [A(t)] − E [∆(t)]+
Note that any λ is a Γ(u)-combination of some {λ(u) ∈ C(u)} and any λ(u) is some convex combination of
elements of C(u). Also, from the allocation constraints it is apparent that if a ∈ C(u) then also a(cid:48) ∈ C(u) if
a(cid:48) ≤ a. These two imply that for any λ ∈ ¯C, if there exists a λ(cid:48) ≤ λ (component-wise) and λ(cid:48) ≥ 0, then λ(cid:48) ∈ ¯C.
That is C is coordinate convex. This in turn implies that for any λO /∈ ¯C there exists an h (cid:54)= 0 ∈ RN
+ such that
hyperplane separation holds for this h. Thus for h ≥ 0:
E(cid:2)hT Q(t + 1)(cid:3) ≥ hT E [Q(t)] + E [A(t)] − E [∆(t)]+
hjE [Qj(t)] + hjE [Aj(t)] − hjE [∆j(t)]+
(hjE [Qj(t)] + hjE [Aj(t)] − hjE [∆j(t)])
=
(cid:88)
≥(cid:88)
≥ E(cid:2)hT Q(t)(cid:3) + hT E(cid:2)λO(cid:3) − sup
≥ E(cid:2)hT Q(t)(cid:3) +
λ∈ ¯C
j
j
hT λ
Thus we have E(cid:2)hT Q(t + 1)(cid:3) → ∞. As h ≥ 0, this implies there exists j such that E[Qj(t)] → ∞ as t → ∞.
Hence, the system is not stable.
B. Proof of Theorem 6
Consider the dynamics of Qj,s(t), the unallocated (j, s) tasks at the end of epoch t.
Qj,s(t + 1) = Qj,s(t) + Aj(t) − Dj,s(t)+
≥ Qj,s(t) + Aj(t) − Dj,s(t)
≥ t(cid:88)
k=0
(Aj(t) − Dj,s(t)) .
As rj,s ≥ 0,
19
E [Aj(t)rj,s − rj,sDj,s(t)]
(λjrj,s − rj,sE [Dj,s(t)])
E [rj,sQj,s(t + 1)] ≥ t(cid:88)
t(cid:88)
k=0
=
k=0
Consider any set J ⊂ [N ], then at any epoch t, to schedule a certain number of tasks of type (j, s), the system
needs that much available usable skill-hours. This follows from conditions (1) and (2) and can be written as:
This in turn implies
hl
i,sU l
i,s.
(cid:88)
(cid:88)
i∈[M l]
hl
i,sµl
i,s.
l∈N (J)
i∈[M l]
j∈J
j∈J
l∈N (J)
(cid:88)
rj,sDj,s(t) ≤ (cid:88)
rj,sE [Dj,s(t)] ≤ (cid:88)
(cid:88)
(cid:88)
≥ t(cid:88)
(cid:88)
≥ t(cid:88)
(cid:105)
j∈J
k=0
(cid:88)
Hence,
E
j∈J
rj,sQj,s(t + 1)
λjrj,s −(cid:88)
λjrj,s − (cid:88)
For any λ /∈ ¯Cout, by definition there exists a J ⊂ [N ] such that (cid:80)
Thus in that case, lim supt→∞ E(cid:104)(cid:80)
j∈J
j∈J
k=0
rj,sE [Dj,s(t)]
(cid:88)
hl
i,sµl
i,s
l∈N (J)
i∈[M l]
j∈J λjrj,s −(cid:80)
(cid:80)
i,s > 0.
= ∞. Note that since J is finite and so is maxj rj,s, there
exists a j ∈ J such that lim supt→∞ E [Qj,s(t + 1)] = ∞. This shows the system is not stable for λ /∈ ¯Cout and
proves CΓ ⊂ Cout.
j∈J rj,sQj,s(t + 1)
i∈[M l] hl
l∈N (J)
i,sµl
(11)
C. Proof of Proposition 7
We consider FD, FND, and IND cases separately.
FD: The MaxWeight part chooses zl
j,s to be integral which implies that integral number of tasks can be allocated
if done appropriately. As hours are allocated from tasks in order, a task later in the order only gets allocated
(partially or fully) after the tasks before it are fully allocated. This leads to no partially-allocated tasks.
FND) and hence if an s-task of a job is chosen then also s(cid:48) is chosen for rj,s, rj,s(cid:48) > 0.
j,s such that aj,s = aj,s(cid:48) (as it satisfies
FND: Same ordering is used for all (j, s) tasks and MaxWeight chooses zl
IND: Same ordering is used for all (j, s) tasks, allocations to different categories are in same order (l = 1 to L)
j,s(cid:48) for all l, l(cid:48) (as it satisfies IND), hence if a task of a job is
j,s such that zl
and MaxWeight chooses zl
allocated to category l then so are the other tasks.
j,s = zl
D. Proof of Theorem 8
Note that MaxWeight chooses an allocation { ∆j,s(t)}. But, the maximum number of (j, s)-tasks that can be served
is Qj,s(t)+Aj(t). By Prop. 7, Task Allocation does a feasible allocation for FD, FND, and IND systems. Also, note
that in the Task Allocation algorithm, the number of allocated (j, s)-tasks is Dj,s = min
.
(cid:16) ∆j,s(t), Qj,s(t) + Aj(t)
(cid:17)
20
j,s
(cid:16)
j,s
j,s(t)
j,s(t)
= E
≤ E
≤ E
= E
Q(t)
Q(t)
j,s
j,s
j,s
j,s
j,s Q2
j,s.
(cid:19)
(cid:19)
j,s(t + 1) − Q2
Qj,s(t)Aj(t)Q(t)
We bound the last term first.
Qj,s(t) ∆j,s(t)Q(t)
Qj,s(t) + Aj(t) − Dj,s
Qj,s(t) + Aj(t) − ∆j,s
(cid:0)Q2
(cid:18)(cid:16)
(cid:18)(cid:16)
E [L (Q(t + 1)) − L (Q(t))Q(t)]
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:16)
Consider the usual Lyapunov function L (Q) =(cid:80)
j,s(t)(cid:1)Q(t)
(cid:17)2 − Q2
(cid:17)2 − Q2
(cid:88)
− E
=
(cid:17)Q(t)
(cid:88)
≤(cid:88)
≤(cid:88)
≤(cid:88)
+ E
(cid:88)
(cid:17)2
(cid:88)
2
2
(cid:88)
(cid:32)(cid:88)
(cid:33)
(cid:17)2(cid:21) M (M + 1)
(cid:20)(cid:16)
To bound the first term, note that if λ + 1 ∈ C, then there exist {ν(u) ∈ C(u)} such that λj ≤(cid:80)
for all j ∈ [J]. Again note that as C(u) is the convex hull of C(u), ν(u) =(cid:80)
and γk ≥ 0 with(cid:80)
(cid:88)
This is a constant B < ∞ independent of Q, as E[A2
j ] +
(cid:16)
j ] and E(cid:104)(cid:0)U l
j ] +
maxl,i,s hl
i,s
min(rj,s > 0)
Qj,s(t)Aj(t)Q(t)
(cid:88)
are finite for all j, l, i.
k γk = 1. So,
A2
j (t) + ∆2
j,s
A2
j (t) + ∆2
j,s
E[A2
j ] + E
min(rj,s > 0)
min(rj,s > 0)
E[A2
j ] +
(cid:1)2(cid:105)
hl
i,s
U l
i
E
max
l,i
rj,s ∆j,s
rj,s ∆j,s
Qj,s(t)λj
E
(cid:17)2
j,s
j,s
j,s
j,s
E[A2
E[A2
1
1
E
j,s
i,l
s
(cid:16)
E
j,s
U l
i
2
j,s
i
(cid:17)Q(t)
E
j,s
(12)
(13)
u Γ(u)ν(u)−
k γkdk(u) for some {dk(u) ∈ C(u)}
j,s
j,s
=
(cid:88)
≤(cid:88)
≤(cid:88)
(cid:88)
≤(cid:88)
(cid:88)
= E
=
j,s
u
u
Qj,s(t)νj −
Qj,s
(cid:88)
(cid:88)
u
j,s
Qj,s(t)
Γ(u)
Γ(u)
Qj,s(t)
j,s
(cid:88)
(cid:88)
(cid:88)
(cid:88)
k
k
j,s
Γ(u) max
d(u)∈C(u)
(cid:88)
(cid:88)
j,s
j,s
Qj,s
Qj,s
(cid:88)
j,s
Qj,s
γkdk
γkdk
j (u) −
j (u) −
Qj,s(t)dj,s(u) −
−
(cid:88)
Qj,s.
Qj,s(t) ∆j,s(t)Q(t)
j,s
j,s
21
(cid:88)
j,s
Qj,s.
Thus, we have a bound on the Lyapunov drift,
E [L (Q(t + 1)) − L (Q(t))Q(t)] ≤ B −
Hence,
E [L (Q(T )) − L (Q(0))] ≤ BT −
T−1(cid:88)
(cid:88)
As L(Q(0)) < ∞ and L(Q) ≥ 0 for all Q, we have that for all T ,
T−1(cid:88)
(cid:88)
1
T
E[Qj,s(t)] ≤ B
+
L(0)
T
< ∞
This in turn implies lim supt→∞(cid:80)
Again note that Qj(t) ≤(cid:80)
that for all (j, s) with rj,s > 0, lim supt→∞ E[Qj,s(t)] < ∞.
lim supt→∞ E[Qj(t)] < ∞ for all j ∈ [N ].
t=0
j,s
j,s
wj,sE[Qj,s(t)].
t=0
j,s
E[Qj,s(t)] < ∞, otherwise the time-average cannot be finite. This implies
s Qj,s, as there can be unallocated jobs with more than one part unallocated. Hence,
E. Proof of Theorem 11
In the GreedyAgent algorithm, as each agent with available skill-hours greedily chooses to serve a task, no (j, s)
task of size r can remain unallocated if there is an agent (or agents) with s skill-hour (total) of at least r. Since at
each allocation epoch a task should either be allocated totally or not at all (i.e., x < r hours cannot be allocated),
it may happen that some agent hours are wasted, as that does not meet the task allocation requirement.
Note that since any job requirement is less than ¯r = maxj,s rj,s, at most ¯r agent-skill-hours can be wasted.
Let Hs(t) be the process of unallocated job-hours for skill s after the allocation at epoch t. Then for all t,
i
(cid:88)
j
Ui(t)hi,s + ¯r
.
Hs(t + 1) ≤
For θ ∈ R, then,
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Hs(t) +
This implies that process Gs(t) given by Gs(t + 1) =
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)+
Aj(t)rj,s −(cid:88)
(cid:12)(cid:12)(cid:12)Gs(t) +(cid:80)
j Aj(t)rj,s −(cid:80)
Gs(t) has dynamics of a queue with arrival process Xs(t) =(cid:80)
i Ui(t)hi,s. Let Aj(θ) = E(cid:2)eθAj(t)(cid:3) and Ui(θ) = E(cid:2)eθUi(t)(cid:3) for j ∈ [N ] and i ∈ [M ].
(cid:80)
e−θYs(t)(cid:105)
eθXs(t)(cid:105) E(cid:104)
= E(cid:104)
j Aj(t)rj,s+¯r(cid:105) E(cid:104)
= E(cid:104)
eθ(cid:80)
e−θ(cid:80)
E(cid:104)
(cid:105)(cid:89)
= eθ¯r(cid:89)
(cid:89)
= eθ¯r(cid:89)
θ¯r +
eθ(Xs(t)−Ys(t))(cid:105)
log Aj(θrj,s) +
E [Ui(t)hi,s]
Ui(−θhi,s)
Aj(θrj,s)
E(cid:104)
i
(cid:88)
(cid:88)
eθAj(t)rj,s
= exp
(cid:105)
i Ui(t)hi,s
j
j
i
.
log Ui(−θhi,s)
(cid:12)(cid:12)(cid:12)+
bounds Hs(t).
j Aj(t)rj,s + ¯r and departure process Ys(t) =
i Ui(t)hi,s + ¯r
(14)
First consider the Gaussian-dominated case. Since the process variance is no more than the mean and the moment
generating function of the variance is upper-bounded by that of a zero-mean Gaussian:
j
i
log Aj(θrj,s) ≤ λjθrj,s + λj
log Ui(−θhi,s) ≤ −µjθhi,s + µj
(θrj,s)2
2
(θhi,s)2
2
.
22
Note that for any two functions k1x2 and k2x, limx→0 k2x/k1x2 = ∞, and hence for any ∈ (0, 1) there exists
i,s > 0, for all i, j, s
x∗ > 0 such that for all x < x∗, k1x2/k2x < . Hence for any ∈ (0, 1), there exist θ∗
such that for all θ < θ∗ = mini,j,s(θ∗
j,s, θ∗
j,s, θ∗
i,s,
log Aj(θrj,s) ≤ λjθ∗rj,s(1 + )
log Ui(−θhi,s) ≤ −µiθhi,s(1 − )
Note that since N, S, and M are finite and θ∗
(15)
(16)
i,s > 0, for all i, j, s, θ∗ > 0. Moreover, note that θ∗ does
not depend on λ, µ since the ratio of the linear and quadratic terms in the log moment generating functions are
independent of λ and µ.
θk
k! , for the Poisson-dominated case we have
As eθ − 1 =(cid:80)∞
j,s, θ∗
k=1
log Aj(θrj,s) ≤ λj
log Ui(−θhi,s) ≤ µj
(cid:88)
(cid:88)
k
k
(θrj,s)k
k!
(−θhi,s)k
k!
Again, by the same argument, we can have a θ∗ for which (26) and (27) are satisfied. Thus, for all θ < θ∗ we
have:
θ¯r +
(cid:88)
(cid:32)
(cid:32)
¯r −(cid:88)
θ
j
λjθ∗rj,s(1 + ) −(cid:88)
(cid:33)(cid:33)
i
µihi,s(α − )
µiθhi,s(1 − )
(18)
(19)
E(cid:104)
eθ(Xs(t)−Ys(t))(cid:105) ≤ exp
≤ exp
j(1 − α)λjrj,s and (cid:80)
i µihi,s(α − ) ≤ −γ(cid:80)
E(cid:104)
Since (cid:80)
i µihi,s >(cid:80)
all j, we have ¯r −(cid:80)
Note (28) follows from the fact λ ∈ (1− α)Cout. As > 0 can be chosen arbitrarily small, we can have α− > 0.
j λjrj,s scales with λ(N ), for sufficiently large λα with λj ≥ λα for
i µihi,s(α − ), for some γ > 0. Thus, we have for some θ > 0,
i
.
(17)
eθ(Xs(t)−Ys(t))(cid:105) ≤ exp (−θK) ,
where K scales with λ.
Now by Loynes’ construction:
Hence,
k=τ
Gs(t) = max
τ <t
t(cid:88)
Pr (Gs > g) ≤ E(cid:2)eθGs(cid:3)
≤ e−θgE(cid:104)
∞(cid:88)
∞(cid:88)
≤ e−θg
eθg
τ =1
≤ e−θg
(Xs(t) − Ys(t)).
t=0(Xs(t)−Ys(t))(cid:105)
(cid:80)τ
eθ(Xs(t)−Ys(t))(cid:105)(cid:17)τ
eθ maxτ
(cid:16)E(cid:104)
exp (−θτ K)
τ =1
≤ e−θg
1
1 − exp (−θK)
.
23
This in turn implies
(cid:16)
Pr
max
s
Gs > c log N
(cid:17) ≤ S Pr (Gs > c log N )
≤ Se−θc log N
1 − exp (−θK)
≤ S
N θc
1 − exp (−θK)
1
Note that the total number of unallocated jobs in the system is upper-bounded by (minj,s:rj,s>0 rj,s)−1(cid:80)
s Gs,
as any unallocated job has at least minj,s:rj,s>0 rj,s skill-hours unallocated. Hence the total number of unallocated
tasks in the system, Q, satisfies:
.
Pr (Q > cS log N ) ≤ Pr
≤ Pr
(cid:32)(cid:88)
(cid:18)
s
max
s
Gs > c min
j,s:rj,s>0
rj,sS log N
(cid:33)
(cid:19)
Note that θ∗ > 0 does not depend on λ, so
that θ∗c minj,s:rj,s>0 rj,s > 3 and hence, Pr (Q > cS log N ) ≤ o(N−2).
Gs > c min
rj,s log N
j,s:rj,s>0
1−exp(−θK) is O(1). As S = O(N ), we can choose a c > 0 such
N θc minj,s:rj,s>0 rj,s
1
S
1
1 − exp (−θK)
.
F. Proof of Proposition 12
We prove this proposition by constructing a simple (but general) system and show that the system is not stable
Consider N = 1, M = 2, and S > 1 being even. Let r = 1, h1 = (1, 1,··· S
2 terms, 0, 0,··· ) and h2 =
via a domination argument with a Markov chain that is not positive recurrent.
(0, 0,··· S
2 terms, 1, 1,··· ). Let the arrival to the system be i.i.d. with
(1 − α − δ)λ, w.p. (1 − )
2λ,
w.p. ,
(cid:40)
S
such that (1 + α)− δ(1− ) = 0, resulting in mean arrival rate (1− λ) and variance < λ. One possible construction
is to take = 1
2λ and δ accordingly. The agent-availability process is considered to be U1 = U 2 = λ. Note that
both arrival and agent availability processes are Gaussian-dominated as well as Poisson-dominated.
First, consider the case where greedy picking of tasks by the agents may be adversarial. Each agent picks all the
tasks that it can take from the job and so agents of different types pick from a different half of the S skill-parts.
Thus if there are ≥ 2λ jobs, in worst case (where adversary gives the job to the agent) agents of type 1 may pick
2 job-parts of λ jobs, while agents of type 2 pick parts of other λ jobs. Hence no job is actually allocated at that
allocation epoch. By the next epoch at least (1 − α)λ jobs have come and if the agents pick jobs in an adversarial
manner, again no job is actually allocated. Thus the number of unallocated jobs keep growing after it hits 2λ once.
Note that since there is a positive probability of ≥ 2λ arrivals, with strictly positive probability, the number of
unallocated jobs grows without bound.
Next, we prove the case where greedy picking of the tasks by the agents is random, i.e., an agent picks all S/2
parts of a randomly selected job (without replacement). As arrivals and availability are i.i.d., we can describe the
number of unallocated jobs by a Markov chain Q(t). Note that for Q ≤ λ, Pr(Q → 0) = 1. On the other hand,
0 < Pr(0 → Q) < 1 for Q < 2λ.
Consider any Q = nλ for n > 1. Note that Pr(Q → x) = for x < (n − 1)λ as no more than λ jobs can be
scheduled, because there are λ agents of each type. Again,
Pr (Q decreases at least by 1) ≤ 1 −
1 −
λ
Q + λ − αλ
because there are at least (1 − α)λ arrivals and each type picks λ agents randomly and this is the probability that
the picked sets have a non-empty intersection. Again, as there are 2λ arrivals w.p. ≥ we have
(cid:18)
(cid:18)
(cid:19)λ
(cid:19)λ
.
Pr (Q increases by at least λ) ≥
1 −
λ
Q + 2λ
Since
(cid:17)λ
(cid:17)λ .
n+1
1 − 1
1 − 1
(cid:16)
1 −(cid:16)
(cid:17)λ
(cid:17)λ → ∞
n+1
π(n + 1) = π(n)
(cid:16)
1 −(cid:16)
n+1
1 − 1
1 − 1
n+1
24
Based on computation of transition probabilities for each transition, it follows that for all k ≥ 1, the probability
of Q decreasing by at least 1 as well as the probability of Q decreasing by k is decreasing with Q, whereas the
probability of Q increasing by k increases.
Hence, we can dominate the above chain by another chain Q on λZ+ with transition probabilities
Pr (λn → λ(n + 1)) =
1 − 1
Pr (λ(n + 1) → λn) = 1 −
(cid:18)
(cid:18)
(cid:19)λ
(cid:19)λ
.
n + 1
1 − 1
n + 1
Now the chain Q(t) is a birth-death chain. If it has a finite (summable over states) invariant, then that is unique.
We first assume that the invariant is π and then show that it is not summable to prove that it is not positive
recurrent.4 As this is a birth-death chain the invariant measure must satisfy:
as n → ∞ for any finite λ > 0, this shows that π is not finite.
G. Proof of Theorem 13
unserved hours of skills s over all jobs.
Consider the different types of unallocated jobs. These are given by {Qj(t) : j ∈ [N ]}.
Consider the following processes: for each s ∈ [S], Qs(t) = (cid:80)
We now construct another process Q s.t. it dominates the process(cid:80)
Q, then the same bound applies for(cid:80)
each time t, (cid:80)
At time t,(cid:80)
s Qs. So, if we can show upper-bound on
s Qs. Hence, in turn we get a bound for {Qj(t)} (as min{rj,s > 0} = Θ(1)
Towards constructing a suitable Q we make the following observation about the dynamics of Qs and {Qj}. At
j Aj,1(t)rj,s amount of s skill hour is brought to add to Qs. Also, this queue gets some service
by the assumption that {rj,s} do not scale with the system size).
j:rj,s>0 Qjrj,s which represent the number of
m Um(t)hm,s s-skill hour of service is brought by the agents.
depending on the available agent hours.
For a job to be allocated, all tasks of it must find an allocation. Hence, for a job in type j-job to find an allocation
it must get rj,s hours of service from each skill s. Thus at any time t any skill s queue gets a service of at least
where ¯r = max{rj,k,s}. This is because of the following. For each skill(cid:80)
min
s∈[S]
Umhm,s − ¯r,
m
s Umhm,s hour is available. Note that
(cid:80)
a job can be allocated if all its tasks find allocations, coverse of which is also true. That is if all tasks of a step
found allocation then the step can be allocated. As mins∈[S]
s Umhm,s hours of service is brought by the agents
s Umhm,s− ¯r of s-skill hours are served (because a maximum of ¯r can be wasted,
for each skill, at least mins∈[S]
as no task is of size more than ¯r).
Also, note that the amount of required service brought to the queue Qs at time t is upper-bounded by
(cid:80)
(cid:88)
(cid:88)
j
max
s∈[S]
Aj(t)rj,s
4An alternate proof follows from noting that if we take a Lyapunov Q itself, then it has bounded jumps and it is easy to check that after
certain Q > 0 the drift is strictly positive, and invoke the Foster-Lyapunov (converse) theorem for irreducible chain with bounded absolute
drift.
Consider a process Qs with evolution
25
(cid:88)
j
Aj(t)rj,s
Qs(t + 1) = max( Qs(t) + max
s∈[S]
(cid:88)
m
− min
s∈[S]
Umhm,s + ¯r, 0).
Note that given Qs(t0) ≥ Qs(t0) at some t0, the same holds true for all t ≥ t0. This is because for x, a, b ≥ 0 and
x(cid:48), a(cid:48), b(cid:48) ≥ 0, with x ≥ x(cid:48), a ≥ a(cid:48) and b ≤ b(cid:48)
max(x + a − b, 0) ≥ max(x(cid:48) + a(cid:48) − b(cid:48), 0),
and hence, the monotonicity propagates over time.
s Qs it is sufficient to bound(cid:80)
s
Thus, to bound(cid:80)
which bounds(cid:80)
let us consider
s Qs.
From the evolution:
Q := S Q1,
Q(t + 1) = max( Q(t) + S max
s∈[S]
(cid:88)
j
Aj(t)rj,s − S min
s∈[S]
(cid:88)
m
Umhm,s + ¯r, 0)
Qs(t). Note that each of Qs has exactly same evolution, so
we can write the Loynes’ construction for this process which has the same distribution as this process (and for
simplicity we use the same notation, as we are interested in the distribution).
(cid:88)
Let us define Xs(t) and Ys(t) as follows: Xs(t) :=(cid:80)
τ≤t≤0
assuming that the process started at −∞.
Q1(0) = max
τ≤0
(cid:88)
(S max
s∈[S]
j
(cid:88)
Umhm,s + ¯r),
Aj(t)rj,s − S min
s∈[S]
m
j Aj(t)rj,s and Ys(t) :=(cid:80)
m Umhm,s. Then,
(20)
(21)
(22)
Now, for any θ > 0
Now,
Q1(0) = max
τ≤0
S(max
s
Xs(t) − min
s
Ys(t) + ¯r).
(cid:88)
τ≤t≤0
Pr(
Qj,1 > ¯rq) ≤ Pr(
(cid:88)
(cid:88)
j
Qs
s
1 > q)
≤ Pr( Q1(0) > q)
= Pr(θ Q1(0) > θq)
= Pr(exp(θ Q1(0)) > exp(θq))
≤ E[exp(−θq)]E[exp(θ Q1(0))].
max
(cid:88)
(cid:88)
τ≤0
τ≤t≤0
τ≤t≤0
≤(cid:88)
τ≤0
E[exp(θ Q1(0))] = E[exp(θS
(max
s
Xs(t) − min
s
Ys(t) + ¯r)
E[exp(θS
(max
s
Xs(t) − min
s
Ys(t) + ¯r))],
)]
where the inequality in (22) follows because for any random variables {Zj}, exp(θZj) are positive random variables
and sum of positives are more than their maximum.
Next, we bound the term within the summation over τ ≤ 0 in (22).
E[exp(θS
(max
s
Xs(t) − min
s
E[exp(θ(max
s
Xs(t) − min
s
Ys(t) + ¯r)))].
(23)
Ys(t) + ¯r))] ≤ (cid:89)
τ≤t≤0
(cid:88)
τ≤t≤0
26
s,s(cid:48)
(24)
E(cid:104)
τ≤t≤0 in (23),
Inequality (24) is due to the same reason as (22).
eθS(Xs(t)−Ys(cid:48) (t)+¯r)(cid:105)
Inequality in (23) follows because Xs(t), Ys(t) are i.i.d. over time.
E(cid:104)
eθ(Xs(t)−Ys(cid:48) (t))(cid:105)
Next we bound the term within the product(cid:81)
Let Aj(θ) = E(cid:2)eθAj(t)(cid:3) and Um(θ) = E(cid:2)eθUm(t)(cid:3) for j ∈ [N ] and m ∈ [M ]. For θ ∈ R, then,
E(cid:104)
eθS(maxs Xs(t)−mins Ys(t)+¯r)(cid:105) ≤(cid:88)
e−θYs(cid:48) (t)(cid:105)
eθXs(t)(cid:105) E(cid:104)
= E(cid:104)
= E(cid:104)
j Aj(t)rj,s+¯r(cid:105) E(cid:104)
(cid:105)
e−θ(cid:80)
eθ(cid:80)
(cid:105)(cid:89)
E(cid:104)
E(cid:104)
e−θUi(t)hi,s(cid:48)(cid:105)
= eθ¯r(cid:89)
= eθ¯r(cid:89)
(cid:89)
θ¯r +
Note that as λ ∈ αC, by the definition of CO, (cid:80)
(cid:80)
m µmhm,s −
m µmhm,s(cid:48) ≤ subpoly(N ) which is used in the following.
First consider the Gaussian-dominated case. Since the process variance is no more than mean and the moment
m µmhm,s and by assumption (cid:80)
(cid:88)
j λjrj,s < α(cid:80)
log Aj(θrj,s) +
log Ui(−θhi,s(cid:48))
Ui(−θhi,s(cid:48))
.
Aj(θrj,s)
(cid:88)
eθAj(t)rj,s
i Ui(t)hi,s
= exp
(25)
j
j
i
j
i
i
generating function of the variance is upper-bounded by that of a zero-mean Gaussian:
log Aj(θrj,s) ≤ λjθrj,s + λj
log Ui(−θhi,s) ≤ −µjθhi,s + µj
(θrj,1,s)2
2
(θhi,s)2
.
2
Note that for any two functions k1x2 and k2x, limx→0 k2x/k1x2 = ∞, and hence for any ∈ (0, 1) there exists
i,s > 0, for all i, j, s
x∗ > 0 such that for all x < x∗, k1x2/k2x < . Hence for any ∈ (0, 1), there exist θ∗
such that for all θ < θ∗ = mini,j,s(θ∗
j,s, θ∗
Note that since N, S, and M are finite and θ∗
(26)
(27)
i,s > 0, for all i, j, s, θ∗ > 0. Moreover, note that θ∗ does
not depend on λ, µ since the ratio of the linear and quadratic terms in the log moment generating functions are
independent of λ and µ.
θk
k! , for the Poisson-dominated case we have
As eθ − 1 =(cid:80)∞
j,s, θ∗
k=1
j,s, θ∗
i,s,
log Aj(θrj,1,s) ≤ λjθ∗rj,s(1 + )
log Ui(−θhi,s) ≤ −µiθhi,s(1 − )
log Aj(θrj,s) ≤ λj
(cid:88)
(cid:88)
k
(θrj,s)k
k!
(−θhi,s)k
Again, by the same argument, we can have a θ∗ for which (26) and (27) are satisfied. As (cid:80)
o(N δ), for all δ > 0, and(cid:80)
i µihi,s = Ω(N c), c > 0, for all θ < θ∗ we have:
log Ui(−θhi,s) ≤ µj
k!
k
E(cid:104)
eθ(Xs(t)−Ys(cid:48) (t))(cid:105) ≤ exp
θ∗¯r +
(cid:88)
(cid:32)
(cid:32)
¯r −(cid:88)
θ
j
i
λjθ∗rj,s(1 + ) −(cid:88)
(cid:33)(cid:33)
i
µihi,s(α − 2)
≤ exp
.
(28)
i µihi,s(cid:48) =
i µihi,s−(cid:80)
µiθhi,s(1 − ) + θ∗o(µiθhi,s)
Note (28) follows from the fact λ ∈ (1− α)Cout. As > 0 can be chosen arbitrarily small, we can have α− 2 > 0.
j λjrj,1,s scales with λ(N ), for sufficiently large λα with λj ≥ λα for
i µihi,s(α − ), for some γ > 0. Thus, we have for some θ > 0,
Since(cid:80)
i µihi,s >(cid:80)
j(1 − α)λjrj,s and(cid:80)
i µihi,s(α − ) ≤ −γ(cid:80)
all j, we have ¯r −(cid:80)
E(cid:104)
where K(N ) scales with N no slower than(cid:80)
eθS(maxs Xs(t)−mins Ys(t))(cid:105) ≤ S2 exp (−θSK(N )) .
eθS(Xs(t)−Ys(cid:48) (t))(cid:105) ≤ exp (−θSK(N )) ,
s:rj,1,s>0 λj(N ) = Ω(N c), c > 0.
E(cid:104)
Thus,
(29)
27
Hence, from (22), (23), and (24) we have that
(cid:88)
s
E[exp(θ
Qs(0))] = E[exp(θS Q1(0))]
= E[exp(θ Q(0))]
≤(cid:88)
S2τ exp(−θSK(N )τ)
τ≤0
≤ c(cid:48),
because S2 < exp(θ∗SK(N )) for all sufficiently large N.
Note that though we proved E[exp(θQ(t))] < c(cid:48) for t = 0, this holds for any finite t with exactly the same
proof. Hence,
Pr(Q > c log N ) ≤ Pr( Q > c log N ) ≤ c(cid:48)e−cθ∗ log N ,
which gives the result for an appropriate choice of c.
H. Proof of Theorem 14
Note that the set {λ : λ =(cid:80)
To prove that any λ ∈ CL is stabilizable it is sufficient to invoke Thm. 15 whose proof is below. To show that
λO /∈ CI is not stabilizable, we take an approach similar to the proof of Thm. 5.
Γ} is convex. Also, if λ(cid:48) ≤ λ (component-wise) for some λ belonging to
the set, then λ(cid:48) ∈ CI. That is, CI is coordinate convex. Hence for any λ /∈ closure of CL, there exists a hyperplane
h ≥ 0 that strictly separates it from CL, i.e., for any {λl ∈ Cl}, for some > 0
l λl, λl ∈ Cl
Note that at any epoch t, number of j jobs allocated ∆j(t) = (cid:80)
l
λl +
allocated to category l agents. Following similar steps as in the proof of Thm. 5, the result follows.
l ∆l
j(t), where ∆l
j(t) is the number of j jobs
hT λO ≥ hT(cid:88)
Let Ql,j,s(t) be the number of unallocated (j, s) tasks in the lth pool and Aj(t) be the number of arrived jobs
l Al,j = Aj(t). At any pool l MaxWeight
I. Proof of Theorem 15
is followed and Dl,j,s(t) is the number of allocated (j, s) tasks in pool l.
of type j. Let Al,j(t) be the number of jobs that are sent to pool l and(cid:80)
Consider a Lyapunov function L(Q) =(cid:80)
+ E
(cid:88)
− E
E [L (Q(t + 1)) − L (Q(t))Q(t)]
Ql,j,s(t) ∆l,j,s(t)Q(t)
Ql,j,s(t)Al,j(t)Q(t)
(cid:88)
l,j,s Q2
l,j,s. Then:
≤ E
l,j,s
l,j,s
(cid:17)Q(t)
(30)
A2
l,j(t) + ∆2
l,j,s
(cid:16)
(cid:88)
l,j,s
28
The last term can be bounded by noting:
A2
l,j(t) + ∆2
l,j,s
Al,j(t) + ∆l,j,s
(cid:17)2
2
(cid:88)
(cid:16)
l,j,s
(cid:88)
(cid:88)
(cid:88)
j,s
(cid:17) ≤
=
≤
j,s
≤ B(cid:48) < ∞.
Aj(t) +
Aj(t) +
∆l,j,s
l,j,s
1
min(rj,s > 0)
2
hl
i,sU l
i,s
(cid:88)
l,i,s
(31)
(32)
(cid:88)
(cid:16)
l,j,s
(cid:88)
l,j,s
E
Note (31) follows similarly as (12), whereas (32) follows because the arrival and agent-availability processes have
bounded second moments.
To bound the first term:
Ql,j,s(t)Al,j(t)Q(t)
Ql,j,s(t)
Al,j(t)Q(t)
Aj(t)Ql,j,s(t)Q(t)
(cid:33)
s
l,j
λl
j
λj
≤ E
= E
(cid:32)(cid:88)
(cid:88)
(cid:88)
(cid:88)
≤(cid:88)
=
(cid:88)
(cid:88)
l,j,s
E
l
,
(33)
(34)
(35)
On the other hand,
E
(cid:88)
where (33) is because of the fact JLTT-MWTA sends all arrivals of type j to the pool l with minimum(cid:80)
j,s
Ql,j,s(t)λl
j
s Ql,j,s(t).
Ql,j,s(t) ∆l,j,s(t)Q(t)
Ql,j,s(t) ∆l,j,s(t)Ql(t)
l,j,s
l
j,s
j,s
(cid:88)
Ql,j,s(t)λl
j
This in turn implies
(cid:88)
(cid:88)
− E
Ql,j,s(t) ∆l,j,s(t)Ql(t)
because at epoch t each pool l runs MaxWeight based on only Ql(t) and { ∆l,j,s : j, s} is independent of
{Al(cid:48),j(t), Ql(cid:48),j,s(t) : l(cid:48) (cid:54)= l} given Ql(t).
first term of (12). If λ + ∈ CL, then λl +
have
For every l, we can compute the difference between the lth term of (34) and that of (35), which is similar to the
L ∈ Cl, hence following the same steps as in the proof of Thm. 8 we
≤ −
(cid:88)
Following similar steps as in proof of Thm. 8, we obtain lim supt→∞(cid:80)
plies that for all j, s, l, lim supt→∞ E [Ql,j,s(t)] < ∞. This proves the theorem since for all j, Qj(t) ≤(cid:80)
It is sufficient to prove that CI ⊆ CO. Consider any λ ∈ CI, then by definition of CI, λ =(cid:80)
λ =(cid:80)
all l, λl ∈ Cl. By the characterization of the outer region of single category systems Cl ⊆ Cout
, for all l. Thus by definition of CO, λ ∈ CO, which completes the proof.
E [Ql,j,s(t)] < ∞, which in turn im-
l,s Ql,j,s(t).
E [L (Q(t + 1)) − L (Q(t))Q(t)] ≤ B(cid:48) −
L
l λl, where for
. Thus
, λl ∈ Cout
J. Proof of Theorem 16
Ql,j,s(t).
Ql,j,s(t).
l λl for λl ∈ Cout
l
l,j,s
l,j,s
L
j,s
j,s
l
l
K. Proof of Theorem 17
29
Consider the dynamics of Q∗
We find a high-probability bound on the number of unallocated jobs of type j and then bound the maximum
number of jobs across type. To bound the number of unallocated jobs, we use stochastic domination based on the
nature of the method of splitting job arrivals across different pools.
j (t) = maxl Ql,j(t). Let Aj,l(t) be the number of jobs of type j that were directed
to pool l. In the Improvised JSQ step, jobs are sent one-by-one with minimum backlog and hence, the queue l∗
with Ql∗
L . On
the other hand, just before allocation at epoch t + 1, the total number of j-jobs in l∗ cannot be less than the number
of jobs in any other l by more than 1. This is because of the Improvised JSQ which allocates jobs one-by-one to
the lowest backlogged (N l
j (t) gets the minimum number of jobs. Since minimum is less than average, Aj,l∗(t) ≤ Aj(t)
j) queue at that time. Hence, we have
j (t) = Q∗
j (t) + (cid:100) Aj(t)
Q∗
L (cid:101) + 1 ≥ Qj,l(t) + Aj,l(t).
i (t).
i U i(t)hi,s. Hence,
agent-availability by U i(t) = minl U l
We can follow similar steps by noting that for θ > 0:
Qs(t + 1) = Qs(t) + X s(t) − Y s(t) + ¯r+ .
Consider the following dynamics Qj(t), j ∈ [N ]. Arrivals for each j are according to (cid:100) Aj(t)
For this we proceed along the lines of the proof of Thm. 13, replacing arrivals by Aj(t) = (cid:100) Aj(t)
infinite) ∆l,j is monotonic in Ul, i.e., if Ul ≥ U(cid:48)l (component-wise) ∆l,j ≥ ∆(cid:48)
useful below.
Given Ql,j, for GreedyJob the number of jobs that can be allocated (assuming number of queued jobs to be
l,j for all j. This property will be
L (cid:101) + 1 and agent-
i (t). This is a single-category system and allocations in this system are
availability is according to U i(t) = minl U l
according to GreedyJob. As U i(t) ≤ U l
i (t) for all i, l, the number of allocations (assuming queues to be infinite)
satisfies ∆j(t) ≤ ∆l,j(t). This implies that for each type j Qj(t) dominates Q∗
j (t) as the first queue at any epoch
has more number of jobs to be allocated and less number of possible allocations (as ∆j is smaller). Thus, it is
sufficient to bound Qj(t).
L (cid:101) + 1 and
j Aj(t)rj,s and the
Hence, for each skill s, the queue of unallocated skill-hours H s(t) is the arrival X s(t) =(cid:80)
possible amount that can be drained is Y s(t) =(cid:80)
= E(cid:104)
e−θY s(cid:105)
e−θ(cid:80)
= E(cid:104)
(cid:80)
(cid:20)
e−θ minl
e−θ(cid:80)
E(cid:104)
≤(cid:88)
e−θ(cid:80)
eθX s(cid:105) ≤ eθ( 1
E(cid:104)
+1)¯r(cid:89)
(cid:105)
(cid:105)
(cid:21)
(cid:105)
Aj(t)(cid:105)
(cid:105) ≤ elog L+O(subpoly(N ))E(cid:104)
(cid:88)
e−θ(cid:80)
(37)
L − subpoly(N ) ∈ Cout. Thus, we can
i = subpoly(N ) for all i, λ ∈ (1 − α)CO implies that λ
For Gaussian-dominated as well as Poisson-dominated cases, following similar steps we can obtain that for a
Since maxl.l(cid:48) µl
use the same steps as we did for single-category systems.
θ∗ > 0 (independent of λ) and sufficiently large λ,
E(cid:104)
e−θ(cid:80)
Making the assumption on agent arrival rates,
On the other hand,
E(cid:104)
E(cid:104)
i − µl(cid:48)
i minl U l
i (t)hi,s
i U 1
i (t)hi,s
i U l
i (t)hi,s
i U l
i (t)hi,s
i U l
i (t)hi,s
.
= E
max
l
i U l
i (t)hi,s
l
L
θrj,s
L
e
.
j
(36)
E(cid:104)
eθ(X s−Y s)(cid:105) ≤ e−θ∗K(λ)+log L+O(subpoly(N ))
large such that −θ∗K(λ) + log L + O (subpoly(N )) is strictly negative and E(cid:2)eθ(X s−Y s)(cid:3) < 1. The rest follows
Note that K(λ) increases as Ω(mini λi) = Ω(N c) for some c > 0, L = O(1), hence there exists λ sufficiently
similarly as the proof of Thm. 11.
l
(cid:105)
30
L. Proof of Theorem 18
(cid:17)
(cid:16)
(cid:16)
=
(cid:88)
≤ (cid:88)
(cid:88)
=
j,s:rj,s>0
We first consider FD, FND, and IND systems.
Let Aj(t) be the number of accepted jobs of type j between starts of epochs t − 1 and t. Let ∆j,s(t) be the
number of allocated (j, s) tasks by the MaxWeight part of MWTA before the execution of Task Allocation. Let
Dj,s(t) be the number of allocated (j, s) tasks at allocation epoch t at the end of Task Allocation. Then:
j,s(t + 1) − Q2
j,s(t)
( Qj,s(t) + Aj(t) − Dj,s(t))2 − Q2
j,s(t)
(cid:88)
(cid:16) Q2
j,s:rj,s>0
j,s(t)
j,s:rj,s>0
Qj,s(t)
(cid:88)
( Qj,s(t) + Aj(t) − ∆j,s(t))2 − Q2
(cid:16) Aj(t) − ∆j,s(t)
(cid:17)
the fact that the arrival processes have bounded second moment. The value E(cid:104)(cid:80)
(cid:88)
Expectation of the second summation (conditioned on Q(t)) can be bounded by noting that A2
j,s(t)
(cid:105)
j (t) ≤ A2
j (t) and
can be bounded
similarly as in the proof of Thm. 8. Hence, we consider the expectation of the second term to be bounded by B
independent of Q.
j,s(t + 1) − Q2
j,s:rj,s>0 ∆2
j (t) + ∆2
j,s:rj,s>0
j,s:rj,s>0
j,s(t)
(cid:17)
E
+
(cid:17)
(cid:17)
(cid:16) A2
j,s(t)
(cid:17)Q(t)
(cid:16) Aj(t) − ∆j,s(t)
(cid:16)
(cid:17)Q(t)
Qj,s(t)
(1 − β(t))Aj(t) − ∆j,s(t)
Qj,s(t)
(cid:17)Q(t), A(t)
j Aj(t)
ν
j Aj(t)
ν
j Aj(t)
ν
+
+
+
(cid:16)
(cid:16)
(cid:16)
j,s:rj,s>0
j,s:rj,s>0
(cid:88)
(cid:88)
(cid:88)
(cid:88)
j,s:rj,s>0
Qj,s(t)
Qj,s(t)
Qj,s(t)
(cid:16)
j,s:rj,s>0
j,s:rj,s>0
(cid:16) Q2
(cid:88)
E
(cid:88)
β(t)(cid:80)
E
β(t)(cid:80)
E
β(t)(cid:80)
E
β∗(cid:80)
E
j,s:rj,s>0
≤ B + E
≤ B + E
≤ B + E
≤ B + E
≤ B + E
≤ B + E
≤ B +
1
ν
(1 − β∗)Aj(t) − ∆j,s(t)
Q(t), A(t)
j Aj(t)
ν
+
(cid:88)
E[(β∗ − β(t))
j,s:rj,s>0
Aj(t)] + E
j
(cid:88)
where (39) follows as β(t) minimizes β(cid:80)
E[(β∗ − β(t))
≤ B +
1
ν
j
Qj,s(t)
(cid:88)
j,s:rj,s>0
Qj,s(t)
(cid:16)
(cid:88)
j Aj(t) − νβ(cid:80)
j,s:rj,s>0
Aj(t)] + E
(1 − β∗)Aj(t) − ∆j,s(t)
(cid:16)
Qj,s(t)
(1 − β∗)λj(t) − ∆j,s(t)
(cid:17)Q(t)
j,s:rj,s>0
Qj,s(t)Aj(t).
(1 − β(t))Aj(t) − ∆j,s(t)
j Aj(t)
ν
Q(t), A(t)
(1 − β(t))E[Aj(t)] − ∆j,s(t)
j Aj(t)
ν
Q(t), A(t)
(1 − β(t))Aj(t) − ∆j,s(t)
j Aj(t)
ν
Q(t), A(t)
Q(t)
(cid:17) − β(t)(cid:80)
(cid:17) − β(t)(cid:80)
(cid:17) − β(t)(cid:80)
(cid:17) − β(t)(cid:80)
(cid:17)Q(t)
j Aj(t)
ν
Q(t)
Q(t)
Q(t)
Q(t)
(38)
(39)
31
Qj,s(t)
≤ −
(cid:88)
(cid:88)
(cid:105)
λj − E
j,s:rj,s>0
j
j,s:rj,s>0
.
Qj,s(t)
< ∞ which implies the crowd
j,s:rj,s>0 Qj,s(t)
< C for some C < ∞.
As (1 − β∗)λ + 1 ∈ C, following the same steps as in the proof of the Thm. 8:
E
E
1
ν
(cid:16)
j,s(t)
j,s:rj,s>0
j,s:rj,s>0
Qj,s(t)
(cid:16) Q2
Thus we can write
j,s(t + 1) − Q2
(cid:88)
(cid:88)
(1 − β∗)λj(t) − ∆j,s(t)
As β∗ − β(t) ≤ 1, we have
(cid:17)Q(t)
(cid:17) ≤ B +
(cid:88)
Following again the same steps we show that lim supt→∞ E(cid:104)(cid:80)
system is stable. Hence, we can assume that E(cid:104)(cid:80)
(cid:105)
(cid:17) ≤ BT +
T(cid:88)
(cid:88)
T(cid:88)
= E
Now consider the ID setting and use the Lyapunov function(cid:80)
(cid:16) Q2
(cid:88)
T(cid:88)
j,s(T ) − Q2
T(cid:88)
(cid:88)
which in turn implies
j,s:rj,s>0 Qj,s(t)
(1 − β(t))
(1 − β(t))
λj − 1
T
(1 − β∗)
(cid:88)
j,s(0)
Aj(t)
j,s:rj,s>0
1
ν
t=1
t=1
t=1
E
j
E
j
j
t=1
j
can be bounded by
E[(β∗ − β(t))
Aj(t)],
(cid:88)
j
Q2
j,s(0)
Aj(t) ≤ νB +
ν
T
j,s:rj,s>0
(cid:88)
.
(cid:88)
Aj(t)
Since B is a constant depending on arrival and availability statistics, ν and T can be chosen to be small and large
respectively to ensure that the left side is arbitrarily small. The desired result follows by noting that
Also note that since the only requirement is the independence of A(t) and U(t) across time, the proof directly
extends to settings with non-stationary arrival and availability processes.
l,j,s Q2
l,j,s. Similar to before, the Lyapunov drift
Note that(cid:80)
as Aj(t) =(cid:80)
l,j,s:rj,s>0
Ql,j,s(t)
B + E
.
(cid:17)Q(t)
(cid:88)
l,j,s:rj,s>0 Ql,j,s(t) Al,j(t) is equal to(cid:88)
Al,j(t) and JLTT ensures that the jobs are sent to the category with minl ((cid:80)
(cid:88)
(cid:88)
(cid:88)
(cid:16) Al,j(t) − ∆l,j,s(t)
(cid:32)(cid:88)
(cid:33)
(cid:16) Al,j(t) − ∆l,j,s(t)
(cid:17)Q(t)
(cid:33)
(cid:32)(cid:88)
Aj(t) − (cid:88)
Aj(t) − (cid:88)
Ql,j,s(t) ∆l,j,s(t)Q(t)
,
Ql,j,s(t)
Ql,j,s(t)
Ql,j,s(t)
Ql,j,s(t)
Aj(t),
≤ E
l,j,s:rj,s>0
l,j,s:rj,s>0
l,j,s:rj,s>0
= E
l,j:rj,s>0
min
min
E
s
s
l
l
l
λl
j
λj
l,j:rj,s>0
j,s:rj,s>0
Ql,j,s(t) ∆l,j,s(t)Q(t)
where (1 − β∗)λ =(cid:80)
FD/FND/IND settings, i.e., the MWTA proof for (1 − β∗)λl + 1 ∈ Cl for each l.
l(1 − β∗)λl for λl + 1 ∈ Cl. The remainder of the proof is similar to the approach for the
s Ql,j,s(t)). So we have
|
1810.11131 | 1 | 1810 | 2018-10-25T23:12:29 | The Impact of Position Errors on Crowd Simulation | [
"cs.MA",
"cs.DC"
] | In large crowd events, there is always a potential possibility that a stampede accident will occur. The accident may cause injuries or even death. Approaches for controlling crowd flows and predicting dangerous congestion spots would be a boon to on-site authorities to manage the crowd and to prevent fatal accidents. One of the most popular approaches is real-time crowd simulation based on position data from personal Global Positioning System (GPS) devices. However, the accuracy of spatial data varies for different GPS devices, and it is also affected by an environment in which an event takes place. In this paper, we would like to assess the effect of position errors on stampede prediction. We propose an Automatic Real-time dEtection of Stampedes (ARES) method to predict stampedes for large events. We implement three different stampede assessment methods in Menge framework and incorporate position errors. Our analysis suggests that the probability of simulated stampede changes significantly with the increase of the magnitude of position errors, which cannot be eliminated entirely with the help of classic techniques, such as the Kalman filter. Thus, it is our position that novel stampede assessment methods should be developed, focusing on the detection of position noise and the elimination of its effect. | cs.MA | cs |
The Impact of Position Errors on Crowd
Simulation
Lei Zhang∗1, Diego Lai2, and Andriy V. Miranskyy†1
1Department of Computer Science, Ryerson University, Toronto, Canada
2Laipac Technology Inc., Toronto, Canada
Abstract
In large crowd events, there is always a potential possibility that
a stampede accident will occur. The accident may cause injuries or
even death. Approaches for controlling crowd flows and predicting
dangerous congestion spots would be a boon to on-site authorities to
manage the crowd and to prevent fatal accidents. One of the most
popular approaches is real-time crowd simulation based on position
data from personal Global Positioning System (GPS) devices. How-
ever, the accuracy of spatial data varies for different GPS devices, and
it is also affected by an environment in which an event takes place.
In this paper, we would like to assess the effect of position errors on
stampede prediction. We propose an Automatic Real-time dEtection
of Stampedes (ARES) method to predict stampedes for large events.
We implement three different stampede assessment methods in Menge
framework and incorporate position errors. Our analysis suggests that
the probability of simulated stampede changes significantly with the
increase of the magnitude of position errors, which cannot be elimi-
nated entirely with the help of classic techniques, such as the Kalman
filter. Thus, it is our position that novel stampede assessment methods
should be developed, focusing on the detection of position noise and
the elimination of its effect.
∗Email address: [email protected]
†Email address: [email protected]
1
1 Introduction
In recent decades, the number of events where a stampede has occurred is
increasing, along with the number of people involved in such accidents. In
2014, 36 people were killed and 47 others were injured in a crush at New
Year's Eve celebrations in Shanghai, China [4]; in 2015, at least 2,177 people
were crushed to death and 934 were injured at the annual Hajj in Saudi
Arabia [50]; in 2017, more than 1,500 soccer fans were injured in stampede
in Turin, Italy [52]; in the same year, at least 22 people were killed and
hundreds were injured in a stampede at Elphinstone road station in Mumbai,
India [51].
To prevent stampede accidents, the Saudi Ministry of Hajj has already
begun considering plans to give each pilgrim an electronic bracelet with in-
dividual identification and Global Positioning System (GPS) [1]. Moreover,
the percentage of the population owning smartphones (having built-in GPS)
is increasing. The recent reports show that ten countries now have smart-
phone penetration1 greater than 70% [38], with the numbers growing each
year [45]. This will provide a technological foundation to enable the gather-
ing of locations of most of the people in the near future, even if no specialized
device (such as the electronic bracelet mentioned above) is present.
In recent years, many researches are leveraging data from GPS handheld
devices to prevent stampedes by providing early warning to authorities [60,
15, 68]. However, all GPS devices suffer from position errors. The main
GPS error source is due to the satellite synchronization [33]. Other errors
arise because of atmospheric disturbances that distort the signals before they
reach a receiver. Reflections from buildings and other large, solid objects
can lead to GPS accuracy problems too. To the best of our knowledge, the
literature is lacking the analysis of the impact of GPS noise on the prediction
of stampedes.
In this paper, we focus on quantifying the impact of GPS measurement er-
rors on stampede probabilities. The main contributions of this paper include:
(1) presentation of a stampede prediction solution by the incorporation of
three different stampede assessment methods into a crowd simulator; (2) in-
troduction of a noise modeling technique in crowd simulation and evaluation
of the probability of simulated stampede with various position errors; and (3)
analysis of the impact of position errors on simulated stampede probabilities.
1That is the percentage of the people in the population that own a smartphone.
2
The remainder of this paper is constructed as follows: Section 2 reviews
the related work in crowd management, stampede assessment methods, and
crowd simulation. Section 3 describes our proposed solution, namely, how
we simulate stampede accidents and measure the position errors. Section 4
covers experimental results and analysis. Finally, Section 5 summarizes our
findings.
2 Related Work
Crowd management based on data from GPS handheld devices (e.g., smart-
phones, smartwatches, and personal fitness trackers) has been widely studied.
Wirz et al. [60] adopted the crowd pressure technique to visualize this in-
formation as heat maps, offered a global view of the crowd situation, and
assessed different crowd conditions instantaneously throughout an event.
Franke et al. [15] implemented a smartphone based crowd management sys-
tem, which also uses a heat map representation of the crowd state and its
evolution. Zhou et al. [68] proposed a solution based on Baidu map and
developed a prediction model to perceive the crowd anomaly and to assess
the risk of the crowd event.
Besides GPS-based solutions, vision-based systems have also been devel-
oped to detect congestion spots and pedestrian behaviors. The accuracy
of vision-based systems can be affected by three factors: camera coverage,
camera resolutions, and level of illumination [68, 25]. The existing computer
vision work focuses on real-time detection of the stampedes [24, 35, 26, 36,
69, 42, 43] rather than proactive prediction, making it complementary to our
goal. While the field of computer vision is rapidly evolving, modern com-
puter vision methods have not perfected crowd analysis yet. For example,
they have difficulty tracking individuals in the crowd [24, 35]. The training
of a computer vision model for each camera has to be individualized [26].
The accuracy of the detection of anomalies fluctuates for various methods,
e.g., going to as low as 57% [36], which may negatively affect the accuracy
of stampede analysis. Last but not least, vision-based detection techniques
face privacy and security related issues [44] (which are conceptually similar to
those based on GPS trackers even though a different technology is leveraged).
Recent research of anomaly detection in crowd focuses on developing deep
learning-based approaches, e.g., Convolutional Neural Networks (CNNs),
combined with video analysis [69, 42, 43]. However, they have not yet been
3
tested on high-density scenarios, which result in stampedes; so far the perfor-
mance of CNNs was benchmarked against the UCSD dataset [54], which fo-
cuses on sparse crowds with non-pedestrian anomalous objects (e.g., skaters,
biker, carts, and wheelchairs) and anomalous pedestrian patterns (e.g., a
pedestrian walking in opposite direction to other pedestrians). Conceptu-
ally, one may create a hybrid stampede detection system that would leverage
GPS trackers and video data feeds synergistically, improving the accuracy of
the stampede analysis [5].
Agent-based pedestrian crowd simulation is a well-known solution for
crowd management. Thus, a variety of simulation frameworks were proposed.
Almeida et al. [2] proposed a multi-agent framework to simulate emergency
evacuation scenarios. Curtis et al. [10] introduced an open-source, cross-
platform and agent-based crowd simulation framework -- Menge. Mahmood
et al. [34] proposed an agent-based crowd simulation and analysis framework
and used a case study of Hajj as an example for the assessment of crowd
evacuation strategies.
To analyze the risk of stampedes, a number of stampede measuring tech-
niques were presented in the literature. Helbing et al. [20], Johansson et
al. [22] derived crowd pressure formulas (by analyzing video recordings of
the crowd disaster during the Hajj in 2006) and discovered crowd turbulence
phenomenon, in which physical contacts among people are transferred and
accumulated. People try to escape the crowd, which causes crowd panic.
Such phenomenon may trigger crowd disasters, such as stampedes. Lee
and Hughes [29] developed a crowd pressure measuring technique using the
standard forward-backward auto-regressive modeling approach. Elliott and
Smith [14] examined some sporting disasters and found the relationship be-
tween the accidents and the inter-person forces. Helbing et al. [19] introduced
social and physical forces among pedestrians and then treated each pedes-
trian as a particle abiding Newton's laws. Lee and Hughes [28] studied the
relationship between crushing situation and people density, which gives the
typical threshold of people density of different nationalities. However, none of
these work has taken into account the impact of GPS noise on the prediction
of stampedes.
Recent research has attempted to combine data from GPS and external
sensors to improve the tracking accuracy. Groves et al. [17] developed a posi-
tioning technique by integrating GPS and Wi-Fi data (to improve positioning
indoors) and accelerometers (to distinguish pedestrian behavior). Corbetta
et al. proposed an approach to utilize overhead Microsoft KinectTMdepth
4
sensors placed in a walkway to provide reliable automatic pedestrian posi-
tioning for tracking [9]. Suzuki et al. [48] proposed to use infrared cameras
to eliminate the multi-path errors which are caused by different positioning
systems. Assisted-GPS (A-GPS) technique is also widely used to improve
the performance of GPS-enabled smartphones [56].
In general, combining
GPS, WiFi, and sensors data improves tracking accuracy indoors (see [47]
for review). Besides employing external sensors, various algorithms were de-
veloped to mitigate GPS measurement noise, e.g., the Kalman filter [13, 61]
and the autoregressive moving average filter [30].
As a rule of thumb, the accuracy of GPS can be hugely affected by the
environment. The U.S. government claimed that GPS-enabled smartphones
are typically accurate to within 4.9 m; however, their accuracy worsens near
buildings or other obstacles [39]. Zandbergen and Barbeau's research showed
that the horizontal error of GPS-enabled smartphones ranges from 5.0 m to
8.5 m [64]. In 2015, Garnett and Stewart tested two personal GPS devices
and two GPS-enabled mobile devices, and they found that the mean relative
accuracy ranges from 3.65 m to 6.50 m [16].
3 The ARES method
In this paper, we propose a simulation-based method to predict a stampede
during large crowd events. The name of the method is ARES, which stands
for Automatic Real-time dEtection of Stampedes. In ARES, we assume that
the venue map is provided before the event, and we can collect all pedestrian
locations via personal GPS devices during the event. We call pedestrians
with wearable GPS devices agents. The word agent is used interchangeably
with pedestrian in this paper. The method employs simulation to monitor
pedestrians' behavior and measure possible stampede accidents. In case of a
potential stampede, on-site authorities will be notified by an alert message,
so that the authorities can intervene proactively and prevent the stampede
from occurring.
The key idea of ARES is to run the simulation and estimate the prob-
ability of a future stampede ahead of time. In reality, agent locations are
collected in a discrete time manner to preserve GPS power. The interval
between two GPS data collection points is denoted by T . During this time
interval, GPS locations are unknown. Thus, we use the latest GPS loca-
tions as starting positions and simulate agents' trajectories for the interval
5
T . To mitigate the uncertainty, we employ Monte Carlo method in simula-
tions to calculate the stampede probabilities. We run simulations in parallel
on a computing platform fast enough (e.g., a computer cluster) to ensure
simulation completion time K is less than T . Once the stampede probabil-
ity is calculated, the result (including position information) is sent to the
authorities, if necessary.
Figure 1: ARES steps to detect stampede accidents in large crowd events.
Figure 1 outlines the major steps of ARES. The details of the steps are
as follows.
1. Collect agents' current locations and velocities. In the beginning
of each interval T , we collect all agents' locations and velocities in the
field. The locations and velocities are measured in the global coordinate
system.
6
2. Convert agents' locations and velocities to the local coordi-
nate system. Agents' longitude and latitude coordinates would be
converted to local coordinates, which are relative locations in the venue
map.
In addition, agents' velocities in the global coordinate system
would also be converted to the ones in the local coordinate system.
3. Run the simulation with a pedestrian model and a stampede
assessment method. Agents' locations and velocities from Steps 2
are input parameters for the pedestrian model. We employ Monte Carlo
method to run simulations and execute M simulations in parallel using
GNU parallel [49] (other parallelization techniques can be used as well).
This is done to ensure that K < T and the time difference is sufficient
for on-site authorities to respond.
4. Compute simulated stampede probabilities. Each realization of
simulation reports if a stampede has happened or not. After all the
simulations complete, we calculate the expected value of the simulated
stampede probability p based on the Law of Large Numbers.
5. Send alerts to on-site authorities if necessary. If the value of p
exceeds a threshold, an alert message can be sent out to notify on-site
authorities, so that they could control the crowd near the specific spot.
6. Repeat the above steps once a new batch of locations and ve-
locities data comes in. Once a new batch of locations and velocities
data has been collected after interval T , we repeat the procedure from
Step 1 to Step 5.
There are two important assumptions in ARES. First, we assume that
each pedestrian has a wearable GPS device. This assumption is reasonable
as mentioned in Section 1. Second, we assume that we know each agent's
next destination based on the current state in the simulation. We believe
that this is also a reasonable assumption in many cases. For instance, we
know that pilgrims will stop and throw pebbles at three pillars on Jamarat
bridge during Hajj, and soccer fans will move towards the nearest exit (or
the one that they came in) to leave a stadium.
7
3.1 Simulation framework
We employ Menge [10] as our crowd simulation framework. Menge has two
built-in pedestrian models: ORCA and PedVO. The former is based on the
Optimal Reciprocal Collision Avoidance (ORCA) algorithm [55], and the lat-
ter is based on the Pedestrian Velocity Obstacle (PedVO) model [11]. As a
successor to the ORCA model, PedVO preserves ORCA's geometric opti-
mization technique and introduces a self-adaptive velocity technique by em-
ploying the fundamental diagram -- the phenomenon that pedestrian speed
reduces as density increases. The PedVO model captures both physiologi-
cal and psychological factors that give rise to the fundamental diagram. It
computes a collision-free velocity for each pedestrian based on a preferred
velocity and its local neighborhood (i.e., other pedestrians and obstacles).
In our work, we choose PedVO to simulate pedestrians' behavior, but
others may choose the most suitable pedestrian model (e.g., ORCA or Jo-
hansson et al.'s model [23]) to address their cases. The default parameters
of the PedVO model in Menge are as follows: agents are represented as
filled circles with radius equals 0.19 m [32], agents' preferred velocity equals
1.04 m/s [8]. Note that the default preferred velocity is lower than the av-
erage preferred walking speed -- 1.4 m/s [7]. However, this setup is adequate
for a crowd of people in a pilgrimage setting. The PedVO model uses the
preferred velocity to compute an actual velocity based on the surroundings
(the maximum velocity is 2 m/s [65]). The maximum acceleration rate is
5 m/s2 [3]. Other calibration parameters of PedVO can be found in B. Note
that these values may be adjusted if necessary (other reasonable values of
the parameters are discussed in [12]).
To simulate a representative crowd congestion scenario, we build a simpli-
fied venue map of the Jamarat bridge based on the map in the supplementary
material of Helbing et al.'s work [21]. The Jamarat bridge is the place where
the stampede happened during Hajj in 2006. A simulated scene on the map
can be seen in Fig. 2.
3.2 Stampede assessment methods
We extend Menge with three stampede assessment methods. The three
threshold-based stampede assessment methods that we employ are crowd
pressure method [20], physical force method [14, 19], and density method [28].
Details of these methods are given below.
8
Figure 2: Simulation venue map. A set of emulated agents are generated on
the right ramp (in a rectangular grid), and they are moving towards the left
in the venue. The width of the ramps is 22 m, and the width of the bridge is
44 m. The three obstacles simulate the three pillars on the Jamarat bridge
where pilgrims throw stones. Agents stop at each pillar for an average of
60 simulation time units (i.e., 60 s in our simulation), and they get off the
bridge after that. Note that this venue is smaller than the actual Jamarat
bridge, but it is sufficient to illustrate the concept.
3.2.1 Crowd pressure method
As crowd density reaches a certain dangerous level (typically seven or eight
people per square metre [20]), the physical contacts transfer and accumu-
late among the crowd, and individual motion gets replaced by mass motion.
When the pressure becomes too large, the mass splits up into different clus-
ters, which causes strong variations of local strengths and velocities. This
can lead to sudden and uncontrolled pressure release. Such a phenomenon is
referred to as crowd turbulence. Crowd turbulence is observed by analyzing
video recordings of the crowd disaster during the Hajj in 2006 [20]. People
try to escape such a situation; therefore, they push each other. This is called
crowd panic as per [20]2. The crowd panic introduces extra power into the
compressed crowd. Eventually, stampede accidents occur when a significant
amount of power accumulates and then releases.
Both crowd turbulence and stampede accident can be quantified by crowd
2Note that a stampede may also occur in the absence of panic.
9
pressure (which is different from physical pressure). Crowd pressure is defined
as local density times the local velocity variance [20].
Formally, the local density at a given place (cid:126)r = (x, y) and time t is defined
as
(cid:16)
(cid:126)rj(t) − (cid:126)r
(cid:17)
,
(cid:88)
j
ρ((cid:126)r, t) =
f
(cid:16)
(cid:126)rj(t) − (cid:126)r
(cid:17)
where (cid:126)rj(t) is the position of pedestrian j in the surrounding of (cid:126)r at time t,
and f
is a Gaussian distance-dependent weight function
(cid:16)
(cid:126)rj(t) − (cid:126)r
(cid:17)
f
=
1
πR2 exp
(cid:104)−(cid:107)(cid:126)rj(t) − (cid:126)r(cid:107)2/R2(cid:105)
,
(2)
where R defines the radius of the local measured area. The local velocity at
position (cid:126)r and time t is defined as
(1)
(3)
(cid:80)
(cid:80)
(cid:16)
(cid:17)
(cid:16)
(cid:17) ,
(cid:126)rj(t) − (cid:126)r
(cid:126)rj(t) − (cid:126)r
j (cid:126)vjf
j f
(cid:126)V ((cid:126)r, t) =
where (cid:126)vj is the velocity of of pedestrian j in the surrounding of (cid:126)r at time t.
Then, we can determine the local crowd pressure based on Eqs. (1) and (3)
as
P ((cid:126)r, t) = ρ((cid:126)r, t) · Var(cid:126)r,t((cid:126)V ),
(4)
where Var(cid:126)r,t((cid:126)V ) is the local velocity variance at time t, and its value is cal-
culated by Var(cid:126)r,t((cid:126)V ) = (cid:104)[V ((cid:126)r, t) − (cid:104)V (cid:105)(cid:126)r]2(cid:105)(cid:126)r, where (cid:104)V (cid:105)(cid:126)r is the average local
velocity over the circular area centered in (cid:126)r and with radius R. Accord-
ing to [20], crowd turbulence starts when the pressure reaches 0.02 s−2, and
stampede starts when the pressure reaches 0.04 s−2.
3.2.2 Physical force method
The second stampede assessment method is based on physical force [14, 19].
The idea is that physical interactions in jammed crowds add up and can lead
to dangerous consequences.
In simulations, we monitor the physical force
among agents, and set a threshold for the force. A stampede is measured if
any monitored physical force exceeds the threshold. Pure force function is
defined by Newton's second law, F = m · a, where m is the mass, and a is
the acceleration. Based on the research in [14, 19], 4500 N is the force that,
10
on average, may cause an adult to fall on the ground, potentially causing
stampede accidents.
For parameterization, we set m to follow a normal distribution with mean
value of 70 kg (which is a typical weight of adults in Europe [57]) and standard
deviation of 10 kg [27]. We set the minimum value of m to 50 kg, and the
maximum value of m to 100 kg. Note that these values can be adjusted
according to different nationalities and occasions. The value of a can be
calculated by
(cid:126)v(t + ∆t) − (cid:126)v(t)
,
(5)
a =
∆t
where ∆t is the change in time, and (cid:126)v(t) is the velocity at time t.
3.2.3 Density method
Our last stampede assessment method is based on a local density thresh-
old. Lee and Hughes [28] found the correlation between crowds involving
stampede accidents and crowd density. Typical stampede accidents involve
a large number of pedestrians, which is generally about seven or eight pedes-
trians per square metre (possibly up to 13 pedestrians per square metre).
The maximum observable densities are naturally dependent on different na-
tionalities [28].
Figure 3: A stampede is observed in the density method. The white filled
circle is the agent whose neighbors exceed the threshold -- 22, and the neigh-
bors are within the unfilled circle with R = 1 m. All the information about
the agent and its neighborhood can be seen in the left part of the figure.
11
Compared to the local density calculation (which is an estimate) in the
crowd pressure method, we employ a classical density estimation. We moni-
tor the number of neighbors (local density) for each agent during simulations,
and set a threshold for the local density, which is calculated as D = L/(πR2),
where L is the number of neighbors within the area where the radius is R.
In our experiments, we set the density threshold to seven agents per square
metre as per [28]. A stampede is measured if any local density exceeds this
threshold.
Figure 3 shows a scenario where a stampede is detected with the density
threshold. In this case, we have L = 22 agents and R = 1 m. Thus, D ≈ 7.01
m−2. In our experiments, the maximum density that we observed is 29 agents
within the radius of one metre, i.e., D ≈ 9.24 m−2.
In this method, the density calculation does not take into account the ob-
stacles (i.e., pillars and walls). We conjecture that if we took obstacles into
account (when computing local densities), the simulated stampede probabili-
ties might increase, making the estimates more sensitive (i.e., increasing false
positive error rate). As the coordinates of agents and obstacles are known,
one may create an algorithm to estimate the impact of obstacles on the local
densities.
In the literature, there exist various approaches to estimate local or global
densities [35, 46, 53], but they are not applicable to our experiments. A
fractal-based image processing technique was proposed to estimate crowd
density in [35]. The work of [46] focused on estimating global density rather
than local density. The method in [53] can be used to estimate local density,
but it is not applicable to our setup, because it assumes that local densities
are homogeneous (which is violated in our case). These methods cannot be
directly applied to our case, but they can be extended to take into consider-
ation only the accessible areas. For example, we know all the coordinates of
obstacles and agents in our case study, and we can remove inaccessible ar-
eas from the density estimators. Because the standard approach is sufficient
for the density estimation, we leave the research of obstacle-aware density
estimation to future work.
3.3 Noise modeling
Now, we introduce the noise modeling technique in crowd simulation.
In
real world, personal GPS devices often report locations with certain amount
of errors. These errors include horizontal distance errors, vertical distance
12
errors, and velocity errors. In this paper, we only consider horizontal errors
as our map is two-dimensional. Velocity errors may have impact on the
two velocity-dependent measurements, i.e., the crowd pressure method and
the physical force method; however, these errors will not affect the density
method (which is velocity independent). More importantly, it was shown that
velocity errors are negligible3 compared to distance errors: velocity errors are
in the order of centimetres while distance errors are in the order of metres [58].
Therefore, we chose to ignore the velocity error.
The empirical distribution of latitude and longitude errors is very simi-
lar to normal distribution [63, 66]. Consequently, the GPS horizontal error,
which is geometric sum of these two errors, is very similar to Rayleigh dis-
tribution [40]. The Cumulative Distribution Function (CDF) for Rayleigh
distribution is:
(cid:33)
(cid:32)
− z2
E2
P (Z ≤ z) = 1 − exp
, E > 0,
(6)
where z is the distance error, and E is the root-mean-squared distance error.
Finally, we could use Eq. (6) to generate random errors z to add to the
starting true positions of agents in the simulation. This simulates the sce-
nario where the real-time simulation starts with GPS position errors.
In
our simulations, we vary the values of E from zero to ten metres to test
the vulnerability of the stampede assessment method. As mentioned earlier,
Zandbergen and Barbeau's research shows that the positional error of smart-
phones ranges from five to eight metres [64]. Wilson's tests on personal GPS
devices suggest a similar estimate [59]. Thus, we believe that it is reasonable
to set the range of the values of E as zero to ten metres, so that we could
cover most of the cases. Note that in the case when E = 0, we use the actual
starting positions of the agents; thus, we do not need to use Eq. (6) in this
case.
Figure 4 shows an example, where agents' starting positions have no noise
added (E = 0). Figures 5 and 6 show two examples, where noise is added to
starting positions. The values of noise are drawn from the Rayleigh distri-
bution as per Eq. (6) with E = 5 and 10 m, respectively. As can be seen in
3The velocity estimate can achieve the accuracy of centimeters per second by using
Doppler frequency shifts of the received signal produced by user-satellite motion [41].
In GPS applications, improved performance is achieved by processing differences of con-
secutive carrier phase measurements, which are better than the "raw" Doppler measure-
ments [41].
13
Figure 4: Starting positions with E = 0 m.
Figure 5: Starting positions with E = 5 m.
Figs. 5 and 6, some agents are shifted outside of the bridge due to the errors.
We deliberately choose not to alter agents' behavior in such simulations to
assess the effect of the spread in simulated stampede probabilities (further
discussion can be found in Section 4.1).
3.4 Simulated stampede probabilities
To estimate the simulated stampede probabilities and the impact of posi-
tion errors, we employ Monte Carlo method. Given probabilistic nature of
14
Figure 6: Starting positions with E = 10 m.
Menge simulation, every realization of the trajectories of agents from start
to finish will be different. By the Law of Large Numbers, once we have suf-
ficient number of realizations of the trajectories, we will be able to compute
the expected value of the probability of simulated stampede along with the
confidence interval for the expected value.
Every realization of the simulation will report if a stampede has occurred
in a given run or not. Let us denote the outcome of the simulation by the
indicator variable I, and the total number of realizations is n. If a stampede
occurred in a given realization, I will be equal to one, otherwise -- to zero.
Then, the expected value of the probability of simulated stampede p is given
by
(cid:80)n
p =
j=1 Ij
n
.
The 95% confidence interval of this expected value is
√
n, p − 1.96σ/
√
≈ [p + 1.96σ/
n],
(7)
(8)
where σ is the standard deviation of the realizations of I.
3.5 Discussion
As discussed in Section 2, there exists a large number of stampede assessment
methods. For this study, we chose three of these methods. The design of this
study is based on the concept of the critical case [62]. That is, the selected
15
models are representative instances of the stampede assessment methods.
While we cannot generalize these results to other methods, the fact that
these models are all affected by the measurement error suggests that other
models may be affected as well. The same experimental examination can
also be applied to other stamped prediction models and methods with well-
designed and controlled experiments.
The GPS accuracy can be improved using various techniques, e.g., A-
GPS, accelerometers, the Kalman filter, and the moving average filter (as
discussed in Section 2). However, the resulting accuracy of the measurements
will not be perfect. To illustrate this, we evaluate the performance of the
standard Kalman filter applied to the noisy coordinates of the agents. We
chose this particular filter, as it has been used for decades to eliminate GPS
position noise [6].
Details of the analysis are given in A. In short, we observe that the stan-
dard Kalman filter reduces the noise significantly when an agent is following
a monotonic trajectory (which has been observed in the past [37]). However,
the filter has difficulty in detecting an abrupt change in the direction or ve-
locity of the move, which replaces random error with systemic error [18]. We
also show that, when applied to a set of agents, the average position error
(after application of the Kalman filter) ranges from 2.6 to 3.0 m. Moreover,
the Kalman filter amplifies noise when the position errors are small, mak-
ing it impractical for such a case. To conclude, the error after the adoption
of the Kalman filter is still significant and may affect model performance.
Therefore, let us explore how the noise affect the performance of the models.
4 Experimental results
We set up a testbed on our server, which has Ubuntu (4.4.0-103-generic)
system, 56 CPUs, and 512 GB memory. We run simulations in the scene
given in Fig. 2; we use GNU parallel [49] to parallelize the simulations. We
vary the values of the root-mean-squared distance error (E) and the number
of agents (N), choose a sufficient number for repeated realizations (in our
case, n = 1000), and calculate the expected value of probabilities (p), as well
as the 95% confidence interval (C) as follows.
1. For E = 0, 1, 2, . . . , 10 m and N = 1, 5, 10, . . . , 10240 (N increases by
a factor of two, starting from N = 5), repeat 1000 simulations with
16
the same set of {E, N}. For each simulation, return one if stampede
occurs; otherwise, return zero.
2. Upon the completion of simulations with one set of {E, N}, compute
p and C using Eqs. (7) and (8), respectively.
The pseudocode of this process is shown in Algorithm 1.
(cid:46) iterate over R if needed
for E = 0, 1, . . . , 10 do
for N = 1, 5, 10, . . . , 10240 do
Algorithm 1 Pseudocode to execute the simulation
Input: E, N
Output: p, C
1: for each stampede assessment method do
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17: end for
end if
i ← i + 1
until i > 1000
Compute p and C using Eqs. (7) and (8)
end for
end for
Run simulation for a given value of E and N
if stampede occurs then
i ← 1
repeat
else
Return 1
Return 0
4.1 Crowd pressure method
Experimental results of simulated stampede probabilities from the crowd
pressure method are shown in Fig. 7. In this method, the values of the crowd
pressure depend on the choice of R in Eq. (2). Figures 7a, 7b, 7c, and 7d
represent the simulated stampede probabilities when R = 1, 2, 3, and 4 m,
respectively. The threshold of the pressure is set to 0.04 s−2; if we detect
a value of the pressure above the threshold -- we record a stampede. In the
17
(a) R = 1 m
(b) R = 2 m
(c) R = 3 m
(d) R = 4 m
Figure 7: Stampede probabilities in the pressure method. Three-dimensional
bar graph of (N, E, p)-tuples. Stampede probabilities with (a) R = 1 m, (b)
R = 2 m, (c) R = 3 m, and (d) R = 4 m.
experiments, we test the sensitivity of this metric to the changes of R and
N. From Fig. 7, we have four observations as follows.
• The crowd pressure method is very sensitive when R = 1 m. It gener-
ates false positive errors when the number of agents is relatively small:
even with 20 agents, the crowd pressure method provides stampede
prediction with p = 1 (almost surely).
• The crowd pressure method becomes less sensitive when R increases:
the greater the R -- the closer the local density ρ to the average of the
global density. As R grows large enough to cover the whole area which
18
includes all the agents, all local densities will result in the same value,
which corresponds to the overall number of agents divided by the area.
• When the values of R and E are fixed, the value of p tends to in-
crease with the growth of N. However, this is not always the case (as
shown in Figs. 7c and 7d). The underlying reason is that the pedestrian
model -- PedVO -- uses the fundamental diagram to model the correla-
tion between velocities and surroundings. Less neighbors means higher
velocities in the beginning of simulations. Consequently, agents may
have to change velocities dramatically when facing the pillars. As a
result, the variance of local velocity in Eq. (4) increases. Experiments
with different distributions of agents' starting positions have been done
(but not presented here to avoid repetition), where the p value can in-
crease with the value of N monotonically.
• Position errors may affect the p values, especially when 0 < p < 1. The
impact of position errors could be positive or negative. For example,
in Fig. 7c, increased position errors decrease the simulated stampede
probability when N = 40. However, the p value rises and then drops
with the growth of E when N = 80. The reason that the p value
decreases, when E increases from seven to ten, is that the crowd is
spread out when the position errors are introduced (as shown in Figs. 5
and 6). Thus, physical contacts among agents are generally small when
E is large. However, the spreading-out positions may randomly gen-
erate some "hot zones", and this explains why the p value increases in
some cases mentioned above.
To estimate the confidence intervals of simulated stampede probabilities
in the crowd pressure method, we show the confidence intervals calculated
by Eq. (8) with R = 1, 2, 3, and 4 m in Fig. 8. To improve comprehension of
the plots, we choose three representative values of E; namely, E = 0, 5, and
10 m. Besides the observations from Fig. 7, we have two new observations
from Fig. 8 as follows.
• The confidence interval is relatively narrow, which means that the sam-
pling size of the method is reasonable and the estimated stampede prob-
abilities are precise. In other words, we are confident that we would
get a similar result when re-running the simulations.
19
(a) R = 1 m
(b) R = 2 m
(c) R = 3 m
(d) R = 4 m
Figure 8: Simulated stampede probabilities with confidence levels in the
crowd pressure method. Two-dimensional line plots of simulated stampede
probabilities (p) and confidence levels (C). Values of p are presented in
dots connected by lines, while values of C are presented in bars around the
corresponding values of p. Plots of E = 0 m are in blue, plots of E = 5 m
are in red, and plots of E = 10 m are in yellow. The values of p and C with
(a) R = 1 m and N = 1, . . . , 40, (b) R = 2 m and N = 1, . . . , 80, (c) R = 3
m and N = 1, . . . , 320, (d) R = 4 m and N = 1, . . . , 10240.
• Generally speaking, the larger the position errors we observe, the less
accurate the estimated stampede probabilities are. For example, the
two curves of E = 0 m and E = 5 m are close to each other in Fig. 8a,
while the curve of E = 10 m departs from them.
20
4.2 Physical force method
(a) Force method
(b) Force method
Figure 9: Simulated stampede probabilities and the corresponding confidence
levels in the physical force method. (a) Simulated stampede probabilities of
the physical force method in three-dimensional bar graph of (N, E, p)-tuples.
(b) Values of p and C of the physical force method in two-dimensional line
plot.
Let us now present the experimental results for the physical force method.
Note that unlike the previous method, it does not need to set the R. In this
method, the threshold is 4500 N. The simulated stampede probabilities can
be seen in Fig. 9a. We also plot the simulated stampede probabilities and
confidence intervals for the physical force method in Fig. 9b. The figures
suggest the following.
• The physical force method is very sensitive to the values of N in the
simulations. The simulated stampede probabilities reach around 0.45
when N = 40.
• The position errors affect the simulated stampede probabilities of the
physical force method, especially when N = 20.
• The confidence intervals are narrow, which indicates that the expected
stampede probabilities would be consistent with repetitive simulations.
21
(a) R = 1 m
(b) R = 1.5 m
(c) R = 1 m
(d) R = 1.5 m
Figure 10: Simulated stampede probabilities and the corresponding confi-
dence levels in the density method. (a) Simulated stampede probabilities
of the density method with R = 1 m in three-dimensional bar graph of
(N, E, p)-tuples. (b) Simulated stampede probabilities of the density method
with R = 1.5 m in three-dimensional bar graph. (c) Values of p and C of the
density method with R = 1 m in two-dimensional line plot. (d) Values of p
and C of the density method with R = 1.5 m in two-dimensional line plot.
4.3 Density method
Results for estimation of p using the density method are given in Fig. 10.
We vary the value of R, setting it to 1 m and 1.5 m to adjust the method's
sensitivity (as was discussed in Section 4.1). The threshold of density is set
to about 7 m−2. Examining Fig. 10 suggests the following.
22
• The larger the value of R -- the less sensitive the density method.
• Similar to the results in Fig. 7, the position errors may affect the value
of p. For example, Figure 10a suggests that the number of agents N is
equal to 320, the p value varies from 8% to 32.7% for different values
of E.
• As for the other models, the confidence intervals are narrow, indicating
high precision and reproducibility of the simulations.
4.4 Comparison
Comparing all three stampede assessment methods, we can arrive at the
following conclusions.
• All the three stampede assessment methods can be largely affected by
changes to their corresponding thresholds, i.e., p = 0.04 s−2 for the
crowd pressure method, N = 4500 N for the physical force method,
and D ≈ 7 m−2 for the density method. Moreover, the results of the
crowd pressure method and the density method can be affected by
changing to the value of R.
• The simulated stampede probabilities p can be affected by position er-
rors (especially in the crowd pressure method and the density method)
if the number of agents is not very small or very large. In the former
case p = 0 and in the latter p = 1, almost certainly.
• In general, the value of p increases with the number of agents.
• In all the experiments, the confidence intervals are narrow, suggesting
high precision and reproducibility of the simulations.
To sum up, the results of the stampede prediction can be affected by the
position errors E, as well as some input parameters for stampede assessments,
e.g., R. Thus, these factors need to be taken into account when implementing
stampede measuring techniques. A practitioner can calibrate and validate
the model with the empirical data gathered from the agents.
23
5 Conclusions and future work
To conclude, in this work we analyze the impact of position errors on the ac-
curacy and precision of stampede prediction. To achieve this goal, we propose
an automatic real-time method, called ARES, for stampede prediction. We
implement a prototype of ARES using Menge simulation framework. Then,
we add position errors to three different stampede assessment methods. We
compute the corresponding probabilities of the simulated stampede and com-
pare those results. Finally, we analyze the impact of position errors. Exper-
imental results show that the position errors can change the probability of
simulated stampede significantly, even when noise reduction technique, such
as the standard Kalman filter, is applied. This implies that analysis of the
impact of position errors on crowd simulation is crucial, especially when the
GPS signal interference is pronounced. Thus, we argue that more research
is needed to take position errors into account when dealing with stampede
assessments. The results in this paper can be seen as a starting point to
control the position errors in stampede assessments. Future research will
focus on exploring potential approaches for mitigating position errors as well
as exploring the distributions of position errors from hybrid sources of data
(e.g., video cameras and GPS) into this framework, to improve the accuracy
of the solution for stampede prediction.
Acknowledgements
The work reported in this paper is supported and funded by Ontario Centres
of Excellence, Natural Sciences and Engineering Research Council of Canada,
and Laipac Technology Inc. We would like to thank Sheik Hoque for per-
forming initial tests on Menge. Last but not least, we would also like to
thank the editor and the anonymous reviewers for their valuable suggestions
and comments that improved the quality of this paper.
Availability of data
Datasets of experimental results are available on Mendeley Data [67] via
http://dx.doi.org/10.17632/tz7gmrzffh.1
24
A Noise reduction using Kalman filter
In order to test if the Kalman filter could improve the tracking accuracy, we
apply a standard Kalman filter to our simulations. The venue map is the
same as shown in Fig. 2, and the pedestrian model is the same as described
in Section 3.1. We compare filtered estimates to noisy positions (which corre-
spond to measured positions of GPS devices) and true positions. In our case,
the state vector of the Kalman filter is (xi,s, yi,s, vx
i,s), where (xi,s, yi,s) is
the coordinate of agent i at time s, and (vx
i,s) is the velocity vector of
agent i at time s.
i,s, vy
i,s, vy
We explore the performance of the Kalman filter in the analysis of a set
of agents in A.1 followed by the analysis of two individual agents in A.2. We
summarize our findings in A.3.
Algorithm 2 Pseudocode to estimate performance of the standard Kalman
Filter
Input: N, S, E
Output: MAE, KDE
1: Simulate and store true positions for each agent at every step, while
N = 10240 and S = 150
2: for E = 0, 1, . . . , 10 do
3:
4:
5:
for each agent do
coordinate at the last step
Add noise to the position at each step
Feed these noisy positions to the Kalman filter and estimate the
6:
Compute the error of the noisy (measured) position and the error
of the estimated position by comparing to the true position
end for
Compute the MAEs and plot the KDEs
7:
8:
9: end for
A.1 Estimation of positions for a set of agents
To explore behaviour of a set of agents, we perform the following set of
experiments summarized by the pseudocode in Algorithm 2. Details of the
experiments are given below.
25
We set the number of agents to N = 10240 and the simulation time
horizon, deemed S, to S = 150 s. Each simulation step takes 0.1 s, which
means that each agent has 1500 simulation steps. For each agent, we collect
and save the position and velocity in each simulation step. We choose 1500
simulation steps, because 1) it provides sufficient prior observations for the
Kalman filter, and 2) it spreads agents over the bridge in the end of the
simulation (some agents have already left the bridge while others have just
reached the first pillar).
Once the simulation is complete, we perform ten experiments (described
below), repeated for ten values of E = 1, 2, . . . , 10 m. For each value of E, we
take the actual positions of the agents and add the noise to the coordinates.
The noise is drawn from the Rayleigh distribution (see Eq. (6), where the
amount of noise is controlled by the value of E. The velocities are kept
accurate. We then apply the standard Kalman filter to the noisy trajectories
(and true velocities) of each agent to estimate the final position of each agent
at the last step (i.e., when S = 150 s).
After that, for each agent, we compute the absolute errors 1) between the
actual and the noisy position at the last step, deemed measured error and 2)
between the actual and the estimated (by the Kalman filter) position at the
last step, deemed estimated error.
E Estimated error Measured error
0.89
1
1.76
2
3
2.66
3.59
4
4.44
5
6
5.30
6.23
7
7.19
8
8.00
9
10
8.83
2.68
2.71
2.74
2.77
2.80
2.83
2.86
2.90
2.93
2.96
Table 1: MAEs of measured and estimated errors for E = 1, 2, . . . , 10 m.
The MAEs are measured in metres.
Finally, we compute the mean absolute errors (MAEs), shown in Table 1.
We also plot kernel density estimates (KDEs) of the measured and estimated
26
(a) E = 1 m
(b) E = 4 m
(c) E = 7 m
(d) E = 10 m
Figure 11: KDEs with E = 1, 4, 7, and 10 m. The x-axis represents the
distance errors in metres, and the y-axis represents the probability density.
The rose shaded area represents the KDEs of the Kalman filter estimates,
and the mint shaded area represents the KDEs of the noisy positions.
errors. To preserve space, in Figure 11, we show KDEs for four representative
values of E, namely, E ∈ {1, 4, 7, 10}.
To compute MAEs, we calculate Euclidean distance errors for each agent's
final coordinates and then compute the average of the Euclidean distances
of all the agents. Table 1 shows the MAEs estimated by the Kalman filter
as well as the amount of error added to the actual coordinates. The Kalman
filter yields an error, ranging, on average, between 2.6 and 3.0 m. The amount
of the estimated error increases monotonically with the growth of E. As was
shown in Figs. 7, 9, and 10, the stampede probabilities will be affected by
this amount of noise.
The measured error is smaller than the one yielded by the Kalman filter
when E < 4 m. Essentially, one should not apply the Kalman filter when
E < 4 m, as it will decrease the accuracy of the estimated coordinates. It
27
may be beneficial to use the filter when E ≥ 4 m. However, an error of ≈ 3
m may lead to significant variation in the positions of the agents relative
to each other, reducing the performance of the stampede prediction models.
Moreover, KDEs plotted in Figure 11 exhibit heavy right tails: while an
average error is ≈ 3 m, some agents will have a position error greater than
20 m.
A.2 Estimation of positions for two individual agents
Why the Kalman filter cannot reduce the MAE below ≈ 3 m (as shown
in Table 1)? It has been observed in the past that the Kalman filter has
difficulty detecting non-monotonic trajectories [18]. We conjecture that the
filter faces this challenge in our experiments too. Too support this conjecture,
let us take a look at the trajectories of two agents from 10240 agents that we
simulated in A.1. We then add noise to the trajectories of the agents using
Eq. (6) and the setting E = 1 m. These trajectories are shown in Fig. 12.
In Fig. 12a, the agent is slowly approaching the first pillar (moving from
right to left), because a lot of agents (which are in front of this agent) block
the agent's path. We can see that the agent's trajectory is monotone. In ad-
dition, the trajectory estimated by the Kalman filter and the true trajectory
starts overlapping from the coordinate around (180, -9.5).
However, the Kalman filter fails to detect abrupt state changes in agent's
trajectory; Figure 12b depicts such a case. The agent (while moving from
right to left) stops around the first two pillars near the coordinates (-10, -6)
and (-30, -6) (changing the velocity). The filter has difficulty detecting this
change and predicts that the agent keeps moving in a straight line. The agent
also changes the direction near the first pillar, which the filter also does not
detect. Thus, the estimates of the Kalman filter have noticeable distances
from the true trajectory.
A.3 Summary
To conclude, the standard Kalman filter may help to improve the positional
accuracy of GPS when the position errors are high (E ≥ 4 m). However,
the stampede probabilities will still be affected by the amount of position
error (≈ 3 m) retained by the filter. This level of noise may be explained
by the difficulty the filter exhibits while dealing with changing movement
patterns. To address this limitation, possible solutions have been proposed
28
(a) An agent is slowly approaching the first pillar.
(b) An agent stopped at the first two pillars, and is approaching the third one.
Figure 12: Examples of Kalman filter performance when E = 1 m. One
coordinate unit is equal to one metre.
in the literature by extending the standard Kalman filter [37] or integrating
other techniques, e.g., mean shift algorithm [31]. We leave the investigation
29
of the applicability of such solutions to future researchers.
B The PedVO model
Parameter
factor
buffer
tau
tauObst
turningBias
density_aware Controls if the preferred speed is adjusted to density (true) or not (false)
Description
The factor capturing the relationship between speed and stride length
The psychological stride buffer required beyond that needed for stride length
The default time horizon for predicting inter-agent collisions
The default time horizon for predicting agent-obstacle collisions
The default turn bias
Value
1.57
0.9
3.0
0.1
1.0
false
Table 2: Default parameters of PedVO adopted in our experiments.
As mentioned in Section 3.1, we use the fundamental diagram as the
velocity modifier, which represents the relationship between speed and den-
sity. The pedestrian model adopted, namely, PedVO [11], also captures the
underlying physiological and psychological factors besides the fundamental
diagram. In our experiments, we use the default settings of PedVO in the
fundamental diagram example provided by Menge. These settings are shown
in Table 2.
References
[1] Al Arabiya News. Saudi Arabia may soon issue e-bracelets for all hajj
pilgrims, 2015. [Online; accessed 1-July-2018].
[2] J. E. Almeida, R. Rosseti, and A. L. Coelho. Crowd simulation model-
ing applied to emergency and evacuation simulations using multi-agent
systems. In Proceedings of the 6th Doctoral Symposium on Informatics
Engineering, pages 93 -- 104. FEUP, 2011.
[3] F. Alonso-Marroquin, J. Busch, C. Chiew, C. Lozano, and Á. Ramírez-
Simulation of counterflow pedestrian dynamics using
Gómez.
spheropolygons. Physical Review E, 90(6):063305, 2014.
[4] BBC News. Shanghai new year crush kills 36, 2015.
1-July-2018].
[Online; accessed
30
[5] E. Bostanci, B. Bostanci, N. Kanwal, and A. F. Clark. Sensor fusion
of camera, GPS and IMU using fuzzy adaptive multiple motion models.
Soft Computing, 22(8):2619 -- 2632, 2018.
[6] R. G. Brown, P. Y. Hwang, et al. Introduction to random signals and
applied Kalman filtering, volume 3. Wiley New York, 1992.
[7] R. C. Browning, E. A. Baker, J. A. Herron, and R. Kram. Effects of
obesity and sex on the energetic cost and preferred speed of walking.
Journal of Applied Physiology, 100(2):390 -- 398, 2006.
[8] M. L. Callisaya, C. P. Launay, V. K. Srikanth, J. Verghese, G. Allali,
and O. Beauchet. Cognitive status, fast walking speed and walking
speed reserve-the gait and alzheimer interactions tracking (GAIT) study.
Geroscience, 39(2):231 -- 239, 2017.
[9] A. Corbetta, J. Meeusen, C. Lee, and F. Toschi. Continuous measure-
ments of real-life bidirectional pedestrian flows on a wide walkway. In
Proceedings of Pedestrian and Evacuation Dynamics 2016, pages 18 -- 24,
2016.
[10] S. Curtis, A. Best, and D. Manocha. Menge: A modular framework for
simulating crowd movement. Collective Dynamics, 1:1 -- 40, 2016.
[11] S. Curtis and D. Manocha. Pedestrian simulation using geometric rea-
soning in velocity space. In U. Weidmann, U. Kirsch, and M. Schrecken-
berg, editors, Proceedings of Pedestrian and Evacuation Dynamics 2012,
pages 875 -- 890, Switzerland, 2014. Springer.
[12] M. H. Dridi. List parameters influencing the pedestrian movement and
pedestrian database. International Journal of Social Science Studies,
3(4):94 -- 106, 2015.
[13] M. Eliasson. A Kalman filter approach to reduce position error for
pedestrian applications in areas of bad GPS reception, 2014.
[14] D. Elliott and D. Smith. Football stadia disasters in the United King-
dom: learning from tragedy? Industrial & Environmental Crisis Quar-
terly, 7(3):205 -- 229, 1993.
31
[15] T. Franke, P. Lukowicz, and U. Blanke. Smart crowds in smart cities:
Real life, city scale deployments of a smartphone based participatory
crowd management platform. Journal of Internet Services and Applica-
tions, 6(1):27, 2015.
[16] R. Garnett and R. Stewart. Comparison of GPS units and mobile Apple
GPS capabilities in an urban landscape. Cartography and Geographic
Information Science, 42(1):1 -- 8, 2015.
[17] P. D. Groves, L. Wang, D. Walter, H. Martin, K. Voutsis, and Z. Jiang.
The four key challenges of advanced multisensor navigation and po-
sitioning.
In Proceedings of 2014 IEEE/ION Position, Location and
Navigation Symposium (PLANS), pages 773 -- 792. IEEE, 2014.
[18] F. Gustafsson and F. Gustafsson. Adaptive filtering and change detec-
tion, volume 1. Citeseer, 2000.
[19] D. Helbing, I. J. Farkas, P. Molnar, and T. Vicsek. Simulation of pedes-
trian crowds in normal and evacuation situations. Pedestrian and evac-
uation dynamics, 21(2):21 -- 58, 2002.
[20] D. Helbing, A. Johansson, and H. Z. Al-Abideen. Dynamics of crowd
disasters: An empirical study. Physical review E, 75(4), 2007.
[21] D. Helbing, A. Johansson, and H. Z. Al-Abideen. The dynamics of
crowd disasters: An empirical study (supplementary information), 2007.
[Online; accessed 7-February-2018].
[22] A. Johansson, D. Helbing, H. Z. Al-Abideen, and S. Al-Bosta. From
crowd dynamics to crowd safety: A video-based analysis. Advances in
Complex Systems, 11(04):497 -- 527, 2008.
[23] A. Johansson, D. Helbing, and P. K. Shukla. Specification of the social
force pedestrian model by evolutionary adjustment to video tracking
data. Advances in complex systems, 10(supp02):271 -- 288, 2007.
[24] J. C. S. J. Junior, S. R. Musse, and C. R. Jung. Crowd analysis
using computer vision techniques. IEEE Signal Processing Magazine,
27(5):66 -- 77, 2010.
32
[25] B. Krausz and C. Bauckhage. Loveparade 2010: Automatic video anal-
ysis of a crowd disaster. Computer Vision and Image Understanding,
116(3):307 -- 319, 2012.
[26] A. V. Kurilkin and S. V. Ivanov. A comparison of methods to detect
people flow using video processing. Procedia Computer Science, 101:125 --
134, 2016.
[27] L. J. Launer and T. Harris. Weight, height and body mass index distri-
butions in geographically and ethnically diverse samples of older persons.
Age and Ageing, 25(4):300 -- 306, 1996.
[28] R. S. Lee and R. L. Hughes. Exploring trampling and crushing in a
crowd. Journal of transportation engineering, 131(8):575 -- 582, 2005.
[29] R. S. Lee and R. L. Hughes. Prediction of human crowd pressures.
Accident analysis & prevention, 38(4):712 -- 722, 2006.
[30] J. Li, K. Miyashita, T. Kato, and S. Miyazaki. GPS time series modeling
by autoregressive moving average method: Application to the crustal
deformation in central Japan. Earth, planets and space, 52(3):155 -- 162,
2000.
[31] Y. Li, Y. Pang, Z. Li, and Y. Liu. An intelligent tracking technology
based on Kalman and mean shift algorithm. In Proceedings of the 2nd
International Conference on Computer Modeling and Simulation (IC-
CMS), volume 1, pages 107 -- 109. IEEE, 2010.
[32] D. Littlefield. Metric handbook: Planning and design data. Routledge,
2012.
[33] J. Liu, B. Priyantha, T. Hart, H. S. Ramos, A. A. Loureiro, and
Q. Wang. Energy efficient GPS sensing with cloud offloading. In Proceed-
ings of the 10th ACM Conference on Embedded Network Sensor Systems,
pages 85 -- 98. ACM, 2012.
[34] I. Mahmood, M. Haris, and H. Sarjoughian. Analyzing emergency evacu-
ation strategies for mass gatherings using crowd simulation and analysis
framework: Hajj scenario.
In Proceedings of the 2017 ACM SIGSIM
Conference on Principles of Advanced Discrete Simulation, pages 231 --
240. ACM, 2017.
33
[35] A. N. Marana, L. D. F. Costa, R. Lotufo, and S. A. Velastin. Estimat-
ing crowd density with Minkowski fractal dimension. In Proceedings of
1999 IEEE International Conference on Acoustics, Speech, and Signal
Processing, pages 3521 -- 3524. IEEE, 1999.
[36] H. Mousavi, S. Mohammadi, A. Perina, R. Chellali, and V. Mur. Ana-
lyzing tracklets for the detection of abnormal crowd behavior. In Pro-
ceedings of 2015 IEEE Winter Conference on Applications of Computer
Vision (WACV), pages 148 -- 155. IEEE, 2015.
[37] B. Moussakhani, J. T. Flåm, T. A. Ramstad, and I. Balasingham. On
change detection in a Kalman filter based tracking problem. Signal
processing, 105:268 -- 276, 2014.
[38] Newzoo. Top countries/markets by smartphone penetration & users,
2018. [Online; accessed 15-August-2018].
[39] Official U.S. government information about the Global Positioning Sys-
[Online; accessed
tem (GPS) and related topics. GPS accuracy, 2017.
15-August-2018].
[40] A. Papoulis and S. U. Pillai. Probability, random variables, and stochas-
tic processes. Tata McGraw-Hill Education, New York, 4th edition,
2002.
[41] M. PETOVELLO. How does a GNSS receiver estimate velocity? Inside
GNSS, pages 38 -- 41, 2015.
[42] M. Sabokrou, M. Fayyaz, M. Fathy, and R. Klette. Deep -- cascade: Cas-
cading 3D deep neural networks for fast anomaly detection and local-
ization in crowded scenes.
IEEE Transactions on Image Processing,
26(4):1992 -- 2004, 2017.
[43] M. Sabokrou, M. Fayyaz, M. Fathy, Z. Moayed, and R. Klette. Deep-
anomaly: Fully convolutional neural network for fast anomaly detection
in crowded scenes. Computer Vision and Image Understanding, 2018.
[44] D. Sharma, A. P. Bhondekar, A. Shukla, and C. Ghanshyam. A re-
view on technological advancements in crowd management. Journal of
Ambient Intelligence and Humanized Computing, 9(3):1 -- 11, 2016.
34
[45] Statista. Number of smartphone users worldwide 2014 - 2020, 2016.
[Online; accessed 15-August-2018].
[46] B. Steffen and A. Seyfried. Methods for measuring pedestrian density,
flow, speed and direction with minimal scatter. Physica A: Statistical
mechanics and its applications, 389(9):1902 -- 1910, 2010.
[47] D. Stojanović and N. Stojanović.
Indoor localization and tracking:
Methods, technologies and research challenges. Facta Universitatis, Se-
ries: Automatic Control and Robotics, 13(1):57 -- 72, 2014.
[48] T. Suzuki, M. Kitamura, Y. Amano, and T. Hashizume. High-accuracy
GPS and GLONASS positioning by multipath mitigation using omni-
directional infrared camera. In Proceedings of 2011 IEEE International
Conference on Robotics and Automation (ICRA), pages 311 -- 316. IEEE,
2011.
[49] O. Tange. GNU parallel 2018, 2018.
[50] The Express Tribune. Hajj stampede death toll rises to 2,177, 2015.
[Online; accessed 1-July-2018].
[51] The New York Times. Stampede at Mumbai railway station kills at least
22, 2017. [Online; accessed 1-July-2018].
[52] The Telegraph News. More than 1,500 Juventus fans injured in stampede
in Turin, 2017. [Online; accessed 1-July-2018].
[53] A. Tordeux, J. Zhang, B. Steffen, and A. Seyfried. Quantitative compar-
ison of estimations for the density within pedestrian streams. Journal of
Statistical Mechanics: Theory and Experiment, 2015(6):P06030, 2015.
[54] UCSD. Anomaly detection dataset, 2016. [Online; accessed 15-August-
2018].
[55] J. Van Den Berg, S. J. Guy, M. Lin, and D. Manocha. Reciprocal n-body
collision avoidance. In Robotics research, pages 3 -- 19. Springer, 2011.
[56] F. S. T. Van Diggelen. A-GPS: Assisted GPS, GNSS, and SBAS. Artech
House, 2009.
35
[57] S. C. Walpole, D. Prieto-Merino, P. Edwards, J. Cleland, G. Stevens,
and I. Roberts. The weight of nations: An estimation of adult human
biomass. BMC public health, 12(1):439, 2012.
[58] F. Wang, X. Zhang, and J. Huang. Error analysis and accuracy assess-
ment of GPS absolute velocity determination without SA. Geo-spatial
Information Science, 11(2):133 -- 138, 2008.
[59] D. L. Wilson. GPS horizontal position accuracy. Global Positioning
System Accuracy webpage, 2002.
[60] M. Wirz, T. Franke, D. Roggen, E. Mitleton-Kelly, P. Lukowicz, and
G. Tröster. Inferring crowd conditions from pedestrians' location traces
for real-time crowd monitoring during city-scale mass gatherings. In Pro-
ceedings of 2012 IEEE 21st International Workshop on Enabling Tech-
nologies: Infrastructure for Collaborative Enterprises (WETICE), pages
367 -- 372. IEEE, 2012.
[61] S. Yamaguchi and T. Tanaka. GPS standard positioning using Kalman
In Proceedings of 2006 International Joint Conference SICE-
filter.
ICASE, pages 1351 -- 1354. IEEE, 2006.
[62] R. K. Yin. Case study research: Design and methods. Sage publications,
5th edition, 2013.
[63] P. A. Zandbergen. Positional accuracy of spatial data: Non-normal
distributions and a critique of the national standard for spatial data
accuracy. Transactions in GIS, 12(1):103 -- 130, 2008.
[64] P. A. Zandbergen and S. J. Barbeau. Positional accuracy of assisted GPS
data from high-sensitivity GPS-enabled mobile phones. The Journal of
Navigation, 64(3):381 -- 399, 2011.
[65] J. Zębala, P. Ciępka, and A. RezA. Pedestrian acceleration and speeds.
Probl. Forensic Sci., 91:227 -- 234, 2012.
[66] A. Zelenkov, A. Kluga, and E. Grab. Accuracy estimation of GPS re-
ceiver parameters with re-reference system in static mode. Telecommu-
nications & Electronics, 8(31):31 -- 36, 2008.
36
[67] L. Zhang, D. Lai, and A. Miranskyy. Datasets for the impact of position
errors on crowd simulation. Mendeley Data, 2018.
[68] J. Zhou, H. Pei, and H. Wu. Early warning of human crowds based on
query data from Baidu maps: Analysis based on Shanghai stampede.
In Big Data Support of Urban Planning and Management, pages 19 -- 41.
Springer, Cham, 2017.
[69] S. Zhou, W. Shen, D. Zeng, M. Fang, Y. Wei, and Z. Zhang. Spatial --
temporal convolutional neural networks for anomaly detection and lo-
calization in crowded scenes. Signal Processing: Image Communication,
47:358 -- 368, 2016.
37
|
1903.01632 | 2 | 1903 | 2019-11-21T15:35:23 | Demonstration of a Time-Efficient Mobility System Using a Scaled Smart City | [
"cs.MA"
] | The implementation of connected and automated vehicle (CAV) technologies enables a novel computational framework to deliver real-time control actions that optimize travel time, energy, and safety. Hardware is an integral part of any practical implementation of CAVs, and as such, it should be incorporated in any validation method. However, high costs associated with full scale, field testing of CAVs have proven to be a significant barrier. In this paper, we present the implementation of a decentralized control framework, which was developed previously, in a scaled-city using robotic CAVs, and discuss the implications of CAVs on travel time. Supplemental information and videos can be found at https://sites.google.com/view/ud-ids-lab/tfms. | cs.MA | cs |
Demonstration of a Time-Efficient Mobility System Using a Scaled
Smart City
L. E. Beaver, B. Chalaki, A. M. I. Mahbub, L. Zhao, R. Zayas, A. A. Malikopoulos
University of Delaware, Department of Mechanical Engineering, Newark, DE, USA 19716
ARTICLE HISTORY
Compiled November 22, 2019
ABSTRACT
The implementation of connected and automated vehicle (CAV) technologies en-
ables a novel computational framework to deliver real-time control actions that
optimize travel time, energy, and safety. Hardware is an integral part of any practi-
cal implementation of CAVs, and as such, it should be incorporated in any val-
idation method. However, high costs associated with full scale, field testing of
CAVs have proven to be a significant barrier. In this paper, we present the im-
plementation of a decentralized control framework, which was developed previ-
ously, in a scaled-city using robotic CAVs, and discuss the implications of CAVs
on travel time. Supplemental
information and videos can be found at https:
//sites.google.com/view/ud-ids-lab/tfms.
KEYWORDS
Connected and automated vehicles; optimal control; emerging mobility systems;
smart city; scaled city.
1. Introduction
Connectivity and automation provide the most intriguing opportunity for enabling
users to better monitor transportation network conditions and make better operating
decisions to reduce energy consumption, greenhouse gas emissions, travel delays, and
improve safety. In the context of a smart city, wireless connectivity provides free-flow
of information among entities, while automation provides precise execution upon such
available information for moving goods and people safely and efficiently (Fig. 1). The
availability of vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communi-
cation has the potential to ease congestion and improve safety by enabling vehicles
to respond rapidly to changes in their mutual environment. Furthermore, vehicle au-
tomation technologies can aim at developing robust vehicle control systems that can
quickly respond to dynamic traffic operating conditions.
With the advent of emerging information and communication technologies, we are
witnessing a massive increase in the integration of our energy, transportation, and
cyber networks. These advances, coupled with human factors, are giving rise to a new
level of complexity in transportation networks [1]. As we move to increasingly com-
plex emerging transportation systems, with changing landscapes enabled by connectiv-
ity and automation, future transportation networks could shift dramatically with the
CONTACT A. A. Malikopoulos. Email: [email protected]
Figure 1. A city enabled by connectivity and automation technologies.
large-scale deployment of connected and automated vehicles (CAVs). On the one hand,
with the generation of massive amounts of data from vehicles and infrastructure, there
are opportunities to develop optimization methods to identify and realize a substantial
energy reduction of the transportation network, and to optimize the large-scale sys-
tem behavior using the interplay among vehicles. On the other hand, evaluation and
validation of new control approaches under different traffic scenarios is a necessity to
ensure successful implementation per vehicle alongside desired system-level outcomes.
The overarching goal of this paper is the experimental demonstration of a decentral-
ized control framework for CAVs presented in [2] using the University of Delaware's
Scaled Smart City (UDSSC). UDSSC is a 1:25 scaled testbed representing an urban
environment with robotic CAVs that can replicate real-world traffic scenarios in a
controlled environment (Fig. 2). UDSSC can be used to explore the acquisition and
processing of V2V and V2I communication. It can also be used to validate control al-
gorithms for CAV coordination in specific transportation segments, e.g., intersections,
merging roadways, and roundabouts, by mitigating the high costs and safety concerns
associated with real-world field testing of CAVs. As an intermediate scale testbed, the
UDSSC is an ideal platform to gain insight into the execution of high-level planning
and coordination on physical hardware with noise, disturbances, and communication
delays. In this paper, our emphasis is on the generation of energy-optimal trajectories,
and as such, we do not consider the problem of describing low-level controllers which
track the optimal trajectories.
The structure of the paper is organized as follows. In Section 2, we discuss related
work on the optimal control for CAVs reported in the literature. In Section 3, we
review the decentralized control framework presented in [2] for the coordination of
CAVs in a transportation network. Then, we describe briefly the UDSSC testbed in
Section 4, and present simulation and experimental results in Section 5. Finally, we
draw concluding remarks from the experiments in Section 6.
2
Figure 2. The University of Delaware's Scaled Smart City.
2. Related Work
CAVs have attracted considerable attention for the potential of improving mobility
and safety along with energy and emission reduction [3,4]. There have been two ma-
jor approaches to utilizing connectivity and automation to improve transportation
efficiency and safety, namely, platooning and traffic smoothing.
The first approach utilizes connectivity and automation to form closely-coupled ve-
hicular platoons to reduce aerodynamic drag effectively, especially at high cruising
speeds. The concept of forming platoons of vehicles was a popular system-level ap-
proach to address traffic congestion, which gained momentum in the 1980s and 1990s
[5,6]. Such automated transportation system can alleviate congestion, reduce energy
use and emissions, and improve safety while increasing throughput significantly. The
Japan ITS Energy Project [7], the Safe Road Trains for the Environment program [8],
and the California Partner for Advanced Transportation Technology [9] are among the
mostly-reported efforts in this area.
The second approach is to smooth the traffic flow by centralized or decentralized ve-
hicle control to reduce spatial and temporal speed variation and braking events, e.g.,
automated intersection crossing [10 -- 14], cooperative merging [2,15,16], and speed har-
monization through optimal vehicle control [17]. In centralized approaches, there is at
least one task in the system that is globally decided for all vehicles by a single central
controller, whereas in decentralized approaches, the vehicles are treated as autonomous
agents that collect traffic information to optimize their specific performance criteria
while satisfying physical constraints. One of the very early efforts in this direction
was proposed by Athans [18] for safe and efficient coordination of merging maneuvers
with the intention to avoid congestion. Since then, numerous approaches have been
proposed on coordinating CAVs to improve traffic flow [19 -- 21], and to achieve safe and
efficient control of traffic through various traffic bottlenecks where potential vehicle
collisions may happen [22 -- 31]. In terms of energy impact, many studies have shown
that significant fuel consumption savings could be achieved through eco-driving and
vehicle optimal control without sacrificing driver safety [2,12,16,32 -- 34]. Considering
near-future CAV deployment, recent research work has also explored both traffic and
energy implications of partial penetration of CAVs under different transportation sce-
3
narios, e.g., [35 -- 37]. Several survey papers that report the research efforts in this area
can be found in [38 -- 40].
Although previous work has shown promising results emphasizing the potential
benefits of coordination of CAVs, validation has been primarily in simulation. Some
progress has been made with constructing experimental testbeds, such as MIT's Duck-
ietown [41], which focuses primarily on local perception and autonomy, and the Cam-
bridge Minicars [42], which is a testbed for cooperative driving in highway conditions.
In contrast, the UDSSC focuses on traffic coordination in an urban system, where a
majority of stop-and-go driving occurs. In previous work, we presented the experi-
mental validation of the solution to the unconstrained merging roadway problem in
UDSSC using 10 robotic CAVs [43]. In this paper, we demonstrate the impact of an
optimal decentralized framework, developed in earlier work [2], for coordinating CAVs
in a transportation network with multiple conflict zones where a lateral collision may
occur.
3. Decentralized Control Framework
We consider a network of CAVs driving in the roadway network in UDSSC, which
consists of several conflict zones, e.g., ramps, roundabouts, and intersections, where
lateral collisions may occur (marked with red boxes in Fig. 3). For each conflict zone,
there is a coordinator that communicates with all CAVs traveling within its com-
munication range. In practice, the coordinator can be stationary roadside units for
general-purpose traffic monitoring and message dissemination or mobile roadside units
(e.g., transit vehicles, drones), which can provide dynamic traffic monitoring and com-
munication support at specific locations or along corridors. Each CAV is retrofitted
with a communication device necessary to interact with other vehicles and local in-
frastructure within their communication range. Upstream of a conflict zone, we define
a control zone in each direction, inside of which, the CAVs coordinate with each other
in order to travel through the conflict zone without any collisions (Fig. 3). The length
of the control zone can be considered to be the maximum communication range of a
V2X device. Outside the control zone, the CAVs behave as human-driven vehicles. For
simplicity, we do not consider multi-lane trajectories or any lane changes within the
control zone. This is the focus of ongoing work [44], and it is not discussed here.
3.1. Modeling Framework and Assumptions
Let z ∈ Z be the index of a conflict zone in the corridor. Let Nz(t) = {1, 2, ..., N (t)} be
a queue of CAVs to be analyzed corresponding to the conflict zone z, where N (t) ∈ N is
the total number of CAVs at the time t ∈ R+. The dynamics of each vehicle i ∈ Nz(t),
are represented with a state equation
x(t) = f (t, xi, ui), xi(tz,0
i ) = xz,0
i
,
(1)
where xi(t), ui(t) are the state of the vehicle and control input, tz,0
is the initial time
of vehicle i ∈ Nz(t) entering the control zone corresponding to the conflict zone z,
and xz,0
is the value of the initial state. For simplicity, we model each vehicle as a
pi = vi(t) and vi = ui(t), where pi(t) ∈ Pi, vi(t) ∈ Vi, and
double integrator, i.e.,
ui(t) ∈ Ui denote the position, speed, and acceleration/deceleration (control input) of
i
i
4
Figure 3. Vehicle routes in the University of Delaware's Scaled Smart City environment.
each vehicle i. Let xi(t) = [pi(t) vi(t)]T denote the state of each vehicle i ∈ Nz(t), with
(t)]T , taking values in the state space Xi = Pi × Vi. The
initial value xz,0
sets Pi,Vi, and Ui, i ∈ N (t), are complete and totally bounded subsets of R. The state
space Xi for each vehicle i is closed with respect to the induced topology on Pi × Vi
and thus, it is compact.
i (t) = [0 vz,0
i
To ensure that the control input and vehicle speed are within a given admissible
range, we impose the constraints
umin ≤ ui(t) ≤ umax, and
0 ≤ vmin ≤ vi(t) ≤ vmax, ∀t ∈ [tz,0
i
, tz,f
i
],
(2)
where umin, umax are the minimum deceleration and maximum acceleration respec-
tively, vmin, vmax are the minimum and maximum speed limits respectively, and tz,0
,
tz,f
i
are the times that each vehicle i enters and exits the conflict zone z.
To avoid a rear-end collision between two consecutive vehicles traveling in the same
lane, the position of the preceding vehicle should be greater than, or equal to the posi-
tion of the following vehicle plus a predefined safe distance δi(t), which is proportional
to the speed of vehicle i, vi(t). Thus, we impose the rear-end safety constraint
i
si(t) = pk(t) − pi(t) ≥ δi(t),
∀t ∈ [tz,0
i
, tz,f
i
],
(3)
where vehicle k is immediately ahead of i on the same lane. The minimum safe distance,
δi(t), is a function of speed, namely
δi(t) = γi + h · vi(t),
(4)
where γi is the standstill distance and h is the minimum safe time gap that CAV i can
5
maintain while following another vehicle.
Definition 1. For each CAV i ∈ Nz(t) approaching a conflict zone z ∈ Z, we define
two subsets of Nz(t) depending on the physical location of i inside the control zone:
1) Lz
i (t) contains all CAVs traveling in the same road and lane as CAV i, which may
cause rear-end collision with CAV i, and 2) Cz
i (t) contains all CAVs traveling on a
different road from i that can cause a lateral collision inside the conflict zone z.
Definition 2. For each vehicle i ∈ Nz(t), we define the set Γz
i that consists of the
positions along the lane where a lateral collision is possible in a conflict zone z, namely
(cid:44)(cid:110)
Γz
i
pi(t) t ∈ [tz,m
i
(cid:111)
, tz,f
i
]
,
(5)
where tz,m
z), and tz,f
i
i
is the time that vehicle i exits the control zone (and enters the conflict zone
is the time that vehicle i exits the conflict zone z.
Consequently, to avoid a lateral collision for any two vehicles i, j ∈ Nz(t) on different
roads we impose the following constraint
j ∈ Cz
i ∩ Γz
Γz
j = ∅,
i (t),
∀t ∈ [tz,m
i
, tz,f
i
].
(6)
The above constraint implies that only one vehicle at a time can be inside the
conflict zone z when there is a potential for a lateral collision. If the length of the
conflict zone is long, then this constraint may dissipate the capacity of the road.
However, the constraint is not restrictive in the problem formulation, and it can be
modified appropriately.
In the modeling framework described above, we impose the following assumptions:
Assumption 1. For each CAV i, none of the constraints is active at tz,0
zone z.
i
for conflict
Assumption 2. Each CAV i can communicate with the coordinator and other CAVs
to receive local information without errors or delays.
The first assumption ensures that the initial state and control input are feasible.
The second assumption might be strong, but it is relatively straightforward to relax
as long as the noise in the measurements and/or delays is bounded. For example,
we can determine upper bounds on the state uncertainties as a result of sensing or
communication errors and delays, and incorporate these into more conservative safety
constraints.
3.2. Communication Structure
When a CAV i ∈ Nz(t) enters a control zone, it communicates with a coordinator,
assigned to the corresponding conflict zone, and the other CAVs inside the control
zone. Note that the coordinator is not involved in any decision for the CAVs, and
it only facilitates the communication of appropriate information among vehicles. The
coordinator handles the information between the vehicles as follows. When a vehicle
enters the control zone of the conflict zone z at time t, the coordinator assigns a unique
identity i = N (t) + 1, which is an integer representing the location of the vehicle in
6
(cid:111)
(cid:111)
(cid:41)
a first-in-first-out queue Nz(t). If any two, or more, vehicles enter the control zone at
the same time, then the coordinator randomly selects their positions in the queue.
Definition 3. For each CAV i ∈ Nz(t) entering the control zone z, the information
set Y z
i (t), is defined as
pi(t), vi(t),Lz
i (t),Cz
i (t), tz,m
i
,∀t ∈ [tz,0
i
, tz,m
i
],
(7)
i (t) (cid:44)(cid:110)
Y z
where pi(t), vi(t) are the position and speed of CAV i inside the control zone z, tz,0
the time when vehicle i enters the control zone for conflict zone z, and tz,m
targeted for vehicle i to enter the conflict zone z. The set Y z
that each vehicle shares.
is
is the time
i (t) includes all information
i
i
The time tz,m
i
that the vehicle i will be entering the conflict zone z maximizes the
throughput while considering the maximum and minimum speed limits. Therefore, to
ensure that (3) and (6) are satisfied at tz,m
, we impose the following conditions which
depend on the subset that the vehicle i − 1 ∈ Nz(t) belongs to.
i
If CAV i − 1 ∈ Lz
i (t),
(cid:40)
(cid:110)
tz,m
i = max
min
tz,m
i−1 +
δ(t)
vz ,
Lz
vmin
,
Lz
vi(tz,0
i )
,
Lz
vmax
.
(8)
If CAV i − 1 ∈ Cz
i (t),
(cid:40)
tz,m
i = max
min
(cid:110)
tz,m
i−1 +
(cid:111)
,
Sz
vz ,
Lz
vmin
Lz
vi(tz,0
i )
,
Lz
vmax
(cid:41)
,
(9)
i
where Sz is the length of conflict zone z, Lz is the length of control zone for zone z, vz
is the constant imposed speed inside the conflict zone z, and vi(tz,0
i ) is the initial speed
of vehicle i when it enters the control zone at tz,0
. The conditions (8) and (9) ensure
that the time tz,m
each vehicle i will be entering the conflict zone is feasible and can
be attained based on the imposed speed limits inside the control zone. In addition, for
low traffic flow where vehicles i − 1 and i might be located far away from each other,
there is no compelling reason for vehicle i to accelerate within the control zone to have
a distance δ(t) from vehicle i − 1, if i − 1 ∈ Lz
i (t), at
the time tz,m
that vehicle i enters the conflict zone z. Therefore, in such cases, vehicle
i can keep cruising within the control zone with the initial speed vi(tz,0
i ) that entered
the control zone at tz,0
i (t), or a distance Sz if i − 1 ∈ Lz
.
i
i
The recursion is initialized when the first vehicle enters the control zone z, i.e., it is
can be externally assigned as the desired exit time of
is fixed and available
1 (t) from vehicle 1 to compute the time
2 (t) from vehicle 2, and the communication process
N (t)−1(t).
assigned i = 1. In this case, tz,m
this vehicle whose behavior is unconstrained. Thus, the time tz,m
through Y1(t). The second vehicle will access Y z
tz,m
2
will continue with the same fashion until vehicle N (t) in the queue accesses Y z
. The third vehicle will access Y z
1
1
i
7
3.3. Optimal Control Problem Formulation
By controlling the entry time of the vehicles, the speed of queue build-up at each
conflict zone decreases. Thus, the congestion recovery time is also reduced -- the lat-
ter results in maximizing the throughput in the conflict zone. We now consider the
problem of deriving the optimal control input (acceleration/deceleration) of each CAV
inside each control zone separately under hard safety constraints to avoid collisions.
Moreover, by optimizing the acceleration/deceleration of each vehicle, we minimize
transient operation. This will have direct benefits in energy consumption since the
vehicles are optimized to travel over steady state operating points (constant torque
and speed) [45].
Since the coordinator for a conflict zone z is not involved in any decision on the
vehicle coordination, we formulate the following optimization problem for each vehicle
in the queue upstream of conflict zone z, the solution of which can be implemented in
real-time
(cid:90) tz,m
i
min
ui
1
2
Subject to : (1), (2), pi(tz,0
i (t)dt, z ∈ Z,
u2
, vi(tz,0
, tz,m
,
i
i ) = vz,0
i
i
tz,0
i
i ) = pz,0
and given tz,0
i
(10)
, pi(tz,m
i
) = pz,
where pz is the location (i.e., entry position) of the conflict zone z, tz,m
is the time that
the vehicle i enters the conflict zone, and pz,0
are the initial position and speed
of vehicle i ∈ Nz(t) when it enters the control zone of conflict zone z. By minimizing
the L2 norm of acceleration we minimize transient engine operation which results in
an overall improvement in energy efficiency.
, vz,0
i
i
i
For the analytical solution and real-time implementation of the control problem (10),
we apply Hamiltonian analysis. The analytical solution of (10) without considering
state and control constraints was presented in earlier work [15,16] for coordinating
CAVs in real-time at highway on-ramps. When the state and control constraints are
not active, the optimal control input (acceleration/deceleration) as a function of time
is given by
u∗
i (t) = ait + bi, tz,0
i ≤ t ≤ tz,m
i
,
and the optimal speed and position for each vehicle are
v∗
i (t) =
1
2
ait2 + bit + ci, tz,0
i ≤ t ≤ tz,m
i
,
p∗
i (t) =
1
6
ait3 +
1
2
bit2 + cit + di, tz,0
i ≤ t ≤ tz,m
i
(11)
(12)
(13)
,
where ai, bi, ci and di are constants of integration that can be computed by using
the initial and final conditions. Similar results to (11)-(13) can be obtained when the
state and control constraints become active within the control zone. In this case, the
constrained and unconstrained arcs need to be pieced together to satisfy the Euler-
Lagrange equations and the necessary conditions of optimality. The different cases
of the state and control constraint activation along with the corresponding solution
8
can be found in [12], whereas the complete analytical solution that includes the rear-
end safety constraint is reported in [46]. In the present work, we do not consider any
constrained optimization cases as none of the constraints in (2) become active within
the optimal control path during the simulation, as shown in Section 5.
4. Simulation and Experimental Environment
4.1. Simulation Setup
To implement the control framework presented in the previous section, and to gen-
erate the input information required for UDSSC, we first use the microscopic multi-
modal simulation platform PTV VISSIM. We create a simulation setup replicating the
UDSSC map and define a network consisting of four different looped routes and five
bottlenecks (one intersection, one roundabout, and three merging scenarios), as shown
in Fig. 3. In order to maintain compatibility with the UDSSC experiment, we design
each of the routes to hold a finite number of vehicles (19 vehicles in total) traveling in
loops for finite simulation run-time. Among the 19 vehicles, we consider 9 vehicles as
the target (ego-vehicles) to evaluate their performance metrics in different scenarios.
We use the rest of the vehicles to increase the traffic volume in the urban network
and create congestion in the baseline scenario. The vehicles maintain a low desired
speed of 7 m/s for their uncontrolled urban commute throughout the network. There-
fore, the desired speed at all exits of the control zones is set to be equal to the urban
speed. We select the maximum and minimum allowable speed of 8.33 m/s and 2 m/s,
respectively. The maximum and minimum acceleration of the vehicles was taken as 3
m/s2 and -3 m/s2, respectively. To evaluate the effectiveness of the proposed optimal
vehicle dynamics control, we consider two different cases:
a. Baseline Scenario: We construct the baseline scenario by considering all vehicles
as human-driven and without any V2V communication capability. The vehicles sub-
scribe to the Wiedemann car following model. The Wiedemann car following model is
a psycho-physical model to emulate the driving behavior of real human-driven vehi-
cles. The model was first presented in 1974 by Wiedemann [47] and has been adopted
by the software PTV-VISSIM as one of its in-built car following models. The complete
detail of this elaborate model can be found in the literature [47]. We do not change the
VISSIM default parameters of the Wiedemann model to study their impact on vehicle
behavior and traffic flow, as such exposition falls outside the scope of this paper. We
build a fixed time signalized intersection for the four-way traffic at the center. The
fixed time signalized intersection in the simulation provides Signal Phase and Tim-
ing (SPaT) messages to the leading vehicle internally through VISSIM. Based on the
SPaT information, the leading vehicle stops at the desired position by adopting the
Wiedemann car following model. We adopt priority-based (yield/stop) movement for
the other four waypoints consisting of the roundabout and merging scenarios, where
the secondary-road vehicles yield to the main-road vehicles. In both cases, VISSIM
uses the approaching point parameter of Wiedemann car following model to detect
any approaching obstacles and apply necessary braking to slow down or stop.
b. Optimal Controlled Scenario: In the optimal scenario, all 19 vehicles follow our
optimal control framework. The vehicles are connected with each other inside the
control zone through V2V communication capability and are automated within the
control zone. Therefore, they can plan their optimal path inside the control zone,
avoiding any lateral or rear-end collisions while optimizing their own travel time and
9
fuel efficiency. In this scenario, we do not consider the fixed-time signal and movement
priorities considered in the baseline case. We consider five isolated coordinators with
a control zone of 45 m for each conflict zone (see Fig. 3). For the uncontrolled paths
in-between the control zones, the vehicles adopt the Wiedemann car following model
[47] to traverse their respective routes.
4.2. University of Delaware's Scaled Smart City
UDSSC is a 1:25 scaled testbed spanning over 400 square feet (see Fig. 2) and is capable
of accommodating scaled robotic CAVs. It is equipped with a VICON motion capture
system that uses eight cameras to track the position of each vehicle with sub-millimeter
accuracy. Each road in the UDSSC is built up from arc or line segments. In order to
track the desired vector position of each CAV, all road segments are parameterized
in terms of their total length. This formulation allows each vehicle to calculate its
desired position in UDSSC based only on the scalar distance along its current path,
which is achieved by numerically integrating the speed profile in real-time on both
the mainframe computer and each CAV. This decoupling of speed and position allows
significant flexibility in UDSSC, especially in dynamic-routing scenarios.
4.2.1. Connected and Automated Vehicles
The CAVs of UDSSC (see Fig. 4) have been designed using off the shelf electrical
components and 3D printed parts created at the University of Delaware. The primary
microcontroller on the CAV is a Raspberry Pi 3B running Ubuntu Mate and ROS
Kinetic. An Arduino Nano is used as a slave processor for the Pi to do low-level
motor control and ad-hoc analog to digital conversion for the state of charge (SOC)
measurements. The CAV's rear-wheel drive train is powered by a Pololu 75.8:1, 6 V
micro metal gearmotor; the motor is controlled using a motor controller, and encoder
for feedback, through the Arduino. Power from the gearmotor is transferred to the rear
axle with two 3D printed gears with a 1:1 ratio, and two rubberized wheels with radius
r = 1.6 cm are mounted directly to the rear axle. The motor controller receives power
through a 5 V regulator, and a pulse-width modulated command from the Arduino
is used to control the motor's speed. Steering is achieved by a custom 3D printed
Ackermann-style steering mechanism actuated by a Miuzei micro servo motor, which
again is controlled directly by the Arduino. The CAVs are also equipped with a Pi
Camera, ultrasonic sensors, and a SOC measurement circuit to collect experimental
data and reduce the overall reliance on VICON. A power regulator manages the voltage
requirement of the Pi and Arduino by supplying a regulated 5 V DC from two 3000
mAh 3.7 V Li-ion batteries configured in series. With this hardware configuration, the
CAV is able to run and collect experimental data at 20 Hz for up to 2 hours.
4.2.2. Control System Architecture
Coordination of the CAVs within the UDSSC is achieved using a multi-level control
framework spanning a central mainframe computer (Processor: Intel Core i7-6950X
CPU @ 3.00 GHz x 20, Memory: 125.8 Gb) and the individual CAVs in the experiment
(Raspberry Pi 3B). The mainframe runs an Ubuntu 16.04.5 LTS Linux distribution
and ROS Kinetic. High-level routing is achieved by a multithreaded C++ program
running on the mainframe computer. For this set of experiments, the mainframe is
initialized with the path information and speed profiles for each CAV. At the start
10
Figure 4. A picture of the connected and automated vehicle's electronics (left) and outer shell with VICON
markers, ultrasonic sensors, and camera visible (right).
of the experiment, each CAV sets its temporal baseline from which it measures all
later times; this avoids the problem of synchronizing CAV clocks, as all information is
calculated relative to the experiment start time. During the experiment, the mainframe
passes a message to each CAV containing its current position and two seconds of
trajectory data using the UDP/IP protocol at 50 Hz. The CAV receives trajectory
information from the mainframe and uses a modified Stanley [48] controller to handle
lane tracking, while a feedforward-feedback [49] PID controller tracks the desired speed
profile.
Medium and low-level control is accomplished onboard each CAV in a purely dis-
tributed manner. Using information from the mainframe, each CAV updates at 50 Hz
to calculate a lateral, heading, and distance error. The lateral and heading errors are
then passed to the Stanley controller to calculate an output steering angle. Meanwhile,
the position error and desired speed are used in a feedforward-feedback controller to
calculate the desired motor speed. The desired speed and steering angle are then passed
to the Arduino Nano, which runs a low-level PID controller to precisely control the
gearmotor and steering servo.
4.3. Experimental Setup
In order to replicate the simulation results in the UDSSC, speed profiles for the 9
ego vehicles are exported from VISSIM to the mainframe. The path information and
speed profile are dispersed to each CAV for the duration of the experiment. Then, the
CAVs at UDSSC numerically integrate the speed profile data in real-time to calculate
their desired position, allowing them to track the desired speed and position in a
decentralized manner. Simultaneously, the mainframe computer integrates the speed
profile in order to send current and future path information to the CAVs.
11
5. Results
5.1. Simulation Results
The speed profiles of the CAVs, making multiple passes through each loop of the
UDSSC map, for the baseline and optimal control scenario are shown in Fig. 5a - 5d.
The baseline speed profiles show significant stop-and-go driving behavior. The effect
of congestion in these cases can induce artificial congestion or phantom traffic jams
outside the corridor, as can be seen in Fig. 5d between 60 and 120 s.
Compared to the baseline speed profiles, we note that the optimal controller has
completely eliminated stop-and-go driving, and the optimal quadratic profile speed is
realized within the control zones. The congestion observed in the baseline scenario in
Fig. 5d has also been mitigated, rendering an overall smooth traffic flow throughout
the network.
Note that the speed profiles are kept within the maximum and minimum limit inside
the control zone as described in Section 4.1. Therefore, none of the state and control
constraints of (2) became active in the unconstrained arc, and our relaxation of (2)
holds. However, we see a few cases of constraint violation of the acceleration profile
outside the control zone in Fig. 5a - 5d, where the Wiedemann car-following model
[47] is applied instead of the optimal control. Note that optimal control is applied only
inside the designated control zones.
5.2. Experimental Validation
To validate the effectiveness and efficiency of our optimal control observed in the
simulation, we compare the travel time of 9 CAVs between the baseline and optimal
scenarios. The travel time is calculated as the time taken for each vehicle to complete
a single loop by sampling the raw VICON data over the 80 s experiment. In particular,
returning to the initial position was defined as the first time the vehicle came within 10
cm of its initial position after a 5 s initial window. These values are presented in Table 1
and Fig. 6 alongside the route each CAV took (as annotated in Fig. 3); a value of greater
than 80 s corresponds to a vehicle not fully completing its loop during the experiment,
which occurred 3 times in the baseline scenario. Videos of the experiment can be found
at the supplemental site, https://sites.google.com/view/ud-ids-lab/tfms.
Table 1. Travel time for each vehicle to complete a single loop.
Vehicle Baseline time [s] Optimal time [s]
Loop Time saved [s] % Decrease
8
18
12
14
4
13
2
5
17
> 80
> 80
66.25
65.95
57.80
60.30
46.61
43.90
> 80
54.10
48.00
45.55
49.90
53.25
59.75
37.40
44.20
41.50
North
North
East
East
South
South
West
West
West
> 25.9
> 32.0
20.70
16.05
4.55
0.55
9.21
−0.30
> 38.5
32.4
40.0
31.2
24.3
7.9
0.9
19.8
-0.7
48.1
From Table 1, an average improvement of at least 16.35 s (25%) over the baseline
travel time was observed in the UDSSC. The marginal improvement was in the South-
12
(a) North Loop
(b) East Loop
(c) South Loop
(d) West Loop
Figure 5.
taken from the first 160s of simulation.
Instantaneous maximum, minimum, and average speed of each vehicle by route. Each data set is
ern loop, where the traffic was effectively free-flowing in the baseline scenario. In the
loops with conflict zones, i.e., north, east, and west, the impact of the coordinator and
optimal control is clear and significant.
The speed profiles of CAVs 2, 14, 17, and 13 are shown in Fig. 7. These profiles
were taken by numerically deriving the VICON position data taken at 100 Hz to get
velocity components. Then, any speeds above 0.8 m/s, well above the maximum speed
achievable by the CAVs, was attributed to occlusion during the experiment and thus
discarded. Finally, the velocity magnitude was run through a moving average filter
with a window of 0.45 s.
We conclude from the above results that 1) almost the entire reduction in transit
time can be attributed to optimal control in the conflict zones, and 2) the optimal
control framework almost entirely eliminates stop-and-go driving.
13
Optimal Range Optimal Avg. Optimal Range Optimal Avg. Optimal Range Optimal Avg. Optimal Range Optimal Avg. Figure 6. Histogram for the arrival time of each vehicle in Table 1 with 6 bins per experiment.
6. Conclusion
In this paper, we presented an experimental demonstration of a decentralized optimal
control framework for CAVs, presented in [2]. We used a 1:25 scaled testbed repre-
senting an urban environment with robotic CAVs that can replicate real-world traffic
scenarios in a controlled environment. We showed that the optimal control frame-
work could contribute a 25% reduction in travel time compared to a baseline scenario
consisting of human-driven vehicles without connectivity. We should note that under
heavy congested traffic conditions, the control framework might not be as effective as
it is under light to medium traffic conditions in improving traffic flow without tuning
some control parameters (e.g., minimum time gap, length of the control zone, etc.)
[50]. In recent work [51], the effectiveness of the proposed optimal control framework
has been investigated under different traffic conditions to confirm its robustness.
Ongoing research includes the formulation of an upper-level optimization problem,
the solution of which yields, for each CAV, the optimal entry time and lane changes
required to cross the intersection [44] and explores the associated tradeoffs between
throughput and energy consumption of each vehicle.
An important direction for future research is to consider different penetrations of
CAVs, which can significantly alter the efficiency of the entire system. For example,
an important question that needs to be addressed is, "what is the minimum number
of CAVs in order to realize potential benefits?" Future work should also consider
the robustness of the control framework and its applicability under various traffic
conditions. The impact of communication errors and delays on safety and optimality
are also areas of future research.
14
Figure 7. Speed vs time profiles for vehicles (clockwise from top left) 2 (west), 14 (east), 17 (west), 13
(south.).
Acknowledgement(s)
The authors would like to acknowledge Michael Lashner, Kunzheng Li, Haley Lloyd,
Thomas Patterson, and the rest of the UDSSC Senior Design team for their effort in
designing, building, and testing the newest generation of CAVs used in this paper. The
authors would also like to thank Ioannis Vasileios Chremos for his valuable comments
and feedback on the manuscript.
Funding
This research was supported in part by ARPAEs NEXTCAR program under the award
number DE- AR0000796 and by the Delaware Energy Institute (DEI). This support
is gratefully acknowledged.
References
[1] Malikopoulos AA. A duality framework for stochastic optimal control of complex systems.
IEEE Transactions on Automatic Control. 2016;61(10):2756 -- 2765.
[2] Zhao L, Malikopoulos AA. Decentralized optimal control of connected and automated
vehicles in a corridor. In: 2018 21st International Conference on Intelligent Transportation
Systems (ITSC); IEEE; 2018. p. 1252 -- 1257.
[3] Spieser K, Treleaven K, Zhang R, et al. Toward a systematic approach to the design
15
and evaluation of automated mobility-on-demand systems: A case study in singapore. In:
Road vehicle automation. Springer; 2014. p. 229 -- 245.
[4] Fagnant DJ, Kockelman KM. The travel and environmental implications of shared au-
tonomous vehicles, using agent-based model scenarios. Transportation Research Part C:
Emerging Technologies. 2014;40:1 -- 13.
[5] Shladover SE, Desoer CA, Hedrick JK, et al. Automated vehicle control developments in
the PATH program. IEEE Transactions on Vehicular Technology. 1991;40(1):114 -- 130.
[6] Rajamani R, Tan HS, Law BK, et al. Demonstration of integrated longitudinal and lateral
control for the operation of automated vehicles in platoons. IEEE Transactions on Control
Systems Technology. 2000;8(4):695 -- 708.
[7] Tsugawa S. An overview on an automated truck platoon within the energy its project.
IFAC Proceedings Volumes. 2013;46(21):41 -- 46.
[8] D´avila A, Nombela M. Sartre: Safe road trains for the environment. In: Conference on
Personal Rapid Transit PRT@ LHR; Vol. 3; 2010. p. 2 -- 3.
[9] Shladover SE. PATH at 20 -- History and major milestones. IEEE Transactions on intelli-
gent transportation systems. 2007;8(4):584 -- 592.
[10] Lee J, Park B. Development and Evaluation of a Cooperative Vehicle Intersection Control
Algorithm Under the Connected Vehicles Environment. IEEE Transactions on Intelligent
Transportation Systems. 2012;13(1):81 -- 90.
[11] Rakha H, Kamalanathsharma RK. Eco-driving at signalized intersections using V2I com-
munication. In: Intelligent Transportation Systems (ITSC), 2011 14th International IEEE
Conference on; IEEE; 2011. p. 341 -- 346.
[12] Malikopoulos AA, Cassandras CG, Zhang YJ. A decentralized energy-optimal control
framework for connected automated vehicles at signal-free intersections. Automatica.
2018;93:244 -- 256.
[13] Mahbub AMI, Zhao L, Assanis D, et al. Energy-Optimal Coordination of Connected and
Automated Vehicles at Multiple Intersections. In: Proceedings of 2019 American Control
Conference; 2019. p. 2664 -- 2669.
[14] Chalaki B, Malikopoulos AA. An optimal coordination framework for connected and
automated vehicles in two interconnected intersections. In: Proceedings of 2019 IEEE
Conference on Control Technology and Applications, 2019; 2019. p. 888 -- 893.
[15] Rios-Torres J, Malikopoulos AA. Automated and Cooperative Vehicle Merging at
Highway On-Ramps. IEEE Transactions on Intelligent Transportation Systems. 2017;
18(4):780 -- 789.
[16] Ntousakis IA, Nikolos IK, Papageorgiou M. Optimal vehicle trajectory planning in the
context of cooperative merging on highways. Transportation Research Part C: Emerging
Technologies. 2016;71:464 -- 488.
[17] Malikopoulos AA, Hong S, Park B, et al. Optimal control for speed harmonization of
automated vehicles. IEEE Transactions on Intelligent Transportation Systems. 2018;.
[18] Athans M. A unified approach to the vehicle-merging problem. Transportation Research.
1969;3(1):123 -- 133.
[19] Kachroo P, Li Z. Vehicle merging control design for an automated highway system. In:
Proceedings of Conference on Intelligent Transportation Systems; 1997. p. 224 -- 229.
[20] Antoniotti M, Deshpande A, Girault A. Microsimulation analysis of automated vehicles
on multiple merge junction highways. In: IEEE International Conference in Systems, Man,
and Cybernetics; 1997. p. 839 -- 844.
[21] Ran B, Leight S, Chang B. A microscopic simulation model for merging control on a
dedicated-lane automated highway system. Transportation Research Part C: Emerging
Technologies. 1999;7(6):369 -- 388.
[22] Dresner K, Stone P. Multiagent traffic management: a reservation-based intersection
control mechanism. In: Proceedings of the Third International Joint Conference on Au-
tonomous Agents and Multiagents Systems; 2004. p. 530 -- 537.
[23] Dresner K, Stone P. A multiagent approach to autonomous intersection management.
Journal of artificial intelligence research. 2008;31:591 -- 656.
16
[24] de La Fortelle A. Analysis of reservation algorithms for cooperative planning at inter-
sections. In: 13th International IEEE Conference on Intelligent Transportation Systems;
2010. p. 445 -- 449.
[25] Huang S, Sadek A, Zhao Y. Assessing the Mobility and Environmental Benefits of
Reservation-Based Intelligent Intersections Using an Integrated Simulator. IEEE Trans-
actions on Intelligent Transportation Systems. 2012;13(3):1201 -- 1214.
[26] Zohdy IH, Kamalanathsharma RK, Rakha H. Intersection management for autonomous
vehicles using iCACC; 2012. p. 1109 -- 1114.
[27] Yan F, Dridi M, El Moudni A. Autonomous vehicle sequencing algorithm at isolated inter-
sections. 2009 12th International IEEE Conference on Intelligent Transportation Systems.
2009;:1 -- 6.
[28] Li L, Wang FY. Cooperative Driving at Blind Crossings Using Intervehicle Communica-
tion. IEEE Transactions in Vehicular Technology. 2006;55(6):1712,1724.
[29] Zhu F, Ukkusuri SV. A linear programming formulation for autonomous intersection
control within a dynamic traffic assignment and connected vehicle environment. Trans-
portation Research Part C: Emerging Technologies. 2015;55.
[30] Wu J, Perronnet F, Abbas-Turki A. Cooperative vehicle-actuator system: a sequence-
based framework of cooperative intersections management. Intelligent Transport Systems,
IET. 2014;8(4):352 -- 360.
[31] Kim KD, Kumar P. An MPC-Based Approach to Provable System-Wide Safety and
Liveness of Autonomous Ground Traffic. IEEE Transactions on Automatic Control. 2014;
59(12):3341 -- 3356.
[32] Barth M, Boriboonsomsin K. Energy and emissions impacts of a freeway-based dynamic
eco-driving system. Transportation Research Part D: Transport and Environment. 2009;
14(6):400 -- 410.
[33] Berry IM. The effects of driving style and vehicle performance on the real-world fuel con-
sumption of us light-duty vehicles [dissertation]. Massachusetts Institute of Technology;
2010.
[34] Wu C, Zhao G, Ou B. A fuel economy optimization system with applications in vehicles
with human drivers and autonomous vehicles. Transportation Research Part D: Transport
and Environment. 2011;16(7):515 -- 524.
[35] Zhao L, Malikopoulos AA, Rios-Torres J. Optimal control of connected and automated
vehicles at roundabouts: An investigation in a mixed-traffic environment. In: 15th IFAC
Symposium on Control in Transportation Systems; 2018. p. 73 -- 78.
[36] Rios-Torres J, Malikopoulos AA. Impact of partial penetrations of connected and auto-
mated vehicles on fuel consumption and traffic flow. IEEE Transactions on Intelligent
Vehicles. 2018;3(4):453 -- 462.
[37] Zhong Z, Joyoung L, Zhao L. Evaluations of Managed Lane Strategies for Arterial De-
ployment of Cooperative Adaptive Cruise Control . In: TRB Annual Meeting; Washington
DC, USA; 2017.
[38] Rios-Torres J, Malikopoulos AA. A Survey on Coordination of Connected and Automated
Vehicles at Intersections and Merging at Highway On-Ramps. IEEE Transactions on
Intelligent Transportation Systems. 2017;18(5):1066 -- 1077.
[39] Guanetti J, Kim Y, Borrelli F. Control of connected and automated vehicles: State of the
art and future challenges. Annual Reviews in Control. 2018;45:18 -- 40.
[40] Wang Y, Li X, Yao H. Review of trajectory optimisation for connected automated vehicles.
IET Intelligent Transport Systems. 2018;13:580 -- 586.
[41] Paull L, Tani J, Ahn H, et al. Duckietown: An open, inexpensive and flexible platform for
autonomy education and research. In: Proceedings of the IEEE International Conference
on Robotics and Automation; 2017. p. 1497 -- 1504.
[42] Hyaldmar N, He Y, Porok A. A fleet of miniature cars for experiments in cooperative driv-
ing. In: Proceedings of the IEEE International Conference on Robotics and Automation;
2019.
[43] Stager A, Bhan L, Malikopoulos A, et al. A Scaled Smart City for Experimental Validation
17
of Connected and Automated Vehicles. IFAC-PapersOnLine. 2018;51(9):130 -- 135.
[44] Malikopoulos AA, Zhao L. Optimal path planning for connected and automated vehicles
at urban intersections. In: Proceedings of the 58th IEEE Conference on Decision and
Control; 2019 (to appear).
[45] Malikopoulos AA, Assanis DN, Papalambros PY. Optimal engine calibration for individ-
ual driving styles. In: SAE Proceedings, Technical Paper 2008-01-1367; 2008.
[46] Malikopoulos AA, Zhao L. A closed-form analytical solution for optimal coordination of
connected and automated vehicles. In: Proceedings of 2019 American Control Conference;
2019. p. 3599 -- 3604.
[47] Wiedemann R. Simulation des strassenverkehrsflusses [dissertation]. Universitat Karl-
sruhe; 1974.
[48] Thrun S, Montemerlo M, Dahlkamp H, et al. Stanley: The robot that won the DARPA
Grand Challenge. Springer Tracts in Advanced Robotics. 2007;.
[49] Spong MW, Hutchinson S, Vidyasagar M. Robot Dynamics and Control Second Edition;
2004.
[50] Zhao L, Malikopoulos AA, Rios-Torres J. On the traffic impacts of optimally controlled
connected and automated vehicles. In: Proceedings of 2019 IEEE Conference on Control
Technology and Applications; 2019. p. 882 -- 887.
[51] Zhao L, Mahbub AMI, Malikopoulos A. Optimal vehicle dynamics and powertrain control
for connected and automated vehicle. In: Proceedings of IEEE Conference on Control
Technology and Applications; 2019. p. 33 -- 38.
18
|
cs/0702091 | 1 | 0702 | 2007-02-16T16:16:49 | Observable Graphs | [
"cs.MA"
] | An edge-colored directed graph is \emph{observable} if an agent that moves along its edges is able to determine his position in the graph after a sufficiently long observation of the edge colors. When the agent is able to determine his position only from time to time, the graph is said to be \emph{partly observable}. Observability in graphs is desirable in situations where autonomous agents are moving on a network and one wants to localize them (or the agent wants to localize himself) with limited information. In this paper, we completely characterize observable and partly observable graphs and show how these concepts relate to observable discrete event systems and to local automata. Based on these characterizations, we provide polynomial time algorithms to decide observability, to decide partial observability, and to compute the minimal number of observations necessary for finding the position of an agent. In particular we prove that in the worst case this minimal number of observations increases quadratically with the number of nodes in the graph.
From this it follows that it may be necessary for an agent to pass through the same node several times before he is finally able to determine his position in the graph. We then consider the more difficult question of assigning colors to a graph so as to make it observable and we prove that two different versions of this problem are NP-complete. | cs.MA | cs |
OBSERVABLE GRAPHS
RAPHAEL M. JUNGERS AND VINCENT D. BLONDEL
Abstract. An edge-colored directed graph is observable if an agent that moves along its
edges is able to determine his position in the graph after a sufficiently long observation
of the edge colors. When the agent is able to determine his position only from time to
time, the graph is said to be partly observable. Observability in graphs is desirable in
situations where autonomous agents are moving on a network and one wants to localize
them (or the agent wants to localize himself) with limited information. In this paper,
we completely characterize observable and partly observable graphs and show how these
concepts relate to observable discrete event systems and to local automata. Based on
these characterizations, we provide polynomial time algorithms to decide observability, to
decide partial observability, and to compute the minimal number of observations neces-
sary for finding the position of an agent. In particular we prove that in the worst case
this minimal number of observations increases quadratically with the number of nodes in
the graph. From this it follows that it may be necessary for an agent to pass through the
same node several times before he is finally able to determine his position in the graph.
We then consider the more difficult question of assigning colors to a graph so as to make
it observable and we prove that two different versions of this problem are NP-complete.
1. Introduction
Consider an agent moving from node to node in a directed graph whose edges are
colored. The agent knows the colored graph but does not know his position in the graph.
From the sequence of colors he observes he wants to deduce his position. We say that an
edge-colored directed graph1 is observable if there is some observation time length after
which, whatever the color sequence he observes, the agent is able to determine his position.
Date: October 8, 2018.
The research reported here was performed while the authors were at MIT, Cambridge. It was par-
tially supported by the "Communaut´e francaise de Belgique - Actions de Recherche Concert´ees", by the
EU HYCON Network of Excellence (contract number FP6-IST-511368), by the Belgian Programme on
Interuniversity Attraction Poles initiated by the Belgian Federal Science Policy Office, and by the DoD
AFOSR URI for "Architectures for Secure and Robust Distributed Infrastructures", F49620-01-1-0365
(led by Stanford University). The scientific responsibility rests with its authors. Raphael Jungers is a
FNRS fellow (Belgian Fund for Scientific Research).
1The property of being observable is a property of directed graphs that have their edges colored. For
simplicity, we will talk in the sequel about observable graphs rather than observable edge-colored directed
graphs.
1
2
RAPHA EL M. JUNGERS AND VINCENT D. BLONDEL
2
2
1
3
1
5
4
5
(a)
(b)
3
4
Figure 1. An agent is moving along directed edges. Edges are "colored"
solid (S) or dashed (D). The graph (a) is observable because the observation
of the last three colors suffices to determine the position of the agent in the
graph. The graph (b) is not observable; the sequence SSDSSD. . .SSD can
be obtained from a path ending at node 2 or at node 4.
Of course, if all edges are of different colors, or if edges with different end-nodes are of
different colors, then the agent is able to determine his position after just one observation.
So the interesting situation is when there are fewer colors than there are nodes. Consider
for instance the two graphs on Figure 1. The graphs differ only by the edge between the
nodes 1 and 3. The edges are colored with two "colors": solid (S) and dashed (D). We
claim that the graph (a) is observable but that (b) is not. In graph (a), if the observed
color sequence is DDS then the agent is at node 1, while if the sequence is SDD he is at
node 3. Actually, it follows from the results presented in this paper that the observation
of color sequences of length three always suffices to determine the exact position of the
agent in this graph. Consider now the graph (b) and assume that the observed sequence
is SSDSS; after these observations are made, the agent may either be at node 1, or at
node 3. There are two paths that produce the color sequence SSDSS and these paths
have different end-nodes. Sequences of arbitrary large length and with the same property
can be constructed and so graph (b) is not observable.
There are of course very natural conditions for a graph to be observable. A first
condition is that no two edges of identical colors may leave the same node. Indeed, if a
node has two outgoing edges with identical colors then an agent leaving that node and
observing that color will not be able to determine his next position in the graph. So this
is clearly a necessary condition. Another condition is that the graph may not have two
cycles with identical color sequences, as for example the cycles 2−1−3−2 and 4−5−1−4
in graph (b) of Figure 1. If such cycles are present in the graph, then an agent observing
that repeated particular color sequence is not able to determine if he is moving on one or
the other cycle. So, these two conditions are clearly necessary conditions for a graph to
OBSERVABLE GRAPHS
3
*
**
(a)
(b)
Figure 2. Both graphs are partly observable but none of them is observ-
able.
In graph (a), complete trajectories can always be reconstructed a
posteriori; which is not possible in graph (b).
be observable. In our first theorem we prove that these two conditions together are also
sufficient.
In an observable graph there is some T ≥ 1 for which an agent is able to determine his
position in the graph for all observations of length T , and so he is also able to do so for
all subsequent times. Some graphs are not observable in this way but in a weaker sense.
Consider for example the graph (a) of Figure 2. This graph is not observable because there
are two dashed edges that leave the top node and so, whenever the agent passes through
that node the next observed color is D and the position of the agent is then uncertain.
This graph is nevertheless partly observable because, even though the agent is not able to
determine his position at all time, there is a finite window length T such that the agent is
able to determine his position at least once every T observations. Indeed, after observing
the sequence SD in this graph (and this sequence occurs in every observation sequence of
length 4), the next observation is either S or D. If it is S, the agent is at the bottom right
node; if it is D, he is at the bottom left node. In both cases the next observed color is S
and the agent is then at the top node. So the agent is able to determine his position at
least once every five observations. In an observable graph, an agent is able to determine
his position at all times beyond a certain limit; in a partly observable graph the agent is
able to determine his position infinitely often (for formal definitions, see below).
Notice that in the graph (a) of Figure 2, the agent is able to reconstruct its entire
trajectory a posteriori, except maybe for its last position. This is however not the case
for all partly observable graphs. Consider for example the star graph (b) of Figure 2.
The color sequences observed in this graph are SDSDS. . . or DSDSDS. . . . When the last
observed color is D, the agent is at the central node. When the last observed color is S,
the agent only knows that he is at one of the extreme nodes. Hence this graph is partly
observable but in this case, contrary to what we have with graph (a), the entire trajectory
cannot be reconstructed a posteriori.
4
RAPHA EL M. JUNGERS AND VINCENT D. BLONDEL
In this paper, we prove a number of results related to observable and partly observable
graphs. We first prove that the conditions described above for observability are indeed
sufficient. From the proof of this result it follows that in an observable graph an agent
can determine his position in the graph after an observation of length at most n2, where
n is the number of nodes. This quadratic increase of the observation length cannot be
avoided: we provide a family of graphs for which Θ(n2) observations are necessary. Based
on some of these properties we also provide polynomial-time algorithms for checking if
a graph is observable or partly observable. We then consider the question of assigning
colors to the edges or nodes of a directed graph so as to make it observable and we prove
that the problem of finding the minimal number of colors is NP-complete.
The concept of observable graphs is related to a number of concepts in graph and
automata theory, control theory and Markov models which we now describe.
Observable graphs as defined here are actually a particular case of local automata. A
finite state automaton is said to be (d, k)-local (with d ≤ k) if any two paths of length k
and identical color sequences pass through the same state at step d. Hence an observable
graph is a (k, k)-local automaton. It is shown in [1] how to recognize local automata in
polynomial time but the motivation in that context is very different from ours and little
attention is given there at the particular values of the integer d and k. In particular, the
algorithm provided in [1] does not allow to recognize (k, k)-local automata.
Related to the notion of observable graphs, Crespi et al. [2] have recently introduced
the concept of trackable graph. An edge-colored directed graph is said to be trackable if
the maximum number of trajectories compatible with a color sequence of length k grows
subexponentially with k. So, in the context of trackability, one cares about the total
number of compatible trajectories, but not about the position of the agent in the graph.
It has recently been proved that the problem of determining if a graph is trackable can
be solved in polynomial time [5]; see also [2]. It is clear that an observable network is
trackable, but the converse is not true in general. For example the graph on Figure 2 (a)
is trackable but not observable. Notice also that, as shown with the graph on Figure 2
(b), partly observable graphs do not need to be trackable.
Observable graphs are also related to Discrete Event Systems (DES). More precisely, our
notion of partly observable graph is similar to what Oszveren and Willsky call observable
DES [6]. These authors define DES as colored graphs, except for the fact that they
allow transitions to be unobservable: some edges have no colors. From a colored graph
with unobservable transitions we can easily construct an equivalent fully colored graph by
removing all unobservable transitions and adding an edge (h, j) of color c whenever there
is an edge (h, i) of color c and an unobservable transition (or a sequence of unobservable
transitions) between i and j. Our results are therefore applicable to observable DES as
defined in [6].
Finally, the results presented in this paper can also be interpreted in the context of
Hidden Markov Processes (HMP) [3, 7]. More precisely, the graphs we consider can be
seen as finite alphabet HMPs (also called aggregated Markov processes), except that no
OBSERVABLE GRAPHS
5
values different from 0 and 1 are given for the transition probabilities. In our context, a
transition is either allowed or it is not; we do not associate transition probabilities. Also,
we can assign to a given transition several possible colors, but again, without considering
their respective probabilities.
The remainder of this paper is organized as follows: in Section 2 we formalize the notions
presented above and give necessary and sufficient conditions for observability. Section 3
deals with algorithmic aspects: we show there how to check the conditions derived in
Section 2 in polynomial time.
In Section 4 we prove NP-completeness results for the
design of observable and partly observable graphs. Finally, in Section 5 we conclude and
describe some open questions.
2. Characterizing observability
Let G = (V, E) be a graph and C a set of colors. To every edge e ∈ E we associate one
(or more) color from C. A word w on C is the concatenation w1 . . . wT of symbols taken
from C; the length w of w is the number of its symbols. A subword w[i,j] : 1 ≤ i ≤ j ≤ w
of w = w1 . . . wT is the concatenation of the symbols wi . . . wj. We say that a path is
allowed by a word w if for all i the ith edge of the path has color wi. Finally, for a word
w and a set S ⊂ V , we denote by δw(S) the set of nodes q for which there exists a path
allowed by w beginning in a state in S and ending in q; δw(V ) is the set of nodes that can
be reached from a node in V with the color sequence w.
A graph is observable if there exists an integer T such that for all words w of length
w ≥ T , δw(V ) ≤ 1. In other words there is at most one possible end-node after a path
of length T or more. We shall also say that a graph is partly observable if there exists
an integer T such that for all word w with w ≥ T , there exists a 1 ≤ t ≤ T such that
δw[1,t](V ) ≤ 1.
The theorem below characterizes observable graphs. For a graph to be observable no color
sequence w can allow two separated cycles, that is, two cycles π, π′ : [0, w] → V such
that π(i) 6= π′(i), ∀i and the edges (π(i), π(i + 1)) and (π′(i), π′(i + 1) have identical colors
for every i. For instance, the graph in Figure 1 (b) is not observable because the color
sequence "solid-dashed-solid" allows both the cycle 1−3−2−1 and the cycle 5−1−4−5.
Another condition for a graph to be observable is that no node may have two outgoing
edges of the same color. This last condition only applies to the nodes of the graph that
are asymptotically reachable, i.e., that can be reached by paths of arbitrary long lengths.
In a strongly connected graph, all nodes are asymptotically reachable. For graphs that
are not strongly connected, asymptotically reachable nodes are easy to identify in the
decomposition of the graph in strongly connected components. We now prove that these
two conditions together are not only necessary but also sufficient:
Theorem 1. An edge-colored graph is observable if and only if no color sequence allows
two separated cycles, and no asymptotically reachable node of the graph has two outgoing
edges that have identical colors.
6
RAPHA EL M. JUNGERS AND VINCENT D. BLONDEL
Proof. ⇒: We prove this part by contraposition. Suppose that the first condition is
violated: if we have a word w allowing two separated cycles, then ww · · · w is an arbitrarily
long word that allows two separated paths. Suppose now that the second condition is
violated and let v be an asymptotically reachable node. For any T , we can find a path
of length T − 1 ending in v and v has two outgoing edges of the same color. Let w be
a color sequence allowing this path, and c the color of the two edges. We have then an
arbitrarily long word wc such that δwc(V ) ≥ 2.
⇐: Let us define T1 such that every path of length larger than T1 has its last node
asymptotically reachable. We consider two arbitrary compatible paths, that is, two paths
allowed by the same color sequence. We will show that they intersect between t = T1 and
t = T1 + n′2, with n′ the number of asymptotically reachable nodes. This will establish
observability of the graph, since two paths cannot split once they intersect (recall that
no node has two outgoing edges of the same color), and so actually all compatible paths
pass through the same node after t = T1 + n′2 steps. Suppose by contradiction that
the two paths π, π′ allowed by the same color sequence do not intersect during the last
n′2 steps, when they only visit asymptotically reachable nodes. Then by the pigeonhole
principle there are two instants t1, t2 such that (π(t1), π′(t1)) = (π(t2), π′(t2)), and the
first condition is violated: w[t1,t2] is a sequence that allows two separated cycles.
(cid:3)
One could think that the first condition in Theorem 1 suffices for the graph to be partly
observable, but this is not the case. Consider indeed the graph in Figure 3. For notational
simplicity we have put colors on the nodes rather than on the edges: one can interpret
these node colors as being given to the edges incoming in the corresponding node. This
graph satisfies the first condition but it is not partly observable. Indeed, the observed
sequence RGGRGG . . . has no subsequence allowing only one ending node.
Actually, the first condition implies a weaker notion of observability: A graph is partly
a-posteriori observable if for any sufficiently long observation, it is possible a posteriori
(that is, knowing the whole observation) to determine the state of the agent at at least
one previous instant. We do not develop this concept any further in this paper but it
should be clear that all the proofs that we provide in this and in the next section can
easily be extended to cover that case as well.
The graph (a) of Figure 1 is observable and the agent can be localized after an obser-
vation of length at most 3. The above proof provides an upper bound on the observation
length that is needed to localize an agent in an observable graph.
Corollary 1. In an observable graph, n2 − n observations suffice to localize an agent (n
is the number of nodes in the graph).
Proof. Let us suppose by contradiction that there exists a color sequence of length T >
n2 − n that allows two paths ending in different nodes. There are n2 − n different couples
of nodes (v1, v2) : v1 6= v2 and one can follow a similar reasoning as in the previous proof
to show that the graph is actually not observable.
(cid:3)
OBSERVABLE GRAPHS
7
R
R
*
G
G
**
R
R
G
G
Figure 3. A graph that is not partly observable, even though no color se-
quence allows two separated cycles. The colors on nodes can be understood
as assigned to the edges incoming in the corresponding nodes.
One could have expected the maximal value of T to be n, and one may wonder whether
the bound n2 −n in the theorem is tight. We give in Section 3 a family of graphs for which
the value of T is n(n − 1)/2 − 1; the quadratic increase can therefore not be avoided.
3. Verifying observability
In this section, we consider two algorithmic problems. The first problem is that of
reconstructing the possible positions of an agent in a graph for a given sequence of color
observations. That problem is easy; we describe a simple solution to it and give an
example that illustrates that the bound in Corollary 1 is tight. The second problem is
that of deciding observability.
Let us consider the first problem: we are given a color sequence and we would like to
identify the nodes that are compatible with the observed sequence of colors. A simple
algebraic solution is as follows. To every color c, there is an associated graph Gc for
which we can construct the corresponding adjacency matrix Ac. To a color sequence
w = w1, . . . , ww we then associate Aw, the corresponding product of matrices Aw =
Aw1 . . . Aww.
It is easy to verify that the (i, j)th entry of Aw is equal to the number
of paths from i to j allowed by w. The set δw(S) is therefore obtained by restricting
the matrix Aw to the lines corresponding to elements in S, and by taking the indices
of the nonzero columns. This simple algorithm is actually nothing else than the well-
known Viterbi algorithm for Hidden Markov Processes [7, 8], except that in our context
the probabilities are all equal to zero or one.
As an application, let us show that the quadratic dependance for the bound in Corollary
1 cannot be improved. We describe a family of colored graphs with n nodes by giving
the adjacency matrices corresponding to the different colors. Our set contains 2(n − 2)
matrices. For every 1 ≤ i ≤ n − 2 we construct two matrices: the first one, Ai, has
n − i entries equal to one, which are the entries (i + 1, i + 2), . . . , (n − 1, n), and (i, i);
8
RAPHA EL M. JUNGERS AND VINCENT D. BLONDEL
the second matrix, Bi, has two entries equal to one: the entries (i, i + 1) and (n, i +
2). A colored graph defined by this set of matrices is observable (the reader can verify
this by using the algorithm given below); but on the other hand the color sequence
An−2
2 B2 . . . An−2Bn−2, of length n(n − 1)/2 − 1, allows paths that have different
end nodes. This is straightforward to check by computing the corresponding product of
matrices and check that it has two columns with nonzero elements.
1 B1An−3
We now turn to the main question of this section: how can one determine efficiently if
a graph is observable? We have the following result.
Theorem 2. The problems of determining whether a colored graph is observable, or partly
observable, are solvable in O(n4m), where n is the number of nodes, and m is the number
of colors.
1, v′
2) if there is a color that allows both edges (v1, v′
Proof. The algorithm we propose uses the conditions given in Theorem 1. We first check
the condition that no color sequence allows two separated cycles. In order to do that, we
construct an auxiliary directed graph that we will denote G2 and whose nodes are couples
of distinct nodes in G: {(v1, v2) ∈ V 2 : v1 6= v2}. In G2 there is an edge from (v1, v2) to
(v′
2). This graph can be
constructed in O(n4m) operations, since there are less than n4 edges in G2 and for each of
these edges we have to check m colors. It is possible to design two separated cycles in G
allowed by the same color sequence if and only if there is one cycle in G2. This we can test
by a simple breadth first search. Now, we can check easily if an asymptotically reachable
node has two outgoing edges of the same color, and determine whether the network is
observable.
1) and (v2, v′
1, v′
1 and one edge from v1 or v2 to v′
We now would like to test whether the graph is partly observable; in order to do that,
we construct a new auxiliary directed graph G2, and once again we check the existence of
cycles in this graph. An easy way to obtain this new graph is to take G2, and to add some
edges: for every couple of nodes ((v1, v2), (v′
2)), we allow an edge if there is a color that
2. The graph G2 can
allows one edge from v1 or v2 to v′
also be constructed in O(n4m) operations. We claim that the graph is partly observable if
and only if this auxiliary graph is acyclic. Indeed, any path in the auxiliary graph ending
in node, say, (v1, v2) corresponds to a color sequence that allows paths ending in each of
the nodes v1 and v2. If the auxiliary graph contains a cycle, consider the path travelling
along the cycle during T steps; it corresponds to a word of length T such that for any
t ≤ T , the subword obtained by cutting the word at the tth character allows two paths
ending in different nodes, and the definition of partial observability is violated. On the
other hand, if the graph is not partly observable, there exist arbitrarily long sequences
such that all beginning subsequence allows at least two paths ending in different nodes.
This implies that G2 has a cycle.
Indeed, let us consider such a color sequence ω of
length T ≥ n2. There are two paths π, π1 : [0, T ] → V allowed by ω and ending in two
different nodes: π(T ) 6= π1(T ). If π(T − 1) 6= π1(T − 1), then there is an edge in G2 from
(π(T − 1), π1(T − 1)) to (π(T ), π1(T )). If π(T − 1) = π1(T − 1), then by construction of
ω there exists a path π2 allowed by ω[1...T −1] that ends in a node vT −1 6= π(T − 1). Thus
OBSERVABLE GRAPHS
9
we can iteratively define nodes vi : i ∈ [1, T ] of decreasing indices i such that there is a
path in G2 from (vi−1, π(i − 1)) to (vi, π(i)). Finally this path is of length T which is more
than the number of nodes in G2, so this graph is cyclic. Since G2 can be constructed in
O(n4m), the proof is complete.
(cid:3)
This algorithm also allows us to compute the time necessary before localization of the
agent:
Theorem 3. In an observable graph, it is possible to compute in polynomial time the
minimal T such that any observation of length T allows to localize the agent. The same
is true for partial observability.
Proof. Let us consider the graphs G2 and G2 in the above proof. The existence of an
observation of length T allowing two different last nodes (respectively, two paths with at
least two different allowed nodes at every time) is equivalent to the existence of a path of
length T in G2 (respectively, G2). Since the auxiliary graphs are directed and acyclic, it
is possible to find a longest path in polynomial time.
(cid:3)
4. Design of observable graphs is NP-complete
In this section, we prove NP-completeness results for the problem of designing observ-
able graphs. The problem is this: we are given an (uncolored) graph and would like
to color it with as few colors as possible so as to make it observable. Unless the graph
is trivial, one color never suffices. On the other hand, if we have as many colors as
there are nodes the problem becomes trivial. A simple question is thus: "What is the
minimal number of colors that are needed to make the graph observable"? We show
that this problem is NP-complete in two different situations: If we color the nodes and
want to make the graph observable, or if we color the edges and want to obtain a partly
observable graph. We haven't been able to derive similar results for the other two combi-
nations of node/edges coloring and observability/partial observability. In particular, we
leave open the natural question of finding the minimal number of edge colors to make a
graph observable. Even though the statements of our two results are similar, the proofs
are different. In particular, we use reductions from two different NP-complete problems
(3-COLORABILITY and MONOCHROMATIC-TRIANGLE).
Theorem 4. The problem of determining, for a given graph G and integer k, if G can be
made observable by coloring its nodes with k colors is NP-complete.
Proof. We have already shown how to check in polynomial time if a colored graph is
observable. So the problem is in NP. In the rest of the proof we provide a polynomial-
time reduction from 3-COLORABILITY to the problem of determining if a graph can be
made observable by coloring its nodes with 3 colors. This will then establish the proof
since 3-COLORABILITY is known to be NP-complete [4].
In 3-COLORABILITY we are given an undirected graph G and the problem is to color
the nodes with at most three different colors so that no two adjacent nodes have the same
10
RAPHA EL M. JUNGERS AND VINCENT D. BLONDEL
G
G
V1
B
e1
V3
G
e2
R
V2
Ve1''
R
G
G
Ve1'
B
V1'
B
V2'
R
G
G
V3'
G
Ve2'
B
R
Ve2''
Figure 4. Construction of the graph G′ and its coloration
color. For the reduction, consider an undirected graph G = (V, E). From G we construct
a directed graph G′ = (V ′, E′) such that G′ can be made observable by node coloring
with 3 colors if and only if G is 3-colorable. Figure 4 illustrates our construction. On the
left hand side we have a 3-colorable undirected graph. The corresponding directed graph
G′ is represented on the right hand side. The graph G′ can be made observable with
three colors. Here is how we construct G′ from G: For every node v in V , we define a
corresponding node v′ in V ′ and call these nodes "real nodes". For every edge e = (v1, v2)
in E, we define a node v′
2). The nodes
v′
e appear with double circles on Figure 4. The main idea of the reduction is as follows:
If G′ is observable by node coloring, then v′
2 cannot possibly have the same color
since there are edges pointing from v′
2, and so the corresponding nodes
in G do indeed have different colors.
e in V ′ and two outgoing edges (v′
e to both v′
1) and (v′
1 and v′
1 and v′
e, v′
e, v′
e and an edge (v′′
e, we add a new node v′′
The graph G′ we have constructed so far is rather simple and does not contain any
cycle; in the remainder of our construction we add nodes and edges to G′ so as to make
it strongly connected but in such a way that the graph G′ remains observable. For every
node v′
e are represented by a
square on Figure 4. We then order the edges in G: E = {e1, . . . , es}, and we add edges
ei+1) in G′ for every 1 ≤ i ≤ s − 1. Finally, for every real node v′ in G′ we add two
(v′′
nodes, with an edge from v′ to the first one, an edge from the first one to the second one,
and an edge from the second one to v′′
e1. These nodes and edges are represented in light
grey on Figure 4. The graph G′ is now strongly connected and we claim that it can be
made observable by node coloring with three colors if and only if G is 3-colorable. Let us
establish this claim.
e). The nodes v′′
ei, v′′
e , v′
If G′ can be made observable by node coloring with three colors then clearly G is 3-
colorable. Indeed, whenever two nodes in G are connected by an edge, their corresponding
real nodes in G′ are being pointed by edges emanating from the same node and so they
need to have different colors. So this part is easy.
OBSERVABLE GRAPHS
11
e2
e5
e1
e3
e6
e4
T1
T2
T3
T4
E1
E2
E3
E4
E5
E6
Figure 5. The main building block for the reduction.
Assume now that G is 3-colorable and let its nodes be colored in red (R), blue (B) or
green (G). We need to show that the graph G′ can be made observable by coloring its
nodes with three colors. We choose the following colors: every real node is colored by
its corresponding color in G, the nodes v′
e are
all colored in red. Finally, the remaining nodes (in light grey on the figure) are colored
in green. This colored graph is clearly observable. Indeed, any sufficiently long directed
path (in particular, any path whose length is larger than the total number of nodes in
G′) will eventually reach a succession of two green nodes followed by a red node. At the
end of such a sequence, the agent is at node v′′
e1 and his position in the graph can then be
observed for all subsequent steps since G′ has no two outgoing edges from a node leading
to nodes of identical colors. Thus G′ is observable and this concludes the proof.
e are all colored in blue and the nodes v′′
(cid:3)
The next result also deals with observability but the reduction is quite different.
Theorem 5. The problem of determining, for a given graph G and integer k, if G can be
made partly observable by coloring its edges with k colors is NP-complete.
Proof. We have shown in Section 3 how to check in polynomial time if a colored graph
is partly observable. Hence the problem is in P and to complete the proof it suffices
to exhibit a polynomial time reduction from some NP-complete problem. Our proof
12
RAPHA EL M. JUNGERS AND VINCENT D. BLONDEL
e3
T1
e1
e2
e4
T2
e5
T1
T1'
T2
T2'
E1
E4
E3
E5
E2
Figure 6. The complete reduction.
proceeds by reduction from MONOCHROMATIC-TRIANGLE. In MONOCHROMATIC-
TRIANGLE we are given an undirected graph G and we are asked to assign one of two
possible colors to every edge of G so that no triangle in G (a triangle in a graph is a
clique of size three) has its three edges identically colored. A graph for which this is
possible is said to be triangle-colorable and the problem of determining if a given graph
is triangle-colorable is known to be NP-complete [4].
Our proof proceeds as follows: From an undirected graph G we construct a directed
graph G that can be made partly observable with two colors if and only if G is triangle-
colorable. Thus we prove the somewhat stronger result that the problem of determining
if a graph can be made partly observable with two colors, is NP-complete.
As is often the case for NP-hardness proofs the idea of the reduction is simple but the
construction of the reduction is somewhat involved. For clarity we construct G in two
steps. The first step is illustrated on Figure 5. In this first step we construct an acyclic
graph G′ that will be used in the second step. The encoding of MONOCHROMATIC-
TRIANGLE is done at this step. The second step is illustrated on Figure 6 and uses
several replica of parts of the construction given in the first step. This second step is
needed in order to make the final graph G connected.
Let us now consider the first step. The left hand side of Figure 5 represents an undi-
rected graph G. The edges of G are colored but we do not pay attention to the colors at
this stage. At the right hand side of the figure is the graph G′ constructed from G. In
order to construct G′, we order the edges of G and for every edge ei in G we define two
nodes in G′ and a directed edge Ei between them; these edges have no nodes in common
OBSERVABLE GRAPHS
13
and we refer to them as "real edges" because they represent edges in the original graph G.
Next, we identify the triangles {eq, er, es : q < r < s} in G, we order them (for example,
according to the lexicographic order), we add a node Ti in G′ for every triangle in G, and
we add in G′ edges from the node Ti to all three real edges it is composed of, as illustrated
on the second level of the graph on Figure 5. Thus the out-degrees of the nodes Ti are
all equal to three. Finally, we construct a binary tree whose leaves are the nodes Ti. In
the example represented on the figure, the graph G has four triangles and so the tree has
four leaves Ti and has depth 2. (In general, the tree has depth equal to N = ⌈log2 S⌉
where S is the number of triangles in G.) The reduction will be based on the following
observation: Let the edges of G′ be colored and let us consider the sequences of colors
observed on paths from the root of the tree to the bottom nodes. We claim that the
edges of G′ can be colored with two colors so that all color sequences observed on these
paths are different if and only if G can be triangle-colored. We now establish this claim.
Assume first that G can not be triangle-colored and let Ti be a node corresponding to a
triangle whose edges are identically colored. There are three outgoing edges from Ti and
so two of them must have identical colors. Since these edges lead to real edges of identical
colors, G′ has two paths that have identical color sequences. Assume now that the graph
G can be triangle-colored. We color the real edges of G′ with their corresponding colors
in G and we color the edges of the tree so that paths from root to leaves define different
color sequences. It remains to color the outgoing edges from the nodes Ti. Every Ti has
three outgoing edges ui, vi, wi and these edges lead to real edges that are not all identically
colored. Since there are only two colors, two of the three edges must have identical colors.
Let ui and vi be the outgoing edges from Ti that lead to the two identically colored real
edges; we give to ui and vi different colors and choose an arbitrary color for the third
outgoing edge wi. When coloring the edges of G′ in this way, all paths from the root to
the bottom nodes define different color sequences and so the claim is established.
We now describe the second step of the construction. This step is illustrated on Figure
6 where we show a graph G (on the left hand side) and the corresponding graph G (on the
right hand side) constructed from G. Notice that the graph G we consider here is different
from the graph used in Figure 5. The construction of G consists in repeating a number of
times the tree structure described in the first step and then making the graph connected
by connecting the end nodes of the real edges to the roots of the trees through a sequence
of edges, as represented on the figure. For the sake of clarity we have represented only
two copies of the tree on the figure but the complete construction has 2S + 1 such copies
(in the example of the figure, 5 copies); the reason for having 2S + 1 copies will appear
in an argument given below. In the sequence of edges that make the graph connected
there are three nodes at every level and there are N + 3 levels where N is the depth of
the trees. The aim of this last construction is to make us able, by appropriately coloring
the edges, to determine when an agent reaches the root of a tree, but without making us
able to determine what particular root the agent has reached.
14
RAPHA EL M. JUNGERS AND VINCENT D. BLONDEL
We are now able to conclude the proof. We want to establish that G can be made
partly observable with two colors iff G is triangle-colorable.
Assume first that G is triangle-colorable and let the coloration be given with the colors
dashed (D) and Solid (S). We color G as follows: For the real edges and for the edges
in the trees we choose the colors described in step 1, for the edges leaving the end nodes
of the real edges and for the incoming edges to the roots we choose S, finally, all the
other connecting edges are colored by D. We claim that the graph G so colored is partly
observable. In order to establish our claim consider an agent moving in G. After some
time the agent will hit a long sequence of N + 3 connecting edges colored by D. The agent
will know that he is at a root as soon as he observes a S. He will however not be able
to determine what particular root he is at. We then use the argument presented in step
1 to conclude that, when the agent subsequently arrives at one of the end node of the
real edges, he knows exactly at what node he is. Therefore the colored graph G is partly
observable because, whenever the agent reaches the end node of a real edge, he knows
where he is.
We now establish the other direction: If G is not triangle-colorable then G cannot be
made partly observable with two colors. We establish this by proving that when G is
not triangle-colorable, then there exist two separated cycles in G that have identical color
sequences. Since the color sequences are identical and the cycles are separated an agent
observing that particular repeated color sequence can never determine on what cycle he
is and the graph is not partly observable. Let us assume that some coloration for G
has been chosen. Since G is not triangle-colorable, some triangles have all their edges
identically colored. Choose one such triangle and consider in every tree in G the path
that leads from the root to the node corresponding to that triangle in the tree. Any such
path defines a color sequence and there are at most S such sequences. Since there are
2S + 1 ≥ S + 1 trees, we may conclude by the pigeonhole principle that there are two
separated but identically colored paths that leave from distinct roots and lead to leaves
corresponding to the same triangle. If we require in addition that both roots have two
incoming edges colored with the same color, then the same result applies since there are
2S + 1 trees. By using the definition of G and the fact that G cannot be triangle-colored,
the construction of two identically colored distinct cycles can be continued by constructing
paths from the two nodes corresponding to the uni-colored triangle to the roots of their
corresponding trees. In this way one can construct two separated but identically colored
cycles in G that prove that G cannot be made observable with two colors and the proof
is complete.
(cid:3)
5. Conclusion and future work
We have introduced and analyzed a simple notion of observability in graphs. Observ-
ability in graphs is desirable in situations where autonomous agents move on a network
and one want to localize them; this concept appears in a number of different areas. We
OBSERVABLE GRAPHS
15
have characterized various forms of observability and have shown how they can be checked
in polynomial time. We have also shown that the time needed to localize an agent in a
graph can be computed in polynomial time. This time is in the worst case quadratic in
the number of nodes in the graph. We have also proved that the design of observable
graphs is NP-complete in two distinct situations. We leave some open questions and prob-
lems: Is the problem of making a graph observable NP-complete if colors are assigned to
the edges? How can one approximate the minimal number of colors? If a graph can be
made observable with a certain number of colors, how can the colors be assigned so as to
minimize the time necessary to localize the agent?
References
[1] M.-P. Beal. Codage symbolique. Masson, 1993.
[2] V. Crespi, G. V. Cybenko, and G. Jiang. The Theory of Trackability with Applications to Sensor
Networks. Technical Report TR2005-555, Dartmouth College, Computer Science, Hanover, NH,
August 2005.
[3] Y. Ephraim and N. Merhav. Hidden markov processes. IEEE Transactions on Information Theory,
48(6):1518 -- 1569, 2002.
[4] M. R. Garey and D. S. Johnson. Computers and Intractability; A Guide to the Theory of NP-
Completeness. W. H. Freeman & Co., New York, NY, USA, 1990.
[5] R. Jungers, V. Protasov, and V.D. Blondel. Efficient algorithms for deciding the type of growth of
products of integer matrices. submitted to publication.
[6] C. M. Ozveren and A.S. Willsky. Observability of discrete event dynamic systems. IEEE Transactions
on automatic and control, 35:797 -- 806, 1990.
[7] L. R. Rabiner. A tutorial on hidden markov models and selected apllications in speech recognition.
In A. Waibel and K.-F. Lee, editors, Readings in Speech Recognition, pages 267 -- 296. Kaufmann, San
Mateo, CA, 1990.
[8] A. J. Viterbi. Error bounds for convolutional codes and an asymptotically optimal decoding algo-
rithm. IEEE Transaction on Information Theory, IT-13:260 -- 269, 1967.
Division of Applied Mathematics, Universit´e catholique de Louvain, 4 avenue Georges
Lemaitre, B-1348 Louvain-la-Neuve, Belgium
E-mail address: [email protected]
Division of Applied Mathematics, Universit´e catholique de Louvain, 4 avenue Georges
Lemaitre, B-1348 Louvain-la-Neuve, Belgium
E-mail address: [email protected]
|
1106.0235 | 1 | 1106 | 2011-06-01T16:25:00 | Robust Agent Teams via Socially-Attentive Monitoring | [
"cs.MA",
"cs.AI"
] | Agents in dynamic multi-agent environments must monitor their peers to execute individual and group plans. A key open question is how much monitoring of other agents' states is required to be effective: The Monitoring Selectivity Problem. We investigate this question in the context of detecting failures in teams of cooperating agents, via Socially-Attentive Monitoring, which focuses on monitoring for failures in the social relationships between the agents. We empirically and analytically explore a family of socially-attentive teamwork monitoring algorithms in two dynamic, complex, multi-agent domains, under varying conditions of task distribution and uncertainty. We show that a centralized scheme using a complex algorithm trades correctness for completeness and requires monitoring all teammates. In contrast, a simple distributed teamwork monitoring algorithm results in correct and complete detection of teamwork failures, despite relying on limited, uncertain knowledge, and monitoring only key agents in a team. In addition, we report on the design of a socially-attentive monitoring system and demonstrate its generality in monitoring several coordination relationships, diagnosing detected failures, and both on-line and off-line applications. | cs.MA | cs | a f Ai(cid:28)
ia e ige
e Reea
h 12 2000 105 147
S bied 10/99; b ihed 3/00
Rb Age Tea via S
ia y Aeive iig
Ga A. aika
ga kii.ed
i id Tabe
abeii.ed
fai S
ie
e i e ad C e S
ie
e Deae
Uiveiy f S he Ca ifia
4676 Adia y Way
Age e CA 90292 USA
Aba
Age i dyai
i age evie i hei ee exe
e i
divid a ad g a. A key e ei i hw
h iig f he age
ae i e ied be e(cid:27)e
ive: The iig Se e
iviy b e. We iveigae hi
ei i he
ex f dee
ig fai e i ea f
eaig age via S
ia y
Aeive iig whi
h f
e iig f fai e i he
ia e aihi
bewee he age. We eii
a y ad aa yi
a y ex e a fai y f
ia y aeive
eawk iig a gih i w dyai
ex i age dai de
vayig
dii f ak diib i ad
eaiy. We hw ha a
ea ized
hee
ig a
ex a gih ade
e
e f
eee ad e ie iig a
eaae.
a a i e diib ed eawk iig a gih e i
e
ad
ee dee
i f eawk fai e deie e yig iied
eai
kw edge ad iig y key age i a ea. addii we e he deig
f a
ia y aeive iig ye ad deae i geea iy i iig ev
ea
diai e aihi diagig dee
ed fai e ad bh ie ad (cid:27) ie
a i
ai.
1. d
i
Age i
ex dyai
i age evie be ab e dee
diage
ad e
ve f fai e a ie Tyaa age 1997. F ia
e a b
gi ay be iey e behavi ay be ieia y di(cid:30)
edi
i
ai ay fai e
. Exa e f
h evie i
de vi a evie f
aiig h Ri
ke 1997; Ca de Sih C ea
he a Ceawi
z 1993
high (cid:28)de iy diib ed i ai Tabe h e aid Reb
S
hwab 1995; ia Tabe Se Ve Cade
hi awa a baa da
Aada 1997 ad i age bi
ake 1993; Ba
h 1998. The (cid:28) key e i
hi
e i exe
i iig Dy e Aki Dhi 1986; Ab ge
See 1988; Che Aa a 1992; Ree
e Tae 1994; Aki D fee Shi
1997; Ve a
k Cx 1998.
iig exe
i i i age eig e ie a age i i ee
i
e i w
e
exe
i deed a he ae f i ee Che eve e
1991; eig 1993; ake 1993; eig 1995; Gz a 1996; Tabe 1997.
iig ee i f ai
a ia
e i ea i
e ea ebe e y ea
h
he ad wk
e y gehe e aed ak:
(cid:13)2000 A A
e F dai ad ga a fa b ihe. A igh eeved.
aika ad Tabe
(cid:15) iig a w ea ebe
diae hei a
i ad a wih ea
ae he eaae ad
eae wih iefee
e. F exa e dive
f
a i a
vy
a dive wih iig he
a i he
vy a
dibad he
vy ad he he dive if
a beak dw.
(cid:15) iig a w ea ebe e ee a dyai
ifai
e f
eaig ew ifai. F ia
e if a dive i a
vy ee ha he he
a
i f f i dde y he ef he
a ife he exie
e f a ba
e
i ee deie die
y eeig i hee f.
evi wk ha iveigaed di(cid:27)ee way f iig i he
ex f ea f
eaig age. F exa e heei
a wk Shaed a Gz a 1999 ha
diig ihed bewee aive iig i whi
h a age i i(cid:28)ed whe a ii
hage e.g. via
i
ai ad a
ive iig i whi
h a age a
ive y eek
(cid:28)d whe a ii
hage e.g. via bevai ad ifee
e f bevab e
aib e. a
i
a i eeai have iveigaed he e f aive iig via
i
ai eig 1995 a
ive iig via a e
gii be D
fee 1995 a
ive i i
i iig via he evie Fee a Re
hei
1995 ad di(cid:27)ee
biai f hee ehd ake 1993; eig 1993; Tabe
1997; eh Ri
h Side 1999. aa
h i
ea y ei ahe: aive
iig i geea y e
eived a beig e
y ha a
ive iig b a e
e iab e Gz a 1999; be D fee 1995; aika Tabe 1998.
Regad e f he iig ehd badwidh ad
aia iiai hibi
a iig age f iig a he age f exe a he ie eig
1995; D fee 1995; Gz a 1996. Th a key e ei i hw
h i
ig f he age i e ied be e(cid:27)e
ive i ea eig 1993; Gz a
1996 1999. We
a hi
ha egig b e he iig Se e
iviy b e i.e. he
b e f e e
iviy i bevig he ad ifeig hei ae baed he be
vai f iig. A h gh i ha bee aied i he a y a faewk ad
iia
ai f awe wee vided eig 1993; Gz a 1996. F
ia
e he hey f Shaed a e ie age veify ha hei iei d
(cid:29)i
wih he f eaae Gz a 1996. weve he ehd by whi
h
h vei(cid:28)
ai
a ake a
e ae ef f f he iveigai Gz a 1996 .
308. Se
i 8 vide e deai e aed wk.
Thi ae begi adde he iig e e
iviy b e i ea by ivei
gaig iig e iee f e(cid:27)e
ive fai e dee
i. We f
iveigai
dee
ig fai e i he
ia e aihi ha idea y h d bewee age i a
ied ea. We
a
h iig f
ia e aihi
ia y aeive iig
di(cid:27)eeiae i f he ye f iig
h a iig f fai e i he
ge f age wad hei ga . ee he e
ia e aihi i ed dee a
e ai aib e f i e age ae. S
ia y aeive iig i he
vy
exa e iv ve veifyig ha age have
deiai ad headig ha hei
be ief i divig a a
vy ae a e
. F ia
e if he age ae beved
head i di(cid:27)ee die
i hey
ea y d have a
headig. Thi i di(cid:27)ee
ha iig whehe hei
he
headig ead wad hei ageed
deiai.
106
Rb Age Tea via S
ia y Aeive iig
iig e aihi i a ea
ia y aeive iig i a
ii
a ak i
iig ea ebe. Fai e aiai he ea e aihi
a fe ead
aahi
fai e he a f he ea a
k f
eaive behavi ad a
k f
diai. S
h fai e ae fe he e f idivid a age fai e
h a fai e
i a age e ad a
a. Th
ia y aeive iig
ve a age
a
f fai e ad e b idivid a eai.
We ex e
ia y aeive iig a gih f dee
ig eawk fai e
de vai
dii f
eaiy. We aa yi
a y hw ha deie he ee
e f
eaiy ab he a
a ae f ied age a
ea ized a
ive iig
hee
a g aaee fai e dee
i ha i eihe d ad i
ee
ee
ad d. weve hi e ie eaig ab i e hyhee a he a
a
ae f ied age ad iig a age i he ea. We hw ha a
ive
diib ed eawk iig e i bh d ad
ee dee
i
aabi iie
deie ig a
h i e a gih. Thi diib ed a gih: a e y a ig e
ib y i
e
hyhei f he a
a ae f ied age ad b iv ve i
ig y key age i a ea e
eai y a ea ebe. Uig a afai
he aa yi
a
we hw aa g e f
ea ized fai e dee
i i
a ex
i
diai e aihi.
We a
d
a eii
a iveigai f
ia y aeive iig i ea.
We ee a i eeed geea
ia y aeive iig faewk i whi
h he
exe
ed idea
ia e aihi ha ae be aiaied by he age ae
aed
he a
a
ia e aihi. Di
ea
ie ae dee
ed a ib e fai e ad diag
ed. We a y hi faewk w di(cid:27)ee
ex dyai
i age dai
i evi
e f iig vai
ia e aihi bh ie ad (cid:27) ie. Bh f hee
dai iv ve i e iea
ig age i
abaive ad adveaia eig wih
eaiie i bh e
ei ad a
i. e dai we vide eii
a e
f a
ive iig whi
h
(cid:28) aa yi
a e . ahe dai we hw hw
(cid:27) ie
ia y aeive iig
a vide aiaive eawk a iy feedba
k
a deige. We a vide iiia diagi
ed e f dee
ed fai e.
f
i hee ex ai i a
i
a a gih ha have g aaee ef
a
e i ea w d a i
ai. The a gih we ee eek
ee he e f
aive
i
ai baed iig whi
h i e iab e i ay dai ad ex e
he e f i ive key h e a e
gii a a a eaive. weve we d e
he e f
i
ai(cid:22)we i y eek vide e
hi e ha
a wk eve
whe
i
ai fai . aa yi
a g aaee f fai e dee
i de ad
eee h d whehe iig i de h gh
i
ai a e
gii.
Thi ae i gaized a f w: Se
i 2 ee ivaig exa e ad ba
k
g d. Se
i 3 ee he
ia y aeive iig faewk. Se
i 4 ex e
iig e e
iviy i
ea ized eawk iig. Se
i 5 ex e iig
e e
iviy i diib ed eawk iig. Se
i 6 deae he geea iy f
faewk by a yig i i a (cid:27) ie
(cid:28)g ai. Se
i 7 ee iveigai f ad
diia e aihi de . Se
i 8 ee e aed wk ad Se
i 9
de. The
w aedi
e
ai he f f hee eeed Aedix A ad e d
de
f he
ia y aeive iig a gih Aedix B.
107
aika ad Tabe
2. ivai ad Ba
kg d
The iig e e
iviy b e hi ae addee(cid:22)hw
h iig i e ied
f fai e dee
i i ea(cid:22)e f gwig f ai wih he igi(cid:28)
a fwae
aiea
e e(cid:27) i w f a i
ai dai. he dSAF dai a high
(cid:28)de iy ba e(cid:28)e d vi a evie Ca de e a . 1993 we have bee iv ved i he
deve e f yhei
he i
e i Tabe e a . 1995.
he RbC
e
i ai dai ia e a . 1997 we have bee iv ved i deve ig yhei
e aye ae a Adibi A aiza aika ea Ta i Tabe 1999. The
evie i bh dai ae dyai
ad
ex ad have ay
eaiie:
he behavi f he age e f i adveaia e
eaive e iab e
i
ai ad e a
i whi
h ay exe
e a ieded e
. Age i hee
evie ae heefe eeed wih
e iie f fai e deie he
deige be e(cid:27).
Se exa e ay eve i ae. The f wig w exa e ae a
a fai e
ha
ed i he dSAF dai. We wi e hee w i ae ad ex e
ia y aeive iig h gh hi ae:
Exa e 1. ee a ea f hee he i
e i age wee (cid:29)y a e
i(cid:28)ed way
i a give ii whee e f he ea ebe he
wa (cid:29)y fwad
wad he eey whi e i eaae aa
ke ad ad wai f i iga . A f he
age ied f he way i. weve d e a exe
ed e fai e e f
he aa
ke fai ed ee he way i. S whi e he he aa
ke
e
y aded
he fai ig aa
ke
i ed (cid:29)y fwad wih he
ee Fig e 1 f a
ee h
i aig hi fai e.
Exa e 2. a di(cid:27)ee afe a hee age ea
hed he way i ad dee
ed i
he
ha ge fwad ad idei(cid:28)ed he eey. he e a eage he waiig
aa
ke ji i ad aa
k he eey. e f he aa
ke did e
eive he eage
ad i eaied behid ide(cid:28)ie y whi e he
ad he he aa
ke
i ed he
ii a e.
We have
e
ed dze f ii a e i bh he dSAF ad RbC dai.
geea
h fai e ae di(cid:30)
ai
iae i deig ie d e he h ge be f
ib e ae. The age heefe eai y (cid:28)d hee ve i ve ae whi
h have
bee feee by he deve e ad he iig
dii ad
i
ai i a
e
ved i (cid:30)
ie: e f he fai e
ae eed did he age iv ved dee
e
a e
e
hei ee behavi. Ea
h age be ieved he he age be a
ig i
diai wih i i
e
i
ai wa e
eived f he he age idi
ae
hewie. weve he age wee vi aig he
abai e aihi bewee he
a he age
ae diagee wha a i beig exe
ed(cid:22)a
abai e aihi
fai e had
ed. e iiay eii
a e hw ha wad f 30 f fai e
eed iv ved e aihi vi ai e aihi fai e.
a beve hweve wee yi
a y i
k i
e hee fai e be
a e f he
ea
ia ibehavi f he age i hee
ae. They wee ab e ife ha a fai e
ha
ed deie kwig wha exa
y haeed. F ia
e eeig a aa
ke
i ig (cid:29)y ahead deie i eaae wi
hig a di(cid:27)ee a whi
h he h a
108
Rb Age Tea via S
ia y Aeive iig
Enemy
Scout (ahead) and failing attacker (trailing)
Landing attacker
Fig e 1: A a view di ay he dSAF dai i aig he fai e i Exa e 1.
The hi
k wavy ie ae
ie.
beve ifeed f he fa
ha e f he eaae he he aa
ke ha aded
i (cid:30)
ie f a beve dee
ha ehig ha ge ai(cid:22)wih kwig wha
he di(cid:27)ee a wa.
aa yi hwed ha he age wee iig ea
h he (cid:30)
ie y. w
eve a aive i f
i
i
ai bewee he age wa
ea y ia
i
a i
e: i he age ae eaig i a hi e evie; ii he
i
ai
vehead w d have bee hibiive; ad iii i fa
i wa he
i
ai e i
e ie f ha bke dw i e
ae. We heefe gh a
i
a way a
hieve
i
k dee
i f fai e baed he iied abig kw edge ha wa avai ab e
a iig age.
3. S
ia y Aeive iig
We begi wih a veview f he geea
e f a
ia y aeive iig y
e hw i Fig e 2.
i f: 1 a
ia e aihi kw edge bae
aiig
de f he e aihi ha h d h d ag he ied age eab ig ge
eai f exe
ed idea behavi i e f e aihi Se
i 3.1; 2 a age ad
ea de ig
e eib e f
e
ig ad eeeig kw edge ab he
ied age a
a behavi Se
i 3.2; 3 a e aihi fai e dee
i
e ha i f vi ai f e aihi ag ied age by
aig
he exe
ed ad a
a behavi Se
i 3.3; ad 4 a e aihi diagi
e
ha vei(cid:28)e he fai e ad vide a ex aai f he Se
i 3.4. The e ig
109
aika ad Tabe
ex aai diagi i he ed f e
vey e.g. by a egiai ye a
Sy
aa Eve
hik 1998 a geea e ae Ab ge See 1988.
Expected
Attribute
Values
Social Relationships Knowledge-Base
Expected Behavior
Relationship Detected
Failure
Diagnosis
Relationship
Failure
Detections
Actual Values
Actual
Behavior
Diagnosis
Agent/Team Modeling Component
Socially-Attentive Monitoring System
What agents to monitor,
What agent attributes
Observations,
Communications
Monitored Agent
Monitored Agent
Fig e 2: The geea
e f a
ia y aeive iig ye.
3.1 A w edge Bae f Re aihi de
We ake a e aihi ag age be a e ai hei ae aib e. A e aihi
de h e
i(cid:28)e hw di(cid:27)ee aib e f a age ae ae e aed he f
he age i a i age ye. Thee aib e
a i
de he be ief he d by he
age hei ga a a
i e
. F exa e ay eawk e aihi de
e ie ha ea ebe have a be ief i a ji ga Che eve e 1991;
eig 1995. A aia fai e aihi ake 1993; Ba
h 1998 e
i(cid:28)e
e aive dia
e ad ve
iie ha ae be aiaied by a g f age i
dai he i
e i . Cdiai e aihi ay e
ify ea e aihi
ha ae h d ag he a
i f w age e.g. b ie
a
a e
Cw 1991. A
h e aihi ae
ia (cid:22)hey ex i
i y e
ify hw i e
age ae a
ad wha hey ae be ieve if hey ae aiai he e aihi
bewee he.
The e aihi kw edge bae
ai de f he e aihi ha ae ed
h d i he ye ad e
i(cid:28)e he age ha ae ai
iaig i he e aihi. The
kw edge bae g ide he age de ig
e i e e
ig age be ied
ad wha aib e f hei ae eed be eeeed f dee
i ad diagi. i
ed by he fai e dee
i
e geeae exe
ai whi
h ae
aed wih
a
a e aihi aiaied by he age. Ad i vide he diagi
e wih
deai ed ifai ab hw age ae aib e ae e aed dive he diagi
e. i eeai f
ia y aeive iig i ea e f ye f
e aihi: fai e ii aiy a ex
i ad eawk.
F eawk iig we e he STEA Tabe 1997 geea dai
ideede de f eawk whi
h i baed Che ad eve e i ei
Faewk eve e Che e 1990; Che eve e 1991 ad Gz Side
ad a Shaed a Gz Side 1990; Gz a 1996 1999. weve
110
Rb Age Tea via S
ia y Aeive iig
he eawk de ay be ed iead f STEA. A h gh STEA i ed by
i ad
e age geeae
abaive behavi i i e ed hee ideede y
i evi
e f iig i.e. ied age ae a ed be a ea ad STEA i
ed i iig hei eawk. STEA ad he eawk de e.g. Che
eve e 1991; eig 1995; Ri
h Side 1997 e ie a be ief by ea e
be i hei ji ga ad a. Thi
haa
eii
i ed i eawk i
ye. The he e aihi de ae ed y i a e
day iig e. They
wi be di
ed i geae egh i Se
i 7.
3.2 w edge f ied Age ad Tea
The age de ig
e i eib e f a
iig ad aiaiig kw edge
ab ied age. Thi kw edge i ed
he a
a e ai ha exi
bewee age ae aib e whi
h ae
aed he idea exe
ed e ai. hi
e
i we de
ibe he a e
gii
aabi iie f he age de ig
e i
i eeai ad exeie i.e. he exe f he kw edge ha
d be aiaied
ab ied age a if e
eay. ae e
i hw ha i fa
iied ib y
ia
ae kw edge i (cid:30)
ie f e(cid:27)e
ive fai e dee
i. Th i eeai ay
e iized age de ig a gih ahe ha hee f
aabi iie. Se
i 3.4 wi
di
addiia age de ig
aabi iie e
eay f diagi.
3.2.1 Reeeai
F iig eawk e aihi we have f d ha eeeig age i e f
hei e e
ed hiea
hi
a ea
ive a eab e i
k iig f hei ae ad a
fa
i iae f he ifee
e f he ied age be ief ga ad bevab e a
i
i
e hey
a e he age de
ii
ee.
hi eeeai ea
ive a Fiby 1987; ewe 1990 f a ig e de
ii hiea
hy a ee ha eee he a eaive
ig
ee f ea
h age.
Ea
h ea
ive a i he hiea
hy heeafe efeed i y a a a ha e e
i
dii a efeed a e
dii f whe i i a i
ab e ad eiai
dii whi
h ae ed eiae ed a. A ea
h give e he age i
exe
ig a ig e ah a eaf h gh he hiea
hy. Thi ah i
ed f
a a di(cid:27)ee eve .
Fig e 3 ee a a i f
h a hiea
hy
eaed f he dSAF dai.
he
ae f Exa e 1 i he way i ea
h f he age wa exe
ig he ah be
giig wih exe
e ii a highe eve a h gh f y f igh a f y e
ave ig ad w eve . U ea
hig he way i hey wee a ed wi
h
f f y f igh a ad i de
ede wai a i. The aa
ke w d he
e e
j wai a a
hi d f wai a i whi e he
w d e e
fwad
ad i de
ede. f
e he fai ig aa
ke did dee
he way i ad he
eiai
dii f f y f igh a ad he e e
i
dii f wai a i
wee ai(cid:28)ed ad he fai ig aa
ke
i ed exe
e f y f igh a ad i
de
ede.
111
aika ad Tabe
Execute-Mission
Fly Flight Plan (F)
Wait at Point
(W)
Join Scout (J)
Ordered Halt (H)
Fly Route
Scout Forward
Traveling
Nap of the Earth
Contour
Low Level
Just Wait
Fig e 3: i f iea
hi
a Rea
ive a ibay f dSAF Dai Tea a
ae bxed. Thee ae ex aied i Se
i 3.3.
3.2.2 A
iii
F a a
i
a ee
ive whi e he age ay
eaive y e he iig
age hei w ae ig
i
ai i e ie
i
ai
hae be
(cid:30)
ie y fa e iab e ad e
e. Thi i f ae y ib e i ay ea ii
dai a exa e deae Se
i 2.
A eaive y a i ay e a e
gii ife he age bevab e ae
f hei bevab e behavi. Thi aa
h i i ive ad b i fa
e f
i
ai fai e. f
e he i ay i bee(cid:28) f f
ed
i
ai wih
he he age b w d be
ii
a y deede he.
T eab e a e
gii ig ea
ive a
he eeeai we have
e yed a ea
ive a e
gii a gih
a ed RES REa ie Si aed ea
ie. The key
aabi iy e ied i a w ex i
i aiea
e f hiea
hi
a
a hyhee a
hig ea
h age beved behavi whi e ig f hyhee whi
h
ae deeed i
e
e e f iig e. RES wk by exadig he
eie a ibay hiea
hy f ea
h de ed age ad aggig a ah a
hig he
beved behavi f he age beig de ed ee Aedix B f e d
de f he
a gih. e ii
ad exea kw edge ay be ed e iiae ah hyhee
whi
h ae deeed iaiae(cid:22)ideed
h he ii
wi be ex ed h y. RES
bai
aa
h i vey ii a evi wk i ea
ive a e
gii Ra 1994 ad
ea a
kig Tabe 1996 whi
h have bee ed
ef y i he dSAF dai
ad hae ay f RES eie. weve RES add be ief ifee
e
aabi iie
whi
h ae ed i he diagi
e di
ed be w Se
i 3.4.
Fig e 4 give a i i(cid:28)ed eeai f he a hiea
hie f a vaiai f Exa e
1 i whi
h a he age
e
y dee
ed he way i i.e. fai e ha
ed e
ha e a a ieediae eve have bee aba
ed i he (cid:28)g e. The
Fig e 4a ad he w aa
ke Fig e 4b 4
wi
hed f he f y f igh a a
deed by F he wai a i a deed by W. A ide beve ig RES
ife ex aai f ea
h age behavi by bevig he age. The
i e
112
Rb Age Tea via S
ia y Aeive iig
(cid:29)y ahead i eed ad a i de a
hig w eve e f he ib e (cid:29)igh ehd
de bh he f y f igh a F ad wai a i W a. Th hey ae bh
agged a ib e hyhee f he
exe
ig a hiea
hy. Sii a y a he
aa
ke ad RES e
gize ha hey ae exe
ig he j wai a. weve
hi a
a be ed i evi
e f eihe he W he deed ha a(cid:22)a a i
whi
h he he i
e ae deed by hei head ae ad iediae y. Th bh
ad W ae agged a ex aai f ea
h f he aa
ke ae a he e
d eve f
he hiea
hie. F a age RES idei(cid:28)e he a exe
e ii a he eve
a. hi i ai he a
a exe
ig ah f he age ae aked wih (cid:28) ed
aw. he idivid a de ig hyhee ha a
h he beved behavi ae aked
ig dahed aw. A ide beve f
e ha way f kwig whi
h f he
ib e hyhee i
e
.
Execute Mission
Execute Mission
Execute Mission
Wait-at-Point (W)
Fly-Flight-Plan (F)
Wait-at-Point (W)
Ordered-Halt (H)
Wait-at-Point (W)
Ordered-Halt (H)
Low-Level
Low-Level
Just-Wait
Just-Wait
Just-Wait
Just-Wait
(a)
(b)
(c)
Fig e 4: S
a ad Aa
ke b
a
a ad e
gized abbeviaed ea
ive a
hiea
hie.
e idivid a de ig hyhee ae a
ied f ea
h idivid a age ig a
e
gii i i eeai b eia y a by
i
ai he iig
age
bie he
eae ea de ig hyhee a he ae f he ea
a a wh e. The iig age e e
a ig e idivid a de ig hyhei f ea
h
idivid a age ad
bie he i a ig e ea de ig hyhei. Sevea
h
ea de ig hyhee ae ib e give i e hyhee f idivid a age. F
ia
e i Fig e 4 whi e a ea hyhee wi have exe
i ii a he
eve a hee ae eigh di(cid:27)ee ea hyhee whi
h
a be di(cid:27)eeiaed by hei
e
d eve a: WWW WW W W W FWW FW F W F . f he
beve i a ebe f he ea i kw wha i i exe
ig ie f b w d i have
i e hyhee ab i eaae ae. F ia
e if he aa
ke i Fig e 4b
i iig i eaae i hyhee a he e
d eve w d be WWW WW
FWW FW .
T avid ex i
i y eeeig a
biaia be f hyhee RES ex i
i y
aiai a
adidae hyhee f ea
h age idivid a y b a
biai
f idivid a de a ea hyhee. ead hee
biai ae i i
i y ee
eed. Th he be f hyhee ex i
i y aiaied gw iea y i he be
f age.
113
aika ad Tabe
3.3 Re aihi Vi ai Dee
i
The fai e dee
i
e dee
vi ai f he
ia e aihi ha h d
h d ag age. Thi i de by
aig he idea exe
ed e aihi hei
a
a aiea
e by he age. F eawk e
i(cid:28)
a y he e aihi de e ie
ea ebe a way agee whi
h ea a i ji y exe
ed by he ea ii a y
i Reibi iy eig 1995 ad Shaed a Gz a 1996. f hi
e iee fai i a
a iy i.e. he age ae exe
ig di(cid:27)ee ea a he a
eawk fai e ha
ed.
The bai
eawk fai e dee
i a gih i a f w. The ied age
a hiea
hie ae
eed i a dw ae. The dee
i
e e he
eawk de ag e
i(cid:28)
a a ea a ex i
i y eeeig ji a
iviy by
he ea hee a ae bxed i Fig e 3 5 ad 4. The ea a i e a deh
f he hiea
hie ae ed
eae ea de ig hyhee. F ea
h hyhei he
a f di(cid:27)ee age ae
aed dee
diageee. Ay di(cid:27)ee
e f d i a
idi
ai f fai e.
f di(cid:27)ee
e ae f d if he
ai ea
he idivid a
a ea heefe bxed i he (cid:28)g e fai e i dee
ed. divid a a
whi
h ay be
he by a age idivid a y i evi
e f ea a ae bxed i
hee (cid:28)g e ad ae had ed ig he e aihi a di
ed i Se
i 7
F ia
e e he fai ig aa
ke f Exa e 1 i iig he he a
a
ke. Fig e 5 hw i view f i w hiea
hi
a a he ef. The ah he
igh eee he ae f he he aa
ke wh ha aded. Thi ae ha bee i
feed i hi exa e f bevai ade by he iig aa
ke hee we ae
a ig ha he a e
gii
e ha e ed i e
e
hyhei f ea
h
age. We wi di
e ea ii
eig be w. Fig e 5 he di(cid:27)ee
e ha w d
be dee
ed i aked by he aw bewee he w a a he e
d eve f he .
Whi e he fai ig aa
ke i exe
ig he f y f igh a ea a he ef he
he aa
ke i exe
ig he wai a i ea a he igh. The diageee
whi
h ea a i be exe
ed i a fai e f eawk.
Execute Mission
Execute Mission
Fly-Flight-Plan
Wait-at-Point
Fly-Route
Traveling
Low-Level
Just-Wait
Fig e 5: Caig w hiea
hi
a a. The di(cid:27)ee
e i a eve 2.
114
Rb Age Tea via S
ia y Aeive iig
Dee
ig diageee i di(cid:30)
wih i e ea de ig hyhee i
e hey
ay i y
adi
y e wih ee
fai e dee
i: Se hyhee ay i
y ha a fai e had
ed i he ea whi e he ay . Uf ae y hi i
be exe
ed i ea ii
a i
ai. F ia
e Fig e 4 Se
i 3.2 hw evea hy
hee ha ae ib e baed he ae bevai. weve e f he hyhee
WWW i ie fai e ha
ed(cid:22)a he age ae i ageee whi
h ea a
i exe
ig(cid:22)whi e ahe hyhei FW i ie fai e have
ed.
T ii eaig y a a be f ea hyhee whi e ei
ig
fai e dee
i
aabi iie we e a diabig ai he ii
ha ak ea de ig
hyhee by he eve f
hee
e hey eee. Thi he ii
i vided a a iiia
i. ae e
i wi exaie addiia he ii
.
De(cid:28)ii 1. The
hee
e eve f a i age de ig hyhei i de(cid:28)ed a he
ai f he be f age de ed he be f a
aied i he hyhei.
Thi de(cid:28)ii e i a aia deig f he hyhee e f he ea
hee
hyhei e ha aig ea
h age a di(cid:27)ee a ha i ea ae he
hee hyhei ha aig he ae a a ea ebe. F ia
e he
hyhei FW w d have he we eve f
hee
e 1 i
e i i ie
ee
beakdw f eawk(cid:22)evey age i exe
ig a di(cid:27)ee a. The hyhei WWW
w d have a
hee
e eve f 3 he highe eve f
hee
e f he g f hee age
i
e hey ae a aiged he ae a. Raked bewee he w d be he hyhei
WW wih a ig e eawk fai e diageee W ad ad a
hee
e eve f
3/2.
The dee
i
e e e
a ig e axia y
hee ea de ig hyhei
ie bke ad y. The i ii f ig
hee
e i ha fai e agee
deie he age ae a eawk. Th we exe
e ageee ha diagee
e i he ea. The
hee
e eve f a ea hyhei i ivee y e aed he
be f eawk fai e i ied by he hyhei. Se e
ig a axia y
hee hy
hei heefe
ed he ii be f fai e he ii
y ed
i diagi a
he C e de ee 1992.
F he
ae dei
ed i Fig e 4 he
ee dee
i
e ay be
e
a ized a f w
. S e ha e f he aa
ke whe hiea
hy i de
ibed i
1
Fig e 4b i iig he ea. Fi i
e
he a hyhee a he f
he hiea
hy f ea
h age i
dig ie f . hi
ae hey ae {exe
e ii}
{exe
e ii} {exe
e ii}. y e ea de ig hyhei
a be b i
f hee: exe
e ii exe
e ii exe
e ii. Si
e hi hyh
ei hw diageee
a hi eve he
e
i e he e
d eve .
ee he hyhee f he (cid:28) age he ef ae {FW} f he iig e
d age
i
e i kw i w ae hee i y e ibi iy {W} ad f he hid age {W }.
A we aw abve he axia y ea
hee hyhei i WWW whi
h i e e
ed. Si
e
i de idi
ae fai e he
e
i e he hid eve . ee he age ae
exe
ig idivid a a ad he
ai
e . A gih 2 i Aedix
B vide geae deai ab hi
e.
1. he i eeai ay ake e f iized a gih i whi
h he he ii
ae iegaed i
he age de ig a gih.
115
aika ad Tabe
Whe b ea ae id
ed a di(cid:27)ee
e bewee ea a ay be ex aied by
he age i ei beig a a f di(cid:27)ee b ea. S b ea ebe i have
agee bewee hee ve he ji b ea a b hee ay di(cid:27)e f e
b ea he ex. F w e a e ha he ea de
ideai ae i e
ea a de(cid:28)ed i De(cid:28)ii 2. We ake hi de(cid:28)ii i evi
e f ae aa yi
a e
i whi
h i wi aea a a
dii. We e he i e f b ea i Se
i 7.1.
De(cid:28)ii 2. We ay ha a ea T i i e if i a hiea
hy iv ve di(cid:27)ee
ea a whi
h ae be exe
ed by di(cid:27)ee b ea.
iive y he idea i ha i a i e ea a ebe f he ea ji y exe
e
ea
h f he ea a i he hiea
hy. Thi de(cid:28)ii i ewha ii a he de(cid:28)ii
f a g d ea i iy j gbeg Ra Sebeg Tidha Wee 1992 b i
de a w b ea ebe f a ea have a ji a whi
h i di(cid:27)ee ha ha
f he ebe.
3.4 Re aihi Diagi
The diagi
e
a ex aai f he dee
ed fai e ideifyig he
fai e ae ad fa
i iaig e
vey. The diagi i give i e f a e f age
be ief di(cid:27)ee
e i
ie
ie ha ex ai he fai e aiai he e aihi.
The aig i f hi
e i he dee
ed fai e e.g. he di(cid:27)ee
e i ea a.
The diagi
e he
ae he be ief f he age iv ved d
e a e f
i
ie be ief ha ex ai he fai e.
Tw b e exi i a
i
a a i
ai f hi
ed e. Fi he iig
age i ike y have a
e a f he be ief he d by he ied age i
e i
i feaib e i a
i
e
i
ae a he age be ief ea
h he. Se
d ea
h
age i a ea w d dai ay have ay be ief ad ay f he wi vay ag he
age h gh f he wi be ie eva he diagi. Th e eva kw edge
ay be i y be a
eib e ay be hidde i ai f ie eva fa
.
T gai kw edge f he be ief f ied age wih e yig
i
ai
he diagi
e e a
e f be ief a
ii. The age de ig
e
ig RES i i eeai aiai kw edge ab he e e
i ad eiai
dii f e
gized a hyhee. F ea
h e
gized a hyhei he d
e ig
e ife ha ay eiai
dii f he a ae be ieved be fa e
by he ied age i
e i ha eiaed he a. We have a f d i ef
e a addiia he ii
ad ife ha he e e
i
dii e
dii f
ay a whi
h ha j beg exe
i ae e. The idea i ha whe a a i e e
ed
f exe
i i e
dii ae ike y h d a ea f a h eid f ie. Thi
he ii
iv ve a ex i
i a i he a f ye ha he ew a i
e
gized a a i begi exe
i. Deige i he dai wi eed veify
ha hi a i h d.
F ea
h age i he ifeed eiai ad e e
i
dii ake a e f be ief
B
f he age. F ia
e e a age i hyheized have j wi
hed f
i
exe
ig f y f igh a wai a i. RES ife ha he age be ieve ha
he way i wa j dee
ed a e e
i
dii f wai a i.
addii
116
Rb Age Tea via S
ia y Aeive iig
RES ife ha he age be ieve ha a eey wa ee ad ha de wa
e
eived f bae ha he ii egaed eiai
dii f wai a i.
T deeie he fa
ha ae e eva he fai e he diagi
e e
he eawk de . The eawk de di
ae whi
h be ief he age h d be
a y be ieved by a he age i he ea. Ay di(cid:27)ee
e ha i dee
ed i he
be ief i a
eai fai e a he ea ebe d agee i e whi
h ageee
i aday ai
iai i he ea. The eawk de h e
i(cid:28)e ha he
be ief
aied i he B
e h d be a ad h d heefe be
ie:
i
[
i
B
6`?
i
f a i
ie
y i dee
ed he diagi
ed e k f
adi
i diagee
e ha w d
a e he di(cid:27)ee
e i ea a e e
i. A di(cid:27)ee
e i be ief eve
a he diagi a wig he iig age iiiae a
e f e
vey e.g. by
egiaig ab he
(cid:29)i
ig be ief a e a . 1998.
F exa e a hw i Se
i 3.3 he w aa
ke i Exa e 1 Se
i 2 di(cid:27)e
i hei
hi
e f a ea a: e aa
ke i
i ig exe
i f he f y f igh a
a i whi
h he he i
e (cid:29)y i fai. The he aa
ke ha dee
ed he way
i eiaed f y f igh a ad ha wi
hed wai a i adig iedi
ae y Fig e 5. Whe he fai ig aa
ke i i ea ae i dee
a di(cid:27)ee
e i
he ea a Se
i 3.3 ad he dee
ed di(cid:27)ee
e i aed diagi. The fai ig
aa
ke ake he f wig ifee
e:
1. F y f igh a ha hee eiai
dii: a eeig he eey b dee
ig
he way i
e
eivig a de ha . The fai ig aa
ke ef hiea
hy i
Fig e 5 kw i w be ief ha e f hee
dii h d ad h
B
= f:W ay i; :Eey ; : a deg
1
2. Wai a i ha e e e
i
dii: he way i ha bee dee
ed.
eiai
dii i ha he
ha e a eage ji i havig idei(cid:28)ed
he eey ii. The diagi
e i hi
ae heefe ife ha f
he he aa
ke igh hiea
hy i Fig e 5
B
= fW ay i; :S
eageRe
eivedg
2
The
B
[ B
= f:W ay i; W ay i; :Eey ; :S
eageRe
eived; : a deg
1
2
whi
h i i
ie.
The i
ie
y diageee bewee he aa
ke i
f:W ay i; W ay ig i.e.
adi
y be ief ab W ayi. Th w he fai ig
aa
ke kw ha i ea ae ha ee he way i.
a
he ie y ada
hi be ief heeby eiaig i w f y f igh a ad e e
ig wai a i
117
aika ad Tabe
i ay
he he e
vey a
i
h a egiaig wih he he aa
ke whehe
he way i ha bee ea
hed.
We have f d hee diagi
ed e be ef i ay f he fai e dee
ed
by
ia y aeive iig ee Se
i 4 f eva ai ad di
i. weve
i
e hi ae f
e he iig e e
iviy b e i dee
i we eave f he
iveigai f he diagi
ed e f e wk.
4. iig Se e
iviy i Cea ized Teawk iig
Uig he
ia y aeive faewk f Se
i 3 we yeai
a y exaie a fai e
e ai f Exa e 1 ad 2 Se
i 2 de a
ea ized eawk iig
(cid:28)g ai whee a ig e ea ebe i iig he ea. We vay he age
fai ig aa
ke aa
ke ad
e
. ad he e f he iig age aa
ke
. We e he eii
a e f dee
ig ad diagig fai e i a
ae.
Uig hee eii
a e a a g ide we ex e
ea ized eawk iig aa y
i
a y. We hw ha eve de iig
eaiy
ea ized eawk iig
a vide eihe d
ee dee
i e b bh.
A a aig i f ex ai he iig age e a ig e axia y
hee ea de ig hyhei a di
ed i Se
i 3.3. We begi wih Exa e 2.
The a de f exe
i i wai a i W f wed by ji
. D ig
he exe
i f W he w aa
ke ad ad wai f he
vi a y ideify he
eey ii. U idei(cid:28)
ai he
ed he a eage ji i whi
h
igge he e e
i f he a ad he eiai f he W a. Whe exe
ig
he
hve a w a i de waiig f he aa
ke ji i. Ay fai e hee ae
he a f he aa
ke hey
a e
eive he eage he a f he
i
a ed i. Thee fai e aie f ia
e if he adi i bke ea ebe
ae f age. Whe a age fai i
i e exe
e W iead f wi
hig .
Tab e 1 aize he e ai f Exa e 2. The e ai be aea
i he (cid:28)
. The ex hee
hw he a
a a e e
ed by he hee
age A1 A2 ad A3 i ea
h e ai. The e
d a
hw whehe a
e aihi fai e ha
ed i ea
h
ae i.e. whehe diageee exi bewee he
age. Fia y he a
deai he hyi
a
dii i ea
h
ae. Thee ae (cid:28)ve
ib e fai e e ai:
ae 1 e f he age fai ed.
ae 2 ad 3 e
aa
ke fai ed.
ae 4 he
fai ed ed a eage bh aa
ke fai ed
e
eive i.
ae 5 he
de ideify he eey ii eage i e
ad a hee age
i e exe
e he W a. he e ai ae ib e
i
e aa
ke
a wi
h he a wih he
.
F ia
e
ae 2 i Tab e 1
ed Exa e 2. The
A3 ha dee
ed
he eey wi
hed a ad e a eage he aa
ke ji i. e aa
ke
A2 e
eived he eage wi
hed a ad bega (cid:29)yig wad he
. weve
he eaiig aa
ke A1 fai ed e
eive he eage ad i aiai i ii
i ig exe
e W ad fai ig wi
h . Si
e he age ae ge i ageee
whi
h ea a h d be ji y exe
ed a eawk fai e ha
ed. Cdii
i wee ed i he igia fai e
ae i f he
eage. weve
fai e i
i
ai e ed i hee iig
dii be edeed e e.
118
Rb Age Tea via S
ia y Aeive iig
Cae
A
a Exe
ig a
Re aihi Fai e
hyi
a
#
Aa
ke A1 Aa
ke A2
S
A3
ed?
Fai e
1
2
W
A1 fai e
eive
3
W
A2 fai e
eive
4
W
W
A3 eage
5
W
W
W
Eey idei(cid:28)ed
Tab e 1: A ib e fai e e ai f he bke adi ik
eai Exa e 2.
e key i e i aied by
ae 5 i Tab e 1. ee d e he
iabi iy ideify
he eey ii eha d e fai e he
a eha be
a e he eey
i i y hee he hee he i
e i eai i ageee ha he eey ha
bee idei(cid:28)ed. ee eve h gh
ea y he i ae fai ig ake ge wad he
ak ga he
i e ea
h f he eey ide(cid:28)ie y e aihi fai e
i akig a
e i
e he age ae aiaiig he eawk e aihi whi e fai ig
ake ge. Thi
ea y deae ha a fai e ae e
eai y e aihi
fai e.
Tab e 2 ee he e f he
iig i eaae i Exa e 2 ig a
axia y ea
hee hyhei a he bai f dee
i. The (cid:28)
agai hw
he
ae be f efee
e i Tab e 1. The ex hee
hw he
A3
hyhei ab wha a ea
h age i exe
ig a
dig he axia
hee
e
he ii
. The ex w
hw whehe a fai e wa dee
ed ad whehe i wa
diaged
e
y. The a
hw he dee
i
a di
ed be w.
Cae
A3 yheized Exe
ig a
Re aihi Fai e Diagi
Dee
i
#
Aa
ke A1 Aa
ke A2
S
A3
Dee
ed?
S
e?
C a
1
/a
T e egaive
2
W
T e iive
3
W
T e iive
4
T e iive
5
W
W
W
/a
T e egaive
Tab e 2: S
A3 iig e i a e ai f Exa e 2.
F exa e
ae 2 i Tab e 2 hw he e f he
iig i he igia
fai e i Exa e 2 Se
i 2. Uig RES ad e e
ig a axia y
hee hyh
ei he
hyheize ha he vig aa
ke i exe
ig W
ae 2
2
whi e he vig aa
ke i exe
ig
ae 2
3. The
f
e kw ha
i w e e
ed a
ae 2
4. A vi ai f he eawk e aihi i h
dee
ed
ae 2
5 i
e A1 W i i ageee wih he e f he ea .
F hee he diagi wa
ef i ideifyig he
a e f he fai e i.e. he
fa
ha he eey ii ha bee idei(cid:28)ed by he
b kw edge f hi
wa aed he fai ig aa
ke
ae 2
6.
119
aika ad Tabe
The a
f Tab e 2 hw he dee
i
a f ea
h fai e. The dee
i
a
f a
ae
a be e f: e iive e egaive fa e iive ad fa e egaive. Thee
ed he f wig ib e iig
e: A e iive i a
e
whee a e aihi fai e ha a
a y
ed ad ha bee dee
ed. A e egaive
i whee fai e ha
ed ad he ye
e
y e e i beig dee
ed.
A fa e iive i whee fai e ha
ed b he ye evehe e i
e
y
dee
e ad a fa e egaive i whee a fai e ha
ed b he ye fai
dee
i. Tab e 2 hw ha i a e ai f Exa e 2 he eawk iig
e
hi e did e
e he b eai
fa e iive fa e egaive
ae.
A
e k a hee e hi a a key
ib i f hi ae i addeig he
iig e e
iviy b e: E(cid:27)e
ive fai e dee
i
a ake a
e deie he e f
eai iied kw edge ab ied age.
ae 4 f Tab e 2 he i
ig age wa ab e dee
he fai e deie beig wg ab he ae f he age
iv ved. The
be ieve ha he w aa
ke ae exe
ig he deed ha
a b hey ae a
a y exe
ig W. i e e
ed whe a
ad i e
eived f
head ae ha exe
i ad hve i a
e. F he
ee
ive a hv
eig aa
ke
a heefe be ifeed be exe
ig W. Th w e a y aked
axia y
hee hyhee exi: he w aa
ke ae eihe bh exe
ig W bh
exe
ig . A ad e e
i wa ade ad i hi
ae e ed i he wg hyhe
i beig e e
ed. evehe e a vi ai f he eawk e aihi wa dee
ed a
eihe W agee wih he
.
weve a he a
f
ae 4 hw i Tab e 2 he diagi
ed e ae
eiive he e e
i f he ea de ig hyhei. The hyhei ed i hi
ae
de
e
y e(cid:29)e
he e ae f he age ad deie he
e
dee
a fai e i hi
ae he diagi
ed e fai vide
e
diagi he
diagi wa
ef i he w he fai e
ae. Thi hee eea i he
eii
a e we vide be w: diagi fai ed wheeve he hyhei
he wa
i
e
a h gh i wa (cid:30)
ie f dee
i. We ied ex e way ive
he a
a
y f he hyhee i f e wk ad heefe wi adde hi i e
i hi ae aye. The fai e dee
i
aabi iie ae a igi(cid:28)
a ivee i
hee ve i
e he age kw wih
eaiy ha a fai e ha
ed eve if hei
diagi f i i i
e
.
ay
ia ad hyi
a fai e ae
ef y
a ed ig he ea
hee
e
he ii
f iig e e
iviy. fa
i e ai f Exa e 2 ae
whi
h e f he age wa he i a fai e aiai e aihi i.e. a hyi
a
fai e ex
e f e i whi
h he ea eai i ageee wee dee
ed e iab y
a h gh eie diagi fai ed. Thi e i ee
ia y iig
ideig ha
y a ig e age did he iig. evi iig ehd
dii i
ad
i
ai wee ab e dee
he fai e deie beig ed by a hee
age .
Tab e 3 ad 4 ee f he eii
a e he bai f Exa e 1. Tab e 3
ee a fai e e ai f Exa e 1 i he ae fa a Tab e 1. The a
de f exe
i f hee a i a f w: A age ji y exe
e he f y f igh a
F a i hey dee
he way i. They he wi
h he wai a i W a
i whi
h he w aa
ke ad whi e he
i e (cid:29)y ahead ideify he eey.
120
Rb Age Tea via S
ia y Aeive iig
The fai e hee ae a he a f he age dee
he way i ad h wi
h
he W a.
Cae
A
a Exe
ig a
Re aihi Fai e
hyi
a
#
Aa
ke A1 Aa
ke A2
S
A3
ed?
Fai e
1
W
W
W
2
F
W
W
A1 vii fai
3
W
F
W
A2 vii fai
4
F
F
W
A1 A2 vii fai
5
W
W
F
A3 vii fai
6
F
W
F
A1 A3 vii fai
7
W
F
F
A2 A3 vii fai
8
F
F
F
A123 vii fai
Tab e 3: A fai e e ai f he dee
ed way i
eai Exa e 1.
Cae
A1 yheized Exe
ig a
Re aihi Fai e
Dee
i
#
Aa
ke A1 Aa
ke A2
S
A3
Dee
ed?
C a
1
W
W
W
T e egaive
2
F
W
F
T e iive
3
W
F
W
T e iive
4
F
F
F
Fa e egaive
5
W
W
W
Fa e egaive
6
F
W
F
T e iive
7
W
F
F
T e iive
8
F
F
F
T e egaive
Tab e 4: Aa
ke A1 iig e i a e ai f Exa e 1.
Tab e 4 ee he iig e f a e ai f Exa e 1. ee he
aa
ke A1 i iig he ea ig agai a axia y ea
hee hyhei i
dee
ig fai e. The e hw ha A1 i
ef i dee
ig a eawk fai e
b w
ae 4 5 high ighed i b d fa
e.
Thee w fa e
e ae bh fa e egaive. bh hee
ae he iig
aa
ke A1 i
ked a i
e
hyhei f he
i
e he
a
i ead
abig ieeai. The
i (cid:29)y fwad
he eey if i dee
ed he
way i a W b a if i did he i w d be (cid:29)yig (cid:16)i fai(cid:17)(cid:21) a F. The
e f he axia ea
hee
e he ii
a e A1 efe a hyhei i whi
h he
i i ageee wih he aa
ke whe i fa
i i . F exa e i
ae 4 he w
aa
ke have fai ed dee
he way i ad ae exe
ig F. bevig he
he
iig aa
ke A1 i e whehe he
i exe
ig F W. weve be ievig
ha he
i exe
ig F e i a axia y
hee ea de ig hyhei a
he age ae i ageee whi e be ievig ha he
i exe
ig W e i a e
121
aika ad Tabe
hee hyhei. Th A1 e e
a wg hyhei whi
h i hi
ae fai dee
he eawk fai e.
The axia ea
hee
e he ii
a dee
fai e deie ig i
e
hy
hee. Uf ae y
h hyhee
a a ead fa e egaive a we have ee i
Tab e 4. weve e f exeie e ed i a fa e iive e i.e. a e
i whi
h he ye dee
ed a fai e b i ea iy e had
ed. Th he he ii
vided d e i hee
ae. We ae ab e fa y ve hi ey h d i
geea whe he axia ea
hee
e he ii
i ed.
Fi we adde a ae f ai. e a age A i a age B whi
h i
exe
ig e a . We dee by A; B = he e f age de ig hyhee
ha A age de ig
e
baed B bevab e behavi d ig he
exe
i f . he wd A; B = i A e f a a ha a
h B bevab e
behavi. e ha whe A i ie f i ha die
a
e i w ae ad
A; A= =f g. Uig he de ig ai we ake he f wig de(cid:28)ii whi
h
g d a i ab he de yig kw edge ed i iig:
De(cid:28)ii 3. Give a iig age A ad a ied age B we ay ha A
age de ig f age B i
ee if f ay a ha ay be exe
ed by B 2
A; B = .
The e A; B = wi yi
a y i
de he a
hig hyhee beide he
e
hyhei b i g aaeed i
de . F wig hi de(cid:28)ii f idivid a age
de ig
eee we
a de(cid:28)e g wide ea de ig
eee:
De(cid:28)ii 4. e A be a age iig a ea T f age B
; ; B
. We ay ha
1
A ea de ig f he ea T i
ee if A age de ig f ea
h f B
; ; B
i
1
ee.
De(cid:28)ii 4 i
ii
a g aaee he
aabi iie we wi ex e aa yi
a y i hi
e
i ad he ex. geea y h d i e f RES i he dSAF ad RbC
dai ad we ake i ex i
i hee i evi
e f a i
ai f he e
hi e i he
dai.
Aed wih hee de(cid:28)ii we w fa ize he fai e dee
i
aabi iie g
geed by he eii
a evide
e i Thee 1.
Thee 1. e a iig age A i a i e ea T . f A ea de ig
f T i
ee ad A e a axia y ea
hee hyhei f dee
i he he
eawk fai e dee
i e ae d.
f. We wi hw ha if fai e ha
ed e wi be dee
ed ad h ha ay
fai e ha i dee
ed i i fa
a fai e. e a
; : : : ; a
be he age ebe f T . Ea
h
1
age a
i exe
ig e a
1 (cid:20) i (cid:20) . Th
e
ive y he g i exe
ig
i
i
; : : : ;
. f fai e ha
ed he a he age ae exe
ig he ae a
1
0
i.e. 8i;
=
. Si
e A ea de ig i
ee he
e
hyhei
; : : : ;
i
0
0
0
i gig be i he e f ea de ig hyhee . Si
e i i a axia y ea
hee hyhei eihe i wi be e e
ed ha a di(cid:27)ee hyhei f he ae
hee
e eve wi be e e
ed. Ay hyhei wih he ae
hee
e eve a he
e
e i ie fai e i dee
ed. Th he dee
i
ed e i d.
122
Rb Age Tea via S
ia y Aeive iig
Deie
eaiy i he kw edge ed d fai e dee
i
a be g aaeed ig
he axia ea
hee
e he ii
. Thi i e awe he iig e e
iviy
b e. weve a we have ee i Tab e 4 e fai e ay a dee
ed ig hi
he ii
i.e. i ay e i fa e egaive. Dee
i ig axia ea
hee
e
ay heefe f ae y be i
ee. We ay efe iig ye be
ee(cid:22)g aaeed dee
a eawk fai e.
We heefe exeieed wih he axia ea i
hee
e he ii
he ivee f
he axia ea
hee
e he ii
. Thi he ii
efe hyhee ha gge e
fai e ahe ha e. Tab e 5 give he iig aa
ke A1 view f he ea
ii a Tab e 4 b ig a axia y ea i
hee hyhei. hw ha ideed
ig a axia y ea i
hee hyhei wi ead he fa e egaive dee
i
i
ae 4 ad 5 i
a hee
ae i Tab e 4.
Cae
A1 yheized Exe
ig a
Re aihi Fai e
Dee
i
#
Aa
ke A1 Aa
ke A2
S
A3
Dee
ed?
C a
1
W
F
Fa e iive
2
F
W
T e iive
3
W
F
F
T e iive
4
F
F
W
T e iive
5
W
F
T e iive
6
F
W
T e iive
7
W
F
F
T e iive
8
F
F
W
Fa e iive
Tab e 5: Aa
ke A1 iig e i a e ai f Exa e 1 ig ea
i
hee
e.
G ided by hee e we fa y hw ha he ea i
hee
e he ii
ead
a dee
i
ed e ha i
ee.
Thee 2. e a iig age A i a i e ea T . f he A ea de ig
f T i
ee ad A e a axia y ea i
hee hyhei f dee
i he he
eawk fai e dee
i e ae
ee.
f. Aa g ha f Thee 1 he f i vided i aedix A.
weve hee
ee ae (cid:27)e by fa e iive
e i
ae 1 ad 8 f Tab e 5.
hee
ae fai e have
ed b he iig ye fa e y eed dee
ed
fai e. a
i
e hi ay ead
y
eig f ay fa e a a.
dea y he dee
i
aabi iie h d be d ad
ee. Uf ae y we
a
hw ha
hee
e baed diabig ai
hee exi ha e i bh d ad
ee dee
i. We hw i Thee 3 ha vide d ad
ee dee
i a
diabig ai ehd wi have be i
ie: Give he ae e f ib e a
hig
hyhee i wi have eie ak e hyhei ad eie ahe.
Thee 3. e be a
ee ea de ig hyhee e de ig a i e ea.
Thee de exi a diabig ai
hee S ha 1 e
hee
e a e a he bai
123
aika ad Tabe
f diabig ai f ad 2 i deeiii
i i e e
i ad 3 e i d
ad
ee fai e dee
i.
f. e S be a diabig ai
hee ha ead
ee ad d dee
i ad
e y kw edge f he
hee
e f he hyhee i e e
ig a diabig aed hyh
ei. S e f
adi
i ha i i deeiii
ad h
ie i i e e
i
f a hyhei f i.e. give a e f
adidae hyhee i a ie e
deeiii
ed e
he e hyhei baed i
hee
e. Si
e i de
e ay he kw edge ide f he
hee
e f he
adidae hyhee give
he ae e f
adidae i wi a way
he he ae hyhei. e A
be he
iig age ig S . e B be a ied age whe a
i ae idei
a whe
exe
ig ea a
;
. Th A
a deeie whehe B i exe
ig
1
2
1
A
; B =
= A
; B =
= f
;
g. f A
ad B ae bh exe
ig
A
2
1
2
1
2
1
hyhee e i
= f
;
;
;
g
1
1
1
2
Si
e S ead
ee ad d dee
i i wi
he
;
. w whe A
ad
1
1
B ae exe
ig
ad
ee
ive y he a
hig hyhei e i agai a de(cid:28)ed
1
2
abve. B w S e e
;
. Si
e he ae e f
adidae hyhei wa
1
2
ed i ea
h
ae ad he ifai wa ied S be deeiii
i
i e e
i f a diabig aed hyhei
adi
ig he a i.
The eii
a ad aa yi
a e hw ha e f a ig e diabig aed hyhe
i ead ived b iefe
fai e dee
i e
aed he iig
dii ad
i
ai evi y ed. The eii
a e i Tab e 2 4 ad
5 eab ih he bee(cid:28) f he eawk iig e
hi e: hyi
a fai e wee
dee
ed. weve he aa yi
a e Thee 1 2 3 hw ha he e ae e
ha efe
. The a gih ae eihe d
ee b bh. F
ee
iig we w d e ie addiia
ed e ha
a di(cid:27)eeiae he e iive f
he fa e e e.g. by f
ed
i
ai. Thee
ed e ae fe vey exeive.
We
a ed
e he eed f
y vei(cid:28)
ai by eig g f iie
e a ig e
hyhei f
ig iead f aiaiig w hyhee: a axia y
hee hyh
ei ad a axia y i
hee hyhei. Tab e 6 hw a i f he f e f
ea hyhee avai ab e whe he aa
ke A1 i iig he ea. The a be
f hyhee eeed i he ab e i 24 wih a ay a 4
exiig i a ig e
ae
ad h aiaiig a f e f hyhee w d be exeive. weve he w ivee
he ii
(cid:22)ea
hee
e ad i
hee
e(cid:22)eee w exee f he a
e f hee
hyhee. f hey agee ha a fai e exi he a fai e a
a y
ed i
e he
ea
hee hyhei g aaee de Thee 1. f hey agee ha fai e
exi he fai e k a
e i
e he ea i
hee hyhei g aaee
eee Thee 2. f hey diagee i.e. he ea
hee hyhei de i y
a fai e b he ea i
hee hyhei de he iig ye
a be e
eihe way ad eve ba
k vei(cid:28)
ai.
Thi evied dee
i a gih (cid:27)e igi(cid:28)
a
aia avig
aed he
ig e ea i
hee hyhei aa
h. i
ee ad d b igi(cid:28)
a y
124
Rb Age Tea via S
ia y Aeive iig
Cae
A1 yheized Exe
ig a
Re aihi Fai e
Dee
i
#
Aa
ke A1 Aa
ke A2
S
A3
Dee
ed?
C a
1
W
F
Fa e iive
W
W
Fa e iive
W
W
F
Fa e iive
W
W
W
T e egaive
2
F
F
T e iive
F
W
T e iive
F
W
F
T e iive
F
W
W
T e iive
3
W
F
F
T e iive
W
F
W
T e iive
4
F
F
W
T e iive
F
F
F
Fa e egaive
5
W
F
T e iive
W
W
T e iive
W
W
F
T e iive
W
W
W
Fa e egaive
6
F
W
T e iive
F
F
T e iive
F
W
W
T e iive
F
W
F
T e iive
7
W
F
F
T e iive
W
F
W
T e iive
8
F
F
W
Fa e iive
F
F
F
T e egaive
Tab e 6: A i f he aa
ke A1 iig hyhee ad i ied e whe
akig i ed e e
a ig e hyhei f ea
h
ae.
ed
e he eed f vei(cid:28)
ai i
e a ea whe he ea
hee hyhei i ie
fai e vei(cid:28)
ai i e
eay. e ie eeeig y w hyhee ad i
h i
aia y
heae ha aiaiig a exeia be f hyhee.
F exa e ig a axia y ea i
hee hyhei e ai f Exa e
1 e i a eed veify i a eigh
ae a we have ee 5. weve whe we
bie
h a hyhei wih a axia y ea
hee hyhei e.g. a i Tab e 4 we y
eed veify f 50 f he
ae.
ae 2 3 6 7 hee i ageee bewee he
w hyhee ha a fai e ha
ed ad h vei(cid:28)
ai i e ied.
A iig age
a heefe adde he iig e e
iviy b e by ba a
ig
i e
e age agai he g aaeed efa
e f he iig a gih ed.
Eihe f he i e ig e hyhei a gih w d i ize y e hyhei i ea
h
ae wih dee
i
aabi iie ha ae g aaeed be d
ee b bh.
he e
ex a gih w hyhee w d be eaed ab i ea
h
ae ad
125
aika ad Tabe
he a gih w d be
ee ad e ie vei(cid:28)
ai i fewe
ae
aed he
i e hyhei
ee a gih.
5. iig Se e
iviy i Diib ed Teawk iig
Thi e
i f
e iig e e
iviy whe ex iig a key iy f exe
i iig i i age evie(cid:22)i i y he ied age ha
ae diib ed b he iig age
a be diib ed a we . We begi wih he
i e
hee f e e
ig a ig e axia y ea
hee hyhei. Si
e
ea ized
eawk iig wa
ef i addeig a e ai f Exa e 2 we f
hee he e ai f Exa e 1 Tab e 3 i whi
h
ea ized eawk iig
by he aa
ke e ed i fa e egaive dee
i
ae 4 5 i Tab e 4.
a diib ed eawk iig
hee y wi a ig e aa
ke i
i eaae b he
ad he he aa
ke wi a egage i iig. Tab e
7 ee he iig e f he ae fai e e ai wih he
a he
iig age. We (cid:28)d ha he
ef y dee
he w fai e
ae ha
he aa
ke fai ed dee
eaig f he aa
ke iig iake. F he
e i
e he
ed he he axia
hee
e he ii
dee
i i d ad
vei(cid:28)
ai i e ied. The ea f he
e i ha he aa
ke a
i i
hi
ae a h gh abig d ay hyhei ha
a be a
hed he
a. he wd egad e f wha a he aa
ke ae exe
ig i hee
w
ae i i di(cid:27)ee ha he a exe
ed by he
.
Cae
A3 yheized Exe
ig a
Re aihi Fai e
Dee
i
#
Aa
ke A1 Aa
ke A2
S
A3
Dee
ed?
C a
1
W
W
W
T e egaive
2
F
W
F
T e iive
3
W
F
W
T e iive
4
F
F
W
T e iive
5
F
T e iive
6
F
F
T e iive
7
F
F
T e iive
8
F
F
F
T e egaive
Tab e 7: S
A3 iig e i a e ai f Exa e 1 ig ea
hee
e.
Th if a age egaged i iig i e ai f Exa e 1 dee
i w d
be d ad
ee. a a
a fai e
ae ad y i he hee w d a ea e
ea ebe wh dee
he fai e. We ae fa y de(cid:28)e he geea
dii
de whi
h hi hee h d.
De(cid:28)ii 5. We ay ha w ea a
;
have bevab y di(cid:27)ee e R
; R
if
1
2
1
2
give a age B wh f (cid:28) he e R
; R
i he w a e. ay iig age
1
2
A di(cid:27)ee ha B wi have A; B =
\ A; B =
= ;. We he ay ha B ha
1
2
bevab y di(cid:27)ee e i
ad
ad
a B a key age.
1
2
126
Rb Age Tea via S
ia y Aeive iig
iive y B i a key age ha ha bevab y di(cid:27)ee e i he w a if a
iig age
a di(cid:27)eeiae bewee B behavi i exe
ig
ad i exe
ig
1
. F ia
e bh aa
ke have bevab y di(cid:27)ee e i F i whi
h hey (cid:29)y ad
2
W i whi
h hey ad. weve hey d have bevab y di(cid:27)ee e i W ad
bh f whi
h e ie he ad. The
ha bevab y di(cid:27)ee e i W (cid:29)yig
ad adig.
The key age i he bai f he
dii de whi
h a e f iig ea wi
dee
a fai e wih ea
h age ig y ea
hee
e. We (cid:28) ve a ea he
dii i whi
h a ig e give age wi dee
a fai e. We he e hi ea
ve he
dii de whi
h a ea e age i a give ea wi dee
a fai e.
ea 1. S e a i e ea T i e f iig a ebe f he ea i
ea
h he ig he axia y ea
hee he ii
ad de he a i ha f
ea
h age ea de ig i
ee. e A
; A
be iig age wh ae ebe f
1
2
T ad ae exe
ig
;
ee
ive y. A
w d dee
a fai e i aiaiig eawk
1
2
1
e aihi wih a age A
if A
i a key age i
;
.
2
2
1
2
f. See aedix A.
A
kw ha i i exe
ig
. f A
i exe
ig
ad i a key age i
ad
1
1
2
2
1
2
he A
i g aaeed i
e ha a di(cid:27)ee
e exi bewee ie f ad A
i
e A
i
1
2
2
a
ig bevab y di(cid:27)ee ha i w d if i had bee exe
ig
. e hweve ha
1
A
ay ay dee
hi di(cid:27)ee
e i
e f A
ee
ive A
behavi ay
2
2
1
ay be ex aied by
. A
wi dee
a di(cid:27)ee
e y if A
e i
ad
2
2
1
1
2
ae a bevab y di(cid:27)ee. weve i
e A
ha dee
ed he fai e i
a w a e
1
i eaae diage he fai e
he
e
ive a
i.
f we wa g aaee ha a eawk fai e wi a way be dee
ed by a ea
e age we ake e ha i ea
h ib e
biai f a hee ha be
a ea e key age whe e ae bevab y di(cid:27)ee. The ea hw ha he
age iig hi age wi i
e a fai e if e
. T hi ai we de(cid:28)e a
bevab y aiied e f a e yed by a ea.
De(cid:28)ii 6. A e f ea a i aid be bevab y aiied if f ay w a
;
2 hee exi a key age A
. The e f hee A
age i
a ed he key age
i
j
ij
ij
e f .
F ia
e he e f ea a he i
e i ea ha bee ig i he
exa e F y F igh a F Wai a i W deed a ad i S
i bevab y aiied. The aa
ke ad i W ad b (cid:29)y i F ad . The
ad i ad b (cid:29)ie i W ad F. Tab e 8 hw whi
h age have bevab y di(cid:27)ee
e i ay w a i he e. F ia
e by (cid:28)dig he
e a he iee
i f he
w ad he W
we (cid:28)d ha he
ha bevab y di(cid:27)ee e i hee w
a. deed he
ad whe a
ad i e
eived ha exe
i b (cid:29)ie
he eey ii whe exe
ig W. ee i
e a age have bevab y
di(cid:27)ee e i a ea w a he key age e f { W F } i
de a ebe
f he ea(cid:22)aa
ke ad
.
127
aika ad Tabe
F y F igh a F
Wai a i W
deed a
i S
F
Aa
ke
Aa
ke
S
W
Aa
ke
S
S
ad Aa
ke
Aa
ke
S
Aa
ke
S
S
ad Aa
ke
Aa
ke
Tab e 8: bevab e aiiig f he he i
e i ea i dSAF.
Thee 4. f a i e ea 1 e y a bevab y aiied e f ea a
ad a ea ebe i ebe f he key age e f 2 ig
ee ea
de ig ad 3 axia y ea
hee hyhee he he eawk fai e dee
i
e ae d ad
ee.
f. F hee 1 we kw ha dee
i w d be d. T hw ha i i
ee
we wi ve a ea e age wi dee
a di(cid:27)ee
e bewee ie f ad he wheeve
ea ebe ae a exe
ig he ae a i.e. a fai e i
ig. S e he
ea i
e y divided he ea a ha be exe
ed i.e. hee ae age
a
; a
i he ea ha ae exe
ig ea a
;
ee
ive y
h ha
6=
.
i
j
i
j
i
j
Th a fai e ha
ed. e be he key age e f . Si
e he ea i bevab y
aiied f
;
hee exi a ea e key age a
2 . Thee ae hee
ae:
i
j
1
ae i. a
i exe
ig
. hi
ae ay age exe
ig
w d dee
a di(cid:27)ee
e
1
i
j
wih a
ad w d heefe dee
he fai e ea 1.
1
ae ii. a
i exe
ig
. hi
ae ay age exe
ig
w d dee
a di(cid:27)ee
e
1
j
i
wih a
ad w d heefe dee
he fai e ea 1.
1
ae iii. a
i exe
ig e he a . e be bevab y di(cid:27)ee i
1
ad
i ad
i bh ad h ay age exe
ig
ad/
w d dee
i
j
j
i
he fai e. The
ae ha a
e ae bevab y di(cid:27)ee i ad
ad i ad
1
i
i iib e i
e he f a iig age A
j
A
; a
=
\ A
; a
= j (cid:19) f g 6= ;
1
i
1
Cadi
ig a
beig a key age f
;
.
1
i
j
Si
e i a hee
ae a ea e age w d dee
a fai e whee e
ed.
Theefe fai e dee
i i
ee. Si
e i i a d a we have ee dee
i i
d ad
ee.
The hee hw ha diib ed eawk iig
a e i d ad
ee
fai e dee
i whi e ig a i e a gih. Ea
h ea ebe i y he key
age
ig a axia y ea
hee hyhei. f i dee
a fai e he
eai y
2
e ha
ed. f age dee
a fai e he ideed fai e ha
ed.
Thi i e diib ed a gih wih i aei f
ig fea e ad g aaeed
de ad
eee
a wih he e
ex
ea ized a gih whi
h
we di
ed i he evi e
i Se
i 4. The a gih e(cid:27)e
ivee e ie he
2. f he iig ea ebe de kw wh he key age ae b kw hey exi i
a
i a he ea ebe. Thi i
eae iig b d ad
ee fai e dee
i
i i g aaeed.
128
Rb Age Tea via S
ia y Aeive iig
dii f a bevab y aiied e f a ad he diib i f he iig.
A
ay f Thee 3 ad 4 i ha if key age ae avai ab e i he diib ed
ae fai e dee
i i eihe d
ee b bh. Ad eve whe key age
ae avai ab e
ea ized eawk iig i i
ee ad d.
F ae y bevab e aiiig i a di(cid:30)
ey deig: Tea ae vey
fe
ed
h ha a age have he ae e i he ae a ad i geea
e d have bevab e di(cid:27)ee
e bewee he. F ia
e he i
e i ea
i he dSAF dai yi
a y exe
e a e f a wih hi ey a Tab e 8
deae.
f he ea hweve i bevab y aiied hee ay be a
ae whee w age
ae ea
h exe
ig a di(cid:27)ee a b age wi be ab e dee
i ig he ea
hee
e he ii
. The iia
ae whee hi
i whe w age A
ad A
ae
1
2
exe
ig a
ad
ee
ive y ad
ad
ae bevab y di(cid:27)ee
h
1
2
1
2
ha
A
; A
=
\ A
; A
= 2 = f
;
g
2
1
1
1
2
1
2
Thi wi e i A
ad A
ea
h be ievig ha he he i i ageee wih he. A
1
2
he
k f
h a i ai
a be ade a a f he a deig
e akig iky
i i he exe
i i whi
h dee
i i eihe d
ee Thee 3 ad
vei(cid:28)
ai e.g. by
i
ai
a be e
ibed a
ive y. he
he
k
d
be ieed i he
f ie aa yi(cid:22)he age w d i ae he he
hyhee a
hig hei w a
i ad dee
iky i dyai
a y.
6. Uig S
ia y Aeive iig i a (cid:27) ie C(cid:28)g ai
T f he deae he geea iy f
ia y aeive iig faewk hi
e
i exaie e e f eawk iig i dai i whi
h diagi ad e
vey
f evey fai e ae ifeaib e d ig exe
i. Exa e f
h dai i
de ea
i iay h a ea aiig V e Ca Bwe Sa a 1996 ad he
i age dai. The dyai
a e f he dai had ea ie dead ie ad
exiy f he age iv ved e.g. h a ea ebe ake diagi ad e
vey
di(cid:30)
. Eve if a fai e
a be diaged i i fe ae f e(cid:27)e
ive e
vey.
h evie he iig age i fe
eed wih ed f efa
e.
Thi ifai i ia f g e deig eva ai ad aa yi ad eed
e
eai y be
a
aed ie. The e f he aa yi ae ea a feedba
k he
age deige
a
h evi f h a.
T hi ed we ae deve ig a (cid:27) ie
ia y aeive iig ye
a ed
Teae TEAwk iig REview. Teae
e y e exe
i a
e f
he ied age ef he iig ahe ha ig a e
gii. Th i
de eed wy ab he
eaiy i a e
gii ab ea ie e
fa
e. ead i kw wih
eaiy ea
h age a d ig exe
i. Teae
a
ae evea aiaive ea e e aed eawk i
dig he Aveage Tie
Ageee ea e ATA f h ad a ea e f he eve f ageee i a ea.
Thee b i d he fai e dee
i a gih b aggegae fai e i aiaive y. We
f
hee he ATA ea e.
129
aika ad Tabe
Teae de(cid:28)e a wi
h a he ie ieva begiig a he i whee ay ea
ebe a ea e e e
a ew ea a f exe
i by he ea ad edig
a he i whee he ea i agai i ageee he ea a beig exe
ed.
efe
eawk a ea ebe e e
a ew ea a ji y ad a way eai
i ageee.
e ea ii
eai e age wi ake ge wi
h ad
iiia y a eawk fai e wi
. The (cid:28) ea ebe e e
a ew a wi be
i diageee wih e f i eaae i eihe i eji he i exe
ig he
igia a hey ji i i e e
ig he ew a. S
h a wi
h begi wih a dee
ed
fai e ad ed whe e fai e ae dee
ed.
Fig e 6 hw a i ai f a wi
h. The hee age begi i a iiia ae f
ageee ji exe
i f a 1 (cid:28) ed ie. Age 1 i he (cid:28) age wi
h
a 2 ded ie ad i f wed by Age 3 ad (cid:28)a y Age 2. The wi
h i he
ieva whi
h begi a he ia
e Age 1 e e
ed a 2 he ie a hee age
egaied hei ageee b hi ie a 2.
Agent 3
Agent 2
Agent 1
A Switch
Time
Legend:
Plan 1
Plan 2
Fig e 6: A i ai f a wi
h. The hee age wi
h f a 1 a 2.
Teae kee a
k f he egh f ie i whi
h fai e ae dee
ed i hey
ae e ved. The ATA ea e i he aveage wi
h egh i ie (cid:16)i
k(cid:17) e a
ee
ea e.g. a ii i dSAF a gae i RbC . A efe
ea w d have
a wi
he f egh ze ad heefe a ATA f 0. The w ea w d be e ha
f he vey begiig f hei ak exe
i he vey ed f i w d agee
he ea a beig exe
ed. F ia
e ea
h RbC gae a f 6000 (cid:16)i
k(cid:17). The
w ib e ea w d have y e wi
h d ig he gae f egh 6000. Th he
ATA
a e i RbC ge f 0 efe
6000 w.
We ed he ATA ea e aa yze a eie f gae f w RbC i ai
eag e ea SS97 ad SS98 ae a e a . 1999 agai a (cid:28)xed e Adhi 97
Ad 1998.
hee gae we vaied he e f
i
ai by ea
eva ae deig de
ii he e f
i
ai. axiae y ha f f he gae
aye wee a wed e
i
ai i evi
e f eawk. he he ha f a
i
ai bewee age wee diab ed. SS97 ayed axiae y 15 gae i
ea
h eig ad SS98 ayed 30 gae i ea
h
i
ai eig.
Tab e 9 hw he ea ATA va e ve hee gae f w b ea ea
h havig
hee ebe f SS97 ad SS98 ATA va e ae
a
aed eaae y f ea
h b
ea. The (cid:28)
hw whi
h b ea he e efe i ea
h w. The e
d
130
Rb Age Tea via S
ia y Aeive iig
hw he ea ATA f ea
h b ea whe
i
ai wee ed. The
hid
hw he ea ATA whe
i
ai wee ed. The ex
hw
he ize f he ATA ed
i(cid:22)he d i he ea ATA va e whe
i
ai ae
id
ed. The a
hw he babi iy f he hyhei i a w ai ed
e f he di(cid:27)ee
e i he ATA ea. Thi i he babi iy ha he di(cid:27)ee
e i d e
ha
e h a e be idi
ae geae igi(cid:28)
a
e.
SS
ea ATA ea ATA
ATA
e b.
b ea
.
C.
Red
i
hyhei
97 Ga ie
32.80
5.79
27.01
7.13e 13
97 Defede
57.5
6.81
50.69
.45e 10
98 Ga ie
13.28
3.65
9.63
9.26e 16
98 Defede
12.99
3.98
9.01
7.13e 5
Tab e 9: Aveage Tie Ageee ATA f gae agai Adhi 97.
C ea y a vey igi(cid:28)
a di(cid:27)ee
e eege bewee he
i
aig ad
i
aig vei f ea
h b ea. The ATA va e idi
ae ha haig if
ai by way f
i
ai igi(cid:28)
a y de
eae he ie i ake ea ebe
e ageee a e e
ed a. Thi e agee wih i ii ab he
e f
i
ai ad i ha ee ay be iig.
weve he ATA ed
i agi de idi
ae ha SS98 ay be
h e ei
ive f
i
ai ha SS97. The di(cid:27)ee
e i ATA va e f SS97 ae
axiae y i e ea y f ie a gea a f SS98. ex aai f hi
hee i ha SS98 i
ed f aye wih ived
aabi iie f i
ig he evie
h ha hey have bee kw edge f he evie. SS98
i heefe a deede
i
ai a ae ea
h a SS97
ed
f aye wih ee evie iig
aabi iie. SS98 aye ae bee ab e
e e
he
e
a wih e yig hei eaae. Th hey w d be ab e
aiai he ae eve f efa
e whe
i
ai ae ed.
a
SS97 aye e y aig ifai ad f ea
h he iig ea
h he
h gh
i
ai ad k
h ge eab ih ageee whe
i
ai wee avai ab e.
We
a va idae he hyhei ggeed by ATA ea ee by kig a he vea
ea efa
e i he gae ea ed by he
e di(cid:27)ee
e a he ed f he gae.
Tab e 10 hw he ea
e di(cid:27)ee
e f he ae eie f gae agai Adhi 97.
The (cid:28)
i he
i
ai eig wih wih . The e
d ad
hid
hw he ea
e di(cid:27)ee
e i he gae f SS97 ad SS98. The
b w aize he e f e ea
h e f gae deeie he
igi(cid:28)
a
e eve f he di(cid:27)ee
e bewee he ea
e di(cid:27)ee
e. The
e di(cid:27)ee
e
e
bae he ATA e . Whi e he di(cid:27)ee
e i ea
e di(cid:27)ee
e i ideed
aii
a y igi(cid:28)
a i SS97 gae i i igi(cid:28)
a i SS98 gae. Thi
ex aai ha he e i aia y awae SS98 i ideed bee ab e had e
f
i
ai ha SS97.
131
aika ad Tabe
SS97
SS98
C i
ai Ued
3.38
1.53
C i
ai Ued
4.36
2.13
e / hyhei
=0.032
=0.13
Tab e 10: SS97 ad SS98 ea
e di(cid:27)ee
e agai Adhi 97 wih
hagig
i
ai eig
The geea e eegig f hee exeie i ha a ade (cid:27) exi i adde
ig he iig e e
iviy b e. The kw edge ha i aiaied ab eaae
hee via
i
ai
a be aded a exe wih kw edge aiaied ab
he evie. A deige heefe ha a age f a eaive
aabi iie ha i
a
he f i age. Di(cid:27)ee dai ay bee fa
i iae i i
i
diai by
iig he evie whi e he e ie age e y
i
ai ex i
i
kw edge f ea ebe had e he
diai.
The ATA e addiia
i ee
ia y whe
bied wih a geea
efa
e ea e
h a he
e di(cid:27)ee
e. T i ae
ide he f he
a
a daa f hee gae. Fig e 7 a he ATA va e f a f vaia f
he Ga ie b ea. The gah axiae y 60 daa i. We ee i Fig e
7 ha whe
i
ai ae ed SS97 ATA va e ae i geea y bee ha
SS98 ATA wih
i
ai. Th deie i ia
e idivid a i aia
awaee i ab e f y
eae f a
k f
i
ai.
)
A
T
A
(
t
n
e
m
e
e
r
g
A
o
t
e
m
T
e
g
a
r
e
v
A
i
90
80
70
60
50
40
30
20
10
0
ISIS98/Comm.
ISIS97/Comm.
ISIS98/No-Comm.
ISIS97/No-Comm.
Goalies Sub-Team
a ATA Va e f Ga ie bea
Fig e 7: ATA va e f he Ga ie b ea i gae agai Adhi 97.
132
Rb Age Tea via S
ia y Aeive iig
Teae deae he e e f he eawk iig e
hi e deve ed i
ea ie e
i i a (cid:27) ie
(cid:28)g ai. The deige f SS97 h d e i age
e
i
ai i
e he wi have igi(cid:28)
a ivee he
e di(cid:27)ee
e.
a wih wih
i
ai SS98 aye ae ab e aiai hei
abai. Th if
i
ai ake e
i e
e i
a be e aive y afe y
e iiaed f he SS98 age deig ad he deve e e(cid:27)
a be die
ed a
e he
e f he age.
7. Beyd Teawk
We have eeed a geea
ia y aeive iig faewk dee
fai e i
aiaiig ageee ji ea a. weve e(cid:27)e
ive eai i ea fe
e ie addiia e aihi whi
h we bie(cid:29)y adde i hi e
i.
7.1 A Ri
he Ageee de : Ageeig Diagee
The eawk de e ie ji exe
i f ea a. evi
e f
h ageed
ji a age ay eie agee exe
e di(cid:27)ee b a idivid a y i
i b ea exe
e di(cid:27)ee b ea a. Tw exa e ay eve i ae.
Exa e 3. he dSAF dai he i
e egage he eey by eeaed y f wig
he f wig hee e: hidig behid a hi ee akig he ig ak
ig he hig ii e a he eey ad ba
k hidig. e vaiai f hi
a hey ae e ied ake e ha w he i
e ae hig a he ae ie.
f
e d e ii f
i
ai he i
e d fai ad ak a he ae ie.
Exa e 4. he RbC dai 11 aye i bh SS97 ad SS98 ae a
e a . 1999 ae divided i f b ea: id (cid:28)e de aa
ke defede ad ga ie
he ga ie ad w
e defede. Thi divii i b ea i de ed by he age
e e
ig e f f ea a i evi
e f he ay ea a ee Fig e 8. id (cid:28)e de
e e
he idfie d a ga ie e e
he defed ga a e
. Agai idea y
a aa
ke w d eve e e
ay he a b aa
k a defede w d e e
he
a b defed e
. weve d e
i
ai fai e aye ay eie
a
ide y abad hei ieded b ea ad exe
e a ea a f ahe b ea.
[Win−Game]
[Play]
[Interrupt]
...
[Attack]
[Defend]
......
[Midfield]
...........
[Defend−Goal]
[Simple
Advance]
.....
[Flank
Attack]
...
[Careful−
defense]
[Simple−goal
defense]
.....
Score−goal
Pass
Intercept
kick−out Reposition
Fig e 8: A i f he a hiea
hy ed by SS RbC age.
bh f hee exa e
eai di(cid:27)ee
e bewee age ae ageed ad ae a
ig f
e
exe
i f fai e. deed i i he a
k f di(cid:27)ee
e i e e
ed a
133
aika ad Tabe
ha w d idi
ae fai e i hee
ae. We e he e a ex
i
diai
efe hee e aihi. Exa e 3 idea y w i ae exe
ig he hig
a a he ae ie.
Exa e 4 w ebe f di(cid:27)ee b ea e.g. a
aa
ke ad a defede ae exe
ig he ae a i evi
e f ay e.g. defed. A he
exa e deae hee i a
ea eed f iig a ex
i
diai.
e f evi e
i ae e ed i evi
e f
ia y aeive iig f
a ex
i e aihi. They e ie a afai bh i i eeai ad
hey. The hiea
hie ae
aed i he a ae ex
e ha fai e ae igi(cid:28)ed
by e a iie ahe ha di(cid:27)ee
e. F ia
e if a aa
ke i ayig i he ea
w ha f f he (cid:28)e d i eaae ay
e e
ha i iake y (cid:16)defe
ed(cid:17) he
aa
ke b ea ad be ieve ie f be a defede.
The aa yi
a e ae iveed a we . The axia ea
hee
e he ii
wi
w ead
eee i
e i efe hyhee ha
ai e a iie ag age
whi
h ae fai e i a ex
i
diai. The axia ea i
hee
e he i
i
wi w ead d dee
i a i efe hyhee ha i y e a iie have
ed. Thee eie
a be ve fa y.
Thee 5. e a iig age A i a ex
i e aihi i a g
f age G. f A de ig f G i
ee ad A e a axia y ea i
hee
hyhei f dee
i he he fai e dee
i e ae d.
f. vided i aedix A.
Thee 6. e a iig age A i a ex
i e aihi i a g
f age G.
f A de ig f G i
ee ad A e a axia y ea
hee
hyhei f dee
i he he fai e dee
i e ae
ee.
f. vided i aedix A.
Th i a ex
i e aihi a i eawk e aihi g aaeed fai e
dee
i e ay i be vided deie he e f iied
eai kw edge
ab ied age. The
ea ized eawk iig a gih
a w be ea
i y afed f iig a ex
i e aihi. Uf ae y he e i
he diib ed
ae Thee 4
a be eai y afed i
e hey e y he
ey f bevab e aiiig whi
h i a
iaed wih di(cid:27)ee
e wih e a iie.
We eave hi i e f f e wk.
7.2 iig Uig R e Sii aiy Re aihi
Thi e
i a ie
ia y aeive iig e ii aiy e aihi f i
ig idivid a efa
e wihi ea. ai
a i evi
e f ea a age
ay e e
idivid a b a whi
h d e
eiae ageee by ea ebe b
ae
aied by he age e. F ia
e i evi
e f exe
ig he ea a
f y f igh a Fig e 3 i idivid a y e e
hei w idivid a a whi
h e
he ve
iy ad headig wihi he
ai f he fai ad (cid:29)igh ehd e
i(cid:28)ed
i he ii.
R e ii aiy e aihi e
ify he way i whi
h give idivid a a ae ii a
ad wha exe. Tw age f he ae e wh ae exe
ig diii a a
a
134
Rb Age Tea via S
ia y Aeive iig
be
ideed be i vi ai f he e ii aiy e aihi. Thi eab e a
ia y
aeive iig ye dee
fai e i e exe
i. T i idivid a a
he age i exe
ig i
ae i e e
i wih ha f he age f he ae e
ii a y he ehd we ed f eawk. f he a ae
ideed ii a by he
e ii aiy e aihi de fai e i dee
ed. hewie a fai e ay have
ed ad he diagi
e i
a ed veify i ad vide a ex aai.
e i ae wih a fai e f he dSAF dai whi
h ye wa ab e
dee
ig he e ii aiy e aihi:
Exa e 5. A ea f hee he i
e wa ake (cid:27) f he bae ad head a
ii. weve e f he i age fai ed
e
y
e he ii aee.
heefe ke i he i
e hveig abve he bae whi e i eaae ef exe
e
he ii by hee ve.
Thi fai e wa dee
ed ig e ii aiy e aihi iig. The ageed
ea a wa e e
ed by a he age ad b e wih eawk e aihi wa
dee
ed. Thi ea a iv ved ea
h age he e e
ig idivid a ehd f (cid:29)igh
whi
h deeie a i de ad ve
iy. ee he age di(cid:27)eed. The fai ig he i
e
eaied hveig whi e i eaae ved fwad. Uig a e ii aiy e aihi
he fai ig he i
e
aed i w e e
ed a ha f i eaae wh haed i
e f a bdiae i he fai ad ea ized ha hei a wee diii a e gh
a
e a ib e fai e.
Uf ae y he a
a ii aiy ei
ee be dai ad ak e
i(cid:28)
ad
h ae a eay e e a
dai. F hee dee
ed fai e ae e
eai y ea fai e d a dee
ed fai e have he ae weigh. We ae
e y
iveigaig way adde hee
ha egig i e.
8. Re aed Wk
iveigai f
ia y aeive iig ad he e aihi bewee kw edge
aiaied f age ae ad iig e(cid:27)e
ivee b i d eea
h i di(cid:27)ee b
(cid:28)e d f i age ye. We adde hee b (cid:28)e d i hi e
i ad ex ai hw
iveigai i e aed exiig iea e.
8.1 Re aed Wk Teawk
evi wk i eawk ha e
gized ha iig he age i
ii
a ea.
a iveigai have aied he iig e e
iviy b e b have addeed
i i deh. B i dig hee iveigai hi ae begi vide e i deh
awe hi b e.
The hey f Shaed a Gz a 1996 1999
he he eawk i
ig e e
iviy b e i evea way b vide y e iiia awe. Fi he
hey e ie age kw ha hei eaae ae
aab e f
ayig hei ak
i he ea. The a h e ha age
i
ae e gh ab hei a
vi
e hei eaae f hei abi iy
ay a
i Gz a 1996
. 314. Se
d he hey e ie age have a be ief i he haed e
ie a
135
aika ad Tabe
ae ha e ie age ea i(cid:28)ie e
i ab he age be ief. U
f ae y aaie f a be ief i de
idab e i hey a e e 1990
ad he
e be axiaed i a
i
e eig 1995; Ri
h Side 1997. S
h
axiai ay i ie g iig e iee. Thid hey id
e
he iei ha
i evi
e f
diai ad he f behavi i yig
iig f he ge ae he he eed f
h behavi Gz a 1996
Axi A5 A7. F h Shaed a e ie ha iei f a age
(cid:29)i
Gz a 1996 Axi A1 ad i
e e f hee iei i ai
a
iei ha ay iv ve he ai de f he age e iig f he
dee
ad avid
(cid:29)i
i i ied. The a h i ha whi e heei
a y a
h
(cid:29)i
a be dee
ed hi i ifeaib e i a
i
e Gz a 1996 . 307. They
gge ha
(cid:29)i
dee
i ad evei be iveigaed i a b e e
i(cid:28)
ae
wihi he iia
ai i.e. iig f
aabi iie a be ief ge a
k
f
(cid:29)i
vided by he Shaed a faewk . 308 ad 314.
i ei eve e e a . 1990; Che eve e 1991 e ie a age wh
ivae y
e be ieve ha a ji ga i eihe a
hieved a
hievab e ie eva
i havig he eie ea a y be ieve i be he
ae. A i he hey
f Shaed a i ei e f a be ief
a y be axiaed i a
i
e
ad ie g iig e iee. Th he iig e e
iviy b e i
aied f a
i
a i eeai f i ei.
eig ha hyheized ha w
ea
i
eaive i age
diai ae
ie ade by he age ad
vei e ed i hee
ie eig 1993. S
h
vei ae ed de
ide wha ifai eed
be ied ab age ad hw i i be ied. F ia
e a
vei ay
e ie a age e i eaae ay
hage i ivae y dee
wih ee
he aaiabi iy f he ea ga . eig aie he iig e e
iviy b e
ad vide a exa e f e
i(cid:28)
vei f high ad w badwidh i ai i
whi
h e kw edge i
i
aed a age if he badwidh i avai ab e.
weve eig de ex e i deh he ei f hw
h
vei ae e
e
ed ad wha ae he ade (cid:27) ad g aaee a
iaed wih he e e
i f ai
a
vei. F ia
e hee ae g aaee he e(cid:27)e
f ig he w badwidh
vei i he exa e.
The heei
a iveigai de
ibed abve a aie he iig e e
iviy b
e i i
i y ex i
i y. wk b i d hee adde hi b e i deh
i he
ex f
ia y aeive iig i ea. Thi ae e de
ad/
eee eie f eawk e aihi fai e dee
i ha
a be a
a yi
a y g aaeed deie
eaiy i kw edge a
ied ab ied age.
The aa yi
a g aaee ae a i
ab e a e
gii ad
i
ai ad ae
baed by eii
a e .
B i dig heei
a wk a
i
a eawk ye i
de eig 1995; Ri
h
Side 1997 ad Tabe 1997. eig iveigai f he i Reibi iy
eawk de i GRATE eig 1995 b i d i ei ad ii a y
w i eeai e ie age agee he ea a whi
h ae exe
e.
weve GRATE i ed i id ia eig i whi
h f f
i
ai
a
be a ed eig 1995 . 211 ad h y aive iig via
i
a
136
Rb Age Tea via S
ia y Aeive iig
i i ed. A h gh eig vide a eva ai f GRATE efa
e wih
ee
i
ai de ay g aaee ae vided wih ee
fai e de
e
i. GRATE aiai kw edge ab he age h gh a
aia
e de
whi
h ae ed kee a
k f wha ea ebe
aabi iie ae i evi
e f fig
ea. weve he ei f hw
h kw edge h d be ed i hee de i
ef addeed.
Ri
h ad Side iveigae CAGE i a
abaive e iefa
e ye i
whi
h
i
ai ae e iab e Ri
h Side 1997. weve f a h a abi iy
ee
ive iiig he a f
i
ai i i deiab e. T adde hi i
e e
e eii
a wk by eh Side ad Ri
h 1999 i ize a e
gii i
CAGE; he f
f ha wk i ig he
abaive eig ake he
a e
gii a
ab e. F ia
e abig iie i a e
gii ay be e ved
by akig he e f
ai(cid:28)
ai. Wk CAGE de iveigae hw
h
kw edge i be aiaied f e(cid:27)e
ive
abaive dia g e wih he e.
a
we ae ab e vide g aaee he fai e dee
i e f a gih. A
aa yizig he dia g e a f iky i ay a w ye
h a CAGE
de
ide whehe e
i
ai f
ai(cid:28)
ai egad e f a e
gii abi
g iy.
STEA Tabe 1997 aiai iied ifai ab he abi iy f ea ebe
ay hei e. STEA a a w ea ebe ea ex i
i y ab he
f
i
ai i de
idig whehe
i
ae . wk igi(cid:28)
a y
exed hee
aabi iie via a e
gii ad vide aa yi
a y g aaeed fa
dee
i e . F hee eawk fai e dee
i
aabi iie
a be ef
igge STEA e aig
aabi iie.
8.2 Re aed Wk Cdiai
be 1995 iveigaed he e f babi ii
a e
gii i evi
e f a
ive ea
wk iig ivaed by he e iabi iy ad
f aive
i
ai baed
iig i i iay a i
ai. Wahig ex e bevai baed
diai
ig akv de Wahig 1998 f
ig akig he
ai a
ab e.
a be ad Wahig wk f
e he iig e e
iviy b e.
We hwed egh ad iiai f
ea ized ad diib ed aa
he ha g a
aeed fai e dee
i e ig
hee
e baed diabig ai f a e
gii
hyhee.
D fee 1995 di
e vai ehd f ed
ig he a f kw edge ha age
eed
ide i
diaig wih he. The ehd di
ed iv ve ig a
f he eed de ig
i
ai ig hiea
hie ad aba
i e
. Whi e
he f
f hi wk i ehd by whi
h de ig
a be iied he f
f
wk i he ei f hw
h de ig i e ied f g aaeed efa
e(cid:21)he
iig e e
iviy b e. We vide aa yi
a g aaee ade (cid:27) iv ved i
ig iied kw edge f age f fai e dee
i e.
S gawaa ad ee 1998 e he e f
aaive eaig/aa yi e
h
i e i evi
e f eaig ad e
ia izig
diai e f a ye i whi
h di
ib ed age
diae i diagig a fa y ewk. The iveigai i f
ed
137
aika ad Tabe
iizig
diai e iiize ie(cid:30)
ie
y ad ed da
y i he age
diaig eage. U dee
ig b ia
diai via a fa de he age
ex
hage ifai hei
a view f he ye ad he b e vig a
iviy
ad
a g ba view. They he
ae he
a view he g ba view (cid:28)d
ii
a va e/aib e whi
h wee iig f he
a view ad heefe gave ie
he b ia efa
e b e. Thee va e ad aib e ae ed i
ig
i ai e
i(cid:28)
e ha iize
diai i ai
a i ai. F exa e
ewk diagi age ay ea a e ha g ide he
he a
diai a
egy i whi
h y e age ef he diagi ad hae i e wih he e f
he diagi age. wk
ia y aeive iig ii a y e
ai
bewee age view dive he iig
e. weve e f
ai i
a d
f he e aihi we ae iig. Whi e S gawaa ad ee wk
a
be viewed a eig he age i
eea y iize hei iig e iee
e aa yi
a y ex e he eve f iig e ied f e(cid:27)e
ive fai e dee
i i
di(cid:27)ee
(cid:28)g ai. eawk iig e
hi e addee
eaiy i he
a
ied ifai ad de
a g ba view f a aib e he ye(cid:22)a
ha w d be exee y exeive. ead e
hi e f
e iggeig fai e de
e
i via
aig f a he i
eea y exadig he ea
h f di(cid:27)ee
e i
he diagi
e.
Rbi
iea e ha a aied he iig e e
iviy b e. ake 1993
iveigaed he iig e e
iviy b e f a di(cid:27)ee ee
ive f a fai
aiea
e ak. She eii
a y exaied he e(cid:27)e
f
biig
ia y aeive i
fai whi
h he efeed a
a ad kw edge f he ea ga ad
de
ha he fa ea aegy i e whee he age i ea
h he a we a
ge wad he ga . iyhi e a . iyhi R gea x hii ia Sakae
akik a 1994 ee a faewk f
eai by bevai i whi
h b
vi a y aed he a a ee iie
diai. The faewk ee evea
adad aeia e ae i.e. wh i wh. They de(cid:28)e a ea aeia
e a e i whi
h a age i ea
h he. wk f
e he
iig e e
iviy b e wihi
ia y aeive iig f eawk e aihi
ad vide aa yi
a a we a eii
a e . We ea he aeia e ae a
a d
f he e aihi ha h d i he ye. e hw ha iig i
ea ay e
eai y e ie iig a ea ebe.
8.3 he Re aed Wk
ig e a .
ig ee Vi
e Bazza X a 1999 ee a diib ed
diagi ye f a i age ie ige he evie.The ye e fa
de ideify fai e ad ie(cid:30)
ie
ie i
e ad g ide e
vey. S
hede
ad Wage 1997 ed a diib ed diagi e
hi e i whi
h
eaig age
e
eive e e f e ad diage ad ed ee he age. They ea
h
a g ba diagi baed he
a e hey d
e ad e
eive(cid:22)wih he
a i ha
(cid:29)i
wi
. Fh i
h ad ejd 1996 iveigae a
hee
i whi
h i e diagi age
eae via a b a
kbad a
hie
e i diagig a
hyi
a ye. The age ay e di(cid:27)ee diagi de ye b a
ea ized
138
Rb Age Tea via S
ia y Aeive iig
(cid:29)i
e i age i e yed had e ay
(cid:29)i
i diage f d. A hee
aa
he d adde he iig e e
iviy b e.
Thee ae a few
ia ea e e aed he ATA. G dbeg ad aai
1997 i
veigae a i b fagig ak ad ea e iefee
e(cid:22)he a f ie b
ed avidig ea
h he. Ba
h 1998 e
ia ey Bai ey 1990 ea e be
havia diveiy i i age ak f
e fagig ad fai aiea
e. Bh
iveigai f
haa
eizig heegeeiy i i age ye ad i e ai
efa
e.
a he f
f wk i vidig ef feedba
k he
deige. ib e
e ai bewee ak efa
e ad ATA va e eai be
iveigaed.
9. C
i ad F e Wk
The wk eeed i hi ae i ivaed by a
i
a
e. We have beg
iveigai f he iig e e
iviy b e a a e f bevai ha fai e
i e
deie age e f iig
dii ad
i
ai.
Aa yi f he fai e evea ed ha age wee (cid:30)
ie y ifed ab ea
h he
ae. Whi e he eed i e eaae ha bee e
gized eeaed y i he a
eig 1993; Gz a 1996; Tabe 1997 he iig e e
iviy b e(cid:22)
he ei f hw
h iig i e ied(cid:22)eaied age y addeed eig
1993; Gz a 1996.
We vide key awe he iig e e
iviy b e. Wihi he
ex f
ia y aeive iig i ea we deae ha eawk e aihi fai e
a be dee
ed e(cid:27)e
ive y eve wih
eai iied kw edge f ea ebe ae.
We hw aa yi
a y ha
ea ized a
ive eawk iig vide fai e dee
i
ha i eihe
ee ad d d ad i
ee. weve
ea ized ea
wk iig e ie i e hyhee ad iig f a ea ebe.
a diib ed a
ive eawk iig e i
ee ad d fai e
dee
i deie ig a i e a gih ad iig y key age i a ea.
Uig a i eeed geea faewk f
ia y aeive iig we eii
a y va idae hee e i he dSAF dai. We a vide iiia e i
iig a ex
i ad e ii aiy e aihi ad iiia diagi
ed e.
We f he deae he geea iy f he faewk by a yig i i he RbC
dai i whi
h we hw hw ef aiaive aa yi
a be geeaed (cid:27) ie. Bh
dSAF ad RbC ae dyai
ex i age dai ha iv ve ay
eaiie i e
ei ad a
i.
We aeed deae hw he e ad e
hi e
a be a ied i he
dai. We have ex i
i y ied e
eay
dii f he hee h d
h a bevab e aiiig ad ea de ig
eee. The eeed diagi
a gih i eiive he a
a
y f he kw edge ed ad ay e ie a ig ha
a
a be e
gized a a hey ae e e
ed. Thee
dii h d be vei(cid:28)ed by
he deige i he age a i
ai dai. Rea
ive a
he eeeai ae
y ed i ay dyai
i age dai. f
iig ageee
ji a e f he
ea iy f ii a i f ageee i age ad h a
eawk iea e eig 1995; Gz a 1996; V e e a . 1996; Tabe 1997.
139
aika ad Tabe
We ade evea efee
e addiia aea i whi
h we w d ike
d
f he
iveigai. e ia i
whi
h we a iveigae i deh i he g
e iee f he diib ed eawk iig a gih i e f bevabi iy.
de vide i de ad
eee g aaee he diib ed a gih e ie
he abi iy f a ea ebe i he key age. We ae iveigaig way
e ax hi e iee whi e i vidig g aaeed e . addii he diagi
ed e h d be exeded ad fa ized ad we w d ike iveigae way
a eviae he eiiviy f hee
ed e he
hi
e f ea de ig hyhei.
A
kw edge
Thi ai
e i aia y baed a AAA 98 ae aika Tabe 1998 ad a
Age 99 ae aika Tabe 1999 by he ae a h. Thi eea
h wa
ed i a by SF Ga S 9711665 ad i a by AFSR
a
#F49620 97 1
0501. We hak e(cid:27) Ri
ke Gege Bekey Vi
ee Da eay ad David yadah
f ay ef
e. The ay eviewe have hak f he ig
y
a ize idea ad
ib i i evii f hi ae.
Aedix A. f
Thee. # 2 age 123. e a iig age A i a i e ea T . f A
ea de ig f T i
ee ad A e a axia y ea
hee hyhei f de
e
i he he eawk fai e dee
i e ae d.
f. We wi hw ha ay fai e ha
i dee
ed ad h ha a fai e wi
be dee
ed. e a
; : : : ; a
be he age ebe f T . Ea
h age a
i exe
ig e
1
i
a
1 (cid:20) i (cid:20) . Th
e
ive y he g i exe
ig
; : : : ;
. f a fai e
i
1
ha
ed he hee ae w age a
; a
; 1 (cid:20) j; k (cid:20)
h ha a
i exe
ig a
k
j
j
ad a
i exe
ig a
ad
6=
. Si
e A ea de ig i
ee he
j
k
k
j
k
e
hyhei
; : : : ;
; : : : ;
; : : :
wi i he e f ea de ig hyhee.
1
j
k
Si
e A wi
he a axia y ea i
hee hyhei eihe i wi
he he
e
hyhei whi
h i e i
hee ha a hyhei i yig fai e ha
ed
ha i wi e e
a hyhei wih geae i
hee
e hyhei e iva e eve .
ay
ae a fai e w d be dee
ed ad he dee
i
ed e i
ee.
ea. # 1 age 127. S e a i e ea T i e f iig a ebe f
he ea i ea
h he ig he axia y ea
hee he ii
ad de he
a i ha f ea
h age ea de ig i
ee. A iig age A
wh i a
1
ebe f T ad i exe
ig
w d dee
a fai e i aiaiig eawk e aihi
1
wih a age A
a a ebe f T exe
ig a di(cid:27)ee a
if A
ha a bevab y
2
2
2
di(cid:27)ee e i
ad
.
1
2
f. A
kw ha i i exe
ig
. Si
e a ebe f T i ea
h he ad
1
1
hee ve A
i iig A
wh ha a bevab y di(cid:27)ee e i
ad
. Si
e A
1
2
1
2
2
i exe
ig
ad f wig he bevab y di(cid:27)ee e
=2 A
; A
=
. Theefe
2
1
1
2
2
f he ee
ive f A
i
a be he
ae ha i aig
i ay age de ig
1
1
hyhei ad heefe ay ea de ig hyhei ha A
ha wi have A
exe
ig
1
1
140
Rb Age Tea via S
ia y Aeive iig
ad A
exe
ig e a he ha
. he wd f A
ee
ive hee i
1
2
1
1
ea
hee hyhei ad a di(cid:27)ee
e w d be dee
ed bewee A
ad A
.
1
2
Thee. # 5 age 134. e a iig age A i a ex
i e ai
hi i a g f age G. f A de ig f G i
ee ad A e a axia y
ea i
hee hyhei f dee
i he he fai e dee
i e ae d.
f. We wi hw ha if fai e ha
ed e wi be dee
ed ad h ha
ay fai e ha i dee
ed i i fa
a fai e. e a
; : : : ; a
be he age ebe f
1
G. Ea
h age a
i exe
ig e a
1 (cid:20) i (cid:20) . Th
e
ive y he g i
i
i
exe
ig
; : : : ;
. f fai e ha
ed he ea
h age i exe
ig a di(cid:27)ee
1
a i 6= j
6=
. Si
e A g de ig i
ee he
e
hyhei i
i
j
gig be i he e f g de ig hyhee . Si
e i i a axia y i
hee
hyhei eihe i wi be e e
ed ha a di(cid:27)ee hyhei f he ae
hee
e
eve wi be e e
ed. Ay hyhei wih he ae
hee
e eve a he
e
e
i ie fai e i dee
ed. Th he dee
i
ed e i d.
Thee. # 6 age 134. e a iig age A i a ex
i e ai
hi i a g f age G. f A de ig f G i
ee ad A e a axia y
ea
hee hyhei f dee
i he he fai e dee
i e ae
ee.
f. We wi hw ha ay fai e ha
i dee
ed ad h ha he
ed e
i
ee. e a
; : : : ; a
be he age ebe f G. Ea
h age a
i exe
ig e
1
i
a
1 (cid:20) i (cid:20) . Th
e
ive y he g i exe
ig
; : : : ;
. f a fai e
i
1
ha
ed he hee ae w age a
; a
; 1 (cid:20) j; k (cid:20)
h ha a
i exe
ig a
k
j
j
ad a
i exe
ig a
ad
=
. Si
e A g de ig i
ee he
j
k
k
j
k
e
hyhei
; : : : ;
; : : : ;
; : : :
wi i he e f g de ig hyhee.
1
j
k
Si
e A wi
he a axia y ea
hee hyhei eihe i wi
he he
e
hyhei whi
h i e
hee ha a hyhei i yig fai e ha
ed
ha i wi e e
a hyhei wih geae
hee
e hyhei e iva e eve .
ay
ae a fai e w d be dee
ed. Theefe he dee
i
ed e i
ee.
Aedix B. S
ia y Aeive iig A gih
We big hee he a gih i e d
de f he RES a e
gii a gih
he
ai e ig dee
i i bh i e ad i e ea ad he
iig a gih f he
ea ized ad diib ed
ae.
B.1 RES
RES wk by (cid:28) exadig he
ee ea hiea
hy f he age beig d
e ed aggig a a a a
hig. A a e
dii ad eiai
dii
ae (cid:29)agged a a
hig a we . A a a
i ae e be ed a exe
ai
behavi. Afe iiia izig he a e
gii hiea
hy f ea
h ied age b
evai f a age ae
i y a
hed agai he a
i exe
ed by he a.
a whe exe
ai a
h bevai ae agged a a
hig ad hee (cid:29)ag ae
agaed a g he hiea
hy ad dw ha
ee ah h gh he hiea
hy
141
aika ad Tabe
ae (cid:29)agged a a
hig . Thee ah e
ify he ib e a
hig ieeai f
he bevai. addii e
dii ad eiai
dii ae (cid:29)agged a e
igifyig he ifeed aiae be ief by he de ed age. Thi
e i
de
ibed i A gih 1.
A gih 1 RES ai a
hig bevai ad akig ifee
e f a give
a e
gii hiea
hy a ig e age.
1. Ge bevai ab age
2. F ea
h a ha ha a e f exe
ed bevai:
a Cae bevai exe
ai
b f
eed (cid:29)ag a a a
hig
ef y hewie (cid:29)ag a a fai ig a
h
3. F ea
h a ha i (cid:29)agged a a
hig
ef y
a F ag i ae a a
hig
ef y // agae a
hig
4. F ea
h a whe
hi de a f he ae (cid:29)agged a fai ig a
h
a F ag i a fai ig a
h // agae a
hig
B.2 Dee
i f Fai e Cea ized ad Diib ed Teawk iig
A gih 2 hw hw
ai f hiea
hi
a a i
aied . We ii e ve
hee i e ea. The a gih a
e a i w e f hiea
hi
a a hyhe
e ad hei w a
iaed age f
aiy he a gih a e y w age.
The geea izai age i aighfwad. The a gih a a
e a i
y (cid:29)ag
i
y. A TSTC i
y
a e he a gih e axia ea
hee
e
vide d b i
ee dee
i. A ESSSTC i
y
a e he a gih e
axia ea i
hee
e vide
ee b d dee
i.
The e f hiea
hi
a a ae aked hiea
hy_1 ad hiea
hy_2. The w age
ae aked age_1 ad age_2. The a gih ake e f he edi
ae S b ea
whi
h i e if he w age Age1 Age2 be g di(cid:27)ee b ea a he give
eve f he hiea
hy Deh.
Wih he aid f A gih 2 we
a w de(cid:28)e he
ea ized ad diib ed fai
e dee
i a gih. The
ea ized eawk iig a gih A gih 3
i ize A gih 2 wi
e
he
kig f fai e wih bh ESSSTC ad TSTC
i
ie. f he e f bh i
ie agee hey ae
eai. f he e d agee
i.e. he ESSSTC i
y
a e a fai e be dee
ed whi e he TSTC i
y
a e fai e be dee
ed he he iig age
a be
eai ha a fai e
ha ake a
e ad heefe eed veify he fai e. A gih 3 heefe e
FAURE _FAURE SSBE_FAURE.
The diib ed iig a gih i give i e d
de f be
a e i i
hig e ha a
a A gih 2 wih a TSTC i
y aaee. we
142
Rb Age Tea via S
ia y Aeive iig
A gih 2 iea
hi
a
ai f w age a wig f b ea.
1. Se Deh 0 // k f di(cid:27)ee
e (cid:28)
2. Whi e bh a a deh Deh ae ea a D
a if i
y == TSTC
i. he e a_1 a_2 be axia y ea
hee a a eve Deh
f hiea
hy_1 ad hiea
hy_2 ee
ive y.
ii. e e e a_1 a_2 be axia y ea i
hee a a eve Deh
f hiea
hy_1 ad hiea
hy_2 ee
ive y.
b f a_1 i e a a_2
i. he e FAURE
ii. e e if b f hiea
hie ea
hed
e _FAURE hewie i
eae Deh ad g 2.
3. f y e a i a ea a e FAURE e e e _FAURE.
A gih 3 Cea ized Teawk iig a yig bh iii
ad eiii
view.
1. e iii
_Re = Dee
age_1 age_2 hiea
hie_1
hiea
hie_2 TSTC
/ a gih 2 /
2. e eiii
_Re = Dee
age_1 age_2 hiea
hie_1
hiea
hie_2 ESSSTC
/ a gih 2 /
3. if iii
_Re == eiii
_Re
4. he e iii
_Re / eihe FAURE _FAURE /
5. e e e SSBE_FAURE
143
aika ad Tabe
i deived f he fa
ha a ebe f he ea ae ig i i he key age
f he ea.
Refee
e
Ab ge . A. See S. 1988. egaig aig exe
i ad iig.
eedig f he Seveh aia Cfee
e Ai(cid:28)
ia e ige
e AAA 88
iea i/S. a . AAA e.
Ad T. 1998. Re(cid:28)ee f
e age ii ig eif
ee eaig.
ia . Ed. RbC 97: Rb
e w d
1 V . A 1395 . 373(cid:21)388.
Sige ve ag.
Aki E. . D fee E. . Shi . G. 1997. Dee
ig ad ea
ig aed
f w d ae.
eedig f he F eeh aia Cfee
e Ai(cid:28)
ia
e ige
e AAA 97 . 571(cid:21)576 vide
e R. AAA e.
Bai ey . D. 1990. S
ia Ey They. Sae Uiveiy f ew Yk e.
Ba
h T. 1998. Behavia Diveiy i eaig Rb Tea. h.D. hei Gegia
i e f Te
h gy.
Ca de R. B. Sih . E. C ea
he A. . a . . F. Ceawi
z A. Z. 1993.
daf behavi i ai ad
.
eedig f he Thid Cfee
e
C e Geeaed F
e ad Behavia Reeeai ad F ida. i e
f Si ai ad Taiig Uiveiy f Cea F ida.
Che . R. Aa R. S. a D. . 1992. Ea y waig f a fai e fa e
iive ad eve e: Exeie ad a de . Te
h. e. CSC Te
hi
a
Re 92 20 Uiveiy f aa
h e.
Che . R. eve e . . 1991. Teawk. 35.
Dy e R. . Aki D. . Dhi R. S. 1986. Geeaig e
ei e e ad
exe
ai veify he exe
i f a.
eedig f he Fifh aia
Cfee
e Ai(cid:28)
ia e ige
e AAA 86.
D fee E. . 1995. B if iga
e: wig j e gh
diae we .
eedig f he Fi eaia Cfee
e iage Sye CAS 95
. 406(cid:21)413.
Fee . a S. Re
hei . S. 1995. Cdiai wih
i
a
i: Exeiea va idai f f
a i e
hi e.
eedig f he Fi
eaia Cfee
e iage Sye CAS 95 . 102(cid:21)108 Ca ifia
USA.
Fiby R. . 1987. A iveigai i ea
ive aig i
ex dai.
eedig f he Sixh aia Cfee
e Ai(cid:28)
ia e ige
e AAA 87.
144
Rb Age Tea via S
ia y Aeive iig
Fh i
h . ejd W. 1996. Re vig
(cid:29)i
i diib ed diagi. Wah e
W. Ed. he 12h E ea
h Cfee
e Ai(cid:28)
ia e ige
e ECA 96. h
Wi ey S
.
G dbeg D. aai
. . 1997. efee
e a a f deigig ad eva a
ig i b
e.
eedig f he F eeh aia Cfee
e
Ai(cid:28)
ia e ige
e AAA 97 . 637(cid:21)642 vide
e R. AAA e.
Gz B. . a S. 1999. The ev i f haed a. W didge . Ra
A. Ed. F dai ad Theie f Raia Age
y . 227(cid:21)262.
Gz B. . a S. 1996. C abaive a f
ex g a
i. Ai(cid:28)
ia
e ige
e 86 269(cid:21)358.
Gz B. . Side C. . 1990. a f di
e. Che . R. ga .
a
k . Ed. ei i C i
ai . 417(cid:21)445. T e Cabidge
A.
a e . Y. e Y. 1990. w edge ad
kw edge i a diib ed
evie. diib ed
ig 37 3 549(cid:21)587.
a
he W. C e . de ee . Ed.. 1992. Readig i de Baed Diag
i. ga a fa b ihe Sa ae CA.
ig B. ee V. R. Vi
e R. Bazza A. X a . 1999. Diagi a a
iega a f i age adaabi iy. Te
h. e. CSC Te
hi
a Re 1999
03 Uiveiy f aa
h e/Ahe.
be . . D fee E. . 1995. a
ig gehe: Wih
i
ai.
Wkig e f he AAA Sig Syi Reeeig ea Sae ad
e
hai . 60(cid:21)71 Safd CA.
eig . R. 1993. Cie ad
vei: he f dai f
diai i
i age ye. w edge Egieeig Review 8 3 223(cid:21)250.
eig . R. 1995. C ig
eaive b e vig i id ia i age
ye ig ji iei. Ai(cid:28)
ia e ige
e 75 2 195(cid:21)240.
h W. . Ri
ke . 1997. STEVE: A aiaed edaggi
a age f
ed a
aiig i vi a evie. SGART B ei 8 1 4 16(cid:21)21.
aika G. A. Tabe . 1998. Wha wg wih ? vig b e h gh
ia diagi.
eedig f he Fifeeh aia Cfee
e Ai(cid:28)
ia
e ige
e AAA 98 . 97(cid:21)104 adi W. AAA e.
aika G. A. Tabe . 1999. Y e Wee : Exeie i
diib ed ad
ea ized
ia iig ad diagi.
eedig f he
Thid eaia Cfee
e A Age Age 99 Sea e WA. AC
e.
145
aika ad Tabe
iy D. j gbeg . Ra A. Sebeg E. Tidha G. Wee E. 1992. aed
ea a
iviy.
Cae fa
hi C. Wee E. Ed. Ai(cid:28)
ia S
ia Sye
e
e e i A 830 . 227(cid:21)256. Sige Ve ag ew Yk.
ia . Tabe . Se . Ve . Cade
hi S. awa E. a baa .
da . Aada . 1997. The RbC yhei
age
ha ege 97.
eedig f he eaia i Cfee
e Ai(cid:28)
ia e ige
e CA 97
agya aa.
a S. Sy
aa . Eve
hik A. 1998. Rea
ig ageee h gh egiai: a
gi
a de ad i eeai. ai(cid:28)
ia ie ige
e 104 1 2 1(cid:21)69.
iyhi Y. R gea x S. hii . ia . Sakae S. akik a . 1994. C
eai by bevai (cid:21) he faewk ad he bai
ak ae. he EEE
eaia Cfee
e Rbi
ad A ai . 767(cid:21)773 Sa Dieg CA.
EEE C e S
iey e.
eh . Ri
h C. Side C. . 1999. Uig a e
gii i h a
e
abai.
eedig f he Seveh eaia Cfee
e Ue de ig
U 99 Ba(cid:27) Caada.
eve e . . Che . R. e . . T. 1990. a
ig gehe.
eedig
f he Eigh aia Cfee
e Ai(cid:28)
ia e ige
e AAA 90 e ak
CA. AAA e.
a e T. W. Cw . 1991. Twad a iedi
i iay hey f
dia
i. Te
h. e. CCS TR#120 SS W# 3294 91 SA aa
h e i e f
Te
h gy.
ae a S. C. Adibi . A aiza Y. aika G. A. ea . Ta i . Tabe
. 1999. beig a eaae: Exeie
e a
ied i he deig f b
ea..
eedig f he Thid eaia Cfee
e A Age
Age 99 Sea e WA. AC e.
ewe A. 1990. Ui(cid:28)ed Theie f Cgii. avad Uiveiy e Cabidge
aa
h e.
ake . E. 1993. Deigig
aw f
eaive age ea. he
eedig
f he EEE Rbi
ad A ai Cfee
e . 582(cid:21)587 A aa GA.
Ra A. S. 1994. ea ed a e
gii (cid:21) wad a hey f ea
ive e
gii.
eedig f he eaia Cfee
e w edge Reeeai ad Rea
ig R 94 . 497(cid:21)508.
Ree
e G. A. Tae A. 1994. Syheizig e
i i f
a a
e.
eedig f Ai(cid:28)
ia e ige
e aig Sye AS 94 Chi
ag .
Ri
h C. Side C. . 1997. CAGE: Whe age
abae wih e e.
h W. . Ed.
eedig f he Fi eaia Cfee
e A
Age Age 97 . 284(cid:21)291 aia de Rey CA. AC e.
146
Rb Age Tea via S
ia y Aeive iig
S
hede . Wage G. 1997. Diib ed diagi by vivid age.
eedig
f he Fi eaia Cfee
e A Age Age 97 . 268(cid:21)275
aia de Rey CA. AC e.
S gawaa T. ee V. R. 1998. eaig ive
diaed a
i i
eaive
diib ed b e vig evie. a
hie eaig 33 2/3 129(cid:21)153.
Tabe . 1996. Ta
kig dyai
ea a
iviy.
eedig f he aia Cfe
e
e Ai(cid:28)
ia e ige
e AAA.
Tabe . 1997. Twad (cid:29)exib e eawk. a f Ai(cid:28)
ia e ige
e Reea
h
7 83(cid:21)124.
Tabe . h W. . e R. F. aid . E. Reb . S. S
hwab
. 1995. e ige age f iea
ive i ai evie. A agazie
16 1.
Tyaa . age G. D. 1997. f a (cid:28) y d
eed....
eedig f he
F eeh aia Cfee
e Ai(cid:28)
ia e ige
e AAA 97 . 3(cid:21)9 vi
de
e R.
Ve . a
k . E. Cx . T. 1998. Raia e baed iig f aig
i dyai
evie.
eedig f Ai(cid:28)
ia e ige
e aig Sye
AS 98 ib gh A.
V e C. E. Ca Bwe . A. Sa a E. 1996. The ia
f
aiig
ea f
iig: A eii
a iveigai. h a fa
38 1 87(cid:21)100.
Wahig R. 1998. akv a
kig f age
diai.
eedig f he
Se
d eaia Cfee
e A Age Age 98 . 70(cid:21)77 i
ea i/S. a . AC e.
147
|
1809.06049 | 3 | 1809 | 2019-06-05T11:56:28 | Erratic Extremism causes Dynamic Consensus (a new model for one-dimensional opinion dynamics) | [
"cs.MA"
] | A society of agents, with ideological positions, or "opinions" measured by real values ranging from $-\infty$ (the "far left") to $+\infty$ (the "far right"), is considered. At fixed (unit) time intervals agents repeatedly reconsider and change their opinions if and only if they find themselves at the extremes of the range of ideological positions held by members of the society. Extremist agents are erratic: they become either more radical, and move away from the positions of other agents, with probability $\varepsilon$, or more moderate, and move towards the positions held by peers, with probability $(1 - \varepsilon)$. The change in the opinion of the extremists is one unit on the real line. We prove that the agent positions cluster in time, with all non-extremist agents located within a unit interval. However, the consensus opinion is dynamic. Due to the extremists' erratic behavior the clustered opinion set performs a "sluggish" random walk on the entire range of possible ideological positions (the real line). The inertia of the group, the reluctance of the society's agents to change their consensus opinion, increases with the size of the group. The extremists perform biased random walk excursions to the right and left and, in time, their actions succeed to move the society of agents in random directions. The "far left" agent effectively pushes the group consensus toward the right, while the "far right" agent counter-balances the push and causes the consensus to move toward the left.
We believe that this model, and some of its variations, has the potential to explain the real world swings in societal ideologies that we see around us. | cs.MA | cs |
Erratic Extremism causes Dynamic Consensus
(a new model for one-dimensional opinion dynamics)
Dmitry Rabinovich and Alfred M. Bruckstein
Center for Intelligent Systems (CIS)
Multi-Agent Robotic Systems (MARS) Laboratory
Computer Science Dept.
Technion Israel Institute of Technology
3200003, Haifa, Israel
June 6, 2019
Abstract
A society of agents, with ideological positions, or "opinions" measured
by real values ranging from −∞ (the "far left") to +∞ (the "far right"), is
considered. At fixed (unit) time intervals agents repeatedly reconsider and
change their opinions if and only if they find themselves at the extremes of
the range of ideological positions held by members of the society. Extrem-
ist agents are erratic: they become either more radical, and move away
from the positions of other agents, with probability ε, or more moderate,
and move towards the positions held by peers, with probability (1 − ε).
The change in the opinion of the extremists is one unit on the real line.
We prove that the agent positions cluster in time, with all non-extremist
agents located within a unit interval. However, the consensus opinion is
dynamic. Due to the extremists' erratic behavior the clustered opinion
set performs a "sluggish" random walk on the entire range of possible
ideological positions (the real line). The inertia of the group, the reluc-
tance of the society's agents to change their consensus opinion, increases
with the size of the group. The extremists perform biased random walk
excursions to the right and left and, in time, their actions succeed to move
the society of agents in random directions. The "far left" agent effectively
pushes the group consensus toward the right, while the "far right" agent
counter-balances the push and causes the consensus to move toward the
left.
We believe that this model, and some of its variations, has the potential
to explain the real world swings in societal ideologies that we see around
us.
1
1
Introduction
Over the years, social psychologists proposed numerous explanations for the
complex behavior emerging in large groups of supposedly intelligent agents, like
tribes and nations. They proposed models and principles of individual behavior
and some of these models were even amenable to mathematical analysis enabling
predictions about long-term behavior and the inevitable emergence of surprising
global economic or political phenomena.
The ideas of balance theory [Cartwright and Harary 1956] and social disso-
nance [Festinger 1962] led to the consideration of several basic mathematical
models, attempting to incorporate the idea that individuals, or agents attempt
to reach an equilibrium between their drives, opinions and "local comfort" and
those in their neighborhood. They do so by adjusting their position (ideological,
political, economic, or spatial) to be similar, or comfortably near the position
of their neighbors.
Simplified mathematical models for multi-agent interaction consider a group,
colony, society or swarm of agents, each agent associated with a quantity which
can be a real number, or a vector, describing the "state", opinion or position of
the agent. The state of the whole group (at time t) is specified by the vector
X(t) (cid:44) [x1(t), x2(t), . . . , xN (t)]T , where xk(t) is the state of agent k at time t,
and the group comprises N agents.
Then, models postulate that, from some initialization X(0) at time t = 0,
the state of the system evolves, and, if we consider that changes happen at
equal intervals (arbitrarily set to one), we obtain general discrete time evolution
models of the following form
(cid:40)X(t + 1) = Ψ(X(t))
X(0) − initial condition
Here Ψ describes the way each agent k, determines its state at time (t + 1) given
the states of all agents at time t.
The inter-agent interaction function Ψ is designed to reflect the assumed
influence of agents on their peers.
[DeGroot 1974] postulated that Ψ should
be a fixed matrix A acting on X with columns displaying the influence each
agent has on every other agent. Rows of the matrix then display how the next
state of agent k at time (t + 1) will be computed as a weighted combination
of the states of all agents at time t. If A is constant (and independent of the
state at all times) the vector X has a linear evolution, with dynamics completely
determined only by the eigen-structure of A and the initial state.
When positive entries and convex combination of states are postulated, A is
a stochastic matrix, and then one readily has, under quite general conditions,
that the system asymptotically achieves consensus, i.e. as t → ∞ we have that
all xk(t)'s will evolve to have the same value.
This model is highly appealing, however it assumes that each agent always
adjusts its state according to a fixed convex combination of its own state and
all other states. Since real individuals in any group are well known to posses a
certain reluctance in considering far-away positions of others, and tend to stick
to their initial opinions, models that took such tendencies into consideration
2
soon emerged. The very popular Hegselmann-Krause (HK) model [Hegselmann
and Krause 2002] postulates that
(cid:88)
1
Nk
xk(t + 1) =
xl(t),
l∈Nk
where the index set Nk (cid:44) {l
: (cid:107)xk(t) − xl(t)(cid:107) < εk}, i.e. Nk defines an εk-
neighborhood of the k-th agent position, xk(t), at time t, and Nk is the size of
the set Nk.
This model leads, in general, to clusters of agents in local consensus at
different state values/positions, a phenomenon often observed in society. Several
variations based on this model were put forth in the literature and a lot of
research is still devoted to study their convergence and properties.
Another interesting variation of the DeGroot model was proposed in [Fried-
kin and Johnsen 1990]. This model assumes that each agent k remains faithful
to its initial position to a certain degree gk, 0 ≤ gk ≤ 1 and has a susceptibility
of 1−gk to be socially influenced by the other agents. The classical linear model
then becomes, in a matrix notation:
X(t + 1) = GX(0) + (I − G)AX(t),
Here G is a diagonal matrix with gk-s on the main diagonal, and I is the
identity matrix. This model leads to a spread of steady state positions that can
be predicted by a simple matrix inversion.
[Reynolds 1987],
[Dudek et al. 1993],
Following the footsteps of [DeGroot 1974],
[Friedkin and Johnsen 1990] and
[Hegselmann and Krause 2002], a considerable number of interesting "opinion
dynamics", "multi-agent" and "consensus"/"gathering" models have been pro-
posed. Over the years the research in the field split into several branches. Today
researchers of Autonomous Swarms and Swarm Intelligence invent local inter-
action models to achieve "gathering", "geometric consensus", "collective area
sweeps" and "cooperative search and pursuit" with simple autonomous mobile
[Vicsek et al. 1995],
agents (see e.g.
[Ando et al. 1999],
[Olfati-
Saber and Murray 2004],
[S¸ahin 2004]). An overview of this field is provided in
[Barel et al. 2016]. Computer Scientists are interested in agreement and com-
mon knowledge in distributed computer networks ( [Halpern and Moses 1990],
[Shoham and Tennenholtz 1995],
[Flocchini et al. 2012]), while communication
engineers consider distributed coordination and collaboration in large, ad-hoc
networks of "cellular-phone" agents ( [Krishna et al. 1997],
[Chen et al. 2002],
[Chong and Kumar 2003]). Biologists analyze and try to understand and model
colonies of ants, flocks of starlings, schools of fish and swarms of locusts [Okubo
1986],
[Jadbabaie et al. 2003],
[Camazine et al. 2001],
[Camazine et al. 2001],
[Couzin and Krause 2003],
[Sumpter 2006].
Social science researchers continue to be interested in simulating and an-
alyzing human agent interactions, voting patterns and social opinion dynam-
ics. A recent survey by [Lorenz 2017] nicely presents the advancements and
clearly describes some of the issues of interest in the field. Stochastic models,
explicitly dealing with random behavior of agents with parameters probabilis-
tically characterizing their open-mindedness (the agent's probability of chang-
ing/reconsidering opinions), are currently being investigated. In his concluding
remarks, Lorentz states
3
"Agent-based models for the evolution of ideological landscapes are
still in infancy and it remains to show if they can add interesting
insight to political dynamics."
[Lorenz 2017], page 265
In this vein, we here propose a new, probabilistic, opinion dynamics model,
in part based on some early ideas of
[Festinger 1954]. He introduced a qual-
itative social psychology theory, supported by a vast corpus of data collected.
The theory suggests that the majority of agents hold neutral opinion on sub-
jects at hand. This majority is rather unmoved by extreme opinions, while the
"extremists" are unstable and tend to fluctuate, moving most probably in the
direction of a social norm.
We model opinions or ideological positions as real numbers and allow only
extreme agents to change opinions at discrete times by a constant quantum
value arbitrarily set to one in any direction. Changes in the positions of the
"extremists" in the direction of the "social norm", (represented by all agents
except the two "extremists"), are assumed to be highly probable. In the opposite
direction the erratic "extremists" may move, but with smaller probabilities. We
show that for any initial spread of agent opinions, a consensus opinion arises.
The "core" group in consensus spreads over an interval of size smaller than
the quantum change in the opinion of the extreme agents. The "core" is not
stationary and, over time, moves at random. In the society of agents "extremist"
is not a sticky label. From time to time one of the "extremists" becomes a part
of the "core" of normal agents; a previously "normative" moderate agent finds
itself to be at one of the extremes. It is these role-changes between "extremists"
and "moderates" that moves the "core" over time.
This paper is organized as follows. Section 2 presents the mathematical
model of opinion dynamics and states our main results. Section 3 reviews and
proves some basic facts about biased random walks. Section 4 analyzes the
gathering process by first considering a unilateral case in which we assume that
only one extremal agent is active, then a decoupling trick enables us to use the
unilateral results for the analysis of the problem when both extremal agents are
in action. Section 5 presents extensive simulation results confirming the theo-
retical predictions and showing that our bounds are quite loose due to the need
to decouple the action of the extremal agents in order to enable the theoreti-
cal results. The final Section 6 discusses possible interesting extensions of the
model presented along with some initial simulation results in two dimensions.
4
2 Model Description
Suppose a set of point agents, the individuals in the society, called p1, p2, . . . , pN
are, at the beginning of time, i.e. at t = 0, on the real line (the range of positions
or opinions) at locations x1(0), x2(0), . . . , xN (0) ∈ R. The agents are identical
and indistinguishable points and perform the following algorithm :
Agent decision rule (ε ∈ [0, 1
1 For pk located at xk(t) at discrete time t define intervals pR (cid:44) (xk(t),∞)
2 )):
and pL (cid:44) (−∞, xk(t)).
xk(t + 1) = xk(t), i.e. pk stays put.
2 if in both intervals pL and pR there are other agents then
3
4 else
5
if pR is empty then
pk makes a probabilistic jump, setting
xk(t + 1) =
xk(t) + 1
xk(t) − 1
, w.p. ε
, w.p. (1 − ε)
(cid:40)
(cid:40)
6
7
8
9
10
if pL is empty then
pk makes a probabilistic jump, setting
, w.p. (1 − ε)
, w.p. ε
xk(t) + 1
xk(t) − 1
xk(t + 1) =
11 end
Under the rule defined above only the two agents with extremal positions
xmin(t) and xmax(t) will move, and their tendency will be to approach the
agents in between. After each jump, carried out at discrete integer times, we
rename the identical agents to have them always indexed in the increasing order
of their x-locations. Hence at all discrete time instances t = 1, 2, . . . we have
the ordered agents {p1, p2, . . . , pN} with x1(t) ≤ x2(t) ≤ . . . ≤ xN (t), where p1
and pN are extremists and probabilistic jumps will be carried out by extremists
only (see Figure 1).
ε (1 − ε)
−1 +1
(1 − ε)
ε
−1 +1
p2
L = p1
x1(t)
. . .
xi(t) = [xi] + {xi}
pi
pi+1
. . .
pn−1 pN = R
xN (t)
Figure 1: N agents on the line
The process defined above evolves the constellation of points in time and we
clearly expect that a gathering of the agents will occur, since extremal agents
are probabilistically "attracted" toward their peers.
Indeed, if ε would be exactly zero, the deterministic jumps carried out by
the extremal agents p1 ≡ pL and pN ≡ pR would always be toward the interior
of the interval (x1, xN ), shortening it while (xN (t) − x1(t)) > 1. However, note
5
that when the [x1(t), xN (t)] interval reaches a value of 1 or less, interesting
things start to happen, since p1 and pN while jumping cross each other, in such
a way that the spread about the (time invariant, in this case) centroid of
points may increase and decrease in a way that depends on the specific spread
of the initial point locations' fractional parts. We therefore expect similar
things to happen when randomness is introduced as ε increases from 0 towards
1/2. For the time being, for simplicity, shall we assume that the fractional parts
of the distinct initial locations x1(0), x2(0), . . . , xN (0) are all different.
If ε = 0 we have the constellation at time t, {p1, p2, . . . , pN}t described by the
ordered set of point locations x1(t) < x2(t) < . . . < xN (t) and their centroid
and variance behave as follows: for the centroid arbitrarily chosen to be 0 at
time 0 we have
C(t + 1) (cid:44) 1
N
(cid:0)x1(t) + 1 + xN (t)− 1(cid:1) = C(t),
N−1(cid:88)
N(cid:88)
xi(t + 1) =
xi(t) +
1
N
1
N
i=1
i=2
and C(t+1) = C(t) = . . . C(0) (cid:44) 0, hence the centroid is an evolution invariant.
The variance of the constellation about the centroid at 0 is σ2(t) = 1
N
x2
i (t),
N(cid:80)
therefore
σ2(t + 1) = 1
N
N−1(cid:80)
N(cid:80)
i=2
(cid:0)x2
(cid:2)(xN (t) − x1(t)) − 1(cid:3)
1(t) + 2x1(t) + 1 + x2
i (t) + 1
x2
N
i (t) − 2 1
= 1
x2
N
= σ2(t) − 2
(cid:2)(xN (t) − x1(t)) − 1(cid:3)
i=1
N
N
i=1
N (t) − 2xN (t) + 1(cid:1)
While (xN − x1) > 1 the variance monotonically decreases, however when (xN −
x1) ≤ 1 we have σ2(t + 1) > σ2(t). Hence after gathering, or reaching consensus
(i.e. when xN − x1 ≤ 1), oscillations in σ2(t) subsequently occur, but the
constellation remains gathered around 0.
For the probabilistic case we expect a somewhat similar behavior. We shall see
that a "dynamic" consensus is reached. Agents on a line behaving according to
the probabilistic rule discussed above evolve to a dynamic constellation that is
"gathered" and the group of agents move on the line as follows:
1) For a given ε, 0 < ε < 1/2, we have
C(t + 1) = C(t) +
with probability ε(1 − ε)
with probability 1 − 2ε(1 − ε)
N ,
0,
N , with probability ε(1 − ε)
− 2
2
2) The "core" group of moderate agents, i.e. {p2, p3, . . . , pN−1} eventually
gathers to reside within a "dynamic" interval of length less than one.
3) The extremal agents p1 and pN perform random excursions to the left
and right of the core group, with motion biased towards the core. Their
bias ensures that they will be mostly near the core, the total distance
between them being a sum of random variables, one always less than 1
and two others bounded by positive random variables with a geometric
distribution.
6
3 Some Basic Facts About Random Walk
In order to analyze the gathering process due to the random behavior of the
extremal points (pL (cid:44) p1 and pR (cid:44) pN ) in case ε > 0 we need to first recall
some basic facts about random walks on the line. Suppose an agent performs
a (biased) random walk from an initial location (denoted by x(0) = 0) on the
real line, making, at discrete time instants t = 0, 1, 2, . . . moves to the left with
probability (1− ε) and to the right with probability ε. If ε = 1/2 the walk is the
unbiased, symmetric random walk, while ε < 1/2 biases the motion of the agent
towards the left. Let us define α as the positive departure of (ε) and (1 − ε)
from 1/2, i.e.
ε =
1
2
− α ⇔ 1 − ε =
1
2
+ α.
Clearly, α ∈ (0, 1/2), since we assume 0 < ε < 1/2. In this notation α quantifies
the bias towards left of the agents' motion and we have the following results.
3.1 The probability of reaching (-1) from 0.
The probability that the agent hits (−1) is given by the following expression:
P (walk hits (−1)) =
=
=
P (walk hits (−1) at (2k + 1) for the first time)
step to the left after making k steps to the
2 + α(cid:1) · Ck
(cid:0) 1
right and k steps to the left in any order,
i.e. returning to 0, without having been at
(−1)
(cid:0) 1
2 − α(cid:1)k(cid:0) 1
2 + α(cid:1)k
P
k=0
∞(cid:80)
∞(cid:80)
∞(cid:80)
k=0
k=0
It is well known [Stanley 2015],
Here Ck counts the number of possible paths of length k from 0 to 0 never
reaching (−1), which is given by the kth Catalan number.
erating function of the series {Ck} is given by :
1 − √
[Hilton and Pedersen 1991] that, the gen-
(cid:18)2k
(cid:19)
1
(1)
1 − 4x
2x
Ckxk =
∞(cid:88)
∞(cid:88)
P (walk hits (−1)) = (1/2 + α)· 1 −(cid:112)
Hence we have, for α > 0,
xk =
k + 1
k=0
k=0
k
1 − 4(1/4 − α2)
2(1/2 + α)(1/2 − α)
(1/2 + α) (1 − 2α)
(1/2 + α)(1 − 2α)
=
= 1
This is totally expected : a left biased random walk will almost surely (i.e. with
probability 1) reach (−1), when started at 0.
7
3.2 The probability of reaching (+1) from 0.
We have, similarly :
P (walk hits (+1)) =
P
P
k=0
time
(cid:16)walk hits (+1) at step 2k + 1 for the first
(cid:17)
∞(cid:80)
last step to the right after making k steps
∞(cid:80)
∞(cid:80)
2 − α(cid:1)
= (cid:0) 1
to the left and k steps to the right (i.e. re-
turning to 0) without having been at (+1)
2 − α) · Ck( 1
2 + α)k( 1
( 1
1−2α
2−α)(1+2α) = 1−2α
2 − α)k
1+2α < 1
k=0
k=0
=
=
( 1
Hence, while the walk almost surely reaches (−1), there is a non-zero probability,
given by 1 − 1−2α
1−ε of never reaching (+1).
1+2α = 1−2ε
3.3 The expected number of steps to first reach (-1).
Using the generating function for {Ck} we can readily calculate the expected
number of steps to reach (−1) from 0. Hence we have the following, (quite well
known) result :
E (steps to first hit (−1)) =
=
∞(cid:80)
∞(cid:80)
k=0
(cid:33)
(2k + 1) · P (walk hits (−1) at step 2k + 1)
(2k + 1)( 1
2 + α) · Ck( 1
(2k + 1)( 1
k=0
= ( 1
2 + α)
2 − α)k( 1
2 + α)k
k=0
∞(cid:80)
(cid:18) 1 − √
4 − α2)kCk
(cid:19)
=
1 − 4x
2x
1 − 2x − √
√
2x2
1 − 4x
1 − 4x
To compute this value explicitly we use
kCkxk−1 =
Ckxk
=
d
dx
∞(cid:88)
k=0
∞(cid:88)
k=0
hence we have,
kCkxk =
(cid:32) ∞(cid:88)
d
dx
∞(cid:88)
k=0
k=0
k
(cid:18)2k
(cid:19)
k + 1
k
1 − 2x − √
√
1 − 4x
2x
1 − 4x
xk =
(2)
(3)
This yields, setting x to ( 1
2 + α)( 1
2 − α), after some algebra,
E (steps to first hit (-1)) =
1
2α
=
1
1 − 2ε
Of course, the expected number of steps to reach (+1) is infinite, since there
is a strictly positive probability given by 1−2ε
1−ε of never getting there. But, we
know for sure that the biased random walk (more likely moving to the left) will
reach (−1) from 0 in the, above calculated, finite expected number of steps.
8
3.4 The expected farthest excursion to the right on the
way from 0 to first reaching (-1).
Another result that we shall need in analyzing the evaluation of the agents'
behavior is the following result on the excursions that biased random walks
make in the direction opposite to their preferred direction : the expected farthest
excursion to the right on the way from 0 to first reaching (−1) in the left biased
random walk is bounded by
E (farthest right excursion) ≤ ∞(cid:80)
∞(cid:80)
k=0
=
k=0
walk makes k right steps
2 + α(cid:1)(cid:0) 1
2 + α(cid:1)k
(cid:0) 1
and (k + 1) left steps to
first reach (-1)
2 − α(cid:1)k(cid:0) 1
k · P
kCk
Using the previously established relation
The above inequality can be explained as follows: any excursion that starts at
0 and eventually ends in −1 is necessarily of the odd length 2k + 1 for some k.
No matter what the actual order of steps is, the walk makes k steps to the right
and k + 1 steps to the left (with obvious limitations on the order of the steps).
Therefore, the farthest to the right such an excursion could get is a distance k
from 0. Hence, left-hand side of the inequality above is a clear upper bound.
1−4x we obtain
∞(cid:80)
kCkxk = 1−2x−√
√
2 − α(cid:1)k(cid:0) 1
(cid:0) 1
2 + α(cid:1)(cid:0) 1
E (farthest right excursion) ≤ ∞(cid:80)
2 + α(cid:1) ∞(cid:80)
2 − α(cid:1)k(cid:0) 1
= (cid:0) 1
(cid:0) 1
2 + α(cid:1)
= (cid:0) 1
2 + α(cid:1)k
2 + α(cid:1)k
1−4x
kCk
1
2−α
2α = ε
1−2ε
kCk
2−α)2
( 1
2−α)( 1
=
2 +α)
(4)
k=0
2α( 1
k=0
k=0
2x
9
4 Analysis of the Dynamic Gathering Process
4.1 Unilateral Action Results
In order to analyze the gathering process, let us first consider a one sided
version where only the rightmost agent moves at each moment and all other
agents stay put. Furthermore assume that to the left of p1 at t = 0 we
put a "beacon agent" p0 at x0(0) < x1(0). The rightmost agent at times
t = 1, 2, 3, . . . makes a unit jump to the left with high probability (1 − ε), or a
jump to the right with probability ε. Suppose the agents are initially located
at: x1(0), x2(0), . . . , xN−1(0), xN (0). Clearly the rightmost agent pR ≡ pN will
first reach, with probability 1, (xN (0) − 1) in
1−2ε expected number of steps,
then from (xN (0) − 1) it will reach a.s. (xN (0) − 2) in further
1−2ε expected
steps etc. until, at some point it will jump over xN−1(0) to land somewhere in
the interval (xN−1(0)− 1, xN−1(0)), making the agent at xN−1(0) the rightmost
agent. This will happen with probability 1, after a number of steps, which we
shall denote as Tjump, having the expected value of
number of steps.
(cid:2)(cid:98)xN (0)−xN−1(0)(cid:99)+1(cid:3)
1−2ε
1
1
1
-1
···
-1
-1
xN−1(0)
(cid:2)(cid:98)xN (0) − xN−1(0)(cid:99) + 1(cid:3)
Figure 2: The agent pN will jump over pN−1 after an expected number of steps
equal to
1
1−2ε
xN (0)
1
Now it will be the turn of the former pN−1 agent, which is now "renamed"
pN ≡ pR, to start its biased random walk and it will reach (xN−1(0) − 1) in
1−2ε expected steps (clearly jumping over at least the "current" position of the
"former" moving agent) to land in the interval (x1(0) − 1, xN (Tjump)) defined
by the "renamed" agents (p1, p2, . . . , pN−1). Clearly the new rightmost agent
(which might be the former random walker or another agent located to the left
of xN−1(0) in the initial configuration) will do the same.
Recall that we assume, for simplicity, that agents' initial locations
have all distinct fractional parts, so that one agent will never land on
top of another!
From the above description it is clear that the "erratic extremist" random
walk of rightmost agents will eventually "sweep" all the agents towards the left,
and in a finite expected number of steps equal to :
E(cid:0)T{x0(0),x1(0),...,xN (0)}(cid:1) =
N(cid:88)
k=1
1
1 − 2ε
((cid:98)xk(0) − x0(0)(cid:99) + 1),
all agents will be to the right of the "beacon" p0 after having jumped over x0(0)
exactly once, making the "beacon" p0 the rightmost agent for the first time!
Indeed, note that while jumping one over the other (to the left) all the
agents to the right of p0 will have carried out (perhaps with interruptions due
to reordering, following jumps over the agent called pN−1) a biased random walk
from their initial locations x1(0), x2(0), . . . , xN (0) until each one of them, for the
first time, jumped over the fixed "beacon" point p0 at x0(0). Subsequently, the
10
agents will stop and wait for the "beacon" p0 to become the rightmost agent.
This will happen when the last of all the agents (that were p0's initial right
neighbors) completes its random walk by jumping over p0.
An important byproduct of this analysis is the fact that, the moment after
the last right neighbor jumps over p0, all the other agents have made exactly
one left jump over p0 at x0(0), hence all the agents will be located in the
interval (x0(0) − 1, x0(0)]. Therefore we proved:
Theorem 1. If p0, p1, . . . , pN are located at t = 0 at x0(0), x1(0), . . . , xN (0)
with (x0(0) < x1(0) < . . . < xN (0)), and the rightmost agent performs random
walk biased toward the left with probability of a left unit jump of (1 − ε), the
agents first gather to the interval (x0(0)−1, x0(0)], with probability 1, in a finite
expected number of steps given by
E(cid:0)T{x0(0),x1(0),...,xN (0)}(cid:1) =
N(cid:88)
k=1
1
1 − 2ε
((cid:98)xk(0) − x0(0)(cid:99) + 1)
Note that we could have chosen in this description the "beacon" to be the
leftmost agent p1 located at x1(0) and then in a finite expected time of
E(cid:0)T{x1(0),x2(0),...,xN (0)}(cid:1) =
N(cid:88)
k=2
1
1 − 2ε
((cid:98)xk(0) − x1(0)(cid:99) + 1)
the agent p1 becomes the rightmost agent. If, beyond the "first gathering" to
the left of p1, the process continues indefinitely, the group of agents will be
pushed to the left due to the rightmost agent's actions with an average speed
of about 1−2ε/N.
Note also that we have the corresponding symmetric result for agent groups
where only the leftmost agent is moving and it sweeps all agents, by the action
of its biased random walk, towards the right, after gathering the group to an
interval of length bounded by 1.
4.2 Bilateral Action Results
So far we have seen that a unilateral random-walk, biased toward the group
of agents, carried out either by the rightmost or by the leftmost agent results
in gathering the agents into a cluster with a span upper bounded by 1 (i.e.
the step size). Something slightly more complex happens when both extremal
agents are jointly herding the group. Of course we expect gathering to happen,
and even faster than in the case when only one extremal agent is at work. This
is indeed the case, however the simultaneous work of the extremal agents leads
to interactions that slightly complicate the proofs.
Suppose we have a constellation of agents p1, p2, . . . , pN located at time
t = 0 at x1(0) < x2(0) < . . . < xN (0), as before. The "erratic extremists", the
leftmost and rightmost agents pL (cid:44) p1 and pR (cid:44) pN perform biased steps by
simultaneously jumping, towards the agents {p2, p3, . . . , pN−1} with probability
(1 − ε) or away from them with probability ε.
The results below represent the main contribution of this paper. Theorem 2
states that if the internal agents are gathered in an interval smaller than the step
size, they never spread beyond this size. Theorem 3 bounds the expected time
11
to shrink the excess distance, beyond one, between p2 and pN−1 (the internal
agent span) by one half. Theorem 4 then uses the fact that, once less than 2,
the distances xN−1(t) − x2(t) can only take a finite set of values, to show that
the inner agents gather to an interval of length less than 1 in finite expected
time. Theorem 5 uses the bounds on the expected excursions of biased random
walks in the direction opposite to the bias to prove that, with high probability,
the total span of all the agents will have a small value as the process continues
to evolve after the "core" gathered.
Theorem 2. Suppose at t = T the internal agents {p2, p3, . . . , pN−1} are all
close, so that xN−1(T ) − x2(T ) ≤ 1, then xN−1(T + 1) − x2(T + 1) ≤ 1. Hence
for all t > T we will have xN−1(t) − x2(t) ≤ 1.
Proof. Assume xN−1(T ) − x2(T ) ≤ 1. Designate by AL and AR the agents
x2(T ) and xN−1(T ), respectively. After jumps by extremal agents we can have
at t = T + 1 the following cases:
• AL and AR both remained internal. Then all the internal agents are still
inside the interval [x2(T ), xN−1(T )] with assumed length of at most one.
• AL and AR both became extremal. This case is even simpler: all the inter-
nal agents at time T +1 are now strictly inside the interval [x2(T ), xN−1(T )]
with assumed length of at most one.
• Either AL or AR only became an extremal agent. Assume w.l.o.g. that
agent AL at location x2(T ) became extremal, i.e. x1(T + 1) = x2(T ). In
this case all the internal agents are contained in either [x2(T ), xN−1(T )]
(because the left extremal agents moved into it, see Figure 3a) or
[x2(T ), x1(T ) + 1] (because the left extremal agent over-jumped all the
previous internal agents, see Figure 3b). In both cases, the interval con-
taining new internal agents is of length at most one.
AL
AR
AL
x1(T )
x2(T )
xN−1(T )
x1(T )
x2(T )
AR
xN−1(T )
(a) Jump inside
(b) Over-jump
Figure 3: Left extremal agent jump (a)into/(b)over the internal agent interval.
Hence in all possible cases the span of the gathered agents at the next step never
exceeds one.
The next theorem demonstrates that the size of internal agents' interval,
if bigger than one, will be reduced in finite expected time by one-half of the
difference between the interval size and 1. We shall then exploit the fact that
the number of agents is finite and that the shrinkage can not be infinitesimal,
to show that the interval indeed will attain a size less than 1, in finite expected
time.
12
Theorem 3. Let agents p1, p2, . . . , pN be initially located at x1(0), x2(0), . . . ,
xN (0), their behavior being governed by the motion model we consider. Suppose
xN−1(0) − x2(0) = 1 + S0 for some S0 > 0, i.e. internal agents are not initially
gathered inside a unit interval. Let T = inf{t
2 } -
be the first time, when all the internal agents are inside an interval bounded by
1 + S0
: xN−1(t) − x2(t) ≤ 1 + S0
2 , then
(cid:18)
(cid:24) S0
(cid:25)
2
E (T ) <
1
1 − 2ε
(N − 2)
+ (xN (0) − x1(0) − 1)
.
(cid:19)
Proof. Locate two fictional "beacon agents" pL
as follows:
F and pR
F at the locations defined
(a) pL
F at xL
(b) pR
F at xR
F (0) = x2(0) + S0
2
F (0) = xN−1(0) − S0
F (0) − xL
2
F (0) = 1 + S0 − 2 S0
Obviously, xR
2 = 1.
Now consider the agents to the right of xR
F (0) and the action of pR and the
agents to the left of xL
F (0) and the action in time by pL. Clearly there will be no
interaction between the two dynamic processes to the left and to the right of
F (0),∞),
agents located in either the interval (−∞, xL
F (0)(cid:3) until one of the agents pR or pL will fully sweep all
F (0)(cid:3). Indeed no agents from the left can cross
the interval(cid:2)xL
into the unit interval(cid:2)xL
F (0)) or in the interval (xR
F (0), xR
F (0), xR
into the right region until all of them "jumped the fence" at xL
happens in the opposite direction!
F (0) and the same
Therefore we have that in a finite expected time upper bounded by
(cid:18)
(cid:24) S0
(cid:25)
2
1
1 − 2ε
(N − 2)
+ (xN (0) − x1(0) − 1)
(cid:19)
the span of the "internal", non-mobile agents will shrink to be at most 1 + S0
2 .
The bound is explained as follows : if we denote by TL - a random time it
takes the agents left of xL
F (0) to "jump the fence" and by TR - the random time
it takes the agents right of xR
F (0) to "jump the fence", then clearly T , the first
moment when one of the S0
2 intervals will be cleared of agents is bounded above
by min{TL, TR}. We have then, that in the worst case, we will need at most all
one must move all the way to reach the fence. Hence E (T = min{TL, TR}) <
E (worst extremal excursion time), which is the expression above.
internal agents to be swept a distance of at most (cid:6) S0
(cid:7), and also an extremal
2
We next prove the following simple fact
Lemma 1. Let x1, x2, . . . xn be a set of real numbers, such that {xi} (cid:54)= {xj}
for all i (cid:54)= j (i.e. their fractional parts are all different). Define
{{xi} − {xj} , 1 − {xi} − {xj}}
d := min
i(cid:54)=j
xi − xj > 1, we must have that: xi − xj ≥ 1 + d.
Then, if for some i, j
13
Proof. Write xi = si + ri, where ri ∈ [0, 1) and si ∈ Z. Then xi − xj =
(si − sj) + (ri − rj) and −1 < ri − rj < 1.
If xi − xj > 1, then two cases are possible:
• ri > rj, then xi − xj = (si − sj) + (ri − rj), si − sj ≥ 1 and ri − rj =
ri − rj = {xi} − {xj} ≥ d. This yields xi − xj ≥ 1 + d.
• ri < rj, then xi − xj = (si − sj − 1) + (1− (rj − ri)), (si − sj − 1) ≥ 1 and
(1 − (rj − ri)) = 1 − rj − ri = 1 − {xj} − {xi} ≥ d. This again yields
xi − xj ≥ 1 + d.
In case xi − xj < −1, it follows xj − xi > 1 and we apply previous argument
by exchanging roles of indexes i and j. Hence in both cases the claim follows.
Note, if {xi(0)} are the fractional parts of the initial locations of the agents
on the line, then these fractional parts are invariant under the evolution process
since agents jump unit steps.
Assuming, as we do, that all initial fractional parts are distinct we have the
following result: Define d as in Lemma 1 to be the smallest fractional difference
of all the initial agent pair locations. If xN−1(t) − x2(t) > 1, then necessarily
xN−1(t) − x2(t) ≥ 1 + d.
Theorem 4. In the setting of Theorem 3, let T = inf{t : xN−1(t)−x2(t) ≤ 1},
i.e. the first time when all the internal agents are inside an interval bounded by
1, then
E (T ) <
1
1 − 2ε
(N · (S0 + (cid:100)log2
S0
d
(cid:101)) + (xN (0) − x1(0) − S0 − 1)).
Proof. From the Theorem 3, given that at time 0, (xN−1(0) − x2(0)) = 1 + S0,
we have at a random time T1 with finite expectation that (xN−1(T1)−x2(T1)) ≤
1 + S0
2 . We next consider the process with the constellation of agents at the
moment where one of the active extremal agents cleared out an interval of length
S0
2 on one side of the span of "internal agents". At this moment (T1, the initial
time for the next phase) all internal agents are spanning an interval of length
at most 1 + S0
2 . Therefore by Theorem 3, after a random time span of T2, again
having finite expectation, we find the internal points gathered within an interval
of 1 + S0
4 , etc.
After k such steps, each with finite expected duration, we shall find the
internal agents within an interval of length at most 1 + S0
2k . The decrease of the
upper bound value on the span of internal agents at step k will be at least S0
2k+1 .
Recall now that d is the smallest fractional difference of all possible agent pair
locations. Suppose at step kf (at time T ∗ := T1 + T2 + . . . + Tkf ), we attain for
2kf < d but, still we have xN−1(T ∗) − x2(T ∗) > 1. By Lemma 1
the first time S0
we must have
xN−1(T
∗
) − x2(T
∗
) ≥ 1 + d
However, since
xN−1(T
∗
) − x2(T
∗
) ≤ 1 +
S0
2kf
< 1 + d
14
Using the upper bound for every T1, T2, . . . we obtain
E (T ) ≤ E (T1) + E (T2) + . . . + E(cid:0)Tkf
(cid:24)
(cid:1)
(cid:25)
N(cid:100) S0
4 (cid:101)
1−2ε + . . . +
N
S0
kf
2
1−2ε + ∆
2 (cid:101)
≤ N(cid:100) S0
1−2ε +
≤ N (S0+kf )
leads to a contradiction, we must have an interval xN−1(T ∗) − x2(T ∗) ≤ 1 and
d (cid:101) all the internal
T ≤ T ∗. This proves that, at some step before kf = (cid:100)log2
points will be gathered in an interval of unit length.
S0
1−2ε + ∆
need to sweep by their biased walk distances of (cid:6) S0
(cid:7) ,(cid:6) S0
4
(cid:7) , . . . respectively,
2
We still need to evaluate ∆ here. Starting at each time T1, T2, . . . extremal agents
with the exception of the first interval T1, when an additional initial "gap" had
to be traversed by one of extremal agents. The possible initial "gaps" were,
x2(0) − x1(0) and xN (0) − xN−1(0). For the upper bound we take the initial
traversal length to be the sum of these quantities. After reordering we have for
∆ = (xN (0)−x1(0))−1−S0
, hence we obtain
1−2ε
N · (S0 + (cid:100)log2
E (T ) <
S0
d (cid:101)) + (xN (0) − x1(0) − S0 − 1)
1 − 2ε
To summarize, we have the following results so far:
• Consider the span of the non-extremal agents' constellation at time t = 0
on R as
L(0) (cid:44) xN−1(0) − x2(0) (cid:44) 1 + S0
and with S0 > 0. Due to the actions of the "erratic extremist" agents,
while the span of the "core" agents is greater than 1 (i.e. it is L(t) = 1+S,
with S > 0), we have that x2(t), the location of the second agent in the
reordered naming of agents, can only increase, and similarly xN−1(t) can
only decrease. Hence, while L(t) is bigger than one, it will be a non-
increasing sequence in time.
In finite expected time L(t) becomes less
than 1 and the subsequent actions of the extremists can never make it
exceed 1.
• Following the gathering of the "core" agents to a consensus interval less
than 1 after a finite expected time, the total distance between p1 and pN
will be a sum of three parts : the interval occupied by "core" agents of
size at most 1 and two distances from the consensus "core" interval to the
left and right "extremists".
15
4.3 The total span of agents after gathering
In subsection 3.4 we provided a bound of ε/(1−2ε) on the expected length of
maximal excursions of an extremal agent from a fixed point. Since expectation
is linear we can provide a rough bound on the total span of agent locations as the
sum of 1, (which upper bounds the span of the gathered "core", or consensus
agents) and expected maximal excursions to the left and right made by the
"extremist" agents. This argument yields, roughly
E (xN (t) − x1(t)) ≤ 1 +
2ε
1 − 2ε
The Markov's inequality (∀a > 0 P (X ≥ a) ≤ E(X)
a ) then provides
P (xN (t) − x1(t) ≥ k) ≤ 1
k
+
·
2ε
k
1
1 − 2ε
≈ 1
k
(5)
Therefore, we have qualitatively that P (xN (t) − x1(t) ∈ [k, k + 1]) = Θ( 1
However, we can do even better.
k2 ).
1 − ε
1
ε
ε
ε
ε
2
3
4
. . .
1 − ε
1 − ε
1 − ε
1 − ε
Figure 4: Left-biased bounded random walk used to bound extremal agent
distance from the internal core agents
Let us introduce a left-biased, partially reflective and bounded-from-the-left
random walk on the state space {1, 2, . . .} (see Figure 4). Further, consider each
state as representing the extremal agent's current distance from the farthest
internal agent rounded to the closest bigger integer. The probability to move
right, i.e. away from the "core" (which is the gathered, internal agents span), at
every state is ε, and the probability to move closer to the "core" is (1− ε). The
right extremal agent can, with probability (1− ε), jump to the left, but at state
1 such a jump constitutes a move over all the internal agents. In this case, a
new extremal agent "emerges", maintaining the distance from the farthest left
internal agent just below 1 (e.g. Theorem 2). This can happen in two ways.
Either the other extremal agent jumps over the "core", or the closest internal
agent becomes "exposed" and turns into the right extremal one.
After the convergence of the internal agents, suppose we couple the (right)
extremal agent's moves to the above defined random walk, i.e. the random walk
proceeds exactly following the decisions of extremal agent.
Claim 4.1. The random walk defined above provides an upper bound on the
distance of the extremal agent (at xN (t)) from the farthest internal agents (at
x2(t)).
Proof. Let X(t) denote the state of random walk at time t. Suppose at time
t = T , X(T ) ≥ xN (T )−x2(T ), i.e. random walk is at state 'at least the distance
of right extremal agent from the farthest internal agent'. At t = T + 1 one of
the following things can happen.
16
• The right extremal agent decides to jump right. In such a case, the dis-
tance to the extremal agent increases by at most 1, which corresponds to
an increase in the random walk position.
X(T + 1) = X(T ) + 1
= xN (T + 1) − x2(T ) ≥ xN (T + 1) − x2(T + 1)
≥ (xN (T ) + 1) − x2(T )
The last inequality follows from the fact, that the left-most internal agent
can only "move" to the right, due to the action of the left "extremist".
• The right extremal agent decides to jump left, but remains the right ex-
tremal agent at t = T + 1, and xN (T ) − x2(T ) > 1. Therefore,
X(T + 1) = X(T ) − 1
= xN (T + 1) − x2(T ) ≥ xN (T + 1) − x2(T + 1)
≥ (xN (T ) − 1) − x2(T )
The last inequality is explained as in the preceding case.
• The right extremal agent decides to jump left, stops being the right ex-
tremal agent at t = T + 1, and xN (T ) − x2(T ) > 1. We assumed, that
X(T ) ≥ xN (T ) − x2(T ) which is equivalent to X(T ) ≥ 2, hence by def-
inition of coupling X(T + 1) ≥ 1. Two situations are possible: the in-
ternal agent at xN−1(T ) "emerged" to be the right extremal agent at
time T + 1 or the left extremal agent at time T + 1 jumps over all the
other agents to the right and becomes the right extremal one. We have
x2(T + 1) ≥ x2(T ), since the right extremal agent becomes the internal
agent. Also xN (T + 1) = x1(T ) + 1 ≤ x2(T ) + 1, since only the extremal
agents actually move. In both cases it follows that
xN (T + 1) − x2(T + 1) ≤ (x2(T ) + 1) − x2(T ) = 1 ≤ X(T + 1)
• The right extremal agent decides to jump left and xN (T ) − x2(T ) ≤ 1. In
such a case X(T ) ≥ 1 and X(T + 1) ≥ 1, because 1 is the lowest value the
random walk could attain. The distinctive difference from the previous
case is that the right extremal agent moves over all the internal agents. We
have then three cases. The first is when the right extremal agent becomes
the left-most internal agent, hence
x2(T + 1) = xN (T ) − 1 ≥ xN−1(T ) − 1 = xN (T + 1) − 1.
In the second and third cases, it becomes the left extremal agent. We dif-
ferentiate between those two cases considering the new role of the previous
left extremal agent. If it becomes a new right extremal agent, we have
xN (T + 1) = x1(T ) + 1 ≤ x2(T ) + 1 = x2(T + 1) + 1.
Otherwise,
xN (T + 1) = xN−1(T )
= x1(T + 1) + 1 ≤ x2(T + 1) + 1
≤ (xN (T ) − 1) + 1
In all the above cases, we conclude xN (T + 1)− x2(T + 1) ≤ 1 ≤ X(T + 1),
as claimed.
17
Returning to analyze the "upper bounding" random walk we have the follow-
ing: if ε < 1/2 the above random walk is positive recurrent, and aperiodic, hence
has a stationary distribution π that is determined by the balance equations
ε · π(k) = (1 − ε) · π(k + 1)
∞(cid:80)
Along with the normalization condition
π(k) = 1, this provides the steady
state distribution π = [π(1) π(2)
(cid:18) ε
1 − ε
(cid:19)k−1 1 − 2ε
1 − ε
π(k) =
k=1
. . .] with
∀k ∈ {1, 2, . . .}.
(6)
The above analysis is symmetrically applicable to the random walk of the
left extremal agent. We then have two independent and identically distributed
walks, upper bounding the distance of the right and left extremal agents from
the "core's" left and right boundaries. Denoting them by X(t) and Y (t), we
have
xN (t) − x1(t) ≤ (xN (t) − x2(t)) + (xN−1(t) − x1(t)) = X(t) + Y (t)
Here we are interested in assessing P (xN (t) − x1(t) ≤ k), hence we can
estimate a lower bound for P (xN (t) − x1(t) ≤ k) by P (X(t) + Y (t) ≤ k).
Therefore, consider
P (X + Y ≥ k) =
P (X = i)P (Y ≥ k − i) + P (X ≥ k − 1)
(7)
In the steady state we have that P (X = k) is just a π(k), and P (X ≥ k) =
π(i). Therefore,
i=1
∞(cid:80)
i=k
k−2(cid:88)
Thus, we obtained the following simple expression
∞(cid:88)
i=k
π(i) =
P (X ≥ k) =
(cid:19)k−1
=
1 − ε
(cid:18) ε
i=k
1 − ε
1 − ε
(cid:19)i−1 1 − 2ε
(cid:18) ε
∞(cid:88)
(cid:18) ε
(cid:19)k−1
(cid:18) ε
(cid:19)i−1 1 − 2ε
k−2(cid:88)
(cid:19)k−2(cid:18)
(cid:18) ε
1 − ε
1 − ε
1 − ε
i=1
which after few algebraic manipulations provides
P (X + Y ≥ k) =
1 − ε
(k − 2) · 1 − 2ε
1 − ε
We summarize these findings as follows :
Using this result in (7) produces for k greater than two
P (X + Y ≥ k) =
(cid:18) ε
1 − ε
(cid:19)k−2
,
+
(cid:18) ε
·
1 − ε
(cid:19)k−i−1
(cid:19)
+ 1
.
Theorem 5. After the internal agents gather in an interval of length 1, the
distribution of interval lengths' containing all the agents is lower bounded by
(cid:19)k−2
(cid:18) ε
1 − ε
.
(8)
P (xN (t) − x1(t) < k) ≤ P (X + Y < k) ≈ 1 − k
18
4.4 On arbitrary initial position of agents
In proving Theorem 4 we have assumed all the agents locations' fractional parts
are different. We can slightly change the model to accommodate for cases in
which some agents may share the same location. Of course, the problem arises
when several agents find themselves sharing extremal locations. In such cases
their motions must be specified and disambiguated. Suppose several agents
share the same place and all other agents are located on exactly one side either
to the left, or to the right. We assume that only one of these extremal agents
will become "erratic", and move at a given time. We can then readily prove a
claim equivalent to Theorem 4 in this new model.
Theorem 6. Let agents p1, p2, . . . , pN be initially located at x1(0), x2(0), . . .,
xN (0), and define
T := inf{t : xN−1(t) − x2(t) ≤ 1}
be the first time, when all the internal agents are inside the interval bounded by
1, then with the modified rule of behavior we have E (T ) < ∞.
Proof. Since we are not assuming that fractional parts are all different, it is
possible that there will be more than one agent with the same fractional part
of their initial (and subsequent) locations. Let ∆ be the minimal fractional
non-zero distance between two agents.
∆ :=
min
{xj (0)}(cid:54)={xk(0)}
{{xj(0) − xk(0)}, 1 − {xj(0) − xk(0)}},
Step 1. Define a new process with the following initial coordinates:
In case all the agents share the same fractional part, simply set ∆ := 1.
yk(0) = xk(0) + (k−1)∆
for all k ∈ {1, 2, . . . , N}. It is not difficult to see that
the "newly defined locations" y1(0), y2(0), . . . , yN (0) fulfill the requirements of
Theorem 4.
N
Step 2. The Theorem 4 proves that all agents (yk)n
k=1 gather in expected
finite time to the interval of unit length. Denote this time by Ty.
• By separately handling cases of same and different initial fractional part
of the location one can show that for all k ≥ 2,
x2(t) ≤ xk(t)
In the same manner, for all k ≤ N − 1
xk(t) ≤ xN−1(t)
We can then conclude, that all the correspondingly indexed x and their
"shadow y-agents" will be called inner and extremal in both models at the
same time.
19
• Suppose y2(Ty) and yN−1(Ty) are agents which originally had the same
fractional part of their respective location. Due to the way, we mapped
the coordinates, we know that y2(Ty) < x2(Ty) + ∆ ≤ x2(Ty) + 1 and that
yN−1(Ty) ≥ xN−1(Ty), hence xN−1(Ty) − x2(Ty) < 2. But, since x2(Ty)
and xN−1(Ty) have the same fractional part, we conclude that we have
xN−1(Ty) − x2(Ty) ≤ 1.
• If y2(Ty) and yN−1(Ty) are not agents which originally had the same
fractional part of their respective location, then one of two cases is possible.
If yN−1(Ty) < x2(Ty) + 1 we have all agents in original model inside
the interval [x2(Ty), x2(Ty) + 1). Otherwise x2(Ty) + 1 ≤ yN−1(Ty) ≤
y2(Ty) + 1. But, due to definition of ∆, only points, which have the same
fractional part, as x2(Ty) could fall between x2(Ty) + 1 and y2(Ty) + 1.
Hence the latter case is impossible.
It follows, that in all cases xN−1(Ty) − x2(Ty) ≤ 1, which implies T ≤ Ty, and
by Theorem 4, Ty has a finite expectation.
5 Simulation results
We next present some simulation results to showcase the validity of the above-
presented theoretical predictions. In Figure 5 we present the result of simulation
runs with a different values for ε and fixed number of agents N located at random
points uniformly distributed in an initial interval of size 1 + S0. The simulations
measured the time to convergence of the inner agents to an interval of length
one. The theory predicts that the expected time to gathering is bounded as
follows
E (T ) ≤ N · (S0 + (cid:100)log2
S0
d (cid:101)) + (xN (0) − x1(0) − S0 − 1)
1 − 2ε
As predicted, the average convergence times exhibit a hyperbolic dependence
on ε. Figure 6 clearly showcases the linear functional dependence between the
convergence time and the initial span of internal agents, and implicitly to the
initial span of all agents. Varying the number of agents N supplies another
linear dependency as can be seen in Figure 7.
20
·106
Actual averages of 100 runs
Theoretical upper bound
1.2
1
0.8
0.6
0.4
0.2
)
s
p
e
t
s
f
o
#
(
)
T
(
E
0
0
5 · 10−2
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
ε
(a) N = 400, S0 = 500
Bound to actual rate
8.02
8
7.98
7.96
7.94
7.92
7.9
)
T
(
E
l
a
u
t
c
A
/
d
n
u
o
b
)
T
(
E
7.88
0
5 · 10−2
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
ε
(b) N = 400, S0 = 500
Figure 5: (a) Convergence times as a function of probability of motion in the wrong
direction(ε), N = 400, S0 = 500.
(b) Theoretical upper bound to measured con-
vergence time ratio vs ε. Each point on the actual results' line is an average of 100
different simulations with the same parameters set.
21
Actual averages of 100 runs
Theoretical upper bound
100
200
300
400
500
S0
600
700
800
900
1,000
(a) N = 400, ε = 0.1
·105
6
5
4
3
2
1
0
0
8
7.9
7.8
7.7
7.6
7.5
7.4
7.3
7.2
7.1
)
s
p
e
t
s
f
o
#
(
)
T
(
E
)
T
(
E
l
a
u
t
c
A
/
d
n
u
o
b
)
T
(
E
7
0
100
200
300
400
500
S0
Bound to actual rate
600
700
800
900
1,000
(b) N = 400, ε = 0.1
Figure 6: (a) Convergence times as a function of initial span (S0). (b) Ratio between
the theoretical upper bound to measured convergence time vs S0. Each point on the
actual results' line is an average of 100 different simulations with the same parameter
set.
22
·106
Actual averages of 100 runs
Theoretical upper bound
1.4
1.2
1
0.8
0.6
0.4
0.2
)
s
p
e
t
s
f
o
#
(
)
T
(
E
0
0
100
200
300
400
600
700
800
900
1,000
500
N
(a) S0 = 1000, ε = 0.1
8
7.95
7.9
7.85
7.8
7.75
7.7
7.65
7.6
7.55
)
T
(
E
l
a
u
t
c
A
/
d
n
u
o
b
)
T
(
E
7.5
0
100
200
300
400
500
N
Bound to actual rate
600
700
800
900
1,000
(b) S0 = 1000, ε = 0.1
Figure 7: (a) Convergence times as a function of number of agents (N ). (b) Theo-
retical upper bound to measured convergence time vs N . Each point on the actual
results' line is an average of 100 different simulations with the same parameters set.
23
In all above experiments we notice that our theoretical bounds are roughly
eight times higher than the actual measurements. Recall, that we derived our
results based on overly cautious assumptions, namely that one extremal agent
is doing the constructive work, while the second extremal agent is randomly
wandering outside the interval containing the internal agents. In reality, this
is not the case: both agents independently and concurrently contribute to con-
vergence. Hence we should focus on the stochastic process, which is in some
sense "the distance between two independent random walks biased towards each
other".
Let X, Y be two independent biased random walks with a probability ε to
jump right. For any one of mentioned random walks E (step length) = 2ε − 1.
On the other hand, for the process Z (cid:44) X + Y , E (step length of Z) = 2(2ε −
1). Note, that the "contraction" process Z describes the distance between two
extremal agents, with each extremal agent sweeping the internal agents in the
direction of its counterpart. Furthermore, on the average, the core convergence
will happen approximately around the middle of the initial interval. And we
should finally recall the assumed uniform initial spread of agents inside the
initial interval at the beginning, implying that each extremal agent will need
"to push" only half of the internal agents. Thus until convergence we have two
stochastic processes of the kind we analyzed in this paper, and each starts with
half the number of agents, and half the initial interval, and the process will
proceed at least twice as fast. Hence, we have 3 factors that each improve by 2
the time to convergence (hence the 8!)
Another aim of our simulations was to asses on the bounds on the total span
of agents after gathering. As can be seen in Figure 8, using a semi-logarithmic
scaling, the probability to find an extremal agent at a specific distance is indeed
decreasing exponentially fast, according to the bound of Theorem 5. The same
is true about the results predicted in subsection 4.3. In Figure 8b we show that
application of Theorem 5 gives a much better bound than the crude evaluation
of (5).
Figure 9 presents a typical behavior of the Gathered Core's center of mass
and of the two extremal agents. The period before gathering is shown in Fig-
ure 9a followed by a display of "post-gathering" typical behavior in Figure 9b.
Simulations for different number of agents are shown in Figure 10 and Figure 11.
Unsurprisingly, Figures 10b and 11b prove a much higher inertia of the Core
center of mass to the actions of extremal agents, when the number of agents is
five times higher.
24
N = 10
N = 20
N = 50
N = 100
N = 200
102
101
100
10−1
10−2
10−3
10−4
10−5
d
>
n
a
p
S
e
m
i
t
f
o
%
10−6
0
2
4
6
10
8
12
Total Distance, d
14
16
18
20
(a) Long-term distribution of agent span
)
d
<
n
a
p
S
(
P
100
10−0.05
10−0.1
10−0.15
10−0.2
10−0.25
10−0.3
N = 10
N = 20
N = 50
N = 100
N = 200
Theoretical lower bound (Theorem 5)
Theoretical lower bound (crude) (Equation (5))
0
2
4
6
8
10
12
14
16
18
20
Total Distance, d
(b) Cumulative distribution of spans
Figure 8: (a) Long-term (steady-state) distribution of agents' total span, (b) Long-
term cumulative distribution of total span with lower bounds from subsection 3.4 and
subsection 4.3. The simulations were done for different values of N and ε = 0.1.
25
50
40
30
20
10
0
t
e
s
ff
O
−10
−20
−30
−40
−50
0
3
2.5
2
1.5
1
0.5
0
−0.5
−1
−1.5
−2
−2.5
−3
−3.5
−4
0
0
0
1
t
e
s
ff
O
Core-Center
Left Extremal Point
Right Extremal Point
0
5
0
0
1
0
5
1
0
0
2
0
5
2
0
0
3
0
5
3
0
0
4
0
5
4
0
0
5
0
5
5
0
0
6
0
5
6
0
0
7
0
5
7
0
0
8
0
5
8
0
0
9
0
5
9
0
0
0
1
(a) First 1,000 iterations : The gathering process
Iteration
Core-Center
Left Extremal Point
Right Extremal Point
0
5
0
1
0
0
1
1
0
5
1
1
0
0
2
1
0
5
2
1
0
0
3
1
0
5
3
1
0
0
4
1
0
5
4
1
0
0
5
1
0
5
5
1
0
0
6
1
0
5
6
1
0
0
7
1
0
5
7
1
0
0
8
1
0
5
8
1
0
0
9
1
0
5
9
1
0
0
0
2
Iteration
(b) Iterations after gathering (T here was 1000 and the gathering happened
at t = 300)
Figure 9: Typical "core" center and extremal agent location vs.
beginning (a) and after the gathering (b). Simulations with N = 21, ε = 0.1.
time. From the
26
t
e
s
ff
O
t
e
s
ff
O
4
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
0
0
0
0
1
4
3
2
1
0
−1
−2
−3
−4
−5
−6
−7
−8
−9
−10
0
0
0
0
1
Core-Center
Left Extremal Point
Right Extremal Point
5
0
0
0
1
0
1
0
0
1
5
1
0
0
1
0
2
0
0
1
5
2
0
0
1
0
3
0
0
1
5
3
0
0
1
0
4
0
0
1
5
4
0
0
1
0
5
0
0
1
5
5
0
0
1
0
6
0
0
1
5
6
0
0
1
0
7
0
0
1
5
7
0
0
1
0
8
0
0
1
5
8
0
0
1
0
9
0
0
1
5
9
0
0
1
0
0
1
0
1
Iteration
(a) "Core" and extremal agents (zoomed). N = 21
Core-Center
Left Extremal Point
Right Extremal Point
5
0
0
0
1
0
1
0
0
1
5
1
0
0
1
0
2
0
0
1
5
2
0
0
1
0
3
0
0
1
5
3
0
0
1
0
4
0
0
1
5
4
0
0
1
0
5
0
0
1
5
5
0
0
1
0
6
0
0
1
5
6
0
0
1
0
7
0
0
1
5
7
0
0
1
0
8
0
0
1
5
8
0
0
1
0
9
0
0
1
5
9
0
0
1
0
0
1
0
1
Iteration
(b) "Core" and extremal agents (zoomed). N = 121
Figure 10: "Core" center location and both extremal agents vs. time (after gathering,
starting from T = 10, 000). Simulations with ε = 0.3. Note that the "core" of
gathered agents is much more easily moved by "extremists" when the population is
small (N = 21)
27
−2
−4
−6
−8
−10
t
e
s
ff
O
−12
−14
−16
−18
0
0
0
0
0
0
1
Core-Center
Left Extremal Point
Right Extremal Point
0
0
0
0
5
0
1
0
0
0
0
0
1
1
0
0
0
0
5
1
1
0
0
0
0
0
2
1
0
0
0
0
5
2
1
0
0
0
0
0
3
1
0
0
0
0
5
3
1
0
0
0
0
0
4
1
0
0
0
0
5
4
1
0
0
0
0
0
5
1
0
0
0
0
5
5
1
0
0
0
0
0
6
1
0
0
0
0
5
6
1
0
0
0
0
0
7
1
0
0
0
0
5
7
1
0
0
0
0
0
8
1
0
0
0
0
5
8
1
0
0
0
0
0
9
1
0
0
0
0
5
9
1
0
0
0
0
0
0
2
Iteration
(a) Evolution of "core" and extremal agents (zoomed out). N = 200
t
e
s
ff
O
8
6
4
2
0
−2
−4
−6
−8
0
0
0
0
0
0
1
Core-Center
Left Extremal Point
Right Extremal Point
0
0
0
0
5
0
1
0
0
0
0
0
1
1
0
0
0
0
5
1
1
0
0
0
0
0
2
1
0
0
0
0
5
2
1
0
0
0
0
0
3
1
0
0
0
0
5
3
1
0
0
0
0
0
4
1
0
0
0
0
5
4
1
0
0
0
0
0
5
1
0
0
0
0
5
5
1
0
0
0
0
0
6
1
0
0
0
0
5
6
1
0
0
0
0
0
7
1
0
0
0
0
5
7
1
0
0
0
0
0
8
1
0
0
0
0
5
8
1
0
0
0
0
0
9
1
0
0
0
0
5
9
1
0
0
0
0
0
0
2
Iteration
(b) Evolution of "core" and extremal agents (zoomed out). N = 1000
Figure 11: Typical evolution of "core"'s center of mass location and extremal agents'
locations in time (after gathering, starting from T = 1, 000, 000. Simulations done for
ε = 0.1). Note that in both cases the initial center of mass of all agents was at 0. Note
that the "inertia" of society is much higher when N = 1000 than in case N = 200.
28
6 Concluding remarks
We here proposed a mathematical model of randomly interacting particles on the
line, that could describe opinion dynamics in a society of presumably intelligent
agents. Equipped with a simple decision rule, agents eventually get together
to a drifting gathered constellation, in finite expected time. All the agents
of the system, except two, constitute a core of moderate agents that remain
closely clustered from that point on. The two "erratic extremist" agents perform
random walks biased toward the "quasi-stationary" core; once in a while the
roles of extremal agents change, when an "erratic extremist" walker joins the
core. We have derived expression for the expected convergence time and the
distribution of distances of extremal agents. Computer simulations support our
findings.
We believe that the model presented will further help analyze two and higher
dimensional models which have a practical importance in a number of areas in
multi-agent studies.
An interesting two-dimensional model corresponding to the random evolu-
tion process analyzed in this paper could be the following. Assume that the
agents' locations are points in the plane R2. For a group of N agents in the
plane the "extremists" are the ones that define the convex hull of the points.
Suppose at each time instant an agent that realizes it is an extreme vertex of the
convex hull (by sensing the bearing only to all other agents!) decides to move
a unit distance along the bisector of the corresponding convex hull angle either
toward the other agents (i.e. into the convex hull), with probability (1 − ε), or
in the opposite direction, with probability ε. (see Figure 12)
Extremal points
Internal points
(1 − ε)
α/2
α/2
1
P
ε
P
Figure 12: Group of Agents, Convex Hull and zoom on Extremal Agent move-
ment options
29
Preliminary simulations with this model show that indeed the population
gathers to a small region in the plane (see Figure 13) and the gathered group
performs a random walk in the plane (Figures 14 and 15). We plan to study
this and several variations of such models in the near future.
(a) At the beginning
(b) After 100 iterations
(c) After 200 iterations
(d) After 400 iterations
Figure 13: Typical evolution of the system (N = 400, ε = 0.1) from (a) beginning till
(d) 400th iteration. (Convex Hull is depicted for convenience)
30
01020304050607080901000102030405060708090100xy01020304050607080901000102030405060708090100xy01020304050607080901000102030405060708090100xy01020304050607080901000102030405060708090100xy (a) After 400 iterations
(b) After 1, 000 iterations
Figure 14: Evolution of the Center of Mass of the system and the last Convex Hull
after (a) 400 and (b) 1000 iterations
31
4647484950515253545556575844464850525456xy4647484950515253545556575844464850525456xy (a) The X direction
(b) The Y direction
Figure 15: Evolution of the Center of Mass of the system split by (a) X direction and
by (b) Y direction
32
01,0002,0003,0004,0005,0006,0007,0008,0009,00010,0000102030405060708090100iterationxcoordinateCenterofMassXcoordinateMinXcoordinateMaxXcoordinate01,0002,0003,0004,0005,0006,0007,0008,0009,00010,000102030405060708090iterationycoordinateCenterofMassYcoordinateMinYcoordinateMaxYcoordinate References
[1] Ando, H., Y. Oasa, I. Suzuki, and M. Yamashita (1999). Distributed mem-
oryless point convergence algorithm for mobile robots with limited visibility.
IEEE Transactions on Robotics and Automation 15 (5), 818 -- 828.
[2] Barel, A., R. Manor, and A. M. Bruckstein (2016, July). COME TO-
GETHER: Multi-Agent Geometric Consensus (Gathering, Rendezvous, Clus-
tering, Aggregation). Technical Report CIS-2016-03, CS Department Tech-
nion IIT.
[3] Camazine, S., J.-L. Deneubourg, N. R. Franks, J. Sneyd, G. Theraulaz,
and E. Bonabeau (2001, August). Self-Organization in Biological Systems.
Princeton University Press.
[4] Cartwright, D. and F. Harary (1956). Structural balance: a generalization
of heider's theory. Psychological Review 63 (5), 277.
[5] Chen, B., K. Jamieson, H. Balakrishnan, and R. Morris (2002). Span: An
energy-efficient coordination algorithm for topology maintenance in ad hoc
wireless networks. Wireless Networks 8 (5), 481 -- 494.
[6] Chong, C.-Y. and S. P. Kumar (2003). Sensor networks: evolution, oppor-
tunities, and challenges. Proceedings of the IEEE 91 (8), 1247 -- 1256.
[7] Couzin, I. and J. Krause (2003). Self-organization and collective behavior
in invertebrates. Advances in the Study of Behavior 32, 1 -- 75.
[8] DeGroot, M. H. (1974). Reaching a consensus. Journal of the American
Statistical Association 69 (345), 118 -- 121.
[9] Dudek, G., M. Jenkin, E. Milios, and D. Wilkes (1993). A taxonomy for
swarm robots. In Intelligent Robots and Systems' 93, IROS'93. Proceedings
of the 1993 IEEE/RSJ International Conference on, Volume 1, pp. 441 -- 447.
IEEE.
[10] Festinger, L. (1954). A theory of social comparison processes. Human
relations 7 (2), 117 -- 140.
[11] Festinger, L. (1962). A theory of cognitive dissonance, Volume 2. Stanford
University Press.
[12] Flocchini, P., G. Prencipe, and N. Santoro (2012). Distributed comput-
ing by oblivious mobile robots. Synthesis lectures on distributed computing
theory 3 (2), 1 -- 185.
[13] Friedkin, N. E. and E. C. Johnsen (1990). Social influence and opinions.
Journal of Mathematical Sociology 15 (3-4), 193 -- 206.
[14] Halpern, J. Y. and Y. Moses (1990). Knowledge and common knowledge
in a distributed environment. Journal of the ACM (JACM) 37 (3), 549 -- 587.
[15] Hegselmann, R. and U. Krause (2002). Opinion dynamics and bounded
confidence models, analysis, and simulation. Journal of Artificial Societies
and Social Simulation 5 (3).
33
[16] Hilton, P. and J. Pedersen (1991). Catalan numbers, their generalization,
and their uses. The Mathematical Intelligencer 13 (2), 64 -- 75.
[17] Jadbabaie, A., J. Lin, and A. S. Morse (2003). Coordination of groups of
mobile autonomous agents using nearest neighbor rules. IEEE Transactions
on Automatic Control 48 (6), 988 -- 1001.
[18] Krishna, P., N. H. Vaidya, M. Chatterjee, and D. K. Pradhan (1997). A
cluster-based approach for routing in dynamic networks. ACM SIGCOMM
Computer Communication Review 27 (2), 49 -- 64.
[19] Lorenz, J. (2017). Modeling the evolution of ideological landscapes through
In Advances in Social Simulation 2015, pp. 255 -- 266.
opinion dynamics.
Springer.
[20] Okubo, A. (1986). Dynamical aspects of animal grouping: Swarms, schools,
flocks, and herds. Advances in Biophysics 22, 1 -- 94.
[21] Olfati-Saber, R. and R. M. Murray (2004). Consensus problems in networks
of agents with switching topology and time-delays. IEEE Transactions on
automatic control 49 (9), 1520 -- 1533.
[22] Reynolds, C. W. (1987). Flocks, herds and schools: A distributed behav-
ioral model. ACM SIGGRAPH computer graphics 21 (4), 25 -- 34.
[23] S¸ahin, E. (2004). Swarm robotics: From sources of inspiration to domains
In International Workshop on Swarm Robotics, pp. 10 -- 20.
of application.
Springer.
[24] Shoham, Y. and M. Tennenholtz (1995). On social laws for artificial agent
societies: off-line design. Artificial intelligence 73 (1-2), 231 -- 252.
[25] Stanley, R. P. (2015). Catalan numbers. Cambridge University Press.
[26] Sumpter, D. J. (2006). The principles of collective animal behaviour.
Philosophical Transactions of the Royal Society of London B: Biological Sci-
ences 361 (1465), 5 -- 22.
[27] Vicsek, T., A. Czir´ok, E. Ben-Jacob, I. Cohen, and O. Shochet (1995).
Novel type of phase transition in a system of self-driven particles. Physical
Review Letters 75 (6), 1226.
34
|
1711.08151 | 1 | 1711 | 2017-11-22T06:43:21 | Multiagent Simple Temporal Problem: The Arc-Consistency Approach | [
"cs.MA",
"cs.AI"
] | The Simple Temporal Problem (STP) is a fundamental temporal reasoning problem and has recently been extended to the Multiagent Simple Temporal Problem (MaSTP). In this paper we present a novel approach that is based on enforcing arc-consistency (AC) on the input (multiagent) simple temporal network. We show that the AC-based approach is sufficient for solving both the STP and MaSTP and provide efficient algorithms for them. As our AC-based approach does not impose new constraints between agents, it does not violate the privacy of the agents and is superior to the state-of-the-art approach to MaSTP. Empirical evaluations on diverse benchmark datasets also show that our AC-based algorithms for STP and MaSTP are significantly more efficient than existing approaches. | cs.MA | cs | Multiagent Simple Temporal Problem: The Arc-Consistency Approach
Shufeng Kong1 and Jae Hee Lee1 and Sanjiang Li1,2
1Centre for Quantum Software and Information, FEIT, University of Technology Sydney, Australia
2UTS-AMSS Joint Research Laboratory, AMSS, Chinese Academy of Sciences, China
[email protected], {jaehee.lee, sanjiang.li}@uts.edu.au
7
1
0
2
v
o
N
2
2
]
A
M
.
s
c
[
1
v
1
5
1
8
0
.
1
1
7
1
:
v
i
X
r
a
Abstract
The Simple Temporal Problem (STP) is a fundamental tem-
poral reasoning problem and has recently been extended to
the Multiagent Simple Temporal Problem (MaSTP). In this
paper we present a novel approach that is based on enforcing
arc-consistency (AC) on the input (multiagent) simple tem-
poral network. We show that the AC-based approach is suf-
ficient for solving both the STP and MaSTP and provide ef-
ficient algorithms for them. As our AC-based approach does
not impose new constraints between agents, it does not vio-
late the privacy of the agents and is superior to the state-of-
the-art approach to MaSTP. Empirical evaluations on diverse
benchmark datasets also show that our AC-based algorithms
for STP and MaSTP are significantly more efficient than ex-
isting approaches.
1
Introduction
The Simple Temporal Problem (STP) (Dechter, Meiri, and
Pearl 1991) is arguably the most well-known quantitative
temporal representation framework in AI. The STP consid-
ers time points as the variables and represents temporal in-
formation by a set of unary or binary constraints, each spec-
ifying an interval on the real line. Since its introduction in
1991, the STP has become an essential sub-problem in plan-
ning or scheduling problem (Bart´ak, Morris, and Venable
2014).
While the STP is initially introduced for a single schedul-
ing agent and is solved by centralized algorithms, many real-
world applications involve multiple agents who interact with
each other to find a solution like the following example:
Example 1. When Alice is looking for a position at com-
pany X, she might need to arrange an interview appointment
with X. Suppose that her colleague Bob is also applying for
the position and Alice and Bob are both applying for another
position at company Y . To represent and solve such an inter-
view scheduling problem, we need a multiagent framework
(see Figure 1 for an illustration).
Recently,
the extension of STP to multiagent STP
(MaSTP) has been provided in (Boerkoel and Durfee 2013),
which presents a formal definition of the MaSTP as well as
Copyright © 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
Alice
X
Y
Bob
Figure 1: An illustration of Example 1. Alice, Bob, company
X, company Y are four agents, each owning a local simple
temporal network. The circles represent variables and edges
constraints. Red edges represent constraints that are shared
by two different agents.
a distributed algorithm, called D(cid:52)PPC, for computing the
complete joint solution space.
However, as D(cid:52)PPC is based on the P3C algo-
rithm (Planken, de Weerdt, and van der Krogt 2008), which
triangulates the input constraint graph, it has the drawback
of creating new constraints between agents that are possibly
not directly connected. In Figure 1, D(cid:52)PPC triangulates the
inner cycle by adding at least one new constraint either be-
tween X and Y or between Alice and Bob. Neither of these
new constraints are desirable, as they introduce constraints
between two previously not directly connected agents and
thus present a threat to the privacy of the relevant agents.
As the recent technological advancements have allowed
for solving larger problems that are highly interwoven and
dependent on each other, efficiency and privacy have be-
come critical requirements. To address this challenge, we
propose a new approach to solve the MaSTP, which is based
on arc-consistency.
A constraint R between two variables x, y is called arc-
consistent (AC), if for every value dx from the domain of x
there is a value dy in the domain of y such that (dx, dy) ∈ R.
While AC is an important tool for solving finite (multi-
agent) constraint satisfaction problems (CSPs) (Montanari
1974; Baudot and Deville 1997; Nguyen and Deville 1998;
Hamadi 2002) at first glance it is not clear how it can be
applied to solving CSPs with real domains such as the STP,
because either the existing AC algorithms are fine-grained
and work with each single element of a domain to enforce
AC, which is impossible for real domains, or they are coarse-
grained, but cannot guarantee their termination, as real do-
mains can be infinitely refined when constraints are propa-
gated.1
Our contributions in this paper are as follows.
• We provide the first AC-based approach for solving STP
and analyze its computational complexity (Section 3).
• We provide the first AC-based approach for solving multi-
agent STP, which preserves the privacy of the agents, and
analyze its computational complexity (Section 4).
• We experimentally show that both our centralized and dis-
tributed algorithms outperform their existing counterparts
for solving STP (Section 5).
The next section gives a formal introduction to STP.
2 The Simple Temporal Problem
This section briefly introduces the STP. Details can be found
in (Dechter, Meiri, and Pearl 1991).
The simple temporal problem (STP) is a constraint satis-
faction problem where each constraint is a set of linear in-
equalities of the form
avw ≤ w − v ≤ bvw,
(1)
where avw, bvw are constants and v, w are variables defined
on a continuous domain representing time points. The con-
straint in (1) is abbreviated as Ivw = [avw, bvw]. As (1)
is equivalent to −bvw ≤ v − w ≤ −avw, we also obtain
vw = [−bvw,−avw]. The domain of each variable
Iwv = I−1
v is an interval Iv = [av, bv], where Iv could be a single-
ton or empty. Assume that o is a special auxiliary variable
that represents the fixed zero temporal point. Then the do-
main Iv can also be regarded as a constraint from o to v and
Iv = [av, bv] = [aov, bov] = Iov.
Algebraic operations on STP constraints are defined as
follows. The intersection of two STP constraints defined on
variables v, w yields a new constraint over v, w that rep-
resents the conjunction of the constraints. It is defined as
Ivw ∩ I(cid:48)
vw := [max{avw, a(cid:48)
vw}, min{bvw, b(cid:48)
The composition of an STP constraint Ivu over variables
v, u and another STP constraint Iuw over u, w yields a new
STP constraint over v, w that is inferred from the other two
constraints and is defined as Ivu⊗ Iuw := [avu + auw, bvu +
buw]. Here we require that [a, b] ⊗ ∅ = ∅ for any a ≤ b.
Remark 1. For STP constraints, the composition and inter-
section are associative and, as noted in (Dechter, Meiri, and
Pearl 1991), composition distributes over non-empty inter-
section for intervals, i.e., I ⊗ (J ∩ K) = (I ⊗ J) ∩ (I ⊗ K)
for any three intervals I, J, K such that J ∩ K (cid:54)= ∅.
vw}].
1Dechter, Meiri, and Pearl (1991) for example, suggest dis-
cretizing the domains to overcome this issue, in which case the
total number of contraint propagations would depend on the sizes
of the domains. The performance of our AC algorithm for (multia-
gent) STP does not depend on the sizes of the domains.
Definition 1. An instance of STP is called a simple temporal
network (STN) and is a tuple (cid:104)V, D, C(cid:105), where V is a finite
set of variables, D = {Iv v ∈ V } is a set of intervals, and
C is a set of STP constraints defined on V .
We assume that all variables in V appear in C and at most
one constraint exists between any pair of variables v and
w. Moreover, if Ivw = [a, b] is the constraint in C from v
to w, we always assume that the constraint Iwv = I−1
vw =
[−b,−a] is also in C. As previously mentioned, the domain
Iv of each variable v can be regarded as either a unary con-
straint, or a binary constraint Iov = Iv, where o is a fixed
variable representing the zero time point.
An STN naturally induces a graph in the following sense.
Definition 2. The constraint graph GN = (V, E) of an STN
N = (cid:104)V, D, C(cid:105) is an undirected graph, where the set E of
edges consists of constrained unordered pairs of variables in
C, i.e.,
E = {{v, w} v, w ∈ V, v (cid:54)= w, Ivw ∈ C}.
(2)
π = Iu0,u1 ⊗ Iu1,u2 ⊗ ... ⊗ Iuk−1,uk
Let GN = (V, E) be the constraint graph of an STN N .
We can use a labelled directed graph to illustrate N , where
for any undirected edge {v, w} ∈ E there is exactly one
directed edge (v, w) that is labelled with the corresponding
interval [avw, bvw].
A path π from v to w in GN is a sequence of variables
u0, u1, ..., uk such that v = u0, w = uk, and {us, us+1}
is an edge in E for each s = 0, . . . , k − 1 (k is called the
length of π). We write(cid:78) π for the composition of all these
Ius,us+1, i.e.,(cid:79)
[a, b] =(cid:78) π. We call π a negative cycle if b < 0.
If v = w, then we call π a cycle at v. For a cycle π, let
Definition 3. A solution of an STN N = (cid:104)V, D, C(cid:105) is an
assignment, that assigns to each variable v ∈ V a time point
from Iv ∈ D such that all constraints in C are satisfied. N
is said to be consistent if N has a solution. Two STNs are
said to be equivalent if they have the same solution set.
Definition 4 (Minimality). Let N = (cid:104)V, D, C(cid:105) be a con-
sistent STN and let v and w be variables in V . A constraint
Ivw from v to w is said to be minimal if every assignment
that assigns time points from domains Iv and Iw to v and w,
respectively, and satisfies Ivw can be extended to a solution
of N . A domain Iv of v ∈ V is said to be minimal if every
assignment of a time point from Iv to v can be extended to a
solution of N . We say N is minimal if every constraint in C
as well as every domain in D is minimal (note that, since we
regard domains as constraints between the zero time point o
and variables, we also require the domains to be minimal).
3 Solving the STP with Arc-Consistency
In this section we show that enforcing arc-consistency is suf-
ficient to solve the STP.
Definition 5. Let N = (cid:104)V, D, C(cid:105) be an STN. Suppose v
and w are two variables in V , Iv and Iw are, respectively,
their domains, and Ivw is a constraint in C from v to w.
An STN N (cid:48) = (cid:104)V, D(cid:48), C(cid:105) with D(cid:48) = {I(cid:48)
We say that Ivw is arc-consistent (AC) (relative to Iv and
Iw) if for any tv ∈ Iv there exists some tw ∈ Iw such that
tw − tv ∈ Ivw, i.e., avw ≤ tw − tv ≤ bvw. We say that
N is AC if both Ivw and Iwv are AC for every constraint
Ivw ∈ C.
v v ∈ V }
is called the AC-closure of N , if N (cid:48) is the largest arc-
consistent STN which is equivalent to N , in the sense that
for every other arc-consistent STN N (cid:48)(cid:48) = (cid:104)V, D(cid:48)(cid:48), C(cid:105) with
v for all v ∈ V .
D(cid:48)(cid:48) = {I(cid:48)(cid:48)
Lemma 1. Let N = (cid:104)V, D, C(cid:105) be an STN and v, w ∈ V
two variables that are constrained by Ivw in C. Then Ivw is
arc-consistent relative to Iv and Iw iff Iv ⊆ Iw ⊗ Iwv.
Proof. It suffices to show that
v ∈ V }, we have that I(cid:48)(cid:48)
v ⊆ I(cid:48)
i
Iw ⊗ Iwv = {x ∈ R ∃y ∈ Iw s.t. y − x ∈ Ivw}
(3)
Let Iv = [a, b], Iw = [c, d] and Ivw = [e, f ]. Then
{x ∈ R ∃y ∈ Iw s.t. y − x ∈ Ivw}
= {x ∈ R ∃c ≤ y ≤ d s.t. e ≤ y − x ≤ f}
= {x ∈ R ∃c ≤ y ≤ d s.t. y − f ≤ x ≤ y − e}
= {x ∈ R c − f ≤ x ≤ d − e}
= [c, f ] ⊗ [−f,−e] = Iw ⊗ Iwv,
which proves Eq. (3).
Lemma 2. Let N = (cid:104)V, D, C(cid:105) be an arc-consistent STN
and v, w ∈ V two variables that are constrained by Ivw in
C. Then Iv ⊆ Iw ⊗ Iwv.
Proof. This follows directly from Lemma 1 and that Ivw is
AC relative to Iv and Iw.
wv is the minimal constraint from w to v.
The following result directly follows from Lemma 2.
Let π be a path in N from w to v. Then Iv ⊆ Iw ⊗(cid:78) π.
Corollary 3. Let N = (cid:104)V, D, C(cid:105) be an arc-consistent STN.
Lemma 4. Let N = (cid:104)V, D, C(cid:105) be an arc-consistent STN
and v, w variables in V . If N is consistent, then Iv ⊆ Iw ⊗
wv, where I m
I m
Proof. Since N is consistent, I m
wv is nonempty. Recall that
wv is the intersection of the compositions along all paths
I m
in N from w to v (cf. (Dechter, Meiri, and Pearl 1991, §3))
and composition distributes over non-empty intersection for
intervals. The result follows directly from Corollary 3.
Lemma 5 ((Shostak 1981)). Suppose N = (cid:104)V, D, C(cid:105) is an
STN. Then N is inconsistent if and only if there exists a neg-
ative cycle.
Lemma 6. Given a consistent STN N = (cid:104)V, D, C(cid:105) with
n = V , for any path π of length ≥ n there is a path π(cid:48) of
length < n such that(cid:78) π(cid:48) ⊆(cid:78) π.
(cid:78) π(cid:48) ⊆ (cid:78) π. Repeating this procedure until there is no
Proof. Since the length of π is ≥ n, π must have a cycle
at a variable v. As the cycle is not negative, removing the
cycle and leaving only v in the path results in a path π(cid:48) with
cycle gives the desired result.
Lemma 7. Let N = (cid:104)V, D, C(cid:105) be an STN and N (cid:48) its AC-
closure. Then N is consistent iff N (cid:48) has no empty domain.
Proof. We prove N is inconsistent iff N (cid:48) has an empty do-
main. As N and N (cid:48) are equivalent, if N (cid:48) has an empty do-
main, then N is inconsistent.
Now suppose N is inconsistent. Then by Lemma 5, there
[l, h] with h < 0. Now let v be a variable in N with Iwv =
[e, f ] and let I(cid:48)
w = [c, d] be the domains of v and
w in N (cid:48), respectively. Choose k ∈ N sufficiently large, such
that kh < b − d − f. Then, by Lemma 4 we have
exists a negative cycle π in N at some w such that(cid:78) π =
v = [a, b], I(cid:48)
w ⊗(cid:16)(cid:79)
πk ⊗ Iwv
v ⊆ I(cid:48)
I(cid:48)
= [c, d] ⊗ ([kl, kh] ⊗ [e, f ])
= [c + kl + e, d + kh + f ],
(cid:17)
(4)
v is empty.
where πk is the concatenation of k copies of path π. Because
kh < b − d − f, (4) is possible only if I(cid:48)
Theorem 8. Let N = (cid:104)V, D, C(cid:105) be a consistent STN and
N (cid:48) its AC-closure. Then all domains in N (cid:48) are minimal.
Proof. If the constraint graph GN is connected, i.e., for any
two variables v, w, there is a path in GN that connects v
to w, then we may replace the constraint from v to w with
vw, if there
the nonempty minimal constraint I m
was no constraint between v and w). We write the refined
STN as N ∗. For any two variables v, w, by Lemma 4, Iv is
contained in Iw ⊗ I m
vw. This
shows that N ∗ is the same as the minimal STN of N , and
thus, establishes the minimality of each Iv.
In case the constraint graph is disconnected, we consider
the restriction of N to its connected components instead.
The same result applies.
wv and Iw is contained in Iv ⊗ I m
vw (or add I m
Two special solutions can be constructed if N is arc-
consistent and has no empty domain.
Proposition 9. Let N = (cid:104)V, D, C(cid:105) be an arc-consistent
STN with D = {Iv v ∈ V } and Iv = [av, bv] for each
v. If no Iv is empty, then the assignments A = {av v ∈ V }
and B = {bv v ∈ V } are two solutions of N .
Proof. Let N (cid:48) = (cid:104)V, D(cid:48), C(cid:48)(cid:105) be the minimal STN of N . By
Theorem 8, we have D(cid:48) = D and N (cid:48) is equivalent to N . The
above claim follows as the assignments A = {av v ∈ V }
and B = {bv v ∈ V } are two solutions of the minimal STN
N (cid:48) (cf. (Dechter 2003, Corollary 3.2)).
Theorem 10. Enforcing AC is sufficient to solve STP.
Proof. Let N be an STN and N (cid:48) its AC-closure. If N (cid:48) has
an empty domain, then N has no solution by Lemma 7. If
N (cid:48) does not have an empty domain, then we can use Propo-
sition 9 to find a solution.
Remark 2. (i) As solving an STN is equivalent to solving a
system of linear inequalities, the solution set of an STN is a
convex polyhedron. Thus any convex combination of the two
solutions A and B is again a solution of the STN. (ii) Enforc-
ing AC can in essence find all solutions of an STN: Suppose
Algorithm 1: ACSTP
Input
: An STN N = (cid:104)V, D, C(cid:105) and its constraint
graph G = (V, E), where V = n.
Output: An equivalent network that is AC, or
"inconsistent".
foreach v ∈ V do
1 Q ← ∅
2 for k ← 1 to n do
3
v ← Iv
I(cid:48)
4
foreach w ∈ V s.t. {v, w} ∈ E do
5
6
Iv ← Iv ∩ Iw ⊗ Iwv
v = Iv then Q ← Q ∪ {v}
if Iv = ∅ then return "inconsistent"
if I(cid:48)
else Q ← Q \ {v}
7
8
9
if #Q = n then return N
10
11 return "inconsistent"
N is arc-consistent and has no empty domain. We pick an
arbitrary variable v that has not been instantiated yet, then
assign any value from Dv to v, and enforce AC on the re-
sulting network. We repeat this process until all variables are
instantiated. (iii) Proposition 9 can also be obtained by first
showing that STP constraints are both max/min-closed, and
then using the result in (Jeavons and Cooper 1995, Thm 4.2),
which states that the AC-closure of a constraint network over
max/min-closed constraints have the maximal and the min-
imal values of the domains as two solutions. As a conse-
quence of this, Theorem 8 can also be obtained, because the
solution set of an STN is convex (cf. Remark 2 (i)).
A Centralized AC Algorithm for the STP
In this section we propose an AC algorithm, called ACSTP,
to solve STNs. The algorithm is presented as Algorithm 1.
Theorem 11. Given an input STN N , Algorithm 1 returns
"inconsistent" if N is inconsistent. Otherwise, it returns the
AC-closure of N .
Proof. We first note that intersection and composition of
constraints do not change the solution set of the input STN
N . This has two implications: First, if a domain Iv becomes
empty during the process of the algorithm, then the solution
set of N is empty and N is inconsistent. Second, if the al-
gorithm terminates and its output N (cid:48) is AC, then N (cid:48) is the
AC-closure of N . Consequently, it suffices to show that if
the algorithm terminates and returns N (cid:48), then N (cid:48) is AC.
We first consider the case, where the algorithm returns
N (cid:48) in line 10 at the kth iteration of the for-loop (lines 2–10)
for some 1 ≤ k ≤ n. We show that N (cid:48) is AC. Let I k
v be
the domain of v obtained after the kth iteration of the for-
loop. Due to lines 6 and 8, we have for all {v, w} ∈ E that
v ⊆ I k−1
w. Thus we have
I k
for {v, w} ∈ E that I k
w ⊗ Iwv, which is by Lemma 1
equivalent to saying that Ivw is AC w.r.t. domains I k
v and
w. Hence, the output N (cid:48) is AC.
I k
w ⊗ Iwv) and I k−1
∩ (I k−1
v ⊆ I k
w = I k
v
Now suppose that the algorithm exited in line 11 returning
"inconsistent". Thus, at the nth iteration of the for-loop we
have #Q < n in line 10. We prove that N is inconsistent by
contradiction. Assume that N is consistent. For any v ∈ V
and any k ≥ 1, we write Πk
v for the set of paths from o (the
auxiliary variable denoting the zero time point) to v with
length ≤ k in the constraint graph of N . We claim
(5)
v
π∈Πk
v being the minimal domain of
(cid:79)
π
(cid:79)
I k−1
v ⊆ (cid:92)
v ⊆ (cid:92)
π∈Πn
v
for any k ≥ 1. Then, with I m
v, we have
v ⊆ I n−1
I m
π = I m
v ,
v
= I m
v = I n−1
because I m
v is the intersection of the compositions along all
paths in N from o to v (cf. (Dechter, Meiri, and Pearl 1991,
§3)), where it suffices to only build compositions along paths
of length ≤ n by Lemma 6. Thus I n
v for all
v ∈ V , which is a contradiction to our assumption that at the
nth iteration of the for-loop we have #Q < n in line 10.
We now prove (5) by using induction on k. First, for k =
v contains only one path of length 1 (i.e., the edge
1, since Π1
{o, v}), we have I 0
is true for k − 2 for all w ∈ V . Then by line 6 and our
induction hypothesis we have
(cid:78) π. Now suppose (5)
(cid:33)
w ⊗ Iwv
I k−2
(cid:79)
v ⊆ I k−2
I k−1
π∈Π1
w
v
v
⊆ I k−2
∩
v = Iv =(cid:84)
(cid:32)(cid:92)
(cid:92)
(cid:92)
∩
(cid:92)
(cid:79)
(cid:92)
(cid:79)
π∈Πk−1
∩
π
w
v
v
π,
π∈Πk−1
w
⊗ Iwv
(cid:92)
(cid:79)
π
π
(π∈Πk
v )∧(π≥2)
⊆
=
π∈Πk
v
which proves (5).
Theorem 12. Algorithm 1 runs in time O(en), where e is
the number of edges of the constraint graph of the input STN
and n is the number of variables.
Proof. There are at most n iterations of the for-loop and
each iteration involves O(e) operations.
Remark 3. Algorithm 1 can also be understood as comput-
ing the shortest path from a source vertex o to every other
vertex v and the shortest path from every other vertex v to
the source vertex o. This can be realized in time O(en) by
using a shortest path tree algorithm with negative cycle de-
tection (cf. (Tarjan 1983, Section 7.2) and (Korte and Vygen
2012, Section 7.1).
4 Solving the MaSTP with Arc-Consistency
In this section we extend ACSTP to a distributed algorithm
DisACSTP to solve multiagent simple temporal networks
(MaSTNs).
Definition 6. (Boerkoel and Durfee 2013) A multiagent
simple temporal network (MaSTN) is a tuple M = (cid:104)P, CX(cid:105),
where
• P = {Ni i = 1, . . . , p} is a set of local STNs, where
each Ni = (cid:104)Vi, Di, Ci(cid:105) is an STN belonging to agent i
and we require that Vi ∩ Vj = ∅ for any two different
agents i, j = 1, . . . , p.
• CX is a set of external constraints, where each constraint
is over two variables belonging to two different agents.
Constraint graphs for MaSTNs can be defined analogously
as that for STNs, where we use EX for the set of edges cor-
responding to constraints in C X. See Figure 1 for an illus-
tration. In Figure 1, the edges in EX are represented as red
lines.
Definition 7. Suppose M = (cid:104)P, CX(cid:105) is an MaSTN. Let
Ivw ∈ CX with v ∈ Vi, w ∈ Vj be an external constraint. We
say that Ivw is an external constraint of agent i, and write CX
i
for the set of external constraints of agent i. We call v and
w a shared and an external variable of agent i, respectively.
We write V X
for the set of external variables of agent i. In
i
Figure 1, the vertices for shared variables are represented as
red circles.
DisACSTP is presented in Algorithm 2. In DisACSTP
each agent i gets as input its portion Ni of the input MaSTN
M and the set CX
i of its external constraints, and runs its
own algorithm. Similar to ACSTP, DisACSTP updates the
domains of Ni at each iteration of the for-loop and main-
tains a queue Qi to record the information about the un-
changed domains. When a domain becomes empty during
the updates, then the agent can terminate the algorithm and
conclude that the input MaSTN M is inconsistent. There
are however aspects in DisACSTP that are different from
ACSTP, which stem from the fact that in MaSTP an agent
cannot have the global knowledge of the states of other
agents' processes without sharing certain information with
other agents. These aspects are the following:
1. The total number n of the variables in the input MaSTN
is initially not known to individual agents. This, however,
can easily be determined using an echo algorithm (Chang
1982). We can therefore regard n as given as an input to
DisACSTP.
2. As the agents may run their processes at different paces, at
each iteration of the for-loop (lines 2–34), they synchro-
nize the domains of their external variables (lines 3–4).
Otherwise, some agents might use stale external domains
and make wrong conclusions.
3. When a domain becomes
running
this
DisACSTP, an agent broadcasts
information to other agents so that they can terminate
their algorithms as soon as possible.
empty while
(lines 9–11)
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35 return "inconsistent"
Algorithm 2: DisACSTP
Input
: Ni: agent i's portion of MaSTN M;
V X
i : the set of agent i's external variables;
CX
i : the set of agent i's external constraints;
parent(i): the parent of agent i w.r.t. T (M);
children(i): the children of agent i w.r.t.
T (M);
n: the number of variables of M.
"inconsistent".
Output: Agent i's portion of the AC-closure of M or
1 Qi ← ∅
2 for k ← 1 to n do
3
Send the domains of the shared variables to the
neighbors.
Receive the domains of the external variables from
the neighbors.
foreach v ∈ Vi do
i s.t. {v, w} ∈ Ei ∪ EX
i do
v ← Iv
I(cid:48)
foreach w ∈ Vi ∪ V X
Iv ← Iv ∩ Iw ⊗ Iwv
if Iv = ∅ then
Broadcast "inconsistent".
return "inconsistent"
v = Iv then Qi ← Qi ∪ {v}
if I(cid:48)
else Qi ← Qi \ {v}
if #Qi = #Vi then
if root(i) then
while true do
Send inquiry ("Are all Qi full?", k) to
children(i)
m ← RECEIVEMESSAGE()
if m is domains of external variables from a
neighbor then
break
if m is inquiry ("Are all Qi full?", k) then
if leaf(i) then
Send feedback ("yes", k) to
parent(i)
else Send m to children(i)
if m is feedback ("yes", k) then
if all feedbacks received from
children(i) then
if root(i) then
Broadcast "arc-consistent"
return Ni
else Send m to parent(i)
if m is "arc-consistent" then
return Ni
if m is "inconsistent" then
return "inconsistent"
4. If the queue Qi of an agent i is full (i.e., it contains all
of the agent's variables in Vi) after an iteration of the for-
loop, then the agent shares this information with all other
agents in M so as to jointly determine whether the queues
of all agents are full and the network is arc-consistent
(lines 15–16 and 21–30).
5. If the queue Q of an agent is not full after an iteration of
the for-loop, then the agent broadcasts this information to
all other agents, so that they can move to the next iteration
of the for-loop as soon as possible.
All the preceding aspects are subject to communication of
certain information between agents. DisACSTP coordinates
this communication while (i) preserving the privacy of each
agent and (ii) reducing the duration of any idle state of an
individual agent. Concretely:
• Each agent shares information only with the agents who
are connected through an external constraint. We call
them the neighbors of the agent. This neighborhood-
relationship among the agents induces a graph that we call
henceforth the agent graph.
• Each agent shares with its neighbors only the domains
of its shared variables. No other information is shared
(such as its network structure, constraints, private vari-
ables and their domains) and only the neighbors w.r.t. the
agent graph can share the information. This property is
a critical advantage over D(cid:52)PPC (Boerkoel and Durfee
2013), as D(cid:52)PPC often creates new external constraints
during the process and reveal more private information of
the agents than necessary.
• Each agent uses a broadcasting mechanism to share
global properties of the input MaSTN, i.e., an agent first
sends a message (e.g., "inconsistent") to its neighbors,
then the neighbors forward the message to their neigh-
bors and so on, until all agents receive the message. To
reduce the number of messages, duplicates are ignored by
the agents.
An agent i broadcasts the following messages: "arc-
consistent", "inconsistent" and "Qi is not full", where the
last message is indirectly broadcasted by agent i skipping
lines 14–34 and moving to the next iteration of the for-
loop and then sending its shared domains to its neighbors.
This initiates a chain reaction among the idle neighbors of
agent i who have not moved to the next iteration yet, as
they quit the idle states (lines 19–20) and move to the next
iteration of the for-loop and then send also their shared
domains to their idle neighbors (lines 3–4).
• There is a dedicated agent who checks at each iteration
of its for-loop (given its queue is full) whether the queues
of all other agents are full at the same iteration. This ded-
icated agent is determined by building a minimal span-
ning tree (e.g., by using an echo algorithm (Chang 1982))
T (M) of the agent graph. The agent who is the root
(henceforth the root agent) of this tree becomes then the
dedicated agent.
The root agent sends an inquiry to its children to
check whether the queues of all its descendants are full
(lines 15–16). The inquiry is then successively forwarded
by the descendants whose queues are full. We have to dis-
tinguish here between two cases:
(1) If all descendants' queues are full, then the inquiry
reaches all the leaf agents and returns back as feedbacks
(lines 22–23) until the root agent receives all the feed-
backs (lines 25–30) and broadcasts "arc-consistency".
(2) If a descendant's queue is not full, then the descendant
moves on to the next iteration of the for-loop and initi-
ates a chain reaction among other agents by sending the
domains of its shared variables to its neighbors (cf. the
second paragraph of the third bullet point).
Due to the properties so far considered, DisACSTP is
guaranteed to simulate the behavior of ACSTP while allow-
ing concurrent domain update operations.
Theorem 13. Let M = (cid:104)P, CX(cid:105) be an MaSTN. Let Nmax be
a network with emax = max(cid:8)ei + eX
ei and eX
i are the number of edges of the constraint graph
of Ni and the number of external constraints of agent i,
respectively. Then Algorithm 2 enforces AC on M in time
O(emaxn).
(cid:12)(cid:12) 1 ≤ i ≤ p(cid:9), where
i
5 Evaluation
In this section we experimentally compare our algorithms
against the state-of-the-art algorithms for solving STNs.
For centralized algorithms, we compare our ACSTP al-
gorithm against Planken, de Weerdt, and van der Krogt's
P3C algorithm (2008); for distributed algorithms, we com-
pare our DisACSTP algorithm against Boerkoel and Dur-
fee's D(cid:52)PPC algorithm (2013). All experiments for dis-
tributed algorithms used an asynchronous simulator in
which agents are simulated by processes which communi-
cate only through message passing and default communica-
tion latency is assumed to be zero. Our experiments were
implemented in Python 3.6 and carried out on a computer
with an Intel Core i5 processor with a 2.9 GHz frequency
per CPU, 8 GB memory 2.
As measures for comaring performances we use the num-
ber of constraint checks and the number of non-concurrent
constraint checks (NCCCs) performed by the centralized al-
gorithms and the distributed algorithms, respectively. Given
an STN N = (cid:104)V, D, C(cid:105), a constraint check is performed
when we compute relation r ← Ivw∩(Ivu⊗Iuw) and check
if r = Ivw or r (cid:54)⊆ Ivw.
ACSTP vs. P3C
Datasets We selected instances from the benchmark
datasets of STNs used in (Planken, de Weerdt, and van der
Krogt 2012) for evaluations. We considered the scale-free
graphs (Scale-free-1) with 1000 vertices and density pa-
rameter varying from 2 to 50. We also considered the scale-
free graphs (Scale-free-2) with varying vertex count. The
scale-free density parameter for this selection is 5. Be-
side these artificially constructed graphs, we also considered
2The source code for our evaluation can be found in https:
//github.com/sharingcodes/MaSTN
P3C
ACSTP
107
106
105
P3C
ACSTP
105
104
P3C
ACSTP
108
107
106
105
5 10
30
20
40
Network density
50
300 400 500 600 700 800
n
(b) Scale-free-2
400
600
n
800
1,024
(c) New York
(a) Scale-free-1
Figure 2: Evaluation of ACSTP and P3C. The y-axes (on the log scale) represent the number constraint checks.
107
106
105
104
107
106
105
D(cid:52)PPC
DisACSTP
105
104
D(cid:52)PPC
DisACSTP
106
105
104
D(cid:52)PPC
DisACSTP
D(cid:52)PPC
DisACSTP
2 4
8
12
16
Number of agents
0
200 400 600 800
Number of ext. constraints
2 4
8
12
16
Number of agents
80 160 240 320 400 480
Number of tasks
(a) DBH
(b) WS
Figure 3: Evaluation of DisACSTP and D(cid:52)PPC. The y-axis (on the log scale) represent the number of NCCCs.
graphs that are based on the road network of New York City
(New York). This dataset contains 170 graphs on 108–3906
vertices, 113–6422 edges.
Results The results are presented in Figure 2, where base-
10 log scales are used for the y-axes. For the scale-free
graphs we observe that ACSTP is 100–1000 times faster
than P3C. The dataset New York only contains very sparse
networks (each network's density is less than 1%), thus both
algorithms could easily solve these networks. However, we
still observe that ACSTP is about 5–12 times faster than
P3C.
DisACSTP vs. D(cid:52)PPC
Datasets We selected instances from the benchmark
datasets of MaSTNs used in (Boerkoel and Durfee 2013)
for evaluations. The first dataset BDH was randomly gener-
ated using the multiagent adaptation of Hunsberger's (2002)
random STN generator. Each MaSTN has N agents each
with start time points and end time points for 10 activi-
ties, which are subject to various local constraints. In ad-
dition, each MaSTN has X external contraints. We evalu-
ated the algorithms by varying the number of agents (N ∈
{2, 4, 8, 12, 16}, X = 50× (N − 1)) and the total number of
external constraints (N = 16, X ∈ {100, 200, 400, 800}).
The second dataset WS is derived from a multiagent fac-
tory scheduling domain (Wilcox and Shah 2012), where N
agents are working together to complete T tasks in a man-
ufacturing environment. We evaluated algorithms by vary-
ing the number of agents (N ∈ {2, 4, 8, 12, 16}, T =
20 × N ) and the total number of tasks (N = 16, T ∈
{80, 160, 240, 320, 400, 480}).
Results The results are presented in Figure 3, where base-
10 log scales are again used for the y-axes. For the DBH
random networks (Figure 3a) we observe that DisACSTP
is 5–30 times faster than D(cid:52)PPC. For the WS scheduling
networks (Figure 3b) DisACSTP is 2–10 times faster than
D(cid:52)PPC. For both datasets we observe that, with increasing
x-values, the y-values (i.e., NCCCs) for DisACSTP grow
slower than those for D(cid:52)PPC.
6 Conclusion
In this paper we presented a novel AC-based approach for
solving the STP and the MaSTP. We have shown that arc-
consistency is sufficient for solving an STN. Considering
that STNs are defined over infinite domains, this result is
rather surprising. Our empirical evaluations showed that the
AC-based algorithms are significantly more efficient than
their PC-based counterparts. This is mainly due to the fact
that PC-based algorithms add many redundant constraints
in the process of triangulation. More importantly, since our
AC-based approach does not impose new constraints be-
tween agents that are previously not directly connected, it
respects as much privacy of these agents as possible. We
should note here that even though our distributed algo-
rithm DisACSTP showed remarkable performance, it can be
further fine-tuned by using different termination detection
mechanisms (cf. (Mattern 1987) and (Raynal 2013, Ch. 14)).
It would be interesting to see how the result in this pa-
per can be used for solving the general disjunctive tempo-
ral problems (Stergiou and Koubarakis 2000). Potential ex-
tensions of our paper also include adapting our AC algo-
rithms to incremental algorithms for the STP (Planken, de
Weerdt, and Yorke-Smith 2010), dynamic situations (Mor-
ris, Muscettola, and Vidal 2001) and uncertainty (Venable
and Yorke-Smith 2005).
Acknowledgments
We thank the anonymous reviewers, who pointed out the
connections to max/min-closed constraints and the shortest-
path problem. The work of SL was partially supported by
NSFC (No. 11671244), and the work of JL was partially
supported by the Alexander von Humboldt Foundation.
References
[Bart´ak, Morris, and Venable 2014] Bart´ak, R.; Morris,
R. A.; and Venable, K. B.
2014. An Introduction to
Constraint-Based Temporal Reasoning. Synthesis Lectures
on Artificial Intelligence and Machine Learning 8(1):1–121.
[Baudot and Deville 1997] Baudot, B., and Deville, Y. 1997.
Analysis of distributed arc-consistency algorithms. Techni-
cal Report 97-07, Universit´e catholique de Louvain.
[Boerkoel and Durfee 2013] Boerkoel, J. C., and Durfee,
E. H. 2013. Distributed reasoning for multi-agent simple
temporal problems. Journal of Artificial Intelligence Re-
search 47:95–156.
[Chang 1982] Chang, E. J. H. 1982. Echo Algorithms: Depth
Parallel Operations on General Graphs. IEEE Transactions
on Software Engineering SE-8(4):391–401.
[Dechter, Meiri, and Pearl 1991] Dechter, R.; Meiri, I.; and
Pearl, J. 1991. Temporal constraint networks. Artificial
Intelligence 49(1-3):61–95.
[Dechter 2003] Dechter, R. 2003. Constraint processing.
San Francisco: Morgan Kaufmann Publishers.
[Hamadi 2002] Hamadi, Y. 2002. Optimal distributed arc-
consistency. Constraints 7(3-4):367–385.
[Hunsberger 2002] Hunsberger, L. 2002. Algorithms for a
Temporal Decoupling Problem in Multi-agent Planning. In
Eighteenth National Conference on Artificial Intelligence,
468–475. Menlo Park, CA, USA: AAAI Press.
[Jeavons and Cooper 1995] Jeavons, P. G., and Cooper,
M. C. 1995. Tractable constraints on ordered domains. Ar-
tificial Intelligence 79(2):327–339.
[Korte and Vygen 2012] Korte, B., and Vygen, J.
2012.
Combinatorial Optimization, volume 21 of Algorithms and
Combinatorics. Springer Berlin Heidelberg.
[Mattern 1987] Mattern, F. 1987. Algorithms for distributed
termination detection. Distributed computing 2(3):161–175.
[Montanari 1974] Montanari, U. 1974. Networks of con-
straints: Fundamental properties and applications to picture
processing. Information Sciences 7:95–132.
[Morris, Muscettola, and Vidal 2001] Morris, P.; Muscet-
tola, N.; and Vidal, T. 2001. Dynamic Control of Plans with
Temporal Uncertainty. In Proceedings of the 17th Interna-
tional Joint Conference on Artificial Intelligence - Volume
1, IJCAI'01, 494–499. San Francisco, CA, USA: Morgan
Kaufmann Publishers Inc.
[Nguyen and Deville 1998] Nguyen, T., and Deville, Y.
1998. A distributed arc-consistency algorithm. Science of
Computer Programming 30(1-2):227–250.
[Planken, de Weerdt, and van der Krogt 2008] Planken, L.;
2008. P3C: A
de Weerdt, M.; and van der Krogt, R.
New Algorithm for the Simple Temporal Problem.
In
Proceedings of
the Eighteenth International Conference
on International Conference on Automated Planning and
Scheduling, ICAPS'08, 256–263. Sydney, Australia: AAAI
Press.
[Planken, de Weerdt, and van der Krogt 2012] Planken,
L. R.; de Weerdt, M. M.; and van der Krogt, R. P. 2012.
Computing all-pairs shortest paths by leveraging low
Journal of Artificial Intelligence Research
treewidth.
43:353–388.
[Planken, de Weerdt, and Yorke-Smith 2010] Planken,
L.;
de Weerdt, M.; and Yorke-Smith, N. 2010. Incrementally
Solving STNs by Enforcing Partial Path Consistency.
In
Proceedings of
the Twentieth International Conference
on International Conference on Automated Planning and
Scheduling, ICAPS'10, 129–136. Toronto, Ontario, Canada:
AAAI Press.
[Raynal 2013] Raynal, M. 2013. Distributed Algorithms
for Message-Passing Systems. Berlin, Heidelberg: Springer
Berlin Heidelberg.
[Shostak 1981] Shostak, R. 1981. Deciding linear inequali-
ties by computing loop residues. J. ACM 28(4):769–779.
and
[Stergiou and Koubarakis 2000] Stergiou,
Koubarakis, M.
Backtracking algorithms for
disjunctions of temporal constraints. Artificial Intelligence
120(1):81–117.
[Tarjan 1983] Tarjan, R. E. 1983. Data Structures and Net-
work Algorithms. Number 44 in CBMS-NSF Regional Con-
ference Series in Applied Mathematics. Philadelphia, Pa:
Society for Industrial and Applied Mathematics.
[Venable and Yorke-Smith 2005] Venable, K. B., and Yorke-
Smith, N. 2005. Disjunctive Temporal Planning with Uncer-
tainty. In Proceedings of the 19th International Joint Con-
ference on Artificial Intelligence, IJCAI'05, 1721–1722. San
Francisco, CA, USA: Morgan Kaufmann Publishers Inc.
[Wilcox and Shah 2012] Wilcox, R., and Shah, J.
2012.
Optimization of Multi-Agent Workflow for Human-Robot
In In-
Collaboration in Assembly Manufacturing.
fotech@Aerospace 2012. American Institute of Aeronautics
and Astronautics.
2000.
K.,
|
1806.02660 | 1 | 1806 | 2018-05-26T00:13:48 | Analyzing Traffic Delay at Unmanaged Intersections | [
"cs.MA"
] | At an unmanaged intersection, it is important to understand how much traffic delay may be caused as a result of microscopic vehicle interactions. Conventional traffic simulations that explicitly track these interactions are time-consuming. Prior work introduced an analytical traffic model for unmanaged intersections. The traffic delay at the intersection is modeled as an event-driven stochastic process, whose dynamics encode microscopic vehicle interactions. This paper studies the traffic delay in a two-lane intersection using the model. We perform rigorous analyses concerning the distribution of traffic delay under different scenarios. We then discuss the relationships between traffic delay and multiple factors such as traffic flow density, unevenness of traffic flows, temporal gaps between two consecutive vehicles, and the passing order. | cs.MA | cs | Analyzing Traffic Delay at Unmanaged Intersections
Changliu Liu and Mykel J. Kochenderfer
1
8
1
0
2
y
a
M
6
2
]
A
M
.
s
c
[
1
v
0
6
6
2
0
.
6
0
8
1
:
v
i
X
r
a
Abstract-At an unmanaged intersection, it is important to
understand how much traffic delay may be caused as a result of
microscopic vehicle interactions. Conventional traffic simulations
that explicitly track these interactions are time-consuming. Prior
work introduced an analytical traffic model for unmanaged
intersections. The traffic delay at the intersection is modeled
as an event-driven stochastic process, whose dynamics encode
microscopic vehicle interactions. This paper studies the traffic
delay in a two-lane intersection using the model. We perform
rigorous analyses concerning the distribution of traffic delay
under different scenarios. We then discuss the relationships
between traffic delay and multiple factors such as traffic flow
density, unevenness of traffic flows, temporal gaps between two
consecutive vehicles, and the passing order.
I. INTRODUCTION
Delay at intersections affect the capacity of a road network.
There are many methods to analyze traffic delay at signalized
intersections [1], [2], [3]. Such analyses are able to allow
better traffic control to minimize delay. With the emergence
of autonomous vehicles, there is a growing interest in leaving
intersections unmanaged, allowing vehicles to resolve conflicts
among themselves [4]. Unmanaged intersections can reduce
infrastructure cost and allow for more flexible road network
designs. Various vehicle policies have been proposed for
distributed conflict resolution at unmanaged intersections [5],
[6], [7].
It is important to understand how these microscopic poli-
cies affect the macroscopic transportation system. Toward the
development of an efficient transportation system, we need to
quantify the traffic delay generated during vehicle interactions
at those intersections.
Delay at intersections is generally evaluated using micro-
scopic traffic simulation [8]. Various evaluation platforms have
been developed [9], including AIMSUN [10] and VISSIM
[11]. However, it is time-consuming to obtain the micro-macro
relationship by simulation. Only "point-wise" evaluation can
be performed in the sense that a single parametric change in
vehicle behaviors requires new simulations. In order to gain
a deeper understanding of the micro-macro relationships, an
analytical model is desirable.
In contrast with microscopic simulation models, macro-
scopic flow models [12] are analytical. Traffic is described
by relations among aggregated values such as flow speed and
density, without distinguishing its constituent parts. The major
advantage of macroscopic flow models is their tractable math-
ematical structure with relatively few parameters to describe
interactions among vehicles. However, it remains challenging
to model intersections. Though intersections can be included
C. Liu and M. Kochenderfer are with the Department of Aero-
nautics and Astronautics, Stanford University, CA 94305 USA (e-mail:
changliuliu, [email protected]).
(a)
(b)
Fig. 1: Intersection scenario. (a) Road topology. (b) Conflict graph.
(a)
(b)
(c)
Fig. 2: The time of occupancy at the intersection. (a) The desired
time of occupancy. (b) The actual time of occupancy under FIFO.
(c) The actual time of occupancy under FO.
in the flow models as boundary constraints [13], [14],
it
is difficult to model policies other than the first-in-first-out
(FIFO) policy. To consider a variety of policies, the vehicles
need to be treated as particles that interact with one another,
which has not been captured by existing flow models.
The authors introduced an analytical
traffic model [15]
to describe delays at unmanaged intersections. The model
is event-driven, whose dynamics encodes equilibria resulting
from microscopic vehicle interactions. It absorbs the advan-
tages of both the microscopic simulation models and the
macroscopic flow models. This paper performs detailed delay
analysis at unmanaged intersections using the model. The fol-
lowing two components in a vehicle policy strongly influence
the traffic delay: 1) determination of the passing order, and 2)
the temporal gap between two consecutive vehicles to pass the
intersection. We will illustrate how these two components as
well as the distribution of incoming traffic flows affect delay.
The major contributions of this paper are:
1) Illustration of the usage of the analytical traffic model to
obtain analytical distributions of delay.
2) Derivation of the analytical distribution of delay under
two different classes of policies (i.e., two different passing
orders) at a two-lane intersection;
3) Analysis of how traffic delay is affected by multiple
factors at the two-lane intersection.
The remainder of the paper is organized as follows. Sec-
Lane 1Lane 2Lane 3Lane 432214Incoming flowOutgoing flowTimeLane1234to1to2to3to4to5to6to7to8to9TimeLane1234¯t1¯t2¯t3¯t4¯t5¯t6¯t7¯t8¯t9TimeLane1234¯t1¯t2¯t3¯t4¯t5¯t6¯t7¯t8¯t9 2
tion II reviews the analytical traffic model and illustrates how
vehicle behaviors are encoded in the model. Section III derives
the analytical distribution of delay under two different classes
of policies. Section IV shows how the traffic delay is affected
by multiple factors using the analytical distribution. Section V
concludes the paper.
II. TRAFFIC MODEL
This section reviews an event-driven stochastic model for
traffic delay at intersections [15]. The following discussion
considers an intersection with K incoming lanes. A conflict
is where two incoming lanes intersect with each other. These
relationships can be described in a conflict graph G with the
nodes being the incoming lanes and the links representing
conflicts. Fig. 1a illustrates one possible road configuration
with four incoming lanes, and Fig. 1b shows the resulting
conflict graph.
N(cid:88)
i=1
N(cid:88)
i=1
The problem of interest is to quantify the average delay
¯d = lim
N→∞
1
N
(¯ti − to
i ) = lim
N→∞
1
N
(¯t(N )
i ).
i − to
(4)
Fig. 2a illustrates the desired time of occupancy for vehicles
from the four lanes in Fig. 1a. The bars represent the moments
that the intersection is occupied by vehicles, which is centered
at to
i . According to the conflict graph, the scenario in Fig. 2a
is infeasible as vehicles 1, 2, 3, and 4 cannot occupy the
intersection at the same time. After some negotiation and
adaptation among vehicles,
time of occupancy
becomes as shown in Fig. 2b or Fig. 2c. For an unmanaged
intersection, the actual time of occupancy depends on the
policies that
the vehicles adopt. Fig. 2b and Fig. 2c are
different as they correspond to different policies, which will
be discussed in detail in Section III-A. This paper quantifies
the effectiveness of the policies based on the resulted average
delay.
the actual
A. Microscopic Interactions
C. Traffic Model at Intersections
It is assumed that the vehicles at intersections have fixed
paths. To respond to others during interactions, the vehicles
only change their speed profiles to adjust the time to pass the
intersection [16], [17]. This paper reduces the high dimen-
sional speed profile for vehicle i to a single state ti, which
denotes the time for that vehicle to pass the center of the
intersection. As the mapping from ti to the speed profile is
surjective, we can analyze interactions using ti's. The desired
traffic-free time for vehicle i to pass the intersection is denoted
i . The vehicles are indexed according to the desired passing
to
time such that to
i ≤ to
At time step k, vehicle i decides its passing time based on
its desired time to
i and its observation of others' passing times
at the last time step t−i(k − 1) := [t1(k − 1), . . . , ti−1(k −
1), ti+1(k − 1), . . .]. The policy of vehicle i is denoted
i+1 for all i.
ti(k) = f (to
i , t−i(k − 1)).
(1)
It is assumed that all vehicles use the same policy f.
B. Equilibria
The equilibrium among the first i vehicles is denoted
1 , . . . , ¯t(i)
(¯t(i)
i ). In an equilibrium, no vehicle is willing to
change the passing time before the arrival of the (i + 1)th
vehicle. Hence, the equilibrium is time-invariant, i.e.,
¯t(i)
j = f (to
j , ¯t(i)−j),∀j ≤ i.
(2)
It is assumed that an equilibrium can be achieved in negligible
time. Hence, the system moves from the ith equilibrium to the
(i + 1)th equilibrium when the (i + 1)th vehicle is included.
The projected passing time for a vehicle may change from
one equilibrium to another equilibrium, but will eventually
converge to the actual passing time. The actual passing time
¯ti for vehicle i is
¯ti = lim
j→∞ ¯t(j)
i
.
(3)
For quantitative analysis, the traffic is modeled as an event-
driven stochastic system with the state being the traffic delay
and the input being the incoming traffic flow. The delay for
lane k considering i vehicles is denoted T k
i , which captures
the difference between the passing time in the ith equilibrium
and the traffic-free passing time of those vehicles, i.e.,
T k
i = max
sj =k,j≤i
¯t(i)
i ,
j − to
(5)
where sj is the lane number of vehicle j. The input to the
traffic model is the random arrival interval xi = to
i+1 − to
i
between vehicle i + 1 and vehicle i, and the lane number si+1
of vehicle i + 1. Define Ti := [T 1
]T . The dynamics
of the traffic delay follow from
i , . . . , T K
i
Ti+1 = F(Ti, xi, si+1),
(6)
where the function F depends on the policy f in (2) and the
road topology defined by the conflict graph G in Fig. 1b.
It is assumed that the desired passing time of the incoming
traffic flow from lane k follows a Poisson distribution with
parameter λk. The traffic flows from different lanes are in-
dependent of each other. Since the combination of multiple
independent Poisson processes is a Poisson process [18],
the incoming traffic from all lanes can be described as one
k λk. The
Poisson process (to
probability density for xi = x is px(x) = λe−λx. The
λ .
probability for si+1 = k is Ps(k) = λk
2, . . .) with parameter λ =(cid:80)
Given (6), the conditional probability density of Ti+1 given
1, to
Ti, xi and si+1 is
pTi+1(t Ti, xi, si+1) = δ(t = F(Ti, xi, si+1)),
(7)
where δ(·) is the Dirac delta function. The total distribution is
(cid:90)
(cid:90)
(cid:90)
x
(cid:88)
(cid:88)
k
k
pTi+1 (t)
Ps(k)
Ps(k)
=
=
pTi+1 (t τ , x, k)pTi (τ )dτ px(x)dx
τ
δ(0)pTi (τ )px(x)dτ dx,
(8)
F (τ ,x,k)=t
involves
cumulative
which
The
PTi(t)
for t = [t1, . . . , tk].
=
(cid:82) (t1)+
integration
probability
(cid:82) (tk)+
over
of Ti
manifold.
denoted
−∞ pTi(τ 1, . . . , τ k)dτ 1 . . . dτ k
is
a
−∞ ···
In this paper, we investigate the steady state distribution
pT := limi→∞ pTi. Necessary conditions for the convergence
of limi→∞ pTi are provided in Section III. For simplicity,
define the functional mapping M as
M(p)(t) =
δ(0)px(x)p(τ )dxdτ .
(cid:88)
Ps(k)
(cid:90)
F (τ ,x,k)=t
(9)
k
The steady state distribution pT is a fixed point under M.
D. Usage of the Model
Under the model, the distribution of vehicle delay can either
be obtained through direct analysis or event-driven simulation.
1) Theoretical Analysis: The vehicle delay introduced by
the (i + 1)th vehicle is
(cid:16)
(cid:88)
j≤i
di+1 =
¯t(i+1)
j
+ ¯t(i+1)
∗
i+1.
i+1 − t
(10)
In the case that the introduction of a new vehicle only affects
the last vehicle in other lanes (which is usually the case),
di+1 = T si+1
i+1 +
(T k
i+1 − T k
i + xi).
k(cid:54)=si+1
(11)
The two classes of policies correspond to two ways to
3
(a) Domain
(b) Value
Fig. 3: Illustration of the mapping (6) under FIFO for si+1 = 1.
III. STEADY STATE DISTRIBUTION OF DELAY
This section derives the steady state distribution of delay
under two classes of frequently used policies in a two-lane
intersection using the method discussed in Section II-D1.
The two policies are the first-in-first-out (FIFO) policy and
the flexible order (FO) policy, which entail different passing
orders. The required temporal gap between vehicles from
different directions is denoted ∆d. The required temporal gap
between vehicles from the same direction is denoted ∆s.
The gap is affected by the following factors: vehicle speed,
uncertainties in perception, and etc.
A. Vehicle Policies
determine the passing order.
1) FIFO: The passing order is solely determined according
to the arrival time (which is taken to be the desired passing
time to
i ). The actual passing time for vehicle i should be after
the actual passing times for all conflicting vehicles j such that
j < i.1 As the passing order is fixed, the actual passing time
will not be affected by later vehicles, i.e., ¯tj = ¯t(i)
for
all i > j. For vehicle i,
j = ¯t(j)
j
¯t(i)
i
:= max{to
(14)
where Di is the earliest passing time considering vehicles from
other lanes, and Si is the earliest passing time considering
vehicles from the ego lane.
i ,Di,Si},
j
Di = max
Si = max
j
(¯t(i)
j + ∆d) s.t. j < i, (sj, si) ∈ G,
(¯t(i)
j + ∆s) s.t. j < i, sj = si.
(15a)
(15b)
The effect of FIFO is illustrated in Fig. 2b.
i
:= max{to
i , maxj<i,sj =si(¯t(i−1)
2) FO: This strategy allows high priority vehicles to yield
to low priority vehicles if low priority vehicles can arrive
earlier. The passing order may change over time. At step
i, let ¯t(i−1)
+ ∆s)} be the
earliest possible time for vehicle i to pass considering its front
i−1 , ¯t(i−1)
vehicles in the ego lane. Sort the list (¯t(i−1)
)
1Some authors define FIFO to be such that vehicle i should yield to vehicle
j for all j < i no matter there is a conflict or not. The FIFO strategy presented
in this paper is similar to the Maximum Progression Intersection Protocol
(MP-IP) [6]. Nonetheless, there is no difference between the two in the two-
lane scenario.
, . . . , ¯t(i−1)
1
j
i
(cid:17)
j
− ¯t(i)
(cid:88)
Hence, to obtain an analytical steady state distribution of
vehicle delay, we need to 1) obtain (6) from microscopic
interactions models, then 2) solve the fixed point problem
M(p) = p for the steady state distribution pT, and finally
3) compute the steady state distribution of vehicle delay pd
from pT by (11). Section III illustrates the procedures for the
derivation.
The relationship between ¯d in (4) and di in (10) is
¯d = lim
N→∞
1
N
di.
(12)
(cid:88)
i
According to the central limit theorem, the system is ergodic
such that the average delay of all vehicles equals the expected
delay introduced by a new vehicle (moving from one equilib-
rium to another equilibrium) in the steady state,
E( ¯d) = lim
i→∞ E(di).
(13)
2) Event-Driven Simulation (EDS): The transition of the
distribution from one equilibrium to another can also be
simulated. Unlike conventional time-driven traffic simulation,
we can perform event-driven simulation, which is more ef-
ficient. Many particles need to be generated for T0, each
corresponding to one traffic scenario. Those particles are
then propagated according to (6) by randomly sampling xi
and si+1. As the particles propagate, either the distribution
diverges or we obtain the steady state distribution of delay.
xixi s d00T1iT2i1234 d132 d0T1i+1T2i+14 4
(a) FIFO
(b) FO
Fig. 4: The steady state distribution pT for λ1 = 0.1 s−1, λ2 = 0.5 s−1,
∆d = 2 s, and ∆s = 1 s from EDS with 10000 particles.
k := max{¯t(i−1)
¯t(i)
k
,Di
k},
k,S i
(16)
Ps(1)
pTi(τ, t2 + x)dτ pxdx.
(19)
For t1 > t2 + ∆d > 0, pTi+1(t1, t2) =
TABLE I: The mapping (6) under FIFO for si+1 = 1.
Region
Value
Condition
i < xi − ∆s
T 1
i < xi − ∆d
T 2
i ≥ xi − ∆s
T 1
i < xi − ∆d
T 2
T 2
i < T 1
i
i ≥ xi − ∆d
T 2
T 2
i < T 1
i
i ≥ xi − ∆d
T 2
T 2
i > T 1
i
1
2
3
4
T 1
i+1 = 0
i+1 = −∆d
T 2
T 1
i+1 = T 1
i + ∆s − xi
i+1 = −∆d
T 2
T 1
i+1 = T 1
T 2
i+1 = T 2
T 1
i+1 = T 2
T 2
i+1 = T 2
i + ∆s − xi
i − xi
i + ∆d − xi
i − xi
in ascending order and record the ranking in Q : N → N. If
there is a tie, the vehicle with a smaller index is given a smaller
Q value. For the first vehicle in Q, i.e., vehicle k = Q−1(1),
the passing time is ¯t(i)
. By induction, assuming that
¯t(i)
j
for Q(j) < Q(k) has been computed, then
k := ¯t(i−1)
k
j
j
(¯t(i)
j + ∆d) s.t. Q(j) < Q(k), (sj, sk) ∈ G, (17a)
(¯t(i)
(17b)
j + ∆s) s.t. Q(j) < Q(k), sj = sk.
where
Di
k = max
S i
k = max
Under FO, the actual passing time may change over time.
There is a distributed algorithm [7] to implement this policy
where the vehicles do not necessarily need to compute the
global passing order. The effect of FO is illustrated in Fig. 2c.
Vehicles in the same direction tend to form groups and pass
together. For a two-lane intersection,
the passing order is
changed if and only if the next vehicle can pass the intersection
earlier than the last vehicle in the other lane.
B. Case 1: Delay under FIFO
Following from (5) and (14), the dynamic equation (6)
for FIFO can be computed, which is listed in Table I and
illustrated in Fig. 3. Only the case for si+1 = 1 is shown.
Define a conjugate operation (·)∗ as i∗ := 3 − i. The case
for si+1 = 2 can be obtained by taking the conjugate of all
superscripts. In order to bound the domain from below, let
i = max{T j
T j
i ,−∆d} for all i and j ∈ {1, 2}. There are four
smooth components in the mapping as illustrated in Fig. 3 and
Table I. Region 1 corresponds to the case that there is enough
gap in both lanes for vehicle i + 1 to pass without any delay.
Regions 2 and 3 correspond to the case that the last vehicle
is from the ego lane and it causes delay for vehicle i + 1.
Region 4 corresponds to the case that the last vehicle is from
the other lane and it causes delay for vehicle i + 1.
Given the dynamic equation, the probability (8) can be
computed. For simplicity, we only show the case for t1 > t2.
The case for t1 < t2 is symmetric. When t2 = −∆d,
PTi+1(t1,−∆d) =
(cid:90) ∞
Ps(1)
0
PTi (t1 + x − ∆s, x − ∆d)pxdx.
(18)
When t1 = t2 + ∆d, pTi+1(t1, t2) =
(cid:90) ∞
(cid:90) t2+x−∆d
(cid:90) ∞
−∆d
0
Ps(1)
0
pTi(t1 − ∆s + x, t2 + x)pxdx.
(20)
Proposition 1 (Necessary Condition for Convergence under
FIFO). The distributions {pTi}i converges for FIFO only if
the following condition holds
2λ1λ2∆d + [λ2
1 + λ2
2]∆s ≤ λ.
(21)
Proof. The convergence of the distribution implies the con-
vergence of the expected delay. Hence, the minimum average
departure interval between two consecutive vehicles should be
smaller than the average arrival interval. For two consecutive
vehicles, the probability that they are from the same lane
is Ps(1)2 + Ps(2)2, and the probability that they are from
different lanes is 2Ps(1)Ps(2). Hence, the minimum average
departure interval is 2Ps(1)Ps(2)∆d +(cid:2)Ps(1)2 + Ps(2)2(cid:3) ∆s.
2Ps(1)Ps(2)∆d +(cid:2)Ps(1)2 + Ps(2)2(cid:3) ∆s ≤
1
λ
Condition (21) can be obtained by rearranging (22).
The average arrival interval is 1
distribution implies
λ. The convergence of the
(22)
.
The proof of the sufficiency of (21) is left as future work.
In the following discussion, we investigate the steady state
distribution pT = M(pT) for ∆s > 0 and ∆s = 0.
Proposition 2 (Steady State Distribution for ∆s > 0 under
FIFO). When ∆s > 0, for t > −∆d and Γ > ∆d, the
following equalities hold,
pT(t + Γ, t) = C 1
n
pT(t, t + Γ) = C 2
n
pT(t + γ, t)e
−λzzn−1dz,
pT(t, t + γ)e
−λzzn−1dz,
(23a)
(23b)
(cid:90) ∞
(cid:90) ∞
0
0
where n is the maximum integer such that γ := Γ − n∆s ∈
(∆d − ∆s, ∆d], C i
(n−1)! , and t = t + z. Moreover,
n = λn
i
05000.02510100.04T1T2050050.0210100.04T1T2 5
(25a)
(25b)
(25c)
(26)
,
(27)
(28)
(29)
(30)
gi(t) = gi(t) +(cid:98)gi(t)δ(t). Moreover, for i ∈ {1, 2}, define the
probability function Gi, value Mi and value Ii as
(cid:90) t
(cid:90) ∞
(cid:90) ∞
0
0
0
Gi(t) :=
Mi :=
Ii :=
gi(τ )dτ,
gi(x)dx,
gi(x)e
−λxdx.
Value Mi is the probability that lane i has larger delay.
Proposition 4 (Steady State Distribution for ∆s = 0 under
FIFO). When ∆s = 0, pT(t1, t2) = 0 if t1 − t2 (cid:54)= ∆d. For
i ∈ {1, 2}, the following equations hold
Gi(t) = (cid:98)gi(0)eλi∗ t, for t ∈ [0, ∆d],
(cid:26) λi
Ii + e−λ∆dIi∗(cid:3)
(cid:2)
(cid:98)gi(t) =
λi(cid:98)gi∗ (0) = gi(∆
−
d ),
d ) − gi(∆+
λ
0
t = 0
t (cid:54)= 0
Mi =
λi
λ
.
Moreover, when t is sufficiently large,
Mi − Gi(t) ∝ e
−at,
(a − λ1)(a − λ2) − λ1e
where a < 0 is the solution of the following equation
−a∆d = 0.
(31)
Proof. We first show that pT(t1, t2) = 0 if t1 − t2 (cid:54)= ∆d.
There are two cases: min{t1, t2} > −∆d or min{t1, t2} =
−∆d. Consider Γ > ∆d and t > −∆d. According to (20),
−a∆d λ2e
(cid:90) ∞
pT(t + Γ, t) = λ1
pT(t + x + Γ, t + x)e
0
−λxdx.
(32)
By Lemma 3, (32) implies pT(t + Γ, t) ≡ 0 for all t > −∆d.
Similarly, pT(t, t + Γ) ≡ 0 for all t > −∆d. Moreover, for
t > 0, according to (18),
(cid:90) ∞
0
pT(t1, t2) = 0 if min{t1, t2} > −∆d and t1−t2 (cid:54)= ∆d+n∆s
for any n ∈ N.
Proof. Since (23a) and (23b) are symmetric, we will only
show the derivation for (23a) for simplicity. By (20),
(cid:90) ∞
pT(t + n∆s + γ, t)
= Ps(1)
0
pT(t + (n − 1)∆s + γ + x1, t + x1)pxdx1
By induction on n,
(cid:90)
pT(t + z + γ, t + z)pxdx,
pT(t + Γ, t) = Ps(1)n
where x = [x1, x2, . . . , xn], z = (cid:80)n
x≥0
0
pT(t + Γ, t) = λn
1
k=1 xk, and px(x) =
λne−λz. By change of variable from x to [z, x2, . . . , xn],
(cid:90) ∞
−λzdz,
V (z, n− 1)pT(t + z + γ, t + z)e
(n−1)! zn−1 is the volume of an (n− 1)-
where V (z, n− 1) = 1
dimensional cone with depth z.2 Hence, (23a) is verified.
If γ ∈ (∆d− ∆s, ∆d), by definition, pT(t + γ, t) = 0. Then
pT(t + n∆s + γ, t) = 0 for any n ∈ N according to (23a).
Similarly, pT(t, t + n∆s + γ) = 0 for any n ∈ N. Hence,
pT(t1, t2) = 0 if min{t1, t2} > −∆d and t1−t2 (cid:54)= ∆d+n∆s
for any n ∈ N.
Proposition 2 implies a unique "zebra" pattern of the steady
state lane delay. This pattern is also observed in EDS shown
in Fig. 4a. The exact solution of pT for ∆s > 0 is left as
future work. In the following discussion, we derive the case
for ∆s = 0. The assumption that ∆s = 0 is valid when the
traffic density is low. Lemma 3 is useful in the derivation of
the steady state delay.
Lemma 3 (Zero Function). For any norm-bounded L1 func-
0 f (t+x)e−λxdx for all t and λ ≤ a > 0,
tion f, if f (t) = a(cid:82) ∞
(cid:90) ∞
then f ≡ 0.
Proof. Multiply e−λt on both sides, then
−λtf (t) = a
−λ(x+t)dx = a
f (t + x)e
(cid:90) ∞
e
0
t
Take derivative with respect to t on both sides, then
−λtf
(cid:48)
e
(t) − λe
−λtf (t) = −af (t)e
−λt,
which implies that f(cid:48)(t) = (λ − a)f (t) and f (t) = Ce(λ−a)t
for some constant C. However, since λ − a ≥ 0, f cannot be
norm bounded if C (cid:54)= 0. Hence, f ≡ 0.
In the following discussion, we derive the steady state
distribution of delay for ∆s = 0. Proposition 4 shows that
when ∆s = 0, the probability density is non trivial only at
pT(t, t − ∆d) or pT(t − ∆d, t) for t ≥ 0. Hence, we define
(24)
g1(t) := pT(t, t − ∆d), g2(t) := pT(t − ∆d, t).
and delta component, denoted gi and(cid:98)gi respectively such that
The function gi for i ∈ {1, 2} contains both finite component
2V (z, n − 1) =(cid:82) z
· · ·(cid:82) z−x2−...−xn−1
(cid:82) z−x2
dxn · · · dx3dx2.
0
0
0
f (x)e
−λxdx.
pT(t,−∆d) = λ1
pT(t + x,−∆d)e
−λxdx.
(33)
By Lemma 3, (33) implies pT(t,−∆d) ≡ 0 for t > 0.
Similarly, pT(−∆d, t) ≡ 0 for t > 0. Hence, the claim is
verified.
Now we compute the steady state distribution gi. In either
g1 or g2, there is only one point mass at 0 by (18) to (20).
According to (18),
(cid:98)gi(0)
(cid:34)(cid:90) x
(cid:90) ∞
(cid:20)(cid:90) ∞
(cid:90) ∞
(cid:20)(cid:90) ∞
0
0
0
t
= λi
=
λi
λ
= λi
gi(t)dt +
−λxdxgi +
e
(cid:35)
gi∗ (t)dt
(cid:90) ∞
e
t+∆d
−λxdx
e
−λxdxgi∗
(cid:21)
dt
,(34)
(cid:21)
(cid:90) x−∆d
(cid:90) ∞
(cid:90) ∞
0
0
0
−λtdt +
gi∗ (t)e
−λ(t+∆d)dt
gi(t)e
0
where the second equality is obtained by changing the order
of integration. By definition (25c), (34) implies (27).
(cid:90) ∞
According to (19) and (20), for t > 0,
−λxdx,
(35)
gi(t) = λi
[gi(t + x) + gi∗ (t + x − ∆d)] e
0
continuity at ∆d is caused by the point mass(cid:98)gi∗ (0). By (35),
which implies that gi is continuous except at ∆d. The dis-
the claim in (28) is verified. By multiplying e−λt on both
sides of (35) and then taking derivatives similar to the proof
in Lemma 3, we obtain the following differential equation
(cid:48)
i(t) = λi∗ gi(t) − λigi∗ (t − ∆d).
g
(36)
For t ∈ (0, ∆d), since gi∗ (t− ∆d) = 0, (36) implies that there
exists ci ∈ R+ such that
(cid:2)
Ii −(cid:98)gi(0) + e
(37)
Plugging (37) back to (35), the constant ci can be computed,
(38)
Then (26) is verified by integrating (37). Moreover, it is easy
λ . Hence, (26) to (29) are all
to verify that Mi = Ps(i) = λi
verified.
−λ∆dIi∗(cid:3) = λi∗(cid:98)gi(0).
gi(t) = cieλi∗ t.
ci = λi
The characteristic equation [19] of the delay differential
(cid:18)
equation (36) for i ∈ {1, 2} satisfies
−a∆d
det
+ e
aI2 −
0
λ1
(cid:20) λ2
0
(cid:21)
(cid:20) 0
λ2
(cid:21)(cid:19)
λ1
0
= 0,
(39)
which is equivalent to the nonlinear eigenproblem (31). There
are three possible solutions with a = 0, a > 0, and a < 0,
respectively. Since limt→∞ gi(t) = 0, we can only take the
solution a < 0. When t → ∞, gi(t) is proportional to eat.
Then (30) is verified.
λi
Gi(t) =
To compute the exact solution of the distribution, the delay
differential equation (DDE) (36) needs to be solved. To solve
the DDE, we need to compute the expression of Gi(t) for
t ∈ ((n − 1)∆d, n∆d] consecutively for all n considering
the boundary constraints (26) to (29). However, as there are
infinitely many segments, the complexity of the problem grows
quickly. In this paper, we approximate the distribution for
t > ∆d using (30). By incorporating (26) and (29),
the
approximated distribution is
(cid:26) (cid:98)gi(0)eλi∗ t
λ (1 − ea(t−∆d)) + Gi(∆d)ea(t−∆d)
There is only one unknown parameter (cid:98)gi(0), which can be
solved by the remaining equations in Proposition 4. However,
the approximated distribution (40) is not simultaneously com-
patible with (27) and (28). We need to relax either condition.
Equation (27) is a global condition as it is related to the
integral of the distribution. Equation (28) is a local condition as
it concerns the discontinuous point of the probability density.
Remark 1 (Approximation 1). In the first approximation, the
local condition (28) is relaxed. Then (cid:98)gi(0) is obtained by
aλiy(cid:0)(λi − a)λi(y2 − 1) + (a − λ)yi [λi∗ + λiy](cid:1)
(cid:98)gi(0) =
solving (27) and (40),
t ≤ ∆d
t > ∆d
.
(40)
Bi
,
(41)
6
y
t
i
l
i
b
a
b
o
r
P
e
v
i
t
a
l
u
m
u
C
0.6
0.4
0.2
0
0
G2 EDS
G2 Approx1
G2 Approx2
G1 EDS
G1 Approx1
G1 Approx2
2
4
6
8
10
12
14
Delay t [s]
Fig. 5: Steady state traffic delay under FIFO. λ1 = 0.3 s−1, λ2 =
0.5 s−1, ∆d = 2 s, and ∆s = 0 s.
where y := e−λ∆d, yi := e−λi∆d, and
Bi = λ2(a2y(y − yi)(1 − yi) + a(a − λ)yi
i + λiλi∗ yi + λi
+(a − λi)λy2(yi − 1) + (2a − λ)λyyi(1 − yi)
2y2yi − aλiy2).
+(a − λ)λiyy2
Remark 2 (Approximation 2). In the second approximation,
we relax the global condition (27). Then(cid:98)gi(0) is obtained by
(cid:98)gi(0) =
aλiyea∆d(cid:0)λi + λi∗ yi − aea∆d(cid:1)
solving (28) and (40),
(43)
(42)
λ (aλea∆d − a2e2a∆d + λiλi∗ (y − 1))
The accuracy of the two approximations against the steady
state distribution obtained from EDS with 10000 particles is
shown in Fig. 5. Though both underestimate the delay, (41)
provides a better approximation because it preserves the global
property. In the following discussion and analysis, we use the
first approximation.
Corollary 5 (Approximated Steady State Vehicle Delay).
When ∆s = 0, under the approximation (40), the steady state
vehicle delay has the distribution
.
Pd(t) =
(cid:26) (cid:98)g1(0)eλ2t +(cid:98)g2(0)eλ1t
1 − ea(t−∆d) + Pd(∆d)ea(t−∆d)
E(d) =(cid:98)g1(0)E(λ2)+(cid:98)g2(0)E(λ1)−
with expected delay
t ≤ ∆d
t > ∆d
,
(44)
(a∆d − 1) (Pd(∆d) − 1)
,
(45)
a
where
1 + e∆dλi (∆dλi − 1)
.
(46)
E(λi) =
λi
expected delay satisfies E(d) = (cid:82) ∞
(cid:82) ∆d
Proof. By (11), the vehicle delay in the steady state satisfies
that Pd(t) = G1(t) + G2(t). So (44) follows from (40). The
tdPd(t). Let E(λi) :=
tdeλit. Then (45) and (46) follow.
0
0
C. Case 2: Delay under FO
Following from (5) and (16), the dynamic equation (6) for
FO can be computed, which is listed in Table II and illustrated
in Fig. 6 for si+1 = 1. There are eight smooth components
in the mapping. Regions 1 to 4 are the same as in the FIFO
case. Vehicle i + 1 is the last one to pass the intersection.
Regions 5 to 8 correspond to the case that vehicle i + 1 passes
the intersection before the last vehicle in the other lane. In
TABLE II: The mapping (6) under FO for si+1 = 1.
7
Domain
i < xi − ∆s
T 1
i < xi − ∆d
T 2
i ≥ xi − ∆s
T 1
i < xi − ∆d
T 2
T 2
i < T 1
i
i ≥ xi − ∆d
T 2
T 2
i < T 1
i
i ∈ [xi − ∆d, xi)
T 2
T 2
i > T 1
i
i ∈ [xi, xi + ∆d)
T 2
i < xi − ∆s
T 1
i ∈ [∆d , ∆d + ∆s ]
i − T 1
T 2
i ≥ xi − ∆s
T 1
i < xi − ∆s
T 1
i ≥ xi + ∆d
T 2
i − T 1
T 2
i > xi + ∆d + ∆s
i ≥ xi − ∆s
T 1
1
2
3
4
5
6
7
8
Value
T 1
i+1 = 0
i+1 = −∆d
T 2
T 1
i+1 = T 1
i + ∆s − xi
i+1 = −∆d
T 2
T 1
i+1 = T 1
T 2
i+1 = T 2
i + ∆s − xi
i − xi
i + ∆d − xi
T 1
i+1 = T 2
T 2
i+1 = T 2
i − xi
T 1
i+1 = 0
T 2
i+1 = ∆d
T 1
i+1 = T 1
i − xi + ∆s
i − xi + ∆s + ∆d
T 2
i+1 = T 1
T 1
i+1 = 0
T 2
i+1 = T 2
i − xi
T 1
i+1 = T 1
i − xi + ∆s
i − xi
T 2
i+1 = T 2
,
Gi(t) =
Proposition 7 (Steady State Distribution for ∆s = 0 under
FO). If ∆s = 0, then pT(t1, t2) = 0 if t1 − t2 (cid:54)= −∆d or
t1 + t2 > ∆d. Moreover, for i ∈ {1, 2},
t ∈ [0, ∆d)
t ≥ ∆d
(cid:26) ci
λiλi∗(cid:0)λiy2 + λiyi∗ + λi∗ y − λiy2yi∗(cid:1)
λi∗ eλi∗ t
Mi
,
Mi =
ci =
λ2 (1 + yyi + yyi∗ − y − y2)
Proof. Similar to the proof of Proposition 4,
it
is easy
to show that pT(t1, t2) = 0 if
(cid:54)= −∆d.
t1 − t2
regions 3 and 6, gi(t) =
For
t > ∆d, consider
[Ps(i)gi(t + x) + Ps(i∗)gi(t + x)] pxdx. Hence,
(cid:82) ∞
. (50)
λi
λ
(48)
(49)
0
(cid:90) ∞
gi(t) = λ
gi(t + x)e
0
−λxdx.
(51)
According to Lemma 3, gi(t) ≡ 0 for t > ∆d. Hence,
pT(t1, t2) = 0 if t1 − t2 (cid:54)= −∆d or t1 + t2 > ∆d.
For t ∈ (0, ∆d), consider regions 3 and 4, (35) holds.
Similar to the proof in Proposition 4 from (35) to (38), we
conclude that gi = cieλi∗ t for some constant ci such that
ci = λi∗(cid:98)gi(0).
(52)
Then (48) is verified. We solve for ci below.
(cid:98)gi(0) =
expression as in the FIFO case,
Consider region 1. The point mass at 0 has the same
−λ∆dIi∗(cid:3) .
(cid:2)
Consider region 5. The point mass at ∆d satisfies(cid:98)gi(∆d) =
(cid:82) ∞
(cid:82) ∆d
(cid:82) τ
(cid:82) ∆d
gration, we have (cid:98)gi(∆d) = λi
gi∗ (τ )dτ e−λxdx. By changing the order of inte-
(cid:82) ∆d
0 e−λxdxgi∗ (τ )dτ =
(1 − e−λτ )gi∗ (τ )dτ. Hence,
Ii + e
λi
λ
(53)
λi
λ
λi
x
0
0
0
(cid:98)gi(∆d) =
λi
λ
[Mi∗ − Ii∗ ].
(54)
(a) Domain
(b) Value
Fig. 6: Illustration of the mapping (6) under FO for si+1 = 1.
regions 5 and 7, vehicle i + 1 arrives earlier than the last
vehicle in the other lane and there is enough gap in the ego
lane. Hence, vehicle i + 1 passes without delay, but the last
vehicle in the other lane yields (with delay in region 5, without
delay in region 7). Regions 6 and 8 correspond to the case that
vehicle i + 1 is delayed by the last vehicle in the ego lane but
can still go before the last vehicle in the other lane. Delay is
caused in the other lane in region 6.
Given the dynamic equation, the probability (8) can be
computed. The distribution obtained from EDS with the same
condition as in the FIFO case is shown in Fig. 4b. FO generates
smaller delay as compared to FIFO, but FO no longer has the
"zebra" pattern shown in FIFO.
In the following discussion, we discuss the necessary condi-
tion for convergence under FO and derive the exact steady state
distribution of delay for ∆s = 0. The distribution for ∆s > 0
is left as future work. Recall the definitions y = e−λ∆d and
yi = e−λi∆d for i ∈ {1, 2}.
Proposition 6 (Necessary Condition for Convergence under
FO). The distributions {pTi}i converges for FO only if the
following condition holds
1+λ2
λ1λ2(y1+y2)∆d+[λ2
2+λ1λ2(2−y1−y2)]∆s ≤ λ. (47)
Proof. Similar to the discussion in Proposition 1, the min-
imum average departure interval between two consecutive
vehicles should be smaller than the average arrival interval.
As FO adjusts the passing order, vehicles from the same
lane may be grouped and pass the intersection together.
For two vehicles consecutively leaving the intersection, they
go to different lanes only if the following two conditions
holds: 1) they come from different
lanes and 2) the last
vehicle in the two has a temporal gap greater than ∆d with
its front vehicle. Hence, the probability that two departure
vehicles are from different lanes is Ps(1)Ps(2)(e−λ1∆d +
e−λ2∆d ), which is smaller than 2Ps(1)Ps(2). The mini-
mum average departure interval is Ps(1)Ps(2)(y1 + y2)∆d +
(cid:2)Ps(1)2 + Ps(2)2 + Ps(1)Ps(2)(2 − y1 − y2)(cid:3) ∆s. The aver-
λ. Condition (21) can be obtained by
age arrival interval is 1
requiring the minimum departure interval be smaller than the
arrival interval.
xixi s d00T1iT2i12345678 d86571432 d0T1i+1T2i+1 Given the definition in (25),
Mi = (cid:98)gi(0) +
Ii = (cid:98)gi(0) +
(cid:2)eλi∗ ∆d − 1(cid:3) +(cid:98)gi(∆d),
(cid:2)1 − e
−λi∆d(cid:3) +(cid:98)gi(∆d)e
ci
λi∗
ci
λi
(55)
−λ∆d. (56)
Moreover, the probability should add up to one,
M1 + M2 = 1.
(57)
Solving (52) to (57), we conclude that Mi = λi
satisfies (50).
λ and ci
Corollary 8 (Steady State Vehicle Delay under FO). The
steady state vehicle delay under FO has the distribution
Pd(t) =
c1
λ2
c2
λ1
+
eλ1t +
c2
(e
2λ1λ2
eλ2t +
λ2
−λ1t) +
−λt − e
(1 − e
c1
(e
λ1y2
−λt)
−λt − e
(58)
−λ2t),
E(−λ),
(59)
E(d) =
λ2y1
with expected delay
c2
λ1E(λ1) +
c2
−
λ2y1E(−λ1) −
c1
λ1y2 −
λ2y1
+
c1
λ2E(λ2)
c1
λ1y2E(−λ2)
(cid:19)
2λ1λ2
λ2
+
(cid:18) c2
i=1,2 Ps(i)(cid:82) ∞
where E(·) follows (46).
Proof. By (11), the steady state distribution of delay satisfies
0 [gi(t+x)+gi∗ (t+x−∆d)+gi∗ (x−
t + ∆d)]pxdx. Using the result from Proposition 7, the steady
state distribution of the vehicle delay satisfies (58). It is easy
to verify Pd(0) = c2
λ1
pd(t) =(cid:80)
The expected mean E(d) =(cid:82) ∆d
=(cid:98)g1(0) +(cid:98)g2(0) and Pd(∆d) = 1.
tdPd(t) satisfies (59).
+ c1
λ2
0
Corollary 8 implies that the distribution of vehicle delay in
FO no longer equals the sum of traffic delay in all lanes. In
FIFO, the two equal by Corollary 5.
IV. ANALYSIS
This section discusses how delay is affected by traffic den-
sity λ, density ratio r := λ1/λ2, passing order (FIFO or FO),
and temporal gap ∆d. ∆s = 0 is assumed. In particular, we
evaluate the probability of zero delay Pd(0) =(cid:98)g1(0) +(cid:98)g2(0)
in Fig. 7, expected delay E(d) in Fig. 8, and steady state
distribution of delay Pd(t) in Fig. 9. The curves are from direct
analysis. Approximation (41) is used for FIFO. The accuracy
of the analytical solutions is verified by EDS in Fig. 8c.
A. Delay and Traffic Density
In general,
larger traffic density results in larger delay.
According to Fig. 7a, the probability of zero delay Pd(0) drops
when the traffic density goes up. In FIFO, it drops linearly and
reaches zero when the equality in (21) holds, where
)
0
(
d
P
y
t
i
l
i
b
a
b
o
r
P
)
0
(
d
P
y
t
i
l
i
b
a
b
o
r
P
)
0
(
d
P
y
t
i
l
i
b
a
b
o
r
P
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
2
3
1
Traffic density λ [s−1]
4
(a) Fix ∆d = 2 s.
0.2
0.4
0.6
0.8
1
Traffic density ratio r
(b) Fix λ = 1 s−1.
8
r = 1, FIFO
r = 0.6, FIFO
r = 0.4, FIFO
r = 0.2, FIFO
r = 0.1, FIFO
r = 1, FO
r = 0.6, FO
r = 0.4, FO
r = 0.2, FO
r = 0.1, FO
∆d = 0.2, FIFO
∆d = 1, FIFO
∆d = 2, FIFO
∆d = 4, FIFO
∆d = 8, FIFO
∆d = 0.2, FO
∆d = 1, FO
∆d = 2, FO
∆d = 4, FO
∆d = 8, FO
λ = 0.1, FIFO
λ = 0.5, FIFO
λ = 1, FIFO
λ = 2, FIFO
λ = 4, FIFO
λ = 0.1, FO
λ = 0.5, FO
λ = 1, FO
λ = 2, FO
λ = 4, FO
2
1
3
Temporal Gap ∆d [s]
4
(c) Fix r = 0.5.
Fig. 7: The probability of zero delay Pd(0).
In FO, Pd(0) drops with decreasing rate. According to Fig. 8a,
the expected delay E(d) grows with the traffic density λ. In
FIFO, it grows exponentially with λ, and goes to infinity when
λ approaches (60). In FO, it grows with decreasing rate when
λ increases. Fig. 9a illustrates the distribution of delay for
λ ∈ {0.1, 0.5, 1, 2, 4}, ∆d = 2, and r = 0.5. The distribution
does not converge for λ > 1.125 in FIFO, while it always
converge in FO. It is easy to verify that the necessary condition
(47) is always satisfied when ∆s = 0.
B. Delay and Density Ratio
λ =
(1 + r)2
2∆dr
.
(60)
In general, there are more delays when the traffic is more
balanced. According to Fig. 7b, Pd(0) drops with decreasing
]
s
[
)
d
(
E
y
a
l
e
d
d
e
t
c
e
p
x
E
]
s
[
)
d
(
E
y
a
l
e
d
d
e
t
c
e
p
x
E
]
s
[
)
d
(
E
y
a
l
e
d
d
e
t
c
e
p
x
E
2
1.5
1
0.5
0
0
6
4
2
0
0
2
1.5
1
0.5
0
0
2
3
1
Traffic density λ [s−1]
4
(a) Fix ∆d = 2 s.
0.2
0.4
0.6
0.8
1
Traffic density ratio r
(b) Fix λ = 1 s−1.
r = 1, FIFO
r = 0.6, FIFO
r = 0.4, FIFO
r = 0.2, FIFO
r = 0.1, FIFO
r = 1, FO
r = 0.6, FO
r = 0.4, FO
r = 0.2, FO
r = 0.1, FO
∆d = 0.2, FIFO
∆d = 1, FIFO
∆d = 2, FIFO
∆d = 4, FIFO
∆d = 8, FIFO
∆d = 0.2, FO
∆d = 1, FO
∆d = 2, FO
∆d = 4, FO
∆d = 8, FO
∆d = 2, FIFO
EDS
∆d = 2, FO
EDS
λ = 0.1, FIFO
λ = 0.5, FIFO
λ = 0.1, FIFO
λ = 2, FIFO
λ = 4, FIFO
λ = 0.1, FO
λ = 0.5, FO
λ = 1, FO
λ = 2, FO
λ = 4, FO
2
1
3
Temporal gap ∆d [s]
4
(c) Fix r = 0.5.
Fig. 8: The expected delay E(d).
rate when the density ratio approaches 1. In FIFO, it reaches
zero when (60) holds. In FO, Pd(0) is relatively constant for
r > 0.5. According to Fig. 8b, the expected delay E(d) grows
with respect to the density ratio r. In FIFO, the expected delay
grows exponentially with r when there is a solution for r ≤ 1
in (60) for fixed λ and ∆d, e.g.,
λ∆d ≥ min
r∈(0,1]
(1 + r)2
2r
= 2.
(61)
)
t
(
d
P
y
t
i
l
i
b
a
b
o
r
p
e
v
i
t
a
l
u
m
u
C
)
t
(
d
P
y
t
i
l
i
b
a
b
o
r
p
e
v
i
t
a
l
u
m
u
C
)
t
(
d
P
y
t
i
l
i
b
a
b
o
r
p
e
v
i
t
a
l
u
m
u
C
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
1
0.8
0.6
0.4
0.2
0
0
9
λ = 0.1, FIFO
λ = 0.5, FIFO
λ = 1, FIFO
λ = 0.1, FO
λ = 0.5, FO
λ = 1, FO
λ = 2, FO
λ = 4, FO
r = 0.1, FIFO
r = 0.2, FIFO
r = 0.4, FIFO
r = 0.6, FIFO
r = 1, FIFO
r = 0.1, FO
r = 0.2, FO
r = 0.4, FO
r = 0.6, FO
r = 1, FO
∆d = 0.2, FIFO
∆d = 1, FIFO
∆d = 2, FIFO
∆d = 0.2, FO
∆d = 1, FO
∆d = 2, FO
∆d = 4, FO
∆d = 8, FO
2
Delay t [s]
4
(a) Fix ∆d = 2 s and r = 0.5.
2
Delay t [s]
4
(b) Fix ∆d = 2 s and λ = 1 s−1.
2
Delay t [s]
4
(c) Fix λ=1 s−1 and r = 0.5.
Fig. 9: Distribution of steady state vehicle delay Pd(t).
The expected delay grows with decreasing rate when there
is no solution for r ≤ 1 in (60), i.e., λ∆d < 2. In FO, the
expected delay grows in decreasing rate when r approaches 1.
When λ∆d is small, the expected delay in FIFO is close to
the expected delay in FO.
C. Delay and Passing Order
For all scenarios in Fig. 7, Fig. 8, and Fig. 9, FO results in
smaller delay than FIFO. The advantage of FO is due to the
fact that the passing order is adaptable to real time scenarios.
They have similar performances when either λ, ∆d, or r is
small. In those cases, the order determined by FO is close to
the order in FIFO. Moreover, it is worth noting that the delay
distribution in Fig. 9 is not computed for a single vehicle,
but for all vehicles on average. Such average delay does not
exceed ∆d in FO, but it is possible for individual vehicles
10
to have delay greater than ∆d. Though FO is efficient in
the sense that it minimizes delay, it sacrifices fairness by not
obeying the passing order determined by the desired passing
time. As a consequence, certain vehicles may experience larger
delay compared to that in the FIFO case. The tradeoff between
fairness and efficiency in different policies will be studied in
the future.
D. Delay and Temporal Gap
In general, a larger temporal gap results in larger delay.
According to Fig. 7c, Pd(0) drops when the temporal gap ∆d
increases. In FIFO, it drops linearly and reaches zero when the
equality in (60) holds. In FIFO, it drops with decreasing rate.
According to Fig. 8c, the expected delay E(d) grows with
respect to the temporal gap ∆d. In FIFO, the expected delay
grows exponentially. In FO, it eventually reaches a constant
growth rate. The temporal gap is a design parameter in vehicle
policies, which is affected by the uncertainty in perceptions.
When there are larger uncertainties in perception, in order
to stay safe, vehicles tend to maintain larger gaps to other
vehicles. The trade-off between safety and efficiency under
imperfect perception will be studied in the future.
V. CONCLUSION
This paper presented a new approach to perform delay anal-
ysis for unmanaged intersections in an event-driven stochastic
model. The model considered the traffic delay at an intersec-
tion as an event-driven stochastic process, whose dynamics
encoded equilibria resulted from microscopic multi-vehicle
interactions. With the model, the distribution of delay can
be obtained through either direct analysis or event-driven
simulation. In particular, this paper performed detailed anal-
yses for a two-lane intersection under two different classes
of policies corresponding to two different passing orders.
The convergence of the distribution of delay and the steady
state delay were derived through direct analysis. The rela-
tionships between traffic delay and multiple factors such as
traffic flow density, unevenness of traffic flows, temporal gaps
between two consecutive vehicles, and the passing order were
discussed. In the future, such analysis will be extended to
more complex vehicle policies, more complex road topologies,
multiple intersections, and heterogeneous traffic scenarios.
REFERENCES
[1] T. V. Mathew, "Signalized intersection delay models," Lecture notes in
Traffic Engineering and Management, 2014.
[2] J. Xi, W. Li, S. Wang, and C. Wang, "An approach to an intersection
traffic delay study based on shift-share analysis," Information, vol. 6,
no. 2, pp. 246–257, 2015.
[3] Y. Jiang, S. Li, and K. Q. Zhu, "Traffic delay studies at signalized
intersections with global positioning system devices," ITE Journal,
vol. 75, no. 8, pp. 30–39, 2005.
[4] M. VanMiddlesworth, K. Dresner, and P. Stone, "Replacing the stop
sign: Unmanaged intersection control for autonomous vehicles," in
International Joint Conference on Autonomous Agents and Multiagent
Systems, vol. 3, pp. 1413–1416, IFAAMS, 2008.
[5] V. Savic, E. M. Schiller, and M. Papatriantafilou, "Distributed algorithm
for collision avoidance at road intersections in the presence of communi-
cation failures," in Intelligent Vehicles Symposium (IV), pp. 1005–1012,
IEEE, 2017.
[6] S. Azimi, G. Bhatia, R. Rajkumar, and P. Mudalige, "Reliable intersec-
tion protocols using vehicular networks," in International Conference
on Cyber-Physical Systems, ICCPS '13, pp. 1–10, ACM, 2013.
[7] C. Liu, C. W. Lin, S. Shiraishi, and M. Tomizuka, "Distributed conflict
resolution for connected autonomous vehicles," IEEE Transactions on
Intelligent Vehicles, vol. 3, no. 1, pp. 18–29, 2018.
[8] P. Gora and I. Rub, "Traffic models for self-driving connected cars,"
Transportation Research Procedia, vol. 14, pp. 2207 – 2216, 2016.
Transport Research Arena TRA2016.
[9] M. Treiber and A. Kesting, "An open-source microscopic traffic simu-
lator," IEEE Intelligent Transportation Systems Magazine, vol. 2, no. 3,
pp. 6–13, 2010.
[10] J. Barcel´o, J. Casas, J. L. Ferrer, and D. Garc´ıa, Modelling Advanced
Transport Telematic Applications with Microscopic Simulators: The
Case of AIMSUN2, pp. 205–221. Springer, 1999.
[11] M. Fellendorf and P. Vortisch, Microscopic Traffic Flow Simulator
VISSIM, pp. 63–93. Springer, 2010.
[12] S. P. Hoogendoorn and P. H. L. Bovy, "State-of-the-art of vehicular
traffic flow modelling," Proceedings of the Institution of Mechanical
Engineers, Part I: Journal of Systems and Control Engineering, vol. 215,
no. 4, pp. 283–303, 2001.
[13] R. Corthout, G. Flotterod, F. Viti, and C. M. Tamp`ere, "Non-unique flows
in macroscopic first-order intersection models," Transportation Research
Part B: Methodological, vol. 46, no. 3, pp. 343 – 359, 2012.
[14] G. Flotterod and J. Rohde, "Operational macroscopic modeling of
complex urban road intersections," Transportation Research Part B:
Methodological, vol. 45, no. 6, pp. 903 – 922, 2011.
[15] C. Liu and M. J. Kochenderfer, "Analytically Modeling Unmanaged
Intersections with Microscopic Vehicle Interactions," arXiv:1804.04746,
Apr. 2018.
[16] F. Altch´e, X. Qian, and A. de La Fortelle, "Time-optimal coordination
of mobile robots along specified paths," in International Conference on
Intelligent Robots and Systems (IROS), pp. 5020–5026, IEEE, 2016.
[17] X. Qian, F. Altch´e, J. Gr´egoire, and A. de La Fortelle, "Autonomous
intersection management systems: criteria, implementation and evalua-
tion," IET Intelligent Transport Systems, vol. 11, no. 3, pp. 182–189,
2017.
[18] C. Gardiner, Stochastic methods, vol. 4. Springer, 2009.
[19] W. Michiels and S.-I. Niculescu, Spectral Properties of Linear Time-
Delay Systems, ch. 1, pp. 3–31. SIAM, 2007.
|
1802.02218 | 1 | 1802 | 2018-02-06T20:56:51 | On the Preliminary Investigation of Selfish Mining Strategy with Multiple Selfish Miners | [
"cs.MA",
"cs.CR",
"cs.GT"
] | Eyal and Sirer's selfish mining strategy has demonstrated that Bitcoin system is not secure even if 50% of total mining power is held by altruistic miners. Since then, researchers have been investigating either to improve the efficiency of selfish mining, or how to defend against it, typically in a single selfish miner setting. Yet there is no research on a selfish mining strategies concurrently used by multiple miners in the system. The effectiveness of such selfish mining strategies and their required mining power under such multiple selfish miners setting remains unknown.
In this paper, a preliminary investigation and our findings of selfish mining strategy used by multiple miners are reported. In addition, the conventional model of Bitcoin system is slightly redesigned to tackle its shortcoming: namely, a concurrency of individual mining processes. Although a theoretical analysis of selfish mining strategy under this setting is yet to be established, the current findings based on simulations is promising and of great interest. In particular, our work shows that a lower bound of power threshold required for selfish mining strategy decreases in proportion to a number of selfish miners. Moreover, there exist Nash equilibria where all selfish miners in the system do not change to an honest mining strategy and simultaneously earn their unfair amount of mining reward given that they equally possess sufficiently large mining power. Lastly, our new model yields a power threshold for mounting selfish mining strategy slightly greater than one from the conventional model. | cs.MA | cs |
On the Preliminary Investigation of Selfish
Mining Strategy with Multiple Selfish Miners
Tin Leelavimolsilp1, Long Tran-Thanh2, and Sebastian Stein3
Electronics and Computer Science, University of Southampton, UK
Email: [email protected], [email protected], [email protected]
Abstract
Eyal and Sirer's selfish mining strategy has demonstrated that Bit-
coin system is not secure even if 50% of total mining power is held
by altruistic miners [2]. Since then, researchers have been investi-
gating either to improve the efficiency of selfish mining, or how to
defend against it, typically in a single selfish miner setting. Yet there
is no research on a selfish mining strategies concurrently used by mul-
tiple miners in the system. The effectiveness of such selfish mining
strategies and their required mining power under such multiple selfish
miners setting remains unknown.
In this paper, a preliminary investigation and our findings of self-
ish mining strategy used by multiple miners are reported. In addi-
tion, the conventional model of Bitcoin system is slightly redesigned
to tackle its shortcoming: namely, a concurrency of individual mining
processes. Although a theoretical analysis of selfish mining strategy
under this setting is yet to be established, the current findings based
on simulations is promising and of great interest. In particular, our
work shows that a lower bound of power threshold required for selfish
mining strategy decreases in proportion to a number of selfish min-
ers. Moreover, there exist Nash equilibria where all selfish miners in
the system do not change to an honest mining strategy and simulta-
neously earn their unfair amount of mining reward given that they
equally possess sufficiently large mining power. Lastly, our new model
yields a power threshold for mounting selfish mining strategy slightly
greater than one from the conventional model.
1
1
Introduction
Originally invented by Nakamoto [5], a blockchain is used to securely record
a ledger of Bitcoin payment transactions amongst Internet users. The great
success of blockchain and Bitcoin is based on an application of cryptographic
puzzle, namely Proof-of-Work, and an economic incentive for miners, whom
are an underlying workforce of Bitcoin system. In other words, anyone on
the Internet with a sufficient amount of computational power can be a miner
and solve the cryptographic puzzle to earn Bitcoin.
Due to Nakamoto's analysis, it has been widely believed that the Bitcoin
system will remain secure as long as at least half of the total mining power
are held by non-malicious miners, who honestly mine blocks according to the
blockchain protocol [5]. In particular, the analysis of an attacker succeeding
in double-spending his Bitcoin was modelled as a 1-dimensional random walk
with an infinite time and one absorbing bound. Intuitively, the attacker will
succeed if his mining power is more than half of the total power in the system.
However, a selfish mining strategy allows a malicious miner who possesses
at least a third of total mining power to gain more than his fair share and
could consequently disrupt Bitcoin system. As first demonstrated in Eyal
and Sirer's work [2], this strategy which secretly builds a private blockchain
and strategically releases blocks causes honest miners to mine on a block that
will eventually be replaced by selfish miner's. As a result, honest miners will
receive less mining profit, and might stop mining or even participate in selfish
mining together with the first selfish miner. In any case, a percentage of total
mining power held by the selfish miner will increase and further improve the
effectiveness of selfish mining strategy. Therefore, Bitcoin system is not safe
against selfish mining strategy even if a half of the total mining power is held
by honest miners.
Despite of such importance, there is yet no research on a selfish min-
ing strategy simultaneously used by multiple miners. To the best of our
knowledge, most works so far have only focused on one malicious miner us-
ing selfish mining strategy and the others employing honest mining strategy
[2, 3, 7, 6, 8]. However, a number of miners in Bitcoin system could use the
selfish mining strategy at the same time. Whether the selfish mining strategy
is still effective in such situation has not yet been investigated.
In addition, none of previous works so far has considered a concurrency
resulted from individually mining and could misestimate a power threshold
of selfish mining strategy as a result. In other words, a mining process in
2
Bitcoin system has been widely modelled as a single entity randomly as-
signing new blocks to their owners according to their mining powers [2, 8].
On the contrary, miners in practical individually mine on their locally-stored
blockchain and consequently create new blocks independently. Therefore, the
widely used model might not well represent Bitcoin system and thus previous
findings of the power threshold might be lower or greater than the actual one.
For these reasons, we have started a preliminary investigation into an
effectiveness of selfish mining strategy employed by multiple miners with our
new model of Bitcoin system. In essence, our model better reflects the con-
currency of individually mining and further assumes that there is always at
least an honest miner in the system. Such honest miner collectively rep-
resents altruistic miners in Bitcoin community who believe in a long-term
economic impact of a security attack and consequently adhere to the original
blockchain protocol [1].
Despite a lack of theoretical analysis and some limitations, our findings is
useful and provides an insight toward an effectiveness of selfish mining strat-
egy used by multiple miners. We also consider our work to be a complement
to Kiayias's game-theoretical analysis of blockchain mining [4]. In particular,
our work provides a number of contributions as follows.
1. Under a system where there are multiple selfish miners, the least amount
of mining power that could allow a selfish miner to earn his unfair
amount of mining reward becomes lesser in proportion to a number of
selfish miners in the system.
2. Given a specific power configuration or a specific allocation of mining
power, there exist Nash equilibria where multiple miners use a selfish
mining strategy and simultaneously gain their mining reward greater
than their percentage of total mining power.
3. From a perspective of security, an amount of mining power required to
completely secure against a selfish mining strategy remains the same
regardless of a number of selfish miners in the system.
4. Due to a simplified model of Bitcoin system that are widely used in
literatures, all estimated mining power to effectively mount a selfish
mining strategy and other attacks so far could be slightly underesti-
mated.
3
In the rest of this paper, our work are presented as follows. Firstly a
literature related to our study is reviewed. Then we briefly describe our
model and specifically point out any difference to the conventional model
of Bitcoin system. Subsequently our simulation settings and their results
are demonstrated and discussed. Finally we concludes this paper with our
findings and future work.
2 Related Work
After Eyal and Sirer's seminal work [2], a number of studies have advanced
the research of selfish mining strategy. For example, Gobel and his colleagues
extended the original study by using a spatial Poisson point process to model
a network of miners to better account for an effect of network delay [3]. Their
findings showed a relatively high number of blocks in a blockchain and conse-
quently a great amount of mining reward for every miner if there is no miner
the using selfish mining strategy. Nonetheless, the selfish mining strategy
still allows a miner to gain his mining reward greater than his percentage of
total mining power once his mining power is sufficiently large.
On the one hand, some studies have improved the original selfish mining
strategy to be more effective and consequently gain higher amount of mining
reward than the original one [7, 6]. In addition to optimising the original
selfish mining strategy, a combination with other attacks such as an eclipse
attack could be used to further increase its success rate [6]. As a result, a
safety level of mining power required for Bitcoin system to be secure against
the selfish mining strategy became lower than 1/3 of the total mining power.
On the other hand, a number of researches proposed various methods
that improve blockchain protocol to better resist the selfish mining strategy,
though they were difficult to carry out in practical [8]. In particular, these
methods increased a power threshold or an amount of mining power required
for successfully mounting selfish mining strategy to some degree. However,
their methods including Zhang and Preneel's required a good coordination
amongst the majority of miners to adopt them at a specific point of time to
prevent undesirable forking of the blockchain. Consequently, they were hard
to implement in Bitcoin system.
Whilst a game-theoretical analysis could provide a great insight into
miner's decision whether to mount a selfish mining strategy, game theory
is based on an assumption of strong rationality and therefore might not re-
4
flect altruistic miners in Bitcoin system. With regard to mining strategies
that strategically publish their hidden blocks, Kiayias and his colleagues
demonstrated that every miner will use an honest mining strategy if no one
possesses mining power greater than 30.8% of the total mining power [4].
However, their model was based on an assumption of rational miners; in
other words, miners always change their mining strategy to the most benefi-
cial one. Such assumption did not reflect altruistic miners, who adhere to the
original mining software due to a fear of long-term economic impact resulted
from any security attack.
To the best of our knowledge, there is no research that considers multiple
miners simultaneously employing a selfish mining strategy. In practical, a
selfish miner might not be able to immediately switch back and forth between
an honest mining strategy and the selfish mining strategy to best maximise
his mining reward. In other words, they might continuously use the selfish
mining strategy for some periods even if it is less profitable than the honest
mining strategy. A possible cause of selfish miner's irrational decision could
be a lack of information; e.g. a current distribution of mining powers which
could be heavily fluctuating.
Furthermore, all works so far have modelled Bitcoin system in a slightly
unrealistic manner and could consequently misestimate a power threshold
required for the selfish mining strategy. That is, their models assumed that
there is always one miner successfully generating a block at each period.
However, miners in practical individually and concurrently mine blocks on
their locally-stored blockchain. As a result, there could be more than one
block generated in each period. Therefore, the actual power threshold might
be greater or lower than the power threshold that has been known so far.
3 Our Model of Bitcoin System
To tackle the shortcomings previously mentioned, the widely used model of
Bitcoin system is redesigned to reflect miners concurrently mining on their
locally-stored blockchains.
In addition, an honest miner who collectively
represents altruistic miners in Bitcoin system is included into our model. As
a consequence, our model better represents Bitcoin system.
In particular,
our model is described as follows.
• Each miner is denoted by i ∈ {1, 2, ..., N}, in which every miner i ∈
{1, 2, ..., N − 1} always uses the original selfish mining strategy [2] and
5
the N-th miner uses an honest mining strategy [5]. For the sake of
simplicity, all altruistic miners are treated as one honest miner and we
focus on the effect of varying number of selfish miners and their mining
powers.
• For each miner, his mining process on locally-stored blockchain is a
Bernoulli trial with a probability of successfully creating a block pi =
mi × d, where:
– mi ∈ [0, 1] denotes a relative mining power or a proportion of
total mining power that miner i possesses, and we refer to an
allocation of mining power to every miner as a power configuration
M = {mii ∈ {1, 2, ..., N}} in which PN
i=1 mi = 1;
– d ∈ [0, 1] represents a difficulty level of creating blocks, i.e., a
proportion of nonces that yield a hash value lower than a target
value given that all miners create the same block. Analogously, a
high value of d represents Bitcoin system with a high target value
and vice versa for a low value of d.
• Utility function Ui of each miner is defined as a relative reward or a
proportion of his blocks in the longest blockchain and thus Ui ∈ [0, 1] .
• Every miner is assumed to be directly connected to all other miners
and there is no communication delay in the network. Consequently,
any broadcast message regarding a discovery of new blocks is instanta-
neously received. However, if there is a number of messages simultane-
ously sent to the same recipient, an ordering of messages that will be
sequentially received is randomly chosen and thus all messages have an
equal probability of being first received.
Note that our model neither regards selfish miner's network capability as
Eyal and Sirer's model [2] nor favours altruistic miner's one. Since network
capability and communication delay could greatly affect the effectiveness
of selfish mining strategy, a study involving them should be systematically
carried out and thus it is considered as a next step of this work.
Finally, it should be pointed out that a major difference between our
model and the conventional model is a mining process.
In particular, the
mining process in the conventional model was modelled as a global entity
which generates new blocks and randomly assign them to miners according
6
to their mining powers. On the other hand, our model mimics all of individual
mining processes and allows a number of miners to simultaneously generate
their blocks. For further details of the conventional model, readers should
refer to Eyal and Sirer's work [2].
4 Simulation and Results
A series of discrete event simulations were carried out to observe an effective-
ness of selfish mining in three different settings: namely one selfish miner, two
selfish miners, and three selfish miners. In particular, a relative mining power
of selfish miner is varied by 0.01. Furthermore, there is one honest miner who
has the rest of mining power left from selfish miners'. Each simulation which
runs for 200,000 timesteps is then repeated 100 times to calculate an average
relative reward E [Ui] and its 95% confidence interval.
It is worth mentioning that parameter d of our model is fixed to 0.5 in
all simulations. Since d = 0.5 allows blocks to be frequently generated, the
resulted blockchain will be sufficiently long to extract meaningful results.
Together with 200,000 timesteps, a length of the resulted blockchain varies
from 40,000 - 90,000 blocks, which is analogous to 11 - 13 months' worth of
Bitcoin's blockchain.
In the following, simulation results will be shown and described for each
setting. Except for the case of three selfish miners, we compare simulation
results of our model and the conventional model's. Furthermore, a simulation
result of our model when there is no selfish miner is included to demonstrate
a baseline behaviour of our model.
4.1 No Selfish Miner
In contrast to the conventional model, our model shows an interesting aspect
of the actual mining process: that is, a miner with the greatest amount of
mining power was given a mining reward slightly larger than his relative
power and vice versa for the other miners. As shown in figure 1, an honest
miner with a relative mining power of 0.4 received an amount of relative
reward 0.414, whereas all other miners (each with a power of 0.3) got slightly
lesser than their relative powers.
Nevertheless, our model captures one of the main characteristics of Bit-
coin mining. That is, a miner receives his mining reward in proportion to his
7
Figure 1: A line plot demonstrating a convergence of miners' reward in our model without
any selfish miner.
mining power.
4.2 One Selfish Miner
As demonstrated in figure 2, a selfish miner who has mining power exceeding a
threshold in both models gains a relative reward more than his relative power,
however our model yields the threshold slightly higher than the conventional
model's. In particular, a selfish miner requires a relative power of at least 0.34
to effectively use selfish mining strategy in the conventional model, whereas
a power of at least 0.38 is needed in our model.
Such difference could be due to our model's concurrent mining processes,
which further results in a greater demand of mining power for a selfish mining
strategy to successfully create the longest blockchain.
4.3 Two Selfish Miners
In contrast to the previous setting, a power threshold of selfish mining strat-
egy under this setting is comparatively low. As shown in figure 3a, a power
configuration that has the least amount of selfish miner 1's power yet still
allows him to gain his unfair amount of reward is M = {0.29, m2, m3} where
8
Figure 2: A line plot comparing selfish miner's average reward and its 95% confidence
interval between the conventional model and our model in one-selfish-miner setting. A
black dashed line indicates an amount of reward the selfish miner would get if he had
followed an honest mining strategy.
m2 ∈ [0.2, 0.31] and m3 has the rest. On the other hand, other power configu-
rations M ′ = {0.29, m2, m3} where m2 /∈ [0.2, 0.31] do not yield extra reward
for selfish miner 1. A plausible cause for the latter might be an insufficient
amount of selfish miner's power to frequently win against either an honest
miner or the another selfish miner in a race to create the longest blockchain.
Surprisingly, both selfish miners can simultaneously gain their extra re-
wards in some specific power configurations. As demonstrated in figure 4a
and 5b, a power configuration M = {m1, m2, m3} where m1 ∈ [0.29, 0.49]
and m1 = m2 allows both selfish miners to earn their relative rewards higher
than their relative powers. However, their reward amounts become unstable
for any power configuration where both selfish miners' power equally exceeds
0.33. Specifically, their reward amounts could be lesser than their relative
powers if they equally possess relative power greater than or equal to 0.41.
It is also worth mentioning that (i) a system under this setting is com-
pletely vulnerable to selfish mining strategy for any power configuration
where an honest miner has relative mining power less than 0.44, (ii) as por-
trayed in figure 3b, there are some power configurations M = {m1, m2, m3},
m3 ∈ [0.44, 0.61] that allow an honest miner to retain his reward in propor-
9
(a)
(b)
(c)
(d)
Figure 3: Heat maps of selfish miner 1's average reward (a), honest miner's average reward
(b), and their 95% confidence intervals respectively (c,d) in our model with two selfish
miners. In each plot of selfish miner 1's (a,c), a black line separates power configurations
that yield mining reward greater than his power to the right side and vice versa to the left
side of the plot. Similarly, a black solid line in each plot of honest miner's (b,d) separates
power configurations that yield reward at least equal to his relative power to the lower
left side and vice versa to the upper right side of the plot. Also, a dashed line separates
power configurations where there is at least 1 selfish miner earning his unfair amount of
reward to the upper right side and vice versa to the lower left side of the plot. Note
that an intersected area between these two lines indicates power configurations where the
honest miner receives his fair share of reward and at least one selfish miner earns his extra
amount of reward as well.
10
(a)
(b)
Figure 4: Contour plots comparing power configurations that a selfish mining strategy
is effective between our model (a) and the conventional model (b) with 2 selfish miners.
A gray area indicates configurations that only 1 selfish miner gains his unfair amount
of mining reward, whereas a black area indicates configurations that both selfish miners
simultaneously earn their extra amounts of mining reward.
tion to his power whilst also let one selfish miner earn his unfair amount of
reward, and (iii) a power threshold to effectively use selfish mining strategy
under the conventional model is slightly underestimated in comparison to
our model's. A simulation result of the latter is demonstrated in appendix
A.
4.4 Three Selfish Miners
Due to our limitation of graphically presenting the whole results under this
setting, only maximum reward and minimum reward of selfish miner 1 are
shown in figure 6, where we vary selfish miner 1's relative power and the
other selfish miners' combining power. Particular results of our interest are
described in this section.
Unexpectedly, a power threshold that is required for selfish mining strat-
egy becomes even lower in comparison to the previous setting. Specifically,
a selfish miner needs a relative mining power of at very least 0.23 to gain
his unfair amount of mining reward. Corresponding power configurations are
listed in table 1 and a visual overview of selfish miner 1's reward is demon-
strated in figure 6. Other configurations M = {0.23, m2, m3, m4} where m2
11
(a)
(b)
(c)
(d)
Figure 5: Line plots demonstrating miners' average reward and their 95% confidence
intervals of a power configuration where all selfish miners possess an equal amount of
mining power. All upper plots (a,b) are simulation results of our model with 2 selfish
miners, whilst all lower plots (c,d) are results of our model with 3 selfish miners. All left
plots (a,c) are simulation results with 100 repetitions each, whereas all right plots (b,d)
are results with 200 repetitions each.
12
Table 1: List of power configurations in our model with 3 selfish miners where selfish miner
1 has the least amount of mining power but still earns his unfair amount of mining reward.
m1 m2 Range of m3
0.23
0.15
0.16
0.17
0.18
0.19
0.20
0.21
0.22
0.23
0.24
0.25
[0.21, 0.23]
[0.18, 0.24]
[0.17, 0.24]
[0.16, 0.25]
[0.16, 0.25]
[0.16, 0.25]
[0.15, 0.25]
[0.15, 0.24]
[0.15, 0.24]
[0.16, 0.23]
[0.18, 0.21]
and m3 are outside of the reported range do not yield extra reward for selfish
miner 1. Under such configurations, it is possible that selfish miner 1 does
not have enough mining power to build the longest blockchain.
Moreover, all selfish miners can simultaneously earn their unfair amounts
of mining reward under some particular power configurations. As portrayed
in figure 5d, a power configuration M = {m1, m2, m3, m4} in which m1 ∈
[0.23, 0.33], m1 = m2 = m3, and m4 has the rest allows all selfish miners to
gain their relative rewards greater than their powers. Although, their reward
amounts are unsteady when their relative powers are more than or equal to
0.25. Furthermore, their relative rewards could be lower than their relative
power once they possess power greater than or equal to 0.26. Note that there
are also power configurations where only two selfish miners gain their extra
amounts of reward but the another does not.
In addition, we noticed that (i) a system under this setting where an hon-
est miner has relative power lower than 0.34 is always prone to selfish mining
regardless of any selfish miner's power configuration, and (ii) there exist
some power configurations M = {m1, m2, m3, m4} where m4 ∈ [0.32, 0.61]
such that an honest miner still earns at least his fair amount of mining re-
ward in proportion to his mining power whilst at least one selfish miner gains
his relative reward more than his relative power.
Although we do not have a simulation result of the conventional model
with 3 selfish miners, it can be expected that a power threshold of selfish
13
Table 2: A summary of selfish mining strategy's power threshold, a range of selfish miner's
equal powers constituting Nash equilibria, and a safety level of honest miner's power to
secure Bitcoin system against selfish mining strategy with respect to a number of selfish
miners in our model.
Number of
Selfish Miners
Power Threshold Range of Selfish Miners' Equal Power
Safety Level
Lower
Bound
Upper
Bound
Stable
Unstable
Nash Equilibria
Nash Equilibria
Lower
Bound
Upper
Bound
1
2
3
0.38*
0.29
0.23
0.38*
0.50
0.50
-
[0.29, 0.40]
[0.23, 0.25]
-
[0.41, 0.49]
[0.26, 0.33]
0.63*
0.44
0.34
0.63*
0.63
0.63
* Due to a low granularity of varying power configurations in our simulation, a relative mining
power 0.38 allows a selfish miner to gain his unfair amount of reward, but a power 0.37 does
not. As a consequence, a safety level becomes 0.63. By a non-linear interpolation, an exact
power threshold for effectively mounting a selfish mining strategy is approximately 0.3786.
mining strategy in our model is slightly greater than one of the conventional
model similarly to the previous setting.
5 Discussion
As summarised in table 2, a lower bound of power threshold or the least
amount of mining power required for selfish mining strategy decreases with
respect to an increase of a number of selfish miners in the system. All power
configurations that result in the lower bound demonstrate the same trait:
a selfish miner has to possess mining power large enough to frequently win
against an honest miner and other selfish miners in a race to create the
longest blockchain.
Generally speaking, the more power a selfish miner has beyond the lower
bound of power threshold, the higher possibility he will earn his unfair
amount of mining reward. That is, the number of power configurations
in which a selfish mining strategy is effective increases in proportion to an
amount of mining power that the selfish miner has.
However, not every power configuration for a specific amount of self-
ish miner's power between the lower bound and the upper bound of power
threshold always allows him to gain his relative reward greater than his rel-
ative power. In other words, a selfish mining strategy might not be effective
if there is another miner who possesses mining power to a certain degree
greater than the selfish miner himself. As a consequence, a low-power miner
14
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
Figure 6: Heat maps of selfish miner 1's minimum and maximum average reward (a,b),
honest miner's minimum and maximum average reward (c,d), and their 95% confidence
intervals respectively (e-h) in our model with 3 selfish miners. In each plot of selfish miner
1's (a,b,e,f), a black line separates power configurations that yields his relative mining
reward greater than his relative power to the right side and vice versa to the left side of
the plot. Similarly, a black line in each plot of honest miner's (b,d,g,h) separates power
configurations that yields a relative mining reward at least equal to his relative power to
the bottom left side and vice versa to the top right side of the plot.
15
using selfish mining strategy is less likely to frequently create the longest
blockchain and gain his extra mining reward.
Interestingly, there always exists a power configuration where all selfish
miners in the system can simultaneously earn their extra amounts of relative
reward and henceforth it constitutes a Nash equilibrium. Particularly, such
configuration requires all selfish miners to have an equal and sufficiently high
amount of mining power. Since all selfish miners under such configuration
would be worse off switching to an honest mining strategy, a Nash equilibrium
in which all miners do not change to another strategy is formed.
Nevertheless, a power configuration where selfish miners have equally yet
excessively high mining power might not form a stable Nash equilibria. Under
such configuration, all selfish miners have an equal probability of creating a
block at each timestep. However, it is entirely up to chance which selfish
miner consecutively get new blocks and consequently build up his blockchain
longer than the others'. Furthermore, a selfish miner who could not create
the longest blockchain at the previous attempt might be successful in the
subsequent attempt. As such, it results in an unsteady amount of mining
reward which could be lower than their relative mining power. Therefore,
the Nash equilibria that are formed under such configurations are unstable.
From a perspective of security, an amount of mining power to make Bit-
coin system completely secure remains the same. As shown in the table 2, an
upper bound of safety level (or the minimum amount of honest miner's power
that prevents any configuration of selfish miners' power) remains constant re-
gardless of any number of selfish miners in the system. An explanation of
the cause is intuitive: there is only one power configuration that requires an
honest miner to possess a relatively great amount of mining power to prevent
selfish mining, namely a configuration where only one selfish miner possesses
mining power but other selfish miners do not. Clearly such configuration
corresponds to a one-selfish-miner setting (or a case where all selfish miners
pool their mining power and work together against an honest miner), and
hence results in the same upper bound of safety level for all other settings.
On the contrary, a lower bound of safety level (or the least amount of
mining power required to prevent selfish mining in at least one specific power
configuration) decreases in proportion to an increase in a number of selfish
miners. By further observation, a power configuration corresponding to the
lower bound is always one where all selfish miners have power equally but
not high enough to gain their unfair amounts of mining reward. It can be
implied that they waste their mining power by trying to create the longest
16
blockchain which is eventually replaced by honest miner's.
Finally, it is worth mentioning that our model results in a power threshold
slightly greater than one from the conventional model. Though our model of
Bitcoin system might not perfectly reflect the real one, it has demonstrated
an impact of a concurrency of individual mining processes. As a consequence,
all power thresholds of selfish mining strategy and other attacks which were
estimated in other works so far could be slightly underestimated.
6 Conclusion
In this work, a preliminary investigation of selfish mining strategy employed
by multiple miners has been carried out. Based on our empirical results, we
have identified that a lower bound of power threshold required to effectively
use a selfish mining strategy decreases in proportion to a number of self-
ish miners in the system. Nevertheless an upper bound of power threshold
remains constant, namely 0.5, regardless of a number of selfish miners.
Another interesting aspect is an existence of Nash equilibria where a num-
ber of miners use selfish mining strategy and simultaneously gain their rela-
tive mining rewards higher than their relative mining powers. The only re-
quired condition for such outcome is an equal amount of every selfish miner's
mining power which must be sufficiently large to frequently create a private
blockchain longer than honest miner's blockchain. Note that the equilibria
could be unstable if their mining powers are excessively high.
On the other hand, a safety level of mining power to be held by non-
malicious miners and completely secure Bitcoin system against selfish mining
remains the same regardless of a number of selfish miners in Bitcoin system.
Last but not least, whilst our model of Bitcoin system is slightly different
to the widely used model, our findings is still valid and demonstrates that a
power threshold for selfish mining strategy and other security attacks might
be slightly underestimated.
A number of interesting questions remain to be further investigated. As
signified in the original study of selfish mining strategy [2], a network capa-
bility of selfish miner is also an important factor that affects how much this
strategy could be effective. Such aspect will be taken into account in our
future work. Moreover, an optimal selfish mining strategy with respect to
multiple selfish miners, similar to one in Sapirshtein's [7], is yet not known.
With the optimal strategy, it remains to be seen whether our findings is still
17
valid.
A Simulation Result of Conventional Model
in Two-Selfish-Miners Setting
As demonstrated in figure 7, a power configuration that has the least amount
of selfish miner 1's relative power yet allows him to earn his unfair amount
of mining reward is M = {0.27, m2, m3} where m2 ∈ [0.16, 0.28] and m3 has
the rest. Other configurations M ′ = {0.27, m2, m3} where m2 /∈ [0.16, 0.28]
do not make selfish miner 1 gain his relative reward higher than his relative
power.
In addition, a power configuration M = {m1, m2, m3} in which m1 ∈
[0.27, 0.49], m1 = m2 and m3 has the rest allows both selfish miners to
simultaneously gain their unfair amounts of mining reward. Similarly to the
simulation results of our model, their amounts of mining reward are unstable
when both of them equally possess more than 0.33 and could be less than
their amounts of relative power if both equally have power greater than or
equal to 0.43. Due to a limitation of space in this paper, a plot of these
power configurations is omitted.
Finally, a system under this setting is completely defenceless against
selfish mining if an honest miner possesses a relative mining power less
than or equal to 0.47.
In addition, there are some power configurations
M = {m1, m2, m3}, m3 ∈ [0.47, 0.54] where an honest miner still gains his
relative reward no less than his relative power and at least one selfish miner
still earns his unfair amount of reward.
References
[1] Vitalik Buterin.
Selfish mining:
bitcoin network, Nov 2013.
https://bitcoinmagazine.com/articles/selfish-mining-a-25-attack-against-the-bitcoin-network-1383578440/.
Retrieved Jan. 26,
A 25% attack against
the
2018 from
[2] Ittay Eyal and Emin Gun Sirer. Majority is not enough: Bitcoin mining
is vulnerable. In Financial Cryptography and Data Security (FC 2014).
Lecture Notes in Computer Science (including subseries Lecture Notes in
18
(a)
(b)
(c)
(d)
(e)
(f)
Figure 7: Heat maps of selfish miner 1's average reward (a), honest miner's average reward
(b), their 95% confidence intervals respectively (d,e); and line plots demonstrating miners'
average reward and their 95% confidence intervals of a power configuration where all
selfish miners possess an equal amount of mining power (c,f) in the conventional model
with 2 selfish miners. In each plot of selfish miner 1's (a,d), a black line separates power
configurations that yield relative mining reward greater than his relative mining power
to the right side and vice versa to the left side of the plot. Similarly, a black solid line
in each plot of honest miner's (b,e) separates power configurations that yield reward at
least equal to his power to the lower left side and vice versa to the upper right side of the
plot. Similarly, a dashed line separates configurations that at least 1 selfish miner earns
his unfair amount of reward to the upper right side of the plot and vice versa to the lower
left side of the plot. As for line plots, plot (c) is a simulation result with 100 repetitions
each, whereas plot (f) is a simulation result with 200 repetitions.
19
Artificial Intelligence and Lecture Notes in Bioinformatics), volume 8437,
pages 436–454, Berlin, Germany, 2014. Springer-Verlag.
[3] Johannes Gobel, H. Paul Keeler, Anthony E. Krzesinski, and Peter G.
Taylor. Bitcoin blockchain dynamics: The selfish-mine strategy in the
presence of propagation delay. Performance Evaluation, 104:23–41, Oct
2016.
[4] Aggelos Kiayias, Elias Koutsoupias, Maria Kyropoulou, and Yiannis Tse-
lekounis. Blockchain mining games.
In Proceedings of the 2016 ACM
Conference on Economics and Computation, EC '16, pages 365–382, New
York, USA, 2016. ACM Press.
[5] Satoshi Nakamoto. Bitcoin: A peer-to-peer electronic cash system, 2008.
https://bitcoin.org/bitcoin.pdf.
[6] Kartik Nayak, Srijan Kumar, Andrew Miller, and Elaine Shi. Stubborn
mining: Generalizing selfish mining and combining with an eclipse attack.
In 2016 IEEE European Symposium on Security and Privacy, pages 305–
320, Los Alamitos, US, 2016. IEEE Press.
[7] Ayelet Sapirshtein, Yonatan Sompolinsky, and Aviv Zohar. Optimal self-
ish mining strategies in bitcoin.
In Financial Cryptography and Data
Security (FC 2016). Lecture Notes in Computer Science (including sub-
series Lecture Notes in Artificial Intelligence and Lecture Notes in Bioin-
formatics), volume 9603, pages 515–532, Berlin, Germany, 2017. Springer-
Verlag.
[8] Ren Zhang and Bart Preneel. Publish or perish: A backward-compatible
defense against selfish mining in bitcoin. In Topics in Cryptology (CT-
RSA 2017). Lecture Notes in Computer Science (including subseries Lec-
ture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics),
volume 10159, pages 277–292, Cham, Switzerland, 2017. Springer.
20
|
1806.00367 | 1 | 1806 | 2018-05-31T08:26:10 | An Ontology to Support Collective Intelligence in Decentralised Multi-Robot Systems | [
"cs.MA",
"cs.RO"
] | In most multi-robot systems, conditions of the floor, battery and mechanical parts are important and impact cost-efficiency. The costs are generally interpreted through performance times. The relation between performance times andthese factors are not directly derivable, though, performance time has a direct correlation with discharge of batteries. Inroute planning, travel time of an edge is the performance time and may be required to be estimated for multiple times.These estimated travel times are different than heuristics costs as they depict the real states which are impossible toknow from heuristics. This facilitates path planning algorithms to choose the edges with least real travel times or coststo form the path. Nevertheless, a good estimation is dependent on historical data which are close in time. But, there aresituations when all the travel times for one or more edge(s) are not available for the entire duration of operation of theMRS to an individual robot. Then, it is imperative for that robot to gather the necessary travel times from others inthe system as a reference observation. The mechanism of information sharing between one robot to others in the systemhas been devised in a form of a common ontology-based knowledge. This ontology helps to fetch and share informationforming a collective knowledge base facilitating a comprehensive control and planning for the system. This greatly helpsthe MR to estimate travel times more accurately and precisely. Also, accurate estimation affects route planning to bemore precise with reduced cost. The total cost of paths generated through the travel times estimated through sharingis 40% less on average than that of paths generated through travel times without sharing. | cs.MA | cs |
An Ontology to Support Collective Intelligence in Decentralised Multi-Robot
Systems
1Pragna Das is a Researcher in Microelectronics and Electronic Systems Department, Universitat Aut`onoma de Barcelona (UAB), Campus
Pragna Das1, Vincent Hilaire2 and Llu´ıs Ribas-Xirgo3
2Vincent Hilaire is responsible for Multi-Agent Systems team of the SeT Laboratory, Univ Bourgogne Franche-Comt, UTBM, IRTES-SET
3Llu´ıs Ribas-Xirgo is with the Faculty of Microelectronics and Electronic Systems Department, Universitat Aut`onoma de Barcelona (UAB),
EA 7274/IMSI F-90010, Belfort cedex, France
[email protected]
UAB, 08193 Bellaterra, Spain
[email protected]
Campus UAB, 08193 Bellaterra, Spain
[email protected]
Abstract
In most multi-robot systems, conditions of the floor, battery and mechanical parts contribute to the costs incurred in
performances of movements and tasks. The time to complete performance is dependent on all these factors and thus
reflects the costs incurred. The relation between performance times and these factors are not directly derivable, though,
performance time has a direct correlation with discharge of batteries. When movement is a performance, travel time
of an edge is the performance time. When travel times can be estimated to obtain close-to-real values, they become
different than heuristics costs and depict the real states which are impossible to obtain from heuristics. This facilitates
path planning algorithms to choose the edges with least real travel times or costs to form the path. Nevertheless, a good
estimation is dependent on historical data which are close in time. But, there are situations when all the travel times
for one or more edge(s) are not available for the entire duration of operation of the MRS to an individual robot. Then,
it is imperative for that robot to gather the necessary travel times from others in the system as a reference observation.
This work involves devising a mechanism of information sharing between one robot to others in the system in a form of
a common ontology-based knowledge. With the help of this ontology, travel time is obtained by any robot, whenever
necessary, to obtain accurate estimates for itself. These obtained travel times are traveling experiences of other robots
in the system. Still, they can be used to estimate travel time in that robot as model of travel time has an exploration
factor which depicts the change of travel time in that robot. This model uses others' travel time as observation and
self-exploration factor to estimate future travel time. This greatly helps the MR to estimate travel times more accurately
and precisely. The accurate estimation affects route planning to be more precise with reduced cost. The total cost of
paths obtained using travel times estimated through sharing is 40% less on average than that of paths generated through
travel times without sharing.
Keywords: ontology, multi-agent, behavior-based system, multi-robot system, decentralized multi-robot system,
decentralized control, control system, collective intelligence
1. Introduction
The organization control in MRS for manufacturing and
logistics is mostly centralized in the state of the art with
few exceptions like [16, 10, 8, 9] where the architecture
is decentralized, as it is relatively easy than a distributed
control. However, the decentralized architecture in [8] is
focused on only the planning of route to generate feasi-
ble, sub-optimal and collision-free paths for multiple MRs.
Thus, the system architecture is not general enough to
handle different kinds of control and planning functions.
In [10], the linear dynamic model is generated for a spe-
cific task of collectively transporting load in automated
factories. The sliding mode controller is provided through
a non-linear terms with bounds. This kind of stochas-
tic control using dynamic model of the robot is useful in
simple cases where the controller depends on minimal in-
formation which is available to the robot, unaware of the
dynamics of the environment.
On the other hand, there are various investigations con-
ducted using partially observable Markov decision pro-
cesses (POMDPs) to solve the general decentralised con-
trol and planning problems in MRS [4, 2], due to the the
improvement of the general concepts of multi-agent sys-
tems (MAS). However, these solutions are computation-
ally expensive and provide sub-optimal solution. Also, the
requirements of scalability and robustness in a smart fac-
tory is not met with these solutions. Likewise, the prob-
Preprint submitted to Robotics and Autonomous Systems
July 4, 2021
lems solved in latter case using POMDP ignore the aspect
of improving the cooperative functions based on perfor-
mances or state of individual MRs and their environment.
Thus, the problem still persists where a decentralized
system architecture is necessary for MRS which will be
scalable and robust, yet computationally inexpensive.
The robot computational architecture in an MRS is clas-
sified into 4 categories, such as, deliberative, reactive, hy-
brid and behavior-based [14]. Though, all the four cat-
egories of control have their advantages and limitations,
each contribute with interesting but different insights and
not a single approach is ideal [14]. Nevertheless, cur-
rent demands of smart factories are adaptability, real-time
response and modularity which is served excellently by
behavior-based control [14]. Moreover, decentralized con-
trol can be efficiently designed, implemented and handled
through behavior-based systems. Numerous investigations
have been carried out recently towards solving the prob-
lems of formation control in MRS using behavior-based
systems [13], while much attention is not paid toward
behavior-based control of MRS in logistics and transporta-
tion tasks. There are few investigations where this di-
rection is investigated [19], [11], but these behavior-based
MRSs lack the much needed involvement of support and
help by other MRs to each MR in decision making. This
help can be successful when cost incurred by one robot for
a particular performance can be shared to other robots to
estimate their costs to carry out the same performance.
These costs arise through the different battery and floor
conditions while performances. Thus, the time to com-
plete performance is dependent on all these factors and
thus reflects the costs incurred. When movement is a per-
formance, travel time of an edge is the performance time.
When travel times can be estimated to obtain close-to-real
values, they become different than heuristics costs and de-
pict the real states which are impossible to obtain from
heuristics. Nevertheless, a good estimation is dependent
on historical data which are close in time. But, there are
situations when all the travel times for one or more edge(s)
are not available for the entire duration of operation of the
MRS to an individual robot. Then, it is imperative for that
robot to gather the necessary travel times from others in
the system as a reference observation. We demonstrate
this concept with the following example.
Figure 1 illustrates a scaled down MRs-based internal
transportation system in a factory. Traversing a path is
considered as a task in this example. Let, at any instance
of time, t0, A1 is assigned to carry some material to P 1
through the computed path marked by the dotted line.
Again, at time tm (m ¿ 0), A1 needs to carry same mate-
rial to P 1. But at tm, A1 will need more time and energy
to reach P 1 than at t0 due to mainly two reasons. First,
the battery capability of A1 has decreased due to execu-
tion of previous tasks. Also, the condition of the part of
the floor, designated by the given path, can get deterio-
rated (as marked by black dotted lines). As said previ-
ously, travel times of edges depict states of battery and
Figure 1: Problem: An Example Scenario
floor condition (author?) [7]. So, the travel times at pre-
vious instances are useful to estimate travel time at tm, if
only battery state has changed. But, condition of floor has
also changed. This can only be anticipated through travel
times at tm if the robot has traversed that part of the floor
in the previous or nearly previous time instance. Never-
theless, travel time from other MRs who has traversed that
part in nearly previous instance can be useful to A1, along
with its own travel times at previous instances to estimate
its travel time at current. Thus, travel times of these two
sources are useful to estimate it's future travel time. This
work addresses this area of investigation where each MR
get information like travel time of edges from other MRs
in order to make better decisions.
The above example explains that the amount of time
and energy required to complete a task has an existing
correlation with state of charge of batteries and environ-
mental conditions. These time and energy can be formed
as cost coefficients to express state of battery's charge and
environment. These cost coefficients can be of various
forms like travel time of edges, rotating time, loading time,
et cetera depending on the functions. Further, they can
serve as a deciding factor in several planning decisions for
better cost efficient decisions. However, these cost coeffi-
cients need to be either known apriori or estimated to be
used in decision making. In case of knowing apriori, ob-
servations of these costs in various forms like travel time
of edges, rotating time, et cetera need to be measured for
all possibilities, which is not only cumbersome but also
impractical. Hence, estimating them during run-time is a
good solution. But, estimation requires observation of the
same at previous instances. The observations values can
be gathered from the beginning of first decision making
and can be used in subsequent calls for estimation. The
first few iterations of decision making is a learning phase
to gather few observations to start the estimation. But, an
MR may need to estimate the travel time of one or more
edge which it did not traverse previously. This can be miti-
gated by sharing the observation value from other MR who
2
has travelled that edge in nearly previous instance. This
way the knowledge sharing can help an MR to estimate
the travel cost for an unexplored edge at current instance.
In the example, A1 has travelled the edges in the region
(marked by dotted line) towards P 1 long back at t0 and
hence it does not have the latest information about its con-
dition at tm. In this case, travel time of edges, annotated
with time stamps, along the region marked by dotted line
from other MRs who has travelled it in nearly previous
instance, must be communicated to A1, so that A1 can
utilize it while estimating it's own cost at tm.
The travel times have inherent contexts like time stamp
and the edge between pair of nodes. This underlying con-
text has been exploited to form a semantic knowledge shar-
ing mechanism to communicate the costs of edges inform
of travel times in this work. This is instrumental in de-
riving more accurate estimates of travel times to ascertain
cost at current time in each MR. This improves decisions
in each MR for efficiency where MRs help each other to
gather states of environment and other factors. Moreover,
all the MRs are autonomous and have their own control
separately, which make the whole system decentralized.
This type of control is implemented using a behavior-based
system, to utilize the benefits of both decentralized archi-
tecture and behavior-based system.
The subsequent sections elaborate on the background
(Section 1), problem statement and contribution (Sec-
tion 2), methodology (Section 3.1 and Section 4.2) and
implementation (Section 3.2 and Section 4.4). Results of
utilising the proposed methodology is tabulated and ana-
lyzed on Section 6, while discussions and conclusions are
put forward in
2. Problem statement and contribution
This work addresses the problem of building a decentral-
ized system architecture where the planning decisions can
be based on the dynamically changing state of MRs and
their environment. This paves the way towards robust-
ness and scalability in the MRS. One of the most suitable
methods of control for MRS is behavior-based system as it
can handle significant dynamic changes with fast response
and enforce adaptability, few of the major requirements of
current smart factories [14]. The system architecture for
the MRS in current work is developed using concepts of
behavior-based system with specific behaviors for planing
and task execution.
Current work implements an MRS for automated lo-
gistics where each robot need to transport materials to
designated placeholders or racks, termed as ports. Also,
reaching a particular port by an MR is considered as
a task, along with route computing being considered as a
decision making process. Thus an MR is required to tra-
verse from one node to another in a floor, described by a
topological map. This enables the MRs to perform single
task at a time.
Figure 2: Problem description
The travel time of each arc (like aa,b, af,d) in a floor
map given in Figure 2 is influenced by energy exhaustion,
condition of floors, physical parameters of robot, among
others, which incurs cost. Thus time to traverse an arc by
an MR or travel time can be conceptualized as its cost co-
efficients. In this work, travel time is considered as weight
or cost for an edge. This is formalized as Xp,q(k) to denote
travel cost from np to nq, where k is the instance of time of
traversing an edge. Xp,q(k) is time-varying from the per-
spective that at a particular instance of the time, the cost
of that particular arc is dependent on battery discharge
and condition of the floor which changes over passage of
time. A path P is formed as a series of arcs between con-
necting nodes for an MR and thus P can be defined as
P = (cid:104)aa,b, ab,e, ae,g, ag,j, ...........(cid:105)
(1)
Now, the cost of traversing P can be written in a form
CP of
CP = (cid:104)Xa,b(k), Xb,e(k), Xe,g(k), ..., ..............(cid:105)
(2)
The elements of CP are required to be identified for each
call of path planning. From now on, Xp,q(k) will be written
as X(k) for simplicity. For continuous performance of the
MR, path needs to be computed for the MR after it reaches
a destination. Let, at ith call of path planning, path cost
was
q,r, .....(cid:105)
(3)
P = (cid:104)X i
C i
a,b, ..., X i
e,g, ....., X i
Now, in any instance, an MR may need to traverse an arc
which it had traversed in previous instances. Let at jth
(j>i+1) call of path planning, estimation of X(k) for ae,g
is required. As , the MR do not have the observation of
X(k) for ae,g at the previous instance. It can only use the
X(k) for ae,g obtained during traversing P after ith call.
In this scenario, the obtained estimate can be significantly
inaccurate which has the potential to produce inaccurately
optimized path. Thus, the observation of X(k) for ae,g at
jth call can be also obtained from other robots performing
3
in the system which has traversed that arc in the previ-
ous instance or in a nearly previous past instance. More-
over, at jth call, estimation of X(k) for some edge may
be needed which that MR has not been yet traversed dur-
ing all its previous traversals. This can be also solved by
fetching the observation data of X(k) for the required edge
from other robots which have traversed that at previous
or nearly previous instances.
Improved estimated values of X(k) can be obtained by
transferring the right knowledge from one robot to an-
other, which will generate more cost efficient decision.
Thus, an information sharing framework is imperative to
be formed in order to improve the estimation and in turn
improve the decisions where robots can support and help
each other in their decisions.
The contribution of this work include the following
• A completely decentralised system architecture is de-
veloped based on behavior-based system in a hierar-
chical model for each robot, which ensures scalability
and improves robustness
• A semantic knowledge sharing mechanism is devised
in each robot to share estimated values of travel times
of one robot to others. This eventually helps in ob-
taining better estimates of travel times in each robot,
which produces for more optimal path with minimum
path costs.
3. Behavior-based decentralised multi-robot sys-
tem
3.1. Methodology
The objective of this work is to form an MRS with a de-
centralized flow of control, suitable to logistics. The flow
of control is based on the concept of sub-sumption, where
each robot has the same sub-sumption model. Each robot
is capable of taking the decision itself, with the capabil-
ity of gathering information about the environment from
other MRs. This sub-sumption model achieves the goal
making each MR autonomous. The sub-sumption model
involves the control structure to be organized in layers one
above the other with increasing level of competencies and
each level can interact with all other levels with messages.
This technique of flow of control is described in Figure 3,
which consists of two major control layers.
The top most layer is L1 level and the L0 level is below
it. The L0 level is divided into two sub levels L0.1 and
L0.0 levels respectively. The L1 level is the agent level
control layer where it functions on all the agents in the
transportation system and is engaged in controlling more
complex functions like finding path, organizing task, find-
ing destination poses, et cetera for each of the robots. The
L0 level functions on each of the agents individually and
controls the movements. Each robot has its own L0.0 and
L0.1 levels respectively which controls the movements in
each of them. Here, the L0.0 level communicates with the
Figure 3: Controller architecture
L0.1 level and have no communication with the L1 level.
The L0.1 level is the intermediate level which communi-
cates both with L0.0 level and L1 level. The control levels
functions in co-ordination with each other to control the
movements of the robots in the environment [15]. Thus,
essentially the MRs in the system are autonomous. More-
over, the top most L1 level is responsible for intelligent
decision making for task assignments and path traversal,
based on the available travel time, which represents knowl-
edge about the individual robot and the environment.
3.2. Implementation
The control technique [5],[6], described in previous sec-
tion, is achieved using behaviors as the building block of
both decision-making level (L1 level) and action execution
level (L0 level). Separate sets of behaviors are designed for
two layers as illustrated in Figure 4, which is based on the
control framework proposed by R. Brooks in [3]. The hier-
archical control framework has three behavioral levels and
each level has an objective and a corresponding output,
formed as commands or replies. Moreover, the process of
execution of all levels start simultaneously. However, the
output in the form of commands from the highest level
(L1) need to pass on to the next priority level (L0.1) for
it to start execution, and similar process is followed in
L0.0 level. This happens because of the hierarchical frame-
work and the command from the highest-priority behav-
ioral level is required as input to process the low-priority
behavioral level. On the other hand, the replies from the
lower level act as a feed back to the control rules of the
higher level which determines the final decision and output
of the control framework.
The general design of an MR is considered in the pro-
totype system which consists of servo-motors to rotate
wheels and camera. The sensing is conducted with infra-
red sensors and camera. The beagle-bone forms the pro-
cessor for the robot. Each MR in the system is autonomous
4
and environment. As behavior in L level conducts the pro-
cess of decision, the knowledge sharing process is realized
in L in each level which provides each robot an opportu-
nity to seek help about states of environment from other
MRs.
As discussed in Section 1, travel time X(k) (Section 2)
for a particular edge provides the necessary representation
of state of robots and environment. A direct correlation
has been found between X(k) with state of charge of bat-
teries and conditions of floor in the prototype system. This
is depicted in Figure 5. Part (a) plots the cell voltage of
Li-ion batteries over time, Part (b) plots the progressive
mean of observed values of X(k) for mth edge with the
change of state of charge of batteries and Part (c) plots
the observed value of X(k) for the same edge with both
the change of state of the charge of batteries and the floor
condition. The floor is changed from rough at the begin-
ning to smooth during the experiment.
Figure 5: change pic
The plot (b) shows that progressive mean of X(k) in-
crease first, then steadily decrease and then increase grad-
ually till complete discharge. Thus values of X(k) first
increase due to sudden fall of cell voltage at beginning,
then decreasing fast due to cell voltage increasing fast to
a steady level and the values gradually decrease towards
complete discharge of batteries. Thus, a correlation be-
tween X(k) is observed through plots (a) and (b). On
the other hand, the increase in progressive mean of X(k)
is longer than that of plot (b) at equal battery capacity.
The longer increase of values of X(k) in (c) can be at-
tributed to the rough floor, as more energy is required to
traverse in rough surface. Plot of X(k) in different condi-
tions of floor demonstrate that travel time can reflect not
only state of charge of batteries [7] but also environmental
conditions.
During the run-time of MRS, the estimation of X(k) is
conducted for all necessary edges while finding the optimal
5
Figure 4: Behaviors for the layers of control
with its own three level of behavioral control framework.
The following are the behaviors developed in each level.
• In L0.0 level, actuation behaviors are developed. This
behavior conducts the starting of motors for wheel ro-
tation, camera movements and infra-red sensor move-
ments. The commands which refer to target poses
are obtained from L0.1 level in this behavior using
extended finite state machines to conduct the move-
ment of the individual robots. The sensor readings
are transferred to L0.1 for processing as feedback of
commands.
• In L0.1 level, three behaviors are developed. They are
generating target poses from high level commands like
destination port, finding obstacle and obstacle avoid-
ance, processing sensor data to be used in planning
and decision in L1. All these behaviors are developed
using extended finite state machine.
• In L1 level, decision making behaviors are developed
which are finding paths and assigning tasks. These
behaviors are developed using extended finite stacked-
state machine. Also, behavior of maintaining and
sharing the knowledge of travel time is developed.
More details about the sharing mechanism is provided
in next sections.
All these three levels of behavior correspond to a sin-
gular behavior for a single robot and this is repeated for
each MR. Thus, the control flow in the MRS is decentral-
ized. The knowledge sharing mechanism is incorporated
in the behavior of L so that each robot can communicate
through them. The highest level (L1 level) is implemented
on the desktop computer in our model to reduce commu-
nication costs among the L1 level agents in each MR. The
next two lower levels are implemented on body on individ-
ual MR using embedded system techniques. The decisions
for planning need the information about the states of itself
that by A2 at any ti because of differently discharged bat-
teries for different previous tasks. Thus, estimated travel
time provides contextual information representing state of
charge, condition of floor, instance of travelling.
These values of X(k) at a particular instance for a par-
ticular edge of one MR provide contextual information
about cost for that edge to other MRs when communi-
cated. Hence, semantics can be built from these knowl-
edge of travel time as they have inherent contextual in-
formation. They convey information about the costs of
traversing through different edges in the topological map,
which describes the factory floor.
4.2. Using semantics for knowledge sharing
Semantic relationships are built in this work to form se-
mantic sentences in order to fetch the values of travel times
with the inherent contextual information. This concept is
illustrated in Figure 7.
path. Thus, estimated values of X(k) will be generated
at every instance of control decisions, producing a pool
of estimated values. More significantly, every estimated
value of travel time has inherent context associated with
it, which when shared with the other MRs help in the
estimation of X(k) in them. This concept is elaborated in
the next section.
4. Information sharing in behavior-based control
4.1. Semantics in travel time
An MRS is dynamic as its states change over time. Also,
it is evolving as it gathers more and more knowledge about
its states through the course of its operation. Moreover,
the source of knowledge of an MRS is distributed to each of
its constituent robots. The behavior in L1 has the role of
finding paths using Dijkstra's algorithm. Dijkstra's algo-
rithm needs to know the estimated X(k) for the concerned
edges to decide the path as X(k) designates the cost of
traveling the edge. Now, there are possibilities when an
MR has not yet traversed many edges. The estimation of
X(k) for these edges depends on the obtained travel cost
of them from other MRs. Thus, knowledge sharing mech-
anism improves the estimation of X(k) for accuracy. This
will be instrumental to Dijkstra's algorithm to produce
better optimal paths with minimum path costs.
The X(k)s originate from each MR depending on the in-
stance of travelling, zone of the floor, previously traversed
edge, neighboring AGVs, state of charge, et cetera. All
these factors provide context to the estimated values of
X(k).
Figure 7: An example of semantic relationship in MRS
From the above example, a semantic sentence can be
derived as
• Cost from node Na to node Nb is Xa,b at time instance
k
where, k is the instance of estimation. Na and Nb refer
to specific nodes, travel time Xa,b refer to specific kind
of cost. Cost refer to specific kind of utility expenditure
while performing the task. Thus, cost establishes the rela-
tionship between nodes Na and Nb and travel time Xa,b.
When the system knows the meaning of nodes, utility
cost, travel time, then the above sentence will convey
some meaning to the system. This is precisely the method
of developing semantics in the MRS in order to convey the
contextual meaning instilled in travel time to the L level
controller.
Figure 6: change pic
For example, in Figure 6, the X(k) of an edge by A1
at tm will be different than that at t0 due to discharg-
ing of batteries as explained in Figure 5. On the other
hand, X(k) for nth edge (n (cid:54)= m) by A2 in a different
zone (marked by double dotted line) will be different than
that of mth edge, though both mth and nth edge can have
same lenghth. This happens due to different states of floor.
Moreover, X(k) for nth edge by A1 will be different than
6
4.3. Ontology to represent semantics
4.4. Application of ontology
The most traditional, flexible and useful method of rep-
resenting knowledge using semantics is expressions based
on subject, predicate and object logic [17]. Positioning the
obtained knowledge is the next progressive step which is
defined in philosophical terms as ontology. Ontology helps
to create order and define relationships among things use-
ful to an application. A domain specific ontology is de-
veloped in this work to efficiently store, access and com-
municate meaningful semantics across all the MRs in the
system regarding the real-time travel costs of edges.
There are significant advantages of implementing ontol-
ogy for the already mentioned application of this work.
• Conceptualization of information: An ontology is
defined explicitly to form a specification for a shared
conceptualization of a pool of knowledge [20], [12],
[18]. Ontologies define the concepts of the domain
formally and explicitly making further modifications
or reversals less cumbersome.
• Data representation: Ontology is based on dy-
namic data representation where a new instance def-
inition is not constrained to a definite rule. Thus
adding new elements is easy and fast as and when
required. This virtue of ontology is essentially bene-
ficial to share the knowledge of travel time in MRS.
The number of travel time grows with the increase of
operation time. Moreover, reasoners in ontology solve
the problem of data parity, integrity and adhering to
constraints. When a new element is added to an on-
tology, the reasoner performs to check the integrity
of the information. This capability of ontology makes
the knowledge sharing method in the MRS flexible yet
robust. Data addition in MRS is not required to be
done on all instances and when it is added the reasoner
checks for data integrity and new information can be
added smoothly without adhering to rules, previously
defined.
• Modeling technique: Ontology possesses the capa-
bility to express semantic concepts. In case of MRS,
conveying the contextual information inherent to any
cost parameter like travel time requires this semantic
expressiveness than just defining or extracting data.
Moreover, the pool of knowledge gathered in the MRS
through travel time or similar parameters need to be
reused which is only possible through the descriptive
logic models of ontology.
In nutshell, ontology provide an unrestricted framework
to represent a machine readable reality, which assumes
that information can be explicitly defined, shared, reused
or distributed. Moreover, information can also be inter-
changed and used to make deductions or queries. Such rep-
resentation is imperative for representing the travel time
for reasons described above.
Semantics is an efficient way to communicate enough
meaning which can actuate some action. The focus on
representing semantic data is through entities. Seman-
tic models are property oriented and semantic entities are
members of a class. Semantic classes are defined on prop-
erties, it is also possible to define classes in terms of value
of a property. A property type is object property when it
signifies some abstract property like character, contribu-
tion, virtue et cetera. A property type is a data property
when it signifies some literal value. On the other hand,
classes can have any of the type of properties. The sub-
classes are defined which can avail all the properties of
the superclass. The properties have range and domain.
Range is the source type of a property, while domain is
the destination type of the property.
Figure 8: Ontology
the knowledge
of
time
(Figure
Based on these concepts,
the ontology stores and
shares
travel
8).
The ontology has two types of classes (owl:Class),
shown in Figure 8.
NS:Edge and NS:Node,
Thus, NS:Edge and NS:Node are
subclasses of
owl:Class. There are two properties a class can possess,
owl:ObjectProperty and owl:DatatypeProperty.
NS:Origin and NS:Destination are of
types of
owl:ObjectProperty, while NS:tt, timeStamped are
of types of owl:DatatypeProperty.
as
The range of NS:Origin is subclass NS:Node, being
the source type of a property, while domain is NS:Edge
being the destination type of the property. Similarly, the
range of NS:Destination is subclass NS:Node, being
the source type of a property, while domain is NS:Edge
being the destination type of the property. On the other
hand, the range of NS:tt is a float, being the source type
of a property, while domain is NS:Edge being the desti-
nation type of the property. Similar is the case for timeS-
tamped. The tupled relationships are formed by using
these domain and range connections. For example, let
mth edge be between nodes ng and nh. X(k) for mth
edge at k can be formed as NS:tt value at timeStamped
value k for the mth individual of subclass NS:Edge whose
NS:Origin is individual ng of subclass NS:Node and
7
NS:Destination is individual nh of subclass NS:Node.
This semantic sentence can be disintegrated into several
subject, predicate and object logic to derive the necessary
X(k). For example,
• individual mth edge is of type NS:Edge
• individual ng is of type NS:Node
• individual nh is of type NS:Node
• mth edge has NS:Origin ng
• mth edge has NS:Destination nh
• mth edge has NS:tt X(k)
• mth edge has timeStamped k
This
way
the
the
owl:ObjectProperty
and
subclass NS:Edge
owl:DatatypeProperty of
provides the X(k) for the mth edge. Also, the X(k) gets
a context about its edge (between a pair of nodes) and
time stamp. The advantage of this ontology lies in this
formation, as discussed in previous Section 4.3, where
any new element can be inserted through these property
formations without being restrained semantically. With
the use of ontology, travel time X(k) can be efficiently
stored annotated with a pair of nodes demarcating the
edge and the time stamp of traversing it.
The structure illustrated in Figure 8 shows the forma-
tion of ontology which is replicated in each robot in the
MRS. Thus, when the information of travel cost for any
edge for any time instance is required by any MR, X(k)
for that edge at the required time stamp can be retrieved
from ontology of other MRs. This shared information from
other MRs can provide as observation or historical data
for those edges which either have not been yet travelled
or have been travelled long back. This helps in achieving
accurate estimates of X(k) of these edges.
For example, in Figure 6, when A2 requires to estimate
X(k) for edges through the marked zone (marked by dot-
ted line), the historical observation data of X(k) in that
zone can be obtained from the ontology of A1 whi h as tra-
versed those edges in previous or nearly previous instance.
The estimated values at current instance become more ac-
curate using X(k) of the same edges by A1 at previous
instances.
This information can be sought by the L1 level behaviors
in any MR to other L1 level behaviors in other MRs. Thus,
this ontology fulfills the mechanism of knowledge sharing
inside the L level behaviors. A co-operative approach in
achieved through this knowledge sharing for better cost
efficient decisions in each MR, which in turn enhances the
cost efficiency of the MRS.
5. Retrieval of travel time and using in estimation
The sharing of travel time to all MRs is implemented
through ontology in each of them to generate better esti-
mate of travel time among all (Section ??). This section
8
describes the methodology of using travel time of others
in the estimation process of an MR.
The travel time of an MR is modelled using bi-linear
state dependent time series [? ], which is described in
Section 6.3 in Chapter ??. This is again produced here for
convenience. The bi-linear model, provided in equation 4,
is used to model the change of travel costs depending
upon all the previous travel costs.
X(k) + a1X(k − 1) + ..... + ajX(k − j)
= ξk + b1ξ(k − 1) + ... + blξ(k − l)
+
crzξ(k − r)X(k − z)
(cid:88)(cid:88)
(4)
The model described in equation 4 is a special case of the
general class of non-linear models called state dependent
model (SDM) [? ]. In equation 4, X(k) denotes the edge
travel cost at k and ξ at k denotes the inherent variation
of the edge travel cost. In equation 4, X(k) depends on
all the previous values of X and ξ, whose number is pro-
vided by the variables j and l. However, a fixed number
of previous values of X and ξ is used for estimation of cur-
rent X like an window which moves with increase of time.
This fixed size of this window is termed as regression num-
ber and it is chosen as a design parameter, designated by
j and l. The double summation factor over X and ξ in
equation 4 provides the nonlinear variation of X due to
state of batteries and changes in environment.
The state space form of the bi-linear model is given in
equation 5 and equation 6.
s(k) = F (s(k − 1))s(k − 1) + V ξk + Gωk−1
Y (k) = Hs(k − 1) + ξk + ηk
(5)
(6)
is of
state.
the next
the state vector s(k)
state from the current
the state equation which pro-
The equation 5 is
In
vides
equation 5,
the form
(1, ξk−l+1, ...., ξk, Xk−j+1, ......, Xk)T . The state vector
contains the edge costs obtained progressively over time
from Xk−j+1 to Xk. The variable ξ provides values of in-
novation or evolution of edge costs over the time as the
exploration proceeds. Here, j denote number of previ-
ous edge costs to be included in the state vector among
all edges included in the path till kth instance. Also, l
denotes the number of previous evolution values of these
edges. The ξ values are specific for each MR and originate
from the changes in travel time of the particular MR. The
values of ξ are obtained by sampling using the observation
data of travel time. This observation data is obtained for
the static online estimation of travel time (Section 6.1 in
Chapter ??). The ξ values obtained through this method
represents the projection of change of travel time. Though,
these sampling method does not produce the perfect data
to represent the change of travel time, this is suitable to
this simple case where cost factor of one task is considered.
This method should be improved for the case where cost
factors of two or more tasks are to be considered.
The matrices of equation 5 are F , V and G which are
explained in the following.
F =
1 0 0
0 0 1
0 0 0
...
0 0 0
...
...
0 0 0
0 0 0
0 0 0
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
0
0
1
0
...
0
0
0
µ ψl ψl−1
. . . ψ1
0 0
0 0
0 0
. . .
. . .
. . .
. . .
0
0
0
... 0
... 0
... 0
... 0
...
...
... 0
... 0
... 0
1
... − φk − φk−1 ··· − φ1
0 0
...
0
...
1
0 1
0 0
. . .
. . .
. . .
0
0
0
. . .
...
particular instance. In equation 5, this factor contributes
not only to the formation of state but also forms the state
equation to predict the next state. In equation 6, this ξk
is added to the product of H and s(k-1) to form observa-
tion. The product of H and s(k-1) is Xk−1 The addition
of Xk−1, ξk and error term ηk produces the observation
Y (k). In this way, the travel time of other MRs are used
in the model for estimation of travel time. The estima-
tion is done by Kalman filtering. The equations obtained
after applying Kalman filtering this bi-linear model are ex-
plained in Section 6.3 in Chapter ??. The same process is
continued to obtain the travel time of relevant edges.
This travel times are the instruments to decide the path
using Dijkstra's algorithm. This whole process is summa-
rized in Algorithm 1.
6. Experiment and Results
This work proposes a behavior-based control method
which uses online estimated travel time as a decision pa-
rameter for computing optimal routes between pairs of
ports.
The number of rows of F depends on the number of
regression no and given by (2*regression no + 1). The
matrix F contains many new terms like ψ , φ , µ. The ψ
terms are denoted as in equation 7
l(cid:88)
ψl = bl +
cliX(k − i)
(7)
i=1
6.1. Experiment-I: Behavior-based decentralized control
All the φ terms in F are constants. The term µ is the aver-
age value of X till kth instance. Thus, the state transition
matrix F depends on the travel times of the previously
traversed edges. Also, the matrix V is denoted as
V =
0 0 0
. . .
1
... 0
0 . . .
1
The number of
rows of V is again given by
(2*regression no + 1). The equation 6 is the observa-
tion equation which forms the observation for the current
instance. The matrix in equation 6 is H which is described
as
H =
0 0 0
. . .
0
0 . . .
1
... 0
(cid:105)
(cid:105)
(cid:104)
(cid:104)
system for MRS based logistics
A prototype multi-robot system for logistics in factory
is developed based on the proposed behavior-based decen-
tralized planning and control method. The experimenta-
tion platform is briefly described in this section to provide
elaborate explanation of the experiment. A scaled down
prototype of automated indoor logistics system is built.
The observation is formed by multiplying the H
matrix with the state vector s and added with the
innovation at the current instance.
In both equa-
tion 5 and equation 6, s(k-1) denotes the state vec-
tor at current instance. Here, s(k-1) is of the form
(1, ξ(k−1)−l+1, ...., ξk−1, X(k−1)−j+1, ......, Xk−1)T . The X
values in this vector are the travel times obtained for the
edges which are already explored and included in the path.
But, the available travel times may not be enough to ful-
fill all the data till the previous instance. Thus, the travel
times for the previous instance which are not available are
gathered from other MRs. In order to gather this data,
the relevant edge costs are queried in the ontology of other
MRs. Then after retrieval of the data they are filled in the
state vector for both equation 5 and equation 6. Both the
equations have ξ which corresponds to the innovation or
change of travel time. Thus, this factor plays the role of
projecting the travel time of the particular MR at some
Figure 9: Environment of MRS
An environment has been developed using uniform sized
boxes as shown in Figure 9 for the robots to work, doing
single task at a time and is named as single-task robot.
The boxes create a closed labyrinth path to navigate. Also
designated ports are marked on the boxes. The floor is
9
Figure 10: The three topology maps
described in three different topological maps. These maps
are provided in Figure 10.
The control structure is same in each MR which consists
of two layers of sub-sumption structure (Section 3). The
lowest L0.0 level is implemented in the body of each MR
inside the beagle board which forms the main processor of
each robot. The middle L0.1 level and L1 are implemented
in desktop PCs where each level is separated for each MR.
Thus, the entire two-layer control architecture is formed
through the designated behaviors (Fig 4) for each MR in
the system.
The MRs carry out the task of pick-up or drop and car-
rying materials between different pair of ports. The L1
level controller in each MR is responsible for planning de-
cisions to make them reach designated ports. The optimal
path between different pair of ports is found out using Di-
jsktra's algorithm. These functions are carried out through
the DECISION BEHAVIOR (Fig 4) in L1 level. Dijsktra
uses X(k) as weight of an edge at every step of forming the
path. X(k) is estimated on real-time by Kalman Filtering
at required k in each MR. It was stated in Section 1 that
online estimation of X(k) requires observation of the same
at k-1. In this experiment, these observations are gathered
from the beginning of first decision making and are used
in subsequent calls for estimation. But, an MR may need
to estimate the travel time of one or more edge which it
did not traverse previously or has traversed long back. In
this case, the observation of X(k) for the concerned edge
is not available. In this experiment, observation of X(k)
for the concerned edge at k-1 could not be obtained al-
ways and thus the available observation for the concerned
edge is used. Thus, X(k)s are estimated solely based on
the historical observation of the concerned MR. Dijsktra's
algorithm uses the estimated X(k)s for each edge. Then
it chooses the predecessor node of the current node, from
which arrival to current node becomes least cost expend-
ing. This way the optimal path is formed using X(k).
These paths are shown in Section 6.4
6.2. Results-I
The resultant paths (Figure 11) form a high level com-
mand or macro-command to be transferred to the next
lower level L0.1.
In L0.1, these high level commands or paths are dis-
integrated into macro-actions like turn-left, turn-right,
move-ahead, stop-left, stop-right, et cetera. This dis-
integration happens due to the functions of three behav-
iors in L0.1 level (Fig 4). Further, these macro-actions are
processed in L0.1 level to produce easy and low-level com-
mands like GO-(cid:104)angle(cid:105)-(cid:104)distance(cid:105) which can be easily
understood by the lowest level controller L0.0. These low-
level commands enable the behaviors in L0.0 for accurate
and prompt servo actions to generate movements in MR.
This process is shown in Figure ?? where the MR traverses
the path. Figure ?? shows different sections of the path in
different steps 1 from beginning (Part (A)) to end (Part
(H)).
In this way, the decentralized control is performed where
each MR is the master of their own decision with the fa-
cility which is enabled due to the two layer subs-sumption
control architecture based on behaviors.
6.3. Experiment-II
In this experiment, ontological data sharing is incorpo-
rated. The MRs are made to traverse repeatedly between
different pairs of nodes. The pairs are designated previ-
ously from a list in order to suit the carriage necessity. The
route computation between different pairs of node are done
similarly as in Experiment 1 in Section 6.1, using online
estimated values of X(k) as weight of edge.
Online estimation of X(k) at k requires observation of
the same at (k-1).
In both Experiment I and Experi-
ment II, these observations are gathered from the begin-
ning of first decision making and are used in subsequent
calls for path planning. However, when an MR needs to
estimate the travel time of one or more edge(s) which it
did not traverse previously, the available observation of X
1Video of path traversal is available at .....
10
Figure 11: change the pic, make it double column
for the concerned edge is at some previous instance which
may not be (k-1) or close to that in many cases. These
observations of distant past are used in experiment I for es-
timating X(k) at k. Thus, this will generate less accurate
estimates.
This is mitigated in this experiment II by sharing the
observation value from other MR who has travelled that
edge in nearly previous instance. This way the observation
of X for the concerned edge at (k-1) or close will be avail-
able during estimation at k. The knowledge sharing con-
tributes to estimate the travel cost for an unexplored edge
at current instance in an MR. The behavior in L1 layer in
each robot can ask the L1 level of other neighboring robots
for observation values of X(k) whenever required. X(k)s
are estimated for the necessary edges using observations
either from the own MR or from neighbors.
Meanwhile, before deployment of the behavior-based
system and ontology, some legacy data for travel times of
different edges at different instances are obtained. These
pool of date gathered by recording the travel times dur-
ing the operation of MRS correspond to experiences of the
MRs in the system. The estimated X(k) values are com-
pared to these legacy data to measure the accuracy which
is discussed in Section 6.4.
On the other hand, estimated X(k) values of the rele-
vant edges are used by Dijsktra's algorithm as weights of
edges. These estimates are the main instrument at every
step of deciding the predecessor to the current node. Dijsk-
tra's algorithm makes a node predecessor to current, when
weight or cost from the former to later becomes minimum.
Thus, accurate estimated value of X(k) plays a vital role in
deciding the predecessor to current node, in turn deciding
the path. More accurate estimates contributes to gener-
ate paths with less total cost. Optimal paths are obtained
with (experiment II) and without (experiment I) sharing
the X(k) values. These paths are compared in Section 6.4.
6.4. Results-II
This section tabulates the results of experiment II.
X(k)s were estimated for each required edge at every step
of Dijkstra's algorithm. These estimates are obtained us-
ing a non-linear model and Kalman filtering, with observa-
tion data being shared from other MRs. These estimates
are compared with that of experiment I where estimation
is done without sharing data among MRs. Figure ?? plots
the comparison of estimated X(k) at different k for differ-
ent edges with that of legacy data obtained.
This section illustrates the comparison of paths and
their costs obtained with and without sharing the travel
times in the MRS. The path planning is done for 100 rep-
etitions while increasing the regressionno from 4 to 7.
Figure 12 illustrates average total path costs of 100 paths
Figure 12: Average of total path costs
obtained in both Experiment I and Experiment II in four
MRs operating in all three maps (Figure 10). The aver-
age path costs of 100 paths obtained by sharing (Experi-
ment II) and not sharing (Experiment I) travel times are
plotted for each regression, namely Reg4, Reg5, Reg6 and
Reg7.
For each regression in Map 1, the average of path costs
obtained through collective intelligence are 40% less than
the average of path costs obtained without it. For each
MR, the average of total path costs is almost same or vary
in small margin with the increase of regression number.
11
The reason of this is the lack of variation in environmental
conditions. The travel times are varying on battery con-
dition and floor. No other factor for affecting travel time
could be incorporated in the laboratory set-up.
On the other hand, the save of total path costs are same
for all MRs in a single map. This signifies that paths found
through collective intelligence in each MR is 40% more
cost efficient than the paths obtained without it. Thus,
collective intelligence using travel time can affect to find
more cost efficient paths in MRS. The average path costs
decrease in case of Experiment II as through collective
intelligence more relevant observation of travel times are
obtained in each MR. These values are instrumental for
obtaining more accurate estimates of travel time. As a
matter of fact, more accurate estimated values result in
more optimal path with less cost than in that of obtained
in Experiment I. Few examples of these paths are discussed
in the next section.
Moreover, the save on total path costs is consistent in all
the maps. Thus, the travel time is estimated better due to
sharing of travel times from other MRs and this is true for
all the representative structures of the floor. This signifies
that more accurate estimation is possible through collec-
tive intelligence and this is independent of the structure of
the floor.
6.5. Analysis of obtained paths
This section illustrates few paths obtained in Experi-
ment I and Experiment II under the same condition of
regression no and MR. Figure 13 plots two paths, PA and
Initialise Single Source (V, E, s)
Input : V -list of nodes, E-list of edges, s-source
node
π[v]-predecessor of each node
Output: d[v]-attribute for each each node,
for each xi ∈ V do
π[xi] = infinity
d[xi] = NIL
end
d[s] = 0
findedgedCost (u, v, j)
Input : u-current node, v- neighbor node
Output: w- estimated travel time (cost) from u to v
f indP redEdge(u)
prevx := x(prevEdge)
w = estimateKF (prevx,j,X)
findPredEdge (u)
Input : u-current node
Output: prevEdge-edge connection u and predU
prevEdge = edge between u and predU
estimateKF (prevx, j, X)
Input : prevx-x(j − 1), j-instance for estimation, X-
observation variable
Output: xj-travel cost at current j for current edge
Apply Kalman filtering to find sj and return xj
Relax (u, v, w)
Input : u-current node, v- neighbor node, w-
estimated travel time (cost) from u to v
Output: d[v]-attribute for each each node,
π[v]-predecessor of each node
if d[v] > d[u] + w(u, v) then
d[v] = d[u] + w(u, v)
π[v] = u
end
Main (V, E, w, s)
Input : V -list of nodes, E-list of edges, w-edge
weight matrix, s-source node
Output: π[v]-predecessor of each node
P := NIL
Q := V
j := 0
while Q! =0 do
P := P (cid:83) u
j = j+1 u := Extract min (Q)
for each v ∈ Adj[u] do
w = f indedgedCost(u,v, j)
relax(u,v,w)
end
end
Algorithm 1: Dijkstra's algorithm using dynamic esti-
mation of travel time
Figure 13: Paths found by MR 1 in Map 1
PB obtained in Map 1 for MR 1. PA and PB both have
same source and destination. PA is obtained in Experi-
ment I in the third iteration of path planning, while PB
is obtained in Experiment II at the same iteration. Thus,
they are both obtained at the same battery level and in
the same map. Still, both the paths are different and have
12
different total path cost. As described in Section2, CP de-
notes the cost of a path. CP A and CP B denote cost of PA
and PB of Figure 13 respectively. The results show CP A
= 66.5326 and CP B = 39.5385. Thus,
CP B < CP A by 40%
Figure 14: Paths found by MR 2 in Map 2
Figure 14 plots two paths, PC and PD obtained in Map 2
for MR 2. PC and PD both have same source and destina-
tion. PC is obtained in Experiment I in the third iteration
of path planning, while PD is obtained in Experiment II
at the same iteration. Thus, they are both obtained at
the same battery level and in the same map. Still, both
the paths are different and have different total path cost.
CP C and CP D denote cost of PC and PD of Figure 14 re-
spectively. The results show CP C = 58.0729 and CP D =
33.5707. Thus,
CP D < CP C by 42%
From these two comparisons, it is evident that after
sharing the travel times among the MRs, the path ob-
tained in each MR have improved and are of less cost than
that obtained without the sharing.
7. Discussion and Conclusion
The new method to compute cost parameter to be used
in transportation and automation industry is proposed.
With this new method, parameters now reflect the states
of individual robots, their batteries and their environment.
They usually arise locally at the robots as a result of per-
formances of task.
In case of planning, the current state of robots and envi-
ronment plays crucial role. The usual practice is to decide
path using Euclidean distance and a path is considered
13
optimal with optimal length or distance. Many indus-
tries (like BlueBotics [1]) use topology maps to describe
the floor and employs a depth-first search to generates a
length-optimal path. However, the true cost of traversing
a path is not accounted in this case. The cost involved in
traversing the path is generated from condition of floor,
state of batteries, mechanical parts of robots. It is intu-
itive that an edge of same length will incur more cost in
a rough floor than in smooth one. Thus, travel time is
a better tool to decide a path than heuristics based on
Euclidean distance.
In this work, the decision making of each robot is based
solely on the travel costs of its own. In the dynamic esti-
mation process, there are possibilities of not being able to
learn observation of few edges due to lack of experience.
Also, the observation gathered for a particular edge is too
old to be relevant at the current instance of estimation.
To address this, the sharing of travel time is incorporated
to be able to share data of travel time from one MR to
others. This enables the MR to generate more accurate
estimation for travel times.
[1] BlueBotics, 2001 (accessed April 12, 2018).
[2] C. Amato, G. Konidaris, G. Cruz, C. A. Maynor, J. P. How, and
L. P. Kaelbling. Planning for decentralized control of multiple
robots under uncertainty. In 2015 IEEE International Confer-
ence on Robotics and Automation (ICRA), pages 1241 -- 1248.
IEEE, 2015.
[3] R. Brooks. A robust layered control system for a mobile robot.
IEEE Journal on Robotics and Automation, 2(1):14 -- 23, 1986.
[4] J. Capitan, M. T.J. Spaan, L. Merino, and A. Ollero. Decentral-
ized multi-robot cooperation with auctioned pomdps. The In-
ternational Journal of Robotics Research, 32(6):650 -- 671, 2013.
[5] I. F. Chaile and L.R. Xirgo. Agent simulator-based control
architecture for rapid development of multi-robot systems. In
International Conference on Systems, Control, Signal Process-
ing and Informatics (SCSI 2015), INASE Joint Conferences-
Barcelona, Spain, 2015.
[6] I. F. Chaile and L.R. Xirgo. Running agent-based-models sim-
ulations synchronized with reality to control transport systems.
Automatika, 57:452 -- 465, 2017.
[7] Pragna Das and Llu´ıs Ribas-Xirgo. Predicting battery level
analysing the behaviour of mobile robot. In XVII Workshop of
Physical Agents Book of Proceedings, pages 91 -- 98, 2016.
[8] V. Digani, L. Sabattini, C. Secchi, and C. Fantuzzi. Towards
decentralized coordination of multi robot systems in industrial
environments: A hierarchical traffic control strategy. In 2013
IEEE 9th International Conference on Intelligent Computer
Communication and Processing (ICCP), pages 209 -- 215. IEEE,
2013.
[9] A. Farinelli, L. Iocchi, and D. Nardi. Multirobot systems: a clas-
sification focused on coordination. IEEE Transactions on Sys-
tems, Man, and Cybernetics, Part B (Cybernetics), 34(5):2015 --
2028, 2004.
[10] H. Farivarnejad, S. Wilson, and S. Berman. Decentralized slid-
ing mode control for autonomous collective transport by multi-
robot systems. In 2016 IEEE 55th Conference on Decision and
Control (CDC), pages 1826 -- 1833. IEEE, 2016.
[11] E. A. B. Gonz´alez, M. R. P. Pereira, and G. R. B. Boh´orquez.
Behavioral control of a lego nxt robot oriented by searching
tasks and avoiding obstacles. Visi´on electr´onica, (2):12, 2016.
[12] Thomas R. Gruber. A translation approach to portable ontology
specifications. Knowledge Acquisition, 5(2):199 -- 220, 1993.
[13] G. Lee and D. Chwa. Decentralized behavior-based formation
control of multiple robots considering obstacle avoidance. In-
telligent Service Robotics, 11(1):127 -- 138, 2018.
[14] F. Michaud and M. Nicolescu. Behavior-Based Systems, pages
307 -- 327. Springer, 2016.
[15] A. Norouzi and C. A. Acosta. An approach to design a ro-
bust software architecture and an intelligent model for multi-
agent systems.
In 2013 3rd Joint Conference of AI Robotics
and 5th RoboCup Iran Open International Symposium, pages
1 -- 7. IEEE, 2013.
[16] D. Panagou, D. M. Stipanovi, and P. G. Voulgaris. Distributed
coordination control for multi-robot networks using lyapunov-
like barrier functions. IEEE Transactions on Automatic Con-
trol, 61(3):617 -- 632, 2016.
[17] Toby Segaran, Colin Evans, Jamie Taylor, Segaran Toby, Evans
Colin, and Taylor Jamie. Programming the Semantic Web.
O'Reilly Media, Inc., 1st edition, 2009.
[18] Rudi Studer, V.Richard Benjamins, and Dieter Fensel. Knowl-
edge engineering: Principles and methods. Data and Knowledge
Engineering, 25(1):161 -- 197, 1998.
[19] D. Sun, A. Kleiner, and B. Nebel. Behavior-based multi-robot
collision avoidance.
In 2014 IEEE International Conference
on Robotics and Automation (ICRA), pages 1668 -- 1673. IEEE,
2014.
[20] Daya C. Wimalasuriya and Dejing Dou. Ontology-based infor-
mation extraction: An introduction and a survey of current ap-
proaches. Journal of Information Science, 36(3):306 -- 323, 2010.
14
|
1310.7951 | 1 | 1310 | 2013-10-29T20:09:45 | IRM4MLS: the influence reaction model for multi-level simulation | [
"cs.MA"
] | In this paper, a meta-model called IRM4MLS, that aims to be a generic ground to specify and execute multi-level agent-based models is presented. It relies on the influence/reaction principle and more specifically on IRM4S. Simulation models for IRM4MLS are defined. The capabilities and possible extensions of the meta-model are discussed. | cs.MA | cs | the influence reaction model for multi-level
IRM4MLS:
simulation
3
1
0
2
t
c
O
9
2
]
A
M
.
s
c
[
1
v
1
5
9
7
.
0
1
3
1
:
v
i
X
r
a
Gildas Morvan1,2 Alexandre Veremme1,3 Daniel Dupont1,3
1Univ Lille Nord de France, F-59000 Lille, France
2UArtois, LGI2A, F-62400, Béthune, France
3HEI, F-59046, Lille, France
http://www.lgi2a.univ-artois.fr/~morvan/
[email protected]
1
Abstract
In this paper, a meta-model called IRM4MLS, that aims to be a generic
ground to specify and execute multi-level agent-based models is presented. It re-
lies on the influence/reaction principle and more specifically on IRM4S (Michel,
2007a,b). Simulation models for IRM4MLS are defined. The capabilities and
possible extensions of the meta-model are discussed.
Keywords: multi-level simulation, influence/reaction model
Introduction
1
The term "multi-level modeling" refers to the modeling of a system considered at
various levels of organization. E.g., a biological system can be considered at different
levels:
... → molecule → cell → tissue → organ → ... ,
that basically correspond to the segmentation of biological research into specialized
communities:
... → molecular biology → cell biology → histology → physiology → ... .
Each research area has developed its own ontologies and models to describe the
same reality observed at different levels. However, this reductionist approach fails
when considering complex systems. E.g., it has been shown that living systems are
co-produced by processes at different levels of organization (Maturana and Varela,
1980). Therefore, an explanatory model of such systems should consider the in-
teractions between levels. Agent-based modeling (ABM) is a paradigm of choice to
study complex systems. But, while it seems more interesting to integrate knowledge
from the different levels studied and their interactions in a single model, ABM often
remains a pure bottom-up approach (Drogoul et al., 2003).
Thus, recently1 various research projects have aimed at developing multi-level
agent-based models (ML-ABM) in various fields such as histology, ethology or so-
ciology (An, 2008; Gil-Quijano et al., 2008; Lepagnot and Hutzler, 2009; Morvan
et al., 2008, 2009; Pumain et al., 2009; Zhang et al., 2009). A good analysis of some
of these models, and the motivations of these works can be found in Gil-Quijano
et al. (2009).
Various issues should be addressed when developing a ML-ABM. For instance
one major problem is the automatic detection of emergent phenomena that could
influence other levels (Chen et al., 2009; David and Courdier, 2009; Prévost and
Bertelle, 2009). Another important problem is the temporal and spatial mappings of
model levels and thus the scheduling of the simulations (Hoekstra et al., 2007). More
exhaustive presentations of these issues can be found in Gil-Quijano et al. (2009);
Morvan et al. (2009).
In the models found in literature, these issues have been addressed according to
the specificity of the problem. Indeed, they are based on ad-hoc meta-models and the
transferability of ideas from one to another seems difficult.
1It has to be noted that the eleven year old model RIVAGE pioneered the field of ML-ABM (Servat
et al., 1998).
2
In this paper, a meta-model that aims to be a generic ground to specify and exe-
cute ML-ABM is presented. It is based on IRM4S (an Influence Reaction Model for
Simulation) proposed in Michel (2007a,b), itself based on IRM (Influences and Reac-
tion model) originally presented in Ferber and Müller (1996). IRM4S is described in
section 2 and its multi-level extension, called IRM4MLS (Influence Reaction Model
for Multi-level Simulation), in section 3. Section 4 introduces two simulation models
for IRM4MLS. The first one is very simple and similar to IRM4S but supposes that
all levels have the same temporal dynamics while the second one has a more gen-
eral scope but relies on temporal constraints and thus, is more complicated and time
consuming.
2 The IRM4S meta-model
IRM was developed to address issues raised by the classical vision of action in Artifi-
cial Intelligence as the transformation of a global state: simultaneous actions cannot be
easily handled, the result of an action depends on the agent that performs it but not
on other actions and the autonomy of agents is not respected (Ferber and Müller,
1996).
While IRM addresses these issues, its complexity makes it difficult to implement.
IRM4S is an adaptation of IRM, dedicated to simulation, that clarifies some ambigu-
ous points. It is described in the following.
Let δ(t)∈ ∆ be the dynamic state of the system at time t:
δ(t) =< σ(t), γ(t) >,
(1)
where σ(t) ∈ Σ is the set of environmental properties and γ(t) ∈ Γ the set of influ-
ences, representing system dynamics. The state of an agent a ∈ A is characterized
by:
∈ Σ (e.g., its position),
• necessary, its physical state φ
• possibly, its internal state sa
a
a with Φ
a
∈ Φ
∈ Sa (e.g., its beliefs).
Thus, IRM4S distinguishes between the mind and the body of the agents.
The evolution of the system from t to t + d t is a two-step process:
1. agents and environment produce a set of influences2 γ(cid:48)(t)∈ Γ(cid:48):
γ(cid:48)(t) = I n f l uence(δ(t)),
2. the reaction to influences produces the new dynamic state of the system:
δ(t + d t) = Reac t ion(σ(t), γ(cid:48)(t)).
(2)
(3)
As Michel (2007b) notes, "the influences [produced by an agent] do not directly
change the environment, but rather represent the desire of an agent to see it changed
in some way". Thus, Reac t ion computes the consequences of agent desires and
environment dynamics.
2the sets of producible influence sets and influences produced at t are denoted respectively Γ(cid:48) and γ(cid:48)(t)
to point out that the latter is temporary and will be used to compute the dynamic state of the system at
t + d t.
3
An agent a ∈ A produces influences through a function Be haviora : ∆ (cid:55)→ Γ(cid:48).
This function is decomposed into three functions executed sequentially:
pa
(t) = Pe r ce pt iona
(δ(t)),
(t + d t) = M e mor i zat iona
(pa
(t), sa
sa
(t)),
(4)
(5)
γ(cid:48)
a
(t) = Deci siona
(6)
The environment produces influences through a function Nat u ralω : ∆(cid:55)→ Γ(cid:48):
(7)
(t) = Nat u ralω(δ(t)).
(t + d t)).
(sa
γ(cid:48)
ω
(8)
(9)
Then the set of influences produced in the system at t is:
(t)}.
γ(cid:48)(t) ={γ(t)∪ γ(cid:48)
γ(cid:48)
ω
(t)∪(cid:91)
a
a∈A
After those influences have been produced, the new dynamic state of the system
is computed by a function Reac t ion : Σ× Γ(cid:48) (cid:55)→ ∆ such as:
δ(t + d t) = Reac t ion(σ(t), γ(cid:48)(t)).
Strategies for computing Reac t ion can be found in Michel (2007b).
3 The influence reaction model for multi-level simula-
tion (IRM4MLS)
3.1 Specification of the levels and their interactions
A multi-level model is defined by a set of levels L and a specification of the rela-
tions between levels. Two kinds of relations are specified in IRM4MLS: an influence
relation (agents in a level l are able to produce influences in a level l(cid:48) (cid:54)= l) and a per-
ception relation (agents in a level l are able to perceive the dynamic state of a level
l(cid:48) (cid:54)= l), represented by directed graphs denoted respectively < L, EI
>,
where EI and EP are two sets of edges, i.e., ordered pairs of elements of L. Influence
and perception relations in a level are systematic and thus not specified in EI and EP
(cf. eq. 10 and 11).
={l l(cid:48)} then the agents of l are able to perceive the dynamic
E.g.,∀l, l(cid:48) ∈ L2, if EP
states of l and l(cid:48) while the agents of l(cid:48) are able to perceive the dynamic state of l(cid:48).
The perception relation represents the capability, for agents in a level, to be "con-
scious" of other levels, e.g., human beings having knowledge in sociology are con-
scious of the social structures they are involved in. Thus, in a pure reactive agent
= (cid:59). EP represents what agents are able to be conscious of, not what
simulation, EP
they actually are: this is handled by a perception function, proper to each agent.
The in and out neighborhood in < L, EI
>) are denoted
−
I and N +
> (respectively < L, EP
> and < L, EP
N
I
(resp. N
∀l ∈ L,N
−
I
−
P and N +
(l) (resp. N
P ) and are defined as follows:
(cid:48) ∈ L : l
(cid:48)
(l)) ={l}∪{l
−
P
l ∈ EI (resp. EP)},
(10)
4
(l) (resp. N
∀l ∈ L,N +
E.g., ∀l, l(cid:48) ∈ L2 if l(cid:48) ∈ N +
(cid:48) ∈ L : l l
produce influences in the level l(cid:48); conversely we have l ∈ N
by l.
(11)
(l) then the environment and the agents of l are able to
(l(cid:48)), i.e., l(cid:48) is influenced
(cid:48) ∈ EI (resp. EP)},
−
I
(l)) ={l}∪{l
−
P
I
I
3.2 Agent population and environments
The set of agents in the system at time t is denoted A(t). ∀l ∈ L, the set of agents
(t) ⊆ A(t). An agent belongs to a level iff a subset
belonging to l at t is denoted Al
of its physical state φ
∀a ∈ A(t),∀l ∈ L,a ∈ Al
a belongs to the state of the level:
(t)⊆ φ
(t)⊆ σ l (t).
(t) iff ∃φl
(t)φl
(12)
a
a
a
Thus, an agent belongs to zero, one, or more levels. An environment can also belong
to different levels.
Influence production
3.3
The dynamic state of a level l ∈ L at time t, denoted δ l (t) ∈ ∆l , is a tuple <
σ l (t), γ l (t) >, where σ l (t) ∈ Σl and γ l (t) ∈ Γl are the sets of environmental prop-
erties and influences of l.
The influence production step of IRM4S is modified to take into account the
influence and perception relations between levels. Thus, the Be havior l
a function of
an agent a ∈ Al is defined as:
Be havior l
.
(13)
a : (cid:89)
∈N +
P
lP
(l)
∆lP (cid:55)→ (cid:89)
(cid:48)
ΓlI
∈N +
I
lI
(l)
This function is described as a composition of functions. As two types of agents
are considered (tropistic agents, i.e., without memory and hysteretic agents, i.e., with
memory3), two types of behavior functions are defined Ferber (1999).
An hysteretic agent ha in a level l acts according to its internal state. Thus, its
behavior function is defined as:
Be havior l
ha
with
◦ Pe r ce pt ionl
ha,
Pe r ce pt ionl
= Deci sionl
ha
◦ M e mor i zat ionha
ha : (cid:89)
∆lP (cid:55)→ (cid:89)
P lP
ha,
(cid:89)
M e mor i zat ionha : (cid:89)
∈N +
lP
(cid:55)→ Sha,
× Sha
P lP
(cid:55)→ (cid:89)
ha
(cid:48)
l∈Lha∈Al
∈N +
∈N +
ΓlI
(l)
(l)
(l)
Deci sionl
lP
lP
.
P
P
P
ha : Sha
∈N +
I
lI
(l)
(14)
(15)
(16)
(17)
There is no memorization function specific to a level. Like in other multi-agent
system meta-models -- e.g., MASQ (Stratulat et al., 2009) -- , we consider that an agent
3While the tropistic/hysteretic distinction is made in IRM, it does not appear clearly in IRM4S. How-
ever, in a multi-level context, it is important if multi-level agents are considered.
5
can have multiple bodies but only one mind (i.e., one internal state). Moreover, the
coherence of the internal state of the agents would have been difficult to maintain
with several memorization functions.
with Pe r ce pt ionl
Be havior l
t a
= Deci sionl
t a
A tropistic agent t a in a level l acts according to its percepts:
◦ Pe r ce pt ionl
t a,
(cid:55)→ (cid:89)
t a : (cid:89)
ω : ∆l (cid:55)→ (cid:89)
t a following the definition of eq. 15 and
ΓlI
Deci sionl
Nat u ral l
∈N +
∈N +
P
P lP
t a
ΓlI
(cid:48)
.
(cid:48)
.
(l)
lP
lI
(l)
I
The environment ω of a level l produces influences through a function:
∈N +
I
lI
(l)
(18)
(19)
(20)
(21)
(22)
(23)
(24)
3.4 Reaction to influences
Once influences have been produced, interactions between levels do not matter any-
more. Thus, the reaction function defined in IRM4S can be re-used:
Reac t ionl : Σl × Γl(cid:48) (cid:55)→ ∆l ,
where Reac t ionl is the reaction function proper to each level.
4 Simulation of IRM4MLS models
In this section, two simulation models for IRM4MLS are proposed. The first one
(section 4.1) is directly based on IRM4S. It supposes that all levels have the same
temporal dynamics. The second one (section 4.2) has a more general scope but is
also more complicated and time consuming. These models are compatible with the
different classical time evolution methods (event-to-event or fixed time step) used in
multi-agent simulation. In the following, t0 and T denote the first and last simulation
times.
4.1 A simple simulation model
In this section, a model with single temporal dynamics is introduced. As there is
no synchronization issue, it is very similar to the model of IRM4S. Eq. 22 to 28
describe this simple temporal model. HA(t) and TA(t) denote respectively the sets
of hysteretic and tropistic agents in the system.
First, behavior sub-functions are executed for each agent:
∀l ∈ L, pa
(< δ lP (t) : lP
∈ N +
(l) >) : a ∈ Al
P
(t) >,
(t) =< Pe r ce pt ionl
a
∀a ∈ HA(t), sa
(t + d t) = M e mor i zat iona
(pa
(t)),
∀l ∈ L,∀a ∈ HAl
(t), < γ lI
a
(cid:48)(t) : lI
∈ N +
I
(l) >= Deci sionl
a
(sa
(t + d t)),
6
∀l ∈ L,∀a ∈ TAl
∈ N +
Then, environmental influences are produced:
(t), < γ lI
a
(cid:48)(t) : lI
I
(l) >= Deci sionl
a
(pa
(t)).
(δ l (t)).
The set of temporary influences in a level l ∈ L at t is defined as:
(l) >= Nat u ral l
ω
ω
(t) : lI
∀l ∈ L, < γ lI
∈ N +
γ l(cid:48)(t) ={γ l (t) (cid:91)
I
(cid:48)(t) (cid:91)
a∈AlI
γ lI
ω
∈N
−
I
lI
(l)
(cid:48)(t)}.
γ lI
a
Finally, the new state of the system can be computed:
∀l ∈ L, δ l (t + d t) = Reac t ionl (σ l (t), γ l(cid:48)(t)).
Algorithm 1 summarizes this simulation model.
(25)
(26)
(27)
(28)
Algorithm 1: simple simulation model of IRM4MLS
Input: < L, EI , EP
Output: δ(T )
1 t = t0;
2 while t ≤ T do
>,A(t0
), δ(t0
)
foreach a ∈ A(t) do
(t) =< Pe r ce pt ionl
pa
if a ∈ HA(t) then
a
end
sa
(t + d t) = M e mor i zat iona
(< δ lP (t) : lP
(pa
(l) >) : a ∈ Al
>;
∈ N +
P
(t));
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
(δ l (t));
(l) >= Deci sionl
a
< γ lI
a
end
foreach l ∈ L do
(cid:48)(t) : lI
∈ N +
(l) >= Nat u ral l
< γ lIω
foreach a ∈ HAl
ω
I
(t) do
∈ N +
(cid:48)(t) : lI
I
end
foreach a ∈ TAl
(t) do
∈ N +
(cid:48)(t) : lI
I
end
γ l(cid:48)(t) ={γ l (t)(cid:83)
end
foreach l ∈ L do
(cid:48)(t)(cid:83)
< γ lI
a
γ lI
a∈AlI
a
δ l (t + d t) = Reac t ionl (σ l (t), γ l(cid:48)(t)) ;
(l) γ lIω
∈N
−
I
lI
(l) >= Deci sionl
a
(sa
(t + d t));
(pa
(t));
(cid:48)(t)};
20
21
22
23 end
end
t = t + d t;
7
4.2 A simulation model with level-dependent temporal dynamics
In this section, a simulation model with level-dependent temporal dynamics is in-
troduced. In the following, t l and t l + d t l denote respectively the current and next
simulation times of a level l ∈ L. Moreover t =< t l : l ∈ L > and t +d t =< t l +d t l :
l ∈ L > denote respectively the sets of current and next simulation times for all levels.
It is mandatory to introduce rules that constraint perceptions, influence production
and reaction computation. These rules rely primarily on the causality principle:
• an agent cannot perceive the future, i.e.,
∀l ∈ L, lP
∈ N +
(l) is perceptible from l if t l ≥ t lP ,
• an agent or an environment cannot influence the past, i.e.,
∀l ∈ L, lI
∈ N +
(l) can be influenced by l if t l ≤ t lI .
P
I
Moreover, a utility principle should also be applied:
• perceptions should be computed at once, i.e.,
∀l ∈ L,∀a ∈ Al , Pe r ce pt ionl
if ∀lP
∈ N +
(l), t l ≥ t lP .
P
a is computed
• as well as influences, i.e.,
∀l ∈ L,Nat u ral l
∈ N +
if ∀lI
I
ω and ∀a ∈ Al , Deci sionl
(l), t l ≤ t lI ∨ t l + d t l < t lI + d t lI .
a are computed
8
However, the causality principle is not sufficient to ensure a good scheduling. A
coherence principle should also guide the conception of the simulation model:
• an agent can only perceive the latest available dynamic states, i.e.,
∀l ∈ L, lP
∈ N +
(31)
• as a hysteretic agent can belong to more than one level, its internal state must
(l) is perceptible from l if t l < t lP + d t lP ,
P
be computed for the next simulation time at which it is considered, i.e.,
∀l ∈ L, sa
(ta
+ d ta
) = M e mor i zat iona
(pa
(t l )),
such as
ta
+ d ta
= t l + d t l∀t l(cid:48) + d t l(cid:48), t l + d t l ≥ t l(cid:48) + d t l(cid:48)
⇒ t l + d t l = t l(cid:48) + d t l(cid:48) ∧ a ∈ Al ,
• an agent or an environment can influence a level according to its latest state,
i.e.,
∀l ∈ L, lI
∈ N +
I
(l) can be influenced by l if t l + d t l > t lI ,
(34)
• reaction must be computed for the next simulation time, i.e.,
∀l ∈ L, Reac t ionl is computed if t l + d t l ∈ min(t + d t).
(29)
(30)
(32)
(33)
(35)
(36)
(37)
It is easy to show that the rule defined in eq. 36 subsums the rule defined in
eq. 29. Moreover, the rule defined in eq. 35 implies the rule defined in eq. 31.
According to eq. 37, influences are not necessarily produced at each time from a
(l). Thus, a function cI , defines influence production from
∈ N +
I
level l to a level lI
the rules defined by the eq. 34 and 36:
∀l,∈ L,∀lI
γ lI
∈ N +
(l), cI
(l, lI
) =
I
(cid:48)(t lI )
(cid:59)
if
else.
t l ≤ t lI ∧ t l + d t l > t lI
(38)
The simulation model can then be defined as follows. First, if the condition
defined in the eq. 36 is respected, agents construct their percepts and consecutively
hysteretic agents compute their next internal state:
∀a ∈ A(t),
pa
+ d ta
(t l ) =< Pe r ce pt ionl
a
) = M e mor i zat iona
∈ N +
(ta
sa
={l ∈ L : a ∈ Al
(t)∧∀lP
(< δ lP (t lP ) : lP
(pa
(l), t l ≥ t lP}.
P
(t l )) if a ∈ HA(t),
P
∈ N +
(l) >) : l ∈ LP
>,
(39)
(40)
Then, if the condition defined in eq. 37 is respected, agents and environments
with LP
produce influences:
∀l ∈ LI ,
I
∈ N +
∈ N +
∈ N +
I
(l) > = Nat u ral l
ω
(l) > = Deci sionl
a
(l) > = Deci sionl
a
(δ l (t l )),
(l, lI
) : lI
< cI
∀a ∈ HAl , < cI
(ta
(sa
(l, lI
) : lI
∀a ∈ TAl , < cI
(pa
(t l )),
(l, lI
) : lI
(l), t l ≤ t lI ∨ t l + d t l < t lI + d t lI}.
∈ N +
={l ∈ L :∀lI
γ l(cid:48)(t l ) ={γ l (t l ) (cid:91)
The set of temporary influences in a level l ∈ L at t l is defined as:
(lI , l)}.
cI
I
I
+ d ta
)),
(41)
(42)
(43)
(44)
with LI
∈N
−
I
lI
(l)
Finally, reactions are computed for levels that meet the condition defined in
eq. 35:
∀l ∈ LR,
with LR
δ l (t l + d t l ) = Reac t ionl (σ l (t l ), γ l(cid:48)(t l )),
(45)
={l ∈ L : t l + d t l ∈ min(t + d t)}.
9
The algorithm 2 summarizes this simulation model.
>,A(t0
Algorithm 2: simulation model of IRM4MLS with level-dependent temporal
dynamics
Input: < L, EI , EP
Output: δ(T )
1 foreach l ∈ L do
2
3 end
4 while ∃t l ≤ T do
t l = t0;
), δ(t0
)
foreach a ∈ A(t) do
(l) >) : l ∈ LP
>;
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
P
(pa
I
(ta
sa
(l, lI
∈ N +
(t)∧∀lP
(< δ lP (t lP ) : lP
) = M e mor i zat iona
(l), t l ≥ t lP};
∈ N +
P
(t l ));
={l ∈ L : a ∈ Al
LP
(t l ) =< Pe r ce pt ionl
pa
if a ∈ HA(t) then
a
+ d ta
end
end
(l), t l ≤ t lI ∨ t l + d t l < t lI + d t lI};
∈ N +
={l ∈ L :∀lI
LI
foreach l ∈ LI do
∈ N +
) : lI
(l) >= Nat u ral l
< cI
foreach a ∈ HAl
ω
I
(t) do
∈ N +
) : lI
I
end
foreach a ∈ TAl
(t) do
∈ N +
) : lI
I
end
end
={l ∈ L : t l + d t l ∈ min(t + d t)};
γ l(cid:48)(t l ) ={γ l (t l )(cid:83)
LR
foreach l ∈ LR do
(lI , l)};
(l) >= Deci sionl
a
(l) >= Deci sionl
a
(δ l (t l )) ;
(sa
(ta
(t l ));
< cI
(l, lI
< cI
(l, lI
(pa
∈N
−
I
lI
(l) cI
δ l (t l + d t l ) = Reac t ionl (σ l (t l ), γ l(cid:48)(t l ));
t l = t l + d t l ;
+ d ta
));
25
26
27
28 end
end
5 Discussion, conclusion and perspectives
In this paper, a meta-model of ML-ABM, called IRM4MLS, is introduced.
It is
designed to handle many situations encountered in ML-ABM: hierarchical or non-
hierarchical multi-level systems with different spatial and temporal dynamics, multi-
level agents or environments and agents that are dynamically introduced in levels.
Moreover, IRM4MLS relies on a general simulation model contrary to the exist-
ing works published in literature. While this model is, in general, complicated, its
implementation could be simplified to be more efficient in specific situations (sin-
gle perception function, reactive simulation, etc.). Afterwards, examples of typical
ML-ABM situations as well as ideas to treat them in the context of IRM4MLS are
10
presented.
In some models an agent can belong to different levels:
• in the model of bio-inspired automated guided vehicle (AGV) systems pre-
sented in Morvan et al. (2009), an AGV (a micro level agent) can become a
conflict solver (a macro level agent) if a dead lock is detected in the system,
• in the SIMPOP3 multi-level model an agent representing a city plays the role
of interface between two models and then is member of two levels (Pumain
et al., 2009).
The simulation of these models has been addressed using different strategies:
• in the first example (a control problem), a top-first approach is used: the higher
level takes precedence over the lower one,
• in the second example (a simulation problem), levels are executed alternately.
These solutions are context-dependent and likely to generate bias.
In IRM4MLS,
the multi-level agent situation is handled by a single simulation model that general-
izes the two previous ones without scheduling bias, thanks to the influence/reaction
principle.
In many multi-level agent-based models, interactions between entities in a level
affect the population of agents in another level. E.g., in RIVAGE, a model of runoff
dynamics, macro level agents (representing water ponds or ravines) emerge from mi-
cro level agents (representing water balls) when conditions are met (Servat et al.,
1998). Then, the quantity and the flow of water become properties of macro level
agents: water balls are no longer considered as agents. Conversely, micro level agents
can emerge from macro level agents. Similar situations can be found in hybrid mod-
eling of traffic flows (El hmam et al., 2006). In IRM4MLS, the introduction of an
agent a in a level l is performed by the reaction function of l that introduces envi-
ronmental properties representing the physical state of a in σ l (t). Conversely, the
reaction function can delete an agent from the level. An agent that does not belong
to any level is inactive but can be reactivated later.
Finally, the definition of IRM4MLS is not closed in order to offer different pos-
sibilities of implementation or extension. E.g., levels could be defined a priori or
discovered during the simulation (Gil-Quijano et al., 2009). While this approach has
never been used in any model so far, it seems particularly promising. In IRM4MLS,
only the first possibility has been handled so far. It would be necessary to consider
L and < L, EI
> as dynamic directed graphs.
The two main perspectives of this work are the design of a modeling and simula-
tion language and a platform that comply to the specifications of IRM4MLS as well
as the re-implementation of existing models to demonstrate the capabilities of the
meta-model and its simulation models.
> and < L, EP
Acknowledgments
Authors would like to thank Javier Gil-Quijano (LIP6 -- Université Paris VI, France),
Fabien Michel (LIRMM -- Université Montpellier 2, France), Daniel Jolly (LGI2A --
Université d'Artois, France) and Luce Desmidt (HEI -- France) for their help and
support.
11
References
An, G. (2008). Introduction of an agent-based multi-scale modular architecture for
dynamic knowledge representation of acute inflammation. Theoretical Biology and
Medical Modelling, 5(11).
Chen, C., Nagl, S., and Clack, C. (2009). A formalism for multi-level emergent
behaviours in designed component-based systems and agent-based simulations. In
Aziz-Alaoui, M. and Bertelle, C., editors, From System Complexity to Emergent
Properties, volume 12 of Understanding Complex Systems, pages 101 -- 114. Springer.
David, D. and Courdier, R. (2009). See emergence as a metaknowledge. a way to reify
In Proceedings of ICAART'09,
emergent phenomena in multiagent simulations?
pages 564 -- 569, Porto, Portugal.
Drogoul, A., Vanbergue, D., and Meurisse, T. (2003). Multi-agent based simulation:
Where are the agents?
In Sichman, J., Bousquet, F., and Davidsson, P., editors,
Multi-Agent-Based Simulation II, volume 2581 of Lecture Notes in Computer Science,
pages 1 -- 15. Springer.
El hmam, M., Abouaissa, H., Jolly, D., and Benasser, A. (2006). Macro-micro
simulation of traffic flow. In Proceedings of the12th IFAC Symposium on Informa-
tion Control Problems in Manufacturing (INCOM06), pages 351 -- 356, Saint Etienne
FRANCE.
Ferber, J. (1999). Multi-Agent Systems: An Introduction to Distributed Artificial Intel-
ligence. Addison-Wesley Longman Publishing Co., Inc.
Ferber, J. and Müller, J.-P. (1996). Influences and reaction: a model of situated multi-
agent systems. In 2nd International Conference on Multi-agent systems (ICMAS'96),
pages 72 -- 79.
Gil-Quijano, J., Hutzler, G., and Louail, T. (2009). De la cellule biologique à la
cellule urbaine : retour sur trois expériences de modélisation multi-échelles à base
d'agents.
In Actes des 17èmes Journées Francophones sur les Systèmes Multi-Agents
(JFSMA'09).
Gil-Quijano, J., Piron, M., and Drogoul, A. (2008). Social Simulation: Technologies,
Advances and New Discoveries, chapter Mechanisms of Automated Formation and
Evolution of Social-Groups: A Multi-Agent System to Model the Intra-Urban
Mobilities of Bogota City. IGI Global.
Hoekstra, A., Lorenz, E., Falcone, J.-L., and Chopard, B. (2007). Towards a complex
automata framework for multi-scale modeling: Formalism and the scale separation
map.
In Computational Science -- ICCS 2007, volume 4487 of Lecture Notes in
Computer Science, pages 922 -- 930. Springer.
Lepagnot, J. and Hutzler, G. (2009). A multiscale agent-based model for the sim-
ulation of avascular tumour growth. Journal of Biological Physics and Chemistry,
9(1):17 -- 25.
Maturana, H. R. and Varela, F. J. (1980). Autopoiesis and Cognition The Realization of
the Living, volume 42 of Boston Studies in the Philosophy of Science. Springer Berlin
Heidelberg.
12
Michel, F. (2007a). The IRM4S model: the influence/reaction principle for mul-
In Proc. of 6th Int. Conf. on Autonomous Agents and
tiagent based simulation.
Multiagent Systems (AAMAS 2007), pages 1 -- 3.
Michel, F. (2007b). Le modèle IRM4S. de l'utilisation des notions d'influence et de
réaction pour la simulation de systèmes multi-agents. Revue d'Intelligence Artifi-
cielle, 21:757 -- 779.
Morvan, G., Jolly, D., Veremme, A., Dupont, D., and Charabidze, D. (2008). Vers
In Actes de la 7ème Conférence de
une méthode de modélisation multi-niveaux.
Modélisation et Simulation MOSIM, Paris, France, volume 1, pages 167 -- 174.
Morvan, G., Veremme, A., Dupont, D., and Jolly, D. (2009). Modélisation et con-
ception multiniveau de systèmes complexes : stratégie d'agentification des organi-
sations. Journal Européen des Systèmes Automatisés, 43(4-5):381 -- 406.
Prévost, G. and Bertelle, C. (2009). Detection and reification of emerging dynamical
ecosystems from interaction networks. In Complex Systems and Self-organization
Modelling, volume 39 of Understanding Complex Systems, pages 139 -- 161. Springer.
Pumain, D., Sanders, L., Bretagnolle, A., Glisse, B., and Mathian, H. (2009). The
future of urban systems: Exploratory models.
In Lane, D., Pumain, D., der
Leeuw, S. V., and West, G., editors, Complexity Perspectives in Innovation and
Social Change, volume 7 of Methodos Series, pages 331 -- 360. Springer.
Servat, D., Perrier, E., Treuil, J.-P., and Drogoul, A. (1998). When agents emerge
from agents: Introducing multi-scale viewpoints in multi-agent simulations.
In
Multi-Agent Systems and Agent-Based Simulation, volume 1534 of Lecture Notes in
Computer Science, pages 183 -- 198. Springer.
Stratulat, T., Ferber, J., and Tranier, J. (2009). Masq: towards an integral approach
to interaction. In AAMAS '09: Proceedings of The 8th International Conference on
Autonomous Agents and Multiagent Systems, pages 813 -- 820, Richland, SC. Interna-
tional Foundation for Autonomous Agents and Multiagent Systems.
Zhang, L., Wang, Z., Sagotsky, J., and Deisboeck, T. (2009). Multiscale agent-based
cancer modeling. Journal of Mathematical Biology, 48(4 -- 5):545 -- 559.
13
|
cs/0509037 | 1 | 0509 | 2005-09-14T07:30:42 | Friends for Free: Self-Organizing Artificial Social Networks for Trust and Cooperation | [
"cs.MA"
] | By harvesting friendship networks from e-mail contacts or instant message "buddy lists" Peer-to-Peer (P2P) applications can improve performance in low trust environments such as the Internet. However, natural social networks are not always suitable, reliable or available. We propose an algorithm (SLACER) that allows peer nodes to create and manage their own friendship networks.
We evaluate performance using a canonical test application, requiring cooperation between peers for socially optimal outcomes. The Artificial Social Networks (ASN) produced are connected, cooperative and robust - possessing many of the disable properties of human friendship networks such as trust between friends (directly linked peers) and short paths linking everyone via a chain of friends.
In addition to new application possibilities, SLACER could supply ASN to P2P applications that currently depend on human social networks thus transforming them into fully autonomous, self-managing systems. | cs.MA | cs | Friends for Free:
Self-Organizing Artificial Social Networks for Trust and Cooperation1
David Hales ([email protected]), Stefano Arteconi ([email protected])
Dept. of Computer Science,
University of Bologna, Italy
August 2005
Abstract
By harvesting friendship networks from e-mail contacts or instant message “buddy lists”
Peer-to-Peer (P2P) applications can improve performance in low trust environments such
as the Internet. However, natural social networks are not always suitable, reliable or
available. We propose an algorithm (SLACER) that allows peer nodes to create and
manage their own friendship networks.
We evaluate performance using a canonical test application, requiring cooperation
between peers for socially optimal outcomes. The Artificial Social Networks (ASN)
produced are connected, cooperative and robust - possessing many of the disable
properties of human friendship networks such as trust between friends (directly linked
peers) and short paths linking everyone via a chain of friends.
In addition to new application possibilities, SLACER could supply ASN to P2P
applications that currently depend on human social networks thus transforming them into
fully autonomous, self-managing systems.
Keywords: Trust, Cooperation, P2P, Social Networks, Self-Organization.
1. Introduction
It is well known that human social relationships form a social network that spans the
entire globe. These networks form a small-world topology, in which highly connected
clusters of mutual friends are linked to other clusters by individuals that form “social
bridges” or hubs. The upshot of this, as has been dramatically demonstrated (Travers et al
1969), is that it is often possible to find a short chain of friends of friends to anyone on
the globe – you are probably linked to the President of the United States of America by
no more than about six hops in your friendship network.
Friendship networks have several desirable properties, they tend to be cooperative, we
trust our friends, and they support a number of social functions that we need to achieve in
everyday life – we can ask friends for favors or cooperatively solve a difficult problem
with work colleagues. We can turn to friends for good advice or discuss with them
1 This work partially supported by the EU within the 6th Framework Program under
contract 001907 (DELIS).
confidential issues. Furthermore, they are constructed and maintained in a completely
distributed way – there’s no central authority decreeing who should be part of your
friendship network.
It is for these reasons that, increasingly, these networks are being “harvested” and used to
solve tricky collaborative tasks in computer networks. For example, to improve P2P
message routing (Martini et al 2005) - because friends computers can be trusted to pass
on messages - or to support collaborative spam filtering (Kong et al 2004) – if my friends
computer says “it’s spam” then my computer believes its spam.
1.1 The Need for Artificial Social Networks
However, these systems import existing social networks that are built and maintained in
the everyday social world by people. This takes time and they will tend to be biased
towards the social goals of the people involved. For example, in the office the majority of
your e-mail contacts are likely to be work colleagues but on a home computer the bias
might be towards friends and family. In either case, who is to say that the trust
established between friends in a human social context necessarily translates into useful
trust in some computer application? My friends computer may say “it’s spam” but his
machine might be incorrectly configured or infected with malicious code that makes its
advice untrustworthy. Personally, I do not make friends based on their eagerness to
download security patches!
Our approach is to import network formation processes so applications can maintain their
own social networks tuned to their own goals. Put another way, we get computers to
make their own trusted friends - “friends for free”.
Researchers have been trying to solve these problems for some time, namely, how to get
cooperation and trust between computers sharing open networks when there is no central
authority or control. One effect of the current lack of such trust are nasty problems that
plague the internet that most of us probably know all too well; proliferation of e-mail
spam, the spread of malicious viruses, worms, trojans, malware and spyware.
1.2 Overview
We introduce a simple algorithm (SLACER) which, when executed in the nodes of a P2P
network, self-organizes the network into a robust trusted ASN with small-world
characteristics and high cooperation. The algorithm is an extension of the SLAC2
algorithm (Hales 2004, Hales et al 2005) based on the “tagging” approach originating
from Computational Sociology (Riolo et al 2001, Hales 2000) that has been shown to
support high-levels of cooperation without the need for central control, reciprocity or
other evaluation mechanisms. It is based on some simple rules of social behavior
observed in human societies.
2 Which bares comparison to SLIC (Sun et al 2004).
We tested SLACER’s ability to produce cooperative and connected ASN by having
nodes play the Prisoner’s Dilemma (PD) game. In the PD there is always a temptation for
players to betray their partner for personal gain, so this is a good test of trust or “healthy
friendship” between nodes.
SLACER produced high trust ASN that connected the entire population of nodes with
almost all nodes reachable from any other via chains of cooperative nodes. The ASN
were robust and scalable even though nodes act in a selfish individualistic way – If they
see another node outperforming them, then they try to copy their behavior and link to
their friends – rather like a greedy social climber.
We briefly introduce P2P overlay networks, describe the SLACER algorithm in detail,
and then present the PD test application used for evaluation. After this we give simulation
results and indicate a number of practical application areas for which, we believe, the
SLACER algorithm provides a sound basis.
2. Peer-to-Peer Overlay Networks
The target infrastructures we have in mind for the SLACER algorithm are unstructured
peer-to-peer (P2P) overlay networks. In a P2P overlay network there is a population of
nodes, typically processes situated within a physical network, which maintain links to
other nodes (often called their neighbors). P2P applications, like Skype3 or BitTorrent4
implement these to provide services. When we say a node is “linked” in an overlay net
we do not mean there is a physical link between them, but rather an address or label is
stored in the nodes that allows them to communicate. So for example, an overlay net
running on the Internet might represent links as IP addresses. Each node can store some
number of IP’s of other nodes in the network. An overlay net therefore allows us to
abstract away from the physical infrastructure supporting the communication between
nodes. We simply assume that such an infrastructure layer exists. Hence, if two nodes are
linked in the overlay this indicates nothing about their physical network proximity, rather,
the overlay describes a logical topology that runs “on-top” of any physical network that
can provide the necessary communication services such as reliable rooting between all
nodes.
In the work presented here, we also make a further abstraction; we assume that such an
overlay network can be modeled as an undirected graph. In such a graph, all links
(termed edges) between nodes (termed vertices) are undirected. In the context of a P2P
we interpret this as a constraint indicating that all links between nodes are symmetric – if
a node i links to a node j then node j also links to node i. This means we only represent
bidirectional communication links in our overlay and we assume the underlying physical
network can provide this service.
3 www.skype.com
4 www.bittorrent.com
One valuable property of the overlay net abstraction is that rewiring nodes or changing
the topology of the network is a logical process in which nodes simply drop, copy or
exchange symbolic links. It is therefore not costly to maintain highly dynamic network
topologies at the overlay layer. This makes the manipulation of dynamic networks in real
time feasible.
3. The SLACER Algorithm
The SLACER algorithm is an extension of the SLAC algorithm (Hales 2004) – a simple
node rewiring algorithm that produces cooperation when nodes behave selfishly. It’s a
Selfish Link-based Adaptation for Cooperation algorithm (SLAC). SLACER follows the
same general approach but is more conservative because it maintains some old links
rather than rewiring all links – it’s a Selfish Link-based Adaptation for Cooperation
Excluding Rewiring algorithm (SLACER). These acronyms actually do mean things!
We assume that peer nodes can change their strategy (i.e. change the way they behave at
the application level) and drop and make links to nodes they know about. It is also
assumed nodes have the ability to discover other nodes randomly from the entire
population, compare their performance against those nodes in some way and copy their
links and strategies5. SLACER implements a simple local adaptation rule: Nodes try to
use their abilities to selfishly increase their own performance (or utility) in a greedy and
adaptive way by changing their links and strategy. They do this by copying nodes that
appear to be performing better and by making randomized changes with low probability.
Figure 1 shows the pseudocode. Over time, nodes engage in some application task and
generate some measure of utility U. This utility is a numeric value that each node needs
to calculate based on the specifics of the particular application domain. For example, this
might be number of files downloaded, jobs processed or an inverse measure of spyware
infections over some period. The higher the value of U the better the node believes it is
performing in its target domain.
5 For a discussion of how we might deal with malicious noise in the exchange of
information between nodes see the conclusion.
Active thread:
Passive thread:
i ← this node
do forever:
Engage in application task
update i.Utility
Periodically (compare utility):
j ← GetRandomNode()
j.GetState(i)
if i.Utility ≤ j.Utility
CopyStatePartial(j)
Mutate(i)
Utility ← 0.0 (reset utility)
j ← this node
do forever:
sleep until a request received
GetState(i) – send j state to node i:
Send j.Utility to i
Send j.Links to i
Send j.Strategy to i
Function CopyStatePartial(j):
Function Mutate(i):
i.Strategy ← j.Strategy
drop each link from i with prob. W
for each link in j.Links:
i.addLink(link)
with prob. M mutate i.Strategy
with prob. MR mutate i.Links:
drop each link with prob. W
i.addLink(SelectRandomNode())
Figure 1. The SLACER protocol runs continuously in each node. It is composed of an
active and passive thread. The active thread is constantly executed; the passive thread
wakes only to serve requests for information from other nodes. The GetRandomNode()
function supplies a random node from the entire population. Three probabilities: W, M,
MR defines a space of protocols. When W = 1 the protocol collapses to the previous
SLAC protocol, of which SLACER is an extension
Periodically, each node (i) compares its performance against another node (j), randomly
selected from the population. If Ui ≤ Uj then node i drops each of its current links to
other nodes with high probability W, and copies all node j links and adds a link to j itself.
Additionally i then copies j’s strategy – the strategy represents some behaviour that nodes
execute during application level interaction. After such a copy operation has occurred,
then, with low probability M, node i adapts its strategy and with probability MR adapts its
links. Adaptation involves the application of a “mutation” operation. Mutation of the
links involves removing each existing link with probability W and adding a single link to
a node randomly drawn from the network. Mutation of the strategy involves applying
some form of change in application behavior with probability M - the specifics of strategy
mutation are dictated by the application domain (see later). After the periodic utility
comparison, whether this resulted in the copying of another node or not, the node resets
its utility to zero - hence, utilities do not accumulate indefinitely.
Each node is limited to a maximum number of links or neighbors (its so-called view
size). If any SLACER operation causes a node to require an additional neighbor above
this limit then a randomly selected existing link is removed to make space for the new
link. Links are always undirected and symmetrical, so that if node i links to node j, then j
must also maintain a link to node i and conversely if node i breaks a link to node j then
node j also breaks its link to node i. For implementation purposes, SLACER requires
additional passive thread functions (not shown in figure 1) that would handle “addLink
and dropLink” requests from other nodes.
Previous “tag” models, on which SLACER is based (Hales 2000) have indicated that the
rate of mutation applied to the links needs to be significantly higher than that applied to
the strategy by about one order of magnitude hence MR >> M.
When applied in a suitably large population, over time, the algorithm follows an
evolutionary process in which nodes with high utility replace nodes with low utility.
However, as will be seen, this does not lead to the dominance of selfish behavior, as
might be intuitively expected, because a form of social incentive mechanism results from
the emergent network topology (a friendship network). This means that high utility but
anti-social strategies, even though favored by the individual nodes, do not dominate the
population. The topology therefore guides the adaptation of the strategy away from anti-
social behaviors.
3.1 Random Sampling via NEWSCAST
In Figure 1, we assume a function that returns a random node from the entire population
of nodes irrespective of the current network topology. This function cannot use the
network maintained by SLACER itself since it can become partitioned . In our
simulations we used the existing NEWSCAST algorithm (Jelasity et al 2004) in order to
provide this service. NEWSCAST provides exactly this function by maintaining its own
scalable and robust random overlay network based on a gossip protocol – in which
random neighbors constantly gossip their network views (the links they maintain to other
nodes). NEWSCAST maintains a random and fully connected topology even under
conditions of high node failure and malicious behavior. Since NEWSCAST maintains its
own overlay it can be deployed to support SLACER in a modular way – the only
interaction necessary is via a GetRandomNode() function invoked from SLACER to the
lower newscast layer and returning a randomly sampled node of the network.
4. The Prisoner’s Dilemma Test Application
The two player single-round Prisoner’s Dilemma (PD) game captures, in an abstract
form, a situation in which there is a contradiction between collective and individual self-
interest. Two players interact by selecting one of two choices: to "cooperate" (C) or
"defect" (D). For the four possible outcomes of the game, players receive specified
payoffs. Both players receive a reward payoff (R) and a punishment payoff (P) for mutual
cooperation and defection respectively. However, when individuals select different
moves, different payoffs of temptation (T) and sucker (S) are awarded to the defector and
the cooperator respectively. Assuming that neither player can know in advance which
move the other will make and wishes to maximize her own payoff, the dilemma is
evident in the ranking of payoffs: T > R > P > S and the constraint that 2R > T + S.
Although both players would prefer T, because its the highest payoff, only one can attain
it in a single game. No player wants S because it’s the lowest payoff. No matter what the
other player does, by selecting a D-move a player ensures she gets either a better (T) or
an equal (P) payoff to her partner. Playing D, then, can't be bettered since this ensures
that the defector cannot be suckered (S). As any modern Economist would no doubt tell
you, this is the so-called "Nash" equilibrium for the single round game – hence an ideally
rational selfish player would always choose D. It is also an evolutionary stable strategy
(ESS) for a population of randomly paired players, where reproduction fitness is
proportional to payoff, so Darwin can’t save us from the Economists this time!
Therefore, the dilemma is that if both individuals selected a cooperative (C) move they
would be jointly better off (getting R each) than if they both select D, but both
evolutionary pressure and ideal rationality result in mutual defection, so the players only
get P each, because there is always an individual incentive to select defection.
We select this game as a minimal abstract basis for a test “application” that captures,
abstractly, a whole range of possible application tasks in which nodes need to establish
cooperation and trust with their neighbors but without central authority or appeal to
external mechanisms. The PD payoff structure gives each node an individual incentive to
defect while specifying a social (joint) incentive to cooperate. We can relate this to a P2P
application task where individual nodes need to behave in a non-selfish way to increase
network level performance. This could include altruistically sharing files and resources or
passing on a message, to facilitate communication between a sender and receiver node, or
using resources warning friends about a virus program and supplying them with a fix
when you’ve already inoculated yourself against the infection.
Our test application involves all nodes playing the PD with randomly selected neighbors
in the SLACER constructed social network. A node can only choose one of two strategies
– cooperate or defect. The utility value required by SLACER is then set by the PD
application as the average payoff the node received from these game interactions. The
SLACER algorithm then adapts the links and strategy of the nodes as discussed
previously.
In past work, using the SLAC algorithm, we found that results form a similar PD test
application were good indicators of results obtained when the same algorithm was tested
in a more realistic P2P file-sharing scenario (Hales et al 2005) – where free-riding (or
leeching) was the analogue of defection and altruistically sharing files represented
cooperation. We therefore use the PD here as a minimal or canonical test from which
initial evaluations of SLACER can be obtained.
5. Measuring Cooperative Connected Paths (CCPs)
Our goal in designing SLACER is to self-organize artificial social networks (ASN) with
some of the disable properties we observe in human social networks that would be of
value in many P2P application domains. These desirable properties comprise: short paths
between all members, high cooperation between neighbors and paths of cooperation
between non-directly connected members.
In order to evaluate the quality of the produced ASNs we measured, over the entire
population of nodes: average path length, proportion of cooperator nodes and the
proportion of nodes linked via chains of cooperators. For the latter we introduce a
Cooperatively Connected Path (CCP) measure that quantifies the amount of “cooperative
connectedness” in an ANS.
Given a population of nodes in which each node is in one of two states (cooperate or
defect) we define the CCP measure as the proportion of all possible node pairs that are
either linked directly or for which there exists at least one path between them in which all
intermediate nodes are in a cooperative state. This definition does not require that the
node pairs themselves are cooperative, just that a cooperative path exists between them,
or a direct link. See figure 2 for an example.
We can illustrate this concept with a hypothetical message passing game in a human
social network. Suppose I wish to pass a confidential message to a distant individual to
which I am not directly known (the President of the United States, say) and I want the
message to arrive without being compromised (read or changed). It would not be possible
to achieve this if I could not find a route of high trust and cooperation through my social
network from myself to the final recipient. That is, I would need to pass the message via
a set of intermediaries who could be trusted to act cooperatively. However, this task does
not require that either the sender or final recipient behave cooperatively if they are passed
messages for others.
In a connected network in which all nodes were cooperators, the CCP measure would
obviously be 1. However, even with a number of defector nodes inhabiting such a
network the CCP can still be 1, so long as the defectors were located in such a way that
they do not obstruct other cooperative routes between pairs of nodes – i.e. there are
alternative cooperative routes around the defectors. Hence, the CCP measure is
determined by the topology of the network and the strategy distribution over nodes in that
topology (the strategic topography).
Figure 2. The CCP measure for a small network.
Here we consider the shaded nodes to be defector
nodes. Since there are 6 nodes in the population the
total number of node pairs is 15, but only the
following node pairs have cooperatively connected
paths: {a,b}, {b,c}, {b,d}, {b,e}, {b,f}, {c,f}, {d,e},
{e,f}, hence the CCP measure = 8/15 ≈ 0.533.
By combining the CCP measure with other measures, we can get an idea of the strategic
topography of the ASN. For example, if the network was connected (not partitioned) and
the proportion of cooperating nodes was high but the CCP was low, then we would
conclude that a few non-cooperative nodes occupy positions in the network that allow
them to block a large number of unique cooperative paths between nodes.
Going back to our design goal for SLACER, our intention is to generate connected
networks of cooperative nodes, so our ideal outcome would be ASN with CCP at or close
to the maxim of 1.
6. Simulation Specifics
We performed simulation experiments using two independent implementations and
verified the results matched (we “docked” the models). We did this to minimize the effect
of possible implementation artifacts. All the observations and conclusions we draw are
supported by the results obtained from both implementations6.
6.1 One Time Cycle
When nodes engage in their application task this involves playing the PD with randomly
selected neighbors then nodes periodically initiate the “compare utility” activity in the
active thread of the SLACER algorithm. We programmed the simulations so that, on
average, over one cycle, each node initiates some application level activity, causing a
6 An implementation in the PEERSIM environment is available at:
http://peersim.sourceforge.net
utility update and executes one “compare utility” call. We simulated this both semi-
asynchronously and fully asynchronously.
In one time cycle , 10N nodes are selected from the population randomly with
replacement. When a node is selected, it chooses a neighbor at random and plays a
single-round game of the PD. If a node is selected that currently has no neighbors then a
link is made to a randomly selected node so a game can be played. A game is played by
both nodes executing the strategy indicated by their current state (cooperate or defect).
Both nodes then increment their utility by the appropriate payoff. If both cooperated then
both increase their utilities by R, if both defect then both increment by P, otherwise the
cooperating node increments by S and the defecting node gets T. When nodes compare
utility in the SLACER algorithm the average utility over the total number of games
played is used rather than the absolute utility.
6.2 Two Simulation Variants
In the semi-asynchronous simulation implementation, after nodes had played their PD
games, N nodes were selected, randomly, with replacement, to execute the “compare
utility” operation (see Figure 1). A new cycle was then started. In the fully asynchronous
version the same operation was executed probabilistically by each node after each PD
game was played – with probability 0.1.
What we model here is the notion that for some possible application tasks utility updates
may be instantaneous and asynchronous, for example, neighbor nodes may be able to
provide requested resource immediately in some application tasks. However, in others
rewards may be delayed and synchronized over a large set of nodes – especially when a
number of nodes need to coordinate together to achieve a collective task and only then
get their reward – consider a P2P search engine pooling its page rankings, for example.
Obviously many different schemes are possible which would be dictated by the specifics
of an application task but for the purposes of the initial PD test application, we
considered these two.
6.3 Parameter Settings
We set the PD payoffs to T = 1.9, R = 1, P = 2d, S = d, where d is some small value (in
this case 10-4). We set the mutation rates to M = 0.001 and MR = 0.01. Nodes were
allowed a maximum view size of 20 (hence each node could link to maximum of 20
neighbors). We ran simulation with the rewire probability W = 0.9 and W = 1 over a
range of network sizes from N = 2,000 to 64,000.
7. Simulation Results
In our simulation experiments, we initialized the entire population of N nodes to the state
of defection and connected them in a random network topology (we found, in fact, that
any initial network topology, including fully disconnected, had no significant effect on
the outcomes of the simulations). We then measured how long it took in time cycles for
high levels of cooperation to emerge - we stopped when 98% of all nodes were in a
cooperative state. We found cooperation and topology measures meta-stabilized after this
point, randomly oscillating around mean values. Complete stability is not possible
because nodes always continue to move and change strategies within the network.
When a state of high-cooperation was reached, we measured the CCP for the entire
population along with the average path length, the clustering coefficient and the largest
single connected component. For each experiment, we performed a number of runs (10)
with identical parameters but different pseudo-random number generator seeds. In our
analysis we took into account the averages and variances over these runs.
We compared two “rewire” probabilities (see figure 1) W = 0.9 and W = 1. When W = 1,
this eliminates completely the storage of old links when a node copy occurs and this
collapses the algorithm to the previous SLAC algorithm – which we know, from previous
work, produces highly partitioned, though cooperative, networks.
7.1 Cycles to High Cooperation
Figure 3 shows the number of cycles to high cooperation from an initial population of all
defector nodes. As can be seen SLACER takes longer than the original SLAC algorithm
to converge to high cooperation but only by a few tens of cycles. It is an important test to
ensure a network can recover high cooperation from total defection quickly, since this
ensures robustness against a possible catastrophic failure in cooperation caused by other
means. We have an existence proof that the algorithm can escape from this dire situation
quickly. We got similar results to Figure 3 when we initialized the nodes without any
links at all; this indicates the ability of SLACER to recover from a total network outage
quickly.
7.2 SLAC Networks
Figure 4 shows the size of the largest connected component and the CCP value when
W=1, for various network sizes, after high cooperation emerges. The largest component
size does not grow as the network size increases and hence this degrades the CCP value.
The CCP value starts small and does not scale. This is because the nodes are partitioned
into numerous small components; although nodes are cooperative to their neighbors, the
partitions stop cooperative paths being made between nodes in different components. The
SLAC algorithm, then, leads to ASN that suffer from something that might be termed
“extreme tribalism”. It’s rather like the human equivalent of a population composed
entirely of self-absorbed cliques or cults that, although internally cooperative, want
nothing to do with each other.
Figure 3. Average number of cycles to reach high cooperation for various network sizes
(N). Values are shown for rewire probability W=0.9 (SLACER) and W=1 (SLAC). Each
point is an average of 10 runs. The variation over individual runs was never more than
about 10 cycles either side of the mean. As can be seen, on average, it takes no more than
an additional 40 cycles for high cooperation to be achieved when W=0.9.
Figure 3. The size of the largest component as a proportion of population size and the
Cooperative Connected Path (CCP) measure for W=1 (the SLAC algorithm) after high
cooperation has been achieved. Each bar shows an average of 10 runs - variance was
negligible. The curve for the largest component indicates that its absolute size remains
approximately constant (about 150 nodes).
Figure 5. The size of the largest component as a proportion of population size and the
Cooperative Connected Path (CCP) measure when W=0.9 (SLACER). Each bar shows an
average of 10 runs. It can be seen that almost all nodes inhabit a giant connected
component (GCC). The high CCP indicates that the vast majority of nodes within the
GCC are connected by cooperative paths. These results can be compared to Figure 4.
Figure 6. The clustering coefficient (C) and average path length (L) when W=0.9 and
after high cooperation is attained for different network sizes (N). Note that while C stays
more-or-less constant, L increases log-linearly with N. Each point is an average value of
10 simulation runs. The variance of results was negligible.
Figure 7. A typical individual simulation run for a 2000 node network. Shown are the
Clustering Coefficient (C), the average path length (L), The Cooperative Connected Paths
(CCP) measure and the average Cooperative Connected Path Length (CCPL). This latter
measure gives the average path length of cooperative connected paths. Notice that it
peaks as CCP starts to rise then settles down to agreement with L. This shows that when
cooperation first emerges it spreads quickly over the network – forming a kind of
extended “cooperative backbone” which soon seeds the entire population, reducing the
CCPL to the L value.
7.3 SLACER Networks
The results in figure 4 can be compared to those given in figure 5, in which W=0.9. Here
we see very high values for both CCP and the largest component size. Almost all nodes
inhabit a giant connected component (GCC) that, although containing some defecting
nodes, provides cooperative routes between the large majority of its members. SLACER
therefore generates cooperative ASN without the “extreme tribalism” produced by
SLAC. Importantly, the CCP scales well because as N increases CCP does not decrease.
This gives us confidence that SLACER would produce high quality ASN for populations
of any practical size.
7.4 Basic Topological Features
In figure 6, we measure some other structural characteristics of the ASN produced by
SLACER. Again, we take these measures over various network sizes after high
cooperation has been achieved so the results are directly comparable to the previous
figures. When we plot the average path length (L) and clustering coefficient (C) we see a
number of interesting results. Firstly, the ASN follow a “small-world like” topology,
since C is relatively high (compared to a random network) yet L is low – meaning most
nodes are connected by only a few hops. We note that L scales-up log-linearly, indicating
that even in very large networks most nodes are connected by short paths. The degree
distributions (not shown) of the networks appear linear with about 10% of nodes having
the maximum number of links (in our case 20, the view size) and almost no nodes
holding zero links. This means that many nodes have many links. This suggests the ASN
are more robus t than some human soc ia l ne tworks wh ich d isp lay power- law
distributions, giving a scale-free topology, where only a few hub nodes have many links.
7.5 Typical Evolution
Figure 7 shows a time series of a typical run for a 2,000 node network executing
SLACER (W=0.9). We can identify a number of distinct stages in the time evolution of
the network before high cooperation is reached. First, the clustering coefficient (C)
increases rapidly and the path length (L) increases prior to cooperation forming. This
results from a randomized rewiring process, since all utilities are identical. This is
sufficient to create the C and L values we find throughout the run. Just before cycle 20,
via random mutation, two linked nodes become cooperative. This “seed tribe”, rather
than growing locally, explodes - creating a kind of sparse “cooperative backbone” over
the entire network. This spreads cooperation quickly, leading to rapid saturation of
cooperation. These four stages; random rewire, seed formation event, seed explosion then
saturation – we observe in all runs we have examined.
We have discussed in detail, in previous works (Hales 2000, 2004, Hales et al 2005), the
basic nature of this process and how, it operates: though a process of “group like”
selection between clusters or “tribes” . This could be termed “tribal selection” .
Essentially, if tribes are organized such that they provide high utility to their members
then they will tend to recruit more members. Tribes that become “infiltrated” with
defectors will tend to die out – since nodes will move to the tribes offering better utility.
Ironically, by defecting and acting selfishly a node sows the seeds of its own tribe’s
destruction since it’s initial high utility leads to it becoming surrounded by copy-cat
defectors reducing it’s payoffs.
7.6 Robustness to Churn
We have also subjected the ASN produced by SLACER to robustness tests by
introducing various amounts of “churn” - where old nodes leave and new nodes join the
network. In these experiments, we reset randomly selected nodes to defect states over
various intervals of cycles. We found that even when 50% of nodes were replaced at one
cycle then high-cooperation, CCP and the topology structures previously observed
quickly reformed within a few cycles. This was expected, since SLACER incorporates
noise in the form of mutation to both links and strategies, driving its evolutionary
dynamics and, as we have shown, can quickly recover from states of complete defection
and link disconnection.
7. 7 Different Rewire Values
We have also experimented with different values of W. We found that with values of W
higher than 0.9 the CCP started to fall off and we were back to the extreme tribalism of
SLAC. However, when we reduced W below 0.9 we found that the amount of
cooperation in the nodes began to fall, as did the clustering coefficient (C), so that when
W = 0.7 cooperation levels above 90% of nodes were never achieved (with mean values
oscillating around 80%) and C at around 2.5. For W = 0.5, cooperation levels never got
above 70%, oscillating widely around 60% with C below 2. This indicates a trade-off; we
need high W to get high cooperation – because this creates the clusters or “tribes” that
drive the process – however, if W is too high we get “extreme tribalism” – i.e. a
disconnected network. Varying W, therefore, controls the strength of tribalism or
“cliqueyness”.
8. Conclusion
We believe that SLACER is a step towards potentially v e r y useful ASN. We are
confident that we can apply it to more realistic and useful application tasks since we have
had earlier success applying the SLAC algorithm to a P2P file-sharing application and we
should be able to follow the same approach.
We have a specific interest in producing ASN for collective P2P virus, spyware and spam
filtering. Another possible target application is supporting trust and cooperative
interactions between nodes in a P2P search engine. In all these cases, although
sophisticated P2P algorithms already exist, they tend to assume all peers will act in a
cooperative way, which is naive and unsafe in open systems, or they require pre-existing
trusted social networks to be supplied as input – which are not always available or
appropriate.
SLACER can plug this gap with a self-organizing, robust, dynamic and modular
algorithm producing ASN, tuned, in real time, directly for specific applications. This
requires that already existing P2P algorithms be adapted such that they periodically
calculate a utility value, indicating the current service and trust quality derived from the
network and provide some adaptation method allowing for different levels of service and
trust to be offered to the network.
We have assumed that nodes will implement SLACER correctly, sending correct utility
values, links and strategies. It could be argued that by assuming this we are merely
begging the question, falling into the naive trap of assuming the trust and cooperation that
we claim to create! However, we always have in mind, that the determination of the
properties of other nodes could follow an indirect method to reduce malicious noise – by,
for example, asking some set of neighbors of a node to verify the information sent or
using other schemes. Additionally, utilities can be reduced to a binary satisfaction
function that obviates the need for utility comparisons or strategy copying completely.
Then nodes move in the network and change their strategies if they are not satisfied
(Singh et al 2005). The fact that SLAC works when nodes are able to copy those with
higher utility indicates that it is robust when nodes are able to scan the population and
select those behaviors that are producing high individual returns – irrespective of their
collective effect. In any implemented open system, such behavior is possible if nodes
collusively exchange information.
Where applications are complex, involving many kinds of trust and interaction tasks
between peers, we envisage the creation of multiple instantiations of SLACER each
supporting different tasks running concurrently. In this way, each ASN would be tuned to
the particular requirements of the specific task associated with it - rather like the way
humans form many networks around the different goals and tasks they need to achieve in
everyday life.
Could it be that by teaching our computers to make friends with each other we could
finally wave goodbye to the much of the anti-social behavior that plague Internet? On the
other hand, can even humans know who their true friends really are?
Acknowledgements
We thank Ozalp Babaoglu, Mark Jelasity, Alberto Montresor and Simon Patarin from the
Un ivers i ty of Bo logna , Dep t of Compu ter Sc ience , for percep t ive d iscuss ions ,
observations and pointers concerning P2P systems. We also thank Andrea Marcozzi and
Gian Paolo Jesi, who produced an initial implementation of SLAC using NEWSCAST in
the PEERSIM environment.
References
D. Hales, “Cooperation without Space or Memory: Tags, Groups and the Prisoner's
Dilemma,” in Proc. 2nd Int. Workshop on Multi-Agent-Based Simulation - LNAI 1979,
Berlin: Springer, 2000.
D. Hales, “From Selfish Nodes to Cooperative Networks – Emergent Link-based
Incentives in Peer-to-Peer Networks.” In proceedings of The Fourth IEEE International
Conference on Peer-to-Peer Computing (P2P2004), IEEE Computer Society Press, 2004.
D. Hales, B. Edmonds, “Applying a socially-inspired technique (tags) to improve
cooperation in P2P Networks,” IEEE Transactions in Systems, Man and Cybernetics -
Part A: Systems and Humans, 35(3):385-395, 2005.
J. S. Kong, P. O. Boykin, B. Rezei, N. Sarshar, and V. Roychowdhury, “Let you
cyberalter ego share information and manage spam,” 2005. Available as pre-print:
http://xxx.lanl.gov/abs/physics/0504026.
M. Jelasity, A. Montresor, and O. Babaoglu, “A modular paradigm for building self-
organizing peer-to-peer applications,” in Engineering Self-Organising Systems, LNAI
2977, Springer, 2004.
S. Marti, P. Ganesan, and H. Garcia-Molina, "DHT Routing Using Social Links,” in
Peer-to-Peer Systems III: Third International Workshop, IPTPS 2004, Revised Selected
Papers, LNCS 3279, Springer, 2005.
R. L. Riolo, M. D. Cohen and R. Axelrod, “Evolution of cooperation without
reciprocity”, Nature 414, 2001, pp. 441-443.
A. Singh and M. Haahr, “Creating an Adaptive Network of Hubs Using Schelling’s
Model”, presented at the IFIP/IEEE International Workshop on Self-Managed Systems &
S e r v i c e s
a t :
A v a i l a b l e
2 0 0 5 .
M a y
F r a n c e ,
N i c e ,
http://madyne.loria.fr/selfman2005/program.html
Q. Sun and H. Garcia-Molina, “SLIC: A Selfish Link-based Incentive Mechanism for
Unstructured Peer-to-Peer Networks,” in Proc. of 24th IEEE Int. Conf. on Distributed
Systems, IEEE Computer Society, 2004.
J. Travers and Milgram, “An Experimental Study of the Small World Problem.
Sociometry”, Vol. 32, No. 4, pp. 425-443, 1969.
|
1308.3324 | 1 | 1308 | 2013-08-15T07:33:28 | History Based Coalition Formation in Hedonic Context Using Trust | [
"cs.MA",
"cs.AI"
] | In this paper we address the problem of coalition formation in hedonic context. Our modelling tries to be as realistic as possible. In previous models, once an agent joins a coalition it would not be able to leave the coalition and join the new one; in this research we made it possible to leave a coalition but put some restrictions to control the behavior of agents. Leaving or staying of an agent in a coalition will affect on the trust of the other agents included in this coalition. Agents will use the trust values in computing the expected utility of coalitions. Three different risk behaviors are introduced for agents that want to initiate a coalition. Using these risk behaviors, some simulations are made and results are analyzed. | cs.MA | cs | International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 4, No. 4, July 2013
HISTORY BASED COALITION FORMATION IN
HEDONIC CONTEXT USING TRUST
Ahmadreza Ghaffarizadeh and Vicki H. Allan
Department of Computer Science, Utah State University, Logan, UT 84341, USA
[email protected], [email protected]
ABSTRACT
In this paper we address the problem of coalition formation in hedonic context. Our modelling tries to be as
realistic as possible. In previous models, once an agent joins a coalition it would not be able to leave the
coalition and join the new one; in this research we made it possible to leave a coalition but put some
restrictions to control the behavior of agents. Leaving or staying of an agent in a coalition will affect on the
trust of the other agents included in this coalition. Agents will use the trust values in computing the
expected utility of coalitions. Three different risk behaviors are introduced for agents that want to initiate a
coalition. Using these risk behaviors, some simulations are made and results are analyzed.
KEYWORDS
Coalition Formation, Trust, Risk Behaviors, Hedonic Context
1. INTRODUCTION
A multiagent system (MAS) is a system composed of various interacting agents. Based on the
type of agents, system can be a simple group or a complicated system doing sophisticated
operations. Agents in a MAS can have different types of objectives they are trying to reach. There
is not any external decision-maker or centralized processing unit in a MAS; agents work
autonomously and communicate with each other to do their tasks [1]. Due to lack of resources or
expertise, agents may not be able to reach their goals by themselves, so they are willing to form
coalition in order to reach their goals jointly. In most applications of multiagents, agents are
considered as cooperative, this means that they form coalition to reach common goals. This
research explores the context of hedonic games; in this context agents are self-interested and their
coalition satisfaction depends on the composition of the coalition they join. Agents only care
about their own profit and try to maximize it; so it is not reasonable for these agents to follow
actions which maximize the global profit not their own profit. Therefore agents in hedonic games
define a preference order over the set of all coalitions they can join and use this preference order
to form a satisfactory coalition. Many important problems in multi-agent systems including
marriage problem, roommate selection, admission proposals in universities, etc. can be
categorized in the context of hedonic games.
The main usage of hedonic games is in the game theory models in Economics; different models
have been studied in this domain and some research has been conducted to employ the hedonic
games concept in Economics models [2,3,4]. These models concentrate on the analysis of agents’
behavior; these research studies do not consider a multiagent framework and usually analyze self-
interested agents theoretically. Hedonic games are also frequently studied in the context of
Artificial Intelligence; these researches basically deals with finding the core solution in Hedonic
DOI : 10.5121/ijaia.2013.4401 1
International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 4, No. 4, July 2013
games, analyzing stability of core and examining its optimality [5,6,7,8]. Some works have been
published about coalition formation under uncertainty [7,9,10,11], however these works are not in
the context of Hedonic games and can’t be adopted.
An interesting research study has been published by Genin and Aknine [12]; this research mainly
deals with autonomous agents that are self-interested, a protocol is defined for coalition formation
process and some strategies are suggested for proposal acceptance and proposal selection
mechanisms. The current research is an extension of this study and tries to consider more
realistic cases by introducing new behaviors, parameters and mechanisms.
2. PROBLEM DESCRIPTION
2.1. Preliminaries
Definition 1: Agent a is an autonomous entity who can sense the environment and make a
sequence of actions upon to these sensations in order to reach a predefined objective.
Definition 2: A multiagent system M is a group of n agents M={a1, a2, …, an}that are joined to do
a/some task(s). These agents may or may not share common resources.
Definition 3: Coalition c is a nonempty subset of M. Let C denotes the set of all coalitions, then Ci
}
{
= ⊂
∈ .
c M a
c
i
f defined over sets of agents; (
Definition 4: A Hedonic game G is a preference order
i
denotes all coalitions which agent i belongs to;
)
,
iM f ;
C
i
this preference order is transitive and asymmetric.
Definition 5: A solution for a Hedonic game is a partition from set of all agents M, {c1, c2,…, ck}
k
=∑
c M
j
=
1
j
∩ = ∅ .
c
j
where
∀ ≠
i
and
, i
j c
2.2. Protocol
We do not have any central decision maker and agents in this research decide for themselves via
negotiation with other agents. During time each agent may receive many proposals for forming
coalition; it also sends some proposals to other agents asking about their willingness to form
coalition. We use a simple two-phase protocol: an agreement phase and commitment phase. The
protocol works as follow:
1- An initiator selects one of its possible coalitions according to its preference operator (it
will be discussed in later sections); suppose that this coalition includes k agents (1
initiator and k-1other agents).
2- Initiator sends the proposal of the selected coalition to all k-1 agents.
3- Agents respond to initiator with yes or no about their interest in joining this coalition.
4- If all agents agreed with the proposed coalition, initiator sends a confirmation request to
all agents and wants them to commit to the coalition, otherwise it sends a cancellation
message.
If all agents commit to the coalition, initiator informs all members that coalition is formed
otherwise it sends a cancellation message.
2
International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 4, No. 4, July 2013
3. PROPOSAL ACCEPTANCE STRATEGY
3.1. Proposal Acceptance
Agents in this research are rational; since proposal acceptance does not have any cost, and there is
always a probability that the current coalitions get broken; agents accept all proposals they
receive except for the proposals in which they have a utility less than when they are alone. Each
proposal has a limited waiting time, i.e. agents should respond by that time; otherwise the
response will be considered as rejection. Three types of agents are considered in this research:
1- Early responders who responds as soon as possible.
2- Lazy responders who waits to the end of deadline.
3- Random responders who have a random behavior in responding to the proposals.
3.2. Proposal Confirmation
On confirming proposals, whenever an agent feels that a proposal for a coalition is better than
current coalition, it can leave current one and join the proposed coalition. Following are some
considerations in confirming proposals:
1- Each coalition has an obligatory staying time Tc, it means that if an agent enrolled in a
coalition, it needs to be in the coalition for at least Tc unit of time. If it wants to leave
coalition sooner, it should pay the penalty Pc. Earned money from penalty will be shared
between all agents of the coalition equally.
2- Each agent has a level of honesty Hi; since leaving a coalition is unpleasant from
viewpoint of other members, the expected utility of leaving current coalition and joining
the other one should be great enough to tempt agent i to join a new coalition; as an
example if utility of current coalition is 1000; an agent with honesty level of 0.1 will be
tempted by a coalition with utility more than 1100 and an agent with level of 0.4 will be
tempted by a coalition with utility more than 1400. There is another factor that
discourages agents to leave one coalition: Bad Reputation Coefficient (BRC). To decide
on leaving current coalition and joining other coalition, agents act in a greedy way: they
only consider the next timestep. Totally, each agent evaluates the following criterion for
the next step and if it holds, then it leaves the current coalition and joins the new one.
Expected utility of current coalition*(1+Hi) < Expected utility of proposed coalition*(1- BRC)
3- There would be an enrollment fee for each solicited agent, CostEnro ll ; the initiator will
receive a part of this fee as a reward for forming the coalition which benefits all.
Since confirmation doesn’t mean that the confirmed proposal by this agent will certainly form a
coalition, the agent may need to confirm some proposals before joining one of them; but there is a
restriction that an agent at each time step can only confirm a limited number of proposals.
3.3. Trust between Agents
Each agent ai has a record for all other agents that shows its trust in them to not leave the
coalition. As an example, trij shows the trust ai has in the agent aj to stay in the next step in the
current coalition; from mathematical point of view, trij is the probability that aj won’t leave the
group in the next time step. All trust values are initialized by 0.5. In each time step all agents
update their trust values for other included agents in the coalition; suppose that agent ai in time
step t wants to update its trust records for agent aj that is in the same coalition as ai is; two cases
may happen:
3
International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 4, No. 4, July 2013
1- If aj stayed in the current coalition in step t-1, then ai increases the value trij by the special
trust reward r.
2- If aj has left the current coalition in step t-1, then ai decreases the value trij by the special
betray punishment pu.
The value for pu should be bigger than r to reflect that other agents hate the betrayer agent that
breaks the coalition. In this research r is 0.01 and pu is considered as 0.05.
3.4. Expected Utility Evaluation
Agent i uses the following formula to calculate the expected utility of current coalition Ccur for the
next step as:
(
ExpectedUtility C
i
cur
)
=
(
Utility C
i
cur
)
× ∏
tr
ij
∈
a C
j
cur
To compute the expected value of a proposed coalition C, agent i uses the following formula:
(
)
=
(
)
(
ExpectedUtility C Utility C
i
stepCoeff
Penalty
+
Cost
leaving
enroll
)
−
×
stepCoeff is a coefficient that lets the agent to consider the likelihood of being more than one time
step in the proposed coalition; so the stepCoeff is less than one.
4. PROPOSAL SELECTION STRATEGY
We consider a cost CostCom for each proposal the initiator sends. The initiator should consider this
cost to avoid sending bad proposals, i.e. the proposals that has a little chance of acceptance by all
agents. Initiator needs a measure to calculate the likelihood of formation of a candidate coalition;
so it calculates degree of interest of solicited agents based on the similarity with previous
proposals.
Similar to work done by Genin and Aknine[12] and based on the Bayesian reinforcement learning
model proposed by Chalkiadakis and Boutiller[11], the normalized distance between two
coalitions is used in this research to estimate their similarity. Suppose that all agents store all sent
and received proposals. Therefore if agent i finds a coalition that is similar to previously formed
coalition, it can hope that this candidate coalition will be formed too. The normalized distance
between two coalitions c1 and c2 is defined as:
(
,
d c c
2
1
)
=
1
n
(
c
1
∪
c
2
) (
\
c
1
∩
c
2
)
For each agent aj in candidate coalition c, initiator calculates
( )
( )
( )
ref
acc
rcv
jd
c as the
jd
c and
jd
c ,
distance between c and all the proposals received from aj, all the accepted proposals sent to aj and
all the refused proposals sent to aj accordingly. Using these three values, initiator computes the
following degree of interest for aj:
δ
j
( )
c
=
min
(
d
rcv
j
( )
,
c d
acc
j
( )
c
)
−
d
ref
j
( )
c
less than zero shows that aj may be interested in c and when
( )
( )
j cδ
j cδ
shows that aj is not. Since a coalition will be formed if and only if all agents are interested, so for
analyzing the likelihood of formation of a coalition, the degree of most unwilling agent should be
is greater than zero it
4
International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 4, No. 4, July 2013
considered. If we map the willingness value to the range [0, 1], we can use this value as the
probability of formation of a coalition (Pformation).
At each step, agents can send a limited number of proposals. If agents were banned to leave a
formed coalition, initiator could decide on its preference order and evaluate proposals one by one;
but in this research, agents can leave coalitions. So it is possible that an agent, who is not
interested in a proposal in time step t, accepts it at time step t+k. Also when the number of agents
is too large, it is not possible to have a preference order due to cost of memory and computation.
Based on these facts, agents in this project do not have a preference order; instead they have a
preference operator based on their risk attitude:
1- Risk seeking: agents in this category weigh the utility of coalition more than the
possibility of formation. They risk proposing the coalitions with high utility that may
have low possibility of formation. They use following measure to choose between the
proposal they can send:
=
(
)
(
Preference C Utility C
α
)
×
(
C
)
P
formation
(
α
)
>
1
2- Risk-averse: these agents prefer to choose the proposal that has the higher chance to be
accepted. Following equation is the measure which agents in this category use to
discriminate different proposals:
(
)
C β β
<
=
×
)
(
)
(
(
)
1
Preference C Utility C P
formation
3- Risk-neutral: these agents consider expected value and chance of acceptance in the same
weight; using this measure they decide which proposal to send:
=
×
(
)
(
)
(
)
Preference C U tility C P
C
forma tion
At each time step, an initiator creates a set that contains a) a fixed number of random coalitions b)
some randomly selected proposals from the previously sent proposals and c) some randomly
selected proposals from previously received proposals. Using this set, initiator selects the best
proposals it can send using its preference operator. Agents do not send the selected proposal if its
expected utility is less than a threshold; this threshold is set to 2*CostCom in our work.
5. EXPERIMENTS AND RESULTS
We conducted 4 different simulations and repeated each simulation 20 times. There are 20 agents
in each simulation running for 100 time steps. Bad Reputation Coefficient is set to 0.15 and
Honesty level for each agent is taken from the range [0, 0.35]. Expiration time for accepting and
confirmation is set to 3 steps. To calculate the utility of coalitions, each agent considers the direct
interaction with other agents and pairwise interaction of each two other agents in the coalition;
each agent has a table of estimation of interaction for each two agents (including itself). It finds
all permutation of size 2 of agents and look up for the value in the estimation table, then sums all
these value to compute the utility of a given coalition. Each entry in the table is randomly chosen
from the range [-100, 100].
For the first simulation, all agents are risk seeking; for the second simulation agents are risk
averse and in the third simulation all the agents are risk neutral. We have considered the equal
number of each type of agents in the fourth simulation. Gained utility for agents with similar
honesty levels are averaged for simulations and are depicted in Figure 1. Note that honesty levels
are divided to 0.05 ranges and each agent is mapped to the closest level in averaging. Results
show that risk-seeking and risk-neutral agents have similar utility behavior over the changes in
honesty factor and have the highest utility with the honesty factor [0.15, 0.20].
Second set of results show the average distribution of initiator, solicited and alone agents per step
in each simulation. Note that axes are scaled to show the distribution clearly. Figure 2 shows the
5
International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 4, No. 4, July 2013
distribution of alone agents, it can be seen that in the simulation with risk-seeking agents the
number of alone agents are less than two other simulations. Figure 3 depicts distribution of
solicited agents that have similar distributions in all three simulations. Distribution of initiator
agents are depicted in Figure 4; since each coalition has only one initiator, Figure 4 shows the
number of formed coalitions as well. As results show, number of formed coalitions in simulation
with risk-neutral agents is less than two other simulations. The average duration of coalitions for
risk-seeking agents is 8 steps, for risk-averse agents is 12 and for risk neutral agents is 17 steps.
6. CONCLUSION
Humans are self-interested and always (with some exceptions and restrictions) follow their
preferences in their interaction with others. That is why context of “hedonic games,” that includes
self-interested agents, has been received a significant attention. In this paper we addressed the
problem of coalition formation in the context of hedonic games. We modelled a realistic case
where agents at each time step can decide to stay in a coalition or leave. As in reality, when an
agent leaves a group, other agents lose their trust in the left agent. Agents have an honesty factor;
the more this honesty factor, the more reliable our agents are. Agents use trust values to estimate
the possibility of continuing with the coalition they currently are in. We analyzed the gained
utility of agents regards to their risk behavior in proposing coalitions to
y
t
i
l
i
t
U
7000
6500
6000
5500
5000
4500
4000
3500
3000
All Risk Seeking
All Risk Averse
All Risk Neutral
Mixture
0
0.1
0.2
0.3
0.4
Honesty Factor
Figure 1: Utility versus Honesty factor regards to agent types in 4 different simulations.
0.6
0.5
0.4
0.3
0.2
0.1
s
t
n
e
g
a
e
n
o
l
a
f
o
n
o
i
t
u
b
i
r
t
s
i
D
Risk Seeking
Risk Averse
Risk Neutral
0
20
40
60
80
100
Step
Figure 2. Distribution of alone agents in 100 time steps for agents with different risk attitudes.
6
International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 4, No. 4, July 2013
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
s
t
n
e
g
a
d
e
t
i
c
i
l
o
s
f
o
n
o
i
t
u
b
i
r
t
s
i
D
Risk Seeking
Risk Averse
Risk Neutral
0
20
40
60
80
100
Step
Figure 3. Distribution of solicited agents in 100 time steps for agents with different risk attitudes.
0.25
0.2
0.15
0.1
0.05
0
s
t
n
e
g
a
r
o
t
a
i
t
i
n
i
f
o
n
o
i
t
u
b
i
r
t
s
i
D
Risk Seeking
Risk Averse
Risk Neutral
0
20
40
60
80
100
Step
Figure 4. Distribution of initiator agents in 100 time steps for agents with different risk attitudes.
other agents. Results show that different levels of honesty can lead to different utility values for
distinct risk attitudes.
7. FUTURE WORK
Some ideas for future work are as follows: one may think about considering the level of
friendship between agents; an agent may accept a proposal from a friend while it doesn’t accept a
similar proposal with a higher utility from a stranger. As in real world, some people like to be
manager, some agents may enjoy of being initiator; it would be interesting to consider a factor to
show the willingness of agents about being an initiator or being solicited. Implementing this
behavior can be done by introducing a value α in the range [
]0,1 showing the willingness of each
agent to be an initiator. An agent with α=0 never propose coalitions to others and an agent with
α=1 never accept any proposals from others; other values will affect the decision of being an
7
International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 4, No. 4, July 2013
initiator or not. As another future work, one may employ history to estimate duration of candidate
coalitions to use in the coalition utility evaluation.
REFERENCES
[1] K. Ahmadi and V. H. Allan, “Efficient Self Adapting Agent Organizations,” in Proceedings of the 5th
International Conference on Agents and Artificial Intelligence, ICAART 2013, Barcelona, Spain.
[2] A. Bogomolnaia, M. O. Jackson, “The stability of hedonic coalition structures,” Games and Economic
Behavior, vol. 38, no. 2, pp. 201-230, Feb. 2002.
[3] S. Banerjee, H. Konishi, and T. Sonmez, “Core in a simple coalition formation game,” Social Choice
and Welfare, vol. 18, no. 1, pp. 135-153, Jan. 2001.
[4] E. Diamantoudi and L. Xue, “Farsighted stability in hedonic games,” Social Choice and Welfare, vol.
21, no. 1, pp. 39-61, Aug. 2003.
[5] S. Sung and D. Dimitrov, “On core membership testing for hedonic coalition formation games,”
Operations Research Letters, vol. 35, no. 2, pp. 155-158, Mar. 2007.
[6] T. Genin and S. Aknine, “Constraining Self-Interested agents to guarantee pareto optimality in
multiagent coalition formation problem,” vol. 2, pp. 369-372, Aug. 2011.
[7] F. Bloch and E. Diamantoudi, “Noncooperative formation of coalitions in hedonic games,”
International Journal of Game Theory, vol. 40, no. 2, pp. 263-280, May 2011.
[8] H. Aziz, F. Brandt and H. G. Seedig, “Stable partitions in additively separable hedonic games,” Proc.
of AAMAS’11, pp. 183-190, 2011
[9] S. Kraus, O. Shehory, and G. Taase, “Coalition formation with uncertain heterogeneous information,”
in Proceedings of the second international joint conference on Autonomous agents and multiagent
systems, ser. AAMAS '03. New York, NY, USA: ACM, 2003, pp. 1-8.
[10] K. Westwood and V. H. Allan, “Heuristics for dealing with a shrinking pie in agent coalition
formation,” pp. 537-546, Dec. 2006.
[11] G. Chalkiadakis and C. Boutilier, “Bayesian reinforcement learning for coalition formation under
uncertainty,” in Proceedings of the Third International Joint Conference on Autonomous Agents and
Multiagent Systems - Volume 3, ser. AAMAS Washington, DC, USA: IEEE Computer Society, 2004,
pp. 1090–1097
[12] T. Genin and S. Aknine, “Coalition formation strategies for multiagent hedonic games,” Tools with
Artificial Intelligence, IEEE International Conference on, vol. 1, pp. 465-472, 2010.
Authors
Ahmadreza Ghaffarizadeh is a research assistant and PhD candidate in Computer Science
department at Utah State University. His research interests include computational biology,
artificial intelligence and evolutionary algorithms.
Vicki H Allan completed her PhD in Computer Science at Colorado State University in
1986. She completed a master’s degree in computer science and a master’s degree in
mathematics from Utah State University. She is currently an associate professor in the
Computer Science department at Utah State University where she teaches courses in
multiagent systems, programming languages, data structures, and algorithms. Her research
is supported by NSF grants 0812039 and 0829563.
8
|
1607.01838 | 2 | 1607 | 2017-10-11T01:38:44 | Coordinate-Descent Diffusion Learning by Networked Agents | [
"cs.MA",
"cs.DC",
"eess.SY"
] | This work examines the mean-square error performance of diffusion stochastic algorithms under a generalized coordinate-descent scheme. In this setting, the adaptation step by each agent is limited to a random subset of the coordinates of its stochastic gradient vector. The selection of coordinates varies randomly from iteration to iteration and from agent to agent across the network. Such schemes are useful in reducing computational complexity at each iteration in power-intensive large data applications. They are also useful in modeling situations where some partial gradient information may be missing at random. Interestingly, the results show that the steady-state performance of the learning strategy is not always degraded, while the convergence rate suffers some degradation. The results provide yet another indication of the resilience and robustness of adaptive distributed strategies. | cs.MA | cs | Coordinate-Descent Diffusion Learning by
Networked Agents
Chengcheng Wang, Student Member, IEEE, Yonggang Zhang, Senior Member, IEEE,
Bicheng Ying, Student Member, IEEE, and Ali H. Sayed, Fellow, IEEE
1
7
1
0
2
t
c
O
1
1
]
A
M
.
s
c
[
2
v
8
3
8
1
0
.
7
0
6
1
:
v
i
X
r
a
Abstract-This work examines the mean-square error perfor-
mance of diffusion stochastic algorithms under a generalized
coordinate-descent scheme. In this setting, the adaptation step
by each agent is limited to a random subset of the coordinates
of its stochastic gradient vector. The selection of coordinates
varies randomly from iteration to iteration and from agent to
agent across the network. Such schemes are useful in reducing
computational complexity at each iteration in power-intensive
large data applications. They are also useful in modeling situa-
tions where some partial gradient information may be missing
at random. Interestingly, the results show that the steady-state
performance of the learning strategy is not always degraded,
while the convergence rate suffers some degradation. The results
provide yet another indication of the resilience and robustness
of adaptive distributed strategies.
Index Terms-Coordinate descent, stochastic partial update,
computational complexity, diffusion strategies, stochastic gradi-
ent algorithms, strongly-convex cost.
I. INTRODUCTION AND RELATED WORK
Consider a strongly-connected network of N agents, where
information can flow in either direction between any two
connected agents and, moreover, there is at least one self-loop
in the topology [2, p. 436]. We associate a strongly convex
differentiable risk, Jk(w), with each agent k and assume in
this work that all these costs share a common minimizer,
wo ∈ RM , where R denotes field of real numbers. This case
models important situations where agents work cooperatively
towards the same goal. The objective of the network is to
determine the unique minimizer wo of the following aggregate
cost, assumed to be strongly-convex:
Jk(w)
(1)
k=1
A short version of this work appears in the conference publication [1].
This work was performed while C. Wang was a visiting student at the
UCLA Adaptive Systems Laboratory. The work of C. Wang was supported
in part by a Chinese Government Scholarship. The work of Y. Zhang was
supported in part by the National Natural Science Foundation of China
(61371173), and Fundamental Research Founds for the Central University
of Harbin Engineering University (HEUCFP201705). The work of B. Ying
and A. H. Sayed was supported in part by NSF grants CCF-1524250 and
ECCS-1407712.
C. Wang and Y. Zhang are with the College of Automation, Harbin
Engineering University, Harbin, Heilongjiang 150001 China. C. Wang is
also with the School of Electrical and Electronic Engineering, Nanyang
Technological University, 639798 Singapore (e-mail: [email protected];
[email protected]).
B. Ying is with the Department of Electrical Engineering, University of
California, Los Angeles, CA 90024 USA (e-mail: [email protected]).
A. H. Sayed is with the Ecole Polytechnique Federale de Lausanne, EPFL,
School of Engineering, CH-1015 Lausanne, Switzerland (e-mail: ali.sayed@
epfl.ch).
J glob(w) (cid:44) N(cid:88)
It is also assumed that the individual cost functions, Jk(w),
are each twice-differentiable and satisfy
0 < νdIM ≤ ∇2
wJk(w) ≤ δdIM
(2)
where ∇2
wJk(w) denotes the M ×M Hessian matrix of Jk(w)
with respect to w, νd ≤ δd are positive parameters, and IM
is the M × M identity matrix. In addition, for matrices A
and B, the notation A ≤ B denotes that B − A is positive
semi-definite. The condition in (2) is automatically satisfied
by important cases of interest, such as logistic regression or
mean-square-error designs [2], [3].
Starting from some initial conditions {wk,−1}, the agents
work cooperatively in an adaptive manner to seek the mini-
mizer wo of problem (1) by applying the following diffusion
strategy [2], [3]:
(cid:88)
(cid:88)
(cid:96)∈Nk
(cid:96)∈Nk
φk,i−1 =
a1,(cid:96)kw(cid:96),i−1
ψk,i = φk,i−1 − µk
a2,(cid:96)kψ(cid:96),i
wk,i =
(cid:92)∇wTJ k(φk,i−1)
(3a)
(3b)
(3c)
where the M-vector wk,i denotes the estimate by agent k
at iteration i for wo, while ψk,i and φk,i−1 are intermediate
estimates. Moreover, an approximation for the true gradient
vector of Jk(w), (cid:92)∇wTJ k(·), is used in (3b) since it is generally
the case that the true gradient vector is not available (e.g.,
when Jk(w) is defined as the expectation of some loss function
and the probability distribution of the data is not known
beforehand to enable computation of Jk(·) or its gradient
vector). The symbol Nk in (3) refers to the neighborhood of
agent k. The N × N combination matrices A1 = [a1,(cid:96)k] and
A2 = [a2,(cid:96)k] are left-stochastic matrices consisting of convex
combination coefficients that satisfy:
aj,(cid:96)k ≥ 0,
aj,(cid:96)k = 1, aj,(cid:96)k = 0, if (cid:96) /∈ Nk
(4)
(cid:96)=1
for j = 1, 2. Either of these two matrices can be chosen as the
identity matrix, in which case algorithm (3) reduces to one of
two common forms for diffusion adaptation: the adapt-then-
combine (ATC) form when A1 = I and the combine-then-
adapt (CTA) form when A2 = I. We continue to work with the
general formulation (3) in order to treat both algorithms, and
other cases as well, in a unified manner. The parameter µk > 0
is a constant step-size factor used to drive the adaptation
in order to enable
process. Its value is set
to a constant
N(cid:88)
continuous adaptation in response to streaming data or drifting
minimizers. We could also consider a distributed implementa-
tion of the useful consensus-type [2], [4]–[11]. However, it has
been shown in [2], [12] that when constant step-sizes are used
to drive adaptation, the diffusion networks have wider stability
ranges and superior performance. This is because consensus
implementations have an inherent asymmetry in their updates,
which can cause network graphs to behave in an unstable
manner when the step-size is constant. This problem does
not occur over diffusion networks. Since adaptation is a core
element of the proposed strategies in this work, we therefore
focus on diffusion learning mechanisms.
The main distinction in this work relative to prior studies
on diffusion or consensus adaptive networks is that we now
assume that, at each iteration i, the adaptation step in (3b)
has only access to a random subset of the entries of the
approximate gradient vector. This situation may arise due to
missing data or a purposeful desire to reduce the computational
burden of the update step. We model this scenario by replacing
the approximate gradient vector by
(cid:92)∇wT J
miss
k
(φk,i−1) = Γk,i · (cid:92)∇wTJ k(φk,i−1)
(5)
where the random matrix Γk,i is diagonal and consists of
Bernoulli random variables {rk,i(m)}; each of these variables
is either zero or one with probability
Prob(rk,i(m) = 0) (cid:44) rk
where 0 ≤ rk < 1 and
Γk,i = diag{rk,i(1), rk,i(2), . . . , rk,i(M )}
(6)
(7)
In the case when rk,i(m) = 0, the m-th entry of the gradient
vector is missing, and then the m-th entry of ψk,i in (3b) is
not updated. Observe that we are attaching two subscripts to
r: k and i, which means that we are allowing the randomness
in the update to vary across agents and also over time.
A. Relation to Block-Coordinate Descent Methods
Our formulation provides a nontrivial generalization of the
powerful random coordinate-descent
technique studied, for
example, in the context of deterministic optimization in [13]–
[15] and the references therein. Random coordinate-descent
has been primarily applied in the literature to single-agent
convex optimization, namely, to problems of the form:
wo = arg min
w
J(w)
(8)
where J(w) is assumed to be known beforehand. The tradi-
tional gradient descent algorithm for seeking the minimizer of
J(w), assumed differentiable, takes the form
wi = wi−1 − µ∇wTJ(wi−1)
(9)
where the full gradient vector is used at every iteration to
update wi−1 to wi. In a coordinate-descent implementation,
on the other hand, at every iteration i, only a subset of the
entries of the gradient vector is used to perform the update.
These subsets are usually chosen as follows. First, a collection
of K partitions of the parameter space w is defined. These
2
partitions are defined by diagonal matrices, {Ωk}. Each matrix
has ones and zeros on the diagonal and the matrices add up
to the identity matrix:
K(cid:88)
Ωk = IM
(10)
k=1
Multiplying w by any Ωk results in a vector of similar size,
albeit one where the only nontrivial entries are those extracted
by the unit locations in Ωk. At every iteration i, one of the
partitions is selected randomly, say, with probability
Prob(Γi = Ωk) = ωk
(11)
where the {ωk} add up to one. Subsequently, the gradient
descent iteration is replaced by
wi = wi−1 − µΓi∇wTJ(wi−1)
(12)
This formulation is known as the randomized block-coordinate
descent (RBCD) algorithm [13]–[15]. At each iteration, the
gradient descent step employs only a collection of coordinates
represented by the selected entries from the gradient vector.
Besides reducing complexity,
this step helps alleviate the
condition on the step-size parameter for convergence.
If we reduce our formulation (3)–(5) to the single agent
case, it will become similar to (12) in that the desired cost
function is optimized only along a subset of the coordinates at
each iteration. However, our algorithm offers more randomness
in generating the coordinate blocks than the RBCD algorithm,
by allowing more random combinations of the coordinates at
each time index. In particular, we do not limit the selection
of the coordinates to a collection of K possibilities pre-
determined by the {Ωk}. Moreover, in our work we use a
random subset of the stochastic gradient vector instead of
the true gradient vector to update the estimate, which is
necessary for adaptation and online learning when the true risk
function itself is not known (since the statistical distribution
of the data is not known). Also, our results consider a general
multi-agent scenario involving distributed optimization where
each individual agent employs random coordinates for its own
gradient direction, and these coordinates are generally different
from the coordinates used by other agents. In other words,
the networked scenario adds significant flexibility into the
operation of the agents under model (5).
B. Relation to Partial Updating Schemes
It is also important to clarify the differences between our
formulation and other works in the literature, which rely on
other useful notions of partial information updates. To begin
with, our formulation (5) is different from the models used
in [16]–[18] where the step-size parameter was modeled as
a random Bernoulli variable, µk(i), which could assume the
values µk or zero with certain probability. In that case, when
the step-size is zero, all entries of ψk,i will not be updated
and adaptation is turned off completely. This is in contrast
to the current scenario where only a subset of the entries are
left without update and, moreover, this subset varies randomly
from one iteration to another.
Likewise, the useful work [19] employs a different notion
of partial sharing of information by focusing on the exchange
of partial entries of the weight estimates themselves rather
than on partial entries of the gradient directions. In other
words, the partial information used in this work relates to the
combination steps (3a) and (3c) rather than to the adaptation
step (3b). It also focuses on the special case in which the risks
{Jk(w)} are quadratic in w. In [19], it is assumed that only
a subset of the weight entries are shared (diffused) among
neighbors and that the estimate itself is still updated fully. In
comparison, the formulation we are considering diffuses all
entries of the weight estimates. Similarly, in [20] it is assumed
that some entries of the regression vector are missing, which
causes changes to the gradient vectors. In order to undo these
changes, an estimation scheme is proposed in [20] to estimate
the missing data. In our formulation, more generally, a random
subset of the entries of the gradient vector are set to zero at
each iteration, while the remaining entries remain unchanged
and do not need to be estimated.
There are also other criteria that have been used in the
literature to motivate partial updating. For example, in [21], the
periodic and sequential least-mean-squares (LMS) algorithms
are proposed, where the former scheme updates the whole
coefficient vector every N−th iteration, with N > 1, and the
latter updates only a fraction of the coefficients, which are pre-
determined, at each iteration. In [22], [23] the weight vectors
are partially updated by following a set-membership approach,
where updates occur only when the innovation obtained from
the data exceeds a predetermined threshold. In [23]–[25], only
entries corresponding to the largest magnitudes in the regres-
sion vector or the gradient vector at each agent are updated.
However, such scheduled updating techniques may suffer from
non-convergence in the presence of nonstationary signals [26].
Partial update schemes can also be based on dimensionality
reduction policies using Krylov subspace concepts [27]–[29].
There are also techniques that rely on energy considerations
to limit updates, e.g., [30].
The objective of the analysis that follows is to examine the
effect of random partial gradient information on the learning
performance and convergence rate of adaptive networks for
general risk functions. We clarify these questions by adapting
the framework described in [2], [3].
Notation: We use lowercase letters to denote vectors, up-
percase letters for matrices, plain letters for deterministic
variables, and boldface letters for random variables. We also
use (·)T to denote transposition, (·)−1 for matrix inversion,
Tr(·) for the trace of a matrix, diag{·} for a diagonal matrix,
col{·} for a column vector, λ(·) for the eigenvalues of a matrix,
ρ(·) for the spectral radius of a matrix, (cid:107)·(cid:107) for the two-induced
norm of a matrix or the Euclidean norm of a vector, (cid:107)x(cid:107)2
Σ for
the weighted square value xTΣx, ⊗ for Kronecker product,
⊗b for block Kronecker product. Besides, we use p (cid:31) 0
to denote that all entries of vector p are positive. Moreover,
α = O(µ) signifies that α ≤ cµ for some constant c > 0,
and α = o(µ) signifies that α/µ → 0 as µ → 0. In addition,
the notation lim supn→∞ a(n) denotes limit superior of the
sequence a(n).
3
II. DATA MODEL AND ASSUMPTIONS
Let F i−1 represent the filtration (collection) of all random
events generated by the processes {wk,j} and {Γk,j} at all
agents up to time i − 1. In effect, the notation F i−1 refers to
the collection of all past {wk,j, Γk,j} for all j ≤ i− 1 and all
agents.
Assumption 1: (Conditions on indicator variables). It is
assumed that the indicator variables rk,i(m) and r(cid:96),i(n) are
independent of each other, for all (cid:96), k, m, n. In addition, the
variables {rk,i(m)} are independent of F i−1 and (cid:92)∇wT J k(w)
for any iterates w ∈ F i−1 and for all agents k.
(cid:3)
Let
sk,i(φk,i−1) (cid:44) (cid:92)∇wTJ k(φk,i−1) − ∇wTJ k(φk,i−1)
(13)
denote the gradient noise at agent k at iteration i, based on
the complete approximate gradient vector, (cid:92)∇wT J k(w). We
introduce its conditional second-order moment
Rs,k,i(w) (cid:44) E[sk,i(w)sT
k,i(w)F i−1].
(14)
The following assumptions are standard and are satisfied by
important cases of interest, such as logistic regression risks or
mean-square-error risks, as already shown in [2], [3]. These
references also motivate these conditions and explain why they
are reasonable.
Assumption 2: (Conditions on gradient noise) [2, pp. 496–
497]. It is assumed that the first and fourth-order conditional
moments of the individual gradient noise processes satisfy the
following conditions for any iterates w ∈ F i−1 and for all
k, (cid:96) = 1, 2, . . . , N:
E[sk,i(w)F i−1] = 0
E[sk,i(w)sT
(cid:96),i(w)F i−1] = 0, k (cid:54)= (cid:96)
E[(cid:107)sk,i(w)(cid:107)4F i−1] ≤ β4
almost surely, for some nonnegative scalars β4
(15)
(16)
(17)
s,k. (cid:3)
Assumption 3: (Smoothness conditions) [2, pp. 552,576].
It is assumed that the Hessian matrix of each individual cost
function, Jk(w), and the covariance matrix of each individual
gradient noise process are locally Lipschitz continuous in a
small neighborhood around w = wo in the following manner:
k(cid:107)w(cid:107)4 + σ4
s,k
k and σ4
wJk(wo + (cid:52)w) − ∇2
(cid:107)∇2
wJk(wo)(cid:107) ≤ κc(cid:107)(cid:52)w(cid:107)
(cid:107)Rs,k,i(wo + (cid:52)w) − Rs,k,i(wo)(cid:107) ≤ κd(cid:107)(cid:52)w(cid:107)γ
(18)
(19)
for any small perturbations (cid:107)(cid:52)w(cid:107) ≤ ε and for some κc ≥ 0,
κd ≥ 0, and parameter 0 < γ ≤ 4.
(cid:3)
III. MAIN RESULTS: STABILITY AND PERFORMANCE
For each agent k, we introduce the error vectors:
(cid:101)wk,i (cid:44) wo − wk,i
(cid:101)φk,i
(cid:101)ψk,i
(cid:44) wo − φk,i
(cid:44) wo − ψk,i
(20)
(21)
(22)
We also collect all errors, along with the gradient noise
processes, from across the network into block vectors:
(cid:101)wi (cid:44) col{(cid:101)w1,i,(cid:101)w2,i, . . . ,(cid:101)wN,i}
(cid:111)
(cid:110)(cid:101)ψ1,i,(cid:101)ψ2,i, . . . ,(cid:101)ψN,i
(cid:101)ψi
(cid:111)
(cid:110)(cid:101)φ1,i,(cid:101)φ2,i, . . . ,(cid:101)φN,i
(cid:101)φi
(cid:44) col
(cid:44) col
si (cid:44) col{s1,i, s2,i, . . . , sN,i}
(23)
(24)
(25)
(26)
For simplicity, in (26) we use the notation sk,i to replace the
gradient noise sk,i(φk,i−1) defined in (13), but note vector
si is dependent on the collection of {φk,i−1} for all k. We
further introduce the extended matrices:
M (cid:44) diag{µ1, µ2, . . . , µN} ⊗ IM
A1 (cid:44) A1 ⊗ IM ,A2 (cid:44) A2 ⊗ IM
Γi (cid:44) diag{Γ1,i, Γ2,i, . . . , ΓN,i}
(27)
(28)
(29)
Note that
the main difference between the current work
and the prior work in [2] is the appearance of the random
matrices {Γk,i} defined by (5). In the special case when
the random matrices are set to the identity matrices across
i.e., {Γk,i ≡ IM}, current coordinate-descent
the agents,
case will reduce to the full-gradient update studied in [2].
The inclusion of the random matrices {Γk,i} adds a non-
trivial level of complication because now, agents update only
random entries of their iterates at each iteration and, impor-
tantly, these entries vary randomly across the agents. This
procedure adds a rich level of randomness into the operation
of the multi-agent system. As the presentation will reveal,
the study of the stability and limiting performance under
these conditions is more challenging than in the stochastic
full-gradient diffusion implementation due to at
two
factors: (a) First, the evolution of the error dynamics will
now involve a non-symmetric matrix (matrix D11,i defined
later in (111)); because of this asymmetry, the arguments of
[2] do not apply and need to be modified; and (b) second,
there is also randomness in the coefficient matrix for the error
dynamics (namely, randomness in the matrix B(cid:48)
i defined by
(39)). These two factors add nontrivial complications to the
stability, convergence, and performance analysis of distributed
coordinate-descent solutions, as illustrated by the extended
derivations in Appendices A and B. These derivations illustrate
the new arguments that are necessary to handle the networked
solution of this manuscript. For this reason, in the presentation
that follows, whenever we can appeal to a result from [2], we
will simply refer to it so that, due to space limitations, we
can focus the presentation on the new arguments and proofs
that are necessary for the current context. It is clear from the
proofs in Appendices A and B that these newer arguments are
demanding and not straightforward.
least
Lemma 1: (Network error dynamics). Consider a network
of N interacting agents running the diffusion strategy (3) with
the gradient vector replaced by (5). The evolution of the error
dynamics across the network relative to the reference vector
wo is described by the following recursion:
2MΓisi
(cid:101)wi = Bi(cid:101)wi−1 + AT
(30)
where
Bi (cid:44) AT
2 (I − MΓiHi−1)AT
(cid:90) 1
wJk(wo − t(cid:101)φk,i−1)dt.
Hi−1 (cid:44) diag{H1,i−1, H2,i−1, . . . , HN,i−1}
Hk,i−1 (cid:44)
∇2
1
0
4
(31)
(32)
(33)
Proof : Refer to [2, pp. 498–504], which is still applicable to
the current context. We only need to set in that derivation
the matrix Ao to Ao = I, and the vector b to b = 0M N .
These quantities were defined in (8.131) and (8.136) of [2].
The same derivation will lead to (30)–(33), with the main
difference being the appearance now of the random matrix
(cid:3)
Γi in (30) and (31).
We assume that the matrix product P = A1A2 is primi-
tive. This condition is guaranteed automatically, for example,
for ATC and CTA scenarios when the network is strongly-
connected. This means, in view of the Perron-Frobenius The-
orem [2], [3], that P has a single eigenvalue at one. We denote
the corresponding eigenvector by p, and normalize the entries
of p to add up to one. It follows from the same theorem that
the entries of p are strictly positive, written as
P p = p, 1Tp = 1, p (cid:31) 0
(34)
with 1 being the vector of size N with all its entries equal to
one.
Theorem 1:
(Network stability). Consider a strongly-
connected network of N interacting agents running the dif-
fusion strategy (3) with the gradient vector replaced by (5).
Assume the matrix product P = A1A2 is primitive. Assume
also that
the individual cost functions, Jk(w), satisfy the
condition in (2) and that Assumptions 1–2 hold. Then, the
second and fourth-order moments of the network error vectors
are stable for sufficiently small step-sizes, namely, it holds, for
all k = 1, 2, . . . , N, that
E(cid:107)(cid:101)wk,i(cid:107)2 = O(µmax)
E(cid:107)(cid:101)wk,i(cid:107)4 = O(µ2
max)
lim sup
i→∞
lim sup
i→∞
(35)
(36)
(37)
for any µmax < µo, for some small enough µo, where
µmax (cid:44) max{µ1, µ2, . . . , µN}.
Proof : The argument requires some effort and is given in
(cid:3)
Appendix A.
Lemma 2: (Long-term network dynamics). Consider a
strongly-connected network of N interacting agents running
the diffusion strategy (3) under (5). Assume the matrix product
P = A1A2 is primitive. Assume also that the individual
cost functions satisfy (2), and that Assumptions 1–2 and (18)
hold. After sufficient iterations, i (cid:29) 1, the error dynamics
of the network relative to the reference vector wo is well-
approximated by the following model:
(cid:101)w(cid:48)
i(cid:101)w(cid:48)
i = B(cid:48)
i−1 + AT
2MΓisi, i (cid:29) 1
(38)
where
i
(cid:44) AT
2 (I − MΓiH)AT
B(cid:48)
H (cid:44) diag{H1, H2, . . . , HN}
Hk (cid:44) ∇2
wJk(wo)
1
(39)
(40)
(41)
More specifically, it holds for sufficiently small step-sizes that
E(cid:107)(cid:101)w(cid:48)
E(cid:107)(cid:101)w(cid:48)
lim sup
i→∞
lim sup
i→∞
i(cid:107)2 = lim sup
i→∞
E(cid:107)(cid:101)w(cid:48)
k,i(cid:107)2 = O(µmax)
k,i(cid:107)4 = O(µ2
max)
E(cid:107)(cid:101)wi(cid:107)2 + O(µ3/2
max).
(42)
(43)
(44)
lim sup
i→∞
Proof : To establish (38), we refer to the derivation in [2, pp.
553–555], and note that, in our case, (cid:107)Γi(cid:107) ≤ 1 and b = 0M N
(which appeared in (10.2) of [2]). Moreover, the results in (42)
and (43) can be established by following similar techniques
to the proof of Theorem 1, where the only difference is that
the random matrix Hi−1 defined in (32) is now replaced with
the deterministic matrix H defined by (40), and by noting that
the matrices {Hk} in (41) still satisfy the condition (2). With
regards to result (44), we refer to the argument in [2, pp. 557–
560] and note again that (cid:107)Γi(cid:107) ≤ 1.
(cid:3)
Result (35) ensures that the mean-square-error (MSE) per-
formance of the network is in the order of µmax. Using the
long-term model (38), we can be more explicit and derive
the proportionality constant that describes the value of the
network mean-square-error to first-order in µmax. To do so,
we introduce the quantity
q (cid:44) diag{µ1, µ2, . . . , µN} A2p
and the gradient-noise covariance matrices:
i→∞ Rs,k,i(wo)
(cid:44) E[Γk,iGkΓk,i].
Gk (cid:44) lim
G(cid:48)
k
(45)
(46)
(47)
Observe that Gk is the limiting covariance matrix of the
gradient noise process evaluated at wo, and is assumed to
be a constant value, while G(cid:48)
k is a weighted version of it.
A typical example for the existence of the limit in (46) is the
MSE network, where the covariance matrix of the gradient
noise is a constant matrix, which is independent of the time
index i [2, p. 372]. It follows by direct inspection that the
entries of G(cid:48)
(cid:26) (1 − rk)2Gk(m, n), m (cid:54)= n
k are given by:
(1 − rk)Gk(m, m), m = n.
(48)
G(cid:48)
k(m, n) =
We also define the mean-square-deviation (MSD) for each
agent k, and the average MSD across the network to first-
order in µmax - see [2] for further clarifications on these
expressions where it is explained, for example, that MSDk
provides the steady-state value of the error variance E(cid:107)(cid:101)wk,i(cid:107)2
to first-order in µmax:
MSDk (cid:44) µmax
MSDav (cid:44) 1
N
(cid:18)
N(cid:88)
k=1
lim
µmax→0
MSDk.
(cid:19)
E(cid:107)(cid:101)wk,i(cid:107)2
lim sup
i→∞
1
µmax
5
(49)
(50)
Likewise, we define the excess-risk (ER) for each agent k as
the average fluctuation of the normalized aggregate cost
¯J glob(w) (cid:44)
qk
qkJk(w)
(51)
with {qk} being entries of the vector q defined by (45), around
its minimum value ¯J glob(wo) at steady state to first-order in
µmax, namely [2, p. 581]:
(cid:32) N(cid:88)
(cid:33)−1 N(cid:88)
k=1
k=1
(cid:18)
ERk (cid:44) µmax
×
lim
µmax→0
lim sup
i→∞
1
µmax
E{ ¯J glob(wk,i) − ¯J glob(wo)}
(cid:19)
.
(52)
(53)
(54)
The average ER across the network is defined by
N(cid:88)
k=1
ERav (cid:44) 1
N
ERk.
By following similar arguments to [2, p. 582], it can be verified
that the excess risk can also be evaluated by computing a
weighted mean-square-error variance:
(cid:18)
(cid:19)
E(cid:107)(cid:101)wk,i(cid:107)2
1
2
¯H
ERk (cid:44) µmax
lim
µmax→0
lim sup
i→∞
1
µmax
where ¯H denotes the Hessian matrix of the normalized aggre-
gate cost, ¯J glob(w), evaluated at the minimizer w = wo:
(cid:32) N(cid:88)
(cid:33)−1 N(cid:88)
¯H (cid:44)
qk
qkHk
(55)
k=1
k=1
with Hk defined by (41). Moreover, we define the convergence
rate as the slowest rate at which the error variances, E(cid:107)(cid:101)wk,i(cid:107)2,
converge to the steady-state region. By iterating the recursion
for the second-order moment of the error vector, we will arrive
at a relation in the following form:
E[(cid:107)(cid:101)wi(cid:107)2] = E(cid:8)(cid:107)(cid:101)w−1(cid:107)2
(cid:9) + c
F i+1
(56)
for some matrix F and constant c, where (cid:101)w−1 denotes the
steady-state region that E[(cid:107)(cid:101)wi(cid:107)2] converges to. Then,
convergence rate of E[(cid:107)(cid:101)wi(cid:107)2] towards its steady-state region is
network error vector at the initial time instant. The first-term
on the right-hand side corresponds to a transient component
that dies out with time, and the second-term denotes the
the
dictated by ρ(F ) [2, p. 395]. The following conclusion is one
of the main results in this work. It shows how the coordinate
descent construction influences performance in comparison to
the standard diffusion strategy where all entries of the gradient
vector are used at each iteration. Following the statement of
the result, we illustrate its implications by considering several
important cases.
Theorem 2: (MSD and ER performance). Under the same
setting of Theorem 1, and assume also that Assumption 3
holds, it holds that, for sufficiently small step-sizes:
MSDcoor,k = MSDcoor,av
=
1
2
Tr
qk(1 − rk)Hk
(cid:32) N(cid:88)
k=1
(57)
(cid:33)−1 N(cid:88)
N(cid:88)
k=1
k=1
(cid:32)
k
kG(cid:48)
q2
(cid:33)
Tcoor
Tgrad
(58)
ERcoor,k = ERcoor,av =
1
2
Tr
X
kG(cid:48)
q2
k
X
(cid:33)
(cid:33)
(cid:32) N(cid:88)
qk(1 − rk)Hk
where the subscript "coor" denotes the stochastic coordinate-
descent diffusion implementation, and matrix X is the unique
solution to the following Lyapunov equation:
(cid:32) N(cid:88)
convergence rate of the error variances, E(cid:107)(cid:101)wk,i(cid:107)2, towards
(cid:17)
(59)
with ¯H defined by (55). Moreover, for large enough i, the
the steady-state region (57) is given by
qk(1 − rk)Hk
(cid:32) N(cid:88)
X = ¯H
(cid:33)
(cid:16)
k=1
k=1
+
αcoor = 1 − 2λmin
qk(1 − rk)Hk
+ O
µ(N +1)/N
max
k=1
Proof : See Appendix B.
(60)
(cid:3)
IV. IMPLICATIONS AND USEFUL CASES
A. Uniform Missing Probabilities
Consider the case when the missing probabilities are iden-
tical across the agents, i.e., {rk ≡ r}.
1) Convergence time: Consider the ATC or CTA forms of
the full-gradient or coordinate-descent diffusion strategy (3a)–
(3c) and (5). From (56), we find that the error variances for
the distributed strategies evolve according to a relation of the
form:
E(cid:107)(cid:101)wk,i(cid:107)2 ≤ αi+1E(cid:107)(cid:101)wk,−1(cid:107)2 + c
(61)
for some constant c > 0, and where the parameter α deter-
mines the convergence rate. Its value is denoted by αgrad for
the full-gradient implementation and is given by [2, p. 584]:
αgrad = 1 − 2λmin
qkHk
+ o (µmax)
(62)
k=1
Likewise, the convergence rate for the coordinate-descent vari-
ant is denoted by αcoor and is given by expression (60). It is
clear that αcoor ≥ αgrad for 0 ≤ r < 1, so that the coordinate-
descent implementation converges at a slower rate as expected
(since it only employs partial gradient information). Thus, let
Tcoor and Tgrad denote the largest number of iterations that
are needed for the error variances, E(cid:107)(cid:101)wk,i(cid:107)2, to converge to
their steady-state regions. The values of Tcoor and Tgrad can be
estimated by assessing the number of iterations that it takes
(cid:32) N(cid:88)
(cid:33)
for the transient term αi+1E(cid:107)(cid:101)wk,−1(cid:107)2 in (61) to assume a
higher-order value in µmax, i.e., for
E(cid:107)(cid:101)wk,−1(cid:107)2 = dµ1+
E(cid:107)(cid:101)wk,−1(cid:107)2 = dµ1+
max
max
αTcoor
coor
αTgrad
grad
6
(63)
(64)
for some proportionality constant d, and small number > 0.
Then, it holds that
=
(a)≈
(b)≈
=
ln αgrad
ln αcoor
(cid:16)
(cid:16)
ln
1 − 2λmin
ln
1 − 2λmin
−2λmin
(cid:17)(cid:17)
k=1 qk(1 − r)Hk
k=1 qkHk
(cid:16)(cid:80)N
(cid:16)(cid:80)N
(cid:16)(cid:80)N
(cid:16)(cid:80)N
k=1 qkHk
(cid:17)
(cid:17)
k=1 qkHk
−2(1 − r)λmin
1
1 − r
(cid:17)(cid:17)
(65)
where in step (a) we ignored the higher-order term in µmax,
and in (b) we used ln(1 − x) ≈ −x as x → 0. Expression
(65) reveals by how much the convergence time is increased
in the coordinate-descent implementation. Note that because
of longer convergence time, the stochastic coordinate-descent
diffusion implementation may require more quantities to be
exchanged across the network compared to the full-gradient
case.
2) Computational complexity: Let us now compare the
computational complexity of both implementations:
the
coordinate-descent and the full-gradient versions. Assume that
the computation required to calculate each entry of the gradient
vector (cid:92)∇wTJ k(φk,i−1) is identical, and let cm ≥ 0 and
ca ≥ 0 denote the number of multiplications and additions,
respectively, that are needed for each entry of the gradient
vector.
Let nk (cid:44) Nk denote the degree of agent k. Then, in the
full-gradient implementation, the adaptation step (3b) requires
cmM + M multiplications and caM + M additions, while the
combination step (3a) or (3c) requires nkM multiplications
and (nk − 1)M additions. In the coordinate-descent imple-
mentation, the adaptation step (3b) with the gradient vector
replaced by (5) requires (1 − r) · (cmM + M ) multiplications
and (1 − r) · (caM + M ) additions on average, while the
combination step (3a) or (3c) requires nkM multiplications
and (nk − 1)M additions. Let mcoor,k and mgrad,k denote the
combined number of multiplications required by the adaptation
and combination steps per iteration at each agent k in the
coordinate-descent and full-gradient cases. Then,
mgrad,k = (cm + nk + 1)M
mcoor,k = mgrad,k − (cm + 1)M r
(66)
(67)
If we now consider that these algorithms take Tcoor and Tgrad
iterations to reach their steady-state regime, then the total
number of multiplications at agent k, denoted by Mcoor,k and
(68)
(69)
(70)
(71)
Mgrad,k, are therefore given by
Mcoor,k = mcoor,kTcoor
Mgrad,k = mgrad,kTgrad
so that using (65):
Mcoor,k
Mgrad,k
=
mcoor,k
mgrad,k
1
1 − r
Now, the first term on the right hand side satisfies
mcoor,k
mgrad,k
= 1 − cm + 1
r
cm + nk + 1
(cid:18)
(cid:19)
so that from (70) and (71):
= (1 − r)−1
1 − cm + 1
1 ≤ Mcoor,k
Mgrad,k
r
cm + nk + 1
implementation will
(72)
since 0 ≤ r < 1. It is clear that when it is costly to compute
the gradient entries, i.e., when cm (cid:29) nk, then Mcoor,k and
Mgrad,k will be essentially identical. This means that while
the coordinate-descent
take longer to
converge, the savings in computation per iteration that it pro-
vides is such that the overall computational complexity until
convergence remains largely invariant (it is not increased).
This is a useful conclusion. It means that in situations where
computations at each iteration need to be minimal (e.g., when
low end sensors are used), then a coordinate-descent variant is
recommended and it will be able to deliver the same steady-
state performance (to first-order in µmax, see (78) ahead)
with the total computational demand spread over a longer
number of iterations. This also means that the complexity and
convergence rate measures, when normalized by the number
of entries that are truly updated at each iteration, remain
effectively invariant. A similar analysis and conclusion holds
if we examine the total number of additions (as opposed to
multiplications) that are necessary.
(cid:18)
be written as
3) MSD performance: The matrix G(cid:48)
k = (1 − r)2Gk +(cid:0)(1 − r) − (1 − r)2(cid:1) diag{Gk}
k defined by (48) can
G(cid:48)
= (1 − r)2
r
1 − r
diag{Gk}
Gk +
(73)
where the term diag{Gk} is a diagonal matrix that consists
of the diagonal entries of Gk. Then, the MSD expression (57)
gives
(cid:19)
(cid:33)−1
qkHk
×
(cid:32)(cid:32) N(cid:88)
(cid:18)
N(cid:88)
k=1
q2
k
k=1
Gk +
r
1 − r
diag{Gk}
(cid:19)(cid:33)
MSDcoor,k
(73)
=
1
2
(1 − r)Tr
k=1
(cid:32) N(cid:88)
(cid:32) N(cid:88)
(cid:32) N(cid:88)
k=1
qkHk
qkHk
qkHk
(cid:33)−1 N(cid:88)
(cid:33)−1 N(cid:88)
(cid:33)−1 N(cid:88)
k=1
k=1
k=1
k=1
=
1
2
r
2
r
2
Tr
Tr
Tr
+
.
q2
kGk
kdiag{Gk}
q2
−
q2
kGk
7
(74)
By recognizing that the first item in (74) is exactly the MSD
expression for the stochastic full-gradient diffusion case [2, p.
594], which is denoted by "MSDgrad,k", we get
MSDcoor,k − MSDgrad,k
(cid:32)(cid:32) N(cid:88)
(cid:33)−1 N(cid:88)
(cid:33)
=
r
2
Tr
qkHk
q2
k
Gk
(75)
k=1
k=1
where
Gk (cid:44) diag{Gk} − Gk.
(76)
We show in Appendix C that the difference in (75) can be
positive or negative, i.e., the MSD performance can be better or
worse in the stochastic coordinate-descent case in comparison
to the stochastic full-gradient case. Recall from (49) that the
MSD performance is evaluated to first-order in µmax. Then,
the MSD gap in (75) is to first-order in the step-size parameter.
Observe that the missing probability r on the right hand side
of that equation is independent of µmax. It thus follows that
(cid:32)(cid:32) N(cid:88)
(cid:33)−1 N(cid:88)
k=1
k=1
(cid:33)
Tr
qkHk
q2
k
Gk
= O(µmax).
(77)
Corollary 1: (Small missing probabilities). Let r =
O(µε
max) for a small number ε > 0. It holds that
MSDcoor,k − MSDgrad,k = O(µ1+ε
(78)
(cid:3)
We proceed to provide a general upper bound for the
Proof : It follows from (75) and (77).
max) = o(µmax).
difference between MSDcoor,k and MSDgrad,k.
Corollary 2: (Upper bound). Under the same conditions of
Theorem 2, and when the missing probabilities are uniform,
namely, {rk ≡ r}, it holds that:
MSDcoor,k − MSDgrad,k ≤
(cid:33)−1(cid:18) 1
(cid:32) N(cid:88)
k=1
r
2
qk
− 1
δd
νd
(cid:19) N(cid:88)
k=1
q2
kTr(Gk)
(79)
where the positive numbers νd ≤ δd are defined in (2), and
the matrices {Gk} are defined by (46). Furthermore, when the
matrices {Hk} or {Gk} are diagonal, it follows that
MSDcoor,k = MSDgrad,k
Proof : See Appendix D.
(80)
(cid:3)
Corollary 3: (Uniform step-sizes). Continuing with the
setting of Corollary 2 by assuming now that the step-sizes
are uniform across all agents and A1 = I or A2 = I
(corresponding to either the ATC or CTA formulations). Let
{pk} be entries of the vector p defined by (34). Then, in view
of (45) and (34), qk = µpk and the {pk} add up to one. In
this case, the sum of the {qk} is equal to µ and expression
(79) simplifies to
MSDcoor,k − MSDgrad,k ≤
(cid:18) 1
r
2
µ
− 1
δd
νd
(cid:19) N(cid:88)
k=1
matrices, i.e., {Ru,k ≡ Ru > 0}, it holds that
0 ≤ MSDcoor,k − MSDgrad,k ≤
(cid:32) N(cid:88)
(cid:33)−1(cid:32) N(cid:88)
r
qk
q2
kσ2
v,k
(cid:33)(cid:18) δd
νd
k=1
k=1
8
(cid:19)
− 1
M (87)
p2
kTr(Gk).
(81)
(cid:3)
Moreover, it holds that MSDcoor,k = MSDgrad,k if, and only
if, Ru is diagonal.
Proof : It follows from Corollary 2 by using Tr (Gk) =
v,kTr (Ru) and noting that νd/2 ≤ λ (Ru) ≤ δd/2 accord-
4σ2
(cid:3)
ing to (85) and (2).
Consider now MSE networks where the risk function that
is associated with each agent k is the mean-square-error:
Jk(w) = E(dk(i) − uk,iw)2
(82)
where dk(i) denotes the desired signal, and uk,i is a (row)
regression vector. In these networks, the data {dk(i), uk,i}
are assumed to be related via the linear regression model
dk(i) = uk,iwo + vk(i)
(83)
where vk(i) is zero-mean white measurement noise with vari-
ance σ2
v,k and assumed to be independent of all other random
variables. The processes {dk(i), uk,i, vk(i)} are assumed to
be jointly wide-sense stationary random processes. Assume
also that the regression data {uk,i} are zero-mean, and white
over time and space with
E uT
k,iu(cid:96),j (cid:44) Ru,kδk,(cid:96)δi,j
(84)
where Ru,k > 0, and δk,(cid:96) denotes the Kronecker delta
sequence. Consider the case when the covariance matrices of
the regressors are identical across the network, i.e., {Ru,k ≡
Ru > 0}. Then, it holds that [2, p. 598]
Hk ≡ 2Ru, Gk = 4σ2
(85)
v,kRu.
qk
k=1
k=1
(86)
q2
kσ2
v,k
= r
≥ 0
(cid:32) N(cid:88)
(cid:33)−1(cid:32) N(cid:88)
Substituting into (75) we have
MSDcoor,k − MSDgrad,k
(cid:33)
Tr(cid:0)R−1
u diag{Ru} − M(cid:1)
u diag{Ru}(cid:1) ≥ M, which
where (86) holds because Tr(cid:0)R−1
can be shown by using the property that Tr (X) Tr(cid:0)X−1(cid:1) ≥
2{Ru}R−1
M 2 for any M ×M symmetric positive-definite matrix X [31,
2{Ru}. In
p. 317], and choosing X = diag 1
the case of MSE networks, by exploiting the special relation
between the matrices {Hk} and {Gk} in (85), we are able
to show that the MSD in the stochastic coordinate-descent
case is always larger (i.e., worse) than or equal to that in
the stochastic full-gradient diffusion case (although by not
more than o(µmax), as indicated by (78)). We are also able
to provide a general upper bound on the difference between
these two MSDs.
u diag 1
(cid:32) N(cid:88)
(cid:33)
4) ER performance: Consider the scenario when the miss-
ing probabilities are identical across the agents, i.e., {rk ≡ r}.
Then, expression (59) simplifies to
(cid:32) N(cid:88)
(cid:33)
(cid:16)(cid:80)N
qk
¯HX = ¯H (88)
(cid:17) ¯H, it
k=1 qk
(1 − r)
X ¯H + (1 − r)
qk
k=1
k=1
follows that
k=1 qkHk =
where we used the equality(cid:80)N
(cid:32) N(cid:88)
(cid:33)−1
(cid:33)−1
(cid:32) N(cid:88)
(cid:33)−1 N(cid:88)
(cid:32) N(cid:88)
(1 − r)−1
(1 − r)−1
ERcoor,k =
X =
1
2
1
4
k=1
k=1
qk
qk
Tr
(a)
=
1
4
qk
q2
kTr (Gk)
k=1
k=1
Thus, the ER expression in (58) can be rewritten as:
IM .
(89)
(cid:32) N(cid:88)
k=1
kG(cid:48)
q2
k
(cid:33)
(90)
which is exactly the same result for the full gradient case
from [2, p. 608], and where the equality (a) holds because
k) = (1 − r)Tr (Gk) according to the definition in (48).
Tr (G(cid:48)
B. Uniform Individual Costs
Consider the case when the individual costs, Jk(w), are
identical across the network, namely, [2, p. 610]
Jk(w) ≡ J(w) (cid:44) EQ(w; xk,i)
(91)
where Q(w; xk,i) denotes the loss function. In this case, it
will hold that the matrices {Hk, Gk} are uniform across the
agents, i.e.,
Hk = ∇2
wJ(wo) ≡ H
(92)
Gk = E∇wT Q(wo; xk,i) [∇wTQ(wo; xk,i)]T ≡ G
(93)
in view of ∇wTJ(wo) = 0. Then, (92) ensures the matrix
¯H = H according to the definition in (55). By referring to
(59), we have
(cid:33)−1
(cid:32) N(cid:88)
k=1
Corollary 4: (MSE networks). Under the same conditions
of Corollary 2, and for MSE networks with uniform covariance
X =
1
2
qk(1 − rk)
IM .
(94)
Then, expressions (57) and (58) reduce to
MSDcoor,k = MSDcoor,av
qk(1 − rk)
kTr(cid:0)H−1G(cid:48)
k
q2
(cid:1)
=
1
2
(cid:32) N(cid:88)
(cid:32) N(cid:88)
k=1
(cid:33)−1 N(cid:88)
(cid:33)−1 N(cid:88)
k=1
ERcoor,k = ERcoor,av
=
1
4
qk(1 − rk)
k=1
k=1
We proceed to compare the MSD and ER performance in the
stochastic full-gradient and coordinate-descent cases. Let
k(1 − rk)Tr (G) .
q2
(cid:80)N
(cid:80)N
(cid:80)N
(cid:80)N
k=1 q2
k
k=1 qk
k=1 q2
k
k=1 qk
−
−
(97)
(cid:80)N
(cid:80)N
k(1 − rk)2
k=1 q2
(cid:80)N
k=1 qk(1 − rk)
(cid:80)N
k(1 − rk)
k=1 q2
k=1 qk(1 − rk)
α (cid:44)
θ (cid:44)
(98)
and note that α ≤ θ, with equality if, and only if, {rk ≡ 0}.
Corollary 5: (Performance comparison). Under the same
conditions of Theorem 2, when the individual costs Jk(w) are
identical across the agents, it holds that:
a) if α ≥ 0:
θ
νd
Tr (G)
(99)
1
2
(cid:18) θ
(cid:19)
(cid:19)
νd
0 ≤ MSDcoor,k − MSDgrad,k ≤ 1
2
b) if α < 0, and θ ≥ (1 − δd/νd) α ≥ 0:
0 ≤ MSDcoor,k − MSDgrad,k ≤
(cid:18) 1
δd
+
− 1
νd
(cid:19)
(cid:19)
α
Tr (G)
(cid:18) θ
(cid:18) 1
c) if α < 0, and θ ≤ (1 − νd/δd) α ≤ 0:
1
2
Tr (G) ≤
− 1
δd
νd
δd
+
α
MSDcoor,k − MSDgrad,k ≤ 0.
Likewise, it holds that
ERcoor,k − ERgrad,k =
θ
4
Tr (G) .
9
(95)
(96)
Fig. 1. Network topology consisting of N = 100 agents.
Fig. 2. MSD learning curves, averaged over 200 independent runs, in the
case of Corollary 1 when {rk = 0.1}. The dashed lines show the theoretical
MSD values from (57).
(100)
(101)
(102)
A. MSE Networks
In the following examples, we will test performance of
the associated algorithms in the case when uniform missing
probabilities are utilized across the agents. We adopt the ATC
formulation, and set the combination matrices A1 = I, and
A2 according to the averaging rule [2, p. 664] in the first two
examples, and Metropolis rule [2, p. 664] in the third example.
In the first example, we test the case when the gradient vectors
are missing with small probabilities across the agents. Figure
1 shows a network topology with N = 100 agents. The
parameter vector wo is randomly generated with M = 10.
Then, in the case when either the missing probabilities or the
quantities {qk} are uniform across the agents, namely, {rk ≡
r} or {qk ≡ q}, it follows that
ERcoor,k = ERgrad,k.
Proof : See Appendix E.
(103)
(cid:3)
Note that for the other choices of parameter θ that are not
indicated in Corollary 5, there is no consistent conclusion on
which MSD (between MSDcoor,k and MSDgrad,k) is lower.
V. SIMULATION RESULTS
In this section, we illustrate the results by considering MSE
networks and logistic regression networks; both settings satisfy
condition (2) and Assumptions 1 through 3.
Fig. 3. MSD learning curves, averaged over 200 independent runs, in the
case of Corollary 2 when {Hk, Gk} are diagonal. The dashed line along the
horizontal axis shows the theoretical MSD value from (57). Those along the
learning curves show the reference recursion at rates formulated by (60).
0.511.522.533.54x 104−60−50−40−30−20−10010IterationsTransient MSD (dB) Theoretical MSD for full updateTheoretical MSD for partial updateMSE network, full updateMSE network, partial update0.20.40.60.811.21.41.61.82x 104−60−50−40−30−20−10010IterationsTransient MSD (dB) Theoretical MSDMSE network, full updateMSE network, partial updateconvergence rate for full updateconvergence rate for partial update 10
(a)
(b)
(c)
Fig. 4. Learning curves, averaged over 10000 independent runs, and theoretical results calculated from (57) and (58) respectively, for a two-agent MSE
network, with parameters {π1 = −0.34, π2 = 0.99} in (a), and {π1 = 0.34, π2 = 0.99} in (b) and (c).
(cid:113)
1 − π2
The regressors are generated by the first-order autoregressive
model
ktk,i(m), 1 ≤ m < M
uk,i(m) = πkuk,i(m − 1) +
(104)
and the variances are scaled to be 1. The processes {tk,i}
are zero-mean, unit-variance, and independent and identically
distributed (i.i.d) Gaussian sequences. The parameters {πk}
are generated from a uniform distribution on the interval
(−1, 1). The noises, uncorrelated with the regression vectors,
are zero-mean white Gaussian sequences with the variances
uniformly distributed over (0.001, 0.1). The step-sizes {µk}
across the agents are generated from a uniform distribution
on the interval (0.0001, 0.0005). We choose a small missing
probability {rk = 0.1}. Figure 2 shows the simulation re-
sults, which are averaged over 200 independent runs, as well
as the theoretical MSD values calculated from (57), which
are −57.72dB and −57.61dB, respectively, for the full and
partial update case. It is clear from the figure that, when the
gradient information is missing with small probabilities, the
performance of the coordinate-descent case is close to that of
the full-gradient diffusion case.
In the second example, we test the case when the regressors
are white across the agents. We randomly generate wo of
size M = 10. The white regressors are generated from zero-
mean white Gaussian sequences, and the powers, which vary
from entry to entry, and from agent to agent, are uniformly
distributed over (0.05, 0.15). The noises {vk(i)}, uncorrelated
with the regressors, are zero-mean white Gaussian sequences,
with the variances {σ2
v,k} generated from uniform distribution
on the interval (0.0001, 0.01). The step-sizes are uniformly
distributed over (0.001, 0.01). The results, including the the-
oretical MSD value from (57) in Theorem 2, the simulated
MSD learning curves, and the convergence rates from (60),
are illustrated by Fig. 3, where the results are averaged over
200 independent runs. It is clear from the figure that, when
white regressors are utilized in MSE networks, the stochastic
coordinate-descent case converges to the same MSD level
as the full-gradient diffusion case, which verifies (80), at a
convergence rate formulated in (60).
In the third example, we revisit the two-agent MSE network
discussed in Appendix C, i.e., N = 2. We randomly generate
wo of size M = 2. The step-sizes µ1 = µ2 = 0.005 are
uniform across the agents, which gives q1 = q2 = 2.5 ×
Fig. 5. Network topology consisting of N = 20 agents.
10−3. The missing probabilities r1 = r2 = 0.5. The noises
{v1(i), v2(i)} are zero-mean white Gaussian sequences with
the variances {σ2
v,2 = 5 × 10−4}. The regressors,
uncorrelated with the noise sequences, are scaled such that the
covariance matrices are of the form
v,1 = 0.5, σ2
(cid:20) π1 π1
(cid:21)
(cid:20) π2 π2
π2
1
(cid:21)
π1
1
Ru,1 =
, Ru,2 =
(105)
with π1 < 1,π2 < 1. Now we select parameters {π1 =
−0.34, π2 = 0.99}, which satisfy condition (169a), and
{π1 = 0.34, π2 = 0.99} to illustrate the cases of MSDcoor,k <
MSDgrad,k and MSDcoor,k > MSDgrad,k respectively. Fig. 4
(a) shows the simulation results with the parameters {π1 =
−0.34, π2 = 0.99}. Figures 4 (b) and 4 (c) show the simu-
lation results with the parameters {π1 = 0.34, π2 = 0.99}.
All results are averaged over 10000 independent runs. It is
clear from the figures that the simulation results match well
with the theoretical results from Theorem 2. In Fig. 4 (a), the
steady-state MSD of the stochastic coordinate-descent case is
slightly lower than that of the full-gradient diffusion case, by
about 0.32dB, which is close to the theoretical MSD difference
of 0.41dB from (75). The MSD performance is better in the
full-gradient diffusion case in Fig. 4 (b), and the difference
between these two MSDs at steady state is 1.71dB, which is
close to the theoretical difference of 1.49dB from (75). The
ER performance for both the stochastic coordinate-descent and
full-gradient diffusion cases are the same as illustrated in Fig.
4 (c), which verifies the theoretical result in (90).
500100015002000−30−25−20−15−10−50IterationsTransient MSD (dB) 1700180019002000−29.5−29−28.5−28 Theoretical MSD for full updateTheoretical MSD for partial updateMSE network, full updateMSE network, partial update100020003000400050006000−30−25−20−15−10−505IterationsTransient MSD (dB) Theoretical MSD for full updateTheoretical MSD for partial updateMSE network, full updateMSE network, partial update1000200030004000500060007000−30−25−20−15−10−505IterationsTransient ER (dB) Theoretical ERMSE network, full updateMSE network, partial update1125171021587496163181411201319 11
(a)
(b)
(c)
Fig. 6. ER learning curves, averaged over 1000 independent runs, and theoretical results from (96) for diffusion learning over a logistic network with full or
partial updates. Corollary 5 is tested in (a) when a uniform step-size and a doubly-stochastic combination matrix are utilized across the network. Corollary 5
is tested when the parameters {µk} and {rk} are scaled to make θ in (98) negative in (b) and positive in (c).
B. Logistic Networks
We now consider an application in the context of pattern
classification. We assign with each agent the logistic risk
(cid:107)w(cid:107)2 + E(cid:110)
(cid:104)
Jk(w) =
ρ
2
k,iw(cid:105)(cid:111)
1 + e−γk(i)hT
ln
(106)
with regularization parameter ρ > 0, and where the labels
{γk(i) = ±1} are binary random and the {hk,i} are feature
vectors. The objective is for the agents to determine a param-
eter vector wo to enable classification by estimating the class
labels via (cid:98)γk(i) = hT
k,iwo.
We proceed to test the theoretical findings in Corollary 5.
Consider the network topology in Fig. 5 with N = 20 agents.
We still adopt the ATC formulation, and set the combination
matrices A1 = I, and A2 according to the Metropolis rule in
[2, p. 664]. The feature vectors and the unknown parameter
vector are randomly generated from uncorrelated zero-mean
unit-variance i.i.d Gaussian sequences, both of size M = 10.
The parameter ρ in (106) is set to 0.01. To generate the
trajectories for the experiments, the optimal solution to (106),
wo, the Hessian matrix H, and the gradient noise covariance
matrix, G, are first estimated off-line by applying a batch
algorithm to all data points.
In the first example, we consider the case when a uniform
step-size {µk = 0.005} is utilized across the agents. All entries
of the stochastic gradient vectors are missing completely at
random with positive probabilities that are uniformly dis-
tributed over (0, 1). Figure 6 (a) shows the transient ER curves
for the diffusion strategies with complete and partial gradients,
where the results are averaged over 1000 independent runs.
The figure also shows the theoretical result calculated from
(96). It is clear from Figure 6 (a) that the same ER performance
is obtained in the stochastic coordinate-descent and full-
gradient diffusion cases, by utilizing a uniform step-size and
a doubly-stochastic combination matrix across the agents (in
which case the parameters {qk} in (45) are identical across the
agents), which is in agreement with the theoretical analysis in
(103).
In the second and third examples, we randomly generate the
step-sizes {µk} and missing probabilities {rk} by following
uniform distributions on the intervals (0.001, 0.01) and (0, 1)
respectively. In Figure 6 (b), the parameters {µk} and {rk}
are scaled to get a negative value for θ in (98), and in
Fig. 6 (c), those parameters are scaled to make θ positive.
Figures 6 (b) and 6 (c) show respectively the transient ER
learning curves in these two cases for the diffusion strategies
with complete and partial gradients, where the results are
averaged over 1000 independent runs. The figures also show
the theoretical results calculated from (96). It is clear from
Figs. 6 (b) and 6 (c) that these learning curves converge to
their theoretical results at steady state. In Fig. 6 (b) where
θ < 0, the stochastic coordinate-descent case converges to
a lower ER level than the full-gradient diffusion case, and
the difference between these two ERs is 0.637dB, which is
close to the theoretical difference of 0.640dB from (102).
In Fig. 6 (c) where θ > 0, the steady-state ER in the full-
gradient diffusion case is lower than that of the stochastic
coordinate-descent case, by about 0.726dB, which is close to
the theoretical difference of 0.750dB from (102).
APPENDIX A
PROOF OF THEOREM 1
Let P = A1A2. It was argued in [2, p.510] that P admits
a Jordan canonical decomposition of the form P = VJV −1
where
V (cid:44)(cid:2) p VR
(cid:3) , V −1
(cid:44)
(cid:20) 1
(cid:21)
, J =
0
0 J
(107)
(cid:21)
(cid:20) 1T
V T
L
p is defined by (34), denotes an arbitrary positive scalar that
we are free to choose, and the matrix J has a Jordan structure
with appearing in the first lower diagonal rather than unit
entries. All eigenvalues of J are strictly inside the unit circle.
Then,
P (cid:44) P ⊗ IM (cid:44) VJ V−1
(108)
where V (cid:44) V ⊗ IM , J (cid:44) J ⊗ IM . Using (108), we can
rewrite Bi from (31) as
(J T − DT
i )V T
(cid:1)T
Bi (cid:44)(cid:0)V−1
(cid:20) DT
(cid:44) V T
AT
=
DT
(cid:21)
2MΓiHi−1AT
11,i DT
12,i DT
21,i
22,i
1
(cid:0)V−1
(cid:1)T
(109)
(110)
DT
i
where
0.511.522.533.5x 104−45−40−35−30−25−20−15−10−5IterationsTransient ER (dB) Theoretical ERlogistic network, full updatelogistic network, partial update0.511.522.533.5x 104−45−40−35−30−25−20−15−10−5 IterationsTransient ER (dB)3.53.63.7x 104−42−40 Theoretical ER for full updateTheoretical ER for partial updatelogistic network, full updatelogistic network, partial update123456x 104−45−40−35−30−25−20−15−10−5IterationsTransient ER (dB) 5.85.825.845.86x 104−42−40 Theoretical ER for full updateTheoretical ER for partial updatelogistic network, full updatelogistic network, partial update and
N(cid:88)
D11,i =
qkHk,i−1Γk,i
(111)
k=1
with the vector q = {qk} defined by (45). With regards to the
norm of D11,i, we observe that contrary to the arguments
in [2, p. 511],
this matrix is not symmetric anymore in
the coordinate-descent case due to the presence of Γk,i. We
therefore need to adjust the arguments, which we do next.
¯D11,i (cid:44) E [D11,iF i−1]
=
k=1
k=1
(6)
=
(cid:3) .
qkHk,i−1E [Γk,i]
qk(1 − rk)Hk,i−1
11,iF i−1
N(cid:88)
N(cid:88)
= E(cid:2)DT
(cid:26) (1 − rk)(1 − rj), k (cid:54)= j
(cid:1)(cid:0)D11,i − ¯D11,i
(cid:3) − ¯D11,iE(cid:2)DT
(cid:3) − ¯D2
11,iF i−1
1 − rk,
11,iF i−1
k = j
11,i
11,i
(112)
(113)
(114)
(cid:1)T F i−1
(cid:105)
(cid:3)−
11,iF i−1
E [D11,iF i−1] ¯D11,i + ¯D2
qkqjHk,i−1E [Γk,iΓj,i] Hj,i−1−
qkqj(1 − rk)(1 − rj)Hk,i−1Hj,i−1
qkqj(1 − rk)(1 − rj)Hk,i−1Hj,i−1−
qkqj(1 − rk)(1 − rj)Hk,i−1Hj,i−1+
Let
Noting that
we introduce
E [Γk,iΓj,i] =
=
j=1
k=1
(112)
RD11,i (cid:44) E(cid:104)(cid:0)D11,i − ¯D11,i
= E(cid:2)D11,iDT
= E(cid:2)D11,iDT
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
k(1 − rk)H 2
q2
k(1 − rk)H 2
q2
j(cid:54)=k=1
(113)
k=1
k=1
k=1
k=1
k=1
j=1
j=1
=
=
k,i−1
k,i−1 − N(cid:88)
k(1 − rk)2H 2
q2
k,i−1
k=1
(115)
=
k(1 − rk)rkH 2
q2
k,i−1.
k=1
Recall, from (2) and (33), that
0 < νdIM ≤ Hk,i−1 ≤ δdIM .
(116)
Then, matrices ¯D11,i and RD11,i are symmetric positive-
definite. Following similar arguments to those in [2, pp. 511–
12
512], we have
(cid:107)IM − ¯D11,i(cid:107) ≤ 1 − σ11µmax, (cid:107)RD11,i(cid:107) ≤ β2
for some positive constants σ11 and β2
µmax.
11µ2
max (117)
11, and sufficiently small
(118)
, we have
2MΓisi
(cid:21)
V T
V T
Now, multiplying both sides of (30) by V T
(cid:101)wi = (J T − DT
(cid:101)wi−1 + V T
i )V T
AT
(cid:20) (cid:0)pT ⊗ IM
(cid:1) (cid:101)wi
(cid:21)
(cid:20) ¯wi
(cid:0)V T
(cid:1) (cid:101)wi
where (109) was used. Let
(cid:0)qT ⊗ IM
(cid:1) Γisi
(cid:20)
(cid:21)
(cid:1)AT
(cid:0)V T
(cid:21)(cid:20) ¯wi−1
(cid:20) IM − DT
(cid:101)wi =
2MΓisi =
AT
V T
(cid:21)
We then rewrite (118) as
2MΓisi
R ⊗ IM
R ⊗ IM
wi
(cid:44)
11,i
(cid:20) ¯wi
(119)
(cid:20) ¯si
(cid:21)
(cid:20) ¯si
si
+
(120)
(cid:21)
(cid:44)
(cid:21)
−DT
− DT
J T
21,i
=
−DT
wi−1
22,i
12,i
wi
si
(121)
where the asymmetry of the matrix D11,i in this case leads
to a difference in the first row, compared to the arguments in
[2, pp. 514–515]. We adjust the arguments as follows. Using
Jensen's inequality, we have [2, p. 515]:
E[(cid:107)(IM − DT
11,i) ¯wi−1(cid:107)2F i−1]
E[(cid:107) ¯wi(cid:107)2F i−1] ≤ 1
1 − t
E[(cid:107)DT
21,i wi−1(cid:107)2F i−1] + E[(cid:107)¯si(cid:107)2F i−1]
(122)
+
1
t
(a)≤ λmax
11,i) ¯wi−1 − DT
11,i) ¯wi−1(cid:107)2F i−1
for any 0 < t < 1, where the expectation of the cross term
between ¯si and (IM − DT
21,i wi−1 vanishes
conditioned on F i−1 and Γi in view of (15), and the result in
(122) follows by taking the expectations again on both sides
over Γi. Then, the first term on the right hand side of (122)
can be bounded by
E(cid:2)(cid:107)(IM − DT
(cid:3)
= ( ¯wi−1)T E(cid:2)(IM − D11,i)(cid:0)IM − DT
(cid:1)F i−1
(cid:3) ¯wi−1
(cid:1)F i−1
(cid:0)E(cid:2)(IM − D11,i)(cid:0)IM − DT
(cid:3)(cid:1)(cid:107) ¯wi−1(cid:107)2
= (cid:13)(cid:13)E(cid:2)(IM − D11,i)(cid:0)IM − DT
(cid:3)(cid:13)(cid:13)(cid:107) ¯wi−1(cid:107)2
(cid:1)F i−1
= (cid:13)(cid:13)IM − 2 ¯D11,i + E(cid:2)D11,iDT
(cid:3)(cid:13)(cid:13)(cid:107) ¯wi−1(cid:107)2
= (cid:13)(cid:13)IM − 2 ¯D11,i + ¯D2
(cid:13)(cid:13)(cid:107) ¯wi−1(cid:107)2
≤(cid:16)(cid:13)(cid:13)(cid:0)IM − ¯D11,i
(cid:13)(cid:13)(cid:17)(cid:107) ¯wi−1(cid:107)2
(cid:1)2(cid:13)(cid:13) +(cid:13)(cid:13)RD11,i
(cid:16)(cid:13)(cid:13)IM − ¯D11,i
(cid:13)(cid:13)(cid:17)(cid:107) ¯wi−1(cid:107)2
(cid:13)(cid:13)2
+(cid:13)(cid:13)RD11,i
≤(cid:0)(1 − σ11µmax)2 + β2
(cid:1)(cid:107) ¯wi−1(cid:107)2
11,iF i−1
11,i + RD11,i
(123)
(c)
=
(112)
(114)
11,i
11,i
11,i
(b)
11µ2
max
where in step (a) we called upon the Rayleigh-Ritz character-
ization of eigenvalues [32], [33], and (b), (c) hold because
(cid:107)A(cid:107) = λmax(A) for any symmetric positive-semidefinite
matrix A, and (cid:107)A2(cid:107) = (cid:107)A(cid:107)2 for any symmetric matrix A.
Computing the expectations again on both sides of (123),
we have
1
1 − t
E(cid:8)E[(cid:107)(IM − DT
(cid:18)
≤ 1
1 − t
1 − σ11µmax +
(cid:0)(1 − σ11µmax)2 + β2
11,i) ¯wi−1(cid:107)2F i−1](cid:9)
(cid:19)
11µ2
β2
11µ2
1 − σ11µmax
(a)
=
max
max
(cid:1) E(cid:107) ¯wi−1(cid:107)2
E(cid:107) ¯wi−1(cid:107)2
(b)≤ (1 − σ(cid:48)
11µmax) E(cid:107) ¯wi−1(cid:107)2
(124)
−1 β2
11 < σ11, and µmax is small enough such that σ(cid:48)
where in step (a) we set t = σ11µmax, and in (b) positive
11 ≤
number σ(cid:48)
σ11 − (1 − σ11µmax)
11µmax. We can now establish (35)
by substituting (124) into (122), and completing the argument
starting from Eq. (9.69) in the proof of Theorem 9.1 in [2, pp.
516–521], where the quantity b = 0M N (appeared in (9.54)
of [2]).
1
(125)
for any 0 < t < 1, which appeared in (9.117) of [2]. Let
We next establish (36). Compared to the proof for Theorem
9.2 in [2], the main difference, apart from the second-order
moments evaluated in (124), is the term
(1 − t)3
E(cid:2)(cid:107)(IM − DT
11,i) ¯wi−1(cid:107)4(cid:3)
Ki (cid:44) (IM − D11,i)(cid:0)IM − DT
(cid:1) ¯wi−1 ( ¯wi−1)T
× (IM − D11,i)(cid:0)IM − DT
(cid:1)
(cid:1)(cid:1)2
Li (cid:44)(cid:0)(IM − D11,i)(cid:0)IM − DT
(cid:3)
E(cid:2)(cid:107)(IM − DT
11,i) ¯wi−1(cid:107)4F i−1
Then, both matrices Ki and Li are symmetric positive semi-
definite. Thus, we have
(126)
(127)
11,i
11,i
11,i
.
=(cid:0) ¯wT
= ( ¯wi−1)T E [KiF i−1] ¯wi−1
≤ λmax (E [KiF i−1])(cid:107) ¯wi−1(cid:107)2
(a)≤ Tr (E [KiF i−1])(cid:107) ¯wi−1(cid:107)2
E [LiF i−1] ¯wi−1
i−1
≤ λmax (E [LiF i−1])(cid:107) ¯wi−1(cid:107)4
= (cid:107)E [LiF i−1](cid:107)(cid:107) ¯wi−1(cid:107)4
(cid:1)(cid:107) ¯wi−1(cid:107)2
(128)
where the inequality (a) holds because λmax(Σ) ≤ Tr(Σ) for
any symmetric positive semi-definite matrix Σ. We proceed to
deal with the term E [LiF i−1]. Note that
Li = IM − L1,i + L2,i − L3,i + L4,i
(129)
where
L1,i (cid:44) 2D11,i + 2DT
L2,i (cid:44) 3D11,iDT
11,i
11,i + DT
11,iD11,i + (D11,i)2 +(cid:0)DT
(130)
(cid:1)2
L3,i (cid:44) (D11,i)2 DT
L4,i (cid:44)(cid:0)D11,iDT
D11,iDT
11,i
(cid:1)2
11,i + D11,i
11,iD11,i + DT
+
11,iD11,iDT
11,i
11,i
and we have E [L1,iF i−1] = 4 ¯D11,i according to (112).
11,i
(131)
(132)
(133)
(cid:0)DT
(cid:1)2
13
Let X be a constant matrix of size M × M. Then,
E [Γk,iXΓj,i] =
(134)
where X(cid:48) has the same form as (48), and we can further
rewrite X(cid:48) as (73). It follows that:
(cid:26) (1 − rk)(1 − rj)X, k (cid:54)= j
(cid:3) −(cid:0) ¯D11,i
(cid:1)2
X(cid:48),
k = j
=
j=1
j=1
k=1
k=1
E(cid:2)DT
11,iD11,iF i−1
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
j(cid:54)=k=1
k(1 − rk)2H 2
q2
N(cid:88)
N(cid:88)
N(cid:88)
(134)
k=1
k=1
k=1
j=1
=
=
k=1
qkqjE [Γk,iHk,i−1Hj,i−1Γj,iF i−1]−
qkqj(1 − rk)(1 − rj)Hk,i−1Hj,i−1
qkqj(1 − rk)(1 − rj)Hk,i−1Hj,i−1+
N(cid:88)
k=1
k,i−1 +
k(1 − rk)rkdiag{H 2
q2
k,i−1}−
qkqj(1 − rk)(1 − rj)Hk,i−1Hj,i−1
k(1 − rk)rkdiag{H 2
q2
k,i−1}.
(135)
k,i−1 > 0}
k,i−1} > 0}. Computing Euclidean norms on
Recall from (116) that {Hk,i−1 > 0}. Then, {H 2
and {diag{H 2
both sides of (135), we have
11,iD11,iF i−1
k(1 − rk)rk
q2
k=1
(cid:13)(cid:13)E(cid:2)DT
(a)≤ N(cid:88)
(b)≤ N(cid:88)
≤ N(cid:88)
(116)≤ N(cid:88)
k=1
k=1
(cid:1)2(cid:13)(cid:13)
(cid:3) −(cid:0) ¯D11,i
k,i−1}(cid:13)(cid:13)
(cid:13)(cid:13)diag{H 2
(cid:0)Tr(cid:2)H 2
(cid:3)(cid:1)
(cid:0)M λmax
(cid:0)H 2
(cid:0)M δ2
(cid:1)
k,i−1
d
k,i−1
k(1 − rk)rk
q2
k(1 − rk)rk
q2
k(1 − rk)rk
q2
(cid:1)(cid:1)
(136)
k=1
where in step (a) we used the property that (cid:107)A + B(cid:107) ≤
(cid:107)A(cid:107) + (cid:107)B(cid:107) [33], and (b) holds because (cid:107)X(cid:107) ≤ Tr(X),
for any symmetric positive semi-definite matrix X, and that
Tr[diag{X}] = Tr[X]. Recall from (45) that [2, p. 509]
qk = µk(eT
k A2p) (cid:44) µmaxτk(eT
k A2p)
(137)
where ek denotes the k-th basis vector, which has a unit entry
at the k-th location and zeros elsewhere, and the parameter τk
satisfies µk = µmaxτk. Then, we have
(cid:3) −(cid:0) ¯D11,i
(cid:1)2(cid:13)(cid:13) = O(µ2
(cid:13)(cid:13)E(cid:2)DT
11,iD11,iF i−1
max)
(138)
11,i
Recall from (117) and (114) that
Likewise, it follows that
(D11,i)2 F i−1
(cid:1)2 F i−1
(cid:13)(cid:13)(cid:13)E(cid:104)
(cid:105) −(cid:0) ¯D11,i
(cid:13)(cid:13)(cid:13)E(cid:104)(cid:0)DT
(cid:105) −(cid:0) ¯D11,i
(cid:13)(cid:13)E(cid:2)D11,iDT
(cid:3) −(cid:0) ¯D11,i
(cid:13)(cid:13)E [L2,iF i−1] − 6(cid:0) ¯D11,i
(cid:13)(cid:13)E [L3,iF i−1] − 4(cid:0) ¯D11,i
(cid:13)(cid:13)E [L4,iF i−1] −(cid:0) ¯D11,i
(cid:1)2(cid:13)(cid:13)(cid:13) = O(µ2
(cid:1)2(cid:13)(cid:13)(cid:13) = O(µ2
(cid:1)2(cid:13)(cid:13) = O(µ2
(cid:1)2(cid:13)(cid:13) = O(µ2
(cid:1)3(cid:13)(cid:13) = O(µ3
(cid:1)4(cid:13)(cid:13) = O(µ4
11,iF i−1
Substituting (138)–(141) into (131), we obtain
Similarly, it can be verified that
max)
max).
max).
max)
max).
max).
(139)
(140)
(141)
(142)
(143)
(144)
It follows that
(cid:107)E [LiF i−1](cid:107) = (cid:107)I − E [L1,iF i−1] + E [L2,iF i−1]−
+
+
(cid:1)3
(cid:1)2(cid:17)
E [L3,iF i−1] + E [L4,iF i−1](cid:107)
=(cid:13)(cid:13)I − 4 ¯D11,i + 6(cid:0) ¯D11,i
(cid:1)2 − 4(cid:0) ¯D11,i
(cid:16)E [L2,iF i−1] − 6(cid:0) ¯D11,i
(cid:1)4
(cid:0) ¯D11,i
−(cid:16)E [L3,iF i−1] − 4(cid:0) ¯D11,i
(cid:1)3(cid:17)
(cid:1)4(cid:17)(cid:13)(cid:13)
(cid:16)E [L4,iF i−1] −(cid:0) ¯D11,i
≤(cid:13)(cid:13)I − 4 ¯D11,i + 6(cid:0) ¯D11,i
(cid:1)2 − 4(cid:0) ¯D11,i
(cid:1)3
(cid:1)2(cid:13)(cid:13)
(cid:1)4(cid:13)(cid:13) +(cid:13)(cid:13)E [L2,iF i−1] − 6(cid:0) ¯D11,i
(cid:0) ¯D11,i
+(cid:13)(cid:13)E [L3,iF i−1] − 4(cid:0) ¯D11,i
(cid:1)3(cid:13)(cid:13)
+(cid:13)(cid:13)E [L4,iF i−1] −(cid:0) ¯D11,i
(cid:1)4(cid:13)(cid:13)
(cid:1)4 (cid:107) + O(µ2
= (cid:107)(cid:0)I − ¯D11,i
+
+
max)
= (cid:107)I − ¯D11,i(cid:107)4 + O(µ2
max)
(117)≤ (1 − σ11µmax)4 + O(µ2
max).
(145)
Substituting into (128), and taking expectations again on both
sides, we have
(1 − t)3
1
11,i) ¯wi−1(cid:107)4(cid:3)
E(cid:2)(cid:107)(IM − DT
= (cid:0)1 − σ11µmax + O(µ2
(1 − t)3
(cid:16)
≤
1
(a)
(1 − σ11µmax)4 + O(µ2
max)
max)(cid:1) E(cid:107) ¯wi−1(cid:107)4
(cid:17) E(cid:107) ¯wi−1(cid:107)4
≤ (1 − σ(cid:48)(cid:48)
11µmax) E(cid:107) ¯wi−1(cid:107)4
(146)
for some positive constant σ(cid:48)(cid:48)
11 < σ11, and for small enough
µmax, where in step (a) we set t (cid:44) σ11µmax. Then, the result
in (36) can be obtained by continuing from Eq. (9.117) (by
choosing t = σ11µmax) in the proof of Theorem 9.2 in [2, pp.
523–530].
14
Then, by following similar techniques shown in the proof of
Lemma 9.5 [2, pp. 542–546], we have
(I − F)−1 = [(p ⊗ p)(1 ⊗ 1)T] ⊗ Z−1 + O(1)
(148)
where
Z (cid:44) N(cid:88)
k=1
qk(1 − rk) [(Hk ⊗ IM ) + (IM ⊗ Hk)] .
(149)
The desired results (57) and (58) in Theorem 2 now follow
by referring to the proofs of Theorem 11.2 and Lemma 11.3
in [2, pp. 583–596], and Theorem 11.4 in [2, pp. 608–609].
(cid:111)
] = E(cid:110)(cid:107)(cid:101)w(cid:48)
Evaluating the squared Euclidean norms on both sides of
(38) and taking expectations conditioned on F i−1, then taking
expectations again we get
E[(cid:107)(cid:101)w(cid:48)
i(cid:107)2
bvec(INM )
i−1(cid:107)2Fbvec(INM )
E(cid:110)(cid:107)si(cid:107)2
(cid:111)
(150)
E[(ΓiMA2)⊗b(ΓiMA2)]bvec(INM )
σ = (cid:107)x(cid:107)2
where we used the weighted vector notation (cid:107)x(cid:107)2
with σ = bvec(Σ) and bvec(·) denoting the block vector
(cid:111)
operation [2, p. 588]. Iterating the relation we get
−1(cid:107)2F i+1bvec(INM )
] = E(cid:110)(cid:107)(cid:101)w(cid:48)
+
+
Σ
E[(cid:107)(cid:101)w(cid:48)
i(cid:107)2
i(cid:88)
bvec(INM )
E(cid:110)(cid:107)si(cid:107)2
E[(ΓiMA2)⊗b(ΓiMA2)]F nbvec(INM )
(151)
n=0
dies out with time, and the convergence rate of E(cid:107)(cid:101)wk,i(cid:107)2
where the first-term corresponds to a transient component that
towards the steady-state regime is seen to be dictated by ρ (F)
[2, p. 592]. Now, let
(cid:111)
Γ (cid:44) EΓi = diag{(1 − rk)}N
M(cid:48) (cid:44) MΓ = diag{µk(1 − rk)}N
k=1 ⊗ IM
k=1 ⊗ IM
2 (I − M(cid:48)H)AT
1
B(cid:48) (cid:44) EB(cid:48)
i = AT
We now rewrite (147) in terms of B(cid:48) as
F (39)
= E(cid:2)(cid:0)AT
= (A1 ⊗b A1)(cid:0)I − I ⊗b (HΓM) − (HΓM) ⊗b I+
E [(HΓiM) ⊗b (HΓiM)](cid:1) (A2 ⊗b A2)
2 (I − MΓiH)AT
(cid:1) ⊗b
2 (I − MΓiH)AT
(cid:0)AT
1
1
max)
= [B(cid:48) ⊗b B(cid:48)]T + ∆F (µ2
(155)
max) is a matrix whose entries are in the order
where ∆F (µ2
of O(µ2
max). Following similar techniques to the proof of
Theorem 9.3 [2, pp. 535–540], we make the same Jordan
canonical decomposition for matrix P = A1A2 as (108), then
substituting into (154) we get
(152)
(153)
(154)
(cid:1)(cid:3)T
(156)
(157)
Define
APPENDIX B
PROOF OF THEOREM 2
F (cid:44) E(cid:2)B(cid:48)
i ⊗b B(cid:48)
i
(cid:3)T
where
(147)
B(cid:48) =(cid:0)V−1
(cid:1)T
(cid:20) D(cid:48)T
D(cid:48)T (cid:44) V T
=
(cid:21)
2M(cid:48)HAT
AT
11 D(cid:48)T
D(cid:48)T
12 D(cid:48)T
22
21
1
(J T − D(cid:48)T)V T
(cid:0)V−1
(cid:1)T
11. Let
T = diag{µ1/N
and
N(cid:88)
k=1
D(cid:48)
11 =
qk(1 − rk)Hk, D(cid:48)
21 = O(µmax).
(158)
(cid:44) U ΛU T for
We now introduce the eigen-decomposition D(cid:48)T
the symmetric positive-definite matrix D(cid:48)T
11, where U is unitary,
and Λ is a diagonal matrix composed of the eigenvalues of
D(cid:48)T
11
maxU, µ2/N
maxIM , . . . , µ(N−1)/N
max
IM , µmaxIM}
(cid:21)
(159)
(160)
then we have
B(cid:48)(cid:0)V−1
(cid:1)T T =
T −1V T
(cid:20) B(cid:48)
11 B(cid:48)
B(cid:48)
21 B(cid:48)
(cid:44) IM − Λ, B(cid:48)
12 = O(µ(N +1)/N
max
12
22
.
11
It follows that B(cid:48)
) [2,
p. 538]. The matrix B(cid:48) has the same eigenvalues as the
block matrix on the right hand side of (160). By referring to
Gershgorin's Theorem [32], [33], it is shown in [2, pp. 539–
540] that the union of the M Gershgorin discs, each centered
at an eigenvalue of B(cid:48)
), is disjoint
from that of the other M (N − 1) Gershgorin discs, centered
at the diagonal entries of B(cid:48)
ρ (B(cid:48)) = ρ (B(cid:48)
22, and therefore
11) + O(µ(N +1)/N
11 with radius O(µ(N +1)/N
(161)
max
max
).
Let
∆F (cid:44)(cid:0)T TV−1
(cid:16)V
(cid:1) ⊗b
(cid:0)T TV−1
(cid:16)V
(cid:0)T −1(cid:1)T(cid:17) ⊗b
(cid:1) ∆F (µ2
(cid:0)T −1(cid:1)T(cid:17)
max)×
.
(162)
=
O(µ2
max)
o(µ1/N
max)
o(µ2
max)
O(µ2
max)
o(µ2/N
max)
...
o(µ1+1/N
max
)
O(µ2
max)
It follows from (162) that all the diagonal blocks of ∆F are
max), the remaining block matrices in the
in the order of O(µ2
first row are in the order of o(µ1/N
max), the remaining block
matrices in the first column are in the order of o(µ2
max), and
the upper and lower triangular blocks in the (2, 2)th block of
∆F are in the order of o(µ2/N
) respectively.
Then, substituting (160) and (162) into (155), we have
max) and o(µ1+1/N
(cid:1) ⊗b
(cid:0)T TV−1
(cid:18)(cid:20) B(cid:48)
(cid:20) F11 F12
(cid:0)T TV−1
(cid:21)
(cid:21)T
11 B(cid:48)
21 B(cid:48)
B(cid:48)
=
12
22
=
F21 F22
max
(cid:0)T −1(cid:1)T(cid:17) ⊗b
(cid:1)F(cid:16)V
(cid:21)(cid:19)T
(cid:20) B(cid:48)
11 B(cid:48)
21 B(cid:48)
B(cid:48)
12
22
⊗b
(cid:16)V
(cid:0)T −1(cid:1)T(cid:17)
+ ∆F
(163)
11 ⊗ B(cid:48)
where
F11 = B(cid:48)
Recall that B(cid:48)
we have
11 + O(µ2
11 is a diagonal matrix, so is B(cid:48)
max), F12 = O(µ(N +1)/N
).
11 ⊗ B(cid:48)
max
(164)
11, then
diag{F11} = diag{ λ (B(cid:48)
11 ⊗ B(cid:48)
11)} + O(µ2
max)
(165)
15
11⊗ B(cid:48)
11 perturbed by a second-order term, O(µ2
which means that the diagonal entries of F11 are the eigenval-
ues of B(cid:48)
max).
Referring to Gershgorin's Theorem,
the union of the M 2
Gershgorin discs, centered at the diagonal entries of F11 with
radius O(µ(N +1)/N
), is disjoint from the union of the other
M 2(N 2−1) Gershgorin discs, centered at the diagonal entries
of F22. Note that F has the same eigenvalues as the block
matrix on the right hand side of (163), and that eigenvalues
are invariant under a transposition operation. It follows from
(161) that
max
ρ (F) = ρ (B(cid:48)
11 ⊗ B(cid:48)
11) + O(µ(N +1)/N
max
Using the fact that ρ (B(cid:48)
the desired result (60).
11 ⊗ B(cid:48)
11) = [ρ (B(cid:48)
).
(166)
11)]2, we arrive at
APPENDIX C
EXAMINING THE DIFFERENCE IN (75)
We revisit the MSE networks discussed in (82). Assume
that there are only two agents in the network as shown in Fig.
7, namely, N = 2, with M = 2. Assume that σ2
v,2. For
v,1 > σ2
Fig. 7. A two-agent MSE network with a doubly-stochastic combination
matrix.
simplicity, uniform parameters {qk ≡ q} are used across the
agents (which may occur, for example, in the ATC or CTA
forms when the step-sizes are uniform across the agents, i.e.,
{µk ≡ µ}, and doubly-stochastic combination matrices are
adopted. In this case, we get {qk ≡ µ/N} [2, pp. 493–494]).
Let
(cid:20) π1 π1
(cid:21)
(cid:20) π2 π2
(cid:21)
1
1
π2
π1
Ru,1 =
, Ru,2 =
(167)
where the numbers π1 < 1,π2 < 1 ensure that Ru,1 >
0, Ru,2 > 0. Then, expression (75) can be rewritten as
MSDcoor,k − MSDgrad,k
(85)
(Ru,1 + Ru,2)
= rqTr
v,1 (diag{Ru,1} − Ru,1) +
(cid:16)
−1(cid:0)σ2
v,2 (diag{Ru,2} − Ru,2)(cid:1)(cid:17)
(cid:32)(cid:20) π1 + π2 π1 + π2
(167)
σ2
(cid:21)−1 ×
= rqTr
π1 + π2
2
−π1σ2
v,1 − π2σ2
v,2
0
−π1σ2
v,1 − π2σ2
v,2
2rq(π1 + π2)(π1σ2
v,1 + π2σ2
v,2)
2 (π1 + π2) − (π1 + π2)2
0
=
(168)
where the denominator is positive for all π1 < 1,π2 <
1. Then, MSDcoor,k < MSDgrad,k if, and only if (π1 +
(cid:21)(cid:33)
(cid:20)
1122(cid:1)(cid:1)(cid:1)(cid:1)1(cid:1)(cid:1)1 π2)(π1σ2
v,1 + π2σ2
v,2) < 0, which implies that
(cid:40)
0 < π2 < 1,−π2 < π1 < −(cid:0)σ2
−1 < π2 < 0,−(cid:0)σ2
(cid:1) π2
(cid:1) π2 < π1 < −π2.
v,2/σ2
v,1
v,2/σ2
v,1
Otherwise, MSDcoor,k ≥ MSDgrad,k.
APPENDIX D
PROOF OF COROLLARY 2
In the case when the matrices {Hk} or {Gk} are diagonal, it
follows from (75) and (76) that MSDcoor,k − MSDgrad,k = 0,
which verifies (80).
More generally, according to (2) and (46), we have
(cid:32) N(cid:88)
(cid:33)−1
N(cid:88)
qkHk
> 0,
kGk ≥ 0,
q2
kdiag{Gk} ≥ 0
q2
k=1
k=1
k=1
Then, applying the inequality [34]:
λmin(A)Tr(B) ≤ Tr(AB) ≤ λmax(A)Tr(B)
(170)
for any symmetric positive semi-definite matrices A and B,
where λmin(A) and λmax(A) represent respectively the largest
and smallest eigenvalues of A, we obtain
N(cid:88)
(cid:33)−1−
qkHk
MSDcoor,k−MSDgrad,k ≤ r
2
k=1
k=1
k=1
λmax
qkHk
(cid:32) N(cid:88)
(cid:32) N(cid:88)
(cid:40)
(cid:33)−1(cid:41) N(cid:88)
(cid:33)
(cid:32) N(cid:88)
≤ N(cid:88)
(cid:33) (a)≤ 1
(cid:46) N(cid:88)
(cid:32) N(cid:88)
(cid:46) N(cid:88)
(cid:33)−1 ≤
(cid:46)(cid:32)
qkλmax (Hk) ≤
qkHk
qkHk
k=1
k=1
k=1
k=1
λ
qkλmin (Hk)
(b)≤ 1
νd
N(cid:88)
k=1
1
0 <
qkλmin (Hk) ≤ λ
N(cid:88)
(cid:46)(cid:32)
k=1
we have
1
δd
qk
qkλmax (Hk)
(172)
N(cid:88)
(cid:33)
qk
(173)
k=1
k=1
where the inequalities (a) and (b) hold due to (2). Substituting
(173) into (171) gives the upper bound for the difference
MSDcoor,k−MSDgrad,k as shown by (79). Then, by following
a similar argument, we obtain the lower bound for the differ-
ence as the opposite number of the upper bound, which leads
to the desired result in Corollary 2.
λmin
q2
kTr(Gk)
(171)
θ − α =
where we substituted (76) into (75) and used the relation
Tr(Gk) = Tr (diag{Gk}). Then, noting that
=
16
APPENDIX E
PROOF OF COROLLARY 5
(169a)
(169b)
We start from the MSD expression in (95) and note first
(cid:18)
(cid:19)
k = (1 − rk)2
G(cid:48)
G +
rk
1 − rk
diag{G}
(174)
that
Substituting into (95) we have:
where (175) holds because
MSDcoor,k =
=
=
×
×
×
k=1
k=1
×
(cid:33)
(cid:33)
diag{G}
rk
1 − rk
qk(1 − rk)
k(1 − rk)2
(cid:19)(cid:19)
q2
(cid:33)−1(cid:32) N(cid:88)
(cid:33)−1(cid:32) N(cid:88)
k(1 − rk)2
(cid:32) N(cid:88)
(cid:33)−1
q2
(cid:33)
Tr(cid:0)H−1diag{G}(cid:1)
(cid:33)
(cid:33)−1(cid:32) N(cid:88)
(θ − α)Tr(cid:0)H−1diag{G}(cid:1)
(cid:80)N
(cid:80)N
k(1 − rk)2
k=1 q2
k=1 qk(1 − rk)
k(1 − rk)2
q2
k=1
k=1
k=1
k=1
k=1
Tr
1
2
1
2
1
2
G +
(cid:18)
H−1
qk(1 − rk)
qk(1 − rk)
k(1 − rk)rk
q2
(cid:32) N(cid:88)
(cid:18)
(cid:32) N(cid:88)
Tr(cid:0)H−1G(cid:1) +
(cid:32) N(cid:88)
(cid:32) N(cid:88)
Tr(cid:0)H−1G(cid:1) +
(cid:80)N
(cid:80)N
k(1 − rk)
k=1 q2
−
(cid:80)N
k=1 qk(1 − rk)
(cid:80)N
k(1 − rk)rk
k=1 q2
k=1 qk(1 − rk)
(cid:32) N(cid:88)
qk(1 − rk)
1
2
1
2
k=1
1
2
(cid:33)−1 N(cid:88)
k=1
k=1
(175)
(176)
(177)
q2
kTr(cid:0)H−1G(cid:1)
(cid:33)
with the numbers α and θ being defined in (97) and (98),
respectively. Recall that
MSDgrad,k =
qk
is the MSD performance for the full-gradient case. Thus,
MSDcoor,k − MSDgrad,k
(cid:32)(cid:80)N
(cid:80)N
(cid:80)N
(cid:80)N
k(1 − rk)2
k=1 q2
(θ − α)Tr(cid:0)H−1diag{G}(cid:1)
k=1 qk(1 − rk)
Tr(cid:0)H−1G(cid:1) +
Tr(cid:0)H−1G(cid:1)
(θ − α)Tr(cid:0)H−1diag{G}(cid:1) .
k=1 q2
k
k=1 qk
−
1
2
=
=
1
2
+
α
2
1
2
(178)
Applying (170) and (2) to (178), we obtain the desired results
for the MSD performance shown in Corollary 5. The result
for the ER performance in Corollary 5 can be shown by
subtracting the ER expression, ERgrad,k, on the both sides
of (96).
REFERENCES
[1] C. Wang, Y. Zhang, B. Ying, and A. H. Sayed, "Coordinate-descent
in Proc. EUSIPCO, Kos Island, Greece,
adaptation over networks,"
Aug. 2017.
[2] A. H. Sayed, "Adaptation, learning, and optimization over networks,"
Foundations and Trends in Machine Learning, vol. 7, no. 4-5, pp. 311–
801, 2014.
[3] A. H. Sayed, "Adaptive networks," Proceedings of the IEEE, vol. 102,
no. 4, pp. 460–497, Apr. 2014.
[4] S. Kar, J. M. F. Moura, and K. Ramanan, "Distributed parameter esti-
mation in sensor networks: Nonlinear observation models and imperfect
communication," IEEE Trans. Information Theory, vol. 58, no. 6, pp.
3575–3605, June 2012.
[5] A. Nedi´c and A. Ozdaglar, "Cooperative distributed multi-agent opti-
mization," in Convex Optimization in Signal Processing and Communi-
cations, D. P. Palomar and Y. C. Eldar, Eds., pp. 340–386. Cambridge
University Press, 2010.
[6] S. Kar and J. M. F. Moura, "Convergence rate analysis of distributed
gossip (linear parameter) estimation: Fundamental limits and tradeoffs,"
IEEE J. Sel. Top. Signal Process., vol. 5, no. 4, pp. 674–690, Aug. 2011.
[7] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione,
"Gossip algorithms for distributed signal processing," Proc. IEEE, vol.
98, no. 11, pp. 1847–1864, Nov. 2010.
[8] S. Sardellitti, M. Giona, and S. Barbarossa, "Fast distributed average
consensus algorithms based on advection-diffusion processes," IEEE
Trans. Signal Process., vol. 58, no. 2, pp. 826–842, Feb. 2010.
[9] P. Braca, S. Marano, and V. Matta, "Running consensus in wireless
sensor networks," in Proc. 11th International Conference on Information
Fusion, Cologne, Germany, June 2008, pp. 1–6.
[10] U. A. Khan and J. M. F. Moura, "Distributing the Kalman filter for
large-scale systems," IEEE Trans. Signal Process., vol. 56, no. 10, pp.
4919–4935, Oct. 2008.
[11] L. Xiao and S. Boyd, "Fast linear iterations for distributed averaging,"
Syst. Control Lett., vol. 53, no. 1, pp. 65–78, Sep. 2004.
[12] S.-Y. Tu and A. H. Sayed, "Diffusion strategies outperform consensus
IEEE
strategies for distributed estimation over adaptive networks,"
Trans. Signal Process., vol. 60, no. 12, pp. 6217–6234, Dec. 2012.
[13] Y. Nesterov, "Efficiency of coordinate descent methods on huge-scale
optimization problems," SIAM Journal on Optimization, vol. 22, no. 2,
pp. 341–362, 2012.
[14] P. Richt´arik and M. Tak´ac, "Iteration complexity of randomized block-
coordinate descent methods for minimizing a composite function,"
Mathematical Programming, vol. 144, no. 1-2, pp. 1–38, 2014.
[15] Z. Lu and L. Xiao, "On the complexity analysis of randomized block-
coordinate descent methods," Mathematical Programming, vol. 152, pp.
615–642, 2015.
[16] X. Zhao and A. H. Sayed, "Asynchronous adaptation and learning over
networks-Part I: Modeling and stability analysis," IEEE Trans. Signal
Process., vol. 63, no. 4, pp. 811–826, Feb. 2015.
[17] X. Zhao and A. H. Sayed, "Asynchronous adaptation and learning over
networks-Part II: Performance analysis," IEEE Trans. Signal Process.,
vol. 63, no. 4, pp. 827–842, Feb. 2015.
[18] X. Zhao and A. H. Sayed, "Asynchronous adaptation and learning over
networks-Part III: Comparison analysis," IEEE Trans. Signal Process.,
vol. 63, no. 4, pp. 843–858, Feb. 2015.
[19] R. Arablouei, S. Werner, Y.-F. Huang, and K. Dogancay, "Distributed
least mean-square estimation with partial diffusion," IEEE Trans. Signal
Process., vol. 62, no. 2, pp. 472–484, Jan. 2014.
[20] M. R. Gholami, E. G. Strom, and A. H. Sayed, "Diffusion estimation
over cooperative networks with missing data," in Proc. IEEE GlobalSIP,
Austin, TX, Dec. 2013, pp. 411–414.
[21] S. C. Douglas, "Adaptive filters employing partial updates," IEEE Trans.
Circuits Syst. II, vol. 44, no. 3, pp. 209–216, Mar. 1997.
[22] S. Werner, M. Mohammed, Y.-F. Huang, and V. Koivunen, "Decentral-
ized set-membership adaptive estimation for clustered sensor networks,"
in Proc. IEEE ICASSP, Las Vegas, NV, 2008, pp. 3573–3576.
[23] S. Werner and Y.-F. Huang, "Time- and coefficient- selective diffusion
in Proc. Asilomar
strategies for distributed parameter estimation,"
Conference on Signals, Systems, and Computers, Pacific Grove, CA,
2010, pp. 696–700.
[24] S. C. Douglas, "A family of normalized LMS algorithms," IEEE Signal
Process. Lett., vol. 1, no. 3, pp. 49–51, Mar. 1994.
[25] K. Dogancay, O. Tanrıkulu, "Adaptive filtering algorithms with selective
partial updates," IEEE Trans. Circuits Syst. II, vol. 48, no. 8, pp. 762–
769, Aug. 2001.
17
[26] M. Godavarti and A. O. Hero, "Partial update LMS algorithms," IEEE
Trans. Signal Process., vol. 53, no. 7, pp. 2382–2399, Jul. 2005.
[27] S. Chouvardas, K. Slavakis, and S. Theodoridis, "Trading off complexity
with communication costs in distributed adaptive learning via Krylov
subspaces for dimensionality reduction," IEEE Journal Selected Topics
Signal Process., vol. 7, no. 2, pp. 257–273, April 2013.
[28] S. Theodoridis, K. Slavakis, and I. Yamada, "Adaptive learning in a
world of projections: A unifying framework for linear and nonlinear
classification and regression tasks," IEEE Signal Process. Magazine,
vol. 28, no. 1, pp. 97–123, Jan. 2011.
[29] S. Chouvardas, K. Slavakis, Y. Kopsinis, and S. Theodoridis, "A sparsity
IEEE Trans.
promoting adaptive algorithm for distributed learning,"
Signal Process., vol. 60, no. 10, pp. 5412–5425, Oct. 2012.
[30] O. N. Gharehshiran, V. Krishnamurthy, and G. Yin, "Distributed energy-
aware diffusion least mean squares: Game-theoretic learning," IEEE J.
Sel. Top. Signal Process., vol. 7, no. 5, pp. 821–836, Oct. 2013.
[31] A. H. Sayed, Adaptive Filters, Wiley, 2008.
[32] G. H. Golub and C. F. Van Loan, Matrix Computations, The John
Hopkins University Press, Baltimore, 3 edition, 1996.
[33] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University
Press, 2003.
[34] Y. Fang, K. A. Loparo, and X. Feng, "Inequalities for the trace of
matrix product," IEEE Trans. Automatic Control, vol. 39, no. 12, pp.
2489–2490, 1994.
Chengcheng Wang (S'17) received the B.Eng. de-
gree in Electrical Engineering and Automation, and
Ph.D degree in Control Science and Engineering,
from the College of Automation, Harbin Engineer-
ing University, Harbin, China, in 2011 and 2017,
respectively. She is currently a Research Fellow in
the School of Electrical and Electronic Engineering,
Nanyang Technological University, Singapore. From
Sep. 2014 to Sep. 2016, she was a visiting graduate
researcher in Adaptive Systems Laboratory at the
University of California, Los Angeles, California,
USA. Her research interests include adaptive and statistical signal processing,
and distributed adaptation and learning.
Yonggang Zhang (S'06–M'07–SM'16) received the
B.S. and M.S. degrees from the College of Au-
tomation, Harbin Engineering University, Harbin,
China, in 2002 and 2004, respectively. He received
his Ph.D. degree in Electronic Engineering from
Cardiff University, UK in 2007 and worked as a
Post-Doctoral Fellow at Loughborough University,
UK from 2007 to 2008 in the area of adaptive signal
processing. Currently, he is a Professor of Con-
trol Science and Engineering in Harbin Engineering
University (HEU) in China. His current research
interests include signal processing, information fusion and their applications
in navigation technology, such as fiber optical gyroscope, inertial navigation
and integrated navigation.
Bicheng Ying (S'15) received his B.S. and M.S.
degrees from Shanghai Jiao Tong University (SJTU)
and University of California, Los Angeles (UCLA)
in 2013 and 2014, respectively. He is currently work-
ing towards the PhD degree in Electrical Engineering
at UCLA. His research interests include multi-agent
network processing, large-scale machine learning,
distributed optimization, and statistical signal pro-
cessing.
Ali H. Sayed (S'90–M'92–SM'99–F'01) is Dean
of Engineering at EPFL, Switzerland. He has also
served as distinguished professor and former chair-
man of electrical engineering at UCLA. An au-
thor of over 500 scholarly publications and six
books, his research involves several areas including
adaptation and learning, statistical signal processing,
distributed processing, network and data sciences,
and biologically-inspired designs. Dr. Sayed has re-
ceived several awards including the 2015 Education
Award from the IEEE Signal Processing Society, the
2014 Athanasios Papoulis Award from the European Association for Signal
Processing, the 2013 Meritorious Service Award, and the 2012 Technical
Achievement Award from the IEEE Signal Processing Society. Also, the 2005
Terman Award from the American Society for Engineering Education, the
2003 Kuwait Prize, and the 1996 IEEE Donald G. Fink Prize. He served as
Distinguished Lecturer for the IEEE Signal Processing Society in 2005 and as
Editor-in-Chief of the IEEE TRANSACTIONS ON SIGNAL PROCESSING
(2003–2005). His articles received several Best Paper Awards from the IEEE
Signal Processing Society (2002, 2005, 2012, 2014). He is a Fellow of
the American Association for the Advancement of Science (AAAS). He is
recognized as a Highly Cited Researcher by Thomson Reuters. He is serving
as President-Elect of the IEEE Signal Processing Society.
18
|
1901.09833 | 1 | 1901 | 2019-01-28T17:29:40 | The Emergence of Complex Bodyguard Behavior Through Multi-Agent Reinforcement Learning | [
"cs.MA"
] | In this paper we are considering a scenario where a team of robot bodyguards are providing physical protection to a VIP in a crowded public space. We show that the problem involves a complex mesh of interactions between the VIP and the robots, between the robots themselves and the robots and the bystanders respectively. We show how recently proposed multi-agent policy gradient reinforcement learning algorithms such as MADDPG can be successfully adapted to learn collaborative robot behaviors that provide protection to the VIP. | cs.MA | cs | The Emergence of Complex Bodyguard Behavior Through Multi-Agent
Reinforcement Learning
Hassam Ullah Sheikh and Ladislau Boloni
Department of Computer Science
University of Central Florida
[email protected], [email protected]
9
1
0
2
n
a
J
8
2
]
A
M
.
s
c
[
1
v
3
3
8
9
0
.
1
0
9
1
:
v
i
X
r
a
Abstract
In this paper we are considering a scenario where
a team of robot bodyguards are providing physical
protection to a VIP in a crowded public space. We
show that the problem involves a complex mesh of
interactions between the VIP and the robots, be-
tween the robots themselves and the robots and
the bystanders respectively. We show how re-
cently proposed multi-agent policy gradient rein-
forcement learning algorithms such as MADDPG
can be successfully adapted to learn collaborative
robot behaviors that provide protection to the VIP.
1 Introduction
One of the most complex challenges in autonomous robots
concerns behaviors that require a complex interaction be-
tween robots and humans, as well as between the robots
themselves. In particular, learning in these settings is chal-
lenging because the learner can not assume that the other
robots/agents it is interacting with follow a stable strategy. To
keep complexity manageable, most research work in this field
is tested on simple problems, such as the cooperative com-
munication or predator-prey, that involve the robot having to
interact with only one or two types of other robots/agents.
In this paper we are considering a more complex problem
with a practical application. A VIP is navigating a crowded
public area. We assume that the bystanders are purposefully
moving from landmark to landmark, and use culture-specific
crowd navigation protocols [Boloni et al., 2013] on giving
way or asserting their right of way. The goal of a team of
robot bodyguards is to physically protect the VIP from as-
sault by appropriately positioning themselves in relation to
the current position of the VIP.
The goal of the bodyguards can be stated as a minimization
of a metric of the threat to the VIP, a problem made challeng-
ing due to the complexity of the environment. The bodyguard
robot must adapt its position to the current position of the
VIP. Second, the movement must also depend on the move-
ment of the bystanders; if a bystander is heading towards the
VIP, the robot needs to position itself to reduce the threat to
the VIP. Finally, the robots need to coordinate between each
other. This can be achieved either through central planning,
explicit communication by exchanging messages or though
implicit communication, by each robot taking actions based
on a world view including the other robots.
While it is possible to hand-engineer robot bodyguard be-
haviors [Bhatia et al., 2016], it is of interest whether such
behavior can be learned through reinforcement learning.
2 Problem Formulation
The reward structure of the robotic bodyguard team problem
can be modeled as a cooperative Markov game where multi-
ple agents are learning a policy to maximize their rewards. A
multi-agent MDP can be defined as a combination of a state
space S, an action space A, agents A1, . . . ,AN , the transition
probabilities that are defined as T = S1×A1×. . .×SN×AN
and the reward function for every agent that is defined as
ri = R (Si,Ai).
For the sake of simplicity, we are assuming that all the
bodyguards have an identical state and action space. We are
considering a finite horizon problem where each episode is
terminated after T steps.
The environment
that provides the rewards is a two-
dimensional space, with the usual rules of physical move-
ment, landmarks, line-of-sight and communication. We used
the Multi-Agent Particle Environment MPE [Mordatch and
Abbeel, 2017] to perform our experiments. MPE is a two-
dimensional physically simulated environment that runs in
discrete time and continuous state and action space. The en-
vironment consists of N agents and M landmarks possessing
physical attributes such as location, velocity and mass etc.
MPE also allows the agents to communicate via verbal ut-
terances over the communication channel embedded in the
environment. The complete observation of the environment
is given by
o = [x1,...N +M , c1,...N ] ∈ O
The state of each agent is the physical state of all the entities
in the environment and verbal utterances of all the agents.
Formally, the state of each agent is defined as
si = [xj,...N +M , ck,...N ]
where xj is the observation of the entity j from the prespec-
tive of agent i and ck is the verbal utterance of the agent k.
2.1 Reward Function
[Bhatia et al., 2016] defined a metric that quantifies the threat
to the VIP from each crowd member bi at each timestep t. We
extended the metric to form a reward function defined as
Rt (B, V IP, xi) = − 1 +
(1 − T L (VIP, bi))
(1)
k(cid:89)
i=1
+ D (VIP, xi) + p
where
and
T L (VIP, bi) = exp−A(Dist(VIP,bi))/B
D (VIP, xi) =
−1
m ≤ (cid:107)xi − VIP(cid:107)2 ≤ d
otherwise
0
where m is the minimum distance the agents have to main-
tain at every timestep, d is the safe distance and p is a small
penalty to the bodyguard for utterance.
Figure 1: Screenshot showing the emerging collaborative bodyguard
behavior. The (brown) VIP aims to reach the green landmark. The
(blue) bodyguard robots surround and separate it from the closest
bystanders (red). The bystanders navigate among the gray land-
marks.
3 Methodology
The bodyguards were trained using multi-agent deep deter-
ministic policy gradients MADDPG [Lowe et al., 2017],
a multi-agent extension of the DDPG algorithm described
in [Lillicrap et al., 2015]. MADDPG allows agents to have
individual policies, but trains them with a centralized Q-
function. The gradient ∇J (θi) of each policy is written as
i (si, a1, . . . , aN )]
∇J (θi) = E [∇θi log πi (aisi) Qπ
where Qπ
i (si, a1, . . . , aN ) is a centralized action-value func-
tion that takes the actions of all the agents in addition to the
state of the agent to estimate the Q-value for agent i. The pri-
mary motivation behind MADDPG is that knowing all the ac-
tions of other agents makes the environment stationary, even
though their policy changes.
4 Experiments
The experiments were conducted in an environment created
in MPE, simulating a crowded mall with 12 landmarks repre-
senting stores. The VIP is following its own policy of moving
towards a goal landmark. 10 bystanders are following their
own policy that involve sequentially visiting randomly cho-
sen landmarks. 3 bodyguards were deployed to protect the
VIP from physical assault.
For training,
the bodyguard agents were trained over
10,000 episodes, each episode being limited to 25 steps using
DDPG and MADDPG. After training, performance was mea-
sures both empirically (by means of observing the movement
of the bodyguards) and by recording the cummulative threat
over the episode as defined in [Bhatia et al., 2016]. An exam-
ple of the achieved behavior is described in Figure 1. As the
screenshot shows, the robots learned to surround and move
with the VIP, concentrating on the sides where bystanders are
closer. Empirical evaluation had shown better performance
with MADDPG compared to DDPG.
5 Conclusion
In this paper, we outlined a technique for training a multi
robot team of bodyguards that is collaborating towards a com-
mon goal of providing physical security to a VIP moving in a
crowded environment. We trained the robot behaviors using
MADDPG RL, and we observed the emergence of recogniz-
able collaborative bodyguard behavior without any explicit
instruction on how to provide security.
References
[Bhatia et al., 2016] T.S. Bhatia, G. Solmaz, D. Turgut, and
L. Boloni. Controlling the movement of robotic body-
In Proc of the
guards for maximal physical protection.
29th International FLAIRS Conference, pages 380 -- 385,
May 2016.
[Boloni et al., 2013] Ladislau Boloni, S Khan, and Saad
Arif. Robots in crowds-being useful while staying out of
trouble. In Proc. of Intelligent Robotic Systems Workshop
(IRS-2013) at AAAI, pages 2 -- 7, 2013.
[Lillicrap et al., 2015] Timothy P. Lillicrap,
Jonathan J.
Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval
Tassa, David Silver, and Daan Wierstra. Continuous con-
trol with deep reinforcement learning. In Proceedings of
the International Conference on Learning Representations
(ICLR), 2015.
[Lowe et al., 2017] Ryan Lowe, Yi Wu, Aviv Tamar, Jean
Harb, Pieter Abbeel, and Igor Mordatch. Multi-agent
actor-critic for mixed cooperative-competitive environ-
ments. Neural Information Processing Systems (NIPS),
2017.
[Mordatch and Abbeel, 2017] Igor Mordatch and Pieter
lan-
arXiv preprint
Abbeel.
guage in multi-agent populations.
arXiv:1703.04908, 2017.
Emergence of grounded compositional
|
1709.07658 | 1 | 1709 | 2017-09-22T09:45:40 | Potentials and Implications of Dedicated Highway Lanes for Autonomous Vehicles | [
"cs.MA"
] | The introduction of autonomous vehicles (AVs) will have far-reaching effects on road traffic in cities and on highways.The implementation of automated highway system (AHS), possibly with a dedicated lane only for AVs, is believed to be a requirement to maximise the benefit from the advantages of AVs. We study the ramifications of an increasing percentage of AVs on the traffic system with and without the introduction of a dedicated AV lane on highways. We conduct an analytical evaluation of a simplified scenario and a macroscopic simulation of the city of Singapore under user equilibrium conditions with a realistic traffic demand. We present findings regarding average travel time, fuel consumption, throughput and road usage. Instead of only considering the highways, we also focus on the effects on the remaining road network. Our results show a reduction of average travel time and fuel consumption as a result of increasing the portion of AVs in the system. We show that the introduction of an AV lane is not beneficial in terms of average commute time. Examining the effects of the AV population only, however, the AV lane provides a considerable reduction of travel time (approx. 25%) at the price of delaying conventional vehicles (approx. 7%). Furthermore a notable shift of travel demand away from the highways towards major and small roads is noticed in early stages of AV penetration of the system. Finally, our findings show that after a certain threshold percentage of AVs the differences between AV and no AV lane scenarios become negligible. | cs.MA | cs |
Potentials and Implications of Dedicated Highway Lanes for
Autonomous Vehicles
Jordan Ivanchev1, Alois Knoll2, Daniel Zehe1, Suraj Nair1, and David Eckhoff1
2Department of Informatics, Technische Universität München, Germany
1TUMCREATE, Singapore
September 25, 2017
Abstract
The introduction of Autonomous Vehicles (AVs) will have far-reaching effects on road traffic in cities
and on highways.The implementation of Automated Highway Systems (AHS), possibly with a dedicated
lane only for AVs, is believed to be a requirement to maximise the benefit from the advantages of AVs.
We study the ramifications of an increasing percentage of AVs on the traffic system with and without the
introduction of a dedicated AV lane on highways. We conduct an analytical evaluation of a simplified
scenario and a macroscopic simulation of the city of Singapore under user equilibrium conditions with a
realistic traffic demand. We present findings regarding average travel time, fuel consumption, throughput
and road usage. Instead of only considering the highways, we also focus on the effects on the remaining
road network. Our results show a reduction of average travel time and fuel consumption as a result of
increasing the portion of AVs in the system. We show that the introduction of an AV lane is not beneficial
in terms of average commute time. Examining the effects of the AV population only, however, the AV
lane provides a considerable reduction of travel time (≈ 25%) at the price of delaying conventional vehicles
(≈ 7%). Furthermore a notable shift of travel demand away from the highways towards major and small
roads is noticed in early stages of AV penetration of the system. Finally, our findings show that after a
certain threshold percentage of AVs the differences between AV and no AV lane scenarios become negligible.
Introduction
1
Autonomous vehicles have ceased to be only a vision but are rapidly becoming a reality as cities around the world
such as Pittsburgh, San Francisco, and Singapore have begun investigating and testing autonomous mobility
concepts [1]. The planned introduction of thousands of autonomous taxis, as currently planned in Singapore [1],
poses a challenge not only to the in-car systems of the AV, but also to the traffic system itself.
Besides the deployment of more efficient ride sharing systems and the reduction of the total number of
vehicles on the road, AVs can traverse a road faster while using less space. To achieve the maximum benefit
in terms of traffic speeds and congestion reduction, however, mixing of AVs and Conventional Vehicles (CVs)
should be avoided [2]. One method to achieve this is the introduction of dedicated AV lanes on highways to
allow AVs to operate more efficiently due to the absence of unpredictable random behaviour introduced by
humans and the use of communication capabilities, e.g. platoon organization [3].
Blocking certain lanes for CVs, and thereby limiting the overall road capacity for human drivers is certainly
a step that can have considerable ramifications, depending on the portion of AVs in the system. While at the
early stages, where only few AVs are on the road, it would constitute an incentive to obtain an AV, it could
also possibly generate traffic congestion and increase travel times for other vehicles
[4]. As the level of AV
penetration in the road transportation system is increased, the total congestion level would likely drop, however,
the advantage of using AVs over CVs would gradually be diminished as well.
Lastly, converting highways into AHS could also affect the rest of the road network, as drivers of CVs may
then choose a different route, caused by the changed capacity of the highway, which can lead to a mismatch
between road network and traffic demand [5].
In this paper, we take a closer look at the benefits and drawbacks of introducing a dedicated AV lane on
major highways Singapore. Focusing particularly on the effect of varying percentages of AVs in the system, we
study the differences in terms of capacity, travel time, fuel consumption, and impact on other roads compared
to a setting where no dedicated AV lane is assigned. In short, our two main contributions are:
• We present an analytical evaluation of the expected benefit from the introduction of dedicated AV lanes.
1
• Using a macroscopic traffic simulation of the city-state of Singapore and based on realistic travel demand,
we show the impact of vehicle automation with and without AV lanes.
The remainder of this paper is organized as follows: In Section 2 we discuss related work on AHS. Section 3
explains the system model used in this study. In Section 4 we present our analytical evaluation of a synthetic
scenario, followed by Section 5 in which results for the entire city of Singapore are presented and discussed.
Section 6 concludes our article.
2 Related Work
AHS and their implications have received wide attention from both researchers and industry around the world.
Investigations include general AHS policies and concepts [2,6], effects on travel times and capacity [6–13], traffic
safety [6, 8, 13, 14], and interactions between conventional human-driven vehicles and autonomous vehicles [15].
Several AHS studies and field trials were conducted as part of the California Partners for Advanced Transit
and Highways (PATH) program: Tsao et al. discuss the relationship between lane changing manoeuvres and
the overall throughput of the highway [9]. Their analytical and simulation results indicate a direct trade-off
between the two. Lateral movement decreases the traffic flow, and higher traffic flow leads to longer lane change
times. In this work, we ignore decreased throughput caused by lane changing and focus on a best-case city-wide
benefits of dedicated AV lanes. Godbole and Lygeros evaluate the increased capacity by the introduction of fully
automated highways by treating the highway as a single-lane AHS pipe [14]. Similar to the work of Harwood
and Reed [10], they study different platoon sizes and speeds but do not consider separated AV and CV lanes.
AHS pipeline capacity was also studied by Michael et al. [11]. Their results show that longer platoons of AVs
are favourable as they increase the capacity of the road due to lower inter-platoon distances. A high mixture of
different vehicle classes, however, leads to a lower capacity. Lastly, the presented analytical model shows that
as highway speed increases, the capacity reaches a saturation point after which the speed decreases again. This
is inline with the findings presented in this paper when studying the throughput on only the AV lane.
Similar to the work presented in this article, Cohen and Princeton investigated the impact of dedicated
lanes on the road capacity [7]. They find that assigning a dedicated lane can create bottlenecks due to increased
lane changing manoeuvres of departing unauthorized vehicles. When the capacity of the remaining lanes is
exceeded, it may even be impossible for allowed vehicles to access the dedicated lane.
In our macroscopic
study, we abstract away from these problems to answer the more general question whether the introduction of
dedicated AV lanes is a feasible approach.
In summary, it appears that while AHS seem to be a well-studied subject, a general evaluation of the
introduction of AV lanes is still missing. Furthermore, automated highways are mostly investigated in an
isolated manner (and often also simplified to a 1-lane road) without taking into considering the rest of the road
network. Using the city-state of Singapore as a case study, we show how travel times of both AVs and CVs are
affected and that the introduction of dedicated AV lanes has a considerable effect on the entire road network
by changing the distribution of traffic demand within the transportation system.
3 System Model
The goal of this study is to evaluate the allocation of one lane on every highway road to be used exclusively
by autonomous vehicles. We investigate an increasing percentage of AVs in the system under two scenarios:
with and without dedicated AV lanes on highways. In the former scenario, all roads that are not highways
will exhibit normal traffic conditions as there will be a mixture of human drivers and AVs. All lanes on the
highway roads will be accessible to AVs, while CVs will be able to utilize all lanes except one lane, which will
be allocated exclusively for AVs usage.
In the second scenario, all vehicles will share all lanes on all roads.
We model the different behaviour of AVs and CVs by means of smaller headway time, that is the time gap
to the vehicle in front. We assume that AVs can afford a much smaller headway than normal vehicles since
their reaction time is orders of magnitude smaller than that of humans [16]. A direct consequence of this is
that effectively the capacity of the road is increased, as AVs need less space. The capacity (in cars per hour per
lane) can then be calculated as follows:
C =
3600
havpav + hcv (1 − pav)
(1)
where hav is the headway time for AVs (values may vary between 0.5 and 1 second, depending on level
of comfort [13]), pav is the percentage of AVs on the road segment, hcv is the headway time for conventional
vehicles set to 1.8 seconds. Equation 1 is based on [17], where it was derived from collected data on highway
roads in Japan for varying percentages of vehicles with AHS.
2
Figure 1 illustrates the change of capacity as a function of the AV percentage for different AV headway
values. It can be observed that the capacity exponentially increases with the portion of AVs on the road. This
means that when there is a small portion of AVs on a road, their impact is only marginal, however, when the
majority of vehicles are AVs there is a significant increase of the capacity of the road. The capacity for the same
AV percentage also increases exponentially as the headway time decreases. For the remainder of the paper we
will assume a conservative AV headway of hav = 1s.
Figure 1: Changes of road capacity as a function of AV percentage and headway time
The primary measure we use to evaluate the impact on traffic caused by the introduction of an AV lane is
i ti of a vehicle is determined by the traverse times of all n
segments (or links) included in its route. The traverse time ti of a segment i can be computed using the Bureau
of Public Roads (BPR) function [18]:
the travel time of cars. The travel time T = (cid:80)n
(cid:32)
(cid:19)βi(cid:33)
(cid:18) Fi
Ciwit
ti =
1 + αi
li
vi
(2)
(3)
(4)
(5)
where li is length of the road segment, vi is the free flow velocity of the segment, Fi is flow, wi is the number of
lanes, t is time duration of the simulated period, Ci is the capacity of road segment i, αi and βi are parameters
from the BPR function. Free flow velocities v are extracted from historical GPS tracking data [19]. Parameters
αi and βi are calibrated for different classes of roads depending on their speed limits using both GPS tracking
data and a travel time distribution of the population for certain periods of the day. For a more detailed
description of the calibration and validation procedures we refer the reader to [20].
4 Analytical Evaluation
In this section, we will illustrate how to analytically determine the travel times of vehicles on highways in both
cases with and without dedicated AV lanes. We examine a simplistic scenario, where the system consists of a
single road with two lanes. Let the overall percentage of AVs be p and the percentage of AVs on the first and
second lane respectively be p1 and p2. Let the number of vehicles on the respective lanes be f1 and f2 such
that:
where F is the total number of vehicles in the system. Therefore, we know that:
f1 + f2 = F
Under the assumption of User Equilibrium (UE) [21] we also know that:
p1f1 + p2f2 = pF
t(f1) = t(f2)
where t(f ) represent the travel time on the road for flow f. UE is achieved when given the current traffic
situation, a user would not change their route, i.e. the user is already on the shortest route in terms of travel
time.
3
00.20.40.60.81Portion of AVs2000400060008000Lane Capacity [vehicles/hour]Headway=1sHeadway=0.75sHeadway=0.5s Substituting t(f ) for the BPR function (Equation 2) we get:
(cid:32)
(cid:19)β(cid:33)
(cid:18) f
C
t(f ) =
l
v
1 + α
Where α and β are parameters of the BPR function, v and l are the free flow velocity, and the length of the
road and C is the capacity of the road. Substituting the expression for capacity from Equation 1 we get for the
BPR function:
(cid:18) f (havpav + hcv (1 − pav))
(cid:19)β(cid:33)
(cid:32)
t(f ) =
l
v
1 + α
3600
Expanding Equation 5 we get:
(cid:32)
l
v
1 + α
(cid:18) f1 (havp1 + hcv (1 − p1))
3600
(cid:19)β(cid:33)
(cid:32)
1 + α
=
l
v
(cid:18) f2 (havp2 + hcv (1 − p2))
(cid:19)β(cid:33)
3600
(6)
(7)
Removing common terms, simplifying and using Equations 3 and 4 we get:
f1 (havp1 + hcv (1 − p1)) = (hav (pF − p1f1) + hcv (F − f1 − (pF − p1f1)))
Let's examine the two scenarios in this simplistic example: In scenario 1 a lane is exclusively reserved for
AVs (p1 = 1) and in scenario 2 all cars are allowed to use all lanes and therefore get distributed equally on the
two lanes (f1 = f2 = F
2 and p1 = p2 = p). For scenario 1 we get:
After simplifying we obtain:
f1hav = hav (pF − f1) + hcvF (1 − p)
f1hav =
F
2
(havp + hcv (1 − p))
(8)
(9)
Under the assumption of user equilibrium it follows that for scenario 1 the travel time for both lanes is the
same:
(cid:19)β(cid:33)
(cid:18) f1hav
3600
t1 =
1 + α
l
v
(cid:32)
F
2
1 + α
t2 =
l
v
β
(10)
(11)
(havp + hcv (1 − p))
3600
The travel time for a single vehicle in scenario 2 is:
It can be observed that because of Equation 9 the expression in Equation 10 is equal to the expression in
Equation 11. This means that in the simple examined case, regardless whether a dedicated AV lane is introduced
or not, the resulting travel time for all commuters is going to be the same after the point where there are enough
AVs to ensure equilibrium can be reached.
The number of vehicles on the first lane in scenario 1 is bounded from above by the total number of
autonomous vehicles: f1 ≤ pF . In the case where f1 computed according to Equation 9 is smaller than pF
equilibrium cannot be achieved. It is trivial to demonstrate that the overall travel time under scenario 1 is
bigger than the travel time under scenario 2 when f1 ≤ pF . The minimum number of AVs in terms of the AV
percentage pτ (the threshold percentage) in order for the solutions of the two scenarios to coincide is:
pτ F hav =
pτ =
F
2
(havpτ + hcv (1 − pτ ))
hcv
hav + hcv
(12)
(13)
The percentage pτ describes the point where for an additional AV it would be equally attractive to use the
AV lane or the normal lane. This is the moment where the AV lane reaches saturation. Before that percentage
is reached, the AVs have a shorter commuting time than the CVs, however, the system is performing worse
4
in terms of overall congestion compared to the case without a dedicated AV lane. Equation 13 can be easily
extended for an arbitrary number of lanes in our minimalistic example. Assume we have 1 dedicated AV lane
and N normal lanes. Then the percentage of AVs after which the dedicated lane is saturated becomes:
Assuming headway times of hcv = 1.8s and hav = 1s, the percentage pτ of AVs on the road for saturation
of the AV lane is 64.3%, 47.4%, 37.5%, 31%, 26.5% for respectively 2, 3, 4, 5, 6 lane roads with one allocated
AV lane.
pτ =
hcv
N hav + hcv
(14)
5 City-wide Simulation Study
We conduct a macroscopic city-wide simulation study of the city-state of Singapore to better understand the
effects to be expected in a complex environment. We take a look at the traffic conditions on the highway, but
also closely investigate the effect dedicated AV lanes have on the rest of the road network.
5.1 Methodology
In order to evaluate the scenarios of AV introduction in a road transportation system, we make use of an agent-
based macroscopic simulation approach. The simulation consists of three steps: 1) Agent generation, 2) route
computation and 3) travel time estimation.
The underlying road network on which the traffic assignment is performed is modelled by means of a uni-
directional graph, where each edge represents a road segment and nodes represent decision points at which a
road may split or merge.
To introduce dedicated AV lanes, we alter the original graph by duplicating start and stop nodes of highway
segments and creating a new edge to represent the AV lane, while removing one of the normal lanes from
the original segment. Furthermore, costless connections are added between the start and end nodes and their
respective copies so that vehicles can change lanes at the start/end of highway road segments.
Every agent is generated with an origin and destination sampled from an OD matrix representing the travel
demand of the system. Realistic traffic volume is modelled by synthesizing a sufficiently large vehicle population
according to the available survey data.
The routes of the agents are computed using an incremental user equilibrium approach [21] aiming at
representing reality in the sense that every driver is satisfied with their route and would not choose a different
one given the current traffic situation. We assume a driver is satisfied when they are on the shortest route from
origin to destination in terms of travel time. Routing is performed on the road network graph, where each edge
has an attached weight representing the current traverse time of this road segment. Weights are updated after
every batch route computation. To disallow conventional vehicles from using the AV lane, the weight of the
edges representing these lanes are set to infinity when the route of a CV is computed and set back to their
traverse time values for AV route computation.
Further assumptions regarding to the macroscopic level of the simulation include:
• Agents want to minimize travel time and all perceive the traffic situation in the same way (there is no
noise in the observations of the traffic states).
• Traffic is spread homogeneously in time during the simulation. In order for the BPR function (Equation
2) to give a reasonable estimation of the traffic conditions, the flow F during the simulation time t needs
to be spread homogeneously in time. If many vehicles try to enter the road at the same time for example,
the traffic congestion generated will likely produce a much higher traverse time for the road than the
expected one.
In order to ensure as much as possible that traffic demand stays static throughout the
simulation period, both the period t and the time of day, which is simulated, have to be chosen with care.
• Agents do not reroute while on their trip.
In reality drivers may change their trip plan dynamically
according to unexpected events or observed traffic conditions ahead. As our traffic assignment provides
information to the drivers about expected traffic conditions for the whole network, we assume that such
events will not occur as drivers are given all the information they need prior to their trip.
5.2 Data and Scenario Description
We examine the city-state of Singapore with a total population of 5.4 million and around 1 million registered
vehicles including taxis, delivery vans and public transportation vehicles. The fact that Singapore is situated
on an island simplifies our scenario as the examined system is relatively closed with only two expressways
5
leading out of the country. There are 3495 km or roads of which 652 km are major arterial roads and 161
km are expressways (these are the roads on which we introduce the dedicated AV lane indicated on Figure 5a)
spreading over 715 squared km of land area. We have used publicly available data to acquire a unidirectional
graph of Singapore, that comprises of 240, 000 links and 160, 000 nodes representing the road network of the
city. The number of lanes, speed limit and length of every link is available allowing us to extract information
about its capacity.
For the purposes of our model we make use of two separate data sets. The first data set consists of GPS tra-
jectories of a 20, 000 vehicle fleet for the duration of one month, providing information about recorded velocities
on the road network during different times of the day [19]. The second one is the Household Interview Travel
Survey (HITS) conducted in 2012 in the city of Singapore, which studies the traffic habits of the population.
Information about the origin destination pairs, their temporal nature, and commuting time distribution during
rush hour periods is extracted from it.
In order to achieve realistic traffic conditions, we estimated the number of agents based on the Singapore
HITS data. We extracted the ratio of people who actively create traffic on the streets (cab drivers, personal
vehicle drivers, lorry drivers) and the total number of people interviewed.
We chose the morning commute hours (7:30am - 8:30am) as the period which seems to be the most stable
in terms of traffic volume and estimated the traffic demand consisting of 309, 000 agents.
Similar to the simplistic example the vehicle population is split into two parts: AV and CV. The percentage
of AV is varied in order to observe the effects in the initial stages of AV introduction, as well as the possible
traffic situation if all vehicles were autonomous.
In order to have a benchmark for measuring the efficiency of the suggested policy, we also evaluate the
scenario where no AV lane is introduced and all vehicles can access all lanes on all roads. The analytical
solution acquired for the simplistic example in the previous section indicates that the scenario with no AV lane
will produce better or at least the same system performance as the introduction of the AV lane. Our case study
will test our analytical results for a more realistic transportation system environment.
5.3 Effects on Average Travel Time
We evaluated the change in average commute time based on the percentage of AVs in the system. We compare
two different settings: with and without a dedicated AV lane. As a baseline we use the average commute time
without the existence of AVs (and thus no exclusive lanes). This time was found to be approx. 18.5 minutes
and can be seen as the status quo.
We simulated both settings with the exact same travel demand, that is, identical origin-destination pairs for
all vehicles. We then gradually increased the percentage of AVs in the system. The number of vehicles in the
simulation was invariant; a higher percentage of AVs means that CVs were replaced with autonomous vehicles.
Figure 2 shows our result.
(a) Travel times with dedicated AV lane
(b) Comparison for all vehicles
Figure 2: Average travel time of whole population, AVs and CVs as a function of AV percentage
The black dashed line serves as an illustration of the status quo. It can be observed that when introducing
a lane exclusively for AVs and thereby taking away capacity from conventional vehicles (Figure 2a), the travel
time for CVs increases, whereas the low percentage of AVs (driving on almost empty lane) travel considerably
faster. This could be used as an incentive by policy-makers to increase the share of AVs in the transportation
system. With an increasing percentage of AVs we observe that travel times for CVs decrease due to the fact
that effectively vehicles move away from the common lane to the AV lane, which in turn increases the travel
time on the AV lane. This behaviour can be observed until the AV lane is saturated (somewhere between 40%
and 50% AV percentage). From this point on, the choice of lane makes no difference for newly added AVs
6
00.20.40.60.81Portion of AVs14151617181920Average Travel Time [mins]AVsCVsOriginal Scenario00.20.40.60.81Portion of AVs14151617181920Average Travel Time [mins]With AV LaneWithout AV LaneOriginal Scenario as they experience equal travel times regardless whether they choose to take the AV lane or the mixed lane.
Therefore, the difference between average travel time of AV and CV decreases. Travel time of both AVs and CVs
still decreases with the increase of the percentage of AVs since the capacity of the road network is effectively
increased.
The average number of lanes in the Singapore road network is 3.14.
If we apply Equation 14 from our
analytical evaluation, we yield a saturation point of 45.2% AV penetration, which seems to be in agreement
with our simulation results.
We conclude that computing the threshold percentage is useful in terms of AV lane introduction planning.
Figure 2b shows the comparison between the average travel times for the entire vehicle population with
and without the introduction of an AV lane. It can be observed that the setting without an AV lane is always
performing better than the AV lane one. After saturation of the AV lane the difference between the two curves
becomes marginal. This finding is in complete agreement with the results of our analysis in Section 4. Before
saturation of the AV lane, the capacity of the highway is not fully utilized and that, on average, the smaller
travel times of the AVs cannot make up for the introduced delays for the CVs. We conclude that adding an AV
lane, while initially being an incentive for early adopters, will noticeably penalize drivers of conventional vehicles
at the early stages of AV introduction. Additional delays will be introduced by lane changing manoeuvres and
other microscopic effects not considered in this macroscopic study [7, 9].
5.4 Analysis of Effect of Headway Time
The headway time of vehicles is a crucial parameter for the computation of the road capacity (see Equation 1)
and therefore an important input for the travel time computation using the BPR function (see Equation 2).
Although, AVs can afford to have smaller headways, this might negatively affect the comfort of the passengers,
as small distances at high velocities can induce anxiety [13]. To this end, we examine the effect of varying the
headway time (from 0.5s to 1s) for the city-scale simulation scenario with a dedicated AV lane on highways. It
is therefore useful to evaluate the amount of overall travel time decrease as a function of the headway time.
Figure 3: Average travel time curves for varying headway time
In Figure 3 we observed that, depending on the headway, the improvement of overall travel time can vary
between 20% and 26%. The difference between the improvement in the case of all AVs seems is bigger between
headway 0.5s and 0.75s than between 1s and 0.75s.
This is due to the non-linear nature of the BPR function. As vehicles approach free flow velocities due to
the increased capacity, any further improvement in capacity plays a smaller role and therefore the gains become
less significant. It must be noted, that this happens only when traffic congestion is such that vehicles travel at
free flow velocities.
If the system was more congested, we would not be able to observe the decrease of travel time improvement
for smaller headway time, since the BPR function will still be in its non-linear part thus providing significant
changes of travel time as the capacity is altered.
5.5 Effects on Fuel Consumption
We have demonstrated that the introduction of AVs decreases the overall travel time in the system. The next
question we are trying to answer is by how much this saved time reduces fuel consumption (or energy in the
7
00.20.40.60.81Portion of AVs1314151617181920Average Travel Time [mins]Headway=0.5sHeadway=0.75sHeadway=1sOriginal Scenario case of electric vehicles). We deploy the Elemental fuel consumption model [22] to evaluate the average fuel
consumption of the vehicle population and subgroups.
The effects of introducing an AV lane on fuel consumption are shown in Figure 4
Figure 4: Average fuel consumption of whole population, AVs and regular vehicles as a function of AV percentage
As expected, we observe a positive effect on fuel consumption. To provide an idea of magnitude, the fuel
saved at 20% AVs is about 9.3 tons for the duration of one Singapore morning rush hour and approx. 45 tons
when every vehicle is autonomous. As for the travel time, the combination of one dedicated lane and a low
percentage of AVs penalizes CVs.
The introduction of more AVs, however, decreases both the overall travel time and fuel consumption. It can
also be noted that the relative effect of the fuel consumption is smaller than the effect on travel time, which
means that fuel consumption is less sensitive with respect to changes in the traffic states for the examined
scenario. Please note that we do not consider fuel or energy saving resulting from platoon organization [12] and
other AV related technology but only seek to provide a order of magnitude with our macroscopic evaluation.
5.6 Effects on Road Network Throughput and Traffic Distribution
Lastly, we examine the traffic distribution on the entire road network to gain a better understanding of the
changes occurring in traffic conditions as a results of the introduction of dedicated AV lanes. To the best our
knowledge, existing literature focuses on the traffic changes on the highways and ramps only, and such a study
has not been conducted before. Figure 5a shows the Singaporean road network with highways drawn in bold.
AV lanes are added only to these highways, the rest of the network remained unchanged.
As a first step, we provide a qualitative measurement by visualizing the impact an introduction of 50% AVs
has on the road network (compared to 0% AVs). To this end, we draw roads experiencing higher throughput
in blue colours and roads experiencing lower throughput in red colours. We show our results for both settings,
i.e., with AV lane (Figure 5b) and without AV lane (Figure 5c).
It can be observed that in both cases the highways exhibit a considerably higher throughput of vehicles, which
is expected since the capacity of the highway is technically increased by the introduction of the autonomous
vehicles. Other roads exhibit a slight decrease of throughput, which may be due to the changes in routing
that are triggered by the AVs, which have a strong preference towards the highways. In other words, the AVs
are, in a way, attracted to the highways, since they can traverse fast there and the capacity is sufficient. This,
however, makes them willing to pass through more congested roads in order to get to the highway, which creates
additional time losses for the regular vehicles. This argument is further strengthened by the fact that the roads
leading to the highways do not exhibit increased throughput, although they exhibit higher demand since more
vehicles want to use the highways. This means that the level of congestion on those roads is too high to allow
an increase of throughput.
Comparing the throughput changes in the two scenarios, it can be observed that the increase of throughput
on the highways for the case with no AV lane is smaller. Therefore, the change of routing triggered by the
introduction of AVs is qualitatively the same but quantitatively different for the two examined scenarios. This
can be observed in Fig. 6. The relative increase of throughput for highways is almost three times higher for the
dedicated AV lane case. Higher level of throughput increase on highways leads to higher level of throughput
decrease on the major roads, which represent alternative routes. As mentioned before, if highways become too
attractive for AVs, there can be negative effects on traffic conditions stemming from overly utilized roads which
lead to the highways. The more balanced distribution of traffic in the no AV lane scenario could be the reason
8
00.20.40.60.81Portion of AVs11.051.11.151.21.251.3Average Fuel Consumption [litres]Whole populationAVsCVsOriginal Scenario (a) Singapore road network with highways featuring dedicated AV
lanes in bold
(b) Change of road throughput at 50% AVs (with AV lane)
(c) Change of road throughput at 50% AVs (without AV lane)
Figure 5: Road throughput change caused by 50% AVs compared to 0% AVs. Colours from the blue gamma
represent higher throughput; colours in the red gamma represent lower throughput
for its slight, however, consistent superiority over the dedicated AV lane case in terms of average travel time
observed in Figure 2b.
Finally, we investigate the change in demand for different types of roads. We define demand for a road as
the number of vehicles that have this particular road in their route under user equilibrium traffic assignment.
Taking a closer look at the difference in travel demand between the scenario with AV lane and the scenario
without the introduction of AV lane, we measure the relative difference of travel demand between the scenarios.
Formally, let the demand for road i for a given AV percentage k be Da
i,k for
the benchmark scenario without an AV lane. Then, for the different classes of road (highways, major roads,
other roads), we compute the difference in demand between the two examined scenarios relative to the AV lane
scenario (a negative value therefore means that this particular type of road experiences less demand when AV
i,k for the AV lane scenario and Db
9
Figure 6: Change of road throughput for different road types
lanes are introduced compared to not introducing AV lanes):
(cid:80)
Da
∆D,k =
i
i,k −(cid:80)
i
Da
i,k
Db
i,k
(15)
Figure 7 shows our results for highways, major roads, and other roads. For highways, we observe that
initially there is lower demand. This is due to the lower road capacity resulting from the dedicated lane, causing
more CVs to avoid the highways. With more AVs in the system, the demand for the highways increases until
the saturation point of the AV lane is reached. From this point on, the difference between the both settings
becomes smaller, eventually converging to zero as the addition of more AVs cause the common lanes to achieve
capacity values close to the one of the AV lane.
For the major roads we observe increased demand as CVs will favour these roads as an alternative to the
more congested highways caused by the dedicated AV lane. This effect decreases until the saturation point of
the AV lane.
At this moment the highways reach their maximum demand difference and therefore a smaller portion of
the population takes the alternative routes utilizing major roads. Following the negative peak, the demand
difference for major roads also converges to zero as the two scenarios become identical.
The difference for smaller roads is less pronounced and can only be observed at a low percentage of AVs in
the system. The higher utilization of major roads will also increase traverse time of these roads. Some vehicles
will therefore choose to use minor roads as a third alternative.
It is interesting to note that after the saturation point of the AV lane the difference in travel times between
the two scenarios is almost negligible, however, the actual assignment of traffic, as can be observed in Figure 7,
is qualitatively different. This finding indicates that the introduction of an AV lane would not just affect the
travel time of the population but also shift the route preferences of commuters.
Figure 7: Demand difference between AV lane setting and no AV lane setting for varying percentage of AVs
10
HighwaysMajor RoadsOther Roads-10-5051015Relative Change of Throughput [%] AV lane no AV lane00.10.20.30.40.50.60.70.80.91Portion of AVs-30-25-20-15-10-50510Relative Demand Difference "D [%]HighwaysMajor RoadsOther Roads 6 Conclusion and Future Work
In this article we demonstrated the effect of assigning one lane on highways exclusively for AVs. We showed
that for lower percentages of AVs, or more precisely, before the dedicated lane is saturated, travel times for AVs
can be significantly shorter, while at the same time CVs are delayed due to the reduced capacity of the highway.
Looking at the entire road network, we observe that also non-highways are affected as CVs will effectively
be drawn away from the highways onto the major roads. This effect is especially pronounced at early stages
of AV adaptation where the AV lane will remain mostly empty. Regardless of an introduction of the AV lane,
we confirmed earlier findings that a larger number of AVs will have a positive impact on travel times and fuel
consumption for all vehicles. The macroscopic simulation study confirmed our analytical evaluation where we
showed how to compute the saturation point of the AV lane and illustrated that after this point is reached,
the benefits for AV users become negligible. We further compared the scenario with AV lane introduction to a
baseline scenario where no changes to traffic regulations are made. The latter scenario outperforms the former
one over the whole range of AV percentages, however, the difference is of considerable amount only before the
saturation point is reached. This finding coincides with our analytical evaluation of the two scenarios.
Future work includes micro (and submicroscopic) studies to further provide insights on the effects of the
introduction of AV lanes. This includes benefits due to smart platooning strategies but also turbulences caused
by lateral vehicle movement, e.g., from the on-ramp towards the AV lane. Furthermore, the authors would like
to exchange the UE traffic assignment of the AV group of vehicles with the BISOS algorithm [23], which looks
for a system optimum assignment, in order to check whether the adoption of an AV lane would become more
beneficial in this case.
Acknowledgements
This work was financially supported by the Singapore National Research Foundation under its Campus for
Research Excellence And Technological Enterprise (CREATE) programme.
References
[1] J. Henderson and J. Spencer, "Autonomous Vehicles and Commercial Real Estate," Cornell Real Estate
Review, vol. 14, no. 1, June 2016.
[2] T. Litman, "Autonomous Vehicle Implementation Predictions – Implications for Transport Planning,"
Victoria Transport Policy Institute, Tech. Rep., November 2016.
[3] M. Segata, B. Bloessl, S. Joerer, F. Dressler, and R. Lo Cigno, "Supporting Platooning Maneuvers through
IVC: An Initial Protocol Analysis for the Join Maneuver," in 11th IEEE/IFIP Conference on Wireless
On demand Network Systems and Services (WONS 2014). Obergurgl, Austria: IEEE, April 2014, pp.
130–137.
[4] J. Ivanchev, D. Zehe, S. Nair, and A. Knoll, "Fast identification of critical roads by neural networks using
system optimum assignment information," in Intelligent Transportation Systems (ITSC), 2017 IEEE 20th
International Conference on.
IEEE, 2017.
[5] J. Ivanchev, H. Aydt, and A. Knoll, "Spatial and temporal analysis of mismatch between planned road
infrastructure and traffic demand in large cities," in 2015 IEEE 18th International Conference on Intelligent
Transportation Systems, Sept 2015, pp. 1463–1470.
[6] A. Kanaris, P. Ioannou, and F.-S. Ho, "Spacing and Capacity Evaluations for Different AHS Concepts," in
Automated Highway Systems. Springer, June 1997, pp. 125–171.
[7] S. Cohen and J. Princeton, "Impact of a Dedicated Lane on the Capacity and the Level of Service of an
Urban Motorway," Procedia-Social and Behavioral Sciences, vol. 16, pp. 196–206, July 2011.
[8] J. Carbaugh, D. N. Godbole, and R. Sengupta, "Safety and Capacity Analysis of Automated and Manual
Highway Systems," Transportation Research Part C: Emerging Technologies, vol. 6, no. 1, pp. 69–99,
February 1998.
[9] H. Tsao, R. Hall, and B. Hongola, "Capacity Of Automated Highway Systems: Effect Of Platooning And
Barriers," Institute of Transportation Studies, University of California, Berkeley, Tech. Rep. UCB-ITS-
PRR-93-26, February 1994.
[10] N. Harwood and N. Reed, "Modelling the impact of platooning on motorway capacity," in Road Transport
Information and Control Conference 2014 (RTIC 2014). London, UK: IET, October 2014.
11
[11] J. B. Michael, D. N. Godbole, J. Lygeros, and R. Sengupta, "Capacity Analysis of Traffic Flow Over a
Single-Lane Automated Highway System," Journal of Intelligent Transportation System, vol. 4, no. 1-2,
pp. 49–80, 1998.
[12] R. Hall and C. Chin, "Vehicle sorting for platoon formation: Impacts on highway entry and throughput,"
Transportation Research Part C: Emerging Technologies, vol. 13, no. 5, pp. 405–420, October 2005.
[13] B. Van Arem, C. J. Van Driel, and R. Visser, "The Impact of Cooperative Adaptive Cruise Control on
Traffic-Flow Characteristics," IEEE Transactions on Intelligent Transportation Systems, vol. 7, no. 4, pp.
429–436, December 2006.
[14] D. N. Godbole and J. Lygeros, "Safety and throughput analysis of automated highway systems," Institute
of Transportation Studies, University of California, Berkeley, Tech. Rep. UCB-ITS-PRR-2000-1, January
2000.
[15] K. M. Dresner and P. Stone, "Sharing the Road: Autonomous Vehicles Meet Human Drivers," in Twentieth
International Joint Conference on Artificial Intelligence (IJCAI'07), vol. 7, Hyderabad, India, January
2007, pp. 1263–1268.
[16] P. A. Ioannou and C.-C. Chien, "Autonomous intelligent cruise control," IEEE Transactions on Vehicular
technology, vol. 42, no. 4, pp. 657–672, November 1993.
[17] T. Yokota, S. Ueda, and S. Murata, "Evaluation of ahs effect on mean speed by static method," in Fifth
World Congress on Intelligent Transport Systems, Seoul, South Korea, October 1998.
[18] S. C. Dafermos and F. T. Sparrow, "The traffic assignment problem for a general network," Journal of
Research of the National Bureau of Standards, Series B, vol. 73, no. 2, pp. 91–118, 1969.
[19] M. Sindhwani and Q. K. Xin, "Singapore Traffic Information Platform: Enabling Traffic-Aware Applications
& Systems," in 17th ITS World Congress, Busa, South Korea, October 2010.
[20] J. Ivanchev, S. Litescu, D. Zehe, M. Lees, H. Aydt, and A. Knoll, "Determining the Most Harmful Roads
in Search for System Optimal Routing," TU Munich, Tech. Rep. TUM-I1632, February 2016.
[21] C. Fisk, "Some developments in equilibrium traffic assignment," Transportation Research Part B: Method-
ological, vol. 14, no. 3, pp. 243–255, September 1980.
[22] W. F. Faris, H. A. Rakha, R. I. Kafafy, M. Idres, and S. Elmoselhy, "Vehicle fuel consumption and emission
modelling: an in-depth literature review," International Journal of Vehicle Systems Modelling and Testing,
vol. 6, no. 3-4, pp. 318–395, 2011.
[23] J. Ivanchev, D. Zehe, V. Viswanathan, S. Nair, and A. Knoll, "Bisos: Backwards incremental system
optimum search algorithm for fast socially optimal traffic assignment," in 2016 IEEE 19th International
Conference on Intelligent Transportation Systems (ITSC), Nov 2016, pp. 2137–2142.
12
|
1806.07123 | 1 | 1806 | 2018-06-19T09:25:53 | Cooperative Queuing Policies for Effective Human-Multi-Robot Interaction | [
"cs.MA"
] | We consider multi-robot applications, where a team of robots can ask for the intervention of a human operator to handle difficult situations. As the number of requests grows, team members will have to wait for the operator attention, hence the operator becomes a bottleneck for the system. Our aim in this context is to make the robots learn cooperative strategies to decrease the time spent waiting for the operator. In particular, we consider a queuing model where robots decide whether or not to join the queue and use multi-robot learning to estimate the best cooperative policy. In more detail, we formalize the problem as Decentralized Markov Decision Process and provide a suitable state representation, so to apply an independent learners approach. We evaluate the proposed method in a robotic water monitoring simulation and empirically show that our approach can significantly improve the team performance, while being computationally tractable. | cs.MA | cs | Autonomous Robots manuscript No.
(will be inserted by the editor)
Cooperative Queuing Policies for Effective Human-Multi-Robot
Interaction
Masoume M. Raeissi · Alessandro Farinelli
8
1
0
2
n
u
J
9
1
]
A
M
.
s
c
[
1
v
3
2
1
7
0
.
6
0
8
1
:
v
i
X
r
a
Received: date / Accepted: date
Abstract We consider multi-robot applications, where a team
of robots can ask for the intervention of a human operator to
handle difficult situations. As the number of requests grows,
team members will have to wait for the operator attention,
hence the operator becomes a bottleneck for the system. Our
aim in this context is to make the robots learn cooperative
strategies to decrease the time spent waiting for the opera-
tor. In particular, we consider a queuing model where robots
decide whether or not to join the queue and use multi-robot
learning to estimate the best cooperative policy. In more de-
tail, we formalize the problem as Decentralized Markov De-
cision Process and provide a suitable state representation, so
to apply an independent learners approach. We evaluate the
proposed method in a robotic water monitoring simulation
and empirically show that our approach can significantly im-
prove the team performance, while being computationally
tractable.
1 Introduction
In many multi-robot scenarios, such as environmental mon-
itoring [23] or search and rescue [24, 11], one or few opera-
tors are required to interact with a team of robots to perform
complex tasks in challenging environments. Robots, spe-
cially at field sites, are often subject to unexpected events,
that can not be managed without the intervention of opera-
tors. For example, in an environmental monitoring applica-
tion, robots might face extreme environmental events (e.g.,
water currents) or moving obstacles (e.g., animal approach-
ing the robots). In such scenarios, the operator often needs
University of Verona
Verona, Italy
E-mail: [email protected] University of Verona
Verona, Italy
E-mail: [email protected]
to interrupt the activities of individual team members to deal
with particular situations.
The operator's monitoring and supervisory role in these
scenarios becomes critical, particularly when the team size
grows larger. To decrease the operator's monitoring task and
give him/her more time to focus on robots that need atten-
tion, several approaches consider the concept of self-reflection
[19], where robots are able to identify their potential issues
and ask for the intervention of the operator by sending a
request. However, large teams can easily overwhelm the op-
erator with several requests, hence hindering the team per-
formance. Consequently, team members have to wait for the
operator's attention, and the operator becomes a bottleneck
for the system. Queuing is a natural way to manage and ad-
dress this problem. Previous research try to enhance the per-
formance of the system (i.e., decreasing the time spent by
robots waiting for the operator) considering various queue
disciplines (e.g. First in First Out (FIFO) and Shortest Job
First (SJF)) [13, 4] or prioritizing such requests [17]. In both
cases, the queue size may grow indefinitely as no robot will
leave the queue before receiving the operator's attention.
To deal with this problem, we focus on balking queue
model [14], in which the users/agents (i.e robots requesting
attention) can decide either to join the queue or balk. Such
decisions are typically based on a threshold value, that is
computed by assigning a generic reward for receiving the
service and a cost for waiting in the queue to each agent.
When applying this model to a robotic application, there is
no clear indication how such thresholds can be computed.
More important, this model does not consider the cost of
balking (i.e. the cost of a potential failure that robots can
suffer by trying to overcome difficult situations without hu-
man intervention). Considering this, our focus is to devise
an approach that allows the team of robots to learn coopera-
tive balking strategies to make better use of a shared queue.
Therefore, we frame the above problem as a Decentralized
2
M. Raeissi, A. Farinelli
resentation [22, 12], multi-modal interaction [20, 8] and mixed
initiative planning [2]. Here, we consider approaches, which
focus on how to allocate operator attention to a set of robots.
Many work in this area consider that, the robots can perceive
their situations and inform or ask the operator for help. For
example, the work by [18] proposes the idea of a single ser-
vice robot asking for the help of humans.
Considering larger multi-robot teams with small number
of operators, several robots may need the operator's atten-
tion at the same time. Hence, the requests must be queued
for being processed later on. Authors in [4, 13] explore dif-
ferent queue disciplines to enhance the performance of the
system. However, keeping robots idle until the operator be-
comes available might decrease the overall team efficiency.
In contrast, we focus on a specific queuing model with balk-
ing property [14], where robots decide either to join the
queue or not.
The concept of Adjustable Autonomy or mixed initiative
has been the basis of many research in the field of human-
multi-robot interaction. The key issue in this setting is to
devise effective techniques to decide whether a transfer of
control should occur and when this should happen. Differ-
ent techniques have been proposed to address this challeng,
for example, [10, 5, 21] consider that the robot will ask for
human help/intervention when the expected utility of do-
ing so is higher than performing the task autonomously, or
when the uncertainty of the autonomous decision is high
[9, 3] or when the autonomous decision can cause significant
harm [6]. However, these decision making solutions usually
have been considered as individual one-step decisions with-
out considering the long-term cost or the consequences of
decisions on other team members (if any). Our work also
belongs to the this category, where autonomous robots in a
multi-robot scenario can decide whether to wait for the op-
erator or not. In particular, we focus on a specific queuing
model with balking property [14]. However, our focus is on
finding cooperative strategies, where all robots learn concur-
rently to optimize the idle time of the system.
2.2 Balking Queue Model
The first mathematical model of a queuing system with ra-
tional users was formulated by Naor [14]. In his model, users
upon their arrival decide according to a threshold value whether
to join the queue or not (balk). The individual's optimizing
strategy is straightforward, a customer will join the queue
while n other customers are already in the system if
R− n·C
≥ 0
1
µ
(1)
where a uniform cost C for staying in the queue and a
similar reward R for receiving service are assigned to each
Fig. 1 Water monitoring simulation tool.
Markov Decision Process (Dec-MDP) in which, the team of
robots must cooperate to optimize the team idle time. Find-
ing optimal decentralized policies is often hard due to the
partial observability and limited communications. Thus, our
goal is to provide a scalable state representation by adding
the state of the queue as an extra feature to the robots' lo-
cal states and solve the underlying Dec-MDP problem using
multiple independent learners 1. We illustrate that, this ad-
ditional feature will improve the team performance over our
main evaluation metric (i.e. team idle time). In more detail,
this paper provides the following contributions to the state
of the art: (i) We model the human-multi-robot interaction
as a balking queue, in which the robots identify their needs
and decide whether to interrupt the human operator or not.
(ii) We formalize the problem as Dec-MDP and provide a
tractable state representation to learn the balking policies for
each robot. (iii) Finally, we evaluate the performance of our
model by comparing the team idle time to the state-of-the-
art queue disciplines. Overall, the experimental results show
that, the use of our model decreases the total idle time upto
68% over FIFO (without balking) and increases the team re-
ward upto 56% comparing to the other learning models.
2 Background
In this section, first we review the state of the art in robotic
studies, where robots ask for operator's attention. Afterwards,
we present a brief introduction to the Balking Queue[14]
model in which, users/agents decide to join the queue or
not. Then, we present a brief review of Dec-MDP[1, 7] as
the basis model for decision making under uncertainty in
multi-robot scenarios.
2.1 Human-Multi-Robot Interaction
Human-multi robot interaction is an active field of research
with many different research issues, including team plan rep-
1 Some part of this work appears in [16]. That work describes basic
ideas and preliminary results, here we provide a more detailed descrip-
tion of the methodologies, and more extensive empirical analysis.
Effective Human-Multi-Robot Interaction
3
user and µ is the intensity parameter of exponentially dis-
tributed service time. Thus, n = (cid:98) Rµ
C (cid:99) serves as a threshold
value for balking, that is if the number of users waiting in
the queue is greater than n, the newly arrived user will not
join the queue. In a multi-robot application, this threshold
and decision must be computed carefully. Our focus is on
showing how the elements (i.e. reward and cost) of balking
strategy should be adjusted according to a practical robotics
scenario.
2.3 Decentralized Markov Decision Process (Dec-MDP)
The decision of whether to join the queue or not for each
situation of each robot will impact the future decisions of the
entire team. As a result, we are concerned here with team
sequential decision making problems, in which the team's
utility depends on a sequence of decisions.
A Dec-MDP is defined by a tuple (cid:104)S,A,P,R(cid:105) where: S is
the set of world states which is factored into n components,
S = S1×...×Sn. In a special case (i.e. Factored n-agent Dec-
MDP), Si refers to the local state of agent i. In Dec-MDP, the
state is jointly fully observable which means that the aggre-
gated observations made by all agents determines the global
state. A = ×iAi is the set of joint actions, where Ai is the set
of actions for agent i. P = S×A×S → [0,1] is the state tran-
sition probability. R = S× A → R is the immediate reward.
The complexity of Dec-MDP is nondeterministic exponen-
tial (NEXP) hard [1], hence learning is crucial.
3 Problem Formulation
We consider a water monitoring scenario, where several au-
tonomous surface vessels are supervised by a human oper-
ator (see Fig. 1). A set of events can happen to the plat-
forms, such events may affect the normal behavior of the
platforms and hinder their performance. Each event has a
different probability of failure (see Table 1), where requests
with the higher probability of failure are more crucial to re-
ceive the operator's attention
Following previous works [4, 13], a central queue is pro-
vided to both the operator and the boats, where the operator
can select one request at a time (i.e., in FIFO order) and as-
signs a specific sub-mission to resolve that request. A sub-
mission is a plan specific recovery procedure, and this of-
ten requires a human interaction (i.e., the human directly
selects which platforms should execute the interrupt sub-
mission). We used three sub-missions, one for each class
of requests, including: (i) sending a boat to the closest sta-
tion to change/charge its battery. (ii) allowing/not-allowing
a boat to go further (to the area that it might loose connec-
tion), and (iii) teleoperating a boat for traversing a specific
area.
Table 1 Different event types used in the experiments.
Event Type (E j)
Battery-Recharge (E1)
Traversing-Dangerous-Area (E2)
Losing-Connection (E3)
Prob. of Fail
0.9
0.4
0.2
We assume that, whenever an event happens, the plat-
form can detect the event. For example, a robot can perceive
that its battery level is in a critical state, it must then decide
whether to join the queue (i.e. sending the request and wait-
ing for the operator) or balk (i.e., not sending the request) 2.
The consequences or costs of balking are problem specific.
In our model, when a failure happens, the operator should
spend more time to fix the problem, hence failure as a result
of balking, increases the idle time of the system. The goal is
to minimize the time spent in the queue.
Our proposal is then to train the robots in a stationary en-
vironment (i.e., stationary distribution functions with fixed
arrival rate and service time), so that the robots can learn
appropriate balking policies. Then, by applying the learned
policies in similar scenarios, they will be able to optimize
the team objective. More specifically, we consider the fol-
lowing model in our domain: The state space S = S1 × S2 ×
... × Sn. n is the number of boats. The local state of each
boat Si is a tuple (cid:104)Sb,Ntasks(cid:105). Ntasks shows the number of
remaining tasks of boat i. In this application domain, each
task is a location that should be visited by a specific boat.
Sb is the current internal state of boat i. More specifically
Sb ∈ {E j,Waiting,Failed,Autonomy}, where E j refers to
on of the request/event type in Table 1. For example, the
state tuple of a boat when it has 3 tasks to finish and the
event Battery Recharge occurs, would be s = (cid:104)E1,3(cid:105). Ai is
the set of actions for boat i where Ai ∈ {Join,Balk}. The re-
ward function is designed to decrease the idle time (i.e. the
time spent waiting for the operator).
In general, there are two major approaches for learning
in multi-robot scenarios [15]. The first approach is called
team learning and uses a single learner to learn the behav-
ior for the entire team. In contrast, the second approach uses
multiple concurrent learners, usually one for each robot, where
each learner tries to learn its behavior. Each of these meth-
ods has its own advantages and disadvantages which make
it preferable in different domains [25, 15]. In particular, the
major problems with team learning approach are the explo-
sion of the state space (i.e., it keeps the states of the en-
tire team), and the centralization of the learning approach
that needs to access the states of all team members. Using
the team learner in our application, the state space will be
very large which decelerates the convergence to the optimal
2 While this may be a significant challenge in some domains, this is
not the focus of our work.
4
M. Raeissi, A. Farinelli
value. For example, for 5 boats with the above state repre-
sentation, the state space will include more than one million
states, hence requiring a prohibitive long time to estimate
the optimal strategies for each state and action permutations.
The main advantage of independent learners in our domain
is that, this domain can be decomposed into subproblems
(e.g. each boat holds its own state space) and each subprob-
lem can be solved by one boat. In general, two main chal-
lenges arise in concurrent learning: credit assignment and
non-stationary dynamics of the environment [15]. However,
our application scenario has some special properties, that
can be exploited to design a tractable model. First, the ac-
tion selection at each step (i.e. when an event happens) only
requires one agent to select either to join or balk. Hence, the
reward can go directly to that agent. It is different from the
situations, where all agents should decide at each step (i.e.
joint actions), which results in the well-known credit assign-
ment issue. However, when each boat considers only its lo-
cal state without knowing the state of the queue, finding the
optimal behavior for the team may become impossible, or
the model may compute lower quality solutions. Therefore,
we add the state of the queue to the local state of each boat,
and then we use independent learners approach. To sum up,
we consider three possible models:
Team Learner (TL): a team learner has access to the joint
state of all robots which is S = S1 × S2 × ...× Sn. When an
event happens to a boat, the action (cid:104)Join,Balk(cid:105) for the cor-
responding boat will be selected and the state of the system
will be updated. The update will only change the part of the
state related to the corresponding boat. The Q-value of the
team learner will be updated accordingly.
Independent Learners - Unobservable Queue (IL-U): an
independent learner) is used for each boat. Each boat ob-
serves only its local state Si = (cid:104)Sb,Ntasks(cid:105). In this model,
each boat updates its local Q-values interacting to the sys-
tem and receiving the reward.
Independent Learners - Observable Queue (IL-O): in this
model, each boat in addition to observing its local state, has
access to the size of the queue. The queue size shows the
number of waiting boats inside the queue. The state repre-
sentation of each boat in this model is: Si = (cid:104)Sb,Ntasks,Sq(cid:105).
These models are different in their state representation, while
the reward structure is the same for all of them:
(i) R(St = Si,At = Join) = RS − (Nq ¯µ +tserv).
(ii) R(St = Si,At = Balk) = RF ( ¯µ
(iii) R(St = Si,At = Balk) = RT ; if St+1 = A.
¯λ ) + Nq; if St+1 = F.
¯µ and ¯λ are average service time and events arrival rate
respectively. Nq is the number of boats waiting, and tserv is
the average time needed to resolve the request. RS = 1, RF =
−2 and RT = 0.3 are application specific parameters that
must be tuned empirically.
Finally, we use Q-Learning as the basis learning approach,
while the same reward structure, same distribution functions
for generating events and same distribution function for the
service time are used for all models. We use Q-Learning
because of its simplicity and good real time performance.
Moreover, our goal is to propose the use of reinforcement
learning in this novel context and not to provide a novel
learning algorithm. In Q-Learning, each learner interacts with
the environment (i.e. selects an action), receives the immedi-
ate reward and updates its state-action values (i.e. Q-values)
as Eq.2:
Qi(si,ai) ← Qi(si,ai) + α(ri + γ maxa(cid:48)∈Ai Qi(s(cid:48),a(cid:48))− Qi(si,ai))
(2)
where ri and s(cid:48) are respectively the reward and the state
observed by robot i after performing action ai in state si; a(cid:48)
is the action in state s(cid:48) that maximizes the future expected
rewards; α is the learning rate and γ is the discount factor.
Notice that, for Team Learner, there is only one Q-table (i.e.,
table of state-action values) to be updated, where, si refers
to full state representation (i.e. state of all robots) and ai is
the action selected by learner for the specific robot with a
request. This action will either add the event to the queue
or not. As mentioned before, by exploiting specific feature
of our domain, we can use an independent learner approach
that is tractable and scalable.
In the simulation, ¯λ and ¯µ in the reward function are
the same as λ and µ for generating the requests and ser-
vice. However, for a field deployment, these two parameters
should be estimated by considering the average number of
events being generated during an interval as lambda (e.g.
20 events have been generated in 1 hour (or 60 minutes)),
and the average time spent to fix each request as 1
µ (e.g. 5
minutes to fix each request, hence 12 events can be fixed in
60 minutes).
The state of the queue, Sq, can be modified by robots'
action (joining the queue) and the operator's action (leav-
ing the queue). However, under the reasonable assumption
that an arrival and a departure cannot happen exactly at the
same time, only one entity can change the value of Sq at
a time. Moreover, the possibility of having more than one
event at the exact same time is very low. In particular, for
our scenario, it is safe to assume that, the time to change
the state of the queue, is much lower than the time for a new
event arrival. Under this assumption, even if two events hap-
pen within a short time interval, the first one will affect the
state of the queue before the second arrives, hence the other
robots will base their decisions on the updated queue size.
4 Experimental Evaluation
4.1 Learning Phase
The learning phase of balking models starts by defining a
list of locations (i.e., to be visited), and assigning those loca-
tions to boats. We consider 30 locations and 5 boats. Events,
Effective Human-Multi-Robot Interaction
5
Fig. 2 Team accumulated reward in each episode of the learning phase
(better viewed in color).
Fig. 3 Team accumulated reward in each episode of the learning phase
(better viewed in color). From episode 2150, the service rate has been
changed from 0.27 to 0.37 and 0.17.
rate λ an event may happen to one boat. We used a realistic
estimation for parameter λ and µ based on some experience
on the total mission time, number of boats and number of lo-
cations. These numbers define well the type of scenarios we
are interested in, where boats can operate most of the time
in autonomy, but frequently need user's intervention.
Fig. 2 shows the team rewards of each model, TL, IL-U
and IL-O, during the learning phase. The oscillation in the
reward is due to the fact that, the robots learn their policies
by trying new potentially sub-optimal actions. The training
time (the sum over 4000+ episodes) for each model was
about 88-95 hours. As we expected, the convergence rate of
IL-O is much faster than the TL, while they both reach a sim-
ilar team reward. This is due to the larger state space of TL
which needs more iterations to estimate the value for each
state and action. Results also clearly show the importance
of having access to the state of the queue to make better de-
cisions. Since, the reward given to each action is related to
the parameters λ and µ, we expect our policy to be depen-
dent on these two parameters. Fig. 3 shows how IL-O adapts
to changes in µ, where we increase and decrease its value
by 40% during the learning phase. A sudden rise and drop
happen respectively for each value, but then the system con-
verges to a stationary state. Fig. 4 shows another experiment,
where we vary the ratio of event types during the learning
phase. The events were generated with a uniform distribu-
tion up to episode 2150. After that, we vary the percentage
of events from type 1(E1), which has the higher probabil-
ity of failure (see Table 1), while the rest of the events are
uniformly distributed among E2 and E3. In more detail, we
consider E1-100, E1-80, E1-50, E1-10 and E1-0, where each
shows the rate of E1. Fig. 4 illustrates that, as there are more
E1, the system will gain less reward due to the increasing
rate of failures. After several iterations, the learning curve
becomes stationary.
Fig. 4 Team accumulated reward in each episode of the learning phase
(better viewed in color). After episode 2150, we vary the rate of each
event type.
as in table 1, will be generated within an exponential distri-
bution with parameter λ = 0.25. The operator's speed, for
resolving a request is selected from an exponential distri-
bution with parameter µ = 0.27. An episode (i.e., a run of
the algorithm beginning from a start state to a final state)
ends after the system encounters 20 events. For action selec-
tion in our model, we use ε greedy method with parameter
ε = 0.1. Our algorithm uses the learning rate α = 0.1 and
discount factor γ = 0.9 throughout the experiments, which
are tuned empirically. Each episode of the learning phase
starts with all boats in their Autonomy state (i.e. they do not
need the attention of the human operator), then with arrival
4.2 Test Phase
After the learning phase, we run 30 simulation executing
the policy learned previously. In this first experiment, we
use the same values for λ and µ as used during the learn-
ing phase. Fig. 5(a) demonstrates the team reward for each
learning models. A comparison on team reward between IL-
O and IL-U, shows 56% gain for IL-O. Besides, a signifi-
cant decrease (i.e. 40%) on average waiting time is shown
in Fig. 5(b) when using IL-O rather than IL-U. One might
expect the same reward value and idle time for IL-O and
TL. However, the results on Fig. 5(a) and 5(b) show bet-
ter performance values for IL-O than TL. This difference is
due to the fact that, IL-O model keeps only the size of the
queue or Sq (i.e. it does not consider which boats are wait-
ing in the queue), while TL maintains the state of all boats
6
M. Raeissi, A. Farinelli
(a) Total team reward
(b) Total idle time
(c) Total idle time
Fig. 5 (a) and (b) show the team performance (together with the standard error of the means) for three learning models, while (c) compares balking
models to non-balking models.
Fig. 6 Total team reward (together with the standard error of the
means) for IL-O (main) with different levels of noise on λ and µ
(noisy).
SJF, we use the same event rate λ and service rate µ. In these
two queuing models, boats always join the queue regardless
of their request types and the queue size. Fig. 5(c) shows
the team idle time for FIFO, SJF and three learning mod-
els. FIFO without balking, has the worst performance, since
boats wait for the operator until he/she becomes available.
In contrast, IL-O approach outperforms all other models. In
more detail, it decreases the time up to 68% comparing to
FIFO. In general, the results in Fig. 5(c) indicate that, using
balking models significantly decreases the idle time of the
team even though, some events may result in failures. This
is acceptable in our domain, since the penalties for failures
are not critical but only result in a finite increase of time.
To validate the noise sensitivity of our proposed model
IL-O, we consider a set of experiments as follow. We con-
sider adding the same level of noise, according to a uni-
form distribution, to both parameters λ and µ during the
test phase. Fig. 6 shows the team reward and Fig. 7 shows
the team idle time for different levels of noise. The results
show that, the approach is able to cope with a significant
amount of noise on both λ and µ.
Fig. 7 Total idle time (together with the standard error of the means)
for IL-O (main) with different levels of noise on λ and µ (noisy).
which are in their Waiting state (i.e. Sb = W). For exam-
ple, whenever two boats waiting in the queue (assuming the
other features, e.g., severity are the same), IL-O will map
the state to Sq = 2, while TL will differentiate the states de-
pending on which two boats are inside the queue. Since, the
boats are homogeneous in our domain, IL-O results in better
performance by abstracting away features that do not have a
significant impact on the reward. This also makes TL to con-
verge slower than IL-O, due to the larger state space of TL
which needs more iterations to estimate the value for each
state and action.
Next, we compare the behavior of queues with and with-
out balking property (e.g., FIFO and SJF). For FIFO and
5 Conclusions
In this paper, we propose the use of balking queue to model
human-multi-robot interactions when the autonomy of robots
allow them to decide whether to wait for the operator or not.
We frame the problem as a Dec-MDP in which, each robot
observes its local state and the state of the queue and cooper-
ates with other agents to optimize the use of a shared queue.
We apply independent Q-Learning to find these cooperative
strategies in a water monitoring multi-robot simulation. We
consider three different models (TL, IL-U, IL-O), and our
results clearly show that an independent learner approach
where the state of the queue is accessible to the platforms
performs best. Furthermore, the empirical results related to
the noisy estimation for λ and µ (Fig. 6 and 7), suggest that
the approach is able to cope with a significant amount of
noise.
Effective Human-Multi-Robot Interaction
References
1. Bernstein, D.S., Givan, R., Immerman, N., Zilberstein,
S.: The complexity of decentralized control of markov
decision processes. Mathematics of operations research
27(4), 819–840 (2002)
2. Bevacqua, G., Cacace, J., Finzi, A., Lippiello, V.:
Mixed-initiative planning and execution for multiple
drones in search and rescue missions. In: Proceedings of
the Twenty-Fifth International Conference on Interna-
tional Conference on Automated Planning and Schedul-
ing, ICAPS'15, pp. 315–323. AAAI Press (2015)
3. Chernova, S., Veloso, M.: Interactive policy learning
through confidence-based autonomy. J. Artif. Int. Res.
34(1), 1–25 (2009). URL http://dl.acm.org/
citation.cfm?id=1622716.1622717
4. Chien, S.Y., Lewis, M., Mehrotra, S., Brooks, N.,
Sycara, K.P.: Scheduling operator attention for multi-
robot control.
In: IEEE/RSJ International Conference
on Intelligent Robots and Systems, IROS, Vilamoura,
Algarve, Portugal, October 7-12, 2012, pp. 473–479
(2012)
5. Collins, J., Bilot, C., Gini, M., Mobasher, B.: Mixed-
initiative decision support in agent-based automated
contracting. In: Proceedings of the Fourth International
Conference on Autonomous Agents, AGENTS '00, pp.
247–254. ACM, New York, NY, USA (2000)
6. Dorais, G.A., Bonasso, R.P., Kortenkamp, D., Pell, B.,
Schreckenghost, D.: Adjustable autonomy for human-
centered autonomous systems. In: on Mars, in First In-
ternational Conference of the Mars Society (1998)
7. Goldman, C.V., Zilberstein, S.: Optimizing information
exchange in cooperative multi-agent systems. In: Pro-
ceedings of the Second International Joint Conference
on Autonomous Agents and Multiagent Systems, AA-
MAS '03, pp. 137–144. ACM, New York, NY, USA
(2003)
8. Gromov, B., M. Gambardella, L., Di Caro, G.: Wear-
able multi-modal interface for human multi-robot inter-
action. In: in: IEEE International Symposium on Safety,
Security, and Rescue Robotics, SSRR2016., pp. 240–
245 (2016)
9. Gunderson, J.P., Martin, W.N.: Effects of uncertainty
In: IN
on variable autonomy in maintenance robots.
WORKSHOP ON AUTONOMY CONTROL SOFT-
WARE, pp. 26–34 (1999)
10. Horvitz, E., Jacobs, A., Hovel, D.: Attention-sensitive
In: Proceedings of the Fifteenth Conference
alerting.
on Uncertainty in Artificial Intelligence, UAI'99, pp.
305–313. Morgan Kaufmann Publishers Inc., San Fran-
cisco, CA, USA (1999). URL http://dl.acm.
org/citation.cfm?id=2073796.2073831
7
11. Hsieh, M.A., Cowley, A., Keller, J.F., Chaimowicz,
L., Grocholsky, B., Kumar, V., Taylor, C.J., Endo, Y.,
Arkin, R.C., Jung, B., Wolf, D.F., Sukhatme, G.S.,
MacKenzie, D.C.: Adaptive teams of autonomous aerial
and ground robots for situational awareness. Journal of
Field Robotics 24(11-12), 991–1014 (2007)
12. Kaminka, G.A., Frenkel, I.: Flexible teamwork in
behavior-based robots. In: Proceedings of the 20th Na-
tional Conference on Artificial Intelligence - Volume 1,
AAAI'05, pp. 108–113. AAAI Press (2005)
13. Lewis, M., Chien, S.Y., Mehortra, S., Chakraborty, N.,
Sycara, K.: Task switching and single vs. multiple
alarms for supervisory control of multiple robots. In: In-
ternational Conference on Engineering Psychology and
Cognitive Ergonomics, pp. 499–510. Springer (2014)
14. Naor, P.: The regulation of queue size by levying tolls.
Econometrica 37(1), 15–24 (1969). URL http://
www.jstor.org/stable/1909200
15. Panait, L., Luke, S.: Cooperative multi-agent learning:
The state of the art. Autonomous agents and multi-agent
systems 11(3), 387–434 (2005)
16. Raeissi, M.M., Farinelli, A.: Learning queuing strate-
gies in human-multi-robot interaction. Has been ac-
cepted as Extended Abstract, AAMAS '18 (2018)
17. Rosenfeld, A., Agmon, N., Maksimov, O., Azaria, A.,
Kraus, S.: Intelligent agent supporting human-multi-
robot team collaboration. In: Proceedings of the 24th
International Conference on Artificial Intelligence, IJ-
CAI'15, pp. 1902–1908 (2015)
18. Rosenthal, S., Veloso, M.: Using symbiotic relation-
ships with humans to help robots overcome limitations.
In: Workshop for Collaborative Human/AI Control for
Interactive Experiences (2010)
19. Scheutz, M., Kramer, J.: Reflection and reasoning
mechanisms for failure detection and recovery in a dis-
tributed robotic architecture for complex robots.
In:
Robotics and Automation, 2007 IEEE International
Conference on, pp. 3699–3704. IEEE (2007)
20. Stoica, A., Theodoridis, T., Hu, H., McDonald-Maier,
K., Barrero, D.: Towards human-friendly efficient con-
trol of multi-robot teams. In: Proceedings of the 2013
International Conference on Collaboration Technolo-
gies and Systems, CTS 2013, pp. 226–231 (2013)
21. Suzanne Barber, K., Goel, A., Martin, C.E.: Dynamic
adaptive autonomy in multi-agent systems.
Journal
of Experimental & Theoretical Artificial Intelligence
12(2), 129–147 (2000)
22. Tambe, M.: Towards flexible teamwork. J. Artif. Int.
Res. 7(1), 83–124 (1997)
23. Valada, A., Velagapudi, P., Kannan, B., Tomaszewski,
C., Kantor, G., Scerri, P.: Development of a Low Cost
Multi-Robot Autonomous Marine Surface Platform.
Springer Berlin Heidelberg (2014)
8
M. Raeissi, A. Farinelli
24. Wang, J., Lewis, M.: Human control for cooperating
robot teams. In: Proceedings of the ACM/IEEE Inter-
national Conference on Human-robot Interaction, HRI
'07, pp. 9–16. ACM (2007)
25. Xuan, P., Lesser, V.: Multi-agent policies: From cen-
tralized ones to decentralized ones. In: Proceedings of
the First International Joint Conference on Autonomous
Agents and Multiagent Systems: Part 3, AAMAS '02,
pp. 1098–1105. ACM (2002)
|
1911.08642 | 1 | 1911 | 2019-11-20T00:39:36 | Scalable Decision-Theoretic Planning in Open and Typed Multiagent Systems | [
"cs.MA"
] | In open agent systems, the set of agents that are cooperating or competing changes over time and in ways that are nontrivial to predict. For example, if collaborative robots were tasked with fighting wildfires, they may run out of suppressants and be temporarily unavailable to assist their peers. We consider the problem of planning in these contexts with the additional challenges that the agents are unable to communicate with each other and that there are many of them. Because an agent's optimal action depends on the actions of others, each agent must not only predict the actions of its peers, but, before that, reason whether they are even present to perform an action. Addressing openness thus requires agents to model each other's presence, which becomes computationally intractable with high numbers of agents. We present a novel, principled, and scalable method in this context that enables an agent to reason about others' presence in its shared environment and their actions. Our method extrapolates models of a few peers to the overall behavior of the many-agent system, and combines it with a generalization of Monte Carlo tree search to perform individual agent reasoning in many-agent open environments. Theoretical analyses establish the number of agents to model in order to achieve acceptable worst case bounds on extrapolation error, as well as regret bounds on the agent's utility from modeling only some neighbors. Simulations of multiagent wildfire suppression problems demonstrate our approach's efficacy compared with alternative baselines. | cs.MA | cs | Scalable Decision-Theoretic Planning in Open and Typed Multiagent Systems
Adam Eck,1 Maulik Shah,2 Prashant Doshi,2, Leen-Kiat Soh3
1Oberlin College, 2University of Georgia, 3University of Nebraska
[email protected], [email protected], [email protected], [email protected]
9
1
0
2
v
o
N
0
2
]
A
M
.
s
c
[
1
v
2
4
6
8
0
.
1
1
9
1
:
v
i
X
r
a
Abstract
In open agent systems, the set of agents that are cooperating
or competing changes over time and in ways that are non-
trivial to predict. For example, if collaborative robots were
tasked with fighting wildfires, they may run out of suppres-
sants and be temporarily unavailable to assist their peers. We
consider the problem of planning in these contexts with the
additional challenges that the agents are unable to commu-
nicate with each other and that there are many of them. Be-
cause an agent's optimal action depends on the actions of oth-
ers, each agent must not only predict the actions of its peers,
but, before that, reason whether they are even present to per-
form an action. Addressing openness thus requires agents to
model each others presence, which becomes computationally
intractable with high numbers of agents. We present a novel,
principled, and scalable method in this context that enables
an agent to reason about others' presence in its shared envi-
ronment and their actions. Our method extrapolates models
of a few peers to the overall behavior of the many-agent sys-
tem, and combines it with a generalization of Monte Carlo
tree search to perform individual agent reasoning in many-
agent open environments. Theoretical analyses establish the
number of agents to model in order to achieve acceptable
worst case bounds on extrapolation error, as well as regret
bounds on the agent's utility from modeling only some neigh-
bors. Simulations of multiagent wildfire suppression prob-
lems demonstrate our approach's efficacy compared with al-
ternative baselines.
Introduction
In both cooperative and competitive multiagent systems
(MAS), a participating agent benefits from reasoning about
how other agents will behave while choosing optimal actions
that maximize its chances of accomplishing shared or self-
interested goals. However, nuances in real-world environ-
ments often challenge straightforward peer modeling. One
of these is agent openness occurring whenever individual
agents join or leave the system (temporarily or permanently)
over time. For example, cooperative robots tasked with sup-
pressing wildfires alongside or in place of human fire fight-
ers would need to periodically leave the environment to
Copyright c(cid:13) 2020, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
recharge their limited suppressants that were spent during
firefighting. Likewise, competitive autonomous ride-sharing
cars can no longer compete for new passengers while trans-
porting a full ride. Consequently, openness requires that an
agent not only predict what actions their neighbors will take,
but also whether they are even present to take actions. Inac-
curately predicting the presence or absence of others may
cause agents to choose actions that miss their intended ben-
efits or other utility maximizing opportunities.
Furthermore, real-world environments such as wildfire
suppression often involve many agents of various types
working together to put out the wildfires. Thus, the model-
ing and reasoning must additionally scale with the number of
agents. As scalability is a general and ever-present challenge
for multiagent planning, investigations have first focused on
multiagent planning for smaller groups of agents, e.g., (Am-
ato and Oliehoek 2015; Claes et al. 2017; Hoang and Low
2013; Nair et al. 2005; Oliehoek et al. 2008). Subsequently,
there has been some focus on scalable planning in large
groups of agents (Oliehoek, Whiteson, and Spaan 2013;
Nguyen, Kumar, and Lau 2017; Sonu, Chen, and Doshi
2017; Velagapudi et al. 2011) albeit, in the absence of open-
ness. Indeed, state-of-the-art multiagent planning in open
environments has demonstrated successful planning with up
to 5 agents (Chandrasekaran et al. 2016; Cohen, Dibangoye,
and Mouaddib 2017). Of course, openness compounds the
challenges of scalable reasoning since not only should the
agent model how its many peers decide their actions, but
also how they dynamically join and leave the environment.
In this paper, we focus on achieving scalable multiagent
reasoning in open environments. We improve on a previous
approach to individual agent planning in the context of agent
openness (Chandrasekaran et al. 2016) to consider many-
agent settings. To promote flexible scalability, our method
enables a subject agent to intelligently select a small subset
of neighbors to explicitly model. Then, it extrapolates their
expected behaviors to the larger set of all agents. This ap-
proach has its roots in surveying and polling theory where
surveyors rarely poll all individuals in their target class. As
such, this approach represents a novel integration of survey
theory in multiagent planning. However, selective modeling
may not be enough for many-agent systems. We observe that
in many open domains such as wildfire suppression, which
particular firefighter is performing an action is not relevant to
the decision making. Similar observations have been made
in other domains (Velagapudi et al. 2011; Nguyen, Kumar,
and Lau 2017; Sonu, Chen, and Doshi 2017). Consequently,
further efficiency is made possible by exploiting the property
of anonymity of the other agents (Sonu, Chen, and Doshi
2017). These modeling approaches are utilized in a new gen-
eralization of Monte Carlo tree search (Silver and Veness
2010) to plan in open, many-agent settings.
Our results in this paper are both theoretical and exper-
imental. We analytically establish (i) the number of neigh-
bors that a subject agent must model to achieve given worst
case extrapolation errors, and (ii) the regret bounds on the
agent's approximate utility function from modeling some
neighbors only. To empirically demonstrate scalability and
evaluate the benefits of our approach, we extend benchmark
simulations of wildfire suppression (Chandrasekaran et al.
2016) to include setups with up to 50 agents -- an order of
magnitude larger than the prior literature -- capable of per-
forming up to 1.27 × 1030 joint actions, for which existing
decision-theoretic planning does not scale.
Background
i,l
We review the planning framework and a previous approach
for integrating anonymity into reasoning about others.
I-POMDP-Lite Framework
A framework for individual agent planning in partially ob-
servable MAS is the I-POMDP-Lite (Hoang and Low 2013).
This framework retains many of the benefits of the more
general I-POMDP framework but mitigates computational
costs by modeling the other agents' reasoning processes as
an approximation of their true decision making. Formally,
(cid:44) (cid:104)Ag, S, A, Ωi, Ti, Oi, Ri, γ, bi,0,
{Mj,l−1,Mk,l−1, . . . ,Mz,l−1}(cid:105)
IPOMDP-LiteL
• Ag is a set of agents, consisting of a subject agent i using
the I-POMDP-Lite to decide how to act and other agents
j, . . . , z modeled by subject i. Each agent has a frame θ
from the discrete set Θ = {θ1, . . . , θΘ}. A frame repre-
sents the agent's capabilities and preferences.
• S is the set of states of the decision-making problem, pos-
sibly factored into variables F1×F2×. . .×Fk, such as the
intensities of the k wildfires the agents need to suppress.
• A = Ai × Aj × . . . × Az is the set of possible joint
actions of the agents; e.g., the individual fires that each
agent chooses to fight. For notational convenience, a−i ∈
Aj × . . . × Az denotes the vector of actions by Ag \ {i}.
• Ωi is the set of observations of agent i.
• Ti(s, ai, a−i, s(cid:48)) = P (s(cid:48)s, ai, a−i) gives the probabili-
• Oi(s(cid:48), ai, a−i, oi) = P (oiai, a−i, s(cid:48)) models the prob-
abilities of stochastic observations revealed to subject
agent i after joint action (ai, a−i).
• Ri(s, ai, a−i) ∈ R is the reward function of agent i de-
ties of stochastic state transitions caused by actions.
pendent on the state and joint actions.
state of i, respectively.
• γ ∈ (0, 1] and bi,0 are the discount factor and initial belief
• {Mj,l−1,Mk,l−1, . . . ,Mz,l−1} is the set of mental
models ascribed to the other agents j, k, . . ., z at level
l − 1. 1 Each mental model Mj,l−1 is a Nested-MDP.
Hoang and Low (2013) define a Nested-MDP as:
Mi,l (cid:44) (cid:104)S, A, Ti, Ri,{πj,d, πk,d, . . . , πz,d}l−1
where {πj,d, πk,d, . . . , πz,d}l−1
d=0 is the set of policies fol-
lowed by other agents j, k, . . ., z at levels 0 to l − 1. The
policies are obtained by solving their own Nested-MDPs.
In solving an I-POMDP-Lite, an agent chooses actions that
maximize the cumulative, discounted reward function over
a finite horizon H, r0 + γr1 + γ2r2 + . . . + γH−1rH−1, by
considering Bellman equations for each belief/action pair:
d=0, γ(cid:105)
Qt
i,l(bi, ai) = ρi(bi, ai) + γ
T ai,oi
i
(cid:48)
(s
, oibi, ai)V t−1
i,l
(cid:48)
i)
(b
(cid:88)
s(cid:48),oi
(1)
(2)
(3)
V t
i,l(b
(cid:48)
i) = max
ai∈Ai
i,l(bi, ai)
Qt
(cid:89)
where
ρi(bi, ai) =
(cid:88)
(cid:88)
s∈S
a−i∈A−i
π−i,l−1(s, a−i)
−i∈{j,k,...,z}
× Ri(s, ai, a−i) bi(s)
s∈S
P r(s(cid:48)oi, ai, a−i, bi) ∝ Oi(s(cid:48), ai, oi)(cid:80)
Policies π−i,l−1(s, a−i) are solutions of the other agents'
mental models M−i,l−1, and b(cid:48)
i denotes the updated belief
Ti(s, ai, a−i, s(cid:48))
bi(s). Hoang and Low (2013) establish regret bounds on the
value function V based on any error introduced through ap-
proximating other agents as reasoning using a Nested-MDP
(instead of their true reasoning process).
Both the I-POMDP-Lite and decentralized POMDP (Dec-
POMDP) (Bernstein et al. 2002) have been proposed for rea-
soning in open environments (Chandrasekaran et al. 2016;
Cohen, Dibangoye, and Mouaddib 2017). The I-POMDP-
Lite (as a relative of the more general I-POMDP) differs
from the Dec-POMDP in several important ways. First, the
I-POMDP-Lite makes no assumptions that a subject agent
observes the actual actions taken by other agents during op-
eration in the environment, nor the observations they re-
ceive. Indeed, each agent can reason in a vacuum, infer-
ring the behaviors of others through changes to the shared
environment state, whereas Dec-POMDPs assume that ev-
eryone's actions and observations are communicated or ob-
served. This makes the I-POMDP-Lite appropriate for a
broader range of cooperative (and even competitive) envi-
ronments, including the wildfire suppression domain where
agents might not be able to pre-coordinate their behav-
iors, directly observe other agents obscured by smoke and
fire, nor share information during operation due to damaged
communication channels.
1The level of a reasoning process represents the level in a hi-
erarchy of agents reasoning about their neighbors: level 1 implies
all others choose actions randomly, and level l > 1 implies that all
others reason at level l − 1.
Frame-Action Anonymity
In many domains, it does not matter which agents perform
actions, rather how many agents of each frame perform each
action. By relaxing agent identities, multiple joint actions
may produce the same action counts, and, in turn, equiva-
lent state transitions, observations, and rewards. For exam-
ple, in wildfire suppression, agents of different frames (e.g.,
ground firefighters, helicopters) extinguish fires at different
rates. But, each agent of the same frame contributes the same
amount of fire suppression. Therefore, any k agents of the
same frame identically impact the fire. Such domains exhibit
frame-action anonymity (Sonu, Chen, and Doshi 2017).
Formally, let configuration C = (cid:10)na1,θ1 , ..., naA,θΘ
where na1,θ1 is the number of agents with frame θ1 per-
forming action a1 in a joint action tuple (ai, a−i). Histories
with configurations (rather than joint actions) are now suffi-
cient statistics (Sonu, Chen, and Doshi 2017) and the repre-
sentation of Ti, Oi, and Ri can be greatly compacted since
the number of possible configurations C ∝(cid:0)Ag+m+1
(cid:1) is
(cid:11),
m+1
polynomial in the number of agents and thus much fewer
than the exponential number of possible joint actions mAg,
where m = max{Ai,Aj, . . . ,Az} (c.f., Figure 1 below
for comparisons of the number of configurations vs joint ac-
tions in our experiments). Therefore, configurations enabled
by frame-action anonymity improve scalability of individ-
ual planning in many-agent environments, as shown in the
context of I-POMDPs (Sonu, Chen, and Doshi 2017).
Notably, in decision-theoretic planning, configurations
are random variables since agents do not directly observe
the actions chosen by other agents, but instead the likeli-
hoods of different configurations (i.e., counts of actions per
frame) are estimated based on the policies π−i,l−1 solved
for each other agent's mental model Mj,l−1,··· , Mz,l−1.
Many-Agent Planning under Openness
In this paper, we limit our attention to systems where agents
may leave the environment at any time and possibly reen-
ter, but new agents do not enter the system. Still, this brings
unique conceptual and computational challenges.
Challenges of Open Agent Systems
In open systems, individual planning is complicated by the
need of each agent to track which other agents are currently
present in the system and to reason about the actions of
present agents only. In wildfire suppression, each firefighter
must know how many others are currently unavailable due to
recharging their suppressant, so as to focus on the behaviors
of those currently fighting the fires.
Let N (i) ⊆ Ag be the neighborhood of subject agent
i, defined as the set of other agents with which i can in-
teract in the environment and that affect its transition, ob-
servation, and reward functions (e.g., the other firefighting
robots that can suppress the same wildfires as subject agent
i). A naive way of tracking the presence of agents in open
environments involves each agent i maintaining an addi-
tional state variable Pj for each neighbor j ∈ N (i), where
Pj ∈ {present, absent}. This increases the state space by
a factor of 2N (i) -- an exponential increase in the size of
the agent's state space! To illustrate, modeling the presence
or absence of 40 neighbors in this way increases the state
space by a factor of 240 (cid:29) 100 billion states. To avoid this,
we place Pj as an internal state variable within a mental
model Mj,l−1 attributed by subject agent i to neighbor j.
Consequently, the overall increase in the size of the problem
is linear in the number of agents (scaled by the number of
models ascribed to each agent), which promotes scalability.
Nonetheless, even a linear increase in the problem
size poses challenges for planning. In offline planning
methods such as the interactive point-based value it-
eration (Chandrasekaran et al. 2016) and interactive
expectation-maximization (Qu and Doshi 2015), the subject
agent must still consider each neighbor's current presence
and future transition, which adds at least a polynomial co-
efficient O(N (i)2) to the complexity of planning. Like-
wise, in online planning approaches such as Monte Carlo
tree search algorithms for POMDPs and I-POMDPs (Claes
et al. 2017; Hula, Montague, and Dayan 2015), the subject
agent must spend time simulating the transitions of each
neighbor when sampling an environment outcome, resulting
in fewer sampled trajectories in the tree and thus lower qual-
ity approximations of the value function V . Overall, these
challenges motivate further steps to alleviate the impact of
openness on computational complexity.
Selectively Modeling Neighbors
It is well known that surveyors and pollsters do not model
every target person individually to understand their attitudes
and behaviors. Instead, the collective attitudes and behav-
iors of the whole are estimated by sampling those of a
few, then extrapolating to everyone. Bounds on the approx-
imation error of this approach can be established through
statistical analysis and random sampling (Neyman 1934;
Frankel and Frankel 1987; Lohr 2010).
Adapting this methodology to open, many-agent settings,
we maintain mental models for a small subset of an agent's
neighbors, then extrapolate their predicted behaviors to the
entire neighborhood (with bounded error). In the wildfire
example, each firefighter models some of the firefighters
with which it interacts, then uses their (fewer) predicted be-
haviors to estimate how many of the full set of firefight-
ers are recharging or choosing to fight each fire. By ex-
plicitly modeling some neighbors only, we reduce the rea-
soning necessary for considering changes to every neigh-
bors' presence. Of course, the agent's planning process no
longer obtains a policy for every neighbor and thereby esti-
mate the distribution of configurations necessary to estimate
the state transitions, observations, and rewards (inside the
(s(cid:48), oibi, ai) of Eq. 1 and product in Eq. 3). To miti-
T ai,oi
i
gate this trade off, we use an alternative way to estimate con-
figuration likelihoods when modeling a subset of the subject
agent's neighborhood.
Estimating configuration probabilities Recall that dur-
ing planning, configurations C are random variables whose
probability distributions typically depend on the policies ob-
tained by the subject agent for its neighbors. Alternately,
since C =(cid:10)na1,θ1, ..., naA,θΘ
(cid:11) is composed of the counts
of the number of agents of each frame performing each ac-
tion, the distribution over C can also be represented as a
multinomial distribution, parameterized as
P (Cst, M t) ∼ M ulti(N (i),{pa1,θ1,N (i), . . . , paA,θΘ,N (i)})
where pa,θ,N (i) is the probability that neighbors in N (i) of
frame θ will perform action a. We may view the configura-
tion C as the concatenation of several multinomial variables,
Cθ =(cid:10)na1,θ, na2,θ, . . . , naA,θ
(cid:11), one for each frame θ. This
allows us to model the multinomial distribution over C using
separate multinomial distributions for each Cθ,
P (Cθst, M t) ∼ M ulti(Nθ(i),{pa1,Nθ (i), . . . , paA,Nθ (i)})
(4)
where Nθ(i) = {j ∈ N (i)j has frame θ} is the set of i's
neighbors with frame θ.
If an agent chooses to model a proper subset of its neigh-
bors Nθ(i) ⊂ Nθ(i), it can still estimate for each action
the pa,Nθ(i) values that parameterize the multinomial dis-
tribution P (Cθst, M t). Let nπ(st)=a, Nθ(i) be the number
of agents in Nθ(i) predicted to perform action a at current
state st based on the policy obtained by solving their Nested-
MDP model and their presence variable Pj ∈ mj,l−1. Then
pa, Nθ(i) =
nπ(st)=a, Nθ(i)
Nθ(i)
(5)
is how likely an arbitrarily modeled neighbor of frame θ will
perform action a.
Proportions pa, N (i) estimated using Eq. 5 allow us to es-
timate the underlying multinomial distribution over config-
uration likelihoods (Eq. 4). We can then estimate the proba-
bility of a given configuration C using the probability mass
function of the modeled multinomial distribution.
How many neighbors to model? Equations 4 and 5 are
crucial in that they determine how we extrapolate the be-
haviors of a few modeled agents to the entire neighborhood
without requiring explicit mental models and policies for
every neighbor. At the same time, this approach approx-
imates the true distribution of configurations, which natu-
rally raises questions such as how many neighbors should a
subject agent model and how does the number of neighbors
impact the error in the estimated configuration probabilities
based on extrapolation. We relate these two questions in the
following theorem and corollary. As Hoang and Low (2013)
previously established bounds on the error introduced by as-
suming that neighbors reason about others using a Nested-
MDP, we focus our attention on errors introduced by mod-
eling a subset of the neighbors.
Let the acceptable worst-case error in each of the pa, Nθ(i)
proportions in Eqs. 4 and 5 be e p. We refer to this henceforth
as the extrapolation error. From statistical analysis (Lohr
2010), we establish the minimum number of neighbors
which an agent needs to explicitly model so that the extrap-
olation error does not exceed the given bound:
Theorem 1 (Number of modeled neighbors). Let Nθ(i) be
a neighborhood of agents with frame θ and whose size is N,
e p be a desired bound on extrapolation error, (1 − α) be a
come from the Stu-
statistical confidence level, and tn−1, α
dent's t-distribution with n − 1 d.o.f. Then if agent i models
2
(cid:17)2
(cid:16) tn−1, α
(cid:16) tn−1, α
2e p
2
2
2e p
N
N − 1 +
(cid:17)2
(6)
nθ = Nθ(i) ≥
neighbors, then it will be confident at the (1 − α) level that
pa, Nθ(i) for each action a will be within e p of the true pro-
portions of all agents choosing action a.
The proof of this theorem is provided in Appendix A of
the supplementary material. To illustrate, let N = 50. To en-
sure that the agent's estimated pa, Nθ(i) extrapolates to within
e p = 0.1 of the true proportion for the entire neighborhood
with frame θ with 95% confidence, we need model at least
34 neighbors. This drops to 18 if instead e p = 0.2.
If the agent indeed models at least the prescribed number
of other agents from Theorem 1, the error in each configura-
tion's probability estimate is bound by the value P (C) given
below (proof in Appendix B of the supplementary material).
Corollary 1 (Error bound on configuration probability ). Let
nθ be the number of neighbors given by Theorem 1 (for a
given confidence level 1 − α) that subject agent i chooses
to model from its neighborhood Nθ(i) for each θ ∈ Θ,
and let pa,θ be the resulting estimated proportions of agents
within those neighborhoods that will choose action a, given
state s and mental models M. Then the estimated probability
P (Cst, M t) that the entire neighborhood will exhibit the
configuration C has error P (C) due to modeling nθ agents
only, which is less than:
P ∗(Cst, M t) − P (Cst, M t) = P (C)
(cid:81)
(cid:81)
θ Nθ(i)!
a,θ C(a, θ)!
<
(cid:89)
(pa,θ + e p)C(a,θ) −(cid:89)
a,θ
a,θ
pC(a,θ)
a,θ
(7)
where P ∗(Cst, M t) denotes the true likelihood that config-
uration C will result from state st and mental model M t.
Regret Bounds As modeling a subset of neighbors pro-
duces a bounded approximation of the true configuration
probabilities considered during planning, the resulting value
function V will also be an approximation. Next, we estab-
lish that any error in the approximate value function (and
hence the discounted cumulative rewards actually earned
by the agent following the corresponding policy πi) is also
bounded. Here, Ji,k represents the actual discounted cumu-
lative rewards the agent would earn following πi for k time
steps, whereas V ∗
Theorem 2. (Regret bound). Maximum regret that agent i
from following a k-horizon optimal
incurs
policy πi (obtained by solving the many-agent I-POMDP-
Lite) due to the approximate likelihoods of other agents'
configurations P (Cst, M t) is bounded from above:
i,k is the optimal value function.
(cid:13)(cid:13)(cid:13)V ∗
i,k − Ji,k
(cid:13)(cid:13)(cid:13)∞
(cid:13)(cid:13)V ∗
i,k − Ji,k
(cid:13)(cid:13)∞ ≤ 2P (C) · C · Rmax
(cid:18)
(cid:20)
×
γk−1 +
1
1 − γ
(cid:19)(cid:21) (8)
Ωi
1 − γ
1 + 3γ
√
√
Notice that the bound is linear in the error P (C) in the
agent's estimation of configuration likelihoods caused by
modeling some neighbors only. The proof is provided in Ap-
pendix C of the supplementary material. Though the result-
ing value function from planning might not be exact, the ap-
proximation error is upper bounded and is proportional to
only
nθ (in the worst case) due to the fact that P (C) is
at worst linear in e p (c.f., Eq. 7 since e p < 1), and e p is
proportional to
Which neighbors to explicitly model? After determining
the number of agents to model for each frame-action pair,
we must select which neighbors will be explicitly modeled,
Nθ(i). This selection is performed only when the agent joins
the environment. Thus, it does not add to the computational
complexity of planning.
nθ (c.f., Eq. 6).
To decide which neighbors to model, we first note
that we can decompose the agent's neighborhood Nθ(i)
into (potentially overlapping) sets based on the actions
the neighbors can perform: Nθ,a(i) = {j ∈
that
Nθ(i)j can perform action a}. Each Nθ,a(i) provides a
sample set from which we select neighbors in order to es-
timate pa, Nθ(i). To construct Nθ(i), we randomly sample n
agents (Eq. 6) for each (θ, a) pair.
Second, many agents may appear in more than one such
sample set because agents can perform more than one action.
Thus, once a neighbor j has been added to Nθ(i), it con-
tributes to the n needed for each (θ, a) pair corresponding to
the actions that agent j can perform. If sampling neighbors
starts with the largest Nθ,a and proceeds to the smallest, then
agents that have the greater number of possible actions (e.g.,
can fight the most fires) will have the largest total chance of
being sampled for modeling since they appear in most sub-
sets Nθ,a(i). This property is beneficial as such agents are
expected to be the most influential agents in the system, as
well as the most necessary to explicitly model because their
reasoning will be more complex than other agents, and thus
their behaviors are more difficult to predict.
Many-Agent MCTS for Open Systems
Based on our approach for modeling a subset of an agent's
neighbors, we next extend the popular POMCP algorithm
(Silver and Veness 2010) to create a MCTS algorithm for
many-agent open environments.
Single-agent MCTS A popular approach to quickly ap-
proximate the value function and policy of a POMDP is
Monte Carlo tree search (MCTS) (Ross et al. 2008; Sil-
ver and Veness 2010; Somani et al. 2013). In MCTS, the
agent iteratively constructs an AND-OR tree representing
the possible beliefs that can be reached by following se-
quences of actions, and obtains the expected discounted util-
ity of performing those action sequences by simulating fu-
ture rewards. MCTS performs online planning for the be-
liefs that an agent actually encounters, avoiding the need
for computationally-expensive offline planning and scales
well to large state spaces due to MC sampling of future
state-action-reward trajectories. Until the time budget is met,
MCTS revises estimated Q(b, a) values along the trajecto-
ries and extending the tree at the leaf using rollout.
Many-agent MCTS We generalize MCTS to open, many-
agent settings in the context of the I-POMDP-Lite frame-
work (also applicable to I-POMDPs), exploiting both frame-
action anonymity and selective neighbor modeling to further
enhance scalability. Our algorithm, shown in Algorithm 1,
is I-POMCPO. It considers agent openness in (i) Simulate,
which additionally simulates the agents leaving and reenter-
ing (in M), and in (ii) SampleConfiguration when obtain-
ing configuration probabilities.
Our I-POMCPO algorithm differs from the classic sin-
gle agent POMCP (Silver and Veness 2010) in the follow-
ing ways. First, the agent's rewards Ri, state transitions Ti,
and observations Oi depend on not only its own actions and
the environment state, but also the configuration of actions
by the other agents. Thus, before the agent can simulate the
environment on line 6 of the Rollout and line 11 of the Up-
dateTree procedures, the agent must sample a configura-
tion C t of its neighbors' actions. This addition to standard
POMCP occurs on lines 4 and 9 of the two procedures.
Second, the process for sampling a configuration exploits
the multinomial probability distribution P (C tst, M t) de-
termined from the models of a select subset of the agent's
neighbors in the SampleConfiguration procedure, so that
the agent need not model every neighbor. Since only some
mental models are consulted, there is opportunity to save
computation time in this procedure, as opposed to modeling
every neighbor2. Moreover, since only some mental mod-
els are needed by this procedure, the Simulate procedure
modeling the environment (called on lines 6 and 11 of Roll-
out and UpdateTree, respectively) needs to process fewer
mental model transitions, saving additional computational
time. Overall, this should require less time per trajectory
through the AND-OR tree in each call of UpdateTree from
the root of the tree, and thus more total trajectories sampled
and closer approximations of the true value function.
This algorithm significantly differs from previous MCTS
algorithms for multiagent POMDPs. Hula et al. (2015) first
adapted POMCP to an I-POMDP setting. However, they
made several assumptions in their adaptation to their particu-
lar problem: the presence of only two agents, agents' actions
occur sequentially rather than simultaneously, and agents di-
rectly observe each other actions. Their approach would not
scale to many-agent environments as it would have com-
putational complexity that is exponential in the number of
agents. Our approach applies to more general settings, as-
suming only frame-action anonymity, and exhibits complex-
ity that is linear in the number of agents: O(IHN ) where
I is the number of sampled MC trajectories, H is the hori-
zon, and N is the total number of modeled agents. Amato
2Provided that sampling an action for each neighbor is less
computationally expensive than interacting with an actual policy.
Algorithm 1 I-POMCPO: Open Many-Agent MCTS
Note: T is the tree (initially empty), p is a path from the root of
the tree (with p = ∅ signifying the root), Bp is the particle filter
signifying the set of state-model pairs encountered at the node at p
in the tree, P F is the root particle filter, N is count of the number
of visits to each node in the tree initialized to some constant ν ≥ 0,
Q is the Q function initialized to 0, c a constant from UCB-1.
1: procedure I-POMDP-MCTS(P F, τ)
2:
3: while time < τ do
4:
time ← 0
s0, M 0 ← SampleP article (P F )
5: UpdateTree(cid:0)s0, M 0, 0,∅(cid:1)
6:
7:
Increment time
return argmax
Q(∅, a)
a∈Ai
10:
i
i
i, ot
i
i, rt
1: procedure UPDATETREE(st, M t, t, p)
2:
3:
if t ≥ H then
return 0
if p /∈ T then
T ← T + leafnode(p)
4: Bp ← Bp ∪ {(cid:0)st, M t(cid:1)}
return Rollout(cid:0)st, M t, t(cid:1)
9: C t ← SampleConfiguration(cid:0)st, M t(cid:1)
5:
6:
7:
8:
else
i) ← Q(p, at
i) + R − Q(p, at
p(cid:48) ← p +(cid:0)at
Q (p, a) + c(cid:112)(log Np) /Np(cid:55)→a
i, C t(cid:1)
i ← Simulate(cid:0)st, M t, at
(cid:1)
← 1 + Np(cid:55)→at
i + γ·UpdateTree(cid:0)st+1, M t+1, t + 1, p(cid:48)(cid:1)
i ← argmax
at
a∈Ai
11:
st+1, M t+1, ot
12: Np ← 1 + Np
13: Np(cid:55)→at
14:
15: R ← rt
16: Q(p, at
17:
return R
1: procedure ROLLOUT(st, M t, t)
2: R ← 0, t(cid:48) ← t
3: while t < H do
4: C t ← SampleConfiguration(st, M t)
5:
6:
i, rt
7: R ← R + γt−t(cid:48) · rt
8:
1: procedure SAMPLECONFIGURATION(st, M t)
2: C(a, θ) ← 0, nπ(st)=a, Nθ (i) ← 0 ∀a, θ
3:
4:
5:
i ← Simulate(st, M t, at
i, t ← t + 1
i ← SampleAction(Ai)
at
st+1, M t+1, ot
i)/Np(cid:55)→at
return R
i, C t)
i
(i) + 1
(i) ← nπ(st)=a, Nθj
for Mj,l−1 ∈ M t do
a ∼ πj,l−1(st)
nπ(st)=a, Nθj
for θ ∈ Θ do
for a ∈ A do
pa, N (i) ← nπ(st)=a, Nθ (i) / Nθ(i)
for j ∈ Nθ(i) do
a ∼ Cat(pa1, Nθ
, pa2, Nθ
C(a, θ) ← C(a, θ) + 1
return C
, . . . , paA, Nθ
)
6:
7:
8:
9:
10:
11:
12:
and Oliehoek (2015) and Best et al. (2019) proposed simi-
lar algorithms for Dec-POMDPs that scale to more than two
agents. However, our algorithm enables planning in settings
when the joint actions of other agents are unknown and must
be estimated through individual models, as well as incorpo-
rates frame-action anonymity and selective neighbor mod-
eling for improved scalability. Claes et al. (2017) proposed
an algorithm that simplified the models of neighbor behav-
ior using domain heuristics to enhance scalability, as well as
social laws to help promote cooperation. Here, however, we
consider more general models of neighbor behavior that will
work across domains without specific heuristics.
Experiments
We evaluate implementations3 of I-POMCPO on the wild-
fire suppression problem. Chandrasekaran et al. (2016) used
this domain to study planning in small open agent environ-
ments; we extend it to include more fires and significantly
more firefighters and corresponding joint actions.
Setups Agents are tasked with putting out fires of dif-
ferent sizes in the absence of inter-agent communication
and prior coordination. Small, large, and huge fires require
at least 10, 20, and 30 agents to act together to reduce
their intensity, respectively. The spread of fires is modeled
on the dynamics of real wildfires (Boychuk et al. 2009;
Ure et al. 2015). Agents have limited amounts of suppres-
sants that stochastically transition from full to half-full to
empty, and back to full on recharging during which they tem-
porarily leave the environment for an average of two steps,
then rejoin when full. Each agent can use its suppressant on a
fire in an adjacent location or take a NOOP action; the latter
also represents actions to take while recharging suppressant.
Agents earn shared rewards of 20, 40, and 60 for putting
out small, large, and huge fires. They receive shared penal-
ties of 1 when a location burns out and individual penalties
of 100 for either fighting a non-existent fire or choosing a
non-NOOP action without suppressant. Figure 1 illustrates
the environments used in our experiments. Notice that our
agents are ground firefighters and helicopters, whose frames
differ: helicopters are twice as effective in their firefighting
ability. Though several agents share the same location, they
start with different suppressant levels. As such, they plan and
behave differently rather than as one collective agent.
Though in seemingly small grids, these decision prob-
lems are non-trivial and substantially complex, especially
compared to the prior literature -- with 40-50 agents, the
number of joint actions ranges from 2.95 × 1021 to 1.27 ×
1030. Even with frame-action anonymity, there are 15,256 to
9,662,576 configurations; thus, these setups are intractable
for prior planning algorithms. We intentionally do not con-
sider smaller domains for which existing algorithms are
tractable (e.g., with only 5 agents (Chandrasekaran et al.
2016)), as there would be no need for modeling only a few
neighbors (Thm. 1 would model every neighbor ∀e p < 1).
Metrics and baselines Agent performance is measured
by: (i) the average rewards earned by agents, (ii) the av-
erage number of fire locations put out, and (iii) the average
amount of suppressant used. These measures evaluate how
how well agents optimize their planning problems, and how
effectively and efficiently they put out fires, respectively.
3Available at https://github.com/OberlinAI/ScalableOASYS
Figure 1: Our environment setups involve a varying number
of fires, fire intensities and positions, agents and their types.
Figure 2: Total reward earned per agent averaged across 100
runs. Error bars represent 95% CIs. I-POMCPO planned for
5 sec. in Setups 1-3 and 5, and for 20 sec. in Setup 4.
To demonstrate both the efficacy and impact of our strat-
egy of intelligently modeling only a subset of neighbors,
we consider variants of I-POMCPO, each planning at level
l = 2, that use different e p ∈ [0, 0.3] with Theorem 1 pro-
viding the numbers of neighbors to model (with α = 0.05).
Here, e p = 0 represents I-POMCPO modeling every neigh-
bor and also provides an upper bound on how the algorithm
of Hula et. al (2015) would perform if generalized to many-
agent environments and Amato and Oliehoek (2015) if ex-
tended to the I-POMDP-Lite with frame-action anonymity.
We compare the performance of our approach with two
baselines (using Kruskal-Wallis followed by post-hoc pair-
wise Mann-Whitney for significance testing). First, Nested-
MDP represents an agent using an adaptation of VI to per-
form -bounded planning over configurations at level l = 1.
This is a strong baseline, which serves as both a scalable,
fully observable approximation that can tractably solve our
decision problems, as well as a comparison to the models
that agent i assumes its neighbors follow. Second, Heuristic
is an agent that randomly chooses an adjacent fire to fight
if one exists and the agent has suppressant, else it takes
a NOOP action. It represents semi-intelligent behavior in
complex environments (similar to domain heuristics in mul-
tirobot dynamic task allocation, e.g., (Lerman et al. 2006)).
Of note, traditional I-POMDP solvers do not scale to many-
agent environments. Indeed, we ran the previous method, in-
teractive point-based value iteration (Chandrasekaran et al.
2016), on Setup 1 with only 10 total agents; after almost a
week of planning, it did not produce a usable policy.
Results We observe from Fig. 2 that I-POMCPO improved
on NestedMDP and Heuristic by achieving significantly
greater rewards in Setups 1-3 and 5, and statistically equiva-
lent reward to NestedMDP in Setup 4. Further investigations
(c.f., Figs. 4 and 5 in Appendix D of the supplementary ma-
terial) reveal that I-POMCPO agents used slightly more sup-
pressant to put out fires at significantly more locations than
NestedMDP (true in all setups). NestedMDP agents earned
comparable rewards in Setup 4 by concentrating all of their
efforts on a single shared fire, whereas I-POMCPO agents
spread their resources across the most complex environment.
Interestingly, in Setup 5, which had the most shared fires,
NestedMDP never put out a fire. This was due to the symme-
try in the problem causing NestedMDP to behave randomly,
whereas I-POMCPO was more successful.
Comparing the different settings of I-POMCPO, increas-
ing e p so that fewer neighbors were modeled led to in-
creased performance in the more complicated Setups 2-5.
Moreover, agents were also more efficient with their sup-
pressants, using less resources to put out more fires as they
modeled fewer neighbors. These results were obtained be-
cause MCTS could then run a greater number of trajectories
in the fixed planning time τ, in essence trading off modeling
accuracy for better value estimates. Furthermore, in Setup
5, modeling only some neighbors biased agents to favor
one shared fire over another, breaking the complicated sym-
metry of the problem. On the other hand, modeling fewer
neighbors reduced performance in only the simplest Setup 1,
which had only one shared fire and one frame of agents. In
this setup, even when modeling all agents, enough trajecto-
ries were sampled to adequately estimate the value function.
Concluding Remarks
Real-world domains often exhibit agent openness, where
agents may leave and then return. Scaling in the number of
agents in such domains continues to remain a challenge for
planning in realistic multiagent environments. Our method
-- consisting of an approach to selectively model neighbors
that is new to planning and a generalized MCTS algorithm
for many-agent settings -- models the presence or absence
of agents as it predicts their behaviors. Our key insight for
scaling is that we may explicitly model just a few agents and
extrapolate the model predictions to others with provable er-
ror bounds. This improves on baselines as we demonstrated
on a spectrum of firefighting scenarios with an order of mag-
nitude more agents than previously considered in open envi-
ronments. The idea of extrapolating modeling is new and not
limited to open agent settings. It could be combined in future
work with other approximations that promote scalability.
Acknowledgments
This research was supported in part by NSF grant #1909513
(to AE, PD, and LS). Experiments were conducted on Ober-
lin College's SCIURus cluster (NSF Grant #1427949).
References
[Amato and Oliehoek 2015] Amato, C., and Oliehoek, F. A.
2015.
Scalable planning and learning for multiagent
POMDPs. In Twenty-Ninth AAAI Conference on Artificial
Intelligence.
[Bernstein et al. 2002] Bernstein, D. S.; Givan, R.; Immer-
man, N.; and Zilberstein, S. 2002. The complexity of decen-
tralized control of Markov decision processes. Mathematics
of Operations Research 27(4):819 -- 840.
[Best et al. 2019] Best, G.; Cliff, O. M.; Patten, T.; Mettu,
R. R.; and Fitch, R. 2019. Dec-mcts: Decentralized planning
for multi-robot active perception. The International Journal
of Robotics Research 38(2-3):316 -- 337.
[Boychuk et al. 2009] Boychuk, D.; Braun, W.; Kulperger,
R.; Krougly, Z.; and Stanford, D. 2009. A stochastic forest
fire growth model. Environmental and Ecological Statistics
16(2):133 -- 151.
[Chandrasekaran et al. 2016] Chandrasekaran, M.; Eck, A.;
Doshi, P.; and Soh, L. 2016. Individual planning in open
In Proceedings of the Thirty-
and typed agent systems.
Second Conference on Uncertainty in Artificial Intelligence,
UAI'16, 82 -- 91. Arlington, Virginia, United States: AUAI
Press.
[Claes et al. 2017] Claes, D.; Oliehoek, F.; Baier, H.; and
Tuyls, K. 2017. Decentralised online planning for multi-
robot warehouse commissioning. In Proceedings of the 16th
Conference on Autonomous Agents and MultiAgent Systems,
AAMAS '17, 492 -- 500.
[Cohen, Dibangoye, and Mouaddib 2017] Cohen, J.; Diban-
goye, J.-S.; and Mouaddib, A.-I. 2017. Open decentralized
In 2017 IEEE 29th International Conference on
pomdps.
Tools with Artificial Intelligence (ICTAI), 977 -- 984.
[Frankel and Frankel 1987] Frankel, M. R., and Frankel,
L. R. 1987. Fifty years of survey sampling in the united
states. Public Opinion Quarterly 51(4):S127 -- S138.
[Hoang and Low 2013] Hoang, T. N., and Low, K. H. 2013.
Interactive POMDP lite: Towards practical planning to pre-
dict and exploit intentions for interacting with self-interested
In 23rd International Joint Conference on AI (IJ-
agents.
CAI).
[Hula, Montague, and Dayan 2015] Hula, A.; Montague, P.;
and Dayan, P. 2015. Monte Carlo planning method estimates
planning horizons during interactive social exchange. PLOS
Computational Biology 11(6).
[Lerman et al. 2006] Lerman, K.; Jones, C.; Galstyan, A.;
and Matar´ıc, M. J. 2006. Analysis of dynamic task alloca-
tion in multi-robot systems. Int. J. Rob. Res. 25(3):225 -- 241.
[Lohr 2010] Lohr, S. L. 2010. Sampling: Design and Analy-
sis. Boston, MA: Brooks/Cole, 2nd edition.
[Nair et al. 2005] Nair, R.; Varakantham, P.; Tambe, M.; and
Yokoo, M. 2005. Networked distributed POMDPs: A syn-
thesis of distributed constraint optimization and POMDPs.
In Twentieth AAAI Conference on Artificial Intelligence,
133 -- 139.
[Neyman 1934] Neyman, J. 1934. On the two different as-
pects of the representative method: The method of stratified
sampling and the method of purposive selection. Journal of
the Royal Statistical Society 97(4):558 -- 625.
[Nguyen, Kumar, and Lau 2017] Nguyen, D. T.; Kumar, A.;
and Lau, H. C. 2017. Collective multiagent sequential deci-
sion making under uncertainty. In Proceedings of the Thirty-
First AAAI Conference on Artificial Intelligence, AAAI'17,
3036 -- 3043.
[Oliehoek et al. 2008] Oliehoek, F. A.; Spaan, M. T.; White-
son, S.; and Vlassis, N. 2008. Exploiting locality of inter-
action in factored Dec-POMDPs. In 7th International Joint
conference on Autonomous agents and Multiagent Systems
(AAMAS), 517 -- 524.
F. A.;
[Oliehoek, Whiteson, and Spaan 2013] Oliehoek,
Whiteson, S.; and Spaan, M. T.
2013. Approximate
solutions for factored Dec-POMDPs with many agents. In
International Joint Conference on Autonomous Agents and
Multi-Agent Systems (AAMAS), 563 -- 570.
[Qu and Doshi 2015] Qu, X., and Doshi, P. 2015. Individual
planning in infinite-horizon multiagent settings: Inference,
structure and scalability.
In Cortes, C.; Lawrence, N. D.;
Lee, D. D.; Sugiyama, M.; and Garnett, R., eds., Advances
in Neural Information Processing Systems 28. Curran Asso-
ciates, Inc. 478 -- 486.
[Ross et al. 2008] Ross, S.; Pineau, J.; Paquet, S.; and Chaib-
draa, B. 2008. Online planning algorithms for POMDPs. J.
Artif. Int. Res. 32(1):663 -- 704.
[Silver and Veness 2010] Silver, D., and Veness, J. 2010.
In 23rd Interna-
Monte-carlo planning in large pomdps.
tional Conference on Neural Information Processing Sys-
tems (NIPS), NIPS'10, 2164 -- 2172.
[Somani et al. 2013] Somani, A.; Ye, N.; Hsu, D.; and Lee,
W. S. 2013. Despot: Online pomdp planning with regu-
larization. In Proceedings of the 26th International Confer-
ence on Neural Information Processing Systems - Volume 2,
NIPS'13, 1772 -- 1780. USA: Curran Associates Inc.
[Sonu, Chen, and Doshi 2017] Sonu, E.; Chen, Y.; and
Doshi, P.
Decision-theoretic planning under
anonymity in agent populations. J. Artif. Int. Res. 59(1):725 --
770.
[Ure et al. 2015] Ure, N.; Omidshafiei, S.; Lopez, B.; Agha-
Mohammadi, A.-A.; How, J.; and Vian, J. 2015. Online
heterogeneous multiagent learning under limited communi-
cation with applications to forest fire management. In Intel-
ligent Robotics and Systems (IROS), 5181 -- 5188.
[Velagapudi et al. 2011] Velagapudi, P.; Varakantham, P.;
Sycara, K.; and Scerri, P. 2011. Distributed model shap-
ing for scaling to decentralized POMDPs with hundreds of
In 10th International Conference on Autonomous
agents.
Agents and Multiagent Systems-Volume 3, 955 -- 962.
2017.
A Proof for Theorem 1
B Proof for Corollary 1
(cid:114)
Theorem 1 (Number of modeled neighbors). Let Nθ(i) be
a neighborhood of agents with frame θ and whose size is N,
e p be a desired bound on extrapolation error, (1 − α) be a
statistical confidence level, and tn−1, α
come from the Stu-
dent's t-distribution with n − 1 d.o.f. Then if agent i models
2
(cid:16) tn−1, α
(cid:17)2
(cid:16) tn−1, α
2e p
2
2
2e p
N
N − 1 +
(cid:17)2
(6)
nθ = Nθ(i) ≥
neighbors, then it will be confident at the (1 − α) level that
pa, Nθ(i) for each action a will be within e p of the true pro-
portions of all agents choosing action a.
Proof Let p denote pa, Nθ(i) and n denote Nθ(i). With
statistical confidence 1− α, the true proportion of neighbors
of frame θ who will choose action a will lie within the range:
p ± tn−1, α
2
p(1 − p)
n
(9)
2
However, the range in Eq. 9 is rather loose as it assumes the
neighborhood Nθ(i) is infinitely sized. For finite neighbor-
hoods, we can tighten this range using the finite population
correction (Lohr 2010). Let N = Nθ(i) be the size of the
finite neighborhood. Then the error bound is:
(cid:114) N − n
N − 1
(cid:114)
So, with confidence 1 − α the error e p of p is:
p ± tn−1, α
(cid:114)
(cid:114)
(cid:114) N − n
p(1 − p)
p(1 − p)
(cid:114) N − n
(10)
n
n
N − 1
N − 1
The last inequality results because the range in Eq. 10 at-
tains maximal width when p = 0.5. Reorganizing this in-
equality yields Eq. 6. (cid:3)
Of note, Fig. 3 demonstrates the relationship between n
and N in Theorem 1 for different values of e p for the stan-
dard 95% significance level.
.
e p ≤ tn−1, α
0.52
≤ tn−1, α
2
n
2
Note: in the following, we drop the timestep t notation as
our theoretical analysis does not depend on it, but instead
generalizes to all time steps.
Corollary 1 (Error bound on configuration probability ). Let
nθ be the number of neighbors given by Theorem 1 (for a
given confidence level 1 − α) that subject agent i chooses
to model from its neighborhood Nθ(i) for each θ ∈ Θ,
and let pa,θ be the resulting estimated proportions of agents
within those neighborhoods that will choose action a, given
state s and mental models M. Then the estimated probability
P (Cst, M t) that the entire neighborhood will exhibit the
configuration C has error P (C) due to modeling nθ agents
only, which is less than:
P ∗(Cs, M ) − P (Cs, M ) = P (C)
(cid:81)
(cid:81)
θ Nθ(i)!
a,θ C(a, θ)!
<
(cid:89)
(pa,θ + e p)C(a,θ) −(cid:89)
a,θ
a,θ
pC(a,θ)
a,θ
(7)
where P ∗(Cs, M ) denotes the true likelihood that configu-
ration C will result from state s and mental model M.
Proof Recall from Section 4.2.1 that each configuration C
can be modeled as the concatenation of a separate Cθ for
each θ ∈ Θ. The likelihood of an arbitrary Cθ is given by
the probability mass function of the multinomial distribution
resulting from the pa,θ estimated from each neighborhood
Nθ(i):
P (Cθs, M ) = Nθ(i)!
pC(a,θ)
a,θ
C(a, θ)!
(11)
(cid:89)
a,θ
Thus, the probability of the fully concatenated configura-
tion C is given by:
P (Cs, M ) =
(cid:89)
θ
(cid:89)
a
Nθ(i)!
pC(a,θ)
a,θ
C(a, θ)!
(12)
Based on the choice of nθ from Theorem 1, each pa,θ
could over-estimate the true proportions by at most p (with
confidence 1 − α). Indeed, if instead the true proportions
of agents from each neighborhood were their lowest values
from within their corresponding margins of errors, then the
true probability mass function would be:
(cid:89)
(cid:89)
P (Cs, M ) =
Nθ(i)!
θ
a
(pa,θ − p)C(a,θ)
C(a, θ)!
(13)
Figure 3: Relationships between n and N
and thus Eq. 12 could over estimate the true probability of
an arbitrary configuration C by at most:
P (Cs, M ) − P (Cs, M )
Nθ(i)!
=
θ
a
(cid:89)
−(cid:89)
(cid:81)
(cid:81)
θ Nθ(i)!
a,θ C(a, θ)!
(cid:89)
(cid:89)
(cid:89)
Nθ(i)!
=
a
θ
pC(a,θ)
a,θ
C(a, θ)!
(pa,θ − p)C(a,θ)
C(a, θ)!
a,θ −(cid:89)
pC(a,θ)
(pa,θ + p)C(a,θ)
a,θ
a,θ
(14)
Analogously, each pa,θ could under-estimate the true pro-
portions by at most p. Indeed, if instead the true proportions
of agents from each neighborhood were their largest values
from within their corresponding margins of errors, then the
true probability mass function would be:
must be strictly less than P ∗(Cs, M ). Similarly, the proba-
bilities {pa,θ + p∀a ∈ A} do not form a valid probability
distribution because they sum to more than 1, so Eq. 15
must be strictly greater than P ∗(Cs, M ).
Finally, combining Eq. 16, Eq. 18, and Eq. 19, we find
that:
P ∗(Cs, M ) − P (Cs, M )
< max{P (Cs, M ) − P (Cs, M ),
P (Cs, M ) − P (Cs, M )}
= max{P (Cs, M ) − P (Cs, M ),
P (Cs, M ) − P (Cs, M )}
= P (Cs, M ) − P (Cs, M )
(cid:81)
(cid:81)
θ Nθ(i)!
a,θ C(a, θ)!
=
(cid:89)
(pa,θ + p)C(a,θ) −(cid:89)
a,θ
a,θ
(20)
pC(a,θ)
a,θ
(cid:89)
(cid:89)
P (Cs, M ) =
Nθ(i)!
θ
a
(pa,θ + p)C(a,θ)
C(a, θ)!
(15)
= P (C)
which establishes Eq. 7. (cid:3)
and thus Eq. 12 could under estimate the true probability
of an arbitrary configuration C by at most:
P (Cs, M ) − P (Cs, M )
Nθ(i)!
=
(pa,θ + p)C(a,θ)
C(a, θ)!
θ
a
(cid:89)
−(cid:89)
(cid:81)
(cid:81)
θ Nθ(i)!
a,θ C(a, θ)!
(cid:89)
(cid:89)
(cid:89)
Nθ(i)!
=
a
θ
pC(a,θ)
a,θ
C(a, θ)!
(pa,θ + p)C(a,θ) −(cid:89)
pC(a,θ)
a,θ
a,θ
a,θ
(16)
Within the largest possible errors in Eq. 14 and Eq. 16,
we note that for arbitrary action a and frame θ, we must
have that:
(pa,θ + p)C(a,θ) − pC(a,θ)
a,θ − (pa,θ − p)C(a,θ)
(17)
> pC(a,θ)
a,θ
by the binomial expansion of the (pa,θ + p)C(a,θ) and
(pa,θ − p)C(a,θ) terms. Therefore, we find that:
P (Cs, M ) − P (Cs, M ) > P (Cs, M ) − P (Cs, M )
(18)
Based on our definitions above of P (Cs, m) and
P (Cs, M ), we have that:
P (Cs, M ) < P ∗(Cs, M ) < P (Cs, M )
(19)
where the strict inequalities come from the fact that the
probabilities {pa,θ − p∀a ∈ A} do not form a valid proba-
bility distribution because they sum to less than 1, so Eq. 13
C Proof for Theorem 2
Through the following, we establish in Theorem 2 a bound
on the regret of the cumulative discounted rewards that the
subject agent i incurs from following a k horizon Many
Agent IPOMDP-Lite optimal policy πi, where the regret is
caused by using the estimated likelihoods of other agents'
configurations P (Cs, M ) based on modeling only a few
neighbors and extrapolating to the total population (as com-
pared to their true likelihoods P ∗(Cs, M )).
Strategy To establish this regret bound, we extend the the-
oretical analysis of Hoang and Low (Hoang and Low 2013),
following much of the same theoretical framework and argu-
ments that they established (reproduced and adapted here for
completeness). In particular, these extensions below in Lem-
mas 1-6, Proposition 1, and Theorem 2 involve (1) consider-
ing more than one other agent in the environment (Hoang
and Low only established theoretical results for environ-
ments with only two agents), and (2) accounting for the ap-
proximation error in P (Cs, M ) given by Corollary 1.
Lemma 1 (Approximation Error Bound on One Step Re-
wards). Let bi be an arbitrary belief and ai be an ar-
bitrary action for the subject agent i. Then the expected
reward Ri(bi, ai) = E[Ri(bi, ai, C)] with respect to the
estimated probabilities of other agents' actions in config-
urations P (Cs, M ), when compared to expected reward
R∗
i (bi, ai) = E[Ri(bi, ai, C)] with respect to the true prob-
abilities of other agents' actions P ∗(Cs, M ), will have
bounded error:
i (bi, ai) − Ri(bi, ai) ≤ P (C)CRmax
(21)
which is linear in the bounded approximation error of the
configuration probabilities P (C), where C is the set of all
possible configurations and Rmax is the largest reward
value for all possible s, a, C.
R∗
P (C)Ri(s, ai, C)
(by Cor. 1)
Fi(bi, ai,C, oi)(s(cid:48))
Proof This result extends Hoang and Low's Lemma 2
(Hoang and Low 2013). Expanding the expected reward val-
ues, we have:
R∗
i (bi, ai) − Ri(bi, ai)
(cid:88)
(cid:88)
(cid:88)
C
C
≤(cid:88)
≤(cid:88)
≤(cid:88)
(cid:88)
s,M
s,M
bi(s, M )
Ri(s, ai, C)
× P ∗(Cs, M ) − P (Cs, M )
s,M
C
bi(s, M )
bi(s, M )
P (C)Rmax
bi(s, M )P (C)CRmax
=
= P (C)CRmax
s,M
(cid:3).
Lemma 2 (Approximation Error Bound on Pi(C, oibi, ai)).
Let bi be an arbitrary belief and ai be an arbitrary action for
the subject agent i. Then the estimated belief-state observa-
tion model Pi(C, oibi, ai), i.e., the combined likelihood that
the other agents perform configuration C and subject agent i
observes observation oi, has approximation error compared
i (C, oibi, ai) at most the bound on
to its true distribution P ∗
the approximation error of P (Cs, M ):
P ∗
i (C, oibi, ai) − Pi(C, oibi, ai) ≤ P (C)
Lemma 3 (Approximation Error Bound on Belief Updates).
Let bi be an arbitrary belief, ai be an arbitrary action and oi
be an arbitrary observation for the subject agent i, and let
C be an arbitrary configuration of the other agents' actions.
Let Bi(bi, ai, C, oi) represent the belief-update function for
subject agent i, using it's estimated likelihoods of configura-
tions and producing distribution P (s(cid:48)oi, ai, C, bi), and let
Fi(bi, ai, C, oi) be its unnormalized form:
(cid:88)
= Oi(s(cid:48), ai, C, oi)
= Bi(bi, ai, C, oi)Pi(C, oibi, ai)
s,M
Ti(s, ai, C, s(cid:48))P (Cs, M )bi(s, M )
(25)
(22)
Then the norm-1 distances between Fi(bi, ai, C, oi) and its
true values F ∗
i (bi, ai, C, oi) calculated with the true proba-
bilities of other agents' actions P ∗(Cs, M ) is also bounded
by P (C), the approximation error of P (Cs, M ):
(cid:107)F ∗
i (bi, ai, C, oi) − Fi(bi, ai, C, oi)(cid:107)1 ≤ P (C)
(26)
Proof This result extends Hoang and Low's Lemma 3
(Hoang and Low 2013). Expanding the probability func-
tions, we have:
P ∗
i (C, oibi, ai) − Pi(C, oibi, ai)
Oi(s(cid:48), ai, C, oi)
Ti(s, ai, C, s(cid:48))bi(s, M )
× P ∗(Cs, M ) − P (Cs, M )
Oi(s(cid:48), ai, C, oi)
Ti(s, ai, C, s(cid:48))bi(s, M )
≤(cid:88)
≤(cid:88)
s(cid:48)
s(cid:48)
× P (C)
≤ P (C)
s,M
(cid:88)
(cid:88)
(cid:88)
(cid:88)
s,M
s(cid:48)
s(cid:48)
(cid:88)
(cid:88)
(cid:88)
s,M
s,M
(by Cor. 1)
bi(s, M )
bi(s, M )
bi(s, M )
Oi(s(cid:48), a, C, o)Ti(s, a, C, s(cid:48))
Ti(s, ai, C, s(cid:48))
≤ P (C)
= P (C)
= P (C)
(cid:3).
s,M
(24)
(cid:3).
(23)
Proof This result extends Hoang and Low's Lemma 5
(Hoang and Low 2013). Expanding the norm, we have:
F ∗
i (bi, ai, C, oi) − Fi(bi, ai, C, oi)1
Oi(s(cid:48), ai, C, oi)
Ti(s, ai, C, s(cid:48))
× bi(s, M )P ∗(Cs, M ) − P (Cs, M )
Ti(s, ai, C, s(cid:48))
Oi(s(cid:48), ai, C, oi)
s,M
(cid:88)
(cid:88)
s,M
s(cid:48)
≤(cid:88)
≤(cid:88)
(cid:88)
≤(cid:88)
s(cid:48)
× P (C)bi(s, M )
=
bi(s, M )
× P (C)
s,M
(cid:88)
(cid:88)
s(cid:48)
s(cid:48)
(cid:88)
s,M
s,M
× P (C)
= P (C)
= P (C)
bi(s, M )
Ti(s, ai, C, s(cid:48))
bi(s, M )
(27)
(by Cor. 1)
Oi(s(cid:48), ai, C, oi)Ti(s, ai, C, s(cid:48))
Lemma 4. Let bi be an arbitrary belief, ai be an arbitrary
action and oi be an arbitrary observation for the subject
agent i, and let C be an arbitrary configuration of the other
agents' actions. Let Bi(bi, ai, C, oi) be the belief update
function as defined in Lemma 3, calculated using the subject
agent i's estimated P (Cs, M ), and let B∗
i (bi, ai, C, oi) be
similar but calculated using the true P ∗(Cs, M ). Also, let
i (C, oibi, ai) be as given in Lemma
Pi(C, oibi, ai) and P ∗
2. Then:
Pi(C, oibi, ai)(cid:107)B∗
i (bi, ai, C, oi) − Bi(bi, ai, C, oi)(cid:107)1
≤ 2P (C)
(28)
i , Fi, P ∗
i , Bi, F ∗
Proof This result extends Hoang and Low's Proposition
1 (Hoang and Low 2013). Following Hoang and Low's
notation, let B∗
i , and Pi be shorthand
i (bi, ai, C, oi), Bi(bi, ai, C, oi), F ∗
for B∗
i (bi, ai, C, oi),
i (C, oibi, ai), and Pi(C, oibi, ai) re-
Fi(bi, ai, C, oi), P ∗
spectively. Then, expanding the norm, we have:
i B∗
P ∗
(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12) F ∗
− Fi(s(cid:48))
i (s(cid:48))
P ∗
i (s(cid:48))Pi − Fi(s(cid:48))P ∗
F ∗
i
Pi
i
(by Eq. 25)
i
s(cid:48)
s(cid:48)
=
=
1
Pi
(cid:88)
i − Bi1
= P ∗
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
=
=
s(cid:48)
s(cid:48)
Pi (F ∗
1
Pi
≤ 1
Pi
(by Triangle Ineq.)
[Pi F ∗
s(cid:48)
(cid:2)Pi F ∗
≤ 1
Pi
(by Lemma 2)
s(cid:48)
i (s(cid:48)) − Fi(s(cid:48)) +
F ∗
F ∗
i (s(cid:48)) − Fi(s(cid:48)) +
P (C)
Pi
P (C)
Pi
s(cid:48)
(by defintion of Fi)
F ∗
i (s(cid:48)) − Fi(s(cid:48)) + P (C)
s(cid:48)
=
≤ P (C) + P (C)
= 2P (C)
(cid:3).
Lemma 5. Let bi be an arbitrary belief, ai be an arbitrary
action and oi be an arbitrary observation for the subject
agent i, and let C be an arbitrary configuration of the other
agents' actions. Let Bi(bi, ai, C, oi) and B∗(bi, ai, C, oi)
be as given in Lemma 4, and let Pi(C, oibi, ai) and
i (C, oibi, ai) be as given in Lemma 2.
P ∗
Given P (Cs, M ), we have:
V ∗
i (B∗
i (bi, ai, C, oi))P ∗
i (C, oibi, ai)
− V ∗
≤ 3P (C)
i (Bi(bi, ai, C, oi))Pi(C, oibi, ai)
Rmax
1 − γ
(30)
Proof This result extends Hoang and Low's Proposition 2
(Hoang and Low 2013). As in the proof for Lemma 4, for
notational convenience, let B∗
i , and Pi be shorthand
i (C, oibi, ai), and
for B∗
Pi(C, oibi, ai) respectively. Then, expanding the left side of
Eq. 30, we have:
i (bi, ai, C, oi), Bi(bi, ai, C, oi), P ∗
i , Bi, P ∗
V ∗
i (B∗
i )P ∗
i − V ∗
≤ V ∗
+ V ∗
= P ∗
≤ P ∗
i (Bi)Pi
i (B∗
i )P ∗
i (Bi)P ∗
i V ∗
i (B∗
Rmax
1 − γ
i − V ∗
i − V ∗
i ) − V ∗
(cid:107)B∗
i
i (Bi)P ∗
i (Bi)Pi
i (Bi) + V ∗
i − Bi(cid:107)1 + V ∗
i (Bi)P ∗
i (Bi)P ∗
(by Hoang and Low (2013)'s Lemma 4)
i
i − Pi
i − Pi
≤ 2P (C)
i (Bi)P ∗
i − Pi
i (Bi)P (C)
≤ 2P (C)
≤ 2P (C)
+ V ∗
+ V ∗
Rmax
1 − γ
(by Lemma 4)
Rmax
1 − γ
(by Lemma 2)
Rmax
1 − γ
i ≤ Rmax
1 − γ
Rmax
1 − γ
(since V ∗
+
= 3P (C)
Rmax
1 − γ
)
P (C)
(31)
i (s(cid:48)) − Fi(s(cid:48))) + Fi(s(cid:48)) (Pi − P ∗
i )
i (s(cid:48)) − Fi(s(cid:48)) + Fi(s(cid:48))P ∗
i − Pi]
(cid:3)
i (s(cid:48)) − Fi(s(cid:48)) + Fi(s(cid:48))P (C)
(cid:88)
(cid:88)
s(cid:48)
s(cid:48)
Fi(s(cid:48))
PiBi(s(cid:48))
(by Lemma 3)
(29)
(cid:3).
Definition 1.
(Maximum Error in Value Function). Let
Vi,k(bi) be the optimal value function calculated using
the subject agent i's estimated P (Cs, M ) likelihoods over
other agents' configurations after k backups of the Bellman
operator, and let V ∗
i.k be the similar optimal value function
calculated instead using the true likelihoods P ∗(Cs, M ).
Then we define the maximal difference between V ∗
i,k and
Vi,k, caused by approximating the distribution of configu-
ration likelihoods as:
(cid:12)(cid:12)V ∗
i,n(b) − Vi,n(b)(cid:12)(cid:12)
δn (cid:44) max
b
(32)
Note that this definition matches that given by Hoang and
Low (Hoang and Low 2013) for the I-POMDP-Lite, except
their difference is due to approximating the other agents'
decision making as a Nested-MDP, whereas our's accounts
for extrapolating the models of only a few agents to the col-
lective behavior of the whole system.
Lemma 6. Let bi be an arbitrary belief and ai be an arbi-
trary action for the subject agent i. Then:
Q∗
i,k(bi, ai) − Qi,k(bi, ai)
≤ γδk−1 + P (C)CRmax
(cid:20)
1 + 3γ
(cid:21)
Ωi
1 − γ
(33)
where Qi,n(bi, ai) is the the expected Q value over all pos-
sible configurations C using the subject agent's estimated
P (Cs, M ) multinomial distribution after k backups of the
Bellman operation; Q∗
i,k is the similar Q value calculated
instead using the true likelihoods P ∗(Cs, M ), and δk−1 is
as defined in Def. 1.
Proof This result extends Hoang and Low's Proposition 3
(Hoang and Low 2013). As in the proofs for Lemmas 3 -
5, let us use Bi and B∗
i as shorthand notation. Analogous to
Hoang and Low's Proposition 3, we define:
Li,k(bi, ai) = R(bi, ai) + γ
i,k−1(Bi)Pi(C, oibi, ai)
V ∗
(cid:88)
(cid:88)
C
oi
(34)
Looking first at Li,k(bi, ai) − Qi,k(bi, ai), we establish:
Li,k(bi, ai) − Qi,k(bi, ai)
i,k−1(Bi) − Vi,k(Bi)(cid:12)(cid:12)
(cid:88)
(cid:88)
C
≤ γ
≤ γ
C
≤ γδk−1
(cid:88)
Pi(C, oibi, ai)(cid:12)(cid:12)V ∗
(cid:88)
(cid:88)
(cid:88)
Pi(C, oibi, ai)δk−1
Pi(C, oibi, ai)
oi
oi
C
oi
= γδk−1
(cid:12)(cid:12)(cid:12), we establish:
Looking next at
(cid:12)(cid:12)(cid:12)Q∗
i,k(bi, ai) − Li,k(bi, ai)
i,k(bi, ai) − Li,k(bi, ai)
(cid:88)
(cid:88)
i (bi, ai) − Ri(bi, ai)
i )P ∗
≤ R∗
+ γ
Q∗
C
oi
V ∗
i,k−1(B∗
− V ∗
i (C, oibi, ai)
i,k(Bi)Pi(C, oibi, ai)
≤ R∗
(cid:88)
(cid:88)
i (bi, ai) − Ri(bi, ai)
Rmax
(cid:88)
(cid:88)
1 − γ
3P (C)
oi
C
≤ P (C)CRmax + γ
+ γ
(by Lemma 1)
C
oi
(by Lemma 5)
3P (C)
Rmax
1 − γ
= P (C)CRmax + 3γP (C)CΩi Rmax
1 − γ
= P (C)CRmax
1 + 3γ
Ωi
1 − γ
(cid:20)
(cid:21)
(36)
Finally, by the Triangle Inequality and combining Eq. 35 and
Eq. 36, we have:
Q∗
i,k(bi, ai) − Qi,k(bi, ai)
≤ Q∗
i,k(bi, ai) − Li,k(bi, ai)
(cid:20)
+ Li,k(bi, ai) − Qi,k(bi, ai)
≤ γδk−1 + P (C)CRmax
1 + 3γ
Ωi
1 − γ
(cid:21)
(37)
(cid:3).
Proposition 1. The maximum difference δk between the op-
timal value function V ∗
i,k calculated using the true distribu-
tion over other agents' configurations P ∗(Cs, M ) and the
approximated optimal value function Vi,k from the estimated
distribution P (Cs, M ) is bounded by:
δk ≤ P (C)CRmax
γk−1 +
1
1 − γ
1 + 3γ
Ωi
1 − γ
(cid:20)
which is bounded linearly by P (C) and decreases as k
(i.e., the number of Bellman backups) increases, approach-
ing an asymptote of P (C)CRmax
1 + 3γ
Ωi
1−γ
1
1−γ
(cid:18)
(cid:16)
(cid:19)(cid:21)
(38)
(cid:17)
Proof This result extends Hoang and Low's Proposition 4
(Hoang and Low 2013). Based on Hoang and Low's Lemma
1 (Hoang and Low 2013), we have that:
(cid:12)(cid:12)Q∗
i,k(bi, ai) − Qi,k(bi, ai)(cid:12)(cid:12)
(39)
(35)
δk ≤ max
bi
max
ai
Thus, we find that:
δk ≤ max
bi
max
ai
(cid:12)(cid:12)Q∗
i,k(bi, ai) − Qi,k(bi, ai)(cid:12)(cid:12)
≤ γδk−1 + P (C)CRmax
1 + 3γ
(cid:20)
(cid:20)
≤ γk−1δ1 + P (C)C Rmax
1 − γ
1 + 3γ
(by unrolling the Bellman recurrence)
≤ γk−1P (C)CRmax + P (C)C Rmax
1 − γ
Ωi
1 − γ
Ωi
1 − γ
(cid:21)
(cid:20)
(cid:18)
(by Lemma 6)
(cid:21)
1 + 3γ
(cid:21)
Ωi
1 − γ
(cid:19)(cid:21)
(cid:20)
(by Lemma 1)
= P (C)CRmax
γk−1 +
1
1 − γ
1 + 3γ
Ωi
1 − γ
(cid:17)
(cid:16)
(40)
This establishes that δk is linearly bounded by P (C) (Eq.
38). Furthermore, since γ ∈ [0, 1), we have that γk−1 ap-
proaches 0 as k increases. Therefore, as k increases, the
γk−1 term disappears and δk approaches the tigher bound
of P (C)CRmax
Definition 2. (Actual Return Earned by Subject Agent). Let
Ji,k(bi) represent the actual expected cumulative discounted
rewards earned by the subject agent i over a horizon of k
if it follows its optimal policy πi calculated using the esti-
mated likelihoods of configurations of other agents' actions
P (Cs, M ), even though they chose their actions in the en-
vironment according to the true likelihood P ∗(Cs, M ):
Ωi
1−γ
1 + 3γ
. (cid:3)
1
1−γ
Ji,k(bi) =
bi(s, M )
P ∗(Cs, M )Ri(s, πi(bi), C)
(cid:88)
(cid:88)
s,M
C
+ γ
(cid:88)
(cid:88)
C
i (C, oibi, πi(bi))
P ∗
bi(s, M )
× Ji,k−1(Bi(bi, πi(bi), C, oi))
oi
(41)
where Bi represents the belief update function for subject
i (C, oibi, πi(bi)) represents the true likeli-
agent i and P ∗
hood that the other agents choose configuration C and the
subject agent observes oi after taking action πi(bi) with be-
lief current bi.
i,k − Ji,k
Theorem 2. (Regret bound). Maximum regret that agent i
from following a k-horizon optimal
incurs
policy πi (obtained by solving the many-agent I-POMDP-
Lite) due to the approximate likelihoods of other agents'
configurations P (Cst, M t) is bounded from above:
(cid:13)(cid:13)(cid:13)V ∗
(cid:13)(cid:13)V ∗
i,k − Ji,k
(cid:13)(cid:13)(cid:13)∞
(cid:13)(cid:13)∞ ≤ 2P (C) · C · Rmax
(cid:18)
(cid:20)
×
γk−1 +
1
1 − γ
(cid:19)(cid:21) (8)
Ωi
1 − γ
1 + 3γ
which is linear in the error P (C) in the subject agent's esti-
mation in configuration likelihoods caused by only modeling
some neighbors.
Proof This result extends Hoang and Low's Theorem 4
(Hoang and Low 2013). For notational convenience, de-
fine the maximal difference between the subject agent i's
computed value function (using its estimated likelihoods
P (Cs, M )) and its actual expected return:
φk (cid:44) max
b
Vi,k(b) − Ji,k(b)
(42)
Finally, as in prior lemmas, we use B∗
notation for B∗
respectively.
i and Bi as shorthand
i (bi, πi(bi), C, oi) and Bi(bi, πi(bi), C, oi),
To begin, we establish several intermediate results that will
be useful to derive Eq. 8. First, we find that:
i (C, oibi, πi(bi))Ji,k−1(B∗
P ∗
i ) − Vi,k−1(Bi)
≤ P ∗
+ P ∗
≤ P ∗
+ P ∗
i (C, oibi, πi(bi))Ji,k−1(B∗
i (C, oibi, πi(bi))Vi,k−1(B∗
i (C, oibi, πi(bi))Ji,k−1(B∗
i (C, oibi, πi(bi))
(cid:107)B∗
i ) − Vi,k−1(B∗
i )
i ) − Vi,k−1(Bi)
i ) − Vi,k−1(B∗
i )
i − Bi(cid:107)1
(by Hoang and Low (2013)'s Lemma 4)
≤ P ∗
i ) − Vi,k−1(B∗
i )
i (C, oibi, πi(bi))Ji,k−1(B∗
(by Lemma 4)
Rmax
1 − γ
Rmax
1 − γ
+ 2P (C)
≤ P ∗
i (C, oibi, πi(bi))φk−1 + 2P (C)
Rmax
1 − γ
(43)
Next, we find that:
P ∗
i (C, oibi, πi(bi))Ji,k−1(B∗
i ) − Pi(C, oibi, πi(bi))Vi,k−1(Bi)
i (C, oibi, πi(bi)) − Pi(C, oibi, πi(bi))
i ) − Vi,k−1(Bi)
i ) − Vi,k−1(Bi)
≤ P ∗
+ Vi,k−1(Bi)P ∗
≤ P ∗
+ Vi,k−1(Bi)P (C)
≤ P ∗
i (C, oibi, πi(bi))Ji,k−1(B∗
i (C, oibi, πi(bi))Ji,k−1(B∗
(by Lemma 2)
i (C, oibi, πi(bi))Ji,k−1(B∗
i ) − Vi,k−1(Bi)
+ P (C)
≤ P ∗
+ P (C)
= P ∗
Rmax
1 − γ
i (C, oibi, πi(bi))φk−1 + 2P (C)
(since Vi,k−1 ≤ Rmax
1 − γ
Rmax
1 − γ
)
Rmax
1 − γ
(by Eq. 43)
i (C, oibi, πi(bi))φk−1 + 3P (C)
Rmax
1 − γ
(44)
Following, we find that:
Vi,k(bi) − Ji,k(bi)
≤(cid:88)
(cid:88)
s,M
+ γ
bi(s, M )
(cid:88)
(cid:88)
C
Ri(s, πi(bi), C)
× P ∗(Cs, M ) − P (Cs, M )
C
oi
(cid:88)
(cid:88)
P ∗
i (C, oibi, πi(bi))Ji,k−1(B∗
i )
− Pi(C, oibi, πi(bi))Vi,k−1(Bi)
C
(by Lemma 1)
≤ P (C)CRmax
P ∗
i (C, oibi, πi(bi))Ji,k−1(B∗
i )
+ γ
− Pi(C, oibi, πi(bi))Vi,k−1(Bi)
≤ P (C)CRmax
i (C, oibi, πi(bi))φk−1 + 3P (C)
P ∗
(cid:88)
(cid:88)
(cid:20)
+ γ
oi
(cid:21)
Rmax
1 − γ
C
oi
(by Eq. 44)
(cid:88)
(cid:88)
= P (C)CRmax
i (C, oibi, πi(bi))φk−1
P ∗
+ γ
C
oi
+ 3γCΩiP (C)
≤ P (C)CRmax + 3γCΩiP (C)
Rmax
1 − γ
Rmax
1 − γ
+ γφk−1
(45)
(by Eq. 47)
= 2P (C)CRmax
γk−1 +
1
1 − γ
1 + 3γ
Since Eq. 48 holds for arbitrary bi, we conclude that:
(cid:13)(cid:13)V ∗
i,k − Ji,k
(cid:13)(cid:13)∞ ≤ 2P (C)CRmax
γk−1 +
1
1 − γ
(cid:20)
(cid:18)
(cid:19)(cid:21)
Ωi
1 − γ
which is a restatement of Eq. 8. (cid:3).
Rewriting this as a recurrence relation based on the defini-
tion above of φk, we have:
φk ≤ P (C)CRmax + 3γCΩiP (C)
(cid:20)
(cid:21)
Rmax
1 − γ
Ωi
1 − γ
(cid:18)
1 + 3γ
+ γφk−1
(cid:19)
Ωi
1 − γ
= γφk−1 + P (C)CRmax
1 + 3γ
≤ γk−1φ1 + P (C)CRmax
1
1 − γ
(by unrolling the recurrence)
≤ γk−1P (C)CRmax
+ P (C)CRmax
1
1 − γ
(by Lemma 1)
= P (C)CRmax
(cid:20)
(cid:18)
1 + 3γ
(cid:19)
Ωi
1 − γ
(cid:18)
γk−1 +
1
1 − γ
1 + 3γ
Ωi
1 − γ
(cid:19)(cid:21)
(46)
Hence, we must have that:
Vi,k(bi) − Ji,k(bi)
≤ φk
≤ P (C)CRmax
(by Eq. 46)
(cid:20)
γk−1 +
(cid:18)
1
1 − γ
1 + 3γ
Ωi
1 − γ
(cid:19)(cid:21)
(47)
Finally, we can establish our main result in Eq. 8. We find
that for arbitrary belief bi:
V ∗
i,k(bi) − Ji,k(bi)
i,k(bi) − Vi,k(bi)(cid:12)(cid:12) + Vi,k(bi) − Ji,k(bi)
≤(cid:12)(cid:12)V ∗
≤ P (C)CRmax
γk−1 +
(by Prop. 1)
+ Vi,k(bi) − Ji,k(bi)
≤ P (C)CRmax
γk−1 +
+ P (C)CRmax
γk−1 +
1
1 − γ
1
1 − γ
1
1 − γ
(cid:20)
(cid:20)
(cid:20)
(cid:20)
1 + 3γ
Ωi
1 − γ
1 + 3γ
1 + 3γ
Ωi
1 − γ
Ωi
1 − γ
(cid:18)
(cid:18)
(cid:18)
(cid:18)
(cid:19)(cid:21)
(cid:19)(cid:21)
(cid:19)(cid:21)
(cid:19)(cid:21)
Ωi
1 − γ
(48)
1 + 3γ
(49)
D Additional Empirical Results
Here we present the second and third performance measures:
the average number of fire locations put out and the average
amount of suppressant used by agents. They are included
here as there was not sufficient room in the manuscript.
Figure 4: Average number of fire locations put out per run
across 100 runs. Error bars represent the 95% confidence
interval. I-POMCP planned for 5 seconds in Setups 1-3 and
5, and for 20 seconds in Setup 4.
Figure 5: Average units of suppressant used per agent av-
eraged across 100 runs. Error bars represent the 95% con-
fidence interval. I-POMCP planned for 5 seconds in Setups
1-3 and 5, and for 20 seconds in Setup 4.
|
1801.10000 | 3 | 1801 | 2019-02-16T11:27:19 | Crowd Behavior Simulation with Emotional Contagion in Unexpected Multi-hazard Situations | [
"cs.MA"
] | In this paper we present a novel crowd simulation method by modeling the generation and contagion of panic emotion under multi-hazard circumstances. Specifically, we first classify hazards into different types (transient and persistent, concurrent and non-concurrent, static and dynamic ) based on their inherent characteristics. Then, we introduce the concept of perilous field for each hazard and further transform the critical level of the field to its invoked-panic emotion. After that, we propose an emotional contagion model to simulate the evolving process of panic emotion caused by multiple hazards in these situations. Finally, we introduce an Emotional Reciprocal Velocity Obstacles (ERVO) model to simulate the crowd behaviors by augmenting the traditional RVO model with emotional contagion, which combines the emotional impact and local avoidance together for the first time. Our experimental results show that this method can soundly generate realistic group behaviors as well as panic emotion dynamics in a crowd in multi-hazard environments. | cs.MA | cs | Crowd Behavior Simulation with Emotional
Contagion in Unexpected Multi-hazard
Situations
Mingliang Xu, Xiaozheng Xie, Pei Lv, Jianwei Niu, Hua Wang
Chaochao Li, Ruijie Zhu, Zhigang Deng and Bing Zhou
Abstract -- Numerous research efforts have been conducted to simulate crowd movements, while relatively few of them are
specifically focused on multi-hazard situations. In this paper, we propose a novel crowd simulation method by modeling the
generation and contagion of panic emotion under multi-hazard circumstances. In order to depict the effect from hazards and
other agents to crowd movement, we first classify hazards into different types (transient and persistent, concurrent and non-
concurrent, static and dynamic) based on their inherent characteristics. Second, we introduce the concept of perilous field for
each hazard and further transform the critical level of the field to its invoked-panic emotion. After that, we propose an emotional
contagion model to simulate the evolving process of panic emotion caused by multiple hazards. Finally, we introduce an Emotional
Reciprocal Velocity Obstacles (ERVO) model to simulate the crowd behaviors by augmenting the traditional RVO model with
emotional contagion, which for the first time combines the emotional impact and local avoidance together. Our experiment results
demonstrate that the overall approach is robust, can better generate realistic crowds and the panic emotion dynamics in a crowd.
Furthermore, it is recommended that our method can be applied to various complex multi-hazard environments.
Index Terms -- crowd simulation, emotional contagion, multi-hazard, emotional reciprocal velocity obstacles
!
in some real-world cases, multiple hazards may occur
in the same area over a period of time, such as the
two sequential bombing attacks in Boston in 2013.
Traditional crowd simulation algorithms with a single
hazard in the scenario cannot be applied to these cases
directly because of the following reasons:
1) A multi-hazard scenario,
including different
types of hazards, different critical
levels of haz-
ards, dynamic changes of hazards, various evacuation
strategies, and so on, is more complex than the case
with a single hazard. The traditional single-hazard
models are very difficult to handle all the above
factors in a unified way.
2) The emotional contagion in multi-hazard envi-
ronment is a complex combining process of emotional
spreading, concerning both direct effects from hazards
and indirect effects from neighboring individuals.
However, existing emotional contagion models are
mainly designed for single-hazard scenes and cannot
be applied to multi-hazard scenes directly.
3) Traditional multi-agent navigation algorithms,
like Reciprocal Velocity Obstacles (RVO) [8], have not
considered the emotion of individuals, which means
they are short of the mechanism to deal with the con-
flict between obstacle avoidance and panic escaping.
Therefore, the simulation results under multi-hazard
circumstance by these algorithms appear less realistic.
In order to tackle the above challenges, in this pa-
per, we propose a novel multi-hazard scene model to
describe different effects of various types of hazards,
which is mainly applied to fire and explosion situa-
9
1
0
2
b
e
F
6
1
]
A
M
.
s
c
[
3
v
0
0
0
0
1
.
1
0
8
1
:
v
i
X
r
a
1 INTRODUCTION
T He advances in the study of typical crowd behav-
iors (such as stampede incidents and terrorist at-
tacks) in various domains including psychology, secu-
rity management, and computer science, have pointed
out that simulating both the sentimental state evolu-
tion and decision-making of a crowd under different
circumstances is an efficient way to show inherent
laws of nature [1]. This problem has been considered
as a system that as a class of multi-input multi-output
systems in the non-strict feedback structure [2]. As
a result,
it is important to accurately model both
the simulation environment and emotional contagion
among individuals for realistic crowd simulation.
Recent research efforts of crowd simulation in emer-
gency circumstances have been mostly focused on
those situations where there is only one hazard in the
area of interest [3], [4], [5], [6], [7], [8], [9]. However,
• Mingliang Xu, Pei Lv, Chaochao Li, Ruijie Zhu and Bing Zhou
are with Center for Interdisciplinary Information Science Research,
ZhengZhou University, 450000. Dr.Pei Lv is the corresponding author.
• Xiaozheng Xie, Jianwei Niu is with State Key Laboratory of Virtual
• Hua Wang is with School of Computer and Communication Engineer-
• Zhigang Deng is with Department of Computer Science, University
Reality Technology and Systems, Beihang University,100091
ing of Zhengzhou University of Light Industry, 450000.
of Houston, Houston, TX, 77204-3010.
E-mail: {iexumingliang, ielvpei, zhuruijie, iebzhou} @zzu.edu.cn;
[email protected];[email protected]; [email protected];
[email protected];
[email protected]
tions. In this model, the hazards are classified into six
different types according to three kinds of inherent
attributes: durations, time of occurence and dynamics.
Based on the definitions of these hazards, we further
propose the concept of perilous field and a conversion
function to map the criticality of the perilous field
to the emotion of individuals. It is noteworthy that
emotion in this paper mainly refers to the panic mood
of individuals in emergency situations.
In order to depict the complex process of panic
spreading, we put forward a new emotional contagion
model specially designed for multi-hazard situations
by combining panic emotions from different hazards
and individuals. Finally, an Emotional Reciprocal Ve-
locity Obstacles(ERVO) model, inspired from the tra-
ditional RVO model, is proposed to drive the crowd
movement. Different from the existing RVO model,
the ERVO model integrates the emotional effect into
velocity decision for the first time.
The contributions of this paper are:
• We propose a novel multi-hazard scene model for
the description of emergency fire and explosion
situations, containing six different types of haz-
ards with their dynamic changing process and an
unified criticality conversion function.
• We propose a new emotional contagion model in
multi-hazard scenarios, which combines different
emotional effects from hazards and individuals in
a crowd.
• We propose a novel crowd behavior simulation
method, the ERVO to simulate how people under
a panic mode choose their paths to safe places or
planned goals in a realistic way.
The rest of this paper are organized as follows.
Background and related work are reviewed in Section
2. The overview of our work is introduced in Section
3. The definition of multiple types of hazards and
emergency scenes are described in Section 4. The
emotional contagion process is explained in detail in
Section 5. The simulation method of crowd movement
is described in Section 6. Our experiments are pre-
sented in Section 7. Finally, this paper is concluded in
Section 8.
2 RELATED WORK
Although numerous research efforts have been con-
ducted to simulate crowd movements, relatively lit-
tle literature has been specifically focused on emer-
gency evacuation simulation involved with multi-
ple hazards. In this section, we will mainly review
recent works that are clearly related to our work.
For more comprehensive review on crowd simulation
techniques, please refer to [10].
2.1 Crowd evacuation with social or physical
model
One kind of important crowd movement scenarios
is to simulate the emergency evacuation. Helbing
2
et al. [3] employ the social force model, combined
with social psychology and physics models for the
first time, to describe the panic behavior in evacu-
ation. After that, the lattice gas model [11], multi-
grid model [12], agent-based model [5], virtual hin-
drance model [13], etc., have also been proposed to
describe the dynamical behaviors of the emergency
crowd. The commonness among these methods is
that they choose some typical characteristics of the
crowd first, and then use corresponding models to
describe different evacuation behaviors. Other studies
considering more factors in crowd evacuation process,
Narain et al. [4] simulate the clustering behaviors
of a high density crowd in a combined macro-micro
perspective. Funge et al. [14] put forward a cognitive
model to direct autonomous characters to perform
specific tasks, which outperforms many traditional
behaviors models. Durupinar et al. [15] analyze the
impact of psychological factors on the crowd move-
ment from the perspective of social psychology. Lai et
al. [16] aim at a problem of adaptive quantized control
for a class of uncertain nonlinear systems preceded
by asymmetric actuator backlash, which is similar
with our motion analysis with agents in unexpected
situations. Wang et al. [17] propose a semantic-level
crowd evaluation metric, which analyze the semantic
information between real and simulated data. Basak
et al. [18] validate and optimize crowd simulation
by using a data-driven approach, which proves the
parameters learned from the real videos can better
represent the common traits of incidents when simu-
lation. Oguz et al. [1] use continuous dynamic model,
to simulate the movements of agents in outdoor emer-
gency situations successfully. In this paper, our crowd
behavior model mainly focuses on the micro-level
behavior simulation. According to different multi-
hazard environments, we divide the crowd movement
into various cases and design crowd behaviors for
each case specially.
2.2 Crowd simulation with psychological model
In the real world, emotional state of an individ-
ual plays a vital role in his/her decision-making,
which fundamentally determines his/her movements
at each time step [19], [20]. Therefore, many recent
works start to consider the psychological factors of
agents, especially during the simulating process of
crowd movement
[21]. Belkaid et al. [22] stress the
important role that emotional modulation plays on
behavior organization by analyzing the relationships
between emotion and cognition. Bosse et al. [23]
propose the absorption model based on the heat
dissipation theory in thermodynamic, which embod-
ies the role of authority figures in the process of
emotional contagion. Tasi et al. [24] devise a multi-
agent evacuation simulation tool ESCAPES, where an
agent will accept the emotion of other agents who
has the strongest mood or has special identity. Le et
al. [25] propose an agent-based evacuation model by
considering emotion propagation among individuals
to make the simulation more realistic. Lhommet et
al. [26] also propose a computational model of emo-
tional contagion based on individual personality and
relationships. Durupinar et al. [15] create a system
that enables the specification of different crowd types
ranging from audiences to mobs based on a compu-
tational mapping from the OCEAN personality traits
to emotional contagion. Tsai et al. [27] combine the
dynamics-based and epidemiologicial-based models
to describe the dynamics of emotional spreading from
the perspective of social psychology. Fu et al. [28] use
a modified SIR model, originally proposes in [29], to
model the emotion evolving in the process of emer-
gency crowd movement. The work in [30] proposes
a stress model to realize the interactive simulation of
dynamic crowd behaviors. Although stress is similar
to our panic emotion in terms of the impact on crowd
behaviors, there are still some inherent differences.
For one certain crowd scene, they mainly model one
type of stress in it and the stress of external envi-
ronment on individuals. The mutual influence impact
among different individuals is ignored. In addition,
their model only focuses on the changes of individu-
als' velocities caused by the magnitude of stress. By
contrast, in our paper, the emotional state of agents
in emergency situations is mainly the panic emotion.
Due to different emotional spreading and reception
for various agents, we analyze the emotional conta-
gion by involving the personality factors. Since the
panic effect is not only coming from various hazards
but also from neighboring individuals, a new micro-
continuous emotion contagion model is designed.
2.3 Crowd path planning
Generally, path planning can be regarded as the multi-
objective optimization [31] and local information in-
teraction [32], [33] problems. In the process of crowd
evacuation, an individual's action decision [34], [35],
[36], [37], [38], [39], [40] is dependent on the evac-
uating directions of nearby agents, the locations of
hazards, and the obstacles in the scene.
Some researches develop a variety of methods to
avoid the collision problem through the calculation
of possible positions of individuals at the next time
step [41], [42]. On the premise of collision avoidance,
Kluge and Prassler [43] use a local obstacle avoidance
approach, combined with individual's emotion states
to calculate the movements of agents iteratively. Van
den Berg et al. [8] propose the well-known Recip-
rocal Velocity Obstacles (RVO) model to drive the
multi-agent navigation without collision. Concretely,
the reactive behavior of one agent at each time step
depends on the behaviors of all the other agents. In
their method, a collision-avoidance velocity for each
3
agent is chosen by taking into account the positions
and reciprocal velocities of all agents in the scenario.
By constructing visual trees, Belkhouche [44] proposes
a shortest path without conflict. Guy et al. [9] propose
an optimization method for collision avoidance on the
basis of the RVO model for real-time simulation of
large-scale crowd movement. In addition, they also
propose an energy-saving simulation method with the
minimum energy consumption as the guidelines [45].
Furthermore, a series of path planning and navigation
algorithms [46], [47], [48], [49] are also described in
mass population under complex background. In this
paper, we enhance the traditional RVO model with
emotional contagion in multi-hazard circumstances.
Panic is used to describe the emotional state of each
agent, which is changed dynamically and affect the
behaviors of individuals.
3 SYSTEM OVERVIEW
As shown in Figure 1, the main methodology of this
work is divided into three parts: multi-hazard envi-
ronment modeling in Section 4; emotional contagion
process under multi-hazard situations in Section 5;
crowd behavior simulation based on emotional recip-
rocal velocity obstacle in Section 6.
Specifically, in order to simulate the crowd behavior
in multi-hazard situations realistically, we analyze dif-
ferent types of hazards according to their properties,
the time of occurrence and duration. After that, we
propose a perilous field consisting of multiple hazards
and define a conversion function to map the intensity
of danger to panic emotion. Besides the direct effects
from hazards, panic propagations also exist among
different agents in emergency scenes. So we build
an emotional contagion model (ECM) to handle the
above cases. The ECM computes the panic emotion
of each agent in the dangerous field according to the
distance between this agent and the hazards using
the above conversion function. At the same time, the
ECM accumulates the contagious panic emotion from
other agents to obtain the final emotion of each agent.
To realize multi-agent navigation with panic emotion
under multi-hazard situations, we propose an ERVO
model to simulate the crowd behaviors. The major
contribution of ERVO is a new mechanism of velocity
decision by integrating both the traditional RVO and
panic emotion.
4 MULTI-HAZARD ENVIRONMENT MODEL-
ING
The characteristics of complexity,
interactivity and
time-varying make crowd behavior simulation chal-
lenging, especially in multi-hazard environments. In
order to achieve realistic simulation results, we first
need to model multi-hazard simulation environment
quantitatively.
4
Fig. 1: The framework of crowd behavior simulation in multi-hazard situations, consisting of three parts: (a)
the estimation of crowd panic in multi-hazard environment, (b) panic propagation in emergency situation
and (c) the impact on crowd movement from panic emotion. The red solid circle represents the hazard in our
circumstance, different blue solid circles in (a) represent agents with different panic emotion values. The darker
the color, the greater the panic emotion. In (b), the panic emotion of one agent (blue solid circle) is affected
by other agents (green solid circles) in its perceiving range. Stress safety directions, emotional directions and
combined directions of agents are annotated by yellow arrows shown in (c).
According to their durations, we divide hazards
into two different types: transient and persistent. The
former only lasts for a moment, while the latter lasts
for relatively long time. Both of them would cause
drastic changes to the psychological state of a crowd,
and individuals in the dangerous area would respond
immediately. The difference between them is that a
transient hazard only threats those individuals at the
time when it is happening. Once it disappears, the
threat will also disappear immediately. By contrast,
a persistent hazard will continue to impact those
individuals in the dangerous area during its existence.
According to their generation time, we divide haz-
ards into concurrent and non-concurrent. Specifically,
when some hazards occur concurrently, their influ-
ences on neighboring agents can be treated as a single
one. These influences should be accumulated together.
For non-concurrent hazards, we need to consider the
status of the crowd each time when a new hazard
happens. If an agent has already been affected by
other hazards before or has its own emotion, the new
effect needs to be accumulated.
More importantly, the static and dynamic char-
acteristics of hazards also play vital effects on the
crowd movement in complex situations. Based on this
fact, we classify the hazards with fixed position and
influence radius as static ones. Other cases, such as
fixed position with variable influence radius, variable
position with fixed or variable influence radius are
regarded as dynamic hazards. For dynamic hazards,
they may have different states over the time, which
determine their position and area of influence dynam-
ically.
The above six basic types of hazards have obviously
different impacts on the crowd movement. Realistic
multi-hazard scenarios usually consist of these basic
types and their combinations.
After analyzing these hazards qualitatively, we give
quantitative descriptions for them. We first define a
perilous field as the circular area with the hazard
position as the center and a radius. Each agent is
aware of the existence of hazards in the scene through
self-perception or neighbor contagion. The influence
of danger is limited in space: the farther the distance
to the hazard, the weaker influence to the crowd. For
different types of hazards, due to the uncertainty of
their location and range, new perilous fields will be
formed constantly along with the time. Defining the
hazard position as Ps, for example, it can affect all
agents in its perilous field with radius, defined as rs,
in the existence time, defined by U. If the diffusion
velocity and diffusion time for the hazard is vs and
ts, respectively, where vs = {vs1, vs2, ..., vsn}, n → ∞
depicts all possible directions for the diffusion, the
new hazard point P (cid:48)
s can be defined as Equation 1:
P
(cid:48)
s = Ps + ts · vs
= {Ps + ts · vs1, Ps + ts · vs2, ..., Ps + ts · vsn}
(1)
The dangerous range As after the diffusion forms a
closed area consisting of Ps as the source point and
all points P (cid:48)
s as the boundary. Then we divide this
area into two parts using a line between 1 and n
2 , and
this area can be expressed as the sum of integration
of these two parts.
As =
(Ps + ts · vs)dvs −
(Ps + ts · vs)dvs
(2)
According to the above description, the dangerous
impact on each agent is related to the dangerous range
of hazard and the distance between the hazard and
an agent. The farther the distance is, the smaller the
impact, all points with the same distance from hazard
share the same dangerous impact. In order to depict
this symmetry and attenuation, which is inspired by
the work in [1], a Gaussian distribution function is
chosen to depict this procedure by Equation 3.
(cid:90) n
2
1
(cid:90) n
2
n
1√
0
2π·rs
Γs (P, t) =
− (P −Ds)2
e
2rs2
if(cid:107) P − Ds(cid:107) < rs and t ∈ U
otherwise
(3)
Here, Γs(P, t) is the strength of danger at the po-
sition P produced by hazard s at time t. U is the
duration of hazard s. Ds is the intersection position
of line P Ps and the hazard area As(Ds can be seen
as the hazard position Ps in static hazard situations),
and rs is its influence radius. It is noteworthy that
danger strength will be 1.0 if position P is within the
dangerous range As.
5 EMOTIONAL CONTAGION MODEL CON-
STRUCTION
The emotional contagion model under multi-hazard
situations needs to consider the panic emotion in-
voked directly by the hazards, panic propagation
among individuals, and panic attenuation. The final
panic emotion of each agent can be obtained by
summing up these three components.
5.1 Emotional impact from multiple hazards
In Section 4, we have defined the perilous field and
the strength of danger of different hazards. Since the
normalized value of the strength of danger is within
the range [0, 1] , which is the same as the property of
emotional value [28], therefore, we adopt the strength
5
n(cid:88)
of danger, perceived by the agent directly, as the panic
value at the current position in Equation 4.
Eh
i (P, t) =
Γs (P, t)
(4)
s=1
Here, Eh
i (P, t) represents the panic value of agent i
affected by all the hazards s at time t and position P ,
where n denotes the total number of hazards.
5.2 Emotional contagion among individuals
In real life, individuals escaping from the perilous
field will carry panic emotion and propagate the
panic continuously to infect other individuals within
a certain distance when they are moving. Individuals
who perceive this panic may also be affected by
them, incorporate into their emotions and then pass
them out. In addition, emotional contagion among
different agents are totally different. The extent of
emotional transmission among agents depend on their
personalities, which affect their ability of expression
and reception.
In order to depict the above process, we use the
emotional contagion model proposed in [15], which
incorporate a complex but easy-to-use psychological
component into agents to simulate various crowd
types. one personality model and two thresholds are
used in this process. Specifically, OCEAN person-
ality model [50] defines a five-dimensional vector
(cid:104)ΨO, ΨC, ΨE, ΨA, ΨN(cid:105) to characterize the individuals'
five kinds of personality: openness, conscientiousness,
extraversion, agreeableness and neuroticism. Each di-
mension takes a value between -1 and 1. Moreover,
personality can also affect the decision of agent in
different situations. The two thresholds are expres-
siveness and susceptibility. Expressiveness correlated
with extroversion, represents the ability to diffuse
emotion. Susceptibility represents the minimum value
of agent be affected by other agents. Taking agent i
and agent j as an example, if the emotional value for
agent j at certain time is higher than its expressive
force threshold, it will express the emotion to others.
At the same time, if all emotions agent i received
exceeds its susceptibility threshlod, agent i can be
affected by this emotion. The expressiveness threshold
for agent j and susceptibility threshlod for agent i are
defined as follows:
eTj ∼ N
0.5 − 0.5ψE
(cid:16)
j ,(cid:0)(cid:0)0.5 − 0.5ψE
(cid:1) /10(cid:1)2(cid:17)
0.5 − 0.5εj, ((0.5 − 0.5εj) /10)2(cid:17)
j
(cid:16)
(5)
(6)
susTi (t) ∼ N
Where N (., .) represents a normal distribution with
the former as mean and the later parameter as a
standard deviation, the empathy value εi(εi ∈ [−1, 1])
in Equation 7 for agent i can be described as follows
[51] :
εi = 0.354ψO + 0.177ψC + 0.135ψE + 0.312ψA + 0.021ψN (7)
6
t(cid:88)
n(cid:88)
t(cid:48)=t−k+1
j=1
Then for the susceptible agent i, all effect caused
by all agent j who is expressive and in the perceived
range of it at time t can be computed by Equation 8:
Ec
i (P, t) =
di (t(cid:48)) Ec(cid:48)
j (Pj, t(cid:48))
(8)
Where di(t(cid:48)) ∼ N(0.1,0.0001) represents the dose
values which agent i accepted from agent j at time t(cid:48),
j (Pj, t(cid:48)) is the panic emotion of agent j within the
Ec(cid:48)
perceiving range of agent i at time t(cid:48). The value of k
is set as 10 based on [15], which means the emotional
accumulation of agent i at time t is determined by the
emotional values in the last 10 consecutive time steps.
5.3 Emotion combination
Based on the documented observations [15], the panic
emotion of individuals will decay over time gradually
until to the normal state. So we define an emotional
attenuation function to describe this process, where
a parameter η is the emotional decay rate. For agent
i at time step t, its new panic can be computed as
following:
i (P, t) = Ei(P pre, t − 1) · η
Ed
η ∈ (0, 1]
(9)
As mentioned at the beginning of this section, the
final panic emotion of each agent can be obtained by
combining all above three components. Considering
the Equations 4, 8 and 9, the incremental panic of the
agent i, who is at the position P and at time t, can
be computed by Equation 10. With this incremental
value, we can obtain the panic emotion by Equation
11. It is noteworthy that the emotional value Ei(P, t)
needs to be normalized after update.
∆Ei (P, t) = Eh
i (P, t) + Ec
i (P, t) − Ed
i (P, t)
Ei (P, t) = Ei (P pre, t − 1) + ∆Ei (P, t)
(10)
(11)
6 EMOTIONAL RECIPROCAL VELOCITY OB-
STACLE
After the panic of each agent in a multi-hazard envi-
ronment is computed during evacuation, the stress-
ful behaviors of these agents affected by the panic
emotion can be determined. The location and mov-
→
ing direction of an agent are denoted as P and
V ,
respectively. When the agent has perceived the impact
from a hazard s at location Ps, it will try to follow
→
the stress safety direction
PsP to escape from the
hazard instinctively. By contrast, those agents who
Fig. 2: Stress safety direction invoked by hazards. Red
solid circles represent hazards, dotted circles are the
perilous fields of the hazards. The original directions
of agents are represented by blue triangles, while the
stress safety directions of affected agents are denoted
by red triangles.
n−1(cid:80)
→
s=0
V
→
are not within the impacted area of any hazard, will
follow their original moving directions. If an agent is
affected by multiple hazards, then all the stress safety
directions of interest will be the result of a weighted
sum. So, the stress safety direction of an agent in
multi-hazard situations can be described by Equation
12:
→
V s
i (P, t) =
Γs (P, t) · →
PsP if(cid:107) P − Ps(cid:107) < rs and t ∈ U
otherwise
Here,
(12)
i (P, t) is defined as the safety evacuation
V s
direction for agent i at the position P and time t. U
is the duration of hazard s. Figure 2 shows different
safety evacuation directions chosen by a group of
individuals.
Besides the direct emotional impact from hazards,
the contagious panic emotion received from its neigh-
bors may also alter agents' original moving directions.
As mentioned in [3], we assume the probability of
agent i following its original direction is pi and the
probability 1−pi to follow the others' directions. Thus,
the new direction can be defined as the addition of
these two direction vectors. In this paper, the proba-
bility pi is equal to the panic value Ei(P, t) of agent
i. The updated moving direction of agent i at time t
is defined as below:
→
V c
i (P, t) = Ei(P, t)
i (P, t) + (1 − Ei(P, t)) (cid:80)
→
V s
j∈R(i)
→
V c
j (Pj, t)
(13)
Here,
of agent
→
i (P, t) represents the moving direction
V c
i who is at
time t.
the position P at
(cid:80)
→
j (Pj, t) is the combined moving directions of
V c
j∈R(i)
those agents who are in the emotional perception
range of agent i. R(i) denotes neighboring agents
within the perception range of agent i. When the
agent is going to change its direction, we assume the
magnitude of its velocity will remain. In other words,
the velocity module of the agent at that time should
be Vc
i .
In Equation 13, the moving direction of agent i
is only influenced by panic emotion. However, in
the actual crowd movement, the final direction of an
agent is also influenced by its planned targets and
other neighboring moving agents. In other words, the
local obstacle avoidance and global path planning for
agents also need to be considered. The RVO model
[8] is an efficient and safe multi-object automatic
navigation algorithm. However, during the obstacle
avoidance, the RVO model focuses on the position and
velocity of the current agent and other agents (refer
to Equation 14), but does not take into account the
emotional impact on speed selection invoked by sur-
rounding obstacles and existing hazards. In Equation
14, RV Oi
j) is the collision area for agent i
caused by agent j (illustrated in the grey area around
the white circle of Figure 3 (RVO)), which means that
agent i and agent j will collide with each other once
the velocity of agent i fall into this area. Vi and Vj
represent the velocity for agent i and agent (or hazard)
j. αi
j is the effort chosen by agent i to avoid the
collision with agent (or hazard) j, which is implicitly
assumed to 1
2 in the original RVO model. For more
details of the RVO model, please refer to [8].
j(Vj, Vi, αi
RV Oi
j(Vj, Vi, αi
j) = {V
(cid:48)
i 1
αi
j
V
(cid:48)
i + (1 − 1
αi
j
)Vi ∈ V Oi
j(Vj)}
(14)
Inspired by the RVO model, we propose a new
ERVO model by integrating emotional contagion into
crowd movement planning. This new model con-
structs a new collision area (shown by the grey tri-
angle areas in Figure 3 (ERVO)) by considering the
current velocity Vi and the updated velocity Vc
i of
the agent, and also the velocity Vj as described in
Equation 15. The effort made by agent i to avoid col-
lision with agent (or hazard) j is defined in Equation
16.
ERV Oi
+ (1 − 1
j) = {V(cid:48)
j(Vj, Vi, Vc
i , αi
)Vi ∈ V Oi
j(Vj)}
i 1
αi
j
αi
j
(cid:48)
i + Vc
i )
(V
αi
j =
Ej(P, t)
Ei(P, t) + Ej(P, t)
(15)
(16)
During the crowd simulation, for agent i, if Vi is
outside of the emotional reciprocal velocity obstacle
of agent (or hazard)j, both of them will never collide.
The ERVO model can be used to navigate a large
number of agents in a complex multi-hazard scenario.
7
For each agent i in the scene, it has a current position
P , a current velocity Vi, an updated velocity Vc
i , a
current panic emotion Ei(P, t), and a goal location
Gi. For a hazard s, it has position Ps and duration t.
For obstacle o, it has current position Po and velocity
Vo. Static obstacles have zero velocity in particular. In
our experiments, we choose a small time step ∆t to
simulate crowd behaviors. Within this time step, we
select a new velocity for each object independently
and update its position according to the surrounding
environment until all of the agents have reached the
safe area or their goals.
7 EXPERIMENT RESULTS
We run a diverse set of crowd simulations in multi-
hazard situations, all experiments are realized by us-
ing C++ in the Visual Studio and Unity 3D platform.
Our experiment results show that our method can
soundly generate realistic movement as well as panic
emotion dynamics in a crowd. In Section 7.1, we
simulate crowd behaviors in four different outdoor
multi-hazard scenes. In Section 7.2, we analyze the
importance of our emotional contagion mechanism
and different influence in different scenarios. Then the
emotional contagion model is proved more suitable
for our multi-hazard situations in Section 7.3. Fur-
thermore, we validate the realism of our simulation
results by comparing them with the crowd movement
in real world in Section 7.4 and the effectiveness of our
method in different virtual environments in Section
7.5.
7.1 Crowd simulation under different multi-hazard
scenarios
As discussed before, different hazard types have
various effects on crowd movement. We simulate
emergency behaviors in a crowd with the following
two-hazard situations: (1) persistent hazards occur
at the same time; (2) transient hazards occur at the
same time; (3) persistent hazards occur at different
moments; (4) transient hazards occur at different mo-
ments. All simulations run in open field, and each
simulation involves forty agents. The persistent haz-
ards and transient hazards are represented by fire
and explosion, respectively. The time step is set to
0.25s, other parameters in our system are set exper-
imentally: the influence radius rs = 10 m, emotional
decay parameter η = 0.01, the personality parameters
ΨO, ΨC, ΨE, ΨA, ΨN are set to the random number
between -1 and 1 for simplify to depict different
agents, and the perceived scope is set to 4 for all
agents.
Path flow maps for all agents are used to depict
the crowd movement differences among this four
conditions. As illustrated in Figure 4, the black points
are the original positions of all agents, lines of dif-
ferent colors are used to depict different paths of
8
Fig. 3: The collision area computed by the traditional RVO model and our ERVO model for agent i. Grey
triangle areas around white circles and red solid circles represent the collision areas caused by agents and
hazards, respectively. The black, blue, green and red arrows are separately the original direction, stress safety
direction, emotional contagion direction and final direction of one agent. The emotional contagion direction of
an agent is determined by combining its safety stress direction with those of its neighbors. The final direction
is determined by combining its original direction and emotional contagion direction.
Fig. 4: Movement trajectories of forty agents in different types of hazard scenarios. In each scenario, black
points represent the initial positions of all agents, the lines drawn by different colors are used to depict different
paths of agents, while trajectories for the same agent use the same color in different conditions. In addition,
the red solid and hollow circles represent persistent and transient hazard positions, respectively. While the
green one represent the positions of the second hazards in the concurrent conditions.
agents, while trace flows for the same agent indicated
by the same color in four conditions, the red solid
and hollow circles represent persistent and transient
hazard positions in our scenarios, respectively. While
the second hazards occur in the concurrent conditions
draw by green.
If two hazards occur at the same time as shown
in Figure 4(a), Figure 4(b). Agents around these two
hazards will change their routes to be distant far
away from them. When compared with the transient
condition, based on the persistent effect from hazards,
more emotional contagion lead to jittery for many
paths of agents (shown in Figure 4(a)), while trajectory
for the agents in transient conditions are smoother
owing to the disappear of hazards in this scenario
(shown in Figure 4(b)).
If
two hazards occur at different moments, as
shown in Figure 4(c), Figure 4(d). When the first haz-
ard occurs, agents in the perilous field of this hazard
will change their movement direction far away from
it, while other agents keep the original movement.
When the second hazard occurs, if the first one does
9
(a) without emotional contagion model
not disappear, agents will escape away from both of
the two hazards (shown in Figure 4(c)). By contrast,
some agents' path may move to or pass through the
area where the first hazard disappeared (shown in
Figure 4(d)).
In addition to that, the panic emotion changes
of agents are also important during this procedure.
Figure 5 illustrate a snapshot in the condition of per-
sistent hazards occur at the same time, where we use
a cylinder to represent an agent and visualize its panic
value using different colors. Despite those two dead
agents drawn by the black cylinders, the white, light
red, red, dark red and red black are used to represent
Ei = 0, Ei ∈ (0, 0.3], Ei ∈ (0.3, 0.5], Ei ∈ (0.5, 0.7] and
Ei ∈ (0.7, 1.0], respectively. The larger the panic value
is, the darker its color. For more dynamic simulation
details in different multi-hazard conditions, we refer
readers to our supplemental video.
(b) with emotional contagion model
Fig. 6: The comparison of crowd movements with and
without emotional contagion.
In the previous section, we have discussed the effect
of emotional propagation on crowd movement quali-
tatively. Here we mainly focus on the change of panic
emotion of each agent during the crowd evacuation,
especially when persistent/transient hazards occur at
the same time. From Figure 7, we can see that the
panic emotion value will increase to the maximum
when a persistent hazard happens. The reason is that
although the agent is moving far away from the
hazard, the agent is still in the perilous field and the
panic value is accumulated. When agents are out of
the perilous field, their panic values will decay and
reach to a similar low level due to the effect of emotion
contagion. For a transient hazard, the panic emotion
will reach to the maximum immediately when the
hazard occurs, then it will decrease gradually.
7.3 Comparisons with another emotional conta-
gion model
In order to validate the effectiveness of our emotional
contagion model among agents, we compare our sim-
ulation results to an agent-based emotional contagion
model proposed in [52]. Same personality and original
state are chosen for fifty agents in this two mod-
els, then the overall difference caused by emotional
contagion can be caught. After bomb occurs, agents
may have different panic emotions and movements
Fig. 5: One snapshot in the condition of persistent
hazards occur at the same time, where the cylinders
are used to depict agents, different colors represent
different panic values of them, the darker the color is,
the larger its panic value.
7.2 Analysis of emotional contagion
In order to validate the effectiveness of emotional
contagion in our method, we run crowd simulations
in a scene with and without this mechanism, re-
spectively. Figure 6 shows the moving trajectories
of three selected agents in the situation with one
transient hazard. Agents with emotional contagion
will adjust their moving directions to escape away
from the hazard even when they have not reached
the nearby region of the hazard. In contrast, agents
without emotional contagion will keep moving along
the original planned directions. The trajectory of one
agent is illustrated by one colorful line. From these
results, we can infer that the crowd movement in a
hazard environment is affected by the panic emotion
significantly.
10
(a) panic changes caused by persistent hazards
(a) panic emotional interval distribution over time in
model [52]
(b) panic changes caused by transient hazards
Fig. 7: The panic emotion changes in a crowd in
different situations. The simulation contains 15 agents
and each colored line represents the panic emotion of
one agent in the scene.
in different time. Panic emotion of all agents and
movements simulation results can be shown in Figure
8, Figure 9.
In Figure 8, the number distribution of agents panic
emotion are illustrrated by five levels defined in Sec-
tion 7.1, Where 0 as the lowest panic emotion values
0 and level 5 represents the highest panic emotion
values from 0.7 to 1.0. We choose the explosion time
at 4s as the start time, which can be seen that all agents
have the high panic emotion almost the whole evacu-
ation process when used emotional contagion model
mentioned in [52], but in our model, the number of
lower emotion levels decrease first and increase as
following, the higher level ones reverses. The reason
of this phenomenon is that [52] considers all agents
in the whole scenes once a hazard occurs, and does
not take emotion decay into account. While in our
emotional contagion model, each agent have a percep-
tion range as well as expressiveness and susceptibility
to accept emotional contagion from others, and their
panic emotion change along with the movement.
In addition, the simulation results in this two condi-
tions shown in Figure 9. The movement of agents after
explosion in our model are more dispersed as labeled
by red ellipses, while in another model, all agents
behave towards an aggregation states. With consider-
ing the different emotion changes in this two models,
(b) panic emotional interval distribution over time in
our model
Fig. 8: The agent numbers in different panic emotional
interval during the evacuation. five panic emotion
levels depicted by 0-4 with different colors, the higher
this value, the higher the panic emotion.
lower panic emotion lead to a more independent
movement direction (shown in Figure 9(a)) instead
of gathered movement based on stronger emotional
contagion(shown in Figure 9(b)). In real world, panic
emotions will decrease when the crowd are away
from hazards. From the results of these two different
models, where be seen that our emotional contagion
model is more realistic and suitable to simulate the
crowd movement in the multi-hazard situations.
7.4 Comparisons with real-world crowd behaviors
In order to validate our approach, we also compare
the simulation results with real-world crowd evacua-
tion video. Two crowd evacuation video are chosen in
this part, first one is chosen from the public available
dataset of normal crowd videos from University of
Minnesota (UMN) [53], which is designed to test the
abnormal detection method originally. In this scene,
movement details are used to verify the similarity
between real-world crowd behaviors and our simu-
lation results. Although no pre-defined goals are set
in advance, agents can still be driven to escape in a
realistic way by our method.Illustrated in Figure 10,
three images in each row are the crowd movement
states at initial random conditions, at the beginning of
TABLE 1: The comparison between our simulation
result and real grassland scene
11
Agent ID
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
mean
(a) with our emotional contagion model
Trajectory length
in simulation result
(pixel)
111.8814
121.1261
117.4920
106.193
94.7727
86.9983
83.7936
110.8647
57.6991
45.4759
35.5255
73.6405
84.6207
90.2147
96.3887
113.0668
95.3169
Trajectory length
in real video
(pixel)
94.1773
121.4284
155.8467
89.5567
50.7417
32.9509
60.6989
110.0192
61.0370
42.4537
39.0000
42.2433
81.6077
98.1849
116.4673
128.4863
88.3267
Maximum speed
in simulation result
(pixel/frame)
6.0828
5.5902
5.8310
5.4083
5.3852
8.0623
5.0000
6.5192
5.5227
3.6401
3.5355
4.7170
3.5355
5.0000
7.5000
4.5277
5.3661
Maximum speed
in real video
(pixel/frame)
8.0156
8.0000
9.0000
9.8234
6.2560
6.5765
3.6056
11.5109
9.8489
7.5664
8.5586
5.3852
9.0000
10.0000
10.0000
13.5370
8.5405
TABLE 2: The comparison between our simulation
result and real square scene
Agent ID
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
mean
Trajectory length
in simulation result
(pixel)
84.9795
36.1059
32.0000
39.0000
65.1402
37.3852
32.3006
85.3429
86.2274
38.7559
136.2666
109.6164
124.5751
159.8465
176.6732
82.9477
Trajectory length
in real video
(pixel)
93.1883
49.7452
56.9671
70.1650
121.4701
64.8521
66.2214
135.1380
80.3388
88.4081
83.5291
120.4088
69.5775
152.3059
191.3627
96.2452
Maximum speed
in simulation result
(pixel/frame)
6.8007
4.6098
4.0000
3.6056
5.4083
3.2016
3.2016
5.5902
6.2650
5.0249
12.2577
7.0711
13.0096
9.1788
12.0934
6.7546
Maximum speed
in real video
(pixel/frame)
5.5902
4.5277
3.5355
7.1589
9.7082
6.5765
3.6056
9.5525
6.1033
6.0208
6.5192
9.0139
13.2004
10.5000
8.0156
7.3086
red arrows, more details can be found in our supple-
mental video.
7.5 Applications in different scenarios
We apply our method to simulate crowd evacuation
simulations in office building ( Figure 12) and cross-
roads ( Figure 13 ) with multiple hazards to check the
effectiveness of our method.
In an office building, we numbered its four rooms
as 1, 2, 3, 4 from left to right and up to bottom.
The corridor in the middle connects all these rooms
together and there are no exits on both sides. At the
beginning, 50 agents located in different rooms move
randomly in Figure 12(a) . At the 8th frame, there
are two bomb explosions in room 1 and room 4 at
the same time in Figure 12(b). At the 64th frame,
there is a fire in room 2 in Figure 12(c). From the
simulation results, we observe the following: when
bomb explosions occur, in order to avoid the danger,
agents in the rooms begin to move to room 2 and
room 3, respectively. When room 2 is on fire, the
agents in or aiming to room 2 try to escape. At last,
all of the agents move to the safe room 3 in Figure
12(d).
The crossroad scene contains 50 pedestrians and
two non-current car bombs. When the simulation
starts, agents cross the road freely in Figure 13(a). At
the 16th frame, one black car bomb explodes in Figure
13(b) and another red car bomb explodes at the 24th
(b) with emotional contagion model proposed in [52]
Fig. 9: The comparison of crowd movements with
another emotional contagion model.
the evacuation, one moment after the hazard occurs.
The trajectories of agents are shown by the blue lines.
From the trajectories we can find that movement
trends of our simulation results are similar with that
in real scenes. Furthermore, we also compare the
trajectory length and the maximum speed of each
agent from the moment when hazard occurs to the
end of simulation, as shown in Table 1 (comparison
with the grassland scene) and Table 2 (comparison
with the square scene), where can be seen that our
mean trajectory length and mean maximum speed are
close to the true video data. Thus we can see both the
overall movement trend of the crowd and individuals'
movement details in the crowd are similar to those in
the recorded real-world crowd video. More animation
comparison details can be found in our supplemental
video.
The second one is the 911 terrorist attacks with
two explosion, while when considering the camera
shaking and crowd occlusion, movement details of
agents cannot be obtained accurately, thus this scene
is mainly used to verify the similarity of group move-
ment trends between our simulation result and true
situation. In this circumstance, two bombs occurred
concurrent on the building, and all agents straight
forward in the whole procedure. As shown in Figure
11, the crowd movement directions are indicated by
12
(a) Recorded video data (ground-truth)
(b) Our simulation result (corresponding to (a))
(c) Recorded video data (ground-truth)
(d) Our simulation result (corresponding to (c))
Fig. 10: Snapshots of ground-truth crowd evacuations on the outdoor ground and our corresponding simulation
results. Three images from left to right is: initial random status, at the beginning of the evacuation, one moment
in the evacuation. The movement trajectories for all agents are drawn by the blue lines.
frame as shown in Figure 13(c). When the first car
bomb occurs, the agents nearby evacuate immediately.
Some agents affected by their neighbors move away
from the black car bomb. Since the dangerous field of
black car bomb is limited, the agents far away from
it continue to move along their original paths. When
the red car bomb occurs, these agents who are in the
perilous field also begin to evacuate, while others just
move in their original directions. Figure 13(d) is the
result at the end time (at the 60th frame). Animation
details can be found in our supplemental video.
8 DISCUSSION AND CONCLUSION
Crowd behavior simulation under multi-hazard en-
vironment is a very challenging problem, and exist-
ing models with a single hazard cannot be applied
to these cases directly. In this paper, we present a
novel evacuation simulation method by modeling the
generation and contagion of panic emotion under
multi-hazard circumstances. First, we model multi-
hazard environment by classifying hazards into differ-
ent types based on their inherent characteristics and
introducing the concept of perilous field for a hazard.
Then, we propose a novel emotion contagion model to
simulate the panic emotion evolving process in these
13
while it may not include all agents or some agents
with special characters. Thus, more factors need to be
considered. Furthermore, our method is sensitive to
some key parameters, such as the strength of danger.
In the future, we want to utilize a large number
of surveillance video clips to calibrate and further
improve our model.
ACKNOWLEDGMENTS
The authors would like to thank all the anonymous
reviewers. This work was supported by National
Natural Science Foundation of China under Grant
Number 61672469, 61772474, 61822701, 61872324.
REFERENCES
[5]
[1] O. Oguz, A. Akaydın, T. Yılmaz, and U. G ud ukbay, "Emer-
gency crowd simulation for outdoor environments," Computers
& Graphics, vol. 34, no. 2, pp. 136 -- 144, 2010.
[2] Y. J. Liu, S. Li, S. Tong, and C. L. Chen, "Neural approximation-
based adaptive control for a class of nonlinear nonstrict
feedback discrete-time systems," IEEE Transactions on Neural
Networks and Learning Systems, vol. 28, no. 7, pp. 1531 -- 1541,
2016.
[3] D. Helbing, I. Farkas, and T. Vicsek, "Simulating dynamical
features of escape panic," Nature, vol. 407, no. 6803, pp. 487 --
490, 2000.
[4] R. Narain, A. Golas, S. Curtis, and M. C. Lin, "Aggregate
dynamics for dense crowd simulation," ACM Transactions on
Graphics, vol. 28, no. 5, pp. 89 -- 97, 2009.
I. Stamatopoulou, I. Sakellariou, and P. Kefalas, "Formal agent-
based modelling and simulation of crowd behaviour in emer-
gency evacuation plans," in Proceedings of the IEEE International
Conference on Tools with Artificial Intelligence, vol. 1, 2012, pp.
1133 -- 1138.
S. Patil, J. Van den Berg, S. Curtis, M. C. Lin, and D. Manocha,
"Directing crowd simulations using navigation fields," IEEE
Transactions on Visualization and Computer Graphics, vol. 17,
no. 2, pp. 244 -- 254, 2011.
S. Paris, J. Pettr´e, and S. Donikian, "Pedestrian reactive navi-
gation for crowd simulation: a predictive approach," Computer
Graphics Forum, vol. 26, no. 3, pp. 665 -- 674, 2007.
J. Van den Berg, M. Lin, and D. Manocha, "Reciprocal velocity
obstacles for real-time multi-agent navigation," in Proceedings
of the IEEE International Conference on Robotics and Automation,
2008, pp. 1928 -- 1935.
S. J. Guy, J. Chhugani, C. Kim, N. Satish, M. Lin, D. Manocha,
and P. Dubey, "Clearpath: highly parallel collision avoid-
ance for multi-agent simulation," in Proceedings of the ACM
SIGGRAPH/Eurographics Symposium on Computer Animation.
ACM, 2009, pp. 177 -- 187.
[6]
[10] M. L. Xu, H. Jiang, X. G. Jin, and Z. Deng, "Crowd simulation
and its applications: Recent advances," Journal of Computer
Science and Technology, vol. 29, no. 5, pp. 799 -- 811, 2014.
[11] M. Muramatsu and T. Nagatani, "Jamming transition in two-
dimensional pedestrian traffic," Physica A: Statistical Mechanics
and its Applications, vol. 275, no. 1, pp. 281 -- 291, 2000.
[12] W. Song, X. Xu, B.-H. Wang, and S. Ni, "Simulation of
evacuation processes using a multi-grid model for pedestrian
dynamics," Physica A: Statistical Mechanics and its Applications,
vol. 363, no. 2, pp. 492 -- 500, 2006.
[13] W. Lv, W.-g. Song, J. Ma, and Z.-m. Fang, "A two-dimensional
optimal velocity model for unidirectional pedestrian flow
based on pedestrian's visual hindrance field," IEEE Transac-
tions on Intelligent Transportation Systems, vol. 14, no. 4, pp.
1753 -- 1763, 2013.
[14] J. Funge, X. Tu, and D. Terzopoulos, "Cognitive modeling:
knowledge, reasoning and planning for intelligent characters,"
vol. 271, no. 34, 1999, pp. 153 -- 164.
[7]
[8]
[9]
(a) True video data (911 terrorist attacks)
(b) Simulation result
Fig. 11: Snapshots of the true video and our simu-
lation result of 911 scenario, where the red arrows
represent the movement trends of crowd.
situations. Finally, we introduce an emotional Recip-
rocal Velocity Obstacles(ERVO) model by augmenting
the traditional RVO model with emotional contagion,
which combines the panic emotion impact and local
avoidance together for the first time. By comparing
our simulation results with the ground-truth data and
applying our algorithm in different virtual environ-
ments, our experiment results show that the overall
approach is robust and can better generate realistic
crowds as well as the panic emotion dynamics in a
crowd in various multi-hazard environments.
There are still several limitations in our current
work. The first one is that our current method relies
on some important assumptions, such as all agents
in our scenario are treated equally in the face of
hazards except the different personalities, thus they
can perceive the danger level and be affected by the
hazards once he/she enter into the influence radius
of them. Besides that, safe exits chosen in the simula-
tion environment in advance, especially in the office
building situations, where the doors are chosen as
the sole exit for each room. In real world, this is not
very common. So we need to improve the sensing
capability of the agents in an unknown multi-hazard
scenario. The second is, in spite of considering agents
personalities, expressiveness and susceptibility, di-
verse crowd movements are shown in our simulation
results, many other complex personality traits and
prior expertise may also affect the emotion changes
and motion choices of each agent. In addition, the
personality parameters in our emotional contagion
model are set randomly to depict different agents,
14
a ( 1st frame )
b ( 8th frame )
c ( 64th frame )
d ( 120th frame )
Fig. 12: Evacuation simulation result by our approach in an office building
a ( 1st frame )
b ( 16th frame )
c ( 24th frame )
d ( 60th frame )
Fig. 13: Evacuation simulation result by our approach in a crossroad.
15
of a motion modification algorithm," IEEE Transactions on
Systems, Man, and Cybernetics - Part A: Systems and Humans,
vol. 34, no. 3, pp. 376 -- 386, 2004.
[35] J. Cui, Y. Liu, Y. Xu, H. Zhao, and H. Zha, "Tracking generic
human motion via fusion of low- and high-dimensional ap-
proaches," IEEE Transactions on Systems Man and Cybernetics
Systems, vol. 43, no. 4, pp. 996 -- 1002, 2013.
[36] N. Garca, J. Rosell, and R. Surez, "Motion planning by demon-
stration with human-likeness evaluation for dual-arm robots,"
IEEE Transactions on Systems Man and Cybernetics Systems,
vol. PP, no. 99, pp. 1 -- 10, 2017.
[37] J. S. Zelek and M. D. Levine, "Local-global concurrent path
planning and execution," systems man and cybernetics, vol. 30,
no. 6, pp. 865 -- 870, 2000.
[38] X. Qu and M. A. Nussbaum, "Simulating human lifting mo-
tions using fuzzy-logic control," systems man and cybernetics,
vol. 39, no. 1, pp. 109 -- 118, 2009.
[39] W. Park, D. B. Chaffin, B. J. Martin, and J. Yoon, "Memory-
based human motion simulation for computer-aided er-
gonomic design," systems man and cybernetics, vol. 38, no. 3,
pp. 513 -- 527, 2008.
[40] P. Duan, K. Liu, N. Huang, and Z. Duan, "Event-based dis-
tributed tracking control for second-order multiagent systems
with switching networks," IEEE Transactions on Systems, Man,
and Cybernetics: Systems, pp. 1 -- 11, 2018.
[41] S. Rodriguez, J. M. Lien, and N. M. Amato, "Planning motion
in completely deformable environments," in IEEE International
Conference on Robotics and Automation, 2006, pp. 2466 -- 2471.
[42] M. Zucker, J. Kuffner, and M. Branicky, "Multipartite rrts for
rapid replanning in dynamic environments," in Proceedings of
the International Conference on Robotics and Automation, 2007, pp.
1603 -- 1609.
[43] B. Kluge and E. Prassler, "Reflective navigation: Individual be-
haviors and group behaviors," in Proceedings of the International
Conference on Robotics and Automation, 2004, pp. 4172 -- 4177.
[44] F. Belkhouche, "Reactive path planning in a dynamic environ-
ment," IEEE Transactions on Robotics, vol. 25, no. 4, pp. 902 -- 911,
2009.
[45] S. J. Guy, J. Chhugani, S. Curtis, P. Dubey, M. Lin, and
D. Manocha, "Pledestrians: a least-effort approach to crowd
simulation," in Proceedings of the ACM SIGGRAPH/Eurographics
Symposium on Computer Animation, 2010, pp. 119 -- 128.
[46] R. Kulpa, A.-H. Olivierxs, J. Ondrej, and J. Pettr´e, "Imper-
ceptible relaxation of collision avoidance constraints in virtual
crowds," ACM Transactions on Graphics, vol. 30, no. 6, p. Article
No. 138, 2011.
[47] R. Gayle, A. Sud, E. Andersen, S. J. Guy, M. C. Lin, and
D. Manocha, "Interactive navigation of heterogeneous agents
using adaptive roadmaps," IEEE Transactions on Visualization
and Computer Graphics, vol. 15, no. 1, pp. 34 -- 48, 2009.
[48] A. Sud, E. Andersen, S. Curtis, M. C. Lin, and D. Manocha,
"Real-time path planning in dynamic virtual environments
using multiagent navigation graphs," IEEE Transactions on
Visualization and Computer Graphics, vol. 14, no. 3, pp. 526 -- 538,
2008.
[49] A. Golas, R. Narain, S. Curtis, and M. C. Lin, "Hybrid long-
range collision avoidance for crowd simulation," IEEE Trans-
actions on Visualization and Computer Graphics, vol. 20, no. 7,
pp. 1022 -- 1034, 2014.
[50] P. H. Lodhi, S. Deo, and V. M. Belhekar, "The five-factor model
of personality," 1999.
[51] D. Jolliffe and D. P. Farrington, "Development and validation
of the basic empathy scale," Journal of Adolescence, vol. 29, no. 4,
pp. 589 -- 611, 2006.
[52] T. Bosse, R. Duell, Z. A. Memon, J. Treur, and C. N. V. D.
Wal, "Agent-based modeling of emotion contagion in groups,"
Cognitive Computation, vol. 7, no. 1, pp. 111 -- 136, 2015.
[53] "Unusual crowd activity dataset of university of minnesota,"
http://mha.cs.umn.edu/movies/crowdactivity-all.avi.
[15] F. Durupinar, U. Gudukbay, A. Aman, and N. I. Badler, "Psy-
chological parameters for crowd simulation: from audiences
to mobs," IEEE Transactions on Visualization and Computer
Graphics, vol. 22, pp. 2145 -- 2159, 2015.
[16] G. Lai, Z. Liu, Y. Zhang, C. L. Chen, and S. Xie, "Asymmetric
actuator backlash compensation in quantized adaptive control
of uncertain networked nonlinear systems," IEEE Transactions
on Neural Networks and Learning Systems, vol. 28, no. 2, pp.
294 -- 307, 2017.
[17] H. Wang, J. Ondej, and C. OSullivan, "Trending paths: A new
semantic-level metric for comparing simulated and real crowd
data," IEEE Transactions on Visualization and Computer Graphics,
vol. 23, no. 5, pp. 1454 -- 1464, May 2017.
[18] F. D. Ahmet Eren Baak, Uur Gdkbay, "Using real life incidents
for creating realistic virtual crowds with data-driven emotion
contagion," Computer and Graphics, vol. 72, pp. 70 -- 81, 2018.
[19] T. Bosse, M. Hoogendoorn, M. C. Klein, J. Treur, C. N. Van
Der Wal, and A. Van Wissen, "Modelling collective decision
making in groups and crowds: Integrating social contagion
and interacting emotions, beliefs and intentions," Autonomous
Agents and Multi-Agent Systems, vol. 27, no. 1, pp. 52 -- 84, 2012.
[20] H. B. Sharbini and A. Bade, "Analysis of crowd behaviour
theories in panic situation," in International Conference on In-
formation and Multimedia Technology, 2009, pp. 371 -- 375.
[21] W. O. Kermack and A. G. McKendrick, "Contributions to
the mathematical theory of epidemics. iii. further studies of
the problem of endemicity," Bulletin of Mathematical Biology,
vol. 53, no. 1, pp. 89 -- 118, 1991.
[22] M. Belkaid, N. Cuperlier, and P. Gaussier, "Autonomous cog-
nitive robots need emotional modulations: Introducing the
emodul model," IEEE Transactions on Systems, Man, and Cy-
bernetics: Systems, pp. 1 -- 10, 2018.
[23] T. Bosse, R. Duell, Z. A. Memon, J. Treur, and C. N. Van
Der Wal, "A multi-agent model for mutual absorption of
emotions," ECMS, vol. 12, pp. 212 -- 218, 2009.
[24] J. Tsai, N. Fridman, E. Bowring, M. Brown, S. Epstein,
G. Kaminka, S. Marsella, A. Ogden, I. Rika, and A. Sheel,
"Escapes: evacuation simulation with children, authorities,
parents, emotions, and social comparison," in The International
Conference on Autonomous Agents and Multiagent Systems, 2011,
pp. 457 -- 464.
[25] V. M. Le, C. Adam, R. Canal, B. Gaudou, H. T. Vinh, and
P. Taillandier, "Simulation of the emotion dynamics in a group
of agents in an evacuation situation," pp. 604 -- 619, 2010.
[26] M. Lhommet, D. Lourdeaux, Barth, and S. Jean-Paul, "Never
alone in the crowd: A microscopic crowd model based on
emotional contagion," in Ieee/wic/acm International Conference
on Intelligent Agent Technology, Iat 2011, Campus Scientifique De
La Doua, Lyon, France, August, 2011, pp. 89 -- 92.
[27] J. Tsai, E. Bowring, S. Marsella, and M. Tambe, "Empirical eval-
uation of computational fear contagion models in crowd dis-
persions," Autonomous Agents and Multi-Agent Systems, vol. 27,
no. 2, pp. 200 -- 217, 2013.
[28] L. Fu, W. Song, W. Lv, and S. Lo, "Simulation of emotional
contagion using modified sir model: A cellular automaton
approach," Physica A: Statistical Mechanics and its Applications,
vol. 405, pp. 380 -- 391, 2014.
[29] W. O. Kermack and A. G. Mckendrick, "A contribution to the
mathematical theory of epidemics," Bulletin of Mathematical
Biology, vol. 53, no. 1, pp. 89 -- 118, 1991.
[30] S. Kim, S. J. Guy, D. Manocha, and M. C. Lin, "Interactive
simulation of dynamic crowd behaviors using general adap-
tation syndrome theory," in ACM SIGGRAPH Symposium on
Interactive 3d Graphics and Games, 2012, pp. 55 -- 62.
[31] B. Wang, H. Li, Q. Zhang, and Y. Wang, "Decomposition-based
multiobjective optimization for constrained evolutionary opti-
mization," IEEE Transactions on Systems, Man, and Cybernetics,
pp. 1 -- 14, 2018.
[32] W. Wang, X. Chen, H. Fu, and M. Wu, "Model-free distributed
consensus control based on actor-critic framework for discrete-
time nonlinear multiagent systems," IEEE Transactions on Sys-
tems, Man, and Cybernetics: Systems, vol. 99, pp. 1 -- 12, 2018.
[33] Y. Zhang, S. Li, and L. Liao, "Consensus of high-order discrete-
time multiagent systems with switching topology," IEEE Trans-
actions on Systems, Man, and Cybernetics, pp. 1 -- 10, 2018.
[34] W. Park, D. B. Chaffin, and B. J. Martin, "Toward memory-
based human motion simulation: development and validation
|
1205.5003 | 1 | 1205 | 2012-05-22T18:50:48 | Ring Exploration with Oblivious Myopic Robots | [
"cs.MA",
"cs.DC"
] | The exploration problem in the discrete universe, using identical oblivious asynchronous robots without direct communication, has been well investigated. These robots have sensors that allow them to see their environment and move accordingly. However, the previous work on this problem assume that robots have an unlimited visibility, that is, they can see the position of all the other robots. In this paper, we consider deterministic exploration in an anonymous, unoriented ring using asynchronous, oblivious, and myopic robots. By myopic, we mean that the robots have only a limited visibility. We study the computational limits imposed by such robots and we show that under some conditions the exploration problem can still be solved. We study the cases where the robots visibility is limited to 1, 2, and 3 neighboring nodes, respectively. | cs.MA | cs |
Ring Exploration with Oblivious Myopic Robots
Ajoy K. Datta1, Anissa Lamani2, Lawrence L. Larmore1, and Franck Petit3
1 School of Computer Science, University of Nevada Las Vegas, USA
2 MIS, Universit´e de Picardie Jules Verne Amiens, France
3 LIP6, INRIA, CNRS, UPMC Sorbonne Universities, France
Abstract: The exploration problem in the discrete universe, using identical oblivious asynchronous
robots without direct communication, has been well investigated. These robots have sensors that allow them
to see their environment and move accordingly. However, the previous work on this problem assume that
robots have an unlimited visibility, that is, they can see the position of all the other robots. In this paper,
we consider deterministic exploration in an anonymous, unoriented ring using asynchronous, oblivious, and
myopic robots. By myopic, we mean that the robots have only a limited visibility. We study the computational
limits imposed by such robots and we show that under some conditions the exploration problem can still be
solved. We study the cases where the robots visibility is limited to 1, 2, and 3 neighboring nodes, respectively.
Keywords: Asynchronous Anonymous Oblivious Robots, Deterministic Exploration, Discrete Environment,
Limited Visibility
1
Introduction
There has been recent research on systems of autonomous mobile entities (hereafter referred to
as robots) that must collaborate to accomplish a collective task. Possible applications for such
multi-robot systems include environmental monitoring, large-scale construction, mapping, urban
search and rescue, surface cleaning, risky area surrounding or surveillance, exploration of unknown
environments, and other tasks in environments where instead of humans, robots are used. For many
of these applications, the larger the number of robots is, the easier the implementation is.
However, the ability of a team of robots to succeed in accomplishing the assigned task greatly
depends on the capabilities that the robots possess, namely, their sensing capabilities. Clearly, the
type of viewing device has a great impact on the knowledge that the robots have of their environment.
For example, if the robots have access to a global localization system (e.g., GPS, egocentric zone-
based RFID technology), then their viewing capabilities are a priori unlimited. By contrast, endowed
with a camera or a sonar, vision capabilities are limited to a certain distance. Obviously, the stronger
the device capabilities are, the easier the problem is solved.
In order to satisfy technological or budget constraints, it may be important to minimize both
aforementioned parameters, i.e., (i) the number of robots and (ii) capacities (or equipment) required
to accomplish a given task.
In this paper, we consider both parameters for the exploration problem. Exploration is a basic
building block for many applications. For instance, mapping of an unknown area requires that the
robots (collectively) explore the whole area. Similarly, to search and rescue people after a disaster,
the team of robots potentially has to explore the whole area. The so called "area" is often considered
to be either the continuous Euclidean space (possibly with obstacles and objects) or a discrete space.
1
In the latter case, space is partitioned into a finite number of locations represented by a graph, where
nodes represent indivisible locations that can be sensed by the robots, and where edges represent the
possibility for a robot to move from one location to the other, e.g., a building, a town, a factory, a
mine, and more generally, zoned areas. In a discrete environment, exploration requires that starting
from a configuration where no two robots occupy the same node, every node to be visited by at least
one robot, with the additional constraint that all robots eventually stop moving.
We assume robots having weak capacities: they are uniform -- meaning that all robots follow
the same protocol -- , anonymous -- meaning that no robot can distinguish any two other robots,
oblivious -- they have no memory of any past behavior of themselves or any other robot, -- , and
disoriented -- they have no labeling of direction. Furthermore, the robots have no (direct) means
of communicating with each other, i.e., they are unable to communicate together. However, robots
are endowed with visibility sensors enabling to see robots located on nodes.
In this paper, we add another constraint: myopia. A myopic robot has limited visibility, i.e.,
it cannot see the nodes located beyond a certain fixed distance φ. The stronger the myopia is, the
smaller φ is. In other words, we consider that the strongest myopia corresponds to φ = 1, when a
robot can only see robots located at its own and at neighboring nodes. If φ = 2, then a robot can
see robots corresponding to φ = 1 and the neighbors of its neighboring nodes. And so on. Note that
the weaker myopia corresponds to φ = ⌈ D
2 ⌉ − 1, D being the diameter of the graph. Infinite visibility
(i.e., each robot is able to see the whole graph) is simply denoted by φ = ∞.
We study the impact of myopia strength (i.e., the size of the visibility radius) on the problem of
exploration. This assumption is clearly motivated by limiting the vision capacities that each robot is
required to have. As a matter of fact, beyond technological or budget constraints, it is more realistic
to assume robots endowed with vision equipments able to sense their local (w.r.t. φ) environment
than the global universe. Furthermore, solutions that work assuming the strongest assumptions also
work assuming weaker assumptions. In our case, any distributed and deterministic algorithm that
solves the exploration problem with robots having severe myopia (φ = 1), also works with robots
with lower myopia (φ > 1), and even no myopia (φ = ∞).
Related Work. Most of the literature on coordinated distributed robots assumes that the robots
move in a continuous two-dimensional Euclidean space, and use visual sensors with perfect accuracy,
permitting them to locate other robots with infinite precision, e.g.,[1, 2, 4, 7, 8, 17]. In each of these
papers, other than [1, 7], the authors assume that each robot is able to see all other robots in the
plane.
In [1], the authors give an algorithm for myopic robots with limited visibility, converging
toward a single point that is not known in advance. The convergence problem is similar to the
gathering problem where the robots must meet in a single location in finite time. In [7], the authors
present a gathering algorithm for myopic robots in the plane, which requires that the robots agree
on a common coordinate system.
In the discrete model, gathering and exploration are the two main problems that have been
investigated so far e.g., [9, 10, 12, 13] for the gathering problem and [3, 5, 6, 14] for the exploration
problem. In [12], the authors prove that the gathering problem is not feasible in some symmetric
configurations and propose a protocol based on breaking the symmetry of the system. By contrast,
in [13], the authors propose a gathering protocol that exploits this symmetry for a large number
of robots (k > 18) closing the open problem of characterizing symmetric situations on the ring
that admit gathering solutions.
In [10, 11], the authors achieve similar results assuming weaker
robot capabilities: robots may not be able to count the number of robots on the same node. In
[9], the authors studied the gathering problem considering robots having only a local visibility ie,
2
they can only see robots located at its own and at adjacent nodes, i.e., φ is assumed to be equal
to 1. Under this assumption, the authors provide a complete solution of the gathering problem for
regular bipartite graphs. They first characterize the class of initial configurations allowing to solve
the gathering problem on such graphs, and they propose a gathering algorithm assuming that the
system starts from a configuration in this class.
In [6], it is shown that, in general, Ω(n) robots are necessary to explore a tree network of n nodes.
In [5], it is proved that no deterministic exploration is possible on a ring when the number of robots
k divides the number of nodes n. In the same paper, the authors proposed a deterministic algorithm
that solves the problem using at least 17 robots, provided that n and k are co-prime.
In [3], it
is shown that no protocol (probabilistic or deterministic) can explore a ring with fewer than four
robots. In the same paper, the authors give a probabilistic algorithm that solves the problem on a
ring of size n > 8 that is optimal for the number of robots. Finally, in [14], the authors reduce the
gap in the deterministic case between a large upper bound (k ≥ 17) and a small lower bound (k > 3)
by showing that 5 robots are necessary and sufficient in the case that the size of the ring is even,
and that 5 robots are sufficient when the size of the ring is odd.
Contribution. To the best of our knowledge, all previous results for discrete versions of the ex-
ploration problem assume unlimited visibility (φ = ∞), i.e., the whole graph is seen by each robot.
In this paper, we consider deterministic algorithms for ring networks of n nodes that uses k myopic
robots for ring networks. Following the same arguments as in [9], it should be emphasized that ex-
ploration would not be solvable starting from any configuration but from configurations where each
robot is at distance at most φ of another one -- at least when φ = 1.
Our contribution is threefold. First, we tackle the case where robots have a visibility φ = 1, i.e.,
they can only see the state of nodes at distance 1. In this case, we show that the exploration problem
is not solvable in both semi-synchronous and asynchronous models, for n > 1 and 1 ≤ k ≤ n. Next,
we show that even in the (fully) synchronous model, the exploration problem cannot be solved with
less than 5 robots when n > 6. We then propose, optimal deterministic algorithms in the synchronous
model for both cases 3 ≤ n ≤ 6 and n > 6. Second, we propose two deterministic solutions in the
asynchronous model when robots have a visibility φ = 2 using respectively 7 (provided n > k) and
9 (provided that n > kφ + 1) robots. Finally, we show that no exploration is possible with less than
5 robots when they have a visibility φ = 3 and n > 13 in both semi-synchronous and asynchronous
model. We then propose two asynchronous solutions that solve the problem using respectively 5 and
7 robots. The former works starting from specific configurations. All our solutions work assuming
that each robot is at distance at most φ of another one. Both solutions for 7 robots with φ = 2 and
5 robots with φ = 3 work starting from specific configurations.
Roadmap. Section 2 presents the system model that we use throughout the paper. We present
our results for the cases φ = 1, φ = 2, and φ = 3 in Section 3, 4, and 5, respectively. Section 6 gives
some concluding remarks.
2 Model
System. We consider systems of autonomous mobile entities called robots moving into a discrete
environment modeled by a graph G = (V, E), V being a set of nodes representing a set of locations (or
stations) where robots are, E being a set of edges that represent bidirectional connections through
3
which robots move from a station to another. We assume that the graph G is a ring of n nodes,
u0, . . . , un−1, i.e., ui is connected to both ui−1 and ui+1 -- every computation over indices is assumed
to be modulo n. The indices 0 . . . i . . . n − 1 are used for notation purposes only; the nodes are
anonymous and the ring is unoriented, i.e., given two neighboring nodes u, v, there is no kind of
explicit or implicit labeling allowing to determine whether u is on the right or on the left of v. On
this ring, k ≤ n anonymous robots r0, . . . , rj , . . . , rk−1 are collaborating together to achieve a given
task. The indices 0 . . . j . . . k − 1 are used for notation purposes only, since they are undistinguished.
Additionally, the robots are oblivious, i.e, they have no memory of their past actions. We assume
the robots do not communicate in an explicit way. However, they have the ability to sense their
environment. Zero, one, or more robots can be located on a node. The number of robots located
on a node ui at instant t is called multiplicity of ui and is denoted by Mi(t) (or simply Mi, if t is
understood). We say a node ui is free at instant t if Mi(t) = 0. Conversely, we say that ui is occupied
at instant t if Mi(t) 6= 0. If Mi(t) > 1 then, we say that there is an Mi(t).tower (or simply a tower)
at ui at instant t.
We assume that each robot is equipped with an abstract device called multiplicity sensor allowing
to measure node multiplicity. Multiplicity sensors are assumed to have limited capacities (so called,
limited visibility), i.e., each robot rj can sense the multiplicity of nodes that are at most at a fixed
distance1 φ (φ > 0) from the node where rj is located. We assume that φ is a common value to
all the robots. The ring being unoriented, no robot is able to give an orientation to its view. More
precisely, given a robot rj located at a node ui, the multiplicity sensor of rj outputs a sequence, sj,
of 2φ + 1 integers x−φ, x1−φ, . . . , x−1, x0, x1, . . . , xφ−1, xφ such that:
either x−φ = Mi−φ, . . . , x0 = Mi, . . . , xφ = Mi+φ,
or
x−φ = Mi+φ, . . . , x0 = Mi, . . . , xφ = Mi−φ.
If the sequence x1, . . . , xφ−1, xφ is equal to the sequence x−1, . . . , x1−φ, x−φ, then the view of rj
is said to be symmetric. Otherwise, it is said to be asymmetric.
Computations. Robots operate in three phase cycles: Look, Compute and Move (L-C-M). During
the Look phase, a robot rj located at ui takes a snapshot of its environment given by the output of its
multiplicity sensors, i,e., a sequence sj of 2φ + 1 integers. Then, using sj, rj computes a destination
to move to, i.e., either ui−1, ui, or ui+1. In the last phase (move phase), rj moves to the target
destination computed in the previous phase.
Given an arbitrary orientation of the ring and a node ui, γ+i(t) (resp., γ−i(t)) denotes the
sequence hMi(t)Mi+1(t) . . . Mi+n−1(t)i (resp., hMi(t) . . . Mi−(n−1)(t)i). We call the sequences γ−i(t)
and γ+i(t) mirrors of each other. Of course, a symmetric sequence is its own mirror. By convention,
we state that the configuration of the system at instant t is γ0(t). Let γ = hM0M1 . . . Mn−1i be a
configuration. The configuration hMi, Mi+1, . . . , Mi+n−1i is obtained by rotating γ of i ∈ [0 . . . n − 1].
Two configurations γ and γ′ are said to be undistinguishable if and only if γ′ is obtained by rotating
γ or its mirror. Two configurations that are not undistinguished are said to be distinguished.
We call a configuration γ = hM0 . . . Mn−1i towerless if Mi ≤ 1.
for all i. A configuration at
which no robot that can move we call terminal.
An inter-distance d refers to the minimum distance taken among distances between each pair of
distinct robots. Given a configuration γ, a d.block is any maximal elementary path in which robots
are at distance d. Each occupied node at the extremity of the d.block is called a border. The size of a
d.block is the number of robots in the d.block. A robot not being into a d.block is said to be isolated.
1 The distance between two vertices in the ring is the number of edges in a shortest path connecting them.
4
A φ.group is any maximal elementary path in which there is one robot every node at distance at
most φ of each other. In other words, a φ.group is a d-block for the particular case of d = φ.
At each step t, a non-empty subset of robots is selected the scheduler, or daemon. The scheduler
is viewed as an abstract external entity. We assume a distributed fair scheduler. Distributed means
that, at every instant, any non-empty subset of robots can be activated. Fair means that every robot
is activated infinitely often during a computation. We consider three computational models: (i)
The semi-synchronous model, (ii) the (fully) synchronous model, and (iii) the asynchronous model.
In the former model, at every time instant t, every robot that is selected instantaneously executes
the full cycle L-C-M between t and t + 1. (This model is known as the ATOM model [17]). The
synchronous model is similar to the semi-synchronous model, except that the scheduler selects all
enabled robots at each step. In the asynchronous model, cycles L-C-M are performed asynchronously
for each robot, i.e., the time between Look, Compute, and Move operations is finite but unbounded,
and is decided by the scheduler for each action of each robot. Note that since each robot is assumed
to be located at a node, the model considered in our case can be seen as CORDA [15] with the
following constraint: the Move operation is atomic, i.e, no robot can be located on an edge. In other
words, whenever a robot takes a snapshot of its environment, the other robots are always seen on
nodes, never on edges. Since the scheduler is allowed to interleave the operations, a robot can move
according to an outdated view, i.e., during the computation phase, some robots may have moved.
We call computation any infinite sequence of configurations γ0, . . . , γt, γt+1, . . . such that (1)
γ0 is a possible initial configuration and (2) for every instant t ≥ 0, γt+1 is obtained from γt after
some non-empty set of robots executes an action. Any transition γt, γt+1 is called a step of the
computation. A computation c terminates if c contains a terminal configuration.
Algorithm. Each rule in the algorithm is presented in the following manner:
< Label > < Guard > :: < Statement >. The guard is a possible sequence s provided by the sensor
of a robot rj:s = xi−φ, . . . , xi−1, (xi), xi+1, . . . , xi+φ. A robot rj at node ui is enabled at time t (or
simply enabled when it is clear from the context) if:
s = Mi−φ(t), . . . , Mi−1(t), (Mi), Mi+1(t), . . . , Mi+φ(t), or
s = Mi+φ(t), . . . , Mi+1(t), (Mi), Mi−1(t), . . . , Mi−φ(t). The corresponding rule < Label > is then
also said to be enabled. The statement describes the action to be performed by rj. There are only
two possible actions: (i) →, meaning that rj moves towards the node ui+1, (ii) ←, meaning that rj
moves towards the node ui−1. Note that when the view of rj is symmetric, the scheduler chooses the
action to be performed. In this case, we write statement: ← ∨ →.
Character '?' in the algorithms means any value.
Problem Specification.
Definition 2.1 (Exploration) Let P be a deterministic protocol designed for a team of k robots
with a given positive visibility φ, evolving on an n-size ring. P is a deterministic (terminating)
exploration protocol if and only if every computation c of P starting from any towerless configuration,
the k robots collectively explore the ring, ı.e., (i) c terminates in finite time, (ii) Every node is visited
by at least one robot during c.
Theorem 2.1 If P is a deterministic exploration protocol, then for every pair of distinct configura-
tions γi and γj in any execution c of P, γi and γj are distinguished.
5
Proof.
Suppose P is a deterministic exploration protocol, and there exists an execution c of P and
a pair of distinct configurations, γi and γj of c, such that γi and γj are undistinguished. Since P is a
deterministic exploration protocol, c terminates in finite time. Let s = γ0γ1 . . . γl−1 be the smallest
prefix of c of length l that contains distinguishable configurations only. Recall that we assume that
no two consecutive configurations of c are identical. Since s is the smallest prefix of c containing only
distinguished configurations by assumption there exists γi in s such that γi and γl are undistinguished.
Such a γl exists since we assume that c contains at least two undistinguished configurations. So,
executing P, the set of actions that led the system from γi to γi+1 are also applicable in γl. Thus,
by executing the same set of actions, there exists a configuration γl+1 reachable from γl such that
γi+1 and γl+1 are undistinguished. Following the same reasoning for every j ∈ [2..l − 1], there exists
a configuration γl+j reachable from γl+j−1 such that γi+j and γl+j are undistinguished. Therefore,
there exists an execution c of P leading to the infinite sequence of actions starting for γi. This
contradicts the assumption that c terminates in finite time.
✷
3 Visibility φ = 1
In this section, we first prove that no deterministic exploration is possible in the semi-asynchronous
model when robots are able to see at distance 1 only (φ = 1). The result holds for any k < n. We
then show that no deterministic exploration solves the problem with four robots, even in the (fully)
synchronous model when n > 6. The above results are also valid for the asynchronous model [16].
Next, we provide optimal deterministic algorithms in the synchronous model for both cases 3 ≤ n ≤ 6
and n > 6.
3.1 Negative Results
Asynchronous Model. Since robots are able to see only at distance 1, only the four following
rules are possible:
Rsgl
Rout
Rin
Rswp
0(1)0 :: → ∨ ←
0(1)1 :: ←
0(1)1 :: →
1(1)1 :: → ∨ ←
In the following, γ0 refers to an initial configuration. Let us first assume that γ0 consists of a
single 1.block of size k. Note that Rule Rsgl is not applicable in an initial configuration. Also,
Rules Rout and Rin (Rule Rswp) implies that k must be greater than 2 (3, respectively). We first
show that no deterministic exploration protocol includes Rules Rsgl, Rout, and Rswp starting from
γ0. We now show that Rules Rout and Rswp cannot be part of the protocol.
Lemma 3.1 Let P be a semi-synchronous protocol for φ = 1 and 2 ≤ k < n.
exploration problem, then P does not include Rule Rout.
If P solves the
Proof.
By contradiction, assume that P includes Rule Rout. Note that Rule Rout is only
enabled on robots that are at the border of a 1.block. Since there is only one φ-group in γ0 (the
initial configuration), there are only two robots that can execute Rule Rout. Let uiui+1 . . . ui+k be
the 1.block in γ0. Denote rj the robot located on node ui+j (0 ≤ i < k). Without lost of generality,
assume that the adversary activates r0 only in γ0. Let us call the resulting configuration by T
(standing for "Trap" configuration). There are two cases to consider:
6
• k = n − 1. In that case, r0 moves to ui+k+1 and T includes a 1.block ui+1 . . . ui+kui+k+1. T is
undistinguishable from γ0. This contradicts Theorem 2.1.
• k < n − 1. In that case, once r0 moves, it becomes an isolated robot on ui−1 in T . Again, there
are two cases. Assume first that k = 2. Then, T includes two isolated robots, r0 and r1. Even
if P includes Rule Rsgl, by activating r0 and r1 simultaneously, T + 1 is undistinguishable
from T . Again, this contradicts Theorem 2.1. Assume that k > 2. Then, T is the configuration
in which ui+1 is the border occupied by r1 of the 1.block ui+1 . . . ui+k. Assume that in T ,
the adversary activates r1 that executes Rule Rout. Then, Configuration T + 1 includes two
1.blocks disjoint by one node. The former includes r0 and r1 (on ui−1 and ui, respectively).
The latter forms the sequence r2 . . . rk, located on ui+2 . . . ui+k. More generally, assume that
∀i ∈ [1..k − 2], in T + i, the adversary activates ri+1. The resulting configuration T + k − 2 is
undistinguishable from T . A contradiction (Theorem 2.1).
✷
Lemma 3.2 Let P be a semi-synchronous protocol for φ = 1 and 2 ≤ k < n.
exploration problem, then Rule Rin and Rule Rswp are mutually exclusive with respect to P.
If P solves the
Proof.
Assume by contradiction that P includes both Rule Rin and Rule Rswp. Note that
Rule Rin (respectively, Rule Rswp) is enabled only on robots that are located at the border of
the 1.block (resp., inside the 1.block).
In the case where k is even, assume that the scheduler
activates all the robots. The resulting configuration is undistinguishable from γ0. A contradiction
(by Theorem 2.1). In the case where k is odd, suppose that the scheduler activates the robots at the
border of the 1.block and their neighbors (in the case where k = 3, only one extremity is activated).
The resulting configuration is undistinguishable from γ0. A contradiction (by Theorem 2.1).
✷
Lemma 3.3 Let P be a semi-synchronous protocol for φ = 1 and 2 ≤ k < n.
exploration problem, then P does not include Rule Rswp.
If P solves the
Proof.
By contradiction, assume that P includes Rule Rswp. From Lemmas 3.1 and 3.2, P
does not include Rules Rout nor Rule Rin. So, in γ0, the robots that are inside a 1.block are the
only robots that are able to execute an action in the initial configuration (and k must be greater
than or equal to 3). If k > 3, then, the adversary activates at each time two neighboring robots
such that once they move, they simply exchange their position and the resulting configuration is
undistinguishable from γ0. A contradiction.
If k = 3 then, when Rule Rswp is executed, a 2.tower is created. Rule 0(2)0:: ← ∨ → cannot be
enabled since once executed, the scheduler can activate only one robot in the 2.tower such that the
configuration reached is undistinguishable from γ0. In the case where Rsgl is enabled, then the
isolated robot becomes either (i) neighbor of the 2.tower (the case where n = 4) or (ii) remains
isolated.
In the first case (i), if either 0(1)2:: ← or 0(2)1:: → is enabled, a 3.tower is created
(Observe that one node of the ring has not been explored). The only rule that can be executed
is 0(3)0:: ← ∨ →. Suppose that the scheduler activates two robots that move in two opposite
directions. The configuration reached is undistinguishable from γ0. If 0(2)1:: ← is enabled, then
suppose that the scheduler activates only one robot. The configuration reached is undistinguishable
from γ0. In the second case (ii) (the isolated robot remains isolated), Rsgl keeps being enabled.
7
Once it it is executed, the isolated robot moves back to its previous position and the configuration
reached is indistinguishable from the previous one.
✷
Corollary 3.1 Let P be a semi-synchronous protocol for φ = 1, 2 ≤ k < n. If P solves the explo-
ration problem, then P includes Rule Rin that is the only applicable rule in the initial configuration.
Proof. Directly follows from Lemmas 3.1, 3.2, and 3.3.
✷
Lemma 3.4 Let P be a semi-synchronous exploration protocol for φ = 1, 2 ≤ k < n. Then, k must
be greater than or equal to 5.
Proof. Assume that the adversary activates both robots r0 and r1 at the border of the 1.block in
γ0. By Corollary 3.1, both r0 and r1 execute Rule Rin in γ0.
If k = 2, the system reaches a configuration γ1 that is undistinguishable from γ0.
If k = 3, then a 3.tower is created on the middle node of the 1.block in γ1. Since the ring is
not explored anymore (at least one node is not visited), P must include at least the following rule:
0(3)0 :: ← ∨ →. Then, the adversary activates only two of them and an opposite destination node
for each of them. Then, the system reaches a configuration γ2 that is undistinguishable from γ0.
If k = 4, then two neighboring 2.towers are created on the two middle nodes of the 1.block in
γ1. There are two cases to consider: Assume first that P includes the rule: 2(2)0 :: ←. Then, by
activating the same number of robots on each tower, the system reaches γ2 such that γ1 and γ2 are two
undistinguishable configurations. Assume that P includes the rule: 2(2)0 :: →. Again, by activating
only one robot on each tower, the system reaches a configuration γ2 that is undistinguishable from
γ0.
✷
Let us call a symmetric x.tower sequence (an Sx-sequence for short), a sequence of occupied
nodes such that the extremities of the sequence contain an x.tower. Observe that a tower containing
at least 2 robots (by definition), x is greater than or equal to 2. Also, since the robots can only see
at distance 1, the tower is only seen by its neighboring robots.
The following lemma directly follows from Corollary 3.1 and Lemma 3.4:
Lemma 3.5 For every semi-synchronous exploration protocol P for φ = 1, then 5 ≤ k < n and
there exists some executions of P leading to a configuration containing an S2-sequence.
Consider robots being located at the border of an 1.block. Let x ≥ 1 be the number of robots
located on the border node and the following generic rules:
0(x)1 :: ←
T α(x)
T β(x)
0(x)1 :: →
T γ(x) x(1)1 :: ←
T δ(x) x(1)1 :: →
Remark that Rule Rin corresponds to Rule T β(1). Also, note that since x robots are located on
the border node of an 1.block, the local configuration for both T γ(x) and T δ(x) is 0x11. Similarly,
define the generic rule Tsgl(y) (y ≥ 2) as follow: 0(y)0 ::← ∨ →.
Lemma 3.6 Let P be a semi-synchronous protocol for φ = 1 and 5 ≤ k < n.
If P solves the
exploration problem, then for every x ≥ 2, Rule Tsgl(x) is mutually exclusive with both T γ(x − 1)
and T δ(x), with respect to P.
8
Proof.
Assume that P includes Tsgl(x) for some x ≥ 2. There are two cases to consider:
Assume by contradiction that P includes either Rule T γ(x − 1) or Rule T δ(x). Then, starting from
a configuration 0(x − 1)11 or 0(x)11, T γ(x − 1) or T δ(x) is enabled, respectively. In both cases,
by executing T γ(x − 1) or T δ(x), the configuration locally becomes 0x01 in which Rule Tsgl(x)
is enabled for each robot belonging to the x.tower. By executing Tsgl(x) on one robot only, the
adversary can lead the system in a configuration that is undistinguishable from the previous one. A
contradiction by Theorem 2.1.
✷
Lemma 3.7 Let P be a semi-synchronous protocol for φ = 1 and 5 ≤ k < n.
exploration problem, then for every odd x ≥ 3, the rules Tsgl(x) and T β(⌊ x
with respect to P.
If P solves the
2 ⌋) are mutually exclusive
Proof. Assume by contradiction that P includes both Tsgl(x) and T β(⌊ x
2 ⌋) for some odd x ≥ 3.
Starting from a configuration in which Tsgl(x) is enabled, by executing Tsgl(x), the adversary can
execute Tsgl(x) on x−1 robots, moving the half of them on the right side and the half of them on the
2 ⌋-sequence, depending
left side. By doing that, the system reaches either a towerless 1.block or an S⌊ x
on whether x = 3 or x ≥ 5, respectively. In both cases, T β(⌊ x
2 ⌋) is enabled on each robot belonging
to an extremity of the 1.block (if x = 3, then T β(⌊ x
2 ⌋) corresponds to Rule Rin). By executing
T β(⌊ x
2 ⌋) on them, the adversary can bring the system in a configuration that is undistinguishable
from the previous one. A contradiction (Theorem 2.1).
✷
Lemma 3.8 Let P be a semi-synchronous protocol for φ = 1 and 5 ≤ k < n.
exploration problem, then for every x ≥ 1, P includes T β(x) only.
If P solves the
Proof. The lemma is proven for x = 1 by Corollary 3.1. Assume by contradiction that there exists
x ≥ 2 such that P does not include Rule T β(x) only. From Lemma 3.5, there exists some executions
of P leading to a configuration containing an S2-sequence. Let us consider the smallest x satisfying
the above condition. So, starting from an S2-sequence, for every y ∈ [1, x − 1], by activating Rule
T β(y) on every robot at the extremity of the Sy-sequence, in x − 1 steps, the adversary can lead
the system into a configuration γx−1 containing an Sx-sequence. Since P does not include only Rule
T β(x), there are two cases:
1. P includes Rule T α(x). In that case, by activating x − 1 robots on each x.tower of the Sx-
sequence, the system reaches an Sx−1-sequence. This configuration is undistinguishable with
the configuration γx−2, the previous configuration of γx−1 in the execution.
2. P does not include Rule T α(x). Then, the neighbors of the towers are the only robots that
can be activated with either T γ(x) or T δ(x). In the first case (T γ(x) is enabled), the system
reaches a configuration similar to the one shown in Figure 3.1, i.e., a 1.block surrounded by two
isolated x + 1.towers. The size of the 1.block is equal to k − 2(x + 1). By Lemma 3.6, P does
not include Tsgl(x + 1). So, none of the robots belonging to a tower is enabled. Furthermore,
the local view of both extremities of the 1-block being similar to the initial configuration γ0,
from Corollary 3.1, both can be activated with Rule Rin by the adversary. Then, the system
reaches a configuration similar to the one shown in Figure 3.2, i.e., an S2-sequence, surrounded
with 2 towers and separated by one single empty node. In the second case (T δ(x) is enabled),
the system reaches a configuration that is also similar to the one shown in Figure 3.2. Following
9
the same reasoning as above, the system reaches a configuration similar to the one shown in
Figures (3.3, 3.4) or in Figures (3.5, 3.6) depending on whether k is odd or even.
x+1
x+1
l
l
Figure 3.1: T γ(x) is executed.
x
x
x
x
Figure 3.2: T δ(x) is executed
x
x
x
x
x+1
x+1
x+1
x+1
x+1
x+1
x+1
x+1
Figure 3.3: The case where κ = 1
x
x
x
x
Figure 3.4: The case where κ is odd and κ ≥ 3
x
x
x
x
x+1
x+1
x+1
x+1
x+1
x+1
x+1
x+1
Figure 3.5: The case where κ = 0
Figure 3.6: The case where κ is even and κ ≥ 2
Let τ , S, and κ be the number of isolated towers, their size, and the number of robots sur-
rounded by τ towers, respectively. By construction, τ = ⌊ k
S ⌋, κ = k mod (τ S), and κ ≤ 2x.
There are two cases to consider:
(a) κ is odd. Again consider two cases:
i. κ ≥ 3. Then, ⌊ κ
2 ⌋ < x and by assumption, for every y ∈ [1, ⌊ κ
2 ⌋], P includes the rules
T β(y). So, the adversary can lead the system into a configuration containing a single
κ+ 1.tower surrounded with τ towers. Since κ
2 < x and since by assumption, for every
y ∈ [2, κ
2 ⌋], P includes Rule T β(y), by Lemma 3.7, P does not include Rule Tsgl(κ).
So, no robot of the κ.tower is enabled. So, the system is at a deadlock and some
nodes are not visited by at least one robot. This contradicts that P is an exploration
protocol.
ii. κ = 1. The only applicable rules on the single robot is either (i) Rule Rsgl or
(ii) Rule Rsp defined as follow: x(1)x ::← ∨ → depending on whether P includes
T γ(x) or T δ(x), respectively. In the first case (i), after executing the rule, the single
robot is located on the neighboring node of one of the towers, i.e., a configuration
containing 0(y)10 where y is equal to x + 1. The only possible action is then to create
10
κ′
κ′
Figure 3.7: The reached configuration when κ > 2.
an y + 1.tower.
If P includes Tsgl(y + 1), then by selecting a single robot of the
y + 1.tower, the adversary can lead the system into the previous configuration. A
contradiction (Theorem 2.1). In the second case (ii), an x + 1.tower is created. As in
Case (i), if P includes Tsgl(x + 1), then by selecting a single robot of the x + 1.tower,
the adversary can lead the system into the previous configuration. A contradiction
(Theorem 2.1).
(b) κ is even. There are three cases.
i. κ = 0. Since none of the robots of τ towers is enabled, the system is at a deadlock.
This contradicts that P is an exploration protocol.
ii. κ = 2.
in this case there is a single 1.block of size 2 having either two neighboring
x.towers at distance 1 or x + 1.towers at distance 2 depending on whether P includes
T γ(x) or T δ(x), respectively. In both cases, the only applicable rule is T δ(x). Assume
that the scheduler activates both robots at the same time. Then, the system remains
into an undistinguishable configuration. A contradiction.
iii. κ > 2. Then, κ
2 < x and by assumption, for every y ∈ [2, κ
2 ⌋], P includes the rules
T β(y), the adversary leads the system into a configuration containing two neighbor-
ing κ′.towers, κ′ = κ
2 -- refer to Figure 3.7. Assume that P includes Rule Tdbl(κ′)
defined as follows: 0(κ′)κ′ ::←. Then by choosing to execute Rule Tdbl(κ′) on κ′ − 1
robots on each tower (since Rule Rout cannot be a part of P, κ > 2), the adversary
brings the system into an Sκ′−1-sequence that is undistinguisable from the previous
configuration. If Rule Tdbl(κ′) is defined as follows: 0(κ′)κ′ ::→. Then by choosing
to execute Rule Tdbl(κ′) on the same number of robots on each κ′.tower, the sys-
tem remains into an undistinguishable configuration. In both cases, this contradicts
Theorem 2.1.
✷
Let us now suppose that γ0 is any arbitrary starting configuration. The following lemma holds:
Lemma 3.9 Let P be a semi-synchronous protocol for φ = 1 and 2 ≤ k < n.
exploration problem, then Rule Rout and Rule Rsgl are mutually exclusive with respect to P.
If P solves the
Proof. Assume by contradiction that P includes both Rule Rout and Rule Rsgl. Note that
Rule Rout (respectively, Rule Rsgl) is enabled only on robots that are located at the border of an
1.block (resp., on isolated robots). Suppose that the initial configuration γ0 contains two isolated
robots, r1 and r2, that is at distance 2 from a 1.block, one at each side. By executing Rsgl on
both r1 and r2, the adversary can bring the system in a configuration γ1 where both r1 and r2 are
located at the border of the 1.block. Then, Rule Rout becomes enabled on both r1 and r2. By
executing Rout on both r1 and r2, the adversary brings the system in a configuration γ2 that is
undistinguishable from γ0. A contradiction.
✷
11
Lemma 3.10 Let P be a semi-synchronous protocol for φ = 1 and 2 ≤ k < n. If P solves the
exploration problem, then P does not include Rule Rout.
Proof.
By contradiction, assume that P includes Rule Rout. Note that Rule Rout is only
enabled on robots that are at the border of a 1.block. Suppose that γ0 contains two 1.blocks at
distance 2 from each other. Assume that the scheduler activates only one extremity of one of the
two 1.blocks. Let us refer to the 1.block from which one extremity is activated by B1 and let refer
to the other 1.block by B2. Let uiui+1 . . . ui+m be the 1.block B1 in γ0 and let ui−2ui−3 . . . ui−m′ be
the 1.block B2 (m and m′ are the size of respectively B1 and B2). Denote rj the robot located on
node ui+j (0 ≤ i < k). Without lost of generality, assume that the adversary activates r0 only in γ0.
Once r0 moves, it joins B2 in γ1. Assume that in γ1 the scheduler activates r1 that executes Rout
again. More generally, assume that ∀ i ∈ [1, . . . , m − 2], in γ1+i, the adversary activates ri+1. Once
the configuration γk−1 is reached, the scheduler activates the same robots in the reverse order (note
that these robots are part of B2). At each time, one robot is activated, it joins back B1. When
m − 2 robots are activated, the configuration reached is undistinguishable from γ1. A contradiction.
✷
Observe that both Lemmas 3.2 and 3.3 are valid when γ0 is any towerless configuration. In the
case where there are isolated robots that are at distance 2 from a 1.block, the scheduler activates
them, such that they join a 1.block by executing Rsgl. The only rule that can be enabled is Rin.
Each 1.block, behaves independently from the other 1.blocks. Lemmas 3.4 to 3.8 are also valid.
Therefore:
Theorem 3.1 No deterministic exploration protocol exists in the semi-synchronous model (ATOM)
for φ = 1, n > 1, and 1 ≤ k < n.
Proof. The theorem is a direct consequence of Lemma 3.4 and Lemma 3.8 and its proof.
✷
Since the set of executions that are possible in CORDA is a strict superset of those that are
possible in ATOM, Theorem 3.1 also holds in the CORDA [16].
Corollary 3.2 No deterministic exploration protocol exists in the asynchronous model (CORDA)
for φ = 1, n > 1, and 1 ≤ k < n.
Synchronous Model.
Theorem 3.2 Let P be a synchronous exploration protocol for φ = 1 and 2 ≤ k < n. If n > 7, then,
k must be greater than or equal to 5.
Proof.
The proof is by contradiction. Assume that there exists a synchronous exploration protocol P
for φ = 1, 2 ≤ k < 5, and n > 7.
Let us start from a possible starting configuration, the idea is to derive all the possible executions
and then show that non of them can ensure the completion of the exploration task. The cases bellow
are possible:
1. k = 1. It is clear that no exploration is possible by a single robot on the ring.
12
2. k = 2 . The robots in this case, either move towards each other exchanging their position or,
they move in the opposite direction and become isolated robots. In the latter case, once they are
activated again, the scheduler brings them back to their previous position. The configuration
reached is in both cases is undistinguishable from the starting one. A contradiction.
3. k = 3.
(i) In the case where only Rule Rout is enabled. Once the rule is executed, all
robots in the configuration become isolated robots. Thus no exploration is possible (all the
configurations that are created after are undistinguishable). (ii) If Rule Rin is enabled alone,
then the configuration reached contains a single 3.tower. All robots part of the 3.tower behave
as a single robot. According to Case (2), no exploration is possible in this case. (iii) Rule
Rswp, is the only one enabled.
In this case the configuration reached contains an isolated
robot and an isolated 2.tower (let refer to this configuration by T ). Note that only robots part
of the tower are enabled. Suppose that the scheduler activates them such that they move in
the opposite direction of the isolated robots, and then, it activates them again but this time,
they move in the opposite direction (towards the isolated robot). The configuration reached is
undistinguishable from T configuration. Thus no exploration is possible. (iv) If Rules Rswp
and Rin are enabled at the same time, each node of the ring is visited by robots by executing
the following two rules: 0(2)1 ::→ and 2(1)0 ::→, however, robots are not able to detect the end
of the exploration. Thus, no exploration is possible in this case. Finally, if (v) Rules Rswp and
Rout are enabled at the same time. Once such rules are executed, the configuration reached
contains one 1.block of size 2 and one isolated robots. By executing Rule Rout (possibly
Rsgl), a configuration with just isolated robots is eventually reached. Thus, no exploration is
possible in this case too.
4. k = 4. Suppose that the initial configuration contains a single 1.block of size 4. The cases
bellow are possible:
(a) Rule Rswp is executed. The two robots inside the 1.block exchange their position and the
configuration reached is undistinguishable from the starting one. A contradiction.
(b) Rule Rin is executed. Two neighboring towers are then created.
If the towers move
towards each other, the configuration reached is undistinguishable from the starting one.
If they move in the opposite direction of each other, then they become both isolated towers
i.e., the robots in the tower cannot see any neighboring robot. When they are activated
again, the scheduler brings them back to their previous position. Thus, in this case too,
the configuration reached is undistinguishable from the starting one. A contradiction.
(c) Rule Rout is executed. The robots that were at the extremity of the 1.block become
isolated robots. if Rule Rsgl is enabled then two isolated towers are created (Note that
there is at least one node that has not been explored yet). Robots in the same tower
behaves as a single robot. Thus, the configuration is similar to the one containing only
two robots. According to Case (2), no exploration is possible in this case. Hence, Rule
Rsgl cannot be enabled. Rule Rout is the only one that is enabled on robots part of the
1.block of size 2. Once they move, two 1.blocks of size 2 are created. All the robots have
now the same view. Note that Rule Rout remains the only one enabled on all the robots.
Once it is executed, a new 1.block of size 2 is created and the configuration contains
either (i) another 1.block of size 2 (n = 8) or (ii) a tower (n = 7) or (iii) two isolated
robots (n > 8). In the first case (i), Rout remains enabled on all the robots. Once it
is executed, the configuration reached is undistinguishable from one of the previous ones.
13
Thus, no exploration is possible. In the second case (ii), Rout remains enabled on robots
part of the 1.block. Observe that the robots in the tower cannot be activated otherwise,
an undistinguishable configuration is reached (the one that contains two 1.blocks). Thus
only Rout is enabled. Once it is executed, the configuration contains two isolated robots.
Such isolated robots can never detect the end of the exploration task since no rule can
be executed anymore. Thus no exploration is possible in this case too. In the third case
(iii) on robots part of the 1.block, Rule Rout is enabled, once they move all the robots
become isolated robots. Hence, the scheduler can choose always the same direction for
all the robots. The configuration reached at each time is then undistinguishable from the
previous one. A contradiction.
(d) All the robots are allowed to move. Note that in this case either two towers are created
and it is clear that no exploration is possible (since each tower behaves as a single robot
and thus, the system behaves as there is two robots on the ring). or we retrieve case 4c.
Thus, no exploration is possible in this case too.
From the cases above, we can deduce that no exploration is possible for n > 7 using four robots even
in the fully-synchronous model.
✷
3.2 Synchronous Algorithms
In the following, we present optimal deterministic algorithms that solve the exploration problem
in the fully-synchronous model for any ring 3 ≤ n ≤ 6 and n ≥ 7, using respectively (n − 1)
robots (except for the case n = 6 that needs only 4 robots) and five robots. Note that the starting
configuration contains a single 1.block (refer to Figure 3.8).
Fully-Synchronous Exploration for k = 5 and n ≥ 7. The idea of Algorithm 1 is as follow: The
robots that are at the border of the 1.block are the only ones that are allowed to move in the initial
configuration, γ0. Their destination is their adjacent occupied node (Rule 1A1). Since the system is
synchronous, the next configuration, γ1, contains a single robot surrounded by two 2.towers. In the
next step, the towers move in the opposite direction of the single robot (Rule 1A3) and the single
robot moves towards one of the two towers (Rule 1A2). Note that the resulting configuration γ2
is 21020n−4, providing an orientation of the ring. From there, the single 2.tower are the landmark
allowing to detect termination and the three other robots explore the ring by perform Rules 1A3
and 2A4 synchronously. After n − 4 steps, n − 4 nodes are visited by the 3 robots and the system
reaches γn−2 that is equal to 2210n−3. Finally, by performing Rule 2A4, the single robot create a
3.tower, marking the end of the exploration in γn−1 in which each robot is awake of the termination.
Algorithm 1 Fully-Synchronous Exploration for n ≥ 7 (visibility 1)
1A1:
1A2:
1A3:
2A4:
0(1)1
2(1)2
0(2)1
2(1)0
// Move towards my occupied neighboring node
:: →
:: → ∨ ← // Move towards one of my neighboring node
// Move towards my neighboring empty node
:: ←
:: ←
// Move to the tower
Fully-Synchronous Exploration for 3 ≤ n ≤ 6. The formal description of the algorithm is
given in Algorithm 2. The robots in this case detect the end of the exploration task if they are
14
either part of a 2.tower or neighbors of a 2.tower. The idea of the algorithm is the following: For
3 ≤ n ≤ 5, k = (n − 1) robots are necessary to perform the exploration task. The robots that are
at the extremities of the 1.block are the ones allowed to move, since k = n − 1, once they move, a
2.tower is created. If the reached configuration contains an isolated robot (the case where n = 4)
then, this robot is the only one allowed to move. Its destination is one of its adjacent empty nodes
(refer to Rule 1A′2). Once it moves it becomes neighbor of the 2.tower. In the case where (n = 6),
4 robots are necessary to solve the exploration problem. In the same manner, the robots that are
at the border of the 1.block are the ones allowed to move, their destination is their adjacent empty
node (refer to Rule 1A′1). Once they move, the configuration reached contains two 1.blocks of size 2.
Observe that on all the robots Rule 1A′1 is enabled. Once it is executed, two 2.towers are created.
The robots can detect the end of the exploration since they are all part of a 2.tower.
Algorithm 2 Fully-Synchronous Exploration for 3 ≤ n ≤ 6 (visibility 1)
1A'1:
1A'2:
0(1)1
0(1)0
:: ←
:: ← ∨ → // Move towards one of my neighboring nodes
// Move towards my neighboring empty node
(a) Fully-synchronous exploration n = 3 and
k = 2
(b) Fully-synchronous exploration n = 4 and
k = 3
(c) Fully-synchronous exploration n = 5 and
k = 4
(d) Fully-synchronous exploration n = 6 and
k = 4
Figure 3.8: Fully-Synchronous Deterministic Solutions for 3 ≤ n ≤ 6.
4 Visibility φ = 2
In this section we present two deterministic algorithms that solve the exploration problem. The first
one uses 9 robots and works for any towerless initial configurations that contains a single φ.group.
The second one uses 7 robots but works only when the starting configuration contains a single 1.block
of size 7. In both algorithms we suppose that n ≥ φk + 1.
4.1 Asynchronous Exploration using 9 robots
Let us first define two special configurations:
Definition 4.1 A configuration is called Middle (refer to Figure 4.9) at instant t if there exists a
sequence of consecutive nodes ui, ui+1, . . . , ui+5, ui+6 such that:
15
T1
T2
R1
T'2
T'1
Figure 4.9: M iddle configuration
Figure 4.10: T erminal configuration
• Mj = 2 for j ∈ {i, i + 1, i + 5, i + 6}
• Mj = 1 for j = i + 2
• Mj = 0 for j ∈ {i + 3, i + 4}
Definition 4.2 A configuration is said Terminal (refer to Figure 4.10) at instant t if there exists a
sequence of nodes ui, ui+1, . . . , ui+4 such that:
• Mj = 2 for j ∈ {i, i + 3}
• Mj = 4 for j = i + 2
• Mj = 0 for j = i + 1
• Mj = 1 for j = i + 4
The algorithm comprises two phases as follow:
1. Organization Phase. The aim of this phase is to build the Middle Configuration. The
initial configuration is any towerless configuration that contains a single φ.group.
2. Exploration Phase. The starting configuration of this phase is the Middle Configuration.
A set of robots are elected to perform the exploration while other robots stay still as a land
mark. Terminal configuration is created at this end of this phase to indicate the end of
exploration task.
The formal description of the algorithm is presented in Algorithms 3.
Algorithm 3 Asynchronous Exploration using 9 robots (φ = 2)
Organization Phase
2A'1:
2A'2:
2A'3:
2A'4:
2A'5:
2A'6:
2A'7:
00(1)01
00(1)1?
00(2)01
02(1)01
21(1)01
21(1)1?
20(1)02
:: →
:: →
:: →
:: →
:: →
:: ←
:: ← ∨ →
Exploration Phase
2A'6:
2A'8:
2A'9:
2A'10:
2A'11:
2A'12:
2A'13:
2A'14:
2A'15:
21(1)1?
00(2)21
01(1)21
20(2)10
20(1)00
20(2)21
21(1)21
02(2)22
02(1)32
:: ←
:: ←
:: ←
:: ←
:: ←
:: ←
:: ←
:: →
:: →
// Move toward the occupied node
// Move to my adjacent occupied node
// Move toward the occupied node
// Move in the opposite direction of the tower
// Move in the opposite direction of the tower
// Move toward the tower
// Move to one of my neighboring nodes
// Move toward the tower
// Move in the opposite direction of my neighboring tower
// Move in the opposite direction of the tower
// Move toward the tower at distance 2
// Move toward the tower
// Move to my adjacent free node
// Move in the opposite direction of my neighboring tower
// Move toward the tower having another tower as a neighbor
// Move toward the 3.tower
16
Proof of correctness We prove in this section the correctness of our algorithm presented above.
In the following, a configuration is called intermediate if Rule 2A′7 is enabled in the configuration.
Lemma 4.1 Starting from any towerless configuration that contains a single φ.group, intermediate
configuration is eventually reached.
Proof. Two cases are possible as follow:
• The initial configuration contains a 1.block of size 9. Rule 2A′2 is enabled on both robots that
are at the extremity of the 1.block. When the scheduler activates such robots, a S2-sequence
is created. Rule 2A′6 becomes enabled on the robots at distance 2 from the 2-towers. When
such robots are activated by the scheduler, a configuration of type intermediate is created.
Since we consider in this paper asynchronous robots, one tower can be created only at one
extremity of the 1.block, Rule 2A′6 becomes enabled on the robots at distance 2 from such a
2-tower. However, Rule 2A′2 keeps being enabled on the other extremity (let the robot at the
other extremity be r1). Since we consider a fair scheduler, r1 will be eventually enabled and
a 2-tower is created. Rule 2A′6 becomes enabled on the robots at distance 2 from the new
2-towers. Once such robot is activated, intermediate configuration is created.
• Other configurations. Two sub-cases are possible:
-- Case (a): Robots at the extremity of the φ.group do not have any neighboring robots:
Rule 2A′1 is enabled on such robots. When the scheduler activates them, they become
neighbor of an occupied node and we retrieve Case (b).
-- Case (b): Robots at the extremity of the φ.group have one occupied node as a neighbor.
Rule 2A′2 becomes enabled on such robots. When the scheduler activates them, a 2-tower
is created at at least one extremity of the φ.group.
When a 2-tower is created, four sub cases are possible as follow:
-- Case (c): The tower does not have a neighboring occupied node (there is a sequence of
nodes with multiplicities equal to 00201). In this case, Rule 2A′3 becomes enabled on the
robots part of the 2-tower. If the scheduler activates both robots at the same time then
the 2-tower becomes neighbor to an occupied node. (i) If the scheduler activates only
one robot, then the tower is destroyed and Rule 2A′2 becomes enabled on the robot that
did not move. When the rule is executed, the tower is built again and the configuration
reached is exactly the same as in (i).
-- Case (d): The Tower is neighbor of an occupied node that is part of a 1.block of size
3 (002111). Rule 2A′6 is enabled in this case on the robot that is in the middle of the
1.block. Once the Rule is executed, another 2-tower is created.
-- Case (e): The Tower is neighbor of an occupied node that is part of a 1.block of size 2
(002110). Rule 2A′5 is enabled in this case on the robot that is in the extremity of the
1.block not having the 2-tower as a neighbor. When the rule is executed the robot that
has moved becomes part of another 1.block and we retrieve Case (f).
-- Case (f ): The Tower is neighbor of an occupied node that has an empty node as a
neighbor (00210). Rule 2A′4 becomes enabled on the robot that is at distance 1 from
the tower. Once it moves, it becomes neighbor of another occupied node and we retrieve
either Case (c).
17
From the cases above we can deduce that a configuration with a sequence of node 2111 is
reached in a finite time (Cases (c), (e) and (f)). In this case, Rule 2A′6 becomes enabled on
the robot that is in the middle of the 1.block of size 3 having a neighboring 2-tower (Case (d)).
Once the robot moves, a 2-tower is created (2201). Note that the single robot in the sequence
is not allowed to move unless it sees a 2-towers at distance 2 at each side. Thus, in the case the
scheduler activates robots in only one extremity of the φ.group, we are sure that robots in the
other extremity will be eventually activated since they are the only one that will be allowed
to move (the scheduler is fair). Hence, a symmetric configuration in which there will be one
robot on the axes of symmetry and two 2-tower at each side of it is reached in a finite time.
intermediate configuration is then reached and the lemma holds.
✷
Lemma 4.2 Starting from intermediate configuration, Middle configuration is eventually reached.
Proof. When the configuration is of type intermediate, only 2A′7 is enabled. Once the rule
is executed, the robot that is not part of any tower becomes neigbor of a tower of size 2. Middle
configuration is reached and the lemma holds.
✷
Lemma 4.3 Starting from any towerless configuration that contains a single φ.group, a configuration
of type Middle is eventually reached.
Proof. Directly follows from Lemmas 4.2 and 4.3.
✷
Lemma 4.4 Starting from a configuration of type Middle, Terminal configuration is eventually
reached.
Proof.
Let T ′1 and T ′2 be the two neigboring 2-towers that cannot see any other robot, and
let T 1 and T 2 be the other 2-towers such as T 2 has an neighboring occupied node other then T 1 (let
refer to the robot that is neighbor to T 1 and not part of a 2-tower by R1) (refer to Figure 4.9). The
sequence of robots containing T 1, T 2 and R1 is called Explorer-Sequence. On such a configuration,
Rule 2A′8 is enabled on T 1. When the rule is executed, either (i) T 1 becomes at distance 2 from T 2
(the scheduler activates both robots part of T 1 at the same time) or (ii) T 1 is destroyed and in this
case Rule 2A′9 becomes the only rule enabled. When 2A′9 is executed, T 1 is restored and becomes
at distance 2 from T 2 (we retrieve Case (i)). On the new configuration Rule 2A′10 is enabled on
T 2. When the rule is executed, either (a) T 2 becomes neighbor to T 1 (the scheduler activates both
robots in T 1) or (b) T 2 is destroyed and Rule 2A′6 becomes the only one enabled (the scheduler
activates only one robot in T 2). When the rule is executed, T 2 is built again and becomes neighbor
of T 1 (we retrieve Case (a)). Rule 2A′11 becomes then enabled on R1. When R1 is activated,
Explorer-Sequence is built again, however, the distance between T 1 and T ′1 has decreased. Since
the view of robots in the Explorer-Sequence is the same as in the Middle configuration, they will
have the same behavior. Thus, at each time the distance between T 1 and T ′1 decreases. Hence T 1
becomes eventually at distance 2 from T ′1 in a finite time. Rule 2A′12 becomes enabled on T 1. If the
scheduler activates only one robot in T 1 then T 1 is destroyed and Rule 2A′13 becomes enabled, when
the rule is executed, T 1 becomes neighbor of T ′1 (it is like the scheduler activates both robots at
the same time). Since robots in T 2 cannot see the 2-tower T ′1, they will move towards T 1, the same
18
holds for R1. Rule 2A′14 is then enabled on T 2, once it is executed, either a 4-tower is created and
in this case the configuration is of type Terminal or 3-tower is created and in this case Rule 2A′15
becomes enabled, when it is executed a 4-tower is created and the configuration becomes T erminal
and the lemma holds.
✷
Lemma 4.5 Starting from any towerless configuration that contains a single φ.group, Terminal
configuration is eventually reached such that all nodes of the ring have been explored.
Proof. From both lemma 4.3 and 4.4 we can deduce that starting from any initial configuration,
Terminal is reached in a finite time. In another hand, from Lemma 4.3 we are sure that Middle
configuration is reached in a finite time. Note that the only way to build the second towers (T 2 and
T 2′) is to have a tower neighbor of a consecutive sequence of three robot. Since the distance between
R1 and T ′2 is equal to 3, we are sure that these nodes have been occupied and thus they have been
explored. In another hand T 1, T 2 and R1 keep moving towards T 1′ until they will become neighbor.
Hence, all the nodes that were between T 1 and T 1′ have been all visited. We can deduce then that
all the nodes of the ring have been explored and the lemma holds.
✷
4.2 Asynchronous Exploration using 7 robots
Recall that the initial configuration in this case contains a 1.block of size 7. The idea is similar to
the one used in Algorithm 3. Let us first define two spacial configurations:
Definition 4.3 A configuration is called Inter if there exists a sequence of seven nodes ui, ui+1, . . .
, ui+5, ui+6 (refer to Figure 4.11) such that:
• Mj = 2 for j ∈ {i, i + 1, i + 6}
• Mj = 0 for j ∈ {i + 3, i + 4, i + 5}
• Mj = 1 for j = i + 2
Definition 4.4 A configuration is called Final if there exists a sequence of four nodes ui, ui+1, . . .
, ui+3 (refer to Figure 4.12) such that:
• Mj = 2 for j = i
• Mj = 4 for j = i + 2
• Mj = 0 for j = i + 1
• Mj = 1 for j = i + 3
The algorithm comprises two phases as follow:
1. Preparation Phase: The aim of this phase is to build an Inter Configuration. The starting
configuration contains a single 1.block of size 7.
2. Exploration Phase: In this phase a set of robots are elected to perform the exploration task.
At the end of the exploration Final Configuration is built to indicate the end of the exploration
task.
19
T1
T2
R1
T'1
Figure 4.11: IN T ER
Figure 4.12: F inal configuration
The formal description of the algorithm is given in Algorithm 4.
Algorithm 4 Asynchronous Exploration using 7 robots (φ = 2)
Preparation Phase
2A1:
2A2:
2A3:
2A4:
2A5:
2A6:
2A7:
01(1)11
20(1)1?
10(1)01
00(1)20
10(1)02
21(1)10
20(1)00
:: ←
:: ←
:: ← ∨ →
:: ←
:: ←
:: ←
:: ←
Exploration Phase
2A6:
2A7:
2A8:
2A9:
2A10:
2A11:
2A12:
2A13:
2A14:
21(1)10
20(1)00
00(2)21
01(1)21
20(2)10
20(2)21
21(1)21
02(2)21
02(1)32
:: ←
:: ←
:: ←
:: ←
:: ←
:: ←
:: ←
:: →
:: →
// Move to my adjacent node having a free node as a neighbor
// Move to my adjacent free node
// Move to one of my neighboring node
// Move to my adjacent free node
// Move in the opposite direction of the tower
// Move toward the 2.tower
// Move towards the 2.tower
// Move toward the 2.tower
// Move toward the 2.tower
// Move to my adjacent free node
// Move to my adjacent node having one free node as a neighbor
// Move to my adjacent free node
// Move to my adjacent free node
// Move in the opposite direction of the neighboring 2.tower
// Move to my adjacent node having one robot as a neighbor
// Move toward the 3.tower
Proof of correctness We prove in the following the correctness of our algorithm.
Lemma 4.6 Starting from a configuration that contains a single 1.block, a configuration of type
Inter is eventually reached.
Proof. When the configuration contains a single 1.block of size 7, Rule 2A1 is enabled on
the robots that are neighbor of the extremity of the 1.block. Once such robots move, a 2-tower
is created at each extremity of the 1.block. Rule 2A2 becomes enabled on the robots that are at
distance 2 from the 2-towers. Once they move, they become neighbor of a 2-tower. Note that in the
case the scheduler activates only robots at one extremity, we are sure that the configuration that
will be reached in this case is exactly the same as when the scheduler activates both robots at each
extremity since no robot will be able to move besides the ones at the other extremity. In another
hand, since robots that did not move in the other extremity are not aware of the changes in the
current configuration (recall that they are only able to see at distance 2), they will have the same
behavior as when they are activated at the same time with the robots at the other extremity. Rule
2A3 becomes then enabled. Since the view of the robot on which Rule 2A3 is enabled, is symmetric
the scheduler is the one that chooses the direction to take. Once such robot moves, it becomes
neighbor of a single robot and it will be able to see one 2-tower.
In the resulting configuration,
Rule 2A4 becomes enabled on the robot that is neighbor of a 2-tower having an empty node as a
neighbor. Once this robot moves, it becomes able to see the other part that it could not see before
moving. This robot continue to move towards the single robot until they become neighbor (refer to
Rule 2A5). Observe that one tower remains unseen by the other robots. The robot that is in the
middle of the 1.block of size 3 that is neighbor of the tower of a 2-tower is the only one that can
20
move (refer to Rule 2A6). When this robot moves, a new 2.tower is created. The robot that is at
distance 2 from the new 2.tower (note that there is only one such robot) is the only robot allowed
to move, its destination is its adjacent empty node towards the 2.tower (refer to Rule 2A7). When
it moves it becomes neighbor of the tower of size 2. Inter configuration is then reached and the
Lemma holds.
✷
Lemma 4.7 Starting from a configuration of type INTER, a configuration of type Final is eventually
reached.
Proof. When Inter configuration is reached, a subset of robots are elected to perform the
exploration task. These robots are T 1, T 2 and R1 (refer to Figure 4.11). Observe that these robots
form exactly the same sequence (Explorer-Sequence) as in Middle configuration (refer to Algorithm
3). It has been shown in Lemma 4.4 that this subset of robots keep moving towards T ′1 until they
become neighbor of T ′1. Rule 2A13 becomes then enabled on the robots part of T 1. If the scheduler
activates both robots at the same time, a 4-tower is created and the configuration reached is Final.
In the case the scheduler activates only one robot from T 1 then a 3-tower is created. There will be
one robot not part of a tower that is neighbor to the 3-tower. This robot is the one allowed to move
(refer to Rule 2A14). Once the rule is executed a 4-tower is created and the configuration is Final.
Thus, the lemma holds.
✷
Lemma 4.8 Starting from a configuration that contains a single 1.block of size 7, a configuration of
type F inal is eventually reached such that all the nodes of the ring have been explored.
Proof.
From Lemma 4.6 and 4.7 we can deduce that starting from a configuration that
contains a single 1.block, Final configuration is reached in a finite time. In another hand, when
Inter configuration is reached, we are sure that the nodes between R1 and T ′1 have been explored
since these nodes have been occupied by robots at the beginning by the 1.block (refer to Figure
4.11). The sub set of robots that have performed the exploration task (T 1, T 2 and R1) keep moving
towards T ′1 until it becomes neighbor of T ′1. Thus we are sure that the nodes that were between T 1
and T ′1 in Inter configuration have been visited. Thus all the nodes of the ring have been explored
and the lemma holds.
✷
5 Visibility φ = 3
In this section we first prove that no exploration is possible with 4 robots when n > 13. We then
present two deterministic algorithms that solves the exploration. The first one uses 7 robots and
works starting from any towerless configuration that contains a single φ.group. The second one uses
5 robots and works only when the starting configuration contains a single 1.block of size 5. In both
solutions, we suppose that n ≥ kφ + 1.
5.1 Negative Results
In the following, let us consider a configuration γt in which there is a 3.group of size 4 (let call
this configuration Locked). Let ui, ui+1, . . . , ui+9 be the nodes part of the 3.block in γt such as
ui, ui+3, ui+6, ui+9 are occupied. Let r1 and r4 be the robots that are at the extremity of the 3.block
such tha r1 is on ui and let r2 and r3 be the robots that are inside the 3.block such that r2 is on
ui+3. The following three rules are the only ones can be enabled:
21
1. 000(1)001 :: ←
2. 000(1)001 :: →
3. 100(1)001 :: ← ∨ →
Lemma 5.1 Let P be a semi-synchronous exploration protocol for φ = 3, n > 13, and k < n. Then,
P does not include Rule 1.
Proof. By contradiction, assume that P includes Rule 1. Note that Rule 1 is only enabled on r1
and r4. Assume that the adversary activates only r1 inγt. Let us refer to the resulting configuration
by ST (standing for "Second Trap" configuration). Once r1 moves, it becomes an isolated robot on
ui−1 in ST . ui+3 becomes the border of the 3.group. Assume that in ST , the adversary activates r2
(on ui+1) that executes Rule 1. Then, Configuration ST + 1 includes two 3.groups disjoint by three
nodes. The former includes r1 and r2 (on ui−1 and ui, respectively). The latter forms the sequence
r3 . . . r4, located on ui+6 . . . ui+9. Suppose that the scheduler activates r3 that is on node ui+6 that
executes Rule 1. Once it moves, the resulting configuration ST + 2 is undistinguishable from ST .
A contradiction (Theorem 2.1).
✷
Lemma 5.2 Let P be a semi-synchronous exploration protocol for φ = 3, n > 13, and k < n. Then,
P does not include Rule 2.
Proof. By contradiction, assume that P includes Rule 2. Note that Rule 2 is only enabled on
r1 and r4. Suppose that the scheduler activates both r1 and r4 at the same time. Let ST ′ be the
reached configuration once such robots move. The following rules are then possible:
a. 010(1)001 :: ←
b. 010(1)001 :: →
c. 000(1)010 :: →
• Case (a): Rule(a) is enabled. Assume that the scheduler activates both r2 and r3 at the same
time. Once they move, all robots on the ring have the same view. (i) If they execute 000(1)100
:: → then suppose that the scheduler activates them all at the same time, then the configuration
reached is undistinguishable from the previous one that contradicts Theorem 2.1. (ii) if they
execute 000(1)100 :: ←. Assume that the scheduler activates only r2 and r3. Once the robots
move the configuration reached is indistinguishable from ST ′. A contradiction (Theorem 2.1).
• Case (b): Rule (b) is enabled. Assume that the scheduler activates both r2 and r3 at the
same time. r2 and r3 become then neighbors. The following Rules are possible:
b1. 000(1)001 :: ←
b2. 000(1)001 :: →
b3. 100(1)100 :: →
-- Case(b1). Rule (b1) is enabled. Suppose that the scheduler activates both r1 and r4
at the same time. Once the robots move they cannot see any occupied node (recall that
n > 13). Thus if they execute rule 000(1)000 :: ← ∨ →, the robot move back to the
previous location. A contradiction (Theorem 2.1).
22
-- Case(b2). Rule (b2) is enabled. Suppose that the scheduler activates both r1 and r4 at
the same time. (i) If r1 and r4 execute rule 000(1)011 :: →, a 1.block is created. If the
robots inside the 1.block are the ones allowed to move, they either exchange their position
(A contradiction, Theorem 2.1). Or two 2-towers are created. Suppose that the scheduler
always activate robots part of the 2-tower at the same time. Thus, they will have the
same behavior and act as a single robots. A contradiction (From [3]: no exploration is
possible with only 2 robots even if the view is infinite). (ii) If rule 101(1)010 :: → enabled
then, if the scheduler activates both robots at the same time no exploration is possible
since whatever the robots that move an indistinguishable configuration is reached. A
contradiction (Theorem 2.1).
-- Case(b3). Rule (b3) is enabled. Suppose that the scheduler activates r2 and r3 at the
same time. They will simply exchange their position. The configuration that is reached
is undistinguishable from the previous one. A contradiction (Theorem 2.1).
• Case (c): Rule (c) is enabled. Once r1 and r4 move, two 1.blocks are created. If r2 and r3
are the ones allowed to move then if they execute 001(1)001:: ←, two towers are created and
no exploration is possible (refer to [3]). If they execute 001(1)001:: →, a new 1.block of size 2
is created. Note that in the reached configuration r2 and r3 cannot move anymore. If r1 and
r4 execute 000(1)011 :: →, a 1.block of size 4 is created and no exploration is possible (refer
to Case (b)).
If they execute 000(1)011 :: ←, they become at distance 3 from the 1.block.
Let refer to the reached configuration by T ′. If r1 and r4 keep being enabled on T ′, they can
only execute 000(1)001 :: ←. Suppose that the scheduler activates both robots at the same
time. Once they move they cannot see any other robot in T ′ + 1. (i) If 000(1)000 :: ← ∨ → is
enabled. Then suppose that the scheduler activates them at the same time such that they move
back to their previous position. Thus, T ′ + 2 is indistinguishable from T ′ + 1. Contradiction
(Theorem 2.1). (ii) if r2 and r3 are the enabled robots then if they execute 100(1)100 :: ←,
then the configuration reached is indistinguishable from the configuration reached when Rule
(b) is executed. Thus no exploration is possible in this case too.
✷
Lemma 5.3 Let P be a semi-synchronous exploration protocol for φ = 3, n > 13, and k < n. Then,
P does not include Rule 3.
Proof.
By contradiction, suppose that the scheduler activates both r2 and r3 at the same time. Once
they execute Rule 3, a 1.block is created. Such that robots in the 1.block cannot see any other
robots.
If r2 and r3 keep being enabled they can only move towards each other (otherwise the
previous configuration is restored). Suppose that the scheduler activates them at the same time, the
configuration reached is indistinguishable with the previous one. Contradiction (Theorem 2.1). If
r1 and r4 are the ones allowed to move then they can only execute 000(1)000 :: ← ∨ → (recall that
n > 13). Suppose that the scheduler activates them both at the same time such that they move in
the opposite direction of the 1.block (let refer to the reached configuration by IMP). Since 000(1)000
:: ← ∨ → keeps being enabled on both r1 and r4. Suppose that the scheduler activates them at
the same time such that they move back to their previous position. The configuration reached is
indistinguishable from IMP. Contradiction (Theorem 2.1).
✷
23
Lemma 5.4 No deterministic exploration is possible in the ATOM model for φ = 3, n > 13 and
k < n.
Proof.
Follows from Lemmas 5.1, 5.2 and 5.3.
✷
Lemma 5.5 No deterministic exploration is possible in the CORDA model for φ = 3, n > 13 and
k < n.
Proof.
Follows directly from Lemma 5.4.
✷
5.2 Two Asynchronous Algorithms
In the following, we present two deterministic algorithms that solve the exploration problem. The
first one uses 7 robots and works for any towerless initial configuration that contains a single φ.group
such that n ≥ kφ + 1. The second one uses 5 robots but works only when the starting configuration
contains a 1.block of size 5.
5.2.1 Exploration using k = 7.
Before detailing our solution, let us first define some special configurations:
Definition 5.1 A configuration is called Set at instant t (refer to Figure 5.13) if there exists a
sequence of 5 nodes ui, ui+1, . . . , ui+4 such that:
• Mj = 2 for j ∈ {i + 1, i + 4}
• Mj = 3 for j = i + 1
• Mj = 0 for j ∈ {i + 2, i + 3}
Definition 5.2 A configuration is called Final at instant t (refer to Figure 5.14) if there exists a
sequence of 4 nodes ui, ui+1, . . . , ui+3 such that:
• Mj = 2 for j ∈ {i, i + 1}
• Mj = 0 for j = i + 2
• Mj = 3 for j = i + 3
Figure 5.13: Set configuration
Figure 5.14: F inal configuration
The algorithm comprises two phases as follow:
24
1. Set-Up Phase. The aim of this phase is to create a Set Configuration. The starting config-
uration contains a φ.group, such that n ≥ k(φ) + 1.
2. Exploration Phase. The starting configuration of this phase is the Set Configuration. A
set of robots are elected to visit the ring's nodes. Final Configuration is created at the end
of this phase.
The formal Description of the algorithm is given in Algorithm 5.
Algorithm 5 Asynchronous Exploration using 7 robots (φ = 3)
Set-Up Phase
3A1:
3A2:
3A3:
3A4:
3A5:
3A6:
3A7:
3A8:
3A9:
3A10:
3A11:
3A12:
000(1)001
000(1)01?
000(1)1??
000(2)001
000(2)01?
002(1)001
002(1)01?
021(1)001
021(1)01?
021(1)120
220(1)200
022(1)020
Exploration Phase
3A13:
000(2)300
001(1)300
3A14:
000(2)030
3A15:
001(1)030
3A16:
200(3)000
3A17:
3A18:
201(2)00?
202(1)00?
3A19:
200(2)030
3A20:
201(1)030
3A21:
000(2)020
3A22:
3A23:
000(1)120
:: →
:: →
:: →
:: →
:: →
:: →
:: →
:: →
:: →
:: ← ∨ →
:: ←
:: ←
// Move toward the robot at distance 3
// Move toward the robot at distance 2
// Move toward my neighboring occupied node
// Move toward the robot at distance 3
// Move toward the robot at distance 2
// Move in the opposite direction of the tower
// Move in the opposite direction of the tower
// Move toward the robot at distance 3
// Move toward the robot at distance 2
// Move toward one of my neighboring node
// Move toward the tower at distance 2
// Move toward the neighboring tower
:: ←
:: ←
:: ←
:: ←
:: ←
:: ←
:: ←
:: ←
:: ←
:: →
:: →
// Move in the opposite direction of the 3.tower
// Move in the opposite direction of the 3.tower
// Move in the opposite direction of the 3.tower
// Move in the opposite direction of the 3.tower
// Move toward the 2.tower at distance 3
// Move to my neighboring occupied node
// Move to my neighboring occupied node
// Move in the opposite direction of the 3.tower
// Move towards the 2.tower at distance 2
// Move towards the 2.tower at distance 2
// Move towards my neighboring occupied node
Proof of correctness. We prove in the following the correctness of Algorithm 5.
Lemma 5.6 During the first phase, if there is a rule that is enabled on robot part of a 2.tower such
that the scheduler activates only one robot in the 2.tower (the tower is destroyed), then the tower is
built again and the configuration reached is exactly the same as when both robots in the 2.tower were
activated at the same time by the scheduler.
Proof.
The rules that can be executed by the robots in a 2.tower in the first phase are
Rules 3A4 and 3A5. In both Rules, robots in the tower can see only one robot at one side. Their
destination is their adjacent empty node towards the robot that can be seen. Suppose that the
scheduler activates only one robot in the 2.tower. In the configuration reached Rule 3A3 is enabled.
Note that the robot on which one of this rule is enabled is the one that was in the 2.tower and did
not move. Once it execute 3A3, it moves to its adjacent occupied node, thus a 2.tower is created
again and the configuration reached is exactly the same as when both robots were activated by the
scheduler. Hence the lemma holds.
✷
25
Lemma 5.7 Starting from any towerless configuration that contains a single φ.group, a configuration
containing an S2-sequence is eventually reached.
Proof. Two cases are possible as follow:
1. The starting configuration contains a single 1.block of size 7. Rule 3A3 is then enabled on the
two robots that are at the extremity of the 1.block. If the scheduler activates both robots that
the same time, two 2.towers are created and the lemma holds. If the scheduler activates only
one robot, then Rule 3A3 is the only rule enabled in the system. When the rule is executed,
the robot that was supposed to move moves to its neighboring occupied node and a 2.tower is
created. Thus in this case too the lemma holds.
2. Other cases. Let consider only one extremity of the φ.group. If the robot at this extremity (let
this robot be r1) does not have a neighboring occupied node, then either Rule 3A1 or 3A2 is
enabled. Once one of these rules is executed on r1, r1 becomes closer to an occupied node. Thus
it becomes neighbor of an occupied node in a finite time. Rule 3A3 becomes then enabled on
r1. Once r1 moves, a 2.tower is created. If the 2.tower does not have any neighboring occupied
node, either Rule 3A4 or 3A5 is enabled on robots in the 2.tower. According to Lemma 5.6,
robots in the same 2.tower move eventually as when they are activated at the same time. By
moving they become closer to an occupied node. Hence they become eventually neighbors of
an occupied node. (i) If the robot that is neighbor of the 2.tower has an empty node as a
neighbor then it moves to its adjacent empty node (refer to Rule 3A6). (ii) in the case it has
a neighboring occupied node besides the 2.tower, then if the robot on its neighboring occupied
node move in the opposite direction of the 2.tower if it has a neighboring empty node (refer to
Rules 3A7 and 3A8). By doing so the tower becomes neighbor of a 1.block of size 3. Robots
at the other extremity have the same behavior since the only rules that can be executed are
the ones that are enabled at the extremity. Thus, a configuration containing an S2-sequence
is reached in a finite time and the lemma holds.
✷
Lemma 5.8 Starting from a configuration containing an S2-sequence, Set configuration is eventu-
ally reached.
Proof.
In a configuration containing an S2-sequence, Rule 3A10 is enabled. When the rule
is executed, a new 2.tower is created. Rule 3A11 becomes then enabled on the robot not part of a
2.tower (let this robot be r1). Once it moves, only Rule 3A12 becomes enabled on r1. When the
robot is activated a 3.tower is created. Set configuration is then reached and the lemma holds. ✷
Lemma 5.9 Starting from a Set configuration, Final configuration is eventually reached.
Proof.
Let T 1 and T 2 be the 2.towers in the Set configuration such that T 1 has a neighboring
3.tower that we call T 3. Rule 3A13 is enabled on the robots that are part of T 1. If the scheduler
activates only one robot, Rule 3A14 becomes enabled on the robot that was supposed to move,
thus the configuration reached is exactly the same as when both robots in T 1 were activated at the
same time. Rule 3A15 becomes enabled on the T 1. Once the robots on T 1 are activated, the tower
becomes at distance 3 from T 3 (in the case the scheduler activates only one robot in T 1, only Rule
26
3A16 becomes enabled on the robot that was supposed to move, once the scheduler activates the
robot, the configuration reached is exactly the same as when both were activated at the same time).
Robots on T 3 are the only ones allowed to move (refer to Rule 3A17). Once they move, T 3 becomes
at distance 2 from T 1 (in the same manner, if the scheduler activates only some robots in T 3 then
either Rule 3A18 or Rule 3A19 is the only one enabled. Thus, T 3 is built again (the configuration
reached is similar to the one that was reached when all robots in the T 3 have moved at the same
time). Robots in T 1 are now the only one allowed to move and so on. Observe that the distance
between T 3 and T 2 increases while the distance between T 1 and T 2 decreases. When T 2 and T 1
becomes at distance 3. Rule 3A20 is then enabled on T 1 (if the scheduler activates only one robot
then Rule 3A21 is enabled and hence T 1 is built again). Robots in T 2 are now the only ones enabled
(refer to Rule 3A22), when they move, Final configuration is reached and the lemma holds (In the
case the scheduler activates only one robot in T 2, Rule 3A23 becomes enabled on the robot that did
not move. Once it is activated, Final configuration is reached and the lemma holds).
✷
Lemma 5.10 Starting from a any towerless configuration that contains a single φ.group, Final
configuration is eventually reached and all the nodes of the ring have been explored.
Proof.
From Lemmas 5.7, 5.8 and 5.9 we can deduce that Final configuration is eventually
reached starting from a towerless configuration that contains a single φ.group. In another hand, the
node between T 1 and T 2 in the Set configuration have been explored since they were occupied when
the configuration contained an S2-sequence. When the set of robots in charge of performing the
exploration task move on the ring, the distance between T 1 and T 2 decreases such that when Final
configuration is reached T 1 and T 2 become neighbor. Thus, we are sure that all the nodes of the
ring has been explored and the lemma holds.
✷
5.2.2 Exploration using k = 5.
In the following, we provide two definitions that are needed in the description of our algorithm:
Definition 5.3 A configuration is called Set2 at instant t (refer to Figure 5.15) if there exists a
sequence of 6 nodes ui, ui+1, . . . , ui+5 such that:
• Mj = 1 for j ∈ {i, i + 5}
• Mj = 3 for j = i + 1
• Mj = 0 for j ∈ {i + 2, i + 3, i + 4}
Definition 5.4 A configuration is called Done at instant t (refer to Figure 5.16) if there exists a
sequence of 4 nodes ui, ui+1, . . . , ui+3 such that:
• Mj = 1 for j ∈ {i, i + 1}
• Mj = 0 for j = i + 2
• Mj = 3 for j = i + 3
27
Figure 5.15: Set2
Figure 5.16: Done configuration
The algorithm comprises also two phases as follow:
1. Set-Up Phase. The initial configuration of this phase contains a φ.group such that n ≥ kφ + 1.
The aim of this phase is to create a Set2 configuration.
2. Exploration Phase. The starting configuration of this phase is Set2 Configuration. A set of
robots are elected to perform the exploration task. At the end of this phase, Done configuration
is created.
The formal description of our algorithm is given in Algorithm 6.
Algorithm 6 Asynchronous Exploration using 5 robots (φ = 3)
Set-Up Phase
3A'1:
3A'2:
3A'3:
3A'4:
3A'5:
3A'6:
011(1)110
120(1)100
012(1)010
300(1)000
120(1)101
300(1)001
Exploration Phase
000(1)300
000(1)030
100(3)000
101(2)000
102(1)000
100(1)030
010(1)030
010(1)300
3A'7:
3A'8:
3A'9:
3A'10:
3A'11:
3A'12:
3A'13:
3A'14:
:: → ∨ ←
:: ←
:: ←
:: →
:: ←
:: →
// Move to one of my neighboring node
// Move to towards the tower at distance 2
// Move to the tower
// Move in the opposite direction of the 3.tower
// Move toward the 2.tower (case n = 6)
// Move toward in the opposite direction of the 3.tower (case n = 7)
:: ←
:: ←
:: ←
:: ←
:: ←
:: ←
:: ←
:: ←
// Move in the opposite direction of the 3.tower
// Move in the opposite direction of the 3.tower
// Move toward the isolated robot
// Move to my neighboring occupied node
// Move to the 2.tower
// Move in the opposite direction of the 3.tower
// Move in the opposite direction of the 3.tower
// Move in the opposite direction of the 3.tower (case n = 6 ∧ n = 7 )
Proof of Correctness We prove in the following the correctness of our solution.
Lemma 5.11 Starting from a configuration that contains a 1.block of size 5 such that n > k + 2,
Set2 configuration is eventually reached.
Proof. When the configuration contains a 1.block of size 5, Rule 3A′1 is enabled. Once the
rule is executed a 2.tower is created. On the robot that is at distance 2 from the 2.tower (let this
robot be r1) Rule 3A′2 is enabled. When the rule is executed r1 becomes neighbor of the 2.tower.
Rule 3A′3 becomes the only one enabled then enabled. When r1 is activated, it executes the rule
and it joins the 2.tower and hence a 3.tower is created. Rule 3A′4 becomes the only rule enabled on
the robot that is at distance 3 from the 3.block, when the robot is activated, it becomes at distance
4 from the 3.tower. The Set2 configuration is then created and the lemma holds.
✷
Lemma 5.12 Starting from a configuration of type Set2, Done configuration is eventually reached.
28
Proof.
Let r1 be the robot that is neighbor of the 3.tower and let r2 be the robot that cannot see any
other robot. r1 is the only robot allowed to move and this until it becomes at distance 3 from the
tower (see Rules 3A′7 and 3A′8). Rule 3A′9 becomes enabled on the robots part of the 3.tower, their
destination is their adjacent empty node towards r1. In the case the scheduler activates some robots
in the tower, the remaining robots (the one that were supposed to move) are the only ones allowed
to move, their destination is their adjacent occupied node (see Rule 3A′10 and 3A′11). Thus the
configuration reached is exactly the same as when all the robots that are in the 3.tower are activated
by the scheduler at the same time. Hence, the tower becomes eventually at distance 2 from r1. r1
is now allowed to move, its destination is its adjacent empty node in the opposite direction of the
3.tower. Note that once it moves it becomes at distance 3 from this tower. Thus, robots in the
3.tower are the ones allowed to move and so on. Both the 3.tower and r1 keep moving in the same
direction such that at each time they move they become closer to r2 (the distance between r1 and
r2 decreases). Thus, r1 and r2 becomes eventually neighbors. Note that the robots in the 3.tower
cannot see r2 yet, so it will continue to move towards r1. When all of them move, Done configuration
is reached and the lemma holds.
✷
Lemma 5.13 Starting from a configuration that contains a 1.block, when Done configuration is
reached, and all the nodes of the ring have been explored.
Proof.
From Lemmas 5.11 and 5.12 we deduce that staring from any towerless configuration
that contains a single 1.block such that n > k + 2, Done configuration is reached in a finite time.
Let T be the 3.tower and let r1 (resp r2) be respectively the robots that is neighbor of the 3.tower
(resp the robot that cannot see any other robot). In Set2 configuration, the nodes between the T
and r2 has been already visited since the starting configuration was a 1.block. T and r1 keep moving
in the same direction such that the distance between r1 and r2 decreases at each time. When Done
configuration is reached, r1 and r2 become neighbor. We can then deduce that all the nodes of the
ring have been explored and the lemma holds.
✷
6 Conclusion
In this paper, we studied the exploration of uniform rings by a team of oblivious robots. The assump-
tions of unlimited visibility made in previous works has enabled them to focus only on overcoming
the computational weaknesses of robots introduced by the simultaneous presence of obliviousness
and asynchrony in the design of exploration algorithms. In this paper, we added one more weakness:
Myopia ie, robots have only a limited visibility. We provided evidences that the exploration problem
can still be solved under some conditions by oblivious robots despite myopia. We studied the problem
for both synchronous and asynchronous settings, and considered three types of visibility capabilities:
φ = 1, φ = 2, and φ = 3.
The complete characterization for which the exploration of the ring by myopic robots is solvable
remains open in general. We conjuncture that the solutions proposed in this paper are optimal with
respect to number of robots. We also believe that the condition n > kφ+1 is a necessary condition to
solve the problem. Also, the problem of exploring other topologies and arbitrary graphs by myopic
robots is a natural extension of this work.
29
References
[1] H Ando, Y Oasa, I Suzuki, and M Yamashita. A distributed memoryless point convergence algo-
rithm for mobile robots with limited visibility. IEEE Transaction on Robotics and Automation,
15(5):818 -- 828, 1999.
[2] Shantanu Das, Paola Flocchini, Nicola Santoro, and Masafumi Yamashita. On the computa-
tional power of oblivious robots: forming a series of geometric patterns. In 29th Annual ACM
Symposium on Principles of Distributed Computing (PODC), pages 267 -- 276, 2010.
[3] St´ephane Devismes, Franck Petit, and S´ebastien Tixeuil. Optimal probabilistic ring exploration
by semi-synchronous oblivious robots. In 16th International Colloquium on Structural Informa-
tion and Communication Complexity (SIROCCO), pages 195 -- 208, 2009.
[4] Yoann Dieudonn´e, Ouiddad Labbani-Igbida, and Franck Petit. Circle formation of weak mobile
robots. ACM Transactions on Adaptive and Autonomous Systems (TAAS), 3(4), 2008.
[5] P. Flocchini, D. Ilcinkas, A. Pelc, and N. Santoro. Computing without communicating: Ring
exploration by asynchronous oblivious robots. In 14th International Colloquium on Structural
Information and Communication Complexity (SIROCCO), pages 105 -- 118, 2007.
[6] Paola Flocchini, David Ilcinkas, Andrzej Pelc, and Nicola Santoro. Remembering without mem-
ory: Tree exploration by asynchronous oblivious robots. Theor. Comput. Sci., 411(14-15):1583 --
1598, 2010.
[7] Paola Flocchini, Giuseppe Prencipe, Nicola Santoro, and Peter Widmayer. Gathering of asyn-
chronous robots with limited visibility. Theor. Comput. Sci., 337(1-3):147 -- 168, 2005.
[8] Paola Flocchini, Giuseppe Prencipe, Nicola Santoro, and Peter Widmayer. Arbitrary pattern
formation by asynchronous, anonymous, oblivious robots. Theor. Comput. Sci., 407(1-3):412 --
447, 2008.
[9] Samuel Guilbault and Andrzej Pelc. Gathering asynchronous oblivious agents with local vision
in regular bipartite graphs. In 18th International Colloquium on Structural Information and
Communication Complexity (SIROCCO), pages 162 -- 173, 2011.
[10] Tomoko Izumi, Taisuke Izumi, Sayaka Kamei, and Fukuhito Ooshita. Mobile robots gathering
algorithm with local weak multiplicity in rings. In 17th International Colloquium on Structural
Information and Communication Complexity (SIROCCO), pages 101 -- 113, 2010.
[11] Sayaka Kamei, Anissa Lamani, Fukuhito Ooshita, and S´ebastien Tixeuil. Asynchronous mobile
robot gathering from symmetric configurations without global multiplicity detection. In 18th In-
ternational Colloquium on Structural Information and Communication Complexity (SIROCCO),
pages 150 -- 161, 2011.
[12] Ralf Klasing, Adrian Kosowski, and Alfredo Navarra. Taking advantage of symmetries: Gather-
ing of asynchronous oblivious robots on a ring. In 12th International Conference on Principles
of Distributed Systems (OPODIS), pages 446 -- 462, 2008.
[13] Ralf Klasing, Euripides Markou, and Andrzej Pelc. Gathering asynchronous oblivious mobile
robots in a ring. Theoretical Computer Science, 390(1):27 -- 39, 2008.
30
[14] Anissa Lamani, Maria Gradinariu Potop-Butucaru, and S´ebastien Tixeuil. Optimal determin-
istic ring exploration with oblivious asynchronous robots. In 17th International Colloquium on
Structural Information and Communication Complexity (SIROCCO), pages 183 -- 196, 2010.
[15] G. Prencipe. Corda: Distributed coordination of a set of autonomous mobile robots. In Proc.
4th European Research Seminar on Advances in Distributed Systems (ERSADS'01), pages 185 --
190, Bertinoro, Italy, May 2001.
[16] G Prencipe. Distributed Coordination of a Set of Autonomous Mobile Robots. PhD thesis,
Dipartimento di Informatica, University of Pisa, 2002.
[17] I. Suzuki and M. Yamashita. Distributed anonymous mobile robots: Formation of geometric
patterns. SIAM Journal on Computing, 28(4):1347 -- 1363, 1999.
31
|
1912.08929 | 1 | 1912 | 2019-11-14T14:15:02 | PFaRA: a Platoon Forming and Routing Algorithm for Same-Day Deliveries | [
"cs.MA",
"cs.SI"
] | Platoons, vehicles that travel very close together acting as one, promise to improve road usage on freeways and city roads alike. We study platoon formation in the context of same-day delivery in urban environments. Multiple self-interested logistic service providers (LSP) carry out same-day deliveries by deploying autonomous electric vehicles that are capable of forming and traveling in platoons. The novel aspect that we consider in our research is heterogeneity of platoons in the sense that vehicles are equipped with different capabilities and constraints, and belong to different providers. Our aim is to examine how these platoons can form and their potential properties and benefits. We present a platoon forming and routing algorithm, called PFaRA, that finds longest common routes for multiple vehicles, while also respecting vehicle preferences and constraints. PFaRA consists of two parts, a speed clustering step and a linear optimisation step. To test the approach, a simulation was used, working with realistic urban network data and background traffic models. Our results showed that the performance of our approach is comparable to a simple route-matching one, but it leads to better utility values for vehicles and by extension the LSPs. We show that the grouping provided is viable and provides benefits to all vehicles participating in the platoon. | cs.MA | cs |
PFARA: A PLATOON FORMING AND ROUTING ALGORITHM
FOR SAME-DAY DELIVERIES
A PREPRINT
Sınziana-Maria Sebe
orcidID: 0000-0002-9435-9879
Institute of Informatics
Clausthal University of Technology
Clausthal-Zellerfeld, Germany
Jorg P. Muller
orcidID: 0000-0001-7533-3852
Institute of Informatics
Clausthal University of Technology
Clausthal-Zellerfeld, Germany
[email protected]
[email protected]
ABSTRACT
Platoons, vehicles that travel very close together acting as one, promise to improve road usage on
freeways and city roads alike. We study platoon formation in the context of same-day delivery in
urban environments. Multiple self-interested logistic service providers (LSP) carry out same-day
deliveries by deploying autonomous electric vehicles that are capable of forming and traveling in
platoons. The novel aspect that we consider in our research is heterogeneity of platoons in the
sense that vehicles are equipped with different capabilities and constraints, and belong to different
providers. Our aim is to examine how these platoons can form and their potential properties and
benefits. We present a platoon forming and routing algorithm, called PFARA, that finds longest
common routes for multiple vehicles, while also respecting vehicle preferences and constraints.
PFARA consists of two parts, a speed clustering step and a linear optimisation step. To test the
approach, a simulation was used, working with realistic urban network data and background traffic
models. Our results showed that the performance of our approach is comparable to a simple route-
matching one, but it leads to better utility values for vehicles and by extension the LSPs. We show
that the grouping provided is viable and provides benefits to all vehicles participating in the platoon.
Keywords Platoon · Heterogeneous Groups · Route Matching · Group Building · Optimisation · Simulation
1
Introduction
Platooning is a topical subject researched by many scientist especially in the context of automated driving. Relying
on a leader to make the driving decisions and then mirroring them is one part of how this automation can be achieved.
Research on platooning has been conducted since the 1950's [1], [2] and has been gaining momentum since. Most
platooning research tends to focus on highway scenarios [3] with large freight transports in mind e.g. SARTRE [4].
However, with recent developments in technology, urban platooning appears promising with respect to the better
utilisation of the scare resource of urban space. However, current research on urban platooning focuses on stability,
manoeuvring, and control (see e.g., [5][6] [7]), and, to a lesser extent, on the traffic management perspective ([8], [9]).
Same-day deliveries in the scope of logistic traffic is a novel use case for platooning. Due to the growth of online
shopping, and the desire of customers for immediate shipping, logistic companies have to restructure the way they
carry out their deliveries. On the one hand, the distribution network has to be adapted. To support such orders,
Crainic et al. developed a two-tier network architecture([10], [11]), with hubs, satellites and customers being the main
transitional points. Goods are first carried from a large hub, located mostly outside the urban area, to multiple smaller
locations called satellites, which are scattered inside the city, each catering to about 20 to 25 customers [12]. From the
satellite, the goods are transported to the customer, the so-called last mile delivery, which accounts for the majority of
the costs of said delivery [13].
On the other hand, the way orders are handled has to change. Usually, logistic companies plan and schedule deliveries
in a way that minimises cost and maximises order completion. This process is costly in time and computational
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
resources, making it an unviable approach for same-day orders. Immediate deployment of received orders, combined
with having small autonomous electric vehicles carry out the costly last-mile stretch would be a possible solution.
These vehicles are flexible and do not contribute as much to traffic congestion and toxic emissions, but sending
multiple such vehicles in an already busy network does not alleviate the situation entirely.
Combining the logistic same-day deliveries into platoons could benefit urban traffic by streamlining deliveries, re-
moving large trucks from the inner city and potentially decongesting some intersections. However, this subject has
not been thoroughly researched thus far, with only a few publications existing (e.g [14],[9]) and to our knowledge,
none that looks into how vehicles are grouped into platoons, especially with vehicles belonging to different service
providers. We refer to Section 2 for a more in depth analysis of the state of the art.
For our research we assume the context of logistic traffic, with multiple service providers, each with their own fleet
of vehicles that carry out deliveries to customers. We also assume that platooning is incentivised by the traffic man-
agement authority, providing platooning vehicles with subsidies. Platoons form in a spontaneous way, at intersections
where more than two vehicles meet. A vehicle only forms/joins a platoon if its perceived utility for joining is better
than if it were to travel alone. In order to gain the most benefit from platoons, we assume that logistic service providers
are willing to form mixed platoons with each other. This leads to a high degree of heterogeneity, not only in the ve-
hicles and destinations sense, but given the self-interested nature of all providers; in valuation functions, restrictions
and preferences as well. This research presents a grouping and routing algorithm that fosters this type of cooperation
while respecting the limitations imposed.
Our previous work [15] presented a first and simple optimisation approach to solve the problem of cross-provider
platoon formation. In this work the optimisation algorithm only included a distance limitation, and since the model
did not have an explicit representation of time, it could not express nor solve dynamic problems, e.g. including speed
preferences or delivery windows. In [15], we showed that the algorithm computes accurate and promising results, but
has very limited expressiveness. This paper presents a substantially enhanced version of the grouping algorithm that
accounts for important aspects of travel, including distance, time, speed, and cost. The model has been finalised, and
simulated movement of vehicles has been introduced allowing us to model dynamic situations, where the grouping
algorithm is run every time one or more vehicles or platoons meet at intersections.
The structure of the paper is as follows; in Section 2 we give an overview of the state of the art regarding platoons as
well as other group building algorithms. Section 3 presents our proposed platoon forming and routing algorithm whose
goal is finding the longest common route, while respecting, length, time, speed and cost restrictions. A simulation
developed to test the approach is also presented here. Section 4 presents multiple simulation experiments, which show
good results regarding solution quality and algorithm runtime performance. We follow with a discussion in Section 5
where we also address the future developments of this research. Lastly we present our conclusion in Section 6.
2 Literature Review
The concept of platooning, where vehicles drive in a line close to each other and behaving as one unit, has been a topic
of great interest in research due to it's promising outlooks for road usage on heavy-duty transports on freeways [16],
[17], [18]. Since vehicles travel more closely together in platoons than in usual traffic, the resource of road space is
better utilised, which in turn has a positive effect on network flow. Another reason why platooning on freeways is seen
as positive is the reduction in fuel consumption [19]. Because of the reduced gap between vehicles, the wind drag does
not have as high of an effect, which means that less fuel is used to maintain a constant speed. However, more recent
studies and experiments have shown that although platooning did increase safety, the effect on the fuel-consumption
was less than what was expected, with just a three to four percent reduction [20]
Other research on platooning deals more with control, manoeuvres (merging, splitting, travelling), stability and safety.
Ploeg et al. [5] calculated the needed time headway to achieve string stability and safety in platoons. This distance
can be used in all scenarios, both urban and highway to calculate the spacing needed between platooning vehicles
depending on speed. Another piece of research looks into cooperative manoeuvres, as an extent to platooning that
again Ploeg et al. [21] studied for highway scenarios. They present a layered control architecture for such manoeuvres,
which was also the basis for the Grand Cooperative Driving Challenge.
Transitioning to an urban environment, the focus still seems to be on control, stability and manoeuvres. To increase
stability on curved roads and to account for the different speeds in an urban scenario, lateral and longitudinal control
have to function independently according to Ali et al. [7]. A third control mechanism responsible for the platoon must
also be put into place thus giving us a controller for each movement type. In [22], Khalifa et al. show that stability is
guaranteed even with limited of communication between vehicles, through a consensus-based control mechanism for
2
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
longitudinal movement. Stability of platoons is addressed in [6] by making use of multiple state machines to ensure
flexibility. There is one state machine for each of the logic elements of platooning: messaging, forming and distance.
Other benefits of platooning arise from a traffic management perspective in slower urban traffic. Lioris et al [8]
have found that bottleneck-prone places, such as intersections, could benefit from platooning. By using adaptive
traffic lights, and allowing platoons to pass uninterrupted, the throughput (measured in number of vehicles) of the
intersection can be doubled.
Most work focusing on platoons does not touch on the aspect of how vehicles decide to platoon and who to platoon
with. The study presented in [6], details just the actions needed to join or form a platoon. The aspects of how the
vehicles would form a logical group and what characteristics are important in making the platooning decision lack in
research.
An interesting use-case for platoons is logistic transport, from large freight shipments, which we have already ad-
dressed, to small singular same-day deliveries. Urban platooning in the case of logistic same-day deliveries was
addressed in [14] with a focus on network design. Assuming that fully autonomous platoons cannot travel on all edges
of a network, human-lead platoons have to drive through those inaccessible areas. The authors goal is to minimise
the costs across all vehicles in the delivery fleet of a single logistic service provider. Given the circumstance, the
algorithm for vehicle grouping is not addressed, perhaps because of the large overlap in the different vehicle capabil-
ities and objectives, as well as due to the strategic level on which the problem operates. In a similar vein, Haas and
Friedrich [9] look at platoons in a city-logistic scenario from the traffic managers point of view. They focus on the
microscopic/operational level focusing on the number of vehicles in platoons and the number of platoons in a network
during the course of a single day. They consider autonomous delivery vans, that rely on platoons to travel through
the network. This means that they are stationary when alone and can only drive autonomously during splitting and
merging operations. They do note that the vans only join a platoon if they can be brought further to their destinations,
if not they wait for another. Otherwise the authors tend to focus on developing an interaction model for platoons in a
roundabout scenario, where determining who has the right-of-way is not as obvious as it is with normal traffic.
Group building algorithms have been studied for a long time and there are many variants each depending on the
field of application. General approaches are presented in [23]. Other publications address specific algorithms such
as [24] for Walds SPRT Algorithm, [25] which identifies important criterion and then sorts accordingly, maximum
likelihood algorithms, [26] for Wertheimers laws of grouping, [27] for hierarchical grouping and algorithms that find
the maximum consistency between the data and the group appointment.
Switching to a more specific field, namely that of traffic, we are faced again with more aspects to consider. A pre-
requisite for any sort of group formation and disbandment is the ability to communicate; so Taleb et al. [28] proposed
that a group of vehicles could only exist if the communication line is maintained. This resulted in vehicles being
grouped by their velocity headings; all going in the same direction, while matching the speed within the group to
maintain inter-vehicular distances, and therefore the communication link. This sole criterion is not enough when
formulating platoons, but does matter when considering the maintenance of one.
Ways to pair drivers with passengers while also considering schedules and routes are presented in [29]. The problem
they formulate is a many-to-many advanced reservation travel problem. Important characteristics upon which potential
pairs can be made are found and studied, namely: the location of origins and destinations, the passenger type and trip
purpose, closeness of desired departure times, number and capacity of available vehicles and the direction of the trips.
With normal routing algorithms the focus is on cost minimisation for the company (in the form of length of route or
lack of empty trips) and not enough focus is given to the passenger and their preferences. To combat this, the authors
use a fuzzy relation to find similar trips. This work can be used as a base for our own criteria definition, although our
focus is not on the match between passenger and vehicle, but rather the compatibility between vehicles.
Sanderson et al. propose a consensus based approach to clustering vehicles, in their work [30]. They use their own
IPCon algorithm [31] to provide collective arrangements taking preferences and constraints into account through role-
associated power of the participants (vehicles in this case). Similarly, Dennisen and Muller [32] propose preference-
based group formation through iterative committee elections. They assume the context of a ride-sharing service that
drives to different sights within a city. Passengers have to choose which sights to visit and be sorted into the au-
tonomous vehicle based on their preference. The algorithm works by removing the unhappiest passenger using the
Minisum or Minimax approval committee rules. The votes are then recounted and the process started again until a
suitable solution is found. When considering our case, there is overlap in the fact that finding the best common route
is the goal, but since we are dealing with autonomous machines and not humans, an informed vote does not seem
obvious.
More related to our research field is the work of Khan and Boloni [33] which spontaneously groups vehicles driving
on freeways using a utility function based on speed. The focus is on the manoeuvres that come with convoy driving:
3
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
joining, staying and splitting; but not the actions needed, but rather the utility that comes with each one. While this
might be sufficient in a highway scenario, where all vehicles are driving in the same direction, it is not the case for an
urban environment where vehicle routes differ greatly.
To conclude, there is only limited research activity at the intersection of the two research fields of platooning and
grouping algorithms; in particular, it does not adequately address highly heterogeneous scenarios like urban logistic
traffic. PFARA, described in the following section aims to address this research gap.
3 The Platoon Forming and Routing Algorithm (PFARA)
Based on the literature, the decisive constraints and aspects of the heterogeneous group building problem need to be
determined. Taking the different elements of traffic, urban environments and logistics into account, the heterogene-
ity of agents and their preferences makes platoon formation and routing highly complex. First the logistic service
providers will want to keep their cost minimal while reaching their target number of orders as well as keep coopera-
tion only with some but not all other companies. The vehicles are bound by their battery life, thus restricting the route,
as well as speed, be it legal limits or individual speed capabilities. And last, but most important, we have the client
which selects a specific delivery window that must respected.
The problem of heterogeneous platoon building can be solved through a linear optimisation problem using the afore-
mentioned preferences. They are used as linear constraints in the optimisation problem that has the goal of minimising
total costs. This is turn has a positive effect of both travel time and length of route. We refer to subsection 3.5 for a a
more detailed description of the algorithm, but we do note that it is deterministic offering multiple best solutions.
Attempting to solve this problem with the aforementioned fuzzy theorem would not yield good results given the
different weights given to the aspects for each vehicle or logistic service provider. For example some might prioritise
battery life over speed, or some might want to wait for platoons rather than drive alone, or completely disregarding
platoons if not readily available to save up on time. So while the restriction criteria are the same for all vehicles, their
weight or importance will differ in the problem.
If our algorithm fails to find a singular best solution, which could happen in networks that follow a grid-type pattern, the
aforementioned voting approaches could act as fail-safe methods of selecting the grouping and route. This guarantees
that the decision is not made by a singular element, but rather all potential vehicles in the platoon.
3.1
Input data
The urban network is transformed into a graph to allow for routing. Intersections are transformed into vertexes and
the streets into edges. Traffic demand is the sum of routes over all origin and destination points in the network, or
how many cars use each edge in a given time-span. This is meant to represent background traffic which will affect our
platoon routing.
∆N
∆t
(1)
where Q(x,t) is the traffic demand, ∆N is the number of vehicles and ∆t the time-span. [34]. From here we can find
the traffic density, by normalising through division over the edges length.
Q(x, t) =
ρ(x, t) =
∆N
Σαdα
(2)
where Σαdα is the length.
The resulting traffic density is given as each edge's weight, and act as the cost to minimise in the optimisation problem.
We consider traffic density the "cost" because of its direct translation into time-savings; the freer an edge, the faster it
is to transverse it, the quicker the vehicle delivers its package and can return to be dispatched again. Given the electric
nature of the vehicles considered, the more time they spend away from the satellite, the longer they will need to charge,
thus reducing the amounts of orders they can fulfil in a day.
Having the environment defined, vehicles can then be added and their routes calculated by PFARA. Each vehicle
has an origin, a destination and a set of preferences; minimum acceptable speed for platooning, maximum speed,
maximum length of route, maximum travel time and maximum cost.
3.2 Assumptions
The vehicles are assumed to function autonomously, all the while attempting to form a platoon. In order for a platoon
to exist, a minimum of two vehicles is necessary. They must be at the same vertex at the same time to do so. For
4
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
Figure 1: Vehicle interaction and group formation.
now vehicles do not wait at intersections for other vehicles or platoons, even if their preferences allow it. Platoons
form spontaneously and organically with vehicles that would benefit from platooning. A vehicle would not participate
in a platoon if the cost to do so is larger than travelling alone, or its distance, time, speed and cost restrictions are
not respected. After a platoon forms, all vehicles in it drive with a uniform speed. As pointed out in the literature
review, vehicles must have communication capabilities, based on [28] and [22]. Each vehicle sends its destination and
restrictions to a local agent which performs PFARA. The results (route, expected cost, expected length and group) are
then communicated back to each vehicle.
For a visual representation of how platoons would form, please see Figure 1. Two platoons approach the same in-
tersection, but the vehicles comprising them follow different headings, denoted by the different opaque colours. The
vehicles communicate their destination, preferences and limitations to the local agent denoted by the broadcast tower,
which runs PFARA and forms two new platoons, denoted by the slightly more transparent vehicle formations.
To encourage the penetration rate of platoons, they are treated as one vehicle, since logically they perform as one. All
members of the platoon share a common route and contribute equally to its cost. This leads to a reduction of individual
costs for all vehicles and encourages the formation of larger platoons. The cost of a route is considered as the sum of
weights for all edges used in it.
3.3 Notations
For ease of understanding, all symbols to be used in the formal description of the algorithm and their meaning is
presented in Table 1.
3.4 PFARA: The Speed-Clustering Step
Given that all vehicles in a platoon must travel with the same speed to maintain stability, an initial speed clustering has
to be put in place. A minimum acceptable speed for platoons as well as the maximum speed is specified in the vehicles
preferences. To form a grouping based on speed, the maximum minimum speed within the group is first selected. All
vehicles whose maximum speed is greater than it are grouped and their destination and preferences directed to the
second step: the optimiser. If not all vehicles in the original group were selected, the next greatest minimum speed is
taken, and the same procedure applied, to be repeated until all vehicles are grouped, or it is determined that they must
travel alone. This can be written as:
5
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
Table 1: Symbols and meaning [15].
Symbol
G=(V, E)
n ∈ V
e ∈ E
d(e)
l(e)
s
v
destv
K∗
v
Ωv
Λv
p(e)v
N P
Definitions
graph
vertices
edges, e ≡ (i, j)
traffic density of edge e
length of edge e
speed of the platoon
vehicle
destination vertex of vehicle v , fixed
maximum costs for vehicle v
maximum delivery time for vehicle v
maximum length of route for vehicle v
price vehicle v pays for edge e
the number of vehicles in the platoon
( a l l v e h i c l e s grouped o r e x c l u d e d )
w h i l e n o t
{
f i n d maximum minimum sp ee d ;
f o r each v e h i c l e
{
i f
( maxspeed >= maxminspeed )
group v e h i c l e ;
}
}
While determination of the platoon's speed could be written as a linear constraint in the following step, we believe
that this approach would be restrictive. The common speed calculated would be lower than an ideal speed and could
possibly deter vehicles from joining. This means that our algorithm has to be repeated with a smaller group of vehicles
which not only takes time, but also provides less cost savings.
3.5 PFARA: The Optimiser Step
The objective for this step is to find the route with the most overlap between vehicles that also respects their restrictions.
This can be achieved by adapting the shortest path algorithm (presented by equations 3,4 and 5) to fit multiple vehicles
instead of one.
min
(i,j)∈E
x(i, j) ∗ d(i, j)
(cid:88)
1 ⇐⇒ i = Origin
−1 ⇐⇒ i = Destination
0 otherwise.
x(i, j) ∈ 0, 1∀edge (i, j)
x(j, i) =
(cid:88)
x(i, j) −(cid:88)
j
j
(3)
(4)
(5)
x is the variable associated with an edge, defined between vertices i and j. It has the value 0 if its edge is not in
the shortest path, or 1 if it is (4). The flow constraint enables the vehicles to pass through intermediary nodes and is
defined with the three cases (5). Naturally the objective is to minimise the costs which we measure in traffic density
(3).
When transposing this to platooning, we introduce a new variable y to signify the platoon. This variable acts in a
similar fashion to how x did in the classic problem: y = 0 if it is not part of the route of any vehicle and y = 1 if it
is (7). y is used in the objective function to be minimised (6). To ensure that the route covers all vehicles, we add a
restriction that states that y can have the value of one only if there is at least an x with the value of one (8). The flow
constraint remains unchanged (9). To these restrictions we add the ones for length of route (10), delivery time (11),
and maximum cost (14). Incentivising platoons is done by giving a reduction in costs. Since all vehicles contribute
to the cost (13), the more vehicles in a platoon, the less they have to pay for each edge. Therefore the platooning
optimisation problem would be formulated as such follows:
6
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
Definition 1. Given a group of vehicles, being at the same vertex O in the graph at the same time; let the routes for
all the vehicles be given by y(i, j), and the individual vehicle routes by x(i, j)v where
min
y(i, j), x(i, j) ∈ {0, 1},∀edge (i, j)
x(i, j)v ≤ y(i, j)∀edge (i, j),∀vehicle v
x(i, j)v −(cid:88)
(cid:88)
j
x(j, i)v =
(i,j)∈E
y(i, j) ∗ d(i, j)
(cid:88)
1 ⇐⇒ i = O
(cid:88)
(cid:88)
−1 ⇐⇒ i = destv
0 otherwise.
l(i, j) ≤ Λv∀ v
l(i, j)
≤ Ωv∀ v
s
x(i,j)v=1
x(i,j)v=1
(cid:88)
p(i, j)v ≥ 0 ∀ v and edge (i, j)
∀ v where N P(i,j) =
d(i, j)
N P(i,j)
v ,∀ v.
p(i, j)v ≤ K∗
(cid:88)
v
j
p(i, j)v =
∀v, ∀(i, j) ∈ E
x(i, j)v
(6)
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
(i,j):x(i,j)v=1
A similar but much less expressive model was presented in [15]. It only addressed the optimisation problem limited
strictly by a length of route constraint. Also the analysis of the grouping algorithm was done statically since the model
did not feature time, and by extension, speed. Without those elements a complete solution accounting for the four
most important aspects of travel (length, time, speed and cost) was not possible.
This work features all the aforementioned aspects and and takes into account the time of meeting, speed matching and
the cost restrictions. This ensures the solution provided is complete and can be applied to all vehicles regardless.
3.6 Simulation
The simulation was formulated to study the effectiveness of PFARA and its effects on the valuation of cooperation
between vehicles. Its performance is also tested to ensure that it is applicable and appropriate for real-life scenarios.
3.6.1 Framework
To analyse PFARA, a simulation was designed using Java. Each separate component has a specific framework. Jung
[35] is used to generate the environment due to its powerful library that models data into a graph, its visualisation
options, routing and analysis capabilities. To visualise the routes and traffic densities, we employed the use of a heat-
map using the colour schemes provided by [36]. The actual drawing of routes was done with a JXMapViewer2 painter,
modelled after a pre-existing examples [37]. For the optimisation part, Gurobi [38] was used, a powerful commercial
solver. It was extremely easy to implement into the simulation as an external jar file and the problem definition process
was also simple. To write the problem we start by defining as many x variables as we have vehicles in the potential
platoon, and one y variable. Then the objective function is written based on the y and weights (traffic density) of the
edges with minimisation being the goal. Then for each of the criteria defined above (equations 8, 9, 10, 11 and 14), a
linear constraint is written. Since speed is already addressed before running the optimiser, we do not need to include it
here. After running and finding a solution, the results are saved in a separate data structure, available at any time and
the optimiser object made redundant to save up on computational resources.
3.6.2 Methodology
Simulation starts with creating the environment and the vehicles from an input file. The vehicles each have their set of
preferences as well as a starting satellite and a destination. Then the internal clock starts and after running instances
of the optimiser for each satellite to group the vehicles, they begin moving through the environment. They go along
their given route, attempting to form new platoons whenever they encounter new vehicles/platoons. If a vehicle or a
platoon intersects with other vehicles or platoons, they go through the grouping process (first by speed and then by
7
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
optimiser). The algorithm finds the best group, their routes (both common and separate until their destination) and
expected cost for each vehicle. After grouping, the vehicles travel together until the end of their common route, where
they split and continue their trip towards their destination, or attempt to form a new platoon with some of their former
co-platooners. To ensure that the selected route is constantly the "best" one, given that urban traffic tends to shift quite
rapidly, PFARA can be performed at some or all intermediary nodes. To account for the distance and time travelled
as well as the cost of the route, and update of the preferences is performed at each intermediary node. The length
of the edge travelled is subtracted from the maximum distance, and the same goes for the maximum time and cost.
This ensures that a new instance of the optimisation part of the algorithm sets the restrictions to a current version of
the preferences, and not to the ones at the start of travel. The simulation completes with the vehicles reaching their
respective destinations. Afterwards, an output file is written detailing each action of each vehicle, their cost, length of
route and time of arrival. Lastly a heat map of the network is generated allowing us to see which edges were used by
the vehicles and of how many of them were on each edge.
3.6.3 Baseline algorithm: Overlapping
To have a benchmark comparison for our approach we decided to conceptualise the most obvious version of a platoon-
ing grouping algorithm, which is simple best route matching. Considering each vehicles current position as the origin
point, the already-existing routing algorithm present in Jung is used to find the fastest route. By aggregating the routes
for all vehicles, we can find where they would travel together (considering constant speed). By counting the number
of times an edge was used, we determine the potential size of the platoon travelling on it as well as knowing which
vehicles make up the platoon. This approach however does not take into account the possible restrictions, re-routings
and above all, vehicle preferences. It is on the other side extremely easy to calculate and straightforward. To ensure we
are just comparing the optimisation part to this simple overlapping approach, all other aspect of the simulation were
used in the overlapping algorithm as well. Therefore we have the time counter, the vehicles' movement throughout
the environment, and most importantly, the preliminary speed clustering.
3.6.4 Input Data
To simulate scenarios we employed data from [39], namely the Berlin Tiergarten neighbourhood. The data set is finely
granular enough to use and comes with a very rough account of traffic demand. It was not given for each edge in
particular, but rather as an aggregation of trips based on zones in the neighbourhood. So to get accurate weights for
our problem, individual trips needed to be generated. The vertices were divided into zones and random ones pulled
to act as origins and destinations. Routes were calculated, aggregated for the whole instance, and then normalised by
dividing through each edges length (according to the definition given by Equation 2), thus giving us traffic density to
act as the edges weights for our algorithm. An example of this process would be having thirteen trips from zone 2
to zone 9, pulling a random node from zone 2 and one from zone 9, finding the best route between them, saving the
results and repeating the process another twelve times aggregating the edges taken for all. This process does not offer
constant results, but they are similar enough between runs to be considered consistent.
We also created a synthetic smaller network of a five-by-five grid network to follow the routes and examine the driving
behaviour more closely. Factors like traffic density, positioning of satellites and vehicle preferences were varied to
ensure the generality of PFARA.
3.6.5 Output Data
As mentioned before, after each run the simulation provides output in the form of a log file and a heat map visualisation.
The colours selected to display it (Inferno palette) were chosen due to the cognitive ease of understanding it by the
viewer [40] (yellow for hot to dark purple for cold). A light yellow edge means it was the most used, whereas dark
purple indicates it was a part of the route of a single vehicle. A colourless edge signifies it not being a part of any
vehicles' route.
The log file consists of the events of all vehicles. The time step and location of each event is specified. The event types
are:
1. Creation. Takes place at the satellite vertex.
2. Departure. Happens for satellite and intermediary vertices.
3. Arrival. Happens for intermediary vertices as well as the destination.
4. Completed. Accompanies the arrival event to denote the completion of the route.
5. Formed. When a vehicle joins a platoon and who they join with.
8
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
6. Split. When a vehicle disbands from the platoon. Can be due to completion or because the end of the common
route was reached.
Besides the events, upon completion the following is specified:
1. the cost accumulated
2. length of route taken
3. the allowance left for length, time and cost for all vehicles.
The overlapping algorithm provides the same result structure; a heat map and a log file.
4 Results
To check the validity of our algorithm, we ran experiments on real and synthetic environments. The real environment
was the Berlin Tiergarten neighbourhood that also has realistic traffic density. The synthetic one is represented by a
five by five Manhattan grid network. To validate all aspects of the algorithm, multiple experiments were run varying
things like traffic density, satellite placements, and vehicle preferences. The algorithm did well in all cases and all
groupings and routes were correct and respected the restrictions.
By looking at the heat maps, we get an idea of how the placement of satellites and traffic density affect the deliveries,
and where joining and splitting could occur across the platoons. The colours show how many vehicles use a specific
edge and a change in colour from one edge do the next may denote vehicle/s splitting (if it shifts colder) or joining (if
it shifts warmer). However, due to fact that the heat map is generated statically at the end of the simulation, it cannot
serve as a definite way of determining said points. The colour of the edge changes with how many vehicles use it
and not necessarily just with platoons. However we will present a case where the split and join points can be clearly
identified.
We take the example of the five by five Manhattan grid network, with the outside edges having higher traffic density,
which decreases towards the centre. We simulate five different logistic service providers, each with a fleet of five
vehicles, each having a random destination node. The locations of the satellites are the four corners of the network
and the central node. All the vehicles start at the same time (the starting time of the simulation) and travel at the same
speed. Split points can be identified by a transition of the edge colour from warm to cooler when looking at Figure 2.
That is the case at node 23 where a vehicle splits from the five vehicle platoon coming from node 24 and continues
alone towards node 18, with the other four continuing their journey in the platoon towards node 22. Joining points like
node 2 can be identified by the transition of the edge colour from cooler to warmer. Two platoons, one coming from
node 3 containing five vehicles, and the other coming from node 1 containing four vehicles, meet and join at node 2
in a larger seven vehicle platoon (node 2 was the destination for two vehicles) travelling toward node 7.
4.1 Numerical results
To analyse the numerical results as best as possible we compared our approach to independent travel and the overlap-
ping algorithm. Pinning the two grouping algorithms agains one another allows us to see how viable our design is. The
real environment of Berlin Tiergarten Neighbourhood was used to guarantee the results are as realistic as possible. To
resemble real traffic we assumed the satellites are located at intersections where most of the traffic takes place (number
of vehicles leaving the intersection is highest) and the vehicle destinations we selected the nodes where the number of
vehicles arriving would be highest, both based on traffic density information. We have two providers, each with one
satellite and 15 and 10 vehicles respectively, each with a different destination. With the origin and destination points
set, both algorithms were run and the resulting costs analysed.
From Figure 3 we can see what sort of cost savings we accrue from using the two grouping algorithms. In the case of
the first provider (pod 1 to pod 24), both the overlapping and the PFARA approach mostly give the same results. The
only two exceptions are pod 13 and 24, which have less costs with PFARA. This shows us that this approach found
an alternative route that allows for longer platooning for at least one of the vehicles, but which benefits both. This is
confirmed when analysing output log file.
In the case of the second provider (pod 30 to pod 44) we have more variety in the results. Again, both algorithms are
far superior than the alone-travel alternative, but when comparing them to each other we have an array of relations.
In some cases, like pod 43, pod 32, the cost is the same. Pod 40 and pod 30 have a significantly better result with
PFARA. The most interesting result is pod 42, whose costs for both grouping algorithms are lower than travelling
alone, but PFARA does considerably worse than the overlapping alternative. This means the route given by PFARA
9
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
Figure 2: Heat map depicting split and join points.
10
Split Point(cid:10)Join Point(cid:10) PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
Figure 3: Cost comparison of the two algorithms with alone-travel on realistic network.
11
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
for this vehicle is longer, but when considering the global cost of this provider, it might be worth it. The total costs for
this logistic service provider are:
• Alone: 198761
• Overlap: 121897.97
• Optimiser: 102558.95
We can see that even if some vehicles take detours, it might be worth for the service provider to go for the PFARA
approach anyway. The influences and weights of each of these factors (cost, length and time) need to be calculated and
specified by each of the service providers, and adjusting the vehicles preferences and restrictions accordingly. Table 2
shows a more detailed view of the results for the specific cases mentioned above.
Table 2: Result comparison of the two algorithms, detailed.
T imeOverlap LengthOverlap CostP F aRA
T imeP F aRA LengthP F aRA
Vehicle CostOverlap
pod 43
pod 32
pod 40
pod 30
pod 42
633.87
13154.73
18482.71
27100.71
26364.71
45
27
81
101
93
3.74
1.29
6.01
6.85
7.40
633.87
13164.73
6013.36
10816.02
36673
45
27
81
101
107
3.74
1.29
6.01
7.44
8.09
Another experiment is ran on the synthetic five-by-five grid network to present an easier to understand set of results,
which is not possible with the Tiergarten neighbourhood due to its size. In this case there are five logistic service
providers each with five vehicles.
We can tell from Figure 4 that the optimiser approach provides a better result than the overlapping approach for all
vehicles. Usually the vehicles do not stray from their best route, which is shown by observing both the length of route
and travel time given in the output file. The only exception is vehicle 43, which takes a rather considerable detour to
travel with a platoon. Even so, the cost it accrues is still lower than the alone travel alternative, which is 12. One could
argue that this cost saving is not worth the extra time and distance traveled, since the vehicle could have returned to
the satellite and been re-dispatched on another delivery. As mentioned before the cost-time-length valuation is up to
each logistic service provider to decide and implement it through the vehicles preferences.
When looking at the results for both scenarios, we can see that sometimes the simplistic overlapping approach does
offer better results than the optimiser, but globally it is not the best solution. The optimiser guarantees us the group
specific best solution, so the minimal cost for all vehicles considered, leading to a Pareto-type solution. Another fault
of the overlapping algorithm is its simplicity, none of the preferences and limitations imposed on the vehicles are
considered, leading to a possible incomplete solution. The optimiser takes limitations and preferences into account,
giving not the universal (impossible to calculate on a large-scale scenario), nor the user (vehicle) best solution, but
somewhere in the middle, with a group specific best.
Besides the improvement in cost, we also tested the influence that the preferences of the vehicle have in PFARA.
Restrictions in speed, length of route and time of travel affect the grouping of vehicles in the form of the preliminary
clustering (speed) as well as in the optimisation problem (time and length). For that we modified the scenario presented
previously, and made vehicle 5 slower, vehicle 41 to have a length limitation and vehicle 43 to have a time limitation.
The results are presented in Table 3 having the first three columns with the cost, time and length of the route for the
PFARA without restrictions and then the last three columns he cost, time and length of the route for the PFARA with
restrictions.
We can see that the vehicles 30 to 34 and 10 to 14 are not affected by the preference modifications. Vehicles 1 through
4 are affected by the lack of vehicle 5 in the platoon, it being excluded due to its lower speed. This is turn also
slightly affects the cost of vehicles 20, 22, 23 and 24. Given vehicle's 43 time and 41's length restrictions the previous
formation and route of that platoon completely changes. Instead of a five vehicle platoon, we have two platoons, of
two and three vehicles respectively.
4.2 Performance
Figure 5 shows both algorithms' performance, representing the runtime of the PFARA and the overlapping algorithm
respectively for the number of vehicles specified. These experiments were run on the realistic environment of the
Berlin neighbourhood to guarantee PFARA could be applicable in a real-life scenario. When looking at the runtime
of the algorithms, an evident increase happens with raising the number of vehicles. Even so, with 25 vehicles being
12
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
Figure 4: Cost comparison of the two algorithms on synthetic network.
13
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
Table 3: Comparison of PFARA on synthetic network, with and without different vehicular preferences.
T ime Length CostP ref
Vehicle
pod1
pod2
pod3
pod4
pod5
pod10
pod11
pod12
pod13
pod 14
pod20
pod21
pod22
pod23
pod24
pod30
pod31
pod32
pod33
pod34
pod40
pod41
pod42
pod43
pod44
Cost
2.27
0.6
1.35
1.77
1.945
3.85
2.35
1.35
0.6
6.85
2.295
4.2
2.295
6.128
6.128
1.3
2
5.3
4.3
5
2.683
4.35
0.6
11.183
5.183
55
11
22
33
44
44
33
22
11
55
55
33
55
77
77
22
22
33
33
33
44
33
11
88
66
500
100
200
300
400
400
300
200
100
500
500
300
500
700
700
200
200
300
300
300
400
300
100
800
600
2.783
0.75
1.75
2.25
9
3.85
2.35
1.35
0.6
6.85
2.4
4.2
2.9
6.23
6.23
1.3
2
5.3
4.3
5
4
7.5
1.5
8
6.5
T imeP ref
55
11
22
33
67
44
33
22
11
55
55
33
55
77
77
22
22
33
33
33
44
33
11
44
66
LengthP ref
500
100
200
300
400
400
300
200
100
500
500
300
500
700
700
200
200
300
300
300
400
300
100
400
600
considered for platooning at the same time, a result is still found in under two seconds. Considering that vehicles must
not only be physically near each other to platoon, but also close time-wise, we can safely assume that in a real traffic
scenario, no more than 20 vehicles would attempt platooning. In the case where a large platoon would meet another,
due to them logically being one vehicle, PFARA is also easily solvable. With the more likely case of having five or
ten vehicles, PFARA does really well, finding the best solution in 0.5 and 0.75 seconds respectively.
5 Discussion and Future Work
The algorithm presented in this paper provides a viable solution to the platoon forming and routing problem, for
vehicles that are heterogeneous in nature and have limited capabilities.
When comparing our approach to the state of the art, PFARA takes into account multiple aspects aspect of travel
(speed, time, length and cost). It does slightly underperform when comparing it to a more straightforward route-
matching approach, but respects individual preferences and restrictions of vehicles and hence allows us to consider
cross-provider platooning. To the best of our knowledge there are no other approaches to group vehicles in a platoon
from a logical standpoint; as mentioned before, focus tends to be on the manoeuvres necessary to form, rather than
what makes the vehicles form.
When comparing to the results in our previous work, having the same input data, we can see that the inclusion of
time, speed and movement has a significant influence on the costs. More limitations on the system means less flexible
grouping; but this also means that the groups are robust and a spontaneous change of route that could lead to dis-
bandment is less likely. The new additions to the model have not affected the runtime of the algorithm negatively; on
the contrary, PFaRA, with its preliminary speed clustering and enhanced optimisation problem performs faster than
its predecessor presented in [15]. This is due to the added restrictions which discards non-plausible routes faster,
significantly reducing computational time.
PFARA performs well, giving solutions for platooning for all vehicles considered; whether they start at the same satel-
lite or meet along the way. The routes provided are the group specific best solutions, thus accounting for heterogeneity
of the vehicles, providing each vehicle with cost savings. A user specific optimum would most likely exclude that
user from the group because it deviates from the Pareto solution provided. A global best solution would be hard to
14
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
Figure 5: Performance analysis.
calculate, due to the size of the network, and would most likely not be viable long term (for example for the entire
length of the route) due to the fact that urban traffic is considerably volatile. This is why we focus on finding a group
specific best solution and due to the structure of the algorithm, the system can be relaxed or restricted by removing or
adding in constraints.
Our approach can be applied to arbitrary platoons and is not restricted to just logistic traffic. It is general enough
for all types of vehicles and environments given that the vehicles preferences are known. One limitation of our
current approach is that it does not consider the physical size of the vehicles. In urban traffic space is limited, and
while heterogeneous platoons also include differently sized vehicles, one must guarantee that there is enough space
to accommodate it. This is where the time headway proposed in [5] can be used to calculate the physical size of the
platoon based on it's speed and the sizes of the vehicles included in it. We are planning to address this aspect in future
research.
A further important extension of our model will include adding a negotiation algorithm that allows for vehicles to
arrange further travel in a platoon through monetary exchanges at the end of their common route. This means that any
vehicle accepting will be taking a detour from their ideal (selfish) route. The trade-off between cost reduction and travel
limitations is up to each vehicle to calculate. Moreover due to the competitive nature of the LSPs, the compensations
offered need to be high enough to be profitable for competitors in order to convince them to participate, while also
being low enough for the offerer to maintain an upper hand in the market. Problems usually associated with this type
of scenario such as deliberately accepting less than favourable routes to financially "hurt" the competition which also
means waste of resources should be avoided.
6 Conclusion
The use of autonomous electric vehicles for last-mile deliveries in urban areas can bring benefits to customers, traffic
managers and logistic service providers alike. They no not create emissions, can navigate through traffic alone and
remove the need of a human driver. But the advantages they bring can be increased by having them travel in platoons
when possible.
To guarantee the formation of platoons, all delivery vehicles from different logistic service providers should be willing
to cooperate, thus creating heterogeneous groups. The vehicles have preferences (the minimum speed they are willing
15
03006009001200510152025Number of vehiclesExecution time in MilisecondsvariableBenchmarkOptimiser PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
to accept for the platoon), restrictions (the end of the delivery time window, or travel distance manageable with their
battery autonomy) and characteristics (maximum speed they can achieve) that have to be taken into account.
Our contribution is a grouping and routing algorithm for heterogeneous vehicles with same-day delivery in mind. The
vehicles go through two grouping stages; one based on speed and another based on route and restrictions. The second
stage is an optimisation problem that looks for the largest route overlap between the vehicles, while respecting their
limitations (in the form on linear constraints). The objective is to minimise the cost of the vehicles considered, which
is measured in traffic density. This approach offers feasible groups and routes, while maintaining the possibility of
further relaxation or limitation of the system.
The proposed PFARA performs well in terms of solution quality (cost) and runtime performance, offering a viable
and robust grouping and routing for all vehicles, independent of when and where they meet.
7 Acknowledgements
This work has been funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under
Grant 227198829 / GRK1931. The focus of the SocialCars Research Training Group is on significantly improving
the city's future road traffic, through cooperative approaches. We also acknowledge Nelly Nyeck for her help with
gathering results, Philipp Kraus for suggesting frameworks and Stephan Westphal for guidance towards a correct
optimisation formulation.
References
[1] Brian J Lewis. Platoon movement of traffic from an isolated signalized intersection. Highway Research Board
Bulletin, (178), 1958.
[2] Andre Schuhl. The probability theory applied to distribution of vehicles on two-lane highways. 1955.
[3] DVAHG Swaroop and J Karl Hedrick. Constant spacing strategies for platooning in automated highway systems.
Journal of dynamic systems, measurement, and control, 121(3):462 -- 470, 1999.
[4] Tom Robinson, Eric Chan, and Erik Coelingh. Operating platoons on public motorways: An introduction to the
sartre platooning programme. In 17th world congress on intelligent transport systems, volume 1, page 12, 2010.
[5] Jeroen Ploeg, Nathan Van De Wouw, and Henk Nijmeijer. Lp string stability of cascaded systems: Application
to vehicle platooning. IEEE Transactions on Control Systems Technology, 22(2):786 -- 793, 2013.
[6] Julian Schindler, Reza Dariani, Michele Rondinone, and Thomas Walter. Dynamic and flexible platooning in
urban areas. In AAET Automatisiertes und vernetztes Fahren conference 2018, 2018.
[7] Alan Ali, Gaetan Garcia, and Philippe Martinet. Urban platooning using a flatbed tow truck model. In Intelligent
Vehicles Symposium (IV), 2015 IEEE, pages 374 -- 379. IEEE, 2015.
[8] Jennie Lioris, Ramtin Pedarsani, Fatma Yildiz Tascikaraoglu, and Pravin Varaiya. Platoons of connected vehicles
can double throughput in urban roads. Transportation Research Part C: Emerging Technologies, 77:292 -- 305,
2017.
[9] Inbal Haas and Bernhard Friedrich. An autonomous connected platoon-based system for city-logistics: develop-
ment and examination of travel time aspects. Transportmetrica A: Transport Science, pages 1 -- 18, 2018.
[10] Teodor Gabriel Crainic. City logistics. In State-of-the-Art Decision-Making Tools in the Information-Intensive
Age, pages 181 -- 212. INFORMS, 2008.
[11] Teodor Gabriel Crainic and Antonino Sgalambro. Service network design models for two-tier city logistics.
Optimization Letters, 8(4):1375 -- 1387, 2014.
[12] Teodor Gabriel Crainic, Guido Perboli, Simona Mancini, and Roberto Tadei. Two-echelon vehicle routing prob-
lem: a satellite location analysis. Procedia-Social and Behavioral Sciences, 2(3):5944 -- 5955, 2010.
[13] Roel Gevaers, Eddy Van de Voorde, and Thierry Vanelslander. Characteristics and typology of last-mile logistics
from an innovation perspective in an urban context. City Distribution and Urban Freight Transport: Multiple
Perspectives, Edward Elgar Publishing, pages 56 -- 71, 2011.
[14] Yannick Oskar Scherr, Bruno Albert Neumann-Saavedra, Mike Hewitt, and Dirk Christian Mattfeld. Service
network design for same day delivery with mixed autonomous fleets. Transportation research procedia, 30:23 --
32, 2018.
16
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
[15] Sınziana-Maria Sebe., Philipp Kraus., Jorg P. Muller., and Stephan Westphal. Cross-provider platoons for same-
day delivery. In Proceedings of the 5th International Conference on Vehicle Technology and Intelligent Transport
Systems - Volume 1: VEHITS,, pages 106 -- 116. INSTICC, SciTePress, 2019.
[16] Mani Amoozadeh, Hui Deng, Chen-Nee Chuah, H Michael Zhang, and Dipak Ghosal. Platoon management
with cooperative adaptive cruise control enabled by vanet. Vehicular communications, 2(2):110 -- 123, 2015.
[17] Hoai Hoang Bengtsson, Lei Chen, Alexey Voronov, and Cristofer Englund. Interaction protocol for highway
In Intelligent Transportation Systems (ITSC), 2015 IEEE 18th International Conference on,
platoon merge.
pages 1971 -- 1976. IEEE, 2015.
[18] Subir Biswas, Raymond Tatchikou, and Francois Dion. Vehicle-to-vehicle wireless communication protocols for
enhancing highway traffic safety. IEEE communications magazine, 44(1):74 -- 82, 2006.
[19] Jeffrey Larson, Kuo-Yun Liang, and Karl H Johansson. A distributed framework for coordinated heavy-duty
vehicle platooning. IEEE Transactions on Intelligent Transportation Systems, 16(1):419 -- 429, 2014.
[20] Stefan Krempl. Lkw platooning; sicher aber nicht so sparsam wie erwartet, 2019.
[21] Jeroen Ploeg, Elham Semsar-Kazerooni, Alejandro I Morales Medina, Jan FCM de Jongh, Jacco van de Sluis,
Alexey Voronov, Cristofer Englund, Reinder J Bril, Hrishikesh Salunkhe, ´Alvaro Arr´ue, et al. Cooperative
IEEE Transactions on Intelligent
automated maneuvering at the 2016 grand cooperative driving challenge.
Transportation Systems, 19(4):1213 -- 1226, 2017.
[22] Ahmed Khalifa, Olivier Kermorgant, Salvador Dominguez, and Philippe Martinet. Vehicles platooning in urban
environment: Consensus-based longitudinal control with limited communications capabilities. In International
Conference on Control, Automation, Robotics and Vision, 2018.
[23] Arnon Amir and Michael Lindenbaum. A generic grouping algorithm and its quantitative analysis. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence, 20(2):168 -- 185, 1998.
[24] Abraham Wald and Jacob Wolfowitz. Optimum character of the sequential probability ratio test. The Annals of
Mathematical Statistics, pages 326 -- 339, 1948.
[25] Parag Havaldar, G´erard Medioni, and Fridtjof Stein. Extraction of groups for recognition. In European Confer-
ence on Computer Vision, pages 251 -- 261. Springer, 1994.
[26] Ian E Gordon. Theories of visual perception. Psychology Press, 2004.
[27] Richard Mojena. Hierarchical grouping methods and stopping rules: An evaluation. The Computer Journal,
20(4):359 -- 363, 1977.
[28] Tarik Taleb, Ehssan Sakhaee, Abbas Jamalipour, Kazuo Hashimoto, Nei Kato, and Yoshiaki Nemoto. A sta-
IEEE Transactions on Vehicular technology,
ble routing protocol to support its services in vanet networks.
56(6):3337 -- 3347, 2007.
[29] Seiichi Kagaya, Shinya Kikuchi, and Robert A Donnelly. Use of a fuzzy theory technique for grouping of trips in
the vehicle routing and scheduling problem. European Journal of Operational Research, 76(1):143 -- 154, 1994.
[30] David Sanderson, Didac Busquets, and Jeremy Pitt. A micro-meso-macro approach to intelligent transportation
systems. In 2012 IEEE Sixth International Conference on Self-Adaptive and Self-Organizing Systems Workshops,
pages 77 -- 82. IEEE, 2012.
[31] David Sanderson and Jeremy Pitt. Institutionalised consensus in vehicular networks: Executable specification
and empirical validation. In 2012 IEEE Sixth International Conference on Self-Adaptive and Self-Organizing
Systems Workshops, pages 71 -- 76. IEEE, 2012.
[32] Sophie L Dennisen and Jorg P Muller. Iterative committee elections for collective decision-making in a ride-
sharing application. In ATT@ IJCAI, 2016.
[33] Majid Ali Khan and Ladislau Boloni. Convoy driving through ad-hoc coalition formation. In 11th IEEE Real
Time and Embedded Technology and Applications Symposium, pages 98 -- 105. IEEE, 2005.
[34] Martin Treiber and Arne Kesting. Trajectory and floating-car data.
Springer, 2013.
In Traffic Flow Dynamics, pages 15 -- 18.
[35] Joshua O'Madadhain, Danyel Fisher, and Tom Nelson. Jung, java universal network/graph framework, 2018.
[36] Philipp Kraus. Colormap, 2018.
[37] Martin Steiger. jxmapviewer2 github, 2012.
[38] Gurobi. Gurobi optimization, 2018.
[39] Ben Stabler, Hillel Bar-Gera, and Elizabeth Sall. Transportation networks for research, 2018.
17
PFARA: A Platoon Forming and Routing Algorithm for Same-day Deliveries
A PREPRINT
[40] Kristen M Thyng, Chad A Greene, Robert D Hetland, Heather M Zimmerle, and Steven F DiMarco. True colors
of oceanography: Guidelines for effective and accurate colormap selection. Oceanography, 29(3):9 -- 13, 2016.
18
|
1510.04967 | 4 | 1510 | 2016-10-06T20:07:22 | A simple agent-based spatial model of the economy: tools for policy | [
"cs.MA",
"q-fin.GN"
] | This study simulates the evolution of artificial economies in order to understand the tax relevance of administrative boundaries in the quality of life of its citizens. The modeling involves the construction of a computational algorithm, which includes citizens, bounded into families; firms and governments; all of them interacting in markets for goods, labor and real estate. The real estate market allows families to move to dwellings with higher quality or lower price when the families capitalize property values. The goods market allows consumers to search on a flexible number of firms choosing by price and proximity. The labor market entails a matching process between firms (location) and candidates (qualification). The government may be configured into one, four or seven distinct sub-national governments. The role of government is to collect taxes on the value added of firms in its territory and invest the taxes into higher levels of quality of life for residents. The model does not have a credit market. The results suggest that the configuration of administrative boundaries is relevant to the levels of quality of life arising from the reversal of taxes. The model with seven regions is more dynamic, with higher GDP values, but more unequal and heterogeneous across regions. The simulation with only one region is more homogeneously poor. The study seeks to contribute to a theoretical and methodological framework as well as to describe, operationalize and test computer models of public finance analysis, with explicitly spatial and dynamic emphasis. Several alternatives of expansion of the model for future research are described. Moreover, this study adds to the existing literature in the realm of simple microeconomic computational models, specifying structural relationships between local governments and firms, consumers and dwellings mediated by distance. | cs.MA | cs |
1
A simple agent-based spatial model of the
economy: tools for policy
Bernardo Alves Furtado, DISET/IPEA, CNPq,
and Isaque Daniel Rocha Eberhardt, DISET/IPEA, PhD candidate UnB,
Abstract-This study simulates the evolution of artificial economies in order to understand the tax relevance of
administrative boundaries in the quality of life of its citizens. The modeling involves the construction of a computational
algorithm, which includes citizens, bounded into families; firms and governments; all of them interacting in markets for
goods, labor and real estate. The real estate market allows families to move to dwellings with higher quality or lower
price when the families capitalize property values. The goods market allows consumers to search on a flexible number
of firms choosing by price and proximity. The labor market entails a matching process between firms (given its location)
and candidates, according to their qualification. The government may be configured into one, four or seven distinct
sub-national governments, which are all economically conurbated. The role of government is to collect taxes on the
value added of firms in its territory and invest the taxes into higher levels of quality of life for residents. The model does
not have a credit market, given the emphasis of the research question on the relevance of municipal administrative
boundaries. The analysis of the markets indicate development paths and data-generating mechanisms for each
territorial approach used. The results suggest that the configuration of administrative boundaries is relevant to the levels
of quality of life arising from the reversal of taxes. The model with seven regions is more dynamic, with higher GDP
values, but more unequal and heterogeneous across regions. The simulation with only one region is more
homogeneously poor. The study seeks to contribute to a theoretical and methodological framework as well as to
describe, operationalize and test computer models of public finance analysis, with explicitly spatial and dynamic
emphasis. Several alternatives of expansion of the model for future research are described. Moreover, this study adds to
the existing literature in the realm of simple microeconomic computational models, specifying structural relationships
between local governments and firms, consumers and dwellings mediated by distance.
Index Terms-Modeling, agent-based models, public finance, taxes, municipalities, quality of life.
!
1 INTRODUCTION AND LITERATURE
T HE Brazilian tax system is paradoxical,
with high taxes, dual tax systems (taxes
and contributions) and fierce fiscal war between
federated members [1]. The complexity of the
tax system becomes more obvious and strik-
ing when considering the subnational entities.
The post-1988 constitutional decentralization
imposes the same competences to very hetero-
• B. A. Furtado is with the Department of Innovation and Produc-
tion Studies at the Institute for Applied Economic Research and
with the National Council of Research (Brazil).
E-mail: [email protected]
I. D. R. Eberhardt is with the Institute for Applied Economic
Research and the Department of Transport at University of
Bras´ılia
•
Manuscript reviewed October 6, 2016.
geneous municipalities [2]. Municipalities that
have different administrative, technical, and
political capacities; besides their inherently dif-
ferentiated borrowing leverage [3]. This hetero-
geneity among municipalities occurs not only
in relation to budgetary magnitude, but also
with regard to the disparity between central
and peripheral municipalities in metropolitan
and regional context [4], [5]. Indeed, Furtado
et al. [6] identified that there are significantly
fewer resources to metropolitan peripheral mu-
nicipalities vis-`a-vis the central city and non-
metropolitan municipalities. The authors also
suggest that such municipalities are administra-
tively inefficient, with higher expenditures and
poorer results. In addition to this reduced ad-
ministrative and financial capacity, peripheral
municipalities still have worse quality of life
and higher levels of violence [7], [8]. There is a
huge amount of literature on public spending
efficiency [1], [9], [10], which contains actual
policy propositions [11], [12], is descriptive [13],
and of a high-quality level. However, few ex-
ercises emphasize the prospective analysis that
simulates future effects of present public policy
change [14], [15], especially for the case of Brazil
and its subnational entities.
Computer simulation models for macroeco-
nomic analysis and taxes [16], [17], [18], bank-
ing and finance [19], [20], [21], stock exchange
[22], [23] and energy market [24], to name a
few applications, abound. These studies were
developed from the seminal works of Ander-
son, Pines and Arrow [25] and Arthur [26].
Recently, advances in this literature includes
models that discusses bank interconnections by
means of network analysis and systemic fault
possibilities [27], [28], [29]
Given this framework, this paper proposes
an agent-based model that is able to replicate
basic elements of an economy, its markets, its
players and its processes as simple as possible,
enabling spatial and dynamic analysis of the
This abundant literature, however, looks at
specific markets (banking, energy or exchange
markets) or seek to represent markets and its
agents and processes in detail, so that they
quickly become complex and demanding high
computing power [30], [31].
Simple models that intend to model the
interaction among actors in short-term spatial
scales are rare. Tesfatsion [32] makes an initial
proposal of a model with two products (hash
and beans) whereas Straatman et al. [33] pro-
posed a framework that simulates a market
auctions linked to a production model that to-
gether result in a simple model, but complete
and micro founded.
Lengnick [34] expands the work of Gaffeo
et al. [35] and proposes a model that simulates
macroeconomic variables, contains elements of
real estate and goods and labor markets. As
detailed below, our proposal is based on Leng-
nick's model, but makes several changes, in-
cluding explicit space in the housing market,
and subnational administrative regions.
2
central economic mechanisms. The intention is
to understand the mechanisms that generates
the observed data, so that prospective analysis
can be made. Specifically, the research question
is to identify whether the change of administra-
tive boundaries and the consequent change of
local tax revenue dynamics, in principle, alters
the quality of life of the citizens.
In addition to answering the research ques-
tion, the contribution of this study is the explicit
construction of a computational algorithm that
can be configured as a 'simulation engine of the
economy'. The paper can be said to be a mod-
ular laboratory on which small changes and
additions can be applied in order to amplify
research possibilities. Thus, the fourth section
includes specific examples of future applica-
tions of the model in addition to the exercise
done in this text.
The model adds to the literature as an adap-
tation and advancement of the approach pro-
posed by Lengnick [34]. The main contribution
is the inclusion of local governments to collect
taxes and provide public services. However,
the proposed model has a different objective
from the original. Whereas Lengnick seeks to
study effects on macroeconomic variables of
small shocks of monetary policy, this model em-
phasizes the spatial differences among different
administrative regions that collect taxes and in-
vest in their own regions through public service
provision, hence promoting the improvement
of quality of life of local people. Moreover, the
design of the model is also innovative, chang-
ing a fixed dwelling structure into one in which
families move in search of homes and regions
either with better quality of life `a la Tiebout [36]
or that best suits their current income status. 1
Another important distinction of our model
is the absence of a network-like structure that
establishes the interactions of the labor market
and the goods market. In our model the interac-
tions in these markets take place through prices
and the distance between the dwelling and the
1. See Pinto's [37, p. 75] discussion: "The decentralization and
fragmentation of the territory poses alternatives for consumers
of collective services who cannot buy services individually,
but may buy a package of services and goods that are more
preferable".
firm. Finally, the entry into the labor market is
restricted only by age and open to all members
of the family, whereas in Lengnick [34] it is
exclusive of the head of the family.
Hence, this paper proposes a simulation
model of the economy which is based on previ-
ous literature, but advances in the specific area
of simple microeconomic models, introducing
local governments and explicit spatiality of
markets.
Besides this introduction, the text includes
the presentation of the model (section 2), fol-
lowed by the presentation and discussion of
the results (section 3) and the proper sensitiv-
ity analysis. section 4 specifies possibilities of
further development of the model that can be
easily applied, justifying the work as a theo-
retical and methodological 'proposal'. section 5
presents the final considerations.
2 THE PROPOSED MODEL: METHODOL-
OGY, FEATURES AND PROCESSES
In order to model the collection of local taxes
and the provision of public goods to evaluate
and compare policy options an agent-based
model of a simple economy is presented. We
propose a model with heterogeneous agents,
dwellings, firms and governments, each with
attributes, location and specific processes at-
tached. After the description of the theoretical
model, a numerical simulation is applied to the
set of parameters, its robustness is verified by
a sensitivity analysis and the results for specific
periods of time are computed.
2.1 Agent-based modeling
2.1.1 Literature
The economic analysis based on agent-based
models has its methodological groundwork
laid by the 'Sugarscape' model, developed by
Epstein and Axtell [38]. Before that, agent-
based models were discussed in the context of
social segregation in the classical work of Nobel
author Thomas Schelling [39]; on the seminal
framework of game theory and cooperation
strategies [40] and social [41] and economic
3
sciences [42], [43]. Furthermore, complete mi-
crosimulations models of the labor market were
reported much earlier by Bergmann [44] and
Eliasson [45].
More recently, agent-based models have
been applied to learning and behavior studies,
coalition and cooperation [46]; artificial inten-
tionality [47] and education and cognition [48],
[49]2. A recent milestone in economics is the
text of Boero et al. [51] which offers a con-
ceptual and methodological description, along
with applications for human capital develop-
ment, network analysis, the interbank payment
systems, consulting firms, insurance systems
in health, ex-ante evaluation of public policies,
governance, tax, and cooperation.
The methodological steps of ABMs with an
emphasis on interpretation of empirical data
are described by Hassan et al. [52]. Two cen-
tral aspects of the methodology are verification
and validation [53], [54]. The verification step
assesses whether the adopted algorithm effec-
tively does what the modeler and the developer
planned. That is, it checks the adequacy of the
intention of the algorithm against its factual
implementation [55].
The validation process refers to the use of
historical data to assess whether the model can
minimally replicate known trajectories. It veri-
fies that the model contains the essence of the
phenomenon. Once validated, the model can be
used to indicate future trajectories. Zhang et al.
[56] illustrate this process for the adoption of
alternative car fuels.
One methodological principle of this mod-
eling process is that the decisions made and the
steps of the model are known, understood and
comparable. The scientific community suggests
two procedures (a) the adoption of protocols,
such as the Overview, Design concepts, and
Details protocol (ODD), described by Grimm
et al. [57], [58]; and (b) the availability of the
source code. The code used in this study is
available and can be requested to the authors.3
The PseudoCodes are available as appendices.
2. For a more detailed review, see Winikoff et al. [50]
3. Upon publication, the code will be made available in
GitHub and OpenABM.
The ODD protocol is also available as Appendix
section D of this work.
2.2 Attributes of the model: processes and
rules
This section describes the model, its charac-
teristics, assumptions, processes, steps, inten-
tions and limitations. Intuitively, we describe
the decision-making processes that govern the
dynamics of the model. The literature that un-
derlies the choices are listed in the processes.
2.2.1 Classes
The model was developed using the concept of
object-oriented programming (OOP) in Python,
version 3.4.4. 4
The following section describes the initial
values and allocation processes; the breakdown
of markets, the government, the spatial and
temporal sequencing of the model. Then, we
present the implementation, parameters and
limitations of the model.
Classes - initial values
The model contains five main classes:
bounded into
families;
agents:
citizens,
dwellings, firms,
and governments. The
agents' features are drawn from a uniform
age, years of
distribution and includes
schooling
initial
monetary amount.
(qualification)
and
an
Dwellings have different sizes. Their prices
are a function of its size and the value of the
square meter given by its location. Firms are
also located randomly in space and start the
simulation with some capital.
Allocation of agents into families
The modeler determines the number of
agents and the number of families of the model
exogenously. The allocation process is random.
An agent who has not been allocated is chosen
along with a family and the link is made. Thus,
the proportion of agents per family is endoge-
nous and variable. Agents maintain the same
age throughout the simulation.
Initial allocation of families into dwellings
Before simulation begins, families are ran-
domly allocated to dwellings that are vacant.
4. For an introduction in Python, see Downey [59]
4
2.2.2 Real estate market
When the simulation is already underway, the
process of modeling the real estate market is
as follows: Given a parameter chosen by the
modeler, say 0.07, that portion of the set of fam-
ilies monthly enters into a randomly composed
list of 'families on the market' in pursuit of
new residence [60]. At the same time, vacant
dwellings are selected.5 Residential prices Pt,i
are monthly updated, given the price in the
previous month Pt−1,i and the percentage of
change in the Quality of Life Index ∆IQVr of
the region where the residence is located.
Pi,t = Pi,t−1 ∗ (1 +
∆IQVr
IQVr,t−1
)
(1)
The quality of the dwelling Qi is based on the
size of the residence si which is a fixed value
and on the current Quality of Life Index [61],
[62] and it serve solely as a choice criteria for
the new residence.6
Qi = si ∗ IQVr,t
(2)
Two alternatives are available for families who
are on the market. Families whose total finan-
cial resources is higher than the median of all
families will look for houses with higher quality
and will conclude the purchase if the value of
the current family home Pi(s,r) added to the
cash available Y is higher than the value of the
better quality house intended Pi(s,r)+Y > Pj(s,r).
On the other hand, families whose available
resources are less than the median of families'
wealth will look for cheaper homes Pi(s,r) >
Pj(s,r), so that they acquire new cash [63], [61].
When observed the conditions, the change
of address is made and the difference, if moving
into more expensive homes Pi(s,r) − Pj(s,r), or
payback Pj(s,r) − Pi(s,r) is computed on the fam-
ily budget. The houses whose families moved
become vacant.
Thus, a portion of the families are always
looking for larger or better quality homes, lo-
cated in the best areas, when they have the
financial resources and the other portion of
5. The number of dwellings should always be larger than the
number of families, for this version of the model.
6. That is, properties' prices are defined by their own features,
plus local attributes.
families are in search of cheaper homes from
which they can capitalize.
discounted by an exogenous factor of propen-
sity to consume β [66].
5
2.2.3 Firms: production function and prices
The firm's production technology is fixed and
the production function depends on the num-
ber of workers lf, their qualification Ek and
an exogenous parameter α that determines pro-
ductivity.7
f (lf , Ek, α) = Σlf ∗ Eα
k
(3)
The production is updated daily according to
the above equation. In this model there is only
one product per firm.
Firms: decision-making about price adjustment
The literature confirms the rigidity of prices
and the difficulty of managerial decision-
making about the process of changing prices
[64], [65]. In the proposed model, the initial
price is set as the cost price. Firms change
their prices according to inventory levels [44].8
When the level in stock q is below the level
given by the exogenous parameter δ, prices are
adjusted upwards, in the amount stipulated by
another chosen parameter φ. This parameter is
exactly the mark-up chosen by the firm. When
the amount is above the chosen level, prices
go back to cost price. That is, when demand is
low the mark-up is zero. This proposal follows
the survey results, conducted by Blinder [65],
which indicates that only a small portion of
firms readjusts prices downwards.
(cid:40)
se q < δ, pt = pt−1 ∗ (1 + φ)
se q > δ, pt = 1
(4)
2.2.4 Goods market
Given that not all
family agents are part
of the working population, the family's total
resources are equally divided among family
members before the decision to consume. Each
customer then chooses a value for consumption
ranging between 0 and their total wealth wi,
7. Adapted from [34], [18], [35].
8. Bergmann [44] uses cost and profit information in addition
to stock levels to determine price changes. Dawid et al. [16] use
stock levels to define production quantity.
Ci = (0, wβ
i )
(5)
The family then carries out two calculations.
Given the market size parameter Γ, for exam-
ple, of five firms, each agent searches among
these firms, the one with lowest price [67],
and the one with the shortest distance of
the agent's residence [68], [69]. Randomly, the
agent chooses between lower price and shorter
distance. Intuitively, sometimes it is worth the
effort to go further to find the lowest price,
sometimes one chooses closer, though not nec-
essarily cheaper.
2.2.5 Labor market
Wages
Wages are defined as a fixed portion k 9
multiplied by the employee qualification Ei
elevated to a parameter of productivity α. The
parameter α is the same parameter of the com-
pany production function. This decision is har-
monious to the fact that more skilled workers
also produce more (in the proposed model).
wi = k ∗ Eα
i
(6)
Thus, better-qualified (and more productive)
employees have better pay.
Labor market
The firm makes decisions regarding hiring
and firing randomly, on average, once every
four months, according to an exogenously set
parameter .
The selection is made through a public ad-
vertising system. Interested companies become
part of a list. Employees between 17 and 70,
who are currently not employed, repeatedly,
apply themselves to the labor supply list.
Then, there is a matching system between
company and employee, so that the randomly
selected firm chooses the most qualified em-
ployee or the one who lives the closest [71].
Once the matching has been made, the firm
and the hired employee are removed from the
list of public announcements and a new round
of wage, distances and qualifications ranking
9. For this specification of the model, k was set to 0.65
is made. And so on, until there is no more in-
terested firms or available employees.10 When
making firing decisions, the firm just randomly
chooses an employee and let him or her go.
2.2.6 Government
Local governments in each region collect a tax
on consumption τ, at the time of purchase in
accordance with the location of the firm con-
ducting the sale. The rate is determined by an
exogenous parameter. Every month, the gov-
ernments of each region r completely transform
the resources per capita collected in linear in-
creases in the Quality of Life Index [66]. That is,
the QLI is a linear result of the summed sales
of firms in a given region, weighted by (ever-
changing) population dynamics (Nr).
IQVr,t = IQVr,t−1 + Σ
τ
Nr
(7)
In the model proposed in this paper, three alter-
natives of government administrative designs
are proposed. They are detailed in item 2.3.
2.2.7 Model sequence
The model follows the temporal distribution
proposed by Lengnick [34], which consists of
21 days to make a month, months are added to
quarters and then to years. The sequence of ac-
tions occurs with the simultaneous interaction
of various classes (see Table 1).
The model sequence can be described as
follows:
1) The modeler defines whether the sys-
tem should be configured with one, 4
or 7 regions. The simulation parameters
(see section 2.4) and the run parameters
can be changed.
2) Regions, agents, families, households
and firms are created, given the param-
eters provided.
3) Agents are allocated to families and
families are allocated to dwellings.
4) Before the actual start of the simulation
time, the initial framework includes the
10. Neugart et al. [72] have discussed matching mechanisms
for the labor market bu ponder that there is insofar no concen-
sus about the best procedure.
6
creation of one product by firm and an
initial round of hiring.
5) When the simulation begins, the pro-
duction function is applied everyday
for all firms.
6) At the end of each month:
Firms pay wages;
a)
b) Households consume and (in
the same transaction) govern-
ments collect taxes;
d)
c) Governments apply their avail-
able resources into the update of
QLI;
Firms update their profits, given
their last quarter capital;
Firms update product prices;
If necessary, firms post job offers
or fire employees;
e)
f)
g) Unemployed workers
offer
themselves for
the vacancies
and the matching process is
carried out;
h) A share of the families enter
the housing market and perform
transactions.
7) Every quarter, companies report profits
for the period
2.2.8 Indicators and iterations
For this paper, the results were obtained with
1,000 iterations for each spatial divisions (one,
4 and 7 regions).
2.3 Spatial emphasis of the model
The model has a clear emphasis on its spa-
tial aspects, as space is central to answer the
research question. Calculation of the distance
(and accessibility) is present in two moments:
(a) at the choice of the employee by the firm
and (b) when the consumer chooses between
price and (easy) location. As a share of fam-
ilies is relocating, these distance calculations
are dynamic and change the relations among
firms and consumers and firms and workers
every month. That is, firms and dwellings are
fixed, but families move constantly, ensuring
the spatial dynamics of the model.
7
Sequencing and interaction between classes and temporal dynamics of the algorithm. Start processes at 'setup' and 'day 0',
followed by days that add up into months and months into quarters, successively, until the period determined by the modeler.
Items with an asterisk indicate the need for exogenous parameters
Table 1
Agents
Families
Firms
Dwellings
Government
Setup Creation *
Allocation of agents
into families
Day 0
Apply for position
Production
Days ⇓
Months ⇓ Wages
Production
Purchase **
Fire
Apply for position
Quarters ⇓
Years ⇓
Creation *
Allocation of agents
into families
Allocation families in
dwellings.
Family per
distribution
capita
Enters Real
market
estate
Creation *
Creation *
Creation *
Allocation families in
dwellings.
Initialization
Address register
Register
Register
Purchase (Tax) *
Update QLI
Update prices
Inform QLI
Creation of product
Offer position
Hire *
Address register
Production
Production
Wages
Purchase
Update profits
Decide on prices
Offer position/Fire
Hire *
Update profits
Furthermore, the QLI is a linear and spa-
tially compartmentalized reflection of firms
sales in each region. This same QLI, in turn, af-
fects the prices of dwellings. Thus, the housing
market and the goods and labor markets are all
spatially linked.
Besides the presence of spatial interaction
in the processes themselves, the model also
differentiates the applied regions design, ac-
cording to the scheme of Figure 1. The figure
shows the coordinates of the central point (0, 0).
Along the four directions, the boundaries can
be established by parameters. This study uses
the parameters 10, -10, 10, -10, for the North,
South, East and West directions, respectively.
Three different design were used and are
applied by changing the number of regions η.
If η is equal to 1, the model runs with only
one region, with code 0, which encompasses the
entire space. With η equal to 4, the model runs
with four regions, with codes 0, 1, 2 and 3, and
region 3 covers the entire area of subregions 3,
4, 5 and 6 in Figure 1. And finally, with seven
regions, the model follows the configuration of
the codes of Figure 1, with four smaller regions
and three larger ones.
2.4 Model implementation and parameters
Running the model is simple and done with
just one command. Optionally, the modeler can
set the parameters for each simulation and for
8
2.5 Limitations
The limitation of this study arises from the
difficulty of finding complete, integrated, and
simple models that could be used as initial
steps to be expanded and adapted by follow-
ing researchers. In fact, despite the models of
Lengnick [34] and Gaffeo et al. [35], all others
are specific to a single market, such as energy
[73], finance [74] or labor market [75]; or are too
complex [76], [30].
Thus, the task of investigating a specific
phenomenon (in this case, the spatial influence
of consumption taxes), under a single inte-
grated model, requires that all processes (firm
production firm, goods and labor market) and
all its associated parameters are explained, the-
oretically sound and numerically compatible.
In practice, this requires that the modeler get
results that are consistent with the literature
for indicators such as inflation, GDP growth,
unemployment, household income, simultane-
ously and temporally consistent.
The question put to modelers is of an epis-
temological nature. How can we determine
which are the central elements of the phe-
nomenon, which must be present, and what
are the accessory elements? At what point, sim-
plifying the process can take place and where
there is significant change of the observed phe-
nomenon?
Besides this general limitation, this version
of the model also does not include the credit
market, demographic changes nor investment
in social capital.
3 RESULTS AND DISCUSSION
The most thriving economy is the model with
seven regions (Figure 3), on average with me-
dian GDP only 30% higher when compared to
the model with four regions. This model with
four regions, in turn, achieves results that are
38% above the model with a single region (Fig-
ure 2). The variability is higher for the model
with seven regions, vis-`a-vis the one region
model (Table 3 and Figure 4).
Figure 1. Division of space into regions, with coordinates and
codes. When running with one region, the simulation space
is complete, from north to south, from east to west, including
all colors. When running with four regions, the regions are
symmetrical, including the codes 0, 1, 2 and the entire region
that consists of subregions 3, 4, 5, and 6. With seven, there are
larger (0, 1 and 2) and smaller (3, 4, 5 and 6) regions.
the run itself, as described below. A system-
atic analysis of the parameters that assesses
whether small changes significantly affect the
results and seek to confirm the robustness of
the model (sensitivity analysis) is made after
the presentation of results.
For each simulation run you can choose
the number of agents, families, households, the
time duration in days, the number of local gov-
ernments in which space is divided (namely,
municipalities, with competence over their ter-
ritory) and the path file to save the results (??).
The proposed model contains a very small
number of exogenous parameters. The param-
eters help understand how the model mech-
anisms influence its results. Parameter α, for
example, can be set to 1 so that its effect is zero.
By reducing the parameter successively by 0.1,
one can observe the effect of increasingly less
productive workers. The same understanding
of relevance can be made with the β parameter,
or the rate of consumption tax. This construc-
tion offers some flexibility to the modeler. This
flexibility is most relevant if the goal
is to
increase understanding of the problem that is
modeled and the model is used as a guiding
tool for decision-making, or as a methodology
to discuss 'what if' questions.
9
Parameters of the simulation (that define each run) and exogenous parameters of the model.
Table 2
Simulation parameters
Values
Possibilities' intervals Observations
Number of days
Number of agents
Number of families
Number of dwellings
Number of firms
Number of regions
Model parameters
Firms
Alpha
Beta
Quantity to change prices δ
Frequency of entry in the la-
bor market
Mark-up
Agents
Market size Γ
Consumption satisfaction
Families
Real estate market
Government
Consumption tax
5,040
1,000
400
440
110
1; 4; 7
0.25
0.87
10
0.28
0.03
100
0.01
0.021
0.21
(63, 12,800)
(10, 10,000)
(4, 2,000)
(5, 2,200)
(2, 1,000)
(1; 4; 7)
(0, 1)
(0, 1)
(100, 2,000)
(0, 1)
(0, 1)
(1, 1,000)
[0, 1)
(0, 1)
(0, 1)
The model was developed to run up to 50 years, however
with loss of explantory power. We ran the model for 20 years.
The growth of the number of agents makes the simulation
slower.
Endogenously, it's used to define the average number of
agents per families. The suggestion is to have 2.5 agents per
family, on average
Necessarily higher than the number of families number.
Vacancy in Brazil is around 11%.
Approximately 10% of the number of agents.
Alternative number of regions to run the model.
Production function expoent. When set to "1", it does not
change the model, when set to "0", the production of the
firm is one unit.
Consumption function exponent. When set to "1" consump-
tion vary from zero to the total of available money.
Threshold to change prices.
Time frequency of decision-making on labor market. When
set to "0", the evaluation is made every month. When set
to "0.25", the firm enters the market three times every four
months, on average.
Percentage added to prices when demand is high (product
level on inventory is below "Quantity to change prices").
Number of firms checked before agents make decision to
consume. Can be set between "1" and the total number of
firms.
Used to measure satisfaction gained with consumption.
Percentage of families' entering real estate market
Tax on consumption.
Figure 2. GDP growth results, 1,000 iterations, for one region.
Figure 3. GDP growth results, 1,000 iterations, for seven re-
gions.
Considering
the
the
economies converge towards full employment,
labor market,
10
Figure 4. BoxPlot of GDP for the last month of the simulation,
1,000 iterations, for the three regions.
Figure 5. Unemployment, 1,000 iterations and one region.
Median, first and third quartiles for the last month of the
simulations (1,000 iterations) of Gini coefficient, GDP, QLI and
Table 3
families' wealth for each regional design
Gini
0.25
0.890
0.925
0.935
GDP
0.25
1,056,341
2,029,562
2,794,786
QLI
0.25
223.2
425.0
562.1
0.25
589
32,126
82,190
Regions
1
4
7
Regions
1
4
7
Regions
1
4
7
Regions
1
4
7
Median
0.916
0.939
0.944
Median
1,568,746
2,904,486
3,788,903
Median
331.8
608.9
761.8
Families' wealth
Median
19,757
253,450
573,573
0.75
0.932
0.946
0.950
0.75
2,314,751
3,897,903
4,501,469
0.75
487.3
820.9
945.9
0.75
115,815
893,763
1,645,296
keeping a cycle of very low unemployment,
throughout the period, for the three regional
designs show similar results (Figure 5 and
Figure 6).
Household income varies
significantly
among the three regional designs, for one re-
gion (Figure 7) variation is of lower magnitude
when compared to the model with seven re-
gions (Figure 8) through 1.000 iterations. The
Figure 6. Unemployment, 1,000 iterations and seven regions
median, first and third quartiles are higher for
the model with seven regions, vis-`a-vis the
model with only one region (Table 3 and Fig-
ure 6). Although the dispersion is higher for the
model with seven regions, the value of the first
quartile is slightly less than the median with a
single region (Figure 9).
The Gini coefficient is computed on the
utility of the families. Utility is directly pro-
portional to the cumulative consumption of the
families. The GINI coefficient reaches a higher
level in the model with 7 regions when com-
pared to the two other models (Figure 10).
Moreover, the behavior of the coefficient in the
1,000 iterations has a similar pattern of vari-
ability (Figure 11 and Figure 12) with slightly
higher standard deviation.
Finally, the basic indicator to compare the
11
Figure 7. Families' wealth, 1,000 iterations, for one region.
Figure 10. BoxPlot of the Gini coefficient for the last month of
the simulation, 1,000 iterations, for the three regional design.
Figure 8. Families' wealth, 1,000 iterations, for seven regions.
Figure 9. BoxPlot of families' wealth for the last month of the
simulations, 1,000 iterations, for the three regional designs
Figure 11. Results for the Gini coefficient, 1,000 iterations, one
region.
performance of the models is the Quality of Life
Index for each simulated design for one region
(Figure 13) and for seven regions (Figure 14).
The median and the quartiles are higher for the
model with seven regions vis-`a-vis the other
models (Figure 15).
The model results indicate that changes in
administrative boundaries have led to robust
changes among the three considered region de-
sign. According to the procedures described,
the dynamism of the real estate market, namely,
household mobility in the simulations with
more than one region, was relevant to the re-
sults.
12
Figure 12. Results for the Gini coefficient, 1,000 iterations, seven
regions.
Figure 13. Results of the Quality of Life Index (QLI), 1,000
iterations, for one region.
Figure 15. BoxPlot of the QLI, for the last month of the simula-
tion, 1,000 iterations for the three regional designs.
with income level below the median become
sellers in the real estate market. Thus, these
families capitalize on the sale of homes whose
prices increased along with the quality of life
in the region, and migrate toward regions with
poorer quality. This movement is partly coun-
terbalanced by families trying to migrate in
search of better quality.
As a result, the models with subdivisions
lead to regions that are less populated and
have better quality of life and at the same time,
more populated regions that have worse QLI
(Table 4).
Median and standard deviation of the maximum and minimum
regional values for each simulation for QLI and population,
Table 4
1,000 iterations, by regional design.
Regions
QLI me-
dian
QLI
std
Pop.
Median
Pop. Std
1
4
7
Max
Min
Max
Min
Max
Min
333.5
333.5
860.2
423.2
1,499.2
343.9
210.9
210.9
430.2
198
2,047.7
195.3
1,000
1,000
207
288
40
83
30.3
29.2
70.2
89.9
Finally, it is noteworthy the trade off be-
tween the results for the three models. While
the model with seven regions is more dynamic,
more productive and wealthier, it is also more
Figure 14. Results of the Quality of Life Index (QLI), 1,000
iterations, for seven regions.
In the absence of a credit market, families
heterogeneous. The model with one region, in
turn, is more harmonic but less vigorous.
The underlying assumption of the authors
that the model with one region would be more
efficient from the standpoint of conurbation
regions was not observed with the present
configuration. Especially given the strength and
mobility of the real estate market that con-
centrates smaller populations in regions with
higher quality and larger populations in ar-
eas with poorer quality. However, the research
question that is, if administrative boundaries
influence the economic and fiscal dynamic of
the regions can be confirmed.
Yet, the results indicate the wealth of possi-
bilities of analysis of the economic system from
heterogeneous agents and firms in an environ-
ment that is continually changing.
Finally, given the process of creating artifi-
cial economies, at each loop iteration the agents
and firms are completely different. Thus, the
next phase of research, which is the model
application to real data, will input actual data
as attributes of the economy and thus reduce
the variability of results.
3.1 Sensitivity analysis and robustness
The sensitivity analysis is central in building
simulation models to ensure that the model
is structurally consistent and does not depend
solely on a particular parameter, which is ad-
justed for a specific value. Furthermore, the sen-
sitivity analysis may serve as an analytical tool
to show how and with which magnitude cer-
tain configurations and model processes change
trends and results.
The sensitivity analysis made was based on
the variation of the model parameters in 10
different values between their minimum and
maximum values (Table 5). As random num-
bers influence the model results, comparing the
results of different iterations (model runs) is
only possible if we use the same seed. Thus,
if the model
is run several times with the
same parameter and the same seed, the same
results will be produced. Therefore, when the
modeler changes the parameters, variations in
sensitivity analysis results will be a result of the
13
model structure and not of the random number
generator.11
The change of the parameters was per-
formed separately (one parameter at a time)
with the other parameters maintained at their
default values, defined in a first exploratory
analysis.
Given that the premise is to create a model
(or simulation machine), the sensitivity analysis
furthers our understanding of the model. Of
course, we also varied the number of regions
(1, 4, 7).
3.1.1 Alpha
The variation of each of the parameters affects
differently the results of the simulations. The
alpha parameter which evaluates worker pro-
ductivity for example, leads to higher values
of total GDP when values are between 0.32 and
0.37.
Considering unemployment, the alpha pa-
rameter provides full employment, conditioned
to other parameter's default value. Figures Fig-
ure 16 and Figure 17 shows that unemployment
converges quickly towards full employment.
The behavior was similar for all regions. The
variation in productivity and unemployment
indicate that when worker productivity is very
high, there is an excess supply that is not ab-
sorbed by the market.
There is greater variation in firms' profits
when alpha is smaller and more stability when
alpha is higher (Figure 18).
The value of alpha at 0.35 results in a good
balance of resources in the economy among
firms and households, on average, for all the
region designs.
The Gini coefficient is slightly higher for
higher values of alpha. Finally, prices rise
slightly to higher values of alpha.
3.1.2 Beta
The beta parameter which controls the propen-
sity to consume of households
influences
strongly the economy. In fact, higher values
of beta (lower discount at maximum limit of
11. The results of the sensitivity analysis were obtained using
a fixed seed.
Simulation parameter values used in the sensitivity analysis
Table 5
14
Parameters Values
Alpha
Beta
Quantity to change prices
Mark-up
Labor market entry
Market size
Real estate entry
Tax consumption
0.1
0.5
10
0.01
0.1
1
0.01
0.01
0.14
0.55
42
0.04
0.14
3
0.02
0.06
0.19
0.61
74
0.06
0.19
5
0.03
0.11
0.23
0.66
107
0.09
0.23
7
0.04
0.16
0.28
0.72
139
0.12
0.28
10
0.05
0.21
0.32
0.77
171
0.14
0.32
15
0.06
0.25
0.37
0.83
203
0.17
0.37
30
0.07
0.3
0.41
0.88
235
0.2
0.41
50
0.08
0.35
0.46
0.94
268
0.22
0.46
70
0.09
0.4
0.5
0.99
300
0.25
0.5
110
0.1
0.45
Figure 16. Results of alpha variation on unemployment, one
region.
Figure 18. Firms' profit variation results for Alpha 0.1 and 0.9,
and one region
Figure 17. Results of alpha variation on unemployment, seven
regions.
Figure 19. Results for the variation of parameter beta on unem-
ployment values for one region.
household spending) or lower levels of beta,
which restricts consumption, lead to high and
persistent levels of unemployment (Figure 19
and Figure 20), especially in the model with one
region, where the dynamics is more dependent
on the goods market. Low values also keep
firms' profits close to zero.
The impact of beta in the Gini coefficient
is relevant and similar among the regional de-
signs used. For low values of beta and low
15
Figure 20. Results for the variation of parameter beta on unem-
ployment values for seven regions.
Figure 21. Variation in the results of the Gini coefficient for beta
values 0.1 and 0.9, one region.
household consumption, Gini rises gradually,
reaching a maximum at about 0.5. However,
when beta has a value of 0.99, inequality rises
steeply to reach values close to 0.90 at the end
of the period (Figure 21).
3.1.3 Tax on consumption
As expected, the value of the tax rate influences
the economy on many levels. Lower rates lead
to lower unemployment, but influence little
when below 0.5 (Figure 22 and Figure 23). Very
high tax rates bring hyperinflation, widespread
unemployment and a significant drop in rev-
enues and profits of firms. However, given that
the amounts collected in taxes are applied in the
same regions where collected, QLI improves,
and consequently, property prices increase ac-
cordingly. Household income (Figure 24 and
Figure 25) and GDP are higher for intermediate
values of tax rate
3.1.4 Other parameters
A sensitivity analysis was performed for each
exogenous parameters of the model, with lower
relative impact compared to parameters alpha,
beta and the tax rate.
The level of the inventory that
triggers
changes in prices, for example, seems to impact
slightly on the evolution of the price index,
delaying its increase.
The frequency with which firms enter the
labor market affects the speed of adjustment
Figure 22. Results of unemployment for various consumption tax
rate values, one region.
Figure 23. Results of unemployment for various consumption tax
rate values, seven regions.
in the labor market. When parameter values
16
Finally, the percentage of families entering
the housing market seems to have little influ-
ence. When all families are on the market all
the time, there is a small reduction in household
income.
Thus, we understand that the variation of
the results of the model given by the variation
of the parameters is in line with the underly-
ing theory. In addition, there is no change of
parameters that cause different or unexpected
behavior of the model. Thus, we believe it indi-
cates the robustness of the model, as described.
THEORETICAL-
4 THE
APPROACH:
METHODOLOGICAL
POSSIBILITIES FOR FUTURE RESEARCH
This section describes various additions to the
model that could be implemented with rela-
tively small and simple changes in the current
code. Given the seminal methodological trait
of this paper, we thought the model should
be developed in its simplest form possible, fol-
lowing the KISS logic ('keep it simple, stupid').
Eventually, it could evolve into the KIDS form
('keep it descriptive, stupid'), formulated by
Edmonds and Moss [77].
The immediate interest of the authors, it
is to apply it for the Federal District region,
in Brazil. Empirical data would be used in
the initial configuration of the model, namely:
actual municipal boundaries, specific spatially-
bound demographic patterns, actual companies
attributes and location, and supply of skilled la-
bor. The following step would be to validate the
empirical model for a given time line, seeking
correlation or similarity between the evolution
of observed indicators and those produced by
the model. Finally, after validation, the model
could be effectively used to implement public
policy alternatives.
The actual realm of research possibilities are
detailed below, following the KIDS argument of
Edmonds and Moss [77].
1)
Implement demographic change, with
processes that describe birth, deaths
and families creation in order to be-
come a more dynamic and real model
Figure 24. Families' wealth results for various consumption tax
rate, one region.
Figure 25. Families' wealth results for various consumption tax
rate, seven regions.
are higher
taking longer to enter the labor
market unemployment is only insignificant at
the end of the period. When the entry of firms
is frequent, full employment is achieved within
months.
The change in the mark-up value, i.e., the
percentage increase in product prices of firms
when their stock is low does not greatly change
the profit levels of the firms. However, very
high mark-up rates, lead to uncontrolled infla-
tion after some time.
The size of the market checked by con-
sumers when they go shopping does not inter-
fere in the results, with little loss of momentum
when size is restricted to only one firm.
2) Another relatively simple alternative is
the inclusion of updating workers qual-
ification (years of study), deducing in-
vestment from their resources;
3) The credit market, with production and
consumer financing possibility is also
relevant to make the model closer to
economic reality. The literature is al-
ready available [19], [20], [21];
4) Currently, the market for goods is re-
stricted to firms and consumers in the
domestic market. However,
it could
also include firms and governments as
buyers (and sellers), enabling analysis
of intermediate sectors, as well as for-
eign buyers allowing the inclusion of
an economic measure of exports and
trade balance;
while enabling results for specific de-
mographic cohorts. In addition, inter-
temporal analyzes involving inheri-
tance (of wealth or social capital), could
also be tested;
5) Although distance is already included
in the model, the formula could be
sophisticated to effectively include the
transport system available in the mu-
nicipalities that are object of study. As
a result, accessibility analysis would be
systemically integrated with the rest of
the economy, as demand and supply of
the transport system (for employment
purposes).
6) The process of imposing a limit by time
or distance to daily commute would
endogenously enable the creation of a
system with several regions, making it
simple to study urban hierarchy anal-
ysis. In such case, the 'employment
areas' would be endogenous to the
model.
Firms and their production technolo-
gies, decision-making processes and
hiring and firing could be drawn from
tacit information specific to a particular
firm or sector.
7)
8) The taxation system of this model is
simplistic, with only one tax applied to
consumption, typically a value-added
17
tax (VAT) levied on the location of the
firm. However, note the reader, that the
implementation of the Territorial Taxes
on property or on income, or changing
VAT to be collected at the destination,
i.e., at the consumer's place of resi-
dence, could be easily implemented.
Thus, specific research questions of fis-
cal interest could be investigated.
Indeed, it is worth mentioning the advan-
tage of modularity within the scope of this
work. Using the basic model
is possible to
detail, build, and expand the model module
by module, according to the research needs,
while ensuring the evolution of the integration
of other processes already implemented and
validated.
Anyway, this list is not exhaustive and only
fulfills the job of informing the model expan-
sion opportunities, through enhanced feature of
this theoretical and methodological proposal.
5 FINAL CONSIDERATIONS
This paper specifies, explains and justifies the
steps and processes of
the construction of
the computational algorithm that prospectively
simulates a spatial economy. It adds to the
literature on the explicit spatiality of the model,
and in achieving a simple model with three
markets and conurbated subnational govern-
ments. Thus, establishing an actual framework
for economic simulation, constituting itself as a
public policy tool.
The model has a dynamic real estate mar-
ket with prices given by the features of the
dwelling and its location; a labor market, with
matching mechanism between skilled workers
and companies; and a goods market with en-
dogenous price adjustment based on stock. The
configuration in different subnational govern-
ments, one, four or seven differentiated regions
allows for explicit spatial analysis.
The results and trends obtained after 1,000
simulation runs indicate that mobility of fam-
ilies among regions is central to the model
with impoverished families migrating to poorly
serviced places and, therefore, lower real estate
prices; and families that are financially well
migrating to better quality areas. Therefore, the
model with only one region has a less dy-
namic economy, although more homogeneous,
whereas the model with seven regions shows
greater dynamism, but also greater heterogene-
ity and inequality.
The research question that asked whether
the change of administrative boundaries and
the consequent change of local tax revenue dy-
namics, in principle, changes the quality of life
of the citizens ' can be answered affirmatively.
Indeed, administrative boundaries understood
as enclosed area of tax collection over economic
base and its investment as collective public
services can alter the quality of life of citizens.
The underlying question faced by this paper is
the efficiency of the return of taxes to taxpayers.
Is there a spatial, political and administrative
configuration that is more efficient? This de-
bate should be further discussed by following
research.
Finally,
contributes
this paper
to the
methodological framework of economic tools,
particularly those flexible and forward-looking,
with applied realm to public policies of subna-
tional entities.
ACKNOWLEDGMENTS
The authors would like to thank the National
Council of Research (CNPq/BR).
REFERENCES
[2]
[1] A. Afonso, A. Romero, and E. Monsalve, "Public Sector
Efficiency: Evidence for Latin America." IDB Publications,
no. 80478, 2013.
F. Rezende, "Federalismo Fiscal: em busca de um novo
modelo," in Educa¸cao e federalismo no Brasil: combater as
desigualdades, garantir a diversidade. Bras´ılia: Unesco, 2010,
pp. 71–88.
[3] O. Canuto and L. Liu, Until debt do us part: Subnational debt,
insolvency, and markets. World Bank Publications, 2013.
[4] M. E. P. B. Antinarelli, "Federalismo, autonomia municipal
e a constitucionalizac¸ ao simb´olica: uma an´alise da de-
pendencia financeira dos pequenos munic´ıpios mineiros."
Revista da Faculdade de Direito da UFMG, no. 61, pp. 445–
472, 2012.
F. Rezende and S. Garson, "Financing Metropolitan Ar-
eas in Brazil: Political, Institutional, Legal Obstacles and
Emergence of New Proposals For Improving Coordina-
tion," Revista de Economia Contemporanea, vol. 10, no. 1,
2006.
[5]
18
[6] B. Furtado, L. Mation, and L. Monasterio, "Fatos es-
tilizados das financ¸as p ´ublicas municipais metropolitanas
brasileiras entre 2000-2010." in Territ´orio metropolitano,
pol´ıticas municipais. Bras´ılia: Furtado, Bernardo; Krause,
Cleandro; Franc¸a, Karla, 2013, pp. 291–312.
[7] L. T. Andrade and F. O. V. d. Figueiredo, "Vulnerabilidade
social e criminalidade na Regiao Metropolitana de Belo
Horizonte," in XII Congresso Brasileiro de Sociologia, 2005,
p. 19.
J. J. Waiselfisz, Mapa da Violencia.
Sangari, 2012.
Sao Paulo: Instituto
[8]
[9] C. E. Gasparini and R. B. Miranda, "Transferencias,
equidade e eficiencia municipal no Brasil." Planejamento
e Pol´ıticas Pblicas, no. 36, Oct. 2011.
[10] R. O. Orair, C. H. M. Santos, W. d. J. Silva, J. M. d. M.
Brito, A. d. S. Ferreira, H. L. Silva, and W. S. Rocha, "Uma
metodologia de construc¸ ao de s´eries de alta frequencia das
financ¸as municipais no Brasil com aplicac¸ ao para o IPTU e
o ISS: 2004-2010." Pesquisa e Planejamento Econmico, vol. 41,
no. 3, pp. 471–508, 2011.
J. R. R. Afonso, "Imposto de renda e distribuic¸ ao
de renda e riqueza: as estat´ısticas fiscais e um
da Receita
debate
aduaneiros,
Federal:
1,
no.
1,
[Online]. Available:
http://www.revistadareceitafederal.receita.fazenda.gov.
br/index.php/revistadareceitafederal/article/view/97
premente
estudos
pp.
no Brasil." Revista
e
2014.
tribut´arios
28–60,
[11]
vol.
[12] S. W. Gobetti, "Ajuste fiscal no Brasil: os limites do
poss´ıvel." Textos para Discussao, vol. 2037, 2015.
[13] C. H. M. d. O. Santos and R. R. O. Gouvea, Finan¸cas
p´ublicas e macroeconomia no Brasil: um registro da reflexao
do Ipea (2008-2014). Bras´ılia: IPEA, 2014.
[14] L. T. Brandalise, C. A. Rojo, D. M. da Mata,
and A. F. de Souza, "Simulac¸ ao de
e
formulac¸ ao de estrat´egias competitivas: o caso do
atacado lideranc¸a." Revista Gestao & Tecnologia, vol. 12,
no. 3, pp. 223–257, 2012.
[Online]. Available: http:
//revistagt.fpl.edu.br/get/article/view/464
cen´arios
[15] G. Carvalho, A. Oliveira, and C. Oliveira, "Cen´arios de
longo prazo para a cafeicultura brasileira: 2006-2015." in
The Proceedings of SOBER. Sober, 2015.
[16] H. Dawid, S. Gemkow, P. Harting, and M. Neugart,
"Labor market integration policies and the convergence
of regions: the role of skills and technology diffusion."
Journal of Evolutionary Economics, vol. 22, no. 3, pp.
543–562, 2012. [Online]. Available: http://link.springer.
com/article/10.1007/s00191-011-0245-1
[17] G. Dosi, G. Fagiolo, M. Napoletano, and A. Roventini,
"Income Distribution, Credit and Fiscal Policies in an
Agent-Based Keynesian Model." SSRN eLibrary, Jan. 2012.
[18] G. Dosi, G. Fagiolo, and A. Roventini, "The microfoun-
dations of business cycles: an evolutionary, multi-agent
model." in Schumpeterian Perspectives on Innovation, Com-
petition and Growth, U. Cantner, J.-L. Gaffard, and L. Nesta,
Eds. Springer Berlin Heidelberg, Jan. 2009, pp. 161–180.
"Possible
long-range dependence in the Brazilian
causes of
and
stock market." Physica A: Statistical Mechanics
its Applications,
635–645,
2005. [Online]. Available: http://www.sciencedirect.com/
science/article/pii/S0378437104010052
[19] D. O. Cajueiro
B. M.
Tabak,
345,
and
vol.
pp.
no.
3,
[20] --, "The role of banks in the Brazilian Interbank
Market: Does bank type matter?" Physica A: Statistical
Mechanics and its Applications, vol. 387, no. 27, pp. 6825–
6836, 2008. [Online]. Available: http://www.sciencedirect.
com/science/article/pii/S0378437108007498
[21] B. M. Tabak, D. O. Cajueiro, and T. R. Serra,
"Topological properties of bank networks:
the case
of Modern Physics
of Brazil."
C, vol. 20, no. 08, pp. 1121–1143, 2009.
[Online].
Available: http://www.worldscientific.com/doi/abs/10.
1142/S0129183109014205
International
Journal
[22] B. LeBaron, "Agent-based Computational Finance." in The
Handbook of Computational Economic, L. T. a. K. Judd, Ed.
Elsevier, 2006, vol. 2, pp. 1187–1233.
[23] R. Palmer, W. Brian Arthur, J. H. Holland, B. LeBaron,
and P. Tayler, "Artificial economic life: a simple model of
a stockmarket." Physica D: Nonlinear Phenomena, vol. 75,
no. 1, pp. 264–274, 1994. [Online]. Available: http://www.
sciencedirect.com/science/article/pii/0167278994902879
[24] B. LeBaron and L. Tesfatsion, "Modeling macroeconomies
as open-ended dynamic systems of interacting agents."
The American Economic Review, pp. 246–250, 2008.
[25] P. W. Anderson, K. Arrow, and D. Pines, "The economy
as an evolving complex system." in The Proceedings of the
Evolutionary Paths of the Global Economy Workshop, Santa
Fe, NW, 1988.
[26] W. B. Arthur, "Inductive reasoning and bounded ratio-
nality." The American Economic Review, vol. 84, no. 2, pp.
406–411, 1994.
[27] L. Bargigli and G. Tedeschi, "Interaction in agent-based
economics: A survey on the network approach." Physica
A: Statistical Mechanics and its Applications, vol. 399, pp.
1–15, 2014. [Online]. Available: http://www.sciencedirect.
com/science/article/pii/S0378437113011539
[28] R. Grilli, G. Tedeschi, and M. Gallegati, "Markets
of
connectivity
Economic Interaction and Coordination, pp. 1–18, 2014.
[Online]. Available: http://link.springer.com/article/10.
1007/s11403-014-0129-1
and financial
contagion."
Journal
[29] W. Ya-Qi, Y. Xiao-Yuan, H. Yi-Liang, and W. Xu-An,
"Rumor spreading model with trust mechanism in
complex social networks." Communications in Theoretical
Physics, vol. 59, no. 4, p. 510, 2013. [Online]. Available:
http://iopscience.iop.org/0253-6102/59/4/21
[30] S. Van Der Hoog, C. Deissenberg, and H. Dawid, "Produc-
tion and Finance in EURACE," in Complexity and artificial
markets. Springer, 2008, pp. 147–158.
[31] W. Guocheng, S. Yuna, W. Jie, and W. Zili, "Application
analysis on large-scale computation for social and eco-
nomic systems." in The Proceedings of IEEE SMC. Hong
Kong: IEEE SMC, 2015.
[32] L. Tesfatsion, "Agent-Based Computational Economics: A
Constructive Approach to Economic Theory." in Handbook
of Computational Economics. Elsevier, 2006, vol. 2, pp. 831–
880.
[33] B. Straatman, D. J. Marceau, and R. White, "A generic
framework for a combined agent-based market and
production model." Computational Economics, vol. 41,
no. 4, pp. 425–445, 2013.
[Online]. Available: http:
//link.springer.com/article/10.1007/s10614-012-9341-z
[34] M.
Lengnick,
"Agent-based macroeconomics: A
of Economic Behavior &
baseline model."
Organization, vol. 86, pp. 102–120, 2013. [Online]. Avail-
able: http://www.sciencedirect.com/science/article/pii/
S0167268112002806
Journal
[35] E. Gaffeo, D. D. Gatti,
S. Desiderio,
and
M. Gallegati, "Adaptive microfoundations for emergent
Journal, vol. 34,
macroeconomics." Eastern Economic
no.
[Online].
Avail-
http://www.palgrave-journals.com/eej/journal/
able:
v34/n4/abs/eej200827a.html
441–463,
2008.
pp.
4,
19
[36] C. M. Tiebout, "A pure theory of local expenditures."
Journal of Political Economy, vol. 64, pp. 416–424, 1956.
[37] V. C. Pinto, Direito urban´ıstico: plano diretor e direito de
propriedade, 4th ed. Sao Paulo: Revista dos Tribunais, 2014.
J. M. Epstein and R. Axtell, Growing artificial societies:
social science from the bottom up. Cambridge, MA: Brook-
ings/MIT Press, 1996.
[38]
[39] T. C. Schelling, "Models of segregation." The American
Economic Review, pp. 488–493, 1969. [Online]. Available:
http://www.jstor.org/stable/1823701
[40] R. Axelrod and W. D. Hamilton, "The evolution of
cooperation." Science, vol. 211, no. 4489, pp. 1390–1396,
1981. [Online]. Available: http://www.sciencemag.org/
content/211/4489/1390.short
J. Holland, "Complex adaptive systems." Daedalus, vol.
121, pp. 17–30, 1992.
[41]
[42] T. Ciarli,
[43]
[44] B.
"Structural
Interactions
and Long Run
Growth." Revue de l'OFCE, vol. 124, no. 5, pp. 295–
345, 2012. [Online]. Available: http://www.cairn.info/
revue-de-l-ofce-2012-5-page-295.htm
J. Holland and J. H. Miller, "Artificial Adaptive Agents in
Economic Theory." The American Economic Review, vol. 81,
no. 2, pp. 65–71, 1991.
Bergmann,
the
macroeconomy with explicitly represented money flows,"
in Annals of Economic and Social Measurement, Volume 3,
number 3. NBER, 1974, pp. 475–489. [Online]. Available:
http://www.nber.org/chapters/c10173.pdf
"A microsimulation
R.
of
[45] G. Eliasson, G. Olavi, and M. Heiman, "A Micro-
Macro Interactive Simulation Model of
the Swedish
Economy: Preliminary Model Specification," IUI Working
Paper, Tech. Rep., 1976.
[Online]. Available: http:
//www.econstor.eu/handle/10419/94789
[46] L. G. Nardin and J. S. Sichman, "Trust-based coalition
formation: A multiagent-based simulation." in Proceedings
of the 4th World Congress on Social Simulation, 2012.
J. S. a. Sichman, and H. Coelho,
[47] D. F. Adamatti,
"An analysis of
the insertion of virtual players in
GMABS methodology using the vip-Jogoman prototype."
and Social Simulation,
Journal
vol. 12, no. 3, p. 7, 2009.
[Online]. Available: http:
//jasss.soc.surrey.ac.uk/12/3/7.html
of Artificial Societies
[48] S. Maroulis, E. Bakshy, L. Gomez, and U. Wilensky, "An
agent-based model of intra-district public school choice."
Nortwwestern University Working Papers, p. 30, 2010.
[49] --, "Modeling the transition to public school choice."
Journal of Artificial Societies and Social Simulation, vol. 17,
no. 2, 2014.
[50] M. Winikoff, N. Desai, and A. Liu, "Principles and Practice
of Multi-Agent Systems," Multiagent and Grid Systems,
vol. 8, no. 2, 2012.
[51] R. Boero, M. Morini, M. Sonnesa, and P. Terna, Agent-based
models of the economy: from theories to applications. Palgrave
Macmillan, 2015.
[52] S. Hassan, J. Pav´on, L. Antunes, and N. Gilbert, "Injecting
data into agent-based simulation." in Simulating Interact-
ing Agents and Social Phenomena. Springer, 2010, pp. 177–
191.
[53] K. M. Carley, Validating computational models. Office of
Naval Research, 1996.
[54] D. Midgley, R. Marks,
and assurance
and D. Kunchamwar,
"Building
of
agent-based models:
An example and challenge to the field." Journal
of Business Research, vol. 60, no. 8, pp. 884–893,
2007. [Online]. Available: http://www.sciencedirect.com/
science/article/pii/S0148296307000458
[55] N. David, J. S. a. Sichman, and H. Coelho, "The logic
of the method of agent-based simulation in the social
sciences: Empirical and intentional adequacy of computer
programs." Journal of Artificial Societies and Social Simula-
tion, vol. 8, no. 4, 2005.
[56] T. Zhang, S. Gensler, and R. Garcia, "A Study of the
Diffusion of Alternative Fuel Vehicles: An Agent-Based
Modeling Approach*." Journal of Product
Innovation
Management, vol. 28, no. 2, pp. 152–168, 2011. [Online].
Available: http://onlinelibrary.wiley.com/doi/10.1111/j.
1540-5885.2011.00789.x/full
J. Giske,
[57] V. Grimm, U. Berger, F. Bastiansen, S. Eliassen,
J. Goss-Custard, T. Grand, S. K.
V. Ginot,
Heinz, G. Huse, and others, "A standard protocol for
describing individual-based and agent-based models."
Ecological modelling, vol. 198, no. 1, pp. 115–126,
2006. [Online]. Available: http://www.sciencedirect.com/
science/article/pii/S0304380006002043
[58] V. Grimm, U. Berger, D. L. DeAngelis, J. G. Polhill, J. Giske,
and S. F. Railsback, "The ODD protocol: a review and first
update." Ecological Modelling, vol. 221, no. 23, pp. 2760–
2768, 2010. [Online]. Available: http://www.sciencedirect.
com/science/article/pii/S030438001000414X
[59] A. B. Downey, Think Complexity: Complexity Science and
Computational Modeling, 1st ed. Sebastopol, CA: O'Reilly
Media, Mar. 2012.
[60] R. Arnott, "Economic theory and housing." in Handbook
of Regional and Urban Economics. Amsterdam: Elsevier
Science Publishers, 1987, pp. 959–988, 4.
[61] D. DiPasquale and W. C. Wheaton, Urban economics and
Englewood Cliffs, NJ: Prentice Hall,
real estate markets.
1996.
[62] V. G. Nadalin, "Tres ensaios sobre economia urbana e
mercado de habitac¸ ao em Sao Paulo." (tese de doutorado),
IPE-USP, Sao Paulo, 2010.
J. Brueckner, "The structure of urban equilibria: a unified
treatment of the Muth-Mills model." in Handbook of Re-
gional and Urban Economics.
Elsevier Science Publishers
B.V., 1987, pp. 821–845.
[63]
[64] A. S. Blinder, "Inventories and sticky prices." American
Economic Review, pp. 334–48, 1982.
[65] --, "On sticky prices: academic theories meet the real
The Univerisity of Chicago
world." in Monetary policy.
Press: N. Gregory Mankiw, 1994, pp. 117–154.
[66] B. P. Schettini, "An´alises da dinamica orc¸ament´aria dos
munic´ıpios brasileiros: uma aplicac¸ ao da metodologia
VAR com dados empilhados." Economia Aplicada, vol. 16,
no. 2, pp. 291–310, 2012.
[67] N. G. Mankiw, Principles of Economics, 6th ed.
Western College Pub, Jun. 2011.
South-
[68] M. Fujita, P. Krugman, and A. Venables, The spatial econ-
Cambridge,
omy: cities, regions and international trade.
Mass.: MIT Press, 1999.
[69] A. Losch, The economics of location. New Haven: Yale
University Press, 1954.
[71]
[70] R. Michaely and M. R. Roberts, "Corporate dividend
policies: Lessons from private firms." Review of Financial
Studies, vol. 25, no. 3, pp. 711–746, 2012.
in
J. W.
agent-based matching markets." Journal of Economic
Behavior & Organization, vol. 75, no. 2, pp. 168–
179, 2010. [Online]. Available: http://www.sciencedirect.
com/science/article/pii/S0167268110000661
"Stratification
Boudreau,
growth
and
[72] M. Neugart, M. Richiardi, and others, "Agent-based
models of the labor market." LABORatorio R. Revelli
20
working papers series, vol. 125, 2012. [Online]. Available:
http://www.laboratoriorevelli.it/ pdf/wp125.pdf
[73] D. Koesrindartoto, J. Sun, and L. Tesfatsion, "An agent-
based computational laboratory for testing the economic
reliability of wholesale power market designs." in Power
Engineering Society General Meeting, 2005. IEEE.
IEEE,
2005, pp. 2818–2823.
[74] L. Feng, B. Li, B. Podobnik, T. Preis, and H. E.
Stanley, "Linking agent-based models and stochastic
models of financial markets." Proceedings of the National
Academy of Sciences, vol. 109, no. 22, pp. 8388–8393, 2012.
[Online]. Available: http://www.pnas.org/content/109/
22/8388.short
[75] P. Seppecher, "Flexibility of wages and macroeconomic
instability in an agent-based computational model with
endogenous money." Macroeconomic Dynamics, vol. 16,
no. S2, pp. 284–297, 2012.
[Online]. Available: http:
//journals.cambridge.org/abstract S1365100511000447
[76] H. Dawid, S. Gemkow, P. Harting, S. Van der Hoog,
and M. Neugart, "Agent-based macroeconomic modeling
and policy analysis: the Eurace@ Unibi model." Bielefeld
Working Papers in Economics and Management, 2014.
[77] B. Edmonds and S. Moss, From KISS to KIDS an 'anti-
[78] R. G.
simplistic' modelling approach. Springer, 2005.
Ibbotson, L. B. Siegel,
of Portfolio Management, vol.
"World wealth: Market
Journal
pp.
//search.ebscohost.com/login.aspx?direct=true&db=
buh&AN=15197236&lang=pt-br&site=ehost-live
[Online].
values
1985.
4–23,
and
and K. S. Love,
returns,"
1,
http:
Available:
12, no.
Bernardo Alves Furtado Tenured re-
searcher at government
think-tank Insti-
tute for Applied Economic Research (Ipea-
Brazil), PhD from University of Utrecht and
UFMG. Background in urban analysis, eco-
nomics, and geography. Presently, holds a
position as a coordinator at DISET/IPEA
and a grant scholarship from CNPq. For-
mer deputy director, editor, professor. Inter-
ested in public policy applications of agent-based modeling.
Daniel
Rocha
Eberhardt
Isaque
Agronomist engineer
from Agronomist
School Eliseu Maciel of Federal University
of Pelotas - UFPel; Master Degree in
Remote Sensing by National Institute for
Space Research - INPE; PhD Candidate
in Transportation by the University of
Bras´ılia - UNB; and researcher assistant
at Institute for Applied Economic Research
- IPEA. Main subjects are Remote Sensing, Geographical
Information System, Spatial Analysis, R language, Data Mining,
Data Analysis, Spatio-Temporal Analysis.
21
APPENDIX A
PSEUDOCODE: REAL ESTATE MARKET
Symbol "#" indicates comments about the procedures or functions.
Function Real Estate market (families, dwellings, government regions)
# Select a sample of families to enter the market, given the parameter
Generate list of families on the market, chosen randomly
Create a market dictionary
# Select vacant houses; update prices and quality for all houses, given QLI in each region
For each dwelling:
Update prices, given the region
Update Quality of Life Index
If dwelling is vacant
Add it to market dictionary
Calculate median of families' financial resources
For each family that are in the market
Update family's dwelling value
If the family has a positive number of members
Set move equal to False
If the family's financial resources are below median value
# Compare all available dwellings, choose the cheapest
Choose a random dwelling (1)
For each dwelling:
If dwelling's (2) price is lower than dwelling's (1)
# Check the difference between values
If current dwelling is more expensive than the new dwelling:
Else
Choose dwelling (2)
Choose dwelling (1)
# Deduce the difference
Calculate the difference
Update families' values
set move to True
Else
# Compare the available dwellings available, chose the best
Chose random dwelling (1)
For each dwelling
If dwelling's quality (2) higher than (1)
Else
Choose dwelling (2)
Choose dwelling (1)
# Check fund availability
If current dwelling price is higher than intended dwelling
# Deduce the difference
Calculate the difference
Update families' dwelling values
Set move True
# Making the move
22
If move is True
Families vacant old dwelling
# New residence
Family register new dwelling, address, value
Dwelling register new family
APPENDIX B
PSEUDOCODE: GOODS MARKET
Function Equalize families' funds (families)
Distribute equally total resources among family members
Function Consumption (firms, regions), processed within agent's class
If monetary available quantity is positive
If resources are lower than unit
Consume resources in a random value from 0 to total
Else
Consumption equals random value between zero and total,
discounted by a β parameter
# Given the size of market, decide among firms,
choosing lowest price or closest firm
Create an empty market list
Add firms randomly, given market size
# Choose the firm with lowest price
Create an empty prices' list
For each firm on selected market
Add prices to the list
Select firm listed with lowest price
# Choosing closest firm
Create an empty distance list
For each firm on market
Calculate the distance, given consumer's location
Add distance of each firm to the list
Select from the list, closest firm
# Choose randomly between closest firm and lowest prices
Choose firm
# Purchase from chosen firm
Chosen firm processes sales' Function,
inputs (amount available to consumption, firm; region); return change
Consumer updates cash values, considering spent money and change
# Utility
Update consumer's utility from consumed value
Function Sales (consumer's resources, regions), processed within firm's environment
If resources are positive
For each product in inventory
If quantity is positive
Purchased quantity equals available resources divided by price
# Verifying if available quantity is enough
23
If demanded quantity is higher than offered quantity
Calculate amount spent, given quantity and prices
Deduce quantity from firms' inventory
# Taxes and balance
Deduce spent amount from agent
Pay firms, deducing owed taxes, given rates
Add collected taxes to each region
# Quantity sold
Generates statistical information
Return change, if necessary
APPENDIX C
PSEUDOCODE: LABOR MARKET
class Matching
Function Firms offer position
Create dictionary for firms offering positions
Function Candidates application
Create dictionary with agents that fill prerequisites
Function Matching
# Rank candidates by qualification
# Matching
While both candidates' list and firms' list contain elements
Choose random firm (1) among those on dictionary
# Higher qualification of candidate
Choose random candidate (3)
For each candidate
If candidate qualification (4) is higher than candidate (3)
Select candidate (4)
Else
Select candidate (3)
# Firm's closest candidate
Choose random candidate (5)
For each candidate (6)
For each firm (7)
If firm's distance is lowest
Select candidate (5)
Else
Select candidate (6)
# Choose randomly between best candidate or the one who lives closest
Choose candidate
Firm hires, candidate register job
# Delete candidate and firm from the dictionary
Delete firm that hired from dictionary
Delete chosen candidate from dictionary
APPENDIX D
ODD PROTOCOL
The model description follows
the ODD
(Overview, Design concepts, Details) protocol
for describing individual
and agent-based
models [57], [58].
1) Purpose
2) Entities, state variables, and scales
The model is a first exercise to observe
the economy in its full spatially explicit
environment,
its markets, and main
agents in order to capture taxes mech-
anisms and their effects as a means
to enable public policy evaluation ex-
ante. The model falls within the simple
category and it is a loosely expansion
of Lengnick [34] and Gaffeo [35] with
the introduction of spatially bounded
government regions. Proven its valid-
ity, Brazilian intricate tax system can
be inserted into the model and sub-
national development in the medium
term could be analyzed. The main hy-
pothesis is that for a given labor mar-
ket area where citizens commute daily;
making changes to political administra-
tive boundary impacts citizens quality
of life. The results of the model suggest
that it may be the case.
The model contains classes of agents,
families, dwellings, firms, and govern-
ment, along with accessory classes. IDs.
All members of all classes have their
own unique identification (IDs). Agents
keep their current workplace ID, fam-
ily and dwelling IDs. Families keep
track of all their members and current
dwelling. Dwellings knows which fam-
ily is currently hosting,
if any, their
government region ID. Firms knows all
their employers IDs and its government
region ID. Government is passive and
money is transferred directly from con-
sumers at the buying moment. Each
government region has its own ID.
The agents have age (in years), qual-
ification (in years of study), utility and
money attributes along with family and
24
dwelling identification. They also have
processes to update money balance and
consumption procedures. Agents buy
and work in firms, are members of fam-
ilies and can move among dwellings
along with their families.
Families are groups of agents. Con-
sumption money is always equally di-
vided among family members. Families
move among dwellings. They register
the current dwelling address and value.
Dwellings are fixed in space, have
prices (proportional to size and Quality
Index of the government region), size
and quality and addresses. Quality and
prices are updated monthly, given that
the Quality Index at the government
region has been updated.
Governments are within sets of pre-
defined boundaries. They have a Qual-
ity Index and collect taxes deducing
amount from firms when selling to
costumers within their territory. They
invest all treasure into Quality Index
updates.
Firms have one product (with price and
quantity), monthly balance, profits, ad-
dresses (x, y), and constantly knows its
employees. They are fixed in space and
process sales, production (with product
quantity update), hire and fire deci-
sions, and make employers payment.
Temporal extent. The model was de-
signed to run for 20 years, but there is
no strict limit applied. It runs in terms
of days, months (of 21 days), quarters
and years, following Lengnick [34]. In
this configuration 5,040 days represent
20 years.
Spatial extent. The spatial boundaries
are determined exogenously through
parameters. We used -10, 10 on hori-
zontal and vertical axis with location
within the square in float precision.
Along with the main classes, accessory
classes include a system of communica-
tion that holds the labor market proce-
dures; statistics, output, plotting, main,
control (that iterates over simulations),
generator that creates the instances of
agents, families, dwellings and firms
before simulation; parameters; space
and time iteration. A products class
was also built to facilitate new products
development. However, in the current
model only one product is in effect.
3) Process overview and scheduling
4) Design concepts
The model runs in a discrete mode
fully using the Object-Oriented Pro-
gramming (OOP) paradigm. A time
schedule procedure is described in sec-
tion 2. The pseudocode of the main
processes are also available as Annexes
section A, section B, and section C.
The model was designed for a Python
3.X environment. Thus, it makes full
use of classes, their variables and meth-
ods with variable updating being pro-
cessed when methods are called. In
order to run one simulation step, in-
stances of the classes are created to
make sure that objects are updated
correctly. According the description of
item 2, we have the following instances:
my agents, my regions, my houses,
my firms, my journal
(communica-
tions), my parameters, my simulation
(TimeControl). R is used to process the
results and produce plots.
Basic principles. The model is an ex-
tension of some principles from Leng-
nick [34] and Gaffeo et al. [35] with
many adaptations. Given the lack of
simple models [30], some processes
are new implementations. Mainly these
processes are the markets of goods, la-
bor and houses and the decision pro-
cesses of firing and hiring, setting of
goods prices and wages, following ear-
lier literature [44], [64], [65], [70].
Prices. Firms
rationalize on prices
given their current stock. When quan-
tity produced is below a given param-
eter threshold, firms raise prices by
some parameter percentage, i.e. price is
given by cost but when there is enough
demand a mark-up value is added
25
to price. Otherwise, when quantity is
above that same parameter threshold
price is given by cost.
Wages. Firms pay a constant wage plus
a qualification additional.
Goods market. Firms offer their prod-
ucts with given prices. Consumers
choose from a subsample of firms;
which is determined by a parameter
of
the model. When considering to
purchase, consumers randomly decides
from firms that either have the mini-
mum price [67] or is the closest from
its residence [68], [69].
Labor market. Candidates of age with
a given qualification offer themselves
repeatedly on the market. Firms offer
posts. Matching happens between most
qualified employee and either (ran-
domly) or closest firm.
Hiring and firing. Firms decide on hir-
ing or firing periodically, given an ex-
ogenous parameter. Typically, once ev-
ery four months. When profit or cash-
flow is positive or the firm has no
employees, they offer one post. When
profit
they fire an em-
ployee. Hiring follows the process de-
scribed in the labor market. The em-
ployee to be fired is chosen randomly
from the pool of employees of the firm.
Housing market. Families enter the
housing market periodically, following
an exogenous parameter [78]. Once in
the market, their decision to move is
for more quality, when families have
wealth above the median of all families;
or to move for cheaper houses and cap-
italizing on the difference, following a
general model of urban economics [63],
[61]. In the first case, for each family in
search, the available houses are ranked
in order of quality. If the familys wealth
plus the value of the current house is
enough to best house, the transaction
and the move occurs. In the second
case, each family goes for the cheapest
house.
Emergence. As output of the model,
is negative,
typical indicators of the economy are
produced. We believe many of the re-
sults to be endogenous and robust on
the parameters of the model. Built-in
the model
is the idea of economies
of agglomeration and disagglomera-
tion [68] and urban economics in
general[63], [61]. That is, the center (or
the region) where there is a concentra-
tion of firms that are performing well
tend to attract agents. Simultaneously,
rent prices rise and such process expels
families to poor suburbs. However, the
central region is an endogenous result
of the model.
Adaptation and Learning. Agents,
families or firms do not adapt in the
sense that they change the process of
decision-making. However, the loose-
ness of families in respect to dwelling
attachment along with the dynamics
of price changing implies that families
have to financially adapt to constant
changing environment.
Objectives. Families' objective is either
try to move to better quality places or
have enough resources to keep con-
suming. Their success rate is measured
by their members utility, which is an
indicator that cumulatively measures
their actual consumption. Firms objec-
tive is to increase in size and keep on
hiring and producing more and more.
However, when out of equilibrium,
objective changes to restore financial
health (firing employees). Government
(implicit) objective is to increase Qual-
ity Index.
Prediction. Decisions of all agents are
based on cross-sectional
information
try to infer the future.
and do not
The model as a whole is intended as
a comprehension mechanism and as
a provider to the specific questions
posed.
Sensing. Sensing is global for agents
when looking for jobs. When consum-
ing, sensing is restricted to the given
number of firms defined (parameter) as
26
this version of
its private market. Dwelling price and
quality changing mechanism is propor-
tional to the observed at their region;
which in turn is proportional to the
sales of firms acting in the region.
Interaction. Interaction happens com-
petitively at the three designed models.
However, agents and firms interact di-
rectly only in the sense that they may
be excluded from the selection process
(of hiring or consuming), given that
they have lower qualification, are offer-
ing lower salaries or expensive prod-
ucts. Implicitly, agents at the same re-
gion share the same Quality Index.
Thus, families benefit from high quality
dwellings that is given by profitable
firms in the region.
Stochasticity. Stochasticity plays an im-
portant role at
the
model. All population of agents, fam-
ilies, firms, dwellings is generated from
a random process. Random decisions
between two alternatives occur when
deciding either for closest or cheap-
est product and for most qualified or
closest living employee. Further, there
is a random process when the agent
decides the amount for consumption
monthly and when the firm makes its
firing decision.
Collectives. The only collective in the
present model are the families. They
have been described as a class above.
Observation. A number of statistics are
collected on a monthly basis. However,
they do not interfere endogenously, ex-
cept for the average families wealth
(interferes on moving decision). The in-
dicators available are absolute produc-
tion sold (GDP), unemployment, aver-
age number of employees per firm, av-
erage utility of agents, average prices,
average firms' balance, sum of firms'
profit, Gini inequality index (based on
families' average utility). Firms indi-
vidually calculate and use their profits
endogenously. Every three months they
record their total balance and then they
27
calculate next months profit in relation
to the recorded value.
Parameters. The following parameters
are requested from the modeler at every
simulation (section 2).
Iteration. Given the artificialness of the
population and the stochasticity de-
scribed, the model was run 1,000 times
for each regional configuration (one,
four or seven regions) and the results
are presented in terms of distributional
statistics.
Initialization
At time 0 of the simulation, a number
of processes has already run and will
not run again (section 2). Given that
this model is an artificial test aimed at
scrutinizing (and proposing) the model
itself and allowing only a hinted indica-
tion of public policy, the population is
always a different one for each different
run, given the parameters discussed.
It is our plan to apply the model to
a metropolitan area with fixed given
population.
Input data
The model does not use input data, as
it is.
Submodels
There are no submodels in this version.
5)
6)
7)
|
1807.07957 | 1 | 1807 | 2018-07-20T17:59:58 | Decentralized Task Allocation in Multi-Robot Systems via Bipartite Graph Matching Augmented with Fuzzy Clustering | [
"cs.MA",
"cs.AI",
"cs.RO"
] | Robotic systems, working together as a team, are becoming valuable players in different real-world applications, from disaster response to warehouse fulfillment services. Centralized solutions for coordinating multi-robot teams often suffer from poor scalability and vulnerability to communication disruptions. This paper develops a decentralized multi-agent task allocation (Dec-MATA) algorithm for multi-robot applications. The task planning problem is posed as a maximum-weighted matching of a bipartite graph, the solution of which using the blossom algorithm allows each robot to autonomously identify the optimal sequence of tasks it should undertake. The graph weights are determined based on a soft clustering process, which also plays a problem decomposition role seeking to reduce the complexity of the individual-agents' task assignment problems. To evaluate the new Dec-MATA algorithm, a series of case studies (of varying complexity) are performed, with tasks being distributed randomly over an observable 2D environment. A centralized approach, based on a state-of-the-art MILP formulation of the multi-Traveling Salesman problem is used for comparative analysis. While getting within 7-28% of the optimal cost obtained by the centralized algorithm, the Dec-MATA algorithm is found to be 1-3 orders of magnitude faster and minimally sensitive to task-to-robot ratios, unlike the centralized algorithm. | cs.MA | cs | Decentralized Task Allocation in Multi-Robot
Systems via Bipartite Graph Matching Augmented
with Fuzzy Clustering
Payam Ghassemi∗, and Souma Chowdhury†
Department of Mechanical and Aerospace Engineering
University at Buffalo
Buffalo, NY, 14260
Email: ∗[email protected], †[email protected]
8
1
0
2
l
u
J
0
2
]
A
M
.
s
c
[
1
v
7
5
9
7
0
.
7
0
8
1
:
v
i
X
r
a
Abstract-Robotic systems, working together as a team, are be-
coming valuable players in different real-world applications, from
disaster response to warehouse fulfillment services. Centralized
solutions for coordinating multi-robot teams often suffer from
poor scalability and vulnerability to communication disruptions.
This paper develops a decentralized multi-agent task allocation
(Dec-MATA) algorithm for multi-robot applications. The task
planning problem is posed as a maximum-weighted matching
of a bipartite graph, the solution of which using the blossom
algorithm allows each robot to autonomously identify the optimal
sequence of tasks it should undertake. The graph weights are
determined based on a soft clustering process, which also plays
a problem decomposition role seeking to reduce the complexity
of the individual-agents' task assignment problems. To evaluate
the new Dec-MATA algorithm, a series of case studies (of
varying complexity) are performed, with tasks being distributed
randomly over an observable 2D environment. A centralized
approach, based on a state-of-the-art MILP formulation of
the multi-Traveling Salesman problem is used for comparative
analysis. While getting within 7-28% of the optimal cost obtained
by the centralized algorithm, the Dec-MATA algorithm is found
to be 1-3 orders of magnitude faster and minimally sensitive to
task-to-robot ratios, unlike the centralized algorithm.
Index Terms-Bipartite Graph Matching, Decentralized,
Fuzzy Clustering, Ground Robots, Multi-agent task allocation,
Multiple-Traveling Salesmen Problem (mTSP)
I. INTRODUCTION
A. Multi-agent Task Allocation
In contrast to the planning of actions for a single agent,
Multi-Agent Task Allocation (MATA) is concerned with the
allocation of tasks and resources among several agents that
act together in the same environment to accomplish a com-
mon mission. In a MATA environment, there may be either
conflicting or common goals pertaining to the different agents'
decisions, which need to be coordinated [1]. While centralized
formulation and solution of MATA problems have traditionally
dominated the fields of robotics, transportation, and IoT [2],
Copyright©2018 ASME. Personal use of this material is permitted. Per-
mission from ASME must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or
promotional purposes, creating new collective works, for resale or redistribu-
tion to servers or lists, or reuse of any copyrighted component of this work
in other works.
[3], decentralized methods have gained increasing prominence
in recent years. This is partly due to increasing concerns
regarding the scalability of purely centralized approaches and
their vulnerability to communication disruptions, and partly
driven by accelerated advancements in system/agent autonomy
and Artificial Intelligence (AI) capabilities. In this paper, we
consider task allocation for a team of robots in 2D space,
and formulate the MATA problem as finding a set of optimal
routes that minimize the overall costs incurred by the team [4].
Both a baseline centralized solution approach (based on multi-
traveling salesman or mTSP problem) and a new decentralized
framework (that combines bipartite graph theory and clustering
concepts) are developed in this paper, and compared over a set
of case studies involving varying numbers of robots and tasks.
The remainder of this section briefly surveys the literature
on centralized mTSP and auction-based (semi)-decentralized
methods, and converges on the objectives of this paper.
B. Multi-Traveling Salesman Perspective of MATA
The MATA problem, when defined by location-based tasks
distributed over 2D space, is analogical to the Multi-Traveling
Salesmen Problem (mTSP) [4], [5] – a generalization of the
Traveling Salesman Problem (TSP) [6], [7]. In mTSP, there
are multiple salesmen or agents that need to visit a set of
geographically distributed cities or locations, such that no city
is visited more than once collectively by the agents. There are
a number of variations on mTSP and we consider single depot
mTSP, which is defined as [5]: given a set of cities, one depot,
and a cost metric, the objective is to find a set of routes for m
salesmen such that it minimizes the total cost of the m routes.
All salesmen are located in the depot and they return to the
depot. In the literature, a number of approaches have been
proposed for finding optimal solutions for the mTSP. Most of
the approaches can be classified into two main classes [5]:
1) algorithms based on Mixed-Integer Linear Programming
(MILP) formulations, and 2) algorithms that transform the
mTSP into the TSP. For solving the problem in each class,
there are both exact solution and heuristic methods [5].
Although one of the earliest direct approaches for solving
the mTSP was proposed by Laporte and Nobert [8], a branch-
and-bound method proposed by Gavish and Srikanth [9] can be
considered to be one of the first attempt to solve a large-scale
mTSP. Gavish and Srikanth [9] defined a Lagrangian problem
for computing the lower bound of the main branch-and-bound
algorithm. This method is limited to the symmetric mTSP.
Gromicho et al. [10] later solved the mTSP using a similar
branch-and-bound algorithm but with a different approach to
obtain the lower bound, namely a quasi-assignment relaxation
instead of the Lagrangian formulation. This method solves the
asymmetric mTSP with a fixed number of salesmen. There are
also heuristic-based methods to solve the mTSP, such as by
Zhang et al. [11] and Ryan et al. [12], who respectively use
Genetic Algorithms (GA) and Tabu search. Recently, Kalmar
et al. [4] proposed a modified GA method to solve the mTSP.
These methods are slow to converge to optimal solutions, and
often not suited for near-real-time applications (such as multi-
robot real-time task-planning).
The second class of the approaches, rather than solving the
mTSP, investigate different strategies to transform the mTSP
into a standard TSP. Gorenstein [13] solved the mTSP with
m salesmen as a standard TSP with (m − 1) additional home
cities, where infinite costs are assigned to home-to-home dis-
tances and zero costs are assigned between the additional home
cities and other cities. Jonker and Volgenant [14] improved the
transformation of the mTSP to a standard TSP with a sparser
edge configuration. This idea reduced the computational cost
of the branch-and-bound scheme and created a less-degenerate
TSP.
The aforementioned approaches are all centralized algo-
rithms, typically characterized by exponentially growing com-
putational effort [15], as the number of agents and tasks
increase. Moreover, if the MATA problem entails cost models
that are highly nonlinear functions of the inter-city connecting
edges, MILP formulations could become unreliable. Decen-
tralized algorithms can, in principle, provide more tractable
and dependable (if not provably optimal) solutions
[16].
Most of the existing decentralized solutions are consensus-
based auction methods, where agents place bids on services
or resources [17]. These approaches are not necessarily fully
decentralized as they rely on a centralized auctioneer. In
order to address this issue, Dai and Chen [18] proposed a
decentralized auction-based approach, where each agent acts
as auctioneer and bidder to resolve any potential conflicts.
In these approaches, multiple (thus time-consuming) biddings
are necessary to make a decision, which also increases the
inter-robot communication burden. However, consensus-based
method do provide decision robustness under uncertainties,
such as partial observance of mutual states within the robot
team [19]; further discussion of these characteristics is how-
ever not within the scope of this paper. A comprehensive
review of methods on consensus-based decentralized auctions
for task allocation can be found in [19].
C. Objectives of this Paper
Given the (above-described) characteristics of existing
there remains an important opportunity
MATA methods,
for developing computationally-efficient and communication-
frugal decentralized approaches to task allocation in multi-
robot teams. While additional realistic considerations, such
as partial observability and agent heterogeneity, can further
complicate this problem, in this paper we focus on solving
the efficient decentralized task planning problem subject to
the following assumptions: i) all robots/agents are identical;
ii) costs of all tasks can be evaluated deterministically; iii)
there are no environmental uncertainties; and iv) each agent
has full observation of all tasks and the state of other agents.
Warehouse robotics [20], such as Amazon's Kiva system [21],
can be considered to be a close match for an example, where
the decentralized task allocation problem could be practically
solved under these assumptions.
While the long-term goal of this research is to develop
stochastic learning-based algorithms that are capable of near-
instantaneous decision-making and are insensitive to these
assumptions, as a first step here we propose to develop a three-
stage deterministic framework to decentralized planning. The
first and second stages would seek to reduce the computational
complexity of the problem being solved individually by each
agent and identify a compact representation of the reduced
MATA problem; and the third stage would allow each agent
to identify optimal task sequences to undertake. The novel con-
tribution of this paper lies in combining fuzzy clustering and
bipartite graph matching to design this three-stage framework.
Thus the objective of this paper is to develop this "clustering–
graph matching" based MATA framework, and investigate its
performance in comparison to a state-of-the-art centralized
mTSP implementation.
The remaining portion of the paper is organized as follows:
The next section presents the problem definition and central-
ized mTSP formulation. Section III describes our proposed
decentralized MATA framework. Results, encapsulating the
performance of these methods on different-sized problems,
are presented in Section IV. The paper ends with concluding
remarks.
II. COLLABORATIVE MULTI-AGENT
FORMULATION
A. MATA: Defining Problem Components
task allocation (MATA) problem is de-
The multi-agent
tasks and resources among
fined as
in the
several agents that act
same environment to accomplish a common mission. Each
agent (robot) shares its state and its view of the world
with other agents;
the MATA is a tuple
T =< R,{Si}, h, J,{Ai},{Ci}, M, G > where:
together without conflict
the allocation of
in other words,
• R = 1, . . . , m is a finite non-empty set of agents.
• Si is a set of state variables that represent the state of
agent i – for example, its current location and battery
state or traveled distance. Each agent can share their state
variables with the other agents; i.e., full observability is
assumed in the preliminary implementation presented in
this paper.
• h is an integer value stating the maximum number of
tasks that any robot can undertake.
each robot is allowed to take.
• J = 1, . . . , n is a finite non-empty set of active tasks that
• Ai : J × h is a set of decisions of agent i at iteration k,
• Ci : Ai × J is a finite set of state variables describing the
i.e., Ak
i .
relationship of each task j with respect to agent i.
• M is a decision function that maps Si, Ci to Ak
i at iteration
k.
• G is a metric that evaluates the total cost of the mission.
While the above definitions provide generic description of the
problem components, in programmed practice most of these
components are represented as pertinent vectors and matrices.
total cost of the mission G and satisfies (cid:83)m
The MATA problem is defined as finding the decision function
(cid:84)m
M that generates the decision set A∗
i minimizes the
i=1 A∗
i = J and
i = ∅. Next, we describe a modified mTSP formulation
i=1 A∗
that will be serving as the benchmark centralized solution
approach.
i , where A∗
B. Centralized mTSP Formulation
The centralized MATA problem is formulated as a Mixed-
Integer Linear Programming (MILP) problem [5], as given
in Eq. (1) to (6). The decision-space comprises a binary
decision variable zij and an integer variable ui. The variable
zij becomes 1 if any robot takes task j after finishing task i,
and becomes 0 otherwise. The variable uk shows the position
of task k in a sequence (tour). In Eq. (1) to (6), m and h
respectively represent the number of salesmen (robots) and the
maximum number of tasks each robot can take in one tour;
cij is the cost metric for taking task j after performing task
i; it can essentially represent monetary cost, distance, time, or
other cost function based on application.
n(cid:88)
n(cid:88)
i=1
j=1
min
zij ,ui
f =
cijzij
subject to
j=2
n(cid:88)
n(cid:88)
n(cid:88)
n(cid:88)
i=1
i=2
z1j = m
zi1 = m
zij = 1;
∀ j = 2, . . . , n
zij = 1;
∀ i = 2, . . . , n
j=1
ui − uj + h · zij ≤ h − 1;
∀ 2 ≤ i (cid:54)= j ≤ n
(1)
(2)
(3)
(4)
(5)
(6)
In the above set of equations, constraints (2) and (3) ensure
that exactly m agents depart from and return back to the
depot (node 1). Constraints (4) and (5) imply that each
task is taken and finished by only one unique robot in the
n(cid:88)
n(cid:88)
team. Constraint
(6), also known as Sub-tour Elimination
Constraints (SCE), is utilized to remove sub-tours (circular
closed routes not containing the depot). There are different
mathematical formulations for the sub-tour elimination; in this
work, we utilize one of the well-known formulations, called
the Miller-Tucker-Zemlin constraints [22]. Given this mTSP
formulation, the overall cost incurred by the robot team can
be estimated as:
ctotal =
cijzij
(7)
i=1
j=1
III. DECENTRALIZED MATA ALGORITHM
A. Decentralized MATA: Overview
Figure 1 depicts the flowchart of our proposed decentralized
MATA or Dec-MATA algorithm. We pose the decentralized
MATA as a maximum-weighted matching of a bipartite graph.
Our MATA algorithm is composed of three components: 1)
performing soft clustering to divide the tasks into groups
(based on task location) to allow task-assignment problem
decomposition; 2) use the clustering information and agents'
states to transform the problem at each decision-making step
into a weighted bipartite graph; and 3) solving a matching
problem for the bipartite graph to assign a sequence of tasks
to individual robots. Component 2 is generally the most chal-
lenging to accomplish. These three components are described
next.
B. Task Clustering
The purpose of clustering and graph construction is to
identify and systematically represent the degree of relevance
of groups of tasks to particular robots, given their current state
– thereby reducing the size of the task assignment problem to
be solved by each robot. In this paper, relevance is judged
based on the current location of tasks and robots. However,
the underlying concept can be extended to apply to more
generalized scenarios where the state space of the robots and
the feature space of the tasks are more complicated. There
is also an implicit assumption here that the number of tasks
is always greater than the number of robots, and hence our
aim is to identify task relevance not individually but in a
group. In order to group the tasks according to their relative
location, while allowing relevance to be defined in a smooth
(as opposed to a binary, relevant/not relevant) manner, the
Fuzzy C-Means clustering algorithm (FCM) [23] is used. FCM
is a soft clustering method, which allows cluster overlap and
is performed by solving the following minimization [24]:
ij ∈ arg min
b∗
bij
ij(cid:107)xj − ri(cid:107)
bγ
(8)
In Eq. 8, n and m are the number of data points (in this
case, number of tasks) and the number of clusters (in this
case, number of robots), respectively; γ > 1 is a user-defined
coefficient that adjusts the degree of fuzzy overlap; xi is the
i-th data point or task, and ri is the location of the i-th robot,
m(cid:88)
n(cid:88)
i=1
j=1
Fig. 2. A representative weighted bipartite graph with 2 nodes in the robots
set (Kiva robots) and 3 nodes in the tasks set (shelves).
in Fig. 2. This current bipartite graph definition (allowing task-
robot one-to-one mapping) is applicable to MATA problems,
where multiple agent collaboration on any single task is not
required (rather prohibited).
In order to fully construct
the representative weighted
bipartite graph, we should determine the weights of edges, a
typically challenging endeavor given the lack of any standard
recommendations to this end. Here, the edge weights can be
perceived as the strength of the potential relationship between
the corresponding task and robot, and thus indicative of the
degree of relevance of task j to robot i. The weight, wij, of
the bigraph edge ij is defined here in terms of the states of
robot i and task j as:
wij = Kbb∗
ij
cc,i + 1
(cij + 1)2
(9)
the parameter b∗
Here,
ij, which relates task-j to agent-i,
is obtained from fuzzy clustering (Eq. 8). The robots are
assumed to start from a single depot (e.g., co-located charging
stations). Hence, initial task assignment (considered as the
zeroth decision-making iteration) is randomized. However,
then onwards, each robot assumes the location of the current
task (and hence membership of its cluster) that it is under-
taking, allowing us to use the b∗
ij relation (which previously
represented task-cluster relatedness). Here, Kb is a coefficient
to adjust b∗
ij with respect to the magnitude of cij, where
cij shows the cost of doing task j (next) for robot i. The
parameter cc,i is the cumulative cost incurred by robot i up
to that decision point. Note that the bigraph reduces by m
tasks, as tasks get done by the m robots in each iteration.
The depot (denoted as task 0) becomes available for selection
iteration 2 onwards. Figure 3 depicts the weighted bigraphs
generated at the first and second decision-making iteration for
a representative 4 robot – 12 task case.
Fig. 1. Flowchart of the new decentralized MATA algorithm
serving as the associative (not necessarily exact geometric)
center of the i-th cluster. Here, bij is the degree of membership
of xj in the i-th cluster and ranging from 0 to 1. For ease of
implementation, in this paper, the number of fuzzy clusters is
set equal to the number of robots. As shown in the algorithm
flowchart (Fig. 1), the clustering step needs to be repeated only
if new tasks get added to the mission.
C. Weighted Bipartite Graph Transformation
In order to represent and analyze the task-robot relations, we
use the concept of bipartite graphs, so-called bigraphs (pop-
ularly used in recommender systems [26] and social network
analysis [27]). A bigraph is a graph whose vertices can be
divided into two sets of vertices such that no two vertices
in the same set are joined by an edge [28]. In this paper, we
define our problem as a weighted bigraph (R, J, E) during each
decision time-period, where R and J are a set of robots (agents)
and a set of tasks, respectively; and E represent a set of
weighted edges that connect robots to available tasks, as shown
(a) First decision-making iteration
(b) Second decision-making iteration
Fig. 3. Bigraph generated by the dec-MATA algorithm for a representative 4 robots / 12 tasks case (i.e., m = 4, n = 12): black and gray circles represent
robots and tasks, respectively; the color of the edges represent the estimated weights of the connections between robots and tasks. The bigraph can be observed
to vary from one decision-making iteration to another; since the robots collectively chose tasks 2, 5, 8 and 9 in the first iteration, these tasks are missing
from the reduced bigraph constructed in the second iteration. However, the latter bigraph has one more task – task 0 basically represents the depot, which
become available for selection only after the first iteration; this is done to ensure that each robot undertakes at least 1 task.
D. Maximum Weighted Matching
Once the weighted bipartite graph has been constructed,
the final step (Fig. 1) is to solve the task assignment or
allocation problem as a matching problem. This problem is
defined as drawing a set of edges such that they do not share
any vertices [29]. A weighted maximum matching method is
used here to determine the optimal task assignment. Note that,
until this point, all robots have taken the same (redundant)
computational steps; however, this final step differs across
robots, since each robot operates only on the portion of the
bipartite graph relevant to itself (e.g., not consider tasks for
which its bigraph connecting edge has a weight of zero).
To perform this final step, an improved maximum matching
algorithm proposed by Galil [30] is adopted. This algorithm
is based on the classical blossom algorithm introduced by
Edmonds [31], which is known to run in polynomial time.
The pseudocode of this algorithm is given in Algorithm 1. This
algorithm produces the optimal decision function, M (i.e. the
task assignment set). It is important to note that the outcomes
of this decentralized decision-making process are deterministic
and synchronized (assuming a deterministic environment with
perfect
localization). Consequently, each robot will arrive
at the same conflict-free optimal task assignment outcomes.
Although, currently, a degree of redundancy of taking the same
decision on-board each robot (w.r.t. the same task cluster) is
allowed, more efficient implementations can be explored in
the future. These future advancements could also help promote
asynchronous multi-agent decision-making, which is a highly-
challenging and an open area of research in itself [32].
IV. NUMERICAL EXPERIMENTS
A. Problem and Framework Settings
We design and execute a set of numerical experiments to
investigate the performance and scalability of the proposed
decentralized MATA algorithm (Section III), and compare it
with the centralized mTSP solution (Section II-B). For solving
the MILP problem resulting from the centralized approach (1),
the"Gurobi V8.0" [33] is run on the MATLAB® R2016b.
A combination of MATLAB libraries (e.g., fuzzy c-means
clustering), Python libraries and original codes are used to
implement our decentralized MATA framework. Specifically,
the "Python" 3.6.0 and the 64-bit distribution of "Anaconda"
4.3.0 are used to implement the decentralized MATA frame-
work, and the "networkx" library is used for graph analysis and
visualization. The simulations are executed on a system with
Intel® i7-6820HQ 2.70 GHz 4 Cores (8 Logical Processors)
CPU and 16 GB RAM. The Gurobi (MILP) solver exploits all
8 logical processors, while the Python implementation of Dec-
MATA does not exploit all the cores. Here the parameters of
the Gurobi's solver are set as: if the computation time is more
than 7200s, the absolute MIP optimality gap ("MIPGapAbs")
equals 1e − 10, otherwise this gap, MIPGapAbs, is set at 6.
Research data and codes related to Dec-MATA implementation
can be found at: http://adams.eng.buffalo.edu/algorithms/multi-robot-algorithms/.
The prescribed parameter, Kb, is set at 1000. We define 7
different case studies, corresponding to different combinations
of numbers of ground robots and tasks; allowed maximum
tasks per robot is kept at p = (cid:98)n/m + 2(cid:101) to promote equitable
distribution of load across the team. These test cases are listed
in Table I. The task locations are randomly distributed in a
10 × 10 sized 2D space. In order to provide a statistically
insightful evaluation and comparison, ten random test scenar-
ios are generated for each case. As the dimensionality of the
robot-task-space increases, so does the computational cost of
clustering. Hence, it is important to explore the impact and
cost of clustering, or more specifically how the algorithm
performs in the presence and in the absence of clustering.
To this end, the following three different implementations of
123412345678910111212341234567891011122040608010012341346710111201234134671011120200400600800100012001400 M ← ∅
do
P ← BLOSSOMALG(G, M)
M ← M ∆P
Algorithm 1 Blossom algorithm (adopted from [25])
1: procedure MATCHGRAPH(G)
2:
3:
4:
5:
6:
7:
8: procedure BLOSSOMALG(G, M)
9:
10:
11:
while P (cid:54)= ∅
return M
G(cid:48) ← G
for all v ∈ V do
for all v ∈ V : v (cid:54)∈ VM , marked(v) = false do
marked(v) ← false
marked(v) ← true
Q ← v, Qn ← ∅
T ← v
while Q (cid:54)= ∅ do
for all u ∈ Q do
for all w ∈ N(cid:48)(u) do
if w (cid:54)∈ VM then
return shortestPath(T, u, v)+(u, w)
else if marked(w) = false then
x ← y ∈ N (w) : (y, w) ∈ EM
marked(w), marked(x) ← true
T ← T + (u, w) + (w, x)
Qn ← x
(cid:46) Graph G(V, E)
(cid:46) Match M initially empty
(cid:46) New augmenting path found with M, G
(cid:46) Symmetric difference between M, h
(cid:46) Graph G(V, E) and matching M (VM , EM )
(cid:46) Modified BFS tree
(cid:46) w is unsaturated
(cid:46) Not in T yet
(cid:46) Vertex w's match
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
blossom found
return ∅
else if abs(level(T, u)-level(T, w)) mod 2 = even then (cid:46) Even level difference on non-tree edge means
B ← shortestPath(T, u, v)+(u, w)
G(cid:48) ← G(cid:48) · B
T ← T · B
b ← contracted blossom vertex
marked(b) ← true
Qn ← b
(cid:46) Contraction of blossom
(a) 4 robots and 12 tasks
(b) 8 robots and 24 tasks
Fig. 4. Clusters given by FCM for the test cases 2 and 3 (Table I). The size of markers is proportional to the corresponding cluster coefficient of each task
w.r.t. the associated cluster. The left plot shows that two tasks at point (2.5,6) are simultaneously classified into Clusters 3 and 4.
0.02.55.07.510.00246810DepotCluster1Cluster2Cluster3Cluster40.02.55.07.510.00246810DepotCluster1Cluster2Cluster3Cluster4Cluster5Cluster6Cluster7Cluster8 the decentralized MATA algorithm are tested:
1) Dec-MATA:
the original algorithm with clustering
(Fig. 1);
2) DM-No-FCM-1.00: a modified algorithm where no
FCM clustering is used, and the the clustering coeffi-
cients (used in bigraph construction) are replaced with
a constant value of b∗
ij = 1.00,∀(i, j) ∈ (R, J); and
3) DM-No-FCM-0.01: same as the previous implementa-
tion (i.e., no FCM clustering), except that the clustering
coefficients are set at b∗
ij = 0.01,∀(i, j) ∈ (R, J).
B. Results: Impact of FCM & Bigraph Construction
Example illustrations showing outcomes of fuzzy clustering
and bigraph construction are given first to provide insights on
the impact of these steps of Dec-MATA. For example, Fig. 4
illustrates how the FCM divides the tasks into a set of m
clusters (m: # robots) for the test cases 2 and 3 (Table I).
The size of the marker in these plots is proportional to the
corresponding cluster coefficient of each task. Since FCM is a
soft clustering method, some tasks have two or more colors –
for example, two tasks at point (2.5,6) in Fig. 4 (a) belong to
both cluster 3 and 4. Figure 3 shows how the bigraph changes
after the first iteration during the application of Dec-MATA for
the test case 2 (Table I). It is readily evident, from the observed
variation in the edge weights across robot–task pairs, how the
bigraph is able to capture and alleviate the complexity of the
task-assignment problem.
C. Case Study Results: Performance Analysis
Table I summarizes the performances of the benchmark
centralized mTSP/MILP algorithm and the three different Dec-
MATA implementations, for the seven test cases. Both the
overall cost (as given by Eq. (7) incurred by the robot team
under each algorithm, and the variance (σ(Agents(cid:48)Cost)) in
the individual costs incurred by the robots within the team
are reported in the table. The reported values are estimated by
averaging over the ten different runs of each case (with ran-
domized depot and task locations). The overall cost incurred
by each algorithm at their individual termination for each case
is also shown as a bar plot in Fig. 7(a). It can be seen from
Table I that the Dec-MATA solutions get promisingly close to
the optimal solutions obtained by the benchmark centralized
MILP algorithm (within 7%–29 % difference in overall cost).
The observed variance in individual robot costs is comparable
between the centralized and our Dec-MATA approach. The
Dec-MATA approach also clearly outperforms the two other
implementations (DM-no-FCM-1.00 and DM-no-FCM-0.01)
where clustering is not used; this observation highlights the
importance of clustering in the Dec-MATA algorithm. When
a fixed value of bij is used (across all robot–task pairs) in
the absence of clustering, the value itself appears to have
a negligible impact, given the (observed) practically similar
performance of DM-no-FCM-1.00 and DM-no-FCM-0.01.
One of the most complex test cases, case 7 (with 50 agents
and 100 tasks) is chosen for further analysis. Figure 5 shows
the across-the-team variation in the cost incurred or task load
experienced by individual robots. While the task load variation
obtained by Dec-MATA is very similar to that obtained by the
benchmark centralized MILP algorithm, it is interesting to note
that the overall variance in individual robot task load reduces
in the absence of clustering. In order to judge the robustness
of Dec-MATA in comparison to the centralized algorithm,
Fig. 6 illustrates the boxplot of the overall costs incurred by
each algorithm across the ten runs of case 7. Variation in
performance is expected, since the depot and task locations
are randomized across the ten runs; what is interesting to note
is that the overall variance in performance of the Dec-MATA
algorithm is comparable to that of the centralized algorithm,
thereby highlighting the robustness of Dec-MATA.
Fig. 5. Boxplots showing the variation in task load per robot across the
team for test case 7 (m = 50 and n = 100), as accomplished by the
following four algorithms: CNT: Centralized, DM: Dec-MATA, DMNF/1:
DM-No-FCM-1.00, and DMNF/0.01: DM-No-FCM-0.01
Fig. 6. Boxplots showing the variation in overall cost across ten runs of test
case 7 (m = 50 and n = 100), as accomplished by the following four
algorithms: CNT: Centralized, DM: Dec-MATA, DMNF/1: DM-No-FCM-
1.00, and DMNF/0.01: DM-No-FCM-0.01
CNTDMDMNF/1DMNF/0.01Algorithm510152025TaskLoadofEachAgentCNTDMDMNF/1DMNF/0.01Algorithm500600700800Overallcostinatestrun (a) The best overall cost at termination of each algorithm (b) Computing time w.r.t. terminal performance state of Dec-MATA
Fig. 7. Performance and computing time of the algorithms for the 7 case studies. The computing time reported for the benchmark centralized algorithm is
the approximate time required by it to reach the performance (cost) achieved by the Dec-MATA method; note that the centralized approach takes significantly
(mostly several orders of magnitude) greater time to achieve its terminal optimal results depicted in (a).
D. Case Study Results: Computational Efficiency
The comparison of the computational efficiency of Dec-
MATA to that of the centralized algorithm is critical to ana-
lyzing (and thus appreciating) the advantages of decentralized
schemes in multi-robot operations. Since the centralized algo-
rithm and Dec-MATA are executed on two different platforms
(MATLAB and the Jupyter/Python environment, respectively),
a direct comparison is practically challenging. Thus, to obtain
an understanding of how the two environments natively com-
pare to each other, a basic for-loop of length 1000 is executed
10 times on each platform, and the recorded computation is
averaged across the 10 runs. This experiment showed that
the MATLAB environment (that exploits all 8 CPU cores)
is approximately 7.5 times faster than the Jupyter/Python
environment. Thereafter, by directly (without any adjustment)
comparing the recorded computation time of the centralized
algorithm (run on the MATLAB environment) and the Dec-
MATA algorithm (run on the Jupyter environment), the com-
putation time advantages of Dec-MATA discussed below can
be perceived as conservative.
The following three different computation times are mea-
sured and analyzed: 1) total computation time of the Dec-
MATA algorithm, termed tdm; 2) computation time of the
centralized MILP algorithm to reach/overtake Dec-MATA's
PERFORMANCE RESULTS OF THE ALGORITHMS ON 7 TEST CASES
TABLE I
Case
# of Tasks
# of Agents Max. Tasks per Robot
n
10
12
24
20
60
80
100
1
2
3
4
5
6
7
m
2
4
8
10
20
4
50
h
7
5
5
4
5
22
4
Centralized
35.56 ( 3.74 )
54.27 ( 4.73 )
92.69 ( 5.26 )
113.54 ( 4.96 )
197.74 ( 5.33 )
97.12 ( 4.27 )
551.74 ( 4.53 )
Overall Cost ( σ(Agents(cid:48)Cost) )
Dec-MATA
DM-No-FCM-1.00
DM-No-FCM-0.01
41.65 ( 2.99 )
64.81 ( 4.15 )
113.25 ( 4.58 )
121.6 ( 5.19 )
254.91 ( 5.13 )
122.81 ( 5.67 )
593.77 ( 4.68 )
43.2 ( 2.21 )
74.25 ( 2.65 )
138.79 ( 3.04 )
149.02 ( 2.94 )
338.24 ( 3.1 )
122.5 ( 3.69 )
742.56 ( 3.2 )
43.36 ( 2.51 )
74.68 ( 3.06 )
139.57 ( 3.16 )
148.03 ( 2.93 )
338.09 ( 3.17 )
122.39 ( 4.32 )
740.18 ( 3.31 )
COMPUTATION TIMES OF DEC-MATA AND MTSP ON 7 TEST CASES. tdm: TOTAL COMPUTATION TIME OF DEC-MATA; tREACH: COMPUTATION TIME OF
CENTRALIZED MILP TO REACH DEC-MATA'S PERFORMANCE; AND tOPT: COMPUTATION TIME OF CENTRALIZED MILP TO REACH THE FINAL/OPTIMAL
TABLE II
SOLUTIONS.
Case
# of Tasks
# of Agents Max. Tasks per Robot
n
10
12
24
20
60
80
100
1
2
3
4
5
6
7
m
2
4
8
10
20
4
50
h
7
5
5
4
5
22
4
Computation Time [sec]
tdm
0.01
0.02
0.06
0.06
0.53
0.08
4.44
treach
0.94
0.63
0.41
0.52
1.4
topt
1.18
0.82
7.67
0.65
129.74
4.62
6851.38
6004.65
4.63
1234567Case0200400600OverallCost[m]CentralizedDec-MATADM-No-FCM-1.00DM-No-FCM-0.011234567Case012345ComputationalTime[sec]CentralizedDec-MATADM-No-FCM-1.00DM-No-FCM-0.016020406080100120ComputationalTime[sec] (a) Computation time versus the number of robots
(b) Computation time versus the number of tasks
Fig. 8. Variation of the computation time (till termination) of the Centralized, Dec-MATA, and the Dec-MATA's component FCM algorithms w.r.t. problem
size. Dec-MATA's computation time is approximately 1-4 orders of magnitude smaller than that of the centralized algorithm at termination; while Dec-MATA's
computation cost increases with problem size, it is minimally sensitive to the task-to-robot ratio unlike the centralized algorithm (which yields very high
computation time for the 4 robot / 80 task case).
performance, termed treach; and 3) total computation time of
the centralized MILP algorithm to reach the final/optimal
solutions (topt). These computation times for the 7 test cases
(averaged across 10 randomized missions each) are summa-
rized in Table II. The computation times of Dec-MATA and the
centralized algorithm at Dec-MATA's terminal performance
level (i.e., tdm and treach) are also illustrated as bar plots
in Fig. 7 (b). For further analysis of how the problem size
and characteristics impact the algorithms' computational ef-
ficiency, Figs. 8 (a) and (b) illustrate the variation in the
algorithms' computation times with number of robots and
number of tasks, respectively.
It can be observed from Table II and Fig. 7 (b) that the cen-
tralized algorithm, in general, required 1-3 orders of magnitude
more computation time to reach the performance level of Dec-
MATA. The total time required by the centralized algorithm
to obtain its final/optimal is also fairly consistent w.r.t. to the
time when it overtakes Dec-MATA's performance, except in
the cases with a high tasks-to-robot ratio. This observation is
also evident from Figs. 8 (a) and (b). Table II and 8 show
that while the cost of Dec-MATA scales sub-linearly with
the problem size, it is minimally sensitive to the task-to-robot
ratio. On the other hand, for the high task-to-robot ratio, the
cost of the centralized algorithm spikes undesirably, e.g., it
takes 129 seconds to reach Dec-MATA's performance level and
104 minutes to reach its optimal solution for the 4 robot / 80
task case. Lastly, Fig. 8 shows that although the computation
cost of the fuzzy clustering scales with the problem size,
it always remains a small fraction of the total computation
cost of Dec-DATA thereby justifying its role as an efficient
intermediate approach to reduce the problem complexity of
MATA.
Overall, the sub-second computation time (as low as 10ms
for the 2 robot / 10 task case) of Dec-MATA, except for the
100 robot case, and its negligible sensitivity to task-to-robot
ratio clearly demonstrates Dec-MATA's promise in serving
as a computationally tractable solution to autonomous multi-
robot planning. Moreover, there remains an opportunity to
further advance the computational efficiency of Dec-MATA
by adopting more efficient (bipartite) maximum weighted
matching implementations.
V. CONCLUSION
In this paper, we developed a three-stage algorithm to
perform decentralized allocation of location-based tasks in
autonomous multi-robot systems, e.g., those catering to ware-
house robotics and last-mile delivery applications. The decen-
tralized multi-agent task-allocation or Dec-MATA algorithm
aims to tackle the issue of exploding computational cost with
increasing problem size (i.e., number of robots/tasks in the
environment), in order to preserve tractability of use in online
planning. To this end,
the first and second stages of the
algorithm respectively seek to reduce the dimensionality of
the problem (to be solved by individual agents) and generate
a compact representation of task-to-robot relevance; and the
third stage of the algorithm allows each agent to identify
optimal task sequences to undertake. A novel combination of
fuzzy clustering, and bipartite graph construction and maxi-
mum weighted matching is investigated to design this three-
stage Dec-MATA algorithm.
To evaluate the performance of the new Dec-MATA algo-
rithm, a benchmark centralized algorithm is formulated based
on a state-of-the-art MILP solution of the multi-traveling sales-
man (mTSP) problem. Both algorithms are tested on a suite of
case studies involving a varying number of robots and tasks (2
robot/10 task to 50 robot/100 task scenarios), and performance
is analyzed in terms of the total cost incurred by the robot team
(total distance traveled in this case) and computation time.
While getting within 7-28% of the minimum cost given by the
centralized MILP algorithm, dec-MATA is observed to provide
1-3 orders of magnitude gain in computational efficiency (at
comparable performance level), relatively similar robustness
01020304050NumberofRobots10−210−1100101102103104ComputationTime[sec]CentralizedDec-MATAFCM20406080100NumberofTasks10−210−1100101102103104ComputationTime[sec]CentralizedDec-MATAFCM [17] Y. Shoham and K. Leyton-Brown, Multiagent systems: Algorithmic,
game-theoretic, and logical foundations. Cambridge University Press,
2008.
[18] B. Dai and H. Chen, "A multi-agent and auction-based framework and
approach for carrier collaboration," Logistics Research, vol. 3, no. 2-3,
pp. 101–120, 2011.
[19] H.-L. Choi, L. Brunet, and J. P. How, "Consensus-based decentralized
task allocation," IEEE transactions on robotics,
auctions for robust
vol. 25, no. 4, pp. 912–926, 2009.
[20] R. D'Andrea, "Guest editorial: A revolution in the warehouse: A
retrospective on kiva systems and the grand challenges ahead," IEEE
Transactions on Automation Science and Engineering, vol. 9, no. 4, pp.
638–639, 2012.
[21] P. R. Wurman, R. D'Andrea, and M. Mountz, "Coordinating hundreds of
cooperative, autonomous vehicles in warehouses," AI magazine, vol. 29,
no. 1, p. 9, 2008.
[22] C. E. Miller, A. W. Tucker, and R. A. Zemlin, "Integer programming
the ACM
formulation of traveling salesman problems," Journal of
(JACM), vol. 7, no. 4, pp. 326–329, 1960.
[23] J. C. Bezdek, R. Ehrlich, and W. Full, "Fcm: The fuzzy c-means
clustering algorithm," Computers & Geosciences, vol. 10, no. 2-3, pp.
191–203, 1984.
[24] J. C. Bezdek, "Objective function clustering," in Pattern recognition with
fuzzy objective function algorithms. Springer, 1981, pp. 43–93.
[25] G. M. Slota, "General graph matching," in Graph Theory, Spring 2017.
Rensselaer Polytechnic Institute, 2017.
[30] Z. Galil, "Efficient algorithms for finding maximum matching in graphs,"
ACM Computing Surveys (CSUR), vol. 18, no. 1, pp. 23–38, 1986.
[31] J. Edmonds, "Maximum matching and a polyhedron with 0, 1-vertices,"
Journal of Research of the National Bureau of Standards B, vol. 69, no.
125-130, pp. 55–56, 1965.
[32] O. Zedadra, N. Jouandeau, H. Seridi, and G. Fortino, "Multi-agent
foraging: state-of-the-art and research challenges," Complex Adaptive
Systems Modeling, vol. 5, no. 1, p. 3, 2017.
[33] I. Gurobi Optimization, "Gurobi optimizer reference manual," 2016.
[Online]. Available: http://www.gurobi.com
[26] Z. Huang, D. D. Zeng, and H. Chen, "Analyzing consumer-product
graphs: Empirical findings and applications in recommender systems,"
Management science, vol. 53, no. 7, pp. 1146–1164, 2007.
[27] Z. Zhu, J. Su, and L. Kong, "Measuring influence in online social
network based on the user-content bipartite graph," Computers in Human
Behavior, vol. 52, pp. 184–189, 2015.
[28] A. S. Asratian, T. M. Denley, and R. Haggkvist, Bipartite graphs and
their applications. Cambridge University Press, 1998, vol. 131.
[29] L. Lov´asz and M. D. Plummer, Matching theory. American Mathe-
matical Soc., 2009, vol. 367.
across randomized missions, and clearly superior scalability
with increasing task-to-robot ratio, when compared to the
centralized algorithm. Dec-Mata solves all the test cases within
a fraction of a second, except the largest case (50 robot/100
task), which requires 4.44 seconds. The latter observation
indicates opportunities to further optimize the Dec-MATA
algorithm by adopting more efficient problem decomposition,
graph matching, and program parallelization approaches; these
are the immediate next steps in this research. Moreover, the
current version of Dec-MATA makes certain simplifying as-
sumptions, which are likely reasonable for warehouse robotics-
type applications. For example, decisions are synchronized
across the team (no active consensus is however required),
the task feature and agent states are assumed to be precisely
known, and each agent has a full observation of all tasks and
the state of other agents. Future work will focus on alleviat-
ing these assumptions, thereby allowing wider application of
Dec-MATA to asynchronous multi-robot problems involving
uncertain environment and partial observability.
REFERENCES
[1] N. Kuhn
and
J. Muller,
"Multi-agent
planning," Deutsches
Forschungszentrum fur Kunstliche Intelligenz, p. 61, 1992.
[2] G. A. Korsah, A. Stentz, and M. B. Dias, "A comprehensive taxonomy
for multi-robot task allocation," The International Journal of Robotics
Research, vol. 32, no. 12, pp. 1495–1512, 2013.
[3] G. Colistra, V. Pilloni, and L. Atzori, "The problem of task allocation
in the internet of things and the consensus-based approach," Computer
Networks, vol. 73, pp. 98–111, 2014.
[4] T. Kalm´ar-Nagy, G. Giardini, and B. D. Bak, "The multiagent planning
problem," Complexity, vol. 2017, 2017.
[5] T. Bektas, "The multiple traveling salesman problem: an overview of
formulations and solution procedures," Omega, vol. 34, no. 3, pp. 209–
219, 2006.
[6] G. Gutin and A. P. Punnen, The traveling salesman problem and its
variations. Springer Science & Business Media, 2006, vol. 12.
[7] D. S. Johnson and L. A. McGeoch, "The traveling salesman problem:
A case study in local optimization," Local search in combinatorial
optimization, vol. 1, pp. 215–310, 1997.
[8] G. Laporte and Y. Nobert, "A cutting planes algorithm for the m-
salesmen problem," Journal of the Operational Research society, vol. 31,
no. 11, pp. 1017–1023, 1980.
[9] B. Gavish and K. Srikanth, "An optimal solution method for large-scale
multiple traveling salesmen problems," Operations Research, vol. 34,
no. 5, pp. 698–717, 1986.
[10] J. Gromicho, J. Paixao, and I. Bronco, "Exact solution of multiple
traveling salesman problems," in Combinatorial Optimization. Springer,
1992, pp. 291–292.
[11] T. Zhang, W. Gruver, and M. H. Smith, "Team scheduling by genetic
search," in Intelligent Processing and Manufacturing of Materials, 1999.
IPMM'99. Proceedings of
the Second International Conference on,
vol. 2.
IEEE, 1999, pp. 839–844.
[12] J. L. Ryan, T. G. Bailey, J. T. Moore, and W. B. Carlton, "Reactive tabu
search in unmanned aerial reconnaissance simulations," in Proceedings
of the 30th conference on Winter simulation.
IEEE Computer Society
Press, 1998, pp. 873–880.
[13] S. Gorenstein, "Printing press scheduling for multi-edition periodicals,"
Management Science, vol. 16, no. 6, pp. B–373, 1970.
[14] R. Jonker and T. Volgenant, "An improved transformation of the
symmetric multiple traveling salesman problem," Operations Research,
vol. 36, no. 1, pp. 163–167, 1988.
[15] R. Nallusamy, K. Duraiswamy, R. Dhanalaksmi, and P. Parthiban,
"Optimization of non-linear multiple traveling salesman problem using
k-means clustering, shrink wrap algorithm and meta-heuristics," Inter-
national Journal of Nonlinear Science, vol. 8, no. 4, pp. 480–487, 2009.
[16] A. Khamis, A. Hussein, and A. Elmogy, "Multi-robot task allocation:
A review of the state-of-the-art," in Cooperative Robots and Sensor
Networks 2015. Springer, 2015, pp. 31–51.
|
1201.2207 | 1 | 1201 | 2012-01-10T23:29:50 | Multi-sensor Information Processing using Prediction Market-based Belief Aggregation | [
"cs.MA"
] | We consider the problem of information fusion from multiple sensors of different types with the objective of improving the confidence of inference tasks, such as object classification, performed from the data collected by the sensors. We propose a novel technique based on distributed belief aggregation using a multi-agent prediction market to solve this information fusion problem. To monitor the improvement in the confidence of the object classification as well as to dis-incentivize agents from misreporting information, we have introduced a market maker that rewards the agents instantaneously as well as at the end of the inference task, based on the quality of the submitted reports. We have implemented the market maker's reward calculation in the form of a scoring rule and have shown analytically that it incentivizes truthful revelation or accurate reporting by each agent. We have experimentally verified our technique for multi-sensor information fusion for an automated landmine detection scenario. Our experimental results show that, for identical data distributions and settings, using our information aggregation technique increases the accuracy of object classification favorably as compared to two other commonly used techniques for information fusion for landmine detection. | cs.MA | cs |
Multi-sensor Information Processing using
Prediction Market-based Belief Aggregation
Janyl Jumadinova (1), Prithviraj Dasgupta (1)
((1) University of Nebraska at Omaha)
November 9, 2018
We consider the problem of information fusion from multiple sensors of differ-
ent types with the objective of improving the confidence of inference tasks, such
as object classification, performed from the data collected by the sensors. We
propose a novel technique based on distributed belief aggregation using a multi-
agent prediction market to solve this information fusion problem. To monitor
the improvement in the confidence of the object classification as well as to dis-
incentivize agents from misreporting information, we have introduced a market
maker that rewards the agents instantaneously as well as at the end of the infer-
ence task, based on the quality of the submitted reports. We have implemented
the market maker's reward calculation in the form of a scoring rule and have
shown analytically that it incentivizes truthful revelation or accurate reporting
by each agent. We have experimentally verified our technique for multi-sensor
information fusion for an automated landmine detection scenario. Our experi-
mental results show that, for identical data distributions and settings, using our
information aggregation technique increases the accuracy of object classification
favorably as compared to two other commonly used techniques for information
fusion for landmine detection.
1
Introduction
Information fusion from multiple sensors has been a central research topic in
sensor-based systems [17] and recently several multi-agent techniques [16] have
been proposed to address this problem. Most of the solutions for multi-sensor
information fusion and processing are based on Bayesian inference techniques
[8, 13, 15]. While such techniques have been shown to be very effective, we in-
vestigate a complimentary problem where sensors can behave in a self-interested
manner. Such self-interested behavior can be motivated by malicious nodes that
might have been planted into the system to subvert its operation, or, by nor-
mal sensor nodes attempting to give an illusion of efficient performance when
they do not have enough resources (e.g., battery power) to perform accurate
measurements. To address this problem, we describe a market-based aggrega-
tion technique called a prediction market for multi-sensor information fusion
that includes a utility driven mechanism to motivate each sensor, through its
associated agent, to reveal accurate reports.
To motivate our problem we describe a distributed automated landmine
1
detection scenario used for humanitarian demining. An environment contains
different buried objects, some of which could potentially be landmines. A set of
robots, each equipped with one of three types of landmine detection sensor such
as a metal detector (MD), or a ground penetrating radar (GPR) or an infra-red
(IR) heat sensor, are deployed into this environment. Each robot is capable
of perceiving certain features of a buried object through its sensor such as the
object's metal content, area, burial depth, etc. However, the sensors give noisy
readings for each perceived feature depending on the characteristics of the object
as well as on the characteristics of the environment (e.g., moisture content,
ambient temperature, sunlight, etc.). Consequently, a sensor that works well
in one scenario, fails to detect landmines in a different scenario, and, instead
of a single sensor, multiple sensors of different types, possibly with different
detection accuracies can detect landmines with higher certainty [5]. Within this
scenario, the central question that we intend to answer is: given an initial set
of reports about the features of a buried object, what is a suitable set (number
and type) of sensors to deploy over a certain time window to the object, so
that, over this time window, the fused information from the different sensors
successively reduces the uncertainty in determining the object's type.
Our work in this paper is based on the insight that the scenario illustrated
above, of fusing information from multiple sources to predict the outcome of an
initially unknown object, is analogous to the problem of aggregating the beliefs
of different humans to forecast the outcome of an initially unknown event. Such
forecasting is frequently encountered in many problems such as predicting the
outcome of geo-political events, predicting the outcome of financial instruments
like stocks, etc. Recently, a market-based model called prediction market has
been shown to be very successful in aiding humans with such predictions and
with decision-making [1, 2, 14, 18]. Building on these models, in this paper, we
describe a multi-agent prediction market for multi-sensor information fusion.
Besides being an efficient aggregation mechanism, using prediction markets gives
us several useful features - a mathematical formulation called a scoring rule
that deters malicious sensors from misreporting information, a regression-based
belief update mechanism for the sensor agents for incorporating the aggregated
beliefs (or information estimates) of other sensors into their own calculation, and
the ability to incorporate an autonomous decision maker that uses expert-level
domain knowledge to make utility maximizing decisions to deploy additional
sensors appropriately to improve the detection of an object. Our experimental
results illustrated with a landmine detection scenario while using identical data
distributions and settings, show that the information fusion performed using
our technique reduces the root mean squared error by 5 − 13% as compared to
a previously studied technique for landmine data fusion using the Dempster-
Shafer theory [10] and by 3 − 8% using distributed data fusion technique [9].
2 Related Work
Multi-agent Information Fusion. Multi-agent systems have been used to
solve various sensor network related problems and an excellent overview is given
in [16]. In the direction of multi-sensor information processing, significant works
include the use of particle filters[15], distributed data fusion (DDF) architecture
along with its extension, the Bayesian DDF [8, 9], Gaussian processes [13] and
2
mobile agent-based information fusion [19]. For most of the application do-
mains described in these works such as water-tide height measurement, wind
speed measurement, robot tracking and localization, etc., self-interested be-
havior by the sensors is not considered a crucial problem. For our illustrated
application domain of landmine detection, decision-level fusion techniques have
been reported to be amenable for scenarios where the sensor types are differ-
ent from each other, and, non-statistical decision-level fusion techniques, such
as Dempster-Shafer theory [10], fuzzy logic[3], and rule-based fusion techniques
[5] have been reported to generalize well. However, in contrast to our work,
these techniques assume that sensors are fully cooperative and never behave
self-interestedly by misreporting information.
In [12], the authors have ob-
served that most sensor-based information aggregation techniques either do not
consider malicious behavior or use high-overhead, cryptographic techniques to
combat it. To deter false reports by sensor nodes in a data aggregation setting,
they propose various lower overhead reputation-based schemes. Our predic-
tion market-based information aggregation technique is complimentary to such
reputation-based aggregation techniques.
Decision-Making using Prediction Markets. A prediction market is
a market-based aggregation mechanism that is used to combine the opinions
on the outcome of a future, real-world event from different people, called the
market's traders and forecast the event's possible outcome based on their ag-
gregated opinion. Recently, multi-agent systems have been used [1, 7, 14] to
analyze the operation of prediction markets, where the behaviors of the mar-
ket's participants are implemented as automated software agents. The seminal
work on prediction market analysis [18] has shown that the mean belief values
of individual traders about the outcome of a future event corresponds to the
event's market price. The basic operation rules of a prediction market are simi-
lar to those of a continuous double auction, with the role of the auctioneer being
taken up by an entity called the market maker that runs the prediction market.
Hanson [6] developed a mechanism, called a scoring rule, that can be used by
market makers to reward traders for making and improving a prediction about
the outcome of an event, and, showed that if a scoring rule is proper or incentive
compatible, then it can serve as an automated market maker. Recently, authors
in [2, 14] have theoretically analyzed the properties of prediction markets used
for decision making. In [14], the authors analyzed the problem of a decision
maker manipulating a prediction market and proposed a family of scoring rules
to address the problem. In [2], the authors extended this work by allowing ran-
domized decision rules, considering multiple possible outcomes and providing
a simple test to determine whether the scoring rule is proper for an arbitrary
decision rule-scoring rule pair. In this paper, we use a prediction market for de-
cision making, but in contrast to previous works we consider that the decision
maker can make multiple, possibly improved decisions over an event's duration,
and, the outcome of an event is decided independently, outside the market,
and not influenced by the decision maker's decisions. Another contribution our
paper makes is a new, proper scoring rule, called the payment function, that
incentivizes agents to submit truthful reports.
3
3 Problem Formulation
Let L be a set of objects. Each object has certain features that determine its
type. We assume that there are f different features and m different object types.
Let Φ = {φ1, φ2, ..., φf } denote the set of object features and Θ = {θ1, θ2, ..., θm}
denote the set of object types. The features of an object l ∈ L is denoted by
lΦ ⊆ Φ and its type is denoted by lθ ∈ Θ. As illustrated in the example given
in Section 1, lΦ can be perceived, albeit with measurement errors, through
sensors, and, our objective is to determine lθ as accurately as possible from
the perceived but noisy values of lΦ. Let ∆(Θ) = {(δ(θ1), δ(θ2), ..., δ(θm)) :
i=1 δ(θi) = 1}, denote the set of probability distributions over
the different object types. For convenience of analysis, we assume that when
the actual type of object l, lθ = θj, its (scalar) type is expanded into a m-
dimensional probability vector using the function vec : Θ → [0, 1]m : vecj =
1, veci6=j = 0, which has 1 as its j-th component corresponding to l's type θj
and 0 for all other components.
δ(θi) ∈ [0, 1],Pm
Let A denote a set of agents (sensors) and At,l
rep ⊆ A denote the subset of
agents that are able to perceive the object l's features on their sensors at time t.
Based on the perceived object features, agent a ∈ At,l
rep at time t reports a belief
as a probability distribution over the set of object types, which is denoted as
ba,t,l ∈ ∆(Θ). The beliefs of all the agents are combined into a composite belief,
(ba,t,l), and let Θt,l : Bt,l → ∆(Θ) denote a function that
Bt,l = Agga∈At,l
computes a probability distribution over object types based on the aggregated
agent beliefs. Within this setting we formulate the object classification problem
as a decision making problem in the following manner: given an object l and an
initial aggregated belief Bt,l calculated from one or more agent reports for that
object, determine a set of additional agents (sensors) that need to be deployed
at object l such that the following constraint is satisfied:
rep
min RM SE(cid:16) Θt,l, vec(lθ)(cid:17) ,
for
t = 1, 2, ....T
(1)
where T is the time window for classifying an object l and RMSE is the root
mean square error given by RM SE(x, y) = x−y√m . In other words, at every
time step t, the decision maker tries to select a subset of agents such that the
root mean square error (RMSE) between the estimated type of object l and its
actual type is successively minimized.
The major components of the object classification problem described above
integrating the reports from the different sensors and
consists of two parts:
making sensor deployment decisions based on those reports so that the objective
function given in Equation 1 is satisfied. To address the first part, we have
used distributed information aggregation with a multi-agent prediction market,
while for the latter we have used an expected utility maximizing decision-making
framework. A schematic showing the different components of our system and
their interactions is shown in Figure 1 and explained in the following sections.
3.1 Sensor Agents
As mentioned in Section 1, there is a set of robots in the scenario and each
robot has an on-board sensor for analyzing the objects in the scenario. Different
robots can have different types of sensors and sensors of the same type can have
4
Figure 1: The different components of the prediction market for decision making
and the interactions between them.
different degrees of accuracy determined by their cost. Every sensor is associated
with a software agent that runs on-board the robot and performs calculations
related to the data sensed by the robot's sensor. In the rest of the paper, we
have used the terms sensor and agent interchangeably. For the ease of notation,
we drop the subscript l corresponding to an object for the rest of this section.
When an object is within the sensing range of a sensor (agent) a at time t, the
sensor observes the object's features and its agent receives this observation in the
form of an information signal ga,t =< g1, ..., gf > that is drawn from the space
of information signals G ⊆ ∆(Θ). The conditional probability distribution of
object type θj given an information signal g ∈ G, P (θjg) : G → [0, 1], is
constructed using domain knowledge [3, 10, 11] within a Bayesian network and
is made available to each agent. Agent a then updates its belief distribution
ba,t using the following equation:
ba,t = wbel · P(Θga,t) + (1 − wbel) · Bt,
(2)
where Bt is the belief value vector aggregated from all sensor reports.
Agent Rewards. Agents behave in a self-interested manner to ensure that
they give their 'best' report using their available resources including sensor,
battery power, etc. However, some agents can behave maliciously, either being
planted or compromised to infiltrate the system and subvert the object classi-
fication process, or, they might be trying to give an illusion of being efficient
when they do not have sufficient resources to give an accurate report. An agent
a that submits a report at time t, uses its belief distribution ba,t to calculate
the report ra,t =< ra,t
m >∈ ∆(Θ). An agent can have two strategies
to make this report - truthful or malicious. If the agent is truthful, its report
1 , ..., ra,t
5
corresponds to its belief, i.e., ra,t = ba,t. But if it is malicious, it manipulates
its report to reveal an inaccurate belief. Each agent a can update its report ra,t
within the time window T by obtaining new measurements from the object and
using Equation 2 to update its belief. The report from an agent a at time t is
analyzed by a human or agent expert [10] to assign a weight wa,t depending on
the current environment conditions and agent a's sensor type's accuracy under
those environment conditions (e.g., rainy weather reduces the weight assigned
to the measurement from an IR heat sensor, or, soil that is high in metal content
reduces the weight assigned to the measurement from an metal detector).
To motivate an agent to submit reports, an agent a gets an instantaneous
reward, ρa,t, from the market maker for the report ra,t it submits at time t, cor-
responding to its instantaneous utility, which is given by the following equation:
ρa,t = V (nt′=1..t) − C a(ra,t),
(3)
where V (nt′=1..t) is the value for making a report with nt′=1..t being the number
of times the agent a submitted a report up to time t, and, C a(ra,t) is the
cost of making report ra,t for agent a based on the robot's expended time,
battery power, etc. We denote the agent's value for each report V (nt′=1..t) as
a constant-valued function up to a certain threshold and a linearly decreasing
function thereafter, to de-incentivize agents from making a large number of
reports. Agent a's value function is given by the following equation:
V (nt′=1..t) =( ν
ν(nt′ =1..t−nmax)
(nthreshold−nmax)
, nt′=1..t ≤ nthreshold
, otherwise
where ν ∈ Z+, is a constant value that a gets by submitting reports up to a
threshold, nthreshold is the threshold corresponding to the number of reports a
can submit before its report's value starts decreasing, and, nmax is the maximum
number of reports agent a can submit before V becomes negative. Finally, to
determine its strategy while submitting its report, an agent selects the strategy
that maximizes its expected utility obtained from its cumulative reward given
by Equation 3 plus an expected value of its final reward payment if it continues
making similar reports up to the object's time window T .
3.2 Decision Maker Agent
The decision maker agent's task is to use the composite belief about an object's
type, Bt, given by the prediction market, and take actions to deploy additional
robots(sensors) based on the value of the objective function given in Equation
1. Let AC denote a set of possible actions corresponding to deploying a certain
number of robots, and D = {d1, ...dh} : di ∈ Ac ⊆ AC denote the decision
set of the decision maker. The decision function of the decision maker is given
by dec : ∆(Θ) → D. Let udec
j ∈ Rm be the utility that the decision maker
receives by determining an object to be of type θj and let P (diθj) be the
probability that the decision maker makes decision di ∈ D given object type θj.
P (diθj) and udec
are constructed using domain knowledge [3, 11, 10]. Given the
aggregated belief distribution Bt at time t, the expected utility to the decision
j=1 P (diθj) ·
· Bt. The decision that the decision maker takes at time t, also called its
maker for taking decision di at time t is then EU dec(di, Bt) = Pm
j
udec
j
6
decision rule, is the one that maximizes its expected utility and is given by:
dt = arg maxdi EU dec(di, Bt).
3.3 Prediction Market
A conventional prediction market uses the aggregated beliefs of the market's
participants or traders about the outcome of a future event, to predict the
event's outcome. The outcome of an event is represented as a binary variable
(event happens/does not happen). The traders observe information related to
the event and report their beliefs, as probabilities about the event's outcome.
The market maker aggregates the traders' beliefs and uses a scoring rule to
determine a payment or payoff that will be received by each reporting trader.
In our multi-agent prediction market, traders correspond to sensor agents, the
market maker agent automates the calculations on behalf of the conventional
market maker, and, an event in the conventional market corresponds to identi-
fying the type of a detected object. The time window T over which an object
is sensed is called the duration of the object in the market. This time window
is divided into discrete time steps, t = 1, 2...T . During each time step, each
sensor agent observing the object submits a report about the object's type to
the market maker agent. The market maker agent performs two functions with
these reports. First, at each time step t, it aggregates the agent reports into
an aggregated belief about the object, Bt ∈ ∆(Θ). Secondly, it calculates and
distributes payments for the sensor agents. It pays an immediate but nominal
reward to each agent for its report at time step t using Equation 3. Finally, at
the end of the object's time window T , the market maker also gives a larger pay-
off to each agent that contributed towards classifying the object's type. The
calculations and analysis related to these two functions of the market maker
agent are described in the following sections.
Final Payoff Calculation. The payoff calculation for a sensor agent is
performed by the market maker using a decision scoring rule at the end of the
object's time window. A decision scoring rule [2] is defined as any real valued
function that takes the agents' reported beliefs, the realized outcome and the
decisions made by the decision maker as input, and produces a payoff for the
agent for its reported beliefs, i.e. S : ∆(Θ) × Θ × D −→ R. We design a scoring
rule for decision making that is based on how much agent a's final report helped
the decision maker to make the right decisions throughout the duration of the
prediction market and by how close the agent a's final report is to actual object
type. Our proposed scoring rule for decision making given that object's true
type is θj is given in Equation 4:
S(ra,t
j
, d[1:t], θj) = (d[1:t], θj)log(cid:0)ra,t
j (cid:1) ,
(4)
j
where, ra,t
is the reported belief that agent a submitted at time t for object type
θj, d[1:t] is the set consisting of all the decisions that the decision maker took
related to the object up to the current time t, θj is the object's true type that was
revealed at the end the object's time window, log(cid:0)ra,t
of the report at time t relative to the true object type θj, and, (d[1:t], θj) is the
weight, representing how good all the decisions the decision maker took up to
time t were compared to the true object type θj. (d[1:t], θj) is determined by
the decision maker and made available to the agents through the market maker.
j (cid:1) measures the goodness
7
We assume that (d[1:t], θj) =Pt
, which gives the expected
utility of the decision maker agent for making decision i when the true type of
the object is θj.
i=1 P (di θj) · udec
j
Aggregation. Since a sensor agent gets paid both through its immediate
rewards for making reports during the object's time window and through the
scoring rule function for decision making at the end of the object's time window,
we define the total payment that the agent has received by the end of the object's
time window as a payment function.
Definition 1. A function Ψ(ra,t, d[1:t], θj, nt′=1..t) is called a payment func-
tion if each agent a's total received payment at the end of the object's time
window (when t = T ) is
Ψ(ra,t, d[1:t], θj, nt′=1..t) =
t
Xk=1
ρa,k + S(ra,t
j
, d[1:t], θj)
(5)
where ρa,k, S(ra,t
and 4.
j
, d[1:t], θj) and their components are defined as in Equations 3
Let Ψave denote a weighted average of the payment function in Equation 5
over all the reporting agents, using the report-weights assigned by the expert in
Section 3.1, as given below:
t
Ψave(rAt
rep,t, d[1:t], θj, nAt
rep,t) =
wa,kρa,k
rep
Xk=1 Xa∈At
+ (d[1:t], θj) Xa∈At
rep
wa,tlog(cid:0)ra,t
j (cid:1) ,
(6)
(7)
(8)
where At
rep is the subset of agents that are able to perceive object feature at
time t and wa,k is the weight assigned to agent a at time k by the expert. To
calculate an aggregated belief value in a prediction market, Hanson [6] used
the generalized inverse function of the scoring rule. Likewise, we calculate the
aggregated belief for our market maker agent by taking the generalized inverse
of the average payment function given in Equation 7:
Bt
j = Agga∈At
rep(ba,t)
exp(cid:16)Ψave−Pt
k=1 Pa∈At
(d[1:t],θj)
rep
wa,kρa,k(cid:17)
=
Pθm
θj =θ1
exp(cid:16)Ψave−Pt
k=1 Pa∈At
rep
(d[1:t],θj)
wa,kρa,k(cid:17)
where Bt
j ∈ Bt is the j-th component of the aggregated belief for object type θj.
The aggregated belief vector, Bt, calculated by the market maker agent is sent
to the decision maker agent so that it can calculate its expected utility given in
Section 3.2, as well as, sent back to each sensor agent that reported the object's
type till time step t, so that the agent can refine its future reports, if any, using
this aggregate of the reports from other agents.
8
4 Payment function: Properties
and Characteristics
In this section we first show that the payment function is proper, or incentive
compatible. Then we show that when the market maker uses this payment
function to reward each agent for its reported beliefs, reporting beliefs truthfully
is the optimal strategy for each agent.
We can characterize a proper payment function similar to a proper scoring
rule.
Definition 2. A payment function Ψ is proper, or incentive compatible, if
Ψ(ba,t, d[1:t], θj, nt′=1..t) ≥ Ψ(ra,t, d[1:t], θj, nt′=1..t),
(9)
∀ba,t, ra,t ∈ ∆(O).
Ψ is strictly proper if Equation 9 holds with equality, i.e., iff ba,t = ra,t.
Payment functions can be shown to be proper by representing them using
convex functions [2, 4]. To show that our payment function in Equation 5 is
proper, we characterize it in terms of a convex function, as shown below:
Theorem 1. A payment function Ψ is proper for decision making if
G′i,j(ra,t)
P (diθj)
Ψ(ra,t, d[1:t], θj, nt′=1..t) = G(ra,t) − G′(ra,t) · (ra,t) +
,
(10)
where G(ra,t) is a convex function and G′(ra,t) is a subgradient of G at point
ra,t and P (diθj) > 0.
Proof. Consider a payment function Ψ satisfying Equation 10. We will show
that Ψ must be proper for decision making. We will drop the agent and time sub-
scripts in this proof, and also we will write Ψ(r, d[1:t]) (or its element Ψ(rj, diθj))
instead of full Ψ(ra,t, d[1:t], θj, nt′=1..t).
EU (b, b) =
=
h
h
Xi=1
Xi=1
m
m
Xj=1
Xj=1
P (diθj)bjΨ(bj, diθj)
P (di)bj(cid:18)G(b) − G′(b) · b +
G′i,j (b)
P (diθj)(cid:19)
h
m
= G(b) − G′(b) · b +
G′i,j (b)bj
Xi=1
Xj=1
= G(b) − G′(b) · b + G′(b) · b = G(b).
Since G is convex and G′ is its subgradient, we have
h
m
EU (b, r) =
=
h
Xi=1
Xi=1
m
Xj=1
Xj=1
P (diθj)bjΨ(rj , diθj)
P (diθj)bj(cid:18)G(r) − G′(r) · r +
G′i,j (r)
P (diθj)(cid:19)
= G(r) − G′(r)(b − r)
≤ G(b) = EU (b, b).
9
Thus, Ψ is a proper payment function for decision making. Ψ is strictly proper
payment function and the inequality is strict if G is a strictly convex function.
Proposition 1. The payment function given in Equation 5 is proper.
Proof. Let G(b) = EU (b, b) and
G′i,j(b) = P (diθj)Ψ(b, d[1:t], θj, nt′=1..t). Then we can write the payment func-
tion as
h
m
Ψ(b, d[1:t], θj, nt′=1..t) =
P (diθj)bjΨ(b, d[1:t], θj, nt′=1..t)
Xj=1
Xi=1
Xj=1
Xi=1
m
h
− b ·
P (diθj)Ψ(b, d[1:t], θj, nt′=1..t)
Ψ(b, d[1:t], θj, nt′=1..t) · P (diθj)
+
P (diθj)
= Ψ(b, d[1:t], θj, nt′=1..t).
We can clearly see that the payment function can be written in the form given
in Equation 10 from Theorem 1. Therefore, the payment function Ψ given in
Equation 5 is a proper payment function.
Agent Reporting Strategy. Assume that agent a's report at time t is
k=1 ρa,k +
, d[1:t], θj). Then, agent a's expected utility for object type θj given its
, and its true belief about object type θj,
its final report, then its utility function can be written as ua,t
S(ra,t
reported belief for object type θj, ra,t
ba,t
j
j = Pt
at time t is
j
j
EU a
j (ra,t
j
, ba,t
j ) =
=
h
h
Xi=1
Xi=1
P (diθj)ba,t
j ua,t
j
(11)
P (diθj)ba,t
j t
Xk=1
ρa,k + S(ra,t
j
, d[1:t], θj)! ,
where P (diθj) is the probability that the decision maker takes decision di when
the object's type is θj.
Proposition 2. If agent a is paid according to Ψ, then it reports its beliefs
about the object types truthfully.
Proof. Sensor agent a wants to maximize its expected utility function and solves
the following program
arg max
h
r
Xi=1
m
Xj=1
P (diθj)ba,t
j " t
Xk=1
j (cid:1)#
ρa,k + (d[1:t], θj)log(cid:0)ra,t
,
10
j=1 ra,t
j = 1.
The Lagrangian is
s.t. Pm
P (diθj)ba,t
h
m
L(r, λ) =
Xj=1
Xi=1
j − 1
− λ
Xj=1
.
ra,t
m
The first order conditions are
j " t
Xk=1
j (cid:1)#
ρa,k + (d[1:t], θj)log(cid:0)ra,t
∂L
∂rj
=
h
m
Xi=1
Xj=1
⇒ ra,∗j =
∂L
∂λ
= −
m
Xj=1
P (diθj)ba,t
j
(d[1:t], θj)
ra,∗j
− λ = 0
i=1 P (diθj)
(d[1:t], θj)ba,t
λ
j Ph
ra,t
j + 1 = 0.
Substituting ra,∗j
into the second equation above, we have
(d[1:t], θj)ba,t
λ
i=1 P (diθj)
= 1
λ = (d[1:t], θj)
P (diθj)
j Ph
Xi=1
h
ra,∗j = ba,t
j
.
5 Experimental Results
We have conducted several experiments using our aggregation technique for
decision-making within a multi-sensor landmine detection scenario described in
Section 1. Our environment contains different buried objects, some of which
are landmines. The true types of the objects are randomly determined at the
beginning of the simulation. Due to the scarcity of real data related to landmine
detection, we have used the domain knowledge that was reported in [3, 10, 11] to
determine object types, object features, sensor agents' reporting costs, decision
maker agent's decision set, decision maker agent's utility of determining objects
of different types, and, to construct the probability distributions for P (θjg) and
P (diθj). We report simulation results for root mean squared error (RMSE)
defined in Section3 and also for number of sensors over time, cost over object
types, and average utility of the sensors over time.
Compared Techniques. For comparing the performance of our prediction
market based object classification techniques, we have used two other well-
known techniques for information fusion: (a) Dempster-Shafer (D-S) theory for
11
Name
Object types
Features
Sensor types
Max no. of sensors
Max no. of decisions
T (object identification window)
ν (agent's value if nt′=1..t ≤ nthreshold)
nmax (max no. of reports before value
is negative)
nthreshold (no. of reports before agent's value
is less than ν)
Value
mine, metallic object(non-mine),
non-metallic object(non-mine)
metallic content, object's area,
object's depth, sensor's position
MD, IR, and GPR
10 (5MD,3IR,2GPR)
14
10
5
20
5
Table 1: Parameters used for our simulation experiments.
landmine classification [10], where a two-level approach based on belief func-
tions is used. At the first level, the detected object is classified according to
its metal content. At the second level the chosen level of metal content is fur-
ther analyzed to classify the object as a landmine or a friendly object. The
belief update of the sensors that we used for D-S method is the same one we
have described in Section 3.1. (b) Distributed Data Fusion (DDF) [9], where
sensor measurements are refined over successive observations using a tempo-
ral, Bayesian inference-based, information filter. To compare DDF with our
prediction market-based technique, we replaced our belief aggregation mecha-
nism given in Equation 8 with a DDF-based information filter. We compare
our techniques using some standard evaluation metrics from multi-sensor infor-
mation fusion [13]: root mean squared error (RMSE) defined as in Section 3,
normed mean squared errors (NMSE) calculated as:
N M SEt( Θt − vec(θj)) = 10 log10
j=1 vec(θj ))2 , and, the in-
formation gain, also known as Kullback-Leibler divergence and relative entropy,
calculated as:
Dt
j=1 vec(θj )2)−( 1
j ,vec(θj))2
m (Pm
1
m Pm
m Pm
j=1( Θt
1
KL( Θtvec(θj ) =Pm
j=1
Θt
jlog(cid:18) Θt
vec(θj )(cid:19).
j
Θt was calculated using D-S, DDF, and our prediction market technique ( Θt =
Bt). Since the focus of our work is on the quality of information fusion, we will
concentrate on describing the results for one object. We assume that there are
three types of sensors, MD (least operation cost, most noisy), IR (intermediate
operation cost, moderately noisy), and GPR (expensive operation cost, most ac-
curate). We also assume that there are a total of 5 MD sensors, 3 IR sensors, and
2 GPR sensors available to the decision maker for classifying this object. Ini-
tially, the object is detected using one MD sensor. Once the object is detected,
the time window in the prediction market for identifying the object's type starts.
The MD sensor sends its report to the market maker in the prediction market
and the decision maker makes its first decision based on this one report. We
assume that decision maker's decision (sent to the robot/sensor scheduling algo-
12
Object type
Mine
Metallic or
Friendly for D-S
Non-metallic
Time PM
steps
1
2
3
4
5
6
7
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
1(1MD)
3(1MD,1IR)
4(1GPR)
5(1MD)
6(1MD)
7(1IR)
-
1(1MD)
3(1MD,1IR)
4(1GPR)
5(1MD)
6(1IR)
7(1MD)
8(1IR)
-
1(1MD)
2(1MD)
3(1IR)
4(1MD)
5(1IR)
6(1MD)
-
DDF
D-S
1(1MD)
3(1IR,1GPR)
4(1MD)
5(1MD)
6(1IR)
7(1IR)
8(1MD)
1(1MD)
3(1IR,1GPR)
4(1MD)
5(1IR)
6(11IR)
7(1MD)
8(1MD)
8(1MD)
1(1MD)
3(1MD,1GPR)
5(1MD,1IR)
6(1IR)
7(11MD)
8(1MD)
9(1IR)
1(1MD)
4(1MD,1IR,1GPR)
5(1MD)
6(1IR)
7(1MD)
8(1IR)
9(1GPR)
9(1MD)
1(1MD)
2(1IR)
3(1MD)
4(1GPR)
5(1MD)
6(1IR)
7(1MD)
Table 2: Different number of sensors and the sensor types deployed over time
by a decision maker to classify different types of objects.
rithm in Figure 1) is how many (0 − 3) and what type (MD,IR,GPR) of sensors
to send to the site of the detected object subsequently. We have considered a
set of 14 out of all the possible decisions under this setting. From [11], we derive
four object features, which are metallic content, area of the object, depth of the
object, and the position of the sensor. Combinations of the values of these four
features constitute the signal set G and at each time step, a sensor perceiving
the object receives a signal g ∈ G. The value of the signal also varies based
on the robot/sensor's current position relative to the object. We assume that
the identification of an object stops and the object type is revealed when either
Bt
j ≥ 0.95, for any j, or after 10 time steps. The default values for all domain
related parameters are shown in Table 1. All of our results were averaged over
10 runs and the error bars indicate the standard deviation over the number of
runs.
For our first group of experiments we analyze the performance of our tech-
nique w.r.t. the variables in our model, such as wbel and time, and, w.r.t. to
sensor and object types. We observe that as more information gets sensed for
the object, the RMSE value, shown in Figure 2(a), decreases over time. It takes
on average 6 − 8 time steps to predict the object type with 95% or greater
accuracy depending on the object type and the value of wbel. We also observe
that our model performs the best with wbel = 0.5 (in Equation 2), when the
13
wbel = 0.8
0.45
0.4
0.35
0.3
E
S
M
R
0.25
0.2
0.15
wbel = 0.2
0.1
wbel = 0.5
0.05
0
0
2
8
Number of Time Steps
4
6
35
30
25
20
15
10
5
0
y
t
i
l
i
t
U
e
g
a
r
e
v
A
10
−5
0
MD
GPR
IR
t
s
o
C
2
8
Number of Time Steps
4
6
10
110
100
90
80
70
60
50
40
30
20
10
0
MD
IR
GPR
E
S
M
R
e
g
a
r
e
v
A
0.5
0.4
0.3
0.2
0.1
IR
MD
GPR
Mine
Metallic Object Non−metallic Object
Object Type
0
0
1
3
2
8
Number of Time Steps
4
5
6
7
9
10
a
b
c
d
Figure 2: RMSE for different values of wbel(a), Average sensors' utilities for
different sensor types(b), Cost for different object types(c), RMSE for sensors'
reports averaged over sensor types(d).
agent equally incorporates its private signal and also the market's aggregated
belief at each time step into its own belief update. Figure 2(b) shows the av-
erage utility of the agents based on their type. We can see that MD sensors
get more utility because their costs of calculating and submitting reports are
generally less, whereas GPR sensors get the least utility because they encounter
the highest cost. This result is further verified in Figure 2(c) where we can see
the costs based on sensor types and also based on object types. We observe that
detecting a metallic object that is not a mine has the highest cost. We posit
that it is because both MD and IR sensors can detect metallic content in the
object and extra cost is due to the time and effort spent differentiating metallic
object from a mine. Although most of the mines are metallic [10, 11], we can
see that the cost of detecting a mine and a non-metallic object are similar be-
cause we require a prediction of at least 95%. Due to the sensitive nature of the
landmine detection problem, it is important to ensure that even a non-metallic
object is not a mine even if we encounter higher costs. However, despite MD's
high utility (Figure 2(b)) and low cost (Figure 2(c)), its error of classifying the
object type is the largest, as can be seen from Figure 2(d).
In Table 2, we show how the decision maker's decisions using our prediction
market technique results in the deployment of different numbers and types of
sensors over the time window of the object. We report the results for the value of
belief update weight wbel = 0.5(used in Equation 2) while using our prediction
market model, as well as using D-S and DDF. We see that non-metallic object
classification requires less number of sensors as both MD and IR sensors can
distinguish between metallic vs. non-metallic objects, and so, deploying just
these two types of sensors can help to infer that the object is not a mine. In
contrast, metallic objects require more time to get classified as not being a mine
because more object features using all three sensor types need to be observed.
We also observe that on average our aggregation technique using prediction
market deploys a total of 6 − 8 sensors and detects the object type with at least
95% accuracy in 6 − 7 time steps, while the next best compared DDF technique
deploys a total of 7 − 9 sensors and detects the object type with at least 95%
accuracy in 7 − 8 time steps.
Our results shown in Figure 3(a) illustrate that the RMSE using our PM-
based technique is below the RMSEs using D-S and DDF by an average of 8%
14
E
S
M
R
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0
DDF
PM
D−S
2
8
Number of Time Steps
4
6
5
0
−5
−10
−15
−20
−25
−30
E
S
M
N
10
−35
0
D−S
DDF
PM
PM
0
−0.02
−0.04
−0.06
−0.08
−0.1
−0.12
−0.14
DDF
D−S
i
n
a
g
n
o
i
t
a
m
r
o
f
n
I
2
8
Number of Time Steps
4
6
10
−0.16
0
2
8
Number of Time Steps
4
6
10
a
b
c
Figure 3: Comparison of our Prediction Market-based information aggregation
with Dempster-Shafer and Distributed Data Fusion Techniques using different
metrics: RMSE(a), NMSE(b), Information gain(c).
and 5% respectively. Figure 2(b) shows that the NMSE values using our PM-
based technique is 18% and 23% less on average than D-S and DDF techniques
respectively. Finally, in Figure 2(c) we observe that the information gain for
our PM-based technique is 12% and 17% more than D-S and DDF methods
respectively.
6 Conclusions
In this paper, we have described a sensor information aggregation technique
for object classification with a multi-agent prediction market and developed a
payment function used by the market maker to incentivize truthful revelation
by each agent. Currently, the rewards given by the market maker agent to the
sensor agents are additional side payments incurred by the decision maker. In
the future we plan to investigate a payment function that can achieve budget
balance. We are also interested in integrating our decision making problem with
the problem of scheduling robots(sensors), and, incorporating the costs to the
overall system into the decision-making costs. Another direction we plan to
investigate in the future is a problem of minimizing the time to detect an object
in addition to the accuracy of detection. Lastly, we plan to incorporate our
aggregation technique into the experiments with real robots.
References
[1] Chen Y., Dimitrov S., Sami R., Reeves D.M., Pennock D.M., Hanson R.D.,
Fortnow L., Gonen R. Gaming Prediction Markets: Equilibrium Strategies
with a Market Maker. Algorithmica J., 58(4):930 -- 969, 2009.
[2] Chen Y., Kash I. Information Elicitation for Decision Making. In AAMAS
2011, pages 175 -- 182, 2011.
[3] Cremer F., Schutte K., Schavemaker J.G.M, den Breejen E. A comparison
of decision-level sensor-fusion methods for anti-personnel landmine detec-
tion. Information Fusion, 2(1):187 -- 208, 2001.
15
[4] Gneiting T. and Raftery A.E. Strictly proper scoring rules, prediction, and
estimation. Journal of the American Stat. Assoc., 102(477):359 -- 378, 2007.
[5] Gros B., Bruschini C. Sensor technologies for the detection of antiper-
sonnel mines: a survey of current research and system developments. In
Proceedings of the International Symposium on Measurement and Control
in Robotics, pages 509 -- 518, 1996.
[6] Hanson R. Logarithmic Market Scoring Rules for Modular Combinatorial
Information Aggregation. Journal of Prediction Markets, 1(1):3 -- 15, 2007.
[7] Jumadinova J., Dasgupta P. A Multi-Agent System for Analyzing the Effect
of Information on Prediction Markets. International Journal of Intelligent
Systems, 26(1):383 -- 409, 2011.
[8] Makarenko A., Durrant-Whyte H. Decentralized Bayesian algorithms for
active sensor networks. In Information Fusion 7, pages 418 -- 433, 2006.
[9] Manyika J., Durrant-Whyte H. Data fusion and sensor management. Pren-
tice Hall, 1995.
[10] Milisavljevi`c N., Bloch I. Sensor Fusion in Anti-Personnel Mine Detection
Using a Two-Level Belief Function Model, Part C: Applications and Re-
views. IEEE Transactions on Systems, Man, and Cybernetics, 33(2):269 --
283, 2003.
[11] Milisavljevi`c N., Bloch I., Acheroy M. Characterization of Mine Detection
Sensors in Terms of Belief Functions and their Fusion, First Results. In
Proc. of the Third International Conference on Information Fusion, FU-
SION 2000, pages 15 -- 22, 2000.
[12] Mukherjee P., Sen S. Comparing Reputation Schemes for Detecting Ma-
licious Nodes in Sensor Networks. The Computer Journal, 54(3):482 -- 489,
2011.
[13] Osborne M., Roberts S., Rogers A., Ramchurn S., Jennings N. Towards
realt-time information processing of sensor network data using computa-
tionally efficient multi-output gaussian processes. In IPSN, pages 109 -- 120,
2008.
[14] Othman A., Sandholm T. Decision Rules and Decision Markets. In AAMAS
2010, pages 625 -- 632, 2010.
[15] Rosencrantz M., Gordon G., Thrun S. Decentralized sensor fusion with
distributed particle filters. In IPSN, pages 55 -- 62, 2003.
[16] Vinyals M., Rodriguez-Aguilar J., Cerquides J. A Survey of Sensor Net-
works from a Multiagent Perspective. Computer Journal, 54(3):455 -- 470,
2011.
[17] Waltz E., Llinas J. Multisensor Datafusion. Artech House, 1990.
[18] Wolfers J. and Zitzewitz E. Prediction Markets. Journal of Econ. Perspec-
tives, 18(2):107 -- 126, 2004.
16
[19] Wu Q., Rao N., Barhen J., Iyengar S., Vaishnavi V., Qi H., Chakrabarty K.
On Computing Mobile Agent Routes for Data Fusion in Distributed Sensor
Networks. IEEE Trans. on Knowledge and Data Engineering, 16(6):740 --
753, 2004.
17
|
1805.08535 | 2 | 1805 | 2019-11-15T11:58:59 | Learning over Multitask Graphs -- Part I: Stability Analysis | [
"cs.MA"
] | This paper formulates a multitask optimization problem where agents in the network have individual objectives to meet, or individual parameter vectors to estimate, subject to a smoothness condition over the graph. The smoothness condition softens the transition in the tasks among adjacent nodes and allows incorporating information about the graph structure into the solution of the inference problem. A diffusion strategy is devised that responds to streaming data and employs stochastic approximations in place of actual gradient vectors, which are generally unavailable. The approach relies on minimizing a global cost consisting of the aggregate sum of individual costs regularized by a term that promotes smoothness. We show in this Part I of the work, under conditions on the step-size parameter, that the adaptive strategy induces a contraction mapping and leads to small estimation errors on the order of the small step-size. The results in the accompanying Part II will reveal explicitly the influence of the network topology and the regularization strength on the network performance and will provide insights into the design of effective multitask strategies for distributed inference over networks. | cs.MA | cs | Learning over Multitask Graphs --
Part I: Stability Analysis
Roula Nassif, Member, IEEE, Stefan Vlaski, Member, IEEE,
C´edric Richard, Senior Member, IEEE, Ali H. Sayed, Fellow Member, IEEE
9
1
0
2
v
o
N
5
1
]
A
M
.
s
c
[
2
v
5
3
5
8
0
.
5
0
8
1
:
v
i
X
r
a
Abstract
This paper formulates a multitask optimization problem where agents in the network have individual objectives to
meet, or individual parameter vectors to estimate, subject to a smoothness condition over the graph. The smoothness
condition softens the transition in the tasks among adjacent nodes and allows incorporating information about the
graph structure into the solution of the inference problem. A diffusion strategy is devised that responds to streaming
data and employs stochastic approximations in place of actual gradient vectors, which are generally unavailable.
The approach relies on minimizing a global cost consisting of the aggregate sum of individual costs regularized by
a term that promotes smoothness. We show in this Part I of the work, under conditions on the step-size parameter,
that the adaptive strategy induces a contraction mapping and leads to small estimation errors on the order of the
small step-size. The results in the accompanying Part II will reveal explicitly the influence of the network topology
and the regularization strength on the network performance and will provide insights into the design of effective
multitask strategies for distributed inference over networks.
Index Terms
Multitask distributed inference, diffusion strategy, smoothness prior, graph Laplacian regularization, gradient
noise, stability analysis.
The work of A. H. Sayed was supported in part by NSF grants CCF-1524250 and ECCS-1407712. A short version of this work appeared
in the conference publication [1].
This work was done while R. Nassif was a post-doc at EPFL. She is now with the American University of Beirut, Lebanon (e-
mail: [email protected]). S. Vlaski and A. H. Sayed are with Institute of Electrical Engineering, EPFL, Switzerland (e-mail:
stefan.vlaski,[email protected]). C. Richard is with Universit´e de Nice Sophia-Antipolis, France (e-mail: [email protected]).
2
I. INTRODUCTION
Distributed inference allows a collection of interconnected agents to perform parameter estimation tasks from
streaming data by relying solely on local computations and interactions with immediate neighbors. Most prior
literature focuses on single-task problems, where agents with separable objective functions need to agree on a
common parameter vector corresponding to the minimizer of an aggregate sum of individual costs [2] -- [11]. Many
network applications require more complex models and flexible algorithms than single-task implementations since
their agents may need to estimate and track multiple objectives simultaneously [12] -- [22]. Networks of this kind
are referred to as multitask networks. Although agents may generally have distinct though related tasks to perform,
they may still be able to capitalize on inductive transfer between them to improve their performance.
Based on the type of prior information that may be available about how the tasks are related to each other, multitask
learning algorithms can be derived by translating the prior information into constraints on the parameter vectors
to be inferred [12] -- [22]. For example, in [18] -- [20], distributed strategies are developed under the assumption that
the parameter vectors across the agents overlap partially. A more general scenario is considered in [21] where it is
assumed that the tasks across the agents are locally coupled through linear equality constraints. In [22], the parameter
space is decomposed into two orthogonal subspaces, with one of the subspaces being common to all agents. There
is yet another useful way to model relationships among tasks, namely, to formulate optimization problems with
appropriate regularization terms encoding these relationships [13] -- [17]. For example, the strategy developed in [13]
adds squared (cid:96)2-norm co-regularizers to the mean-square-error criterion to promote task similarities, while the
strategy in [14] adds (cid:96)1-norm co-regularizers to promote piece-wise constant transitions.
In this paper, and the accompanying Part II [23], we consider multitask inference problems where each agent in
the network seeks to minimize an individual cost expressed as the expectation of some loss function. The minimizers
of the individual costs are assumed to vary smoothly on the topology captured by the graph Laplacian matrix. The
smoothness property softens the transition in the tasks among adjacent nodes and allows incorporating information
about the graph structure into the solution of the inference problem. In order to exploit the smoothness prior, we
formulate the inference problem as the minimization of the aggregate sum of individual costs regularized by a term
promoting smoothness, known as the graph-Laplacian regularizer [24], [25]. A diffusion strategy is devised that
responds to streaming data and employs stochastic approximations in place of actual gradient vectors, which are
generally unavailable. We show in this Part I of the work, under conditions on the step-size learning parameter
µ, that the adaptive strategy induces a contraction mapping and that despite gradient noise, it is able to converge
in the mean-square-error sense within O(µ) from the solution of the regularized problem, for sufficiently small
µ. The analysis in the current part also reveals how the regularization strength η can steer the convergence point
of the network toward many modes starting from the non-cooperative mode where each agent converges to the
minimizer of its individual cost and ending with the single-task mode where all agents converge to a common
parameter vector corresponding to the minimizer of the aggregate sum of individual costs. We shall also derive in
3
Part II [23] a closed-form expression for the steady-state network mean-square-error relative to the minimizer of
the regularized cost. This closed form expression will reveal explicitly the influence of the regularization strength,
network topology, gradient noise, and data characteristics, on the network performance. Additionally, a closed-form
expression for the steady-state network mean-square-error relative to the minimizers of the individual costs will be
also derived in Part II [23]. This expression will provide insights into the design of effective multitask strategies
for distributed inference over networks.
There have been many works in the literature studying distributed multitask adaptive strategies and their con-
vergence behavior. Nevertheless, with few exceptions [20], most of these works focus on mean-square-error costs.
This paper, and the accompanying Part II [23], generalize distributed multitask inference over networks and applies
it to a wide class of individual costs. Furthermore, previous works in this domain tend to show the benefit of
multitask learning empirically by simulations. Following some careful and demanding analysis, we establish in Part
II [23], which builds on the results of this Part I, a useful expression for the network steady-state performance. This
expression provides insights into the learning behavior of multitask networks and clarifies how multitask distributed
learning may improve the network performance.
Notation. All vectors are column vectors. Random quantities are denoted in boldface. Matrices are denoted in
capital letters while vectors and scalars are denoted in lower-case letters. The operator (cid:22) denotes an element-wise
inequality; i.e., a (cid:22) b implies that each entry of the vector a is less than or equal to the corresponding entry of
b. The symbol diag{·} forms a matrix from block arguments by placing each block immediately below and to the
right of its predecessor. The operator col{·} stacks the column vector entries on top of each other. The symbol ⊗
denotes the Kronecker product.
II. DISTRIBUTED INFERENCE UNDER SMOOTHNESS PRIORS
A. Problem formulation and adaptive strategy
defined as L = D − A, where the degree matrix D is a diagonal matrix with k-th entry [D]kk =(cid:80)N
We refer to Fig. 1 and consider a connected network (or graph) G = {N ,E, A}, where N is a set of N agents
(nodes), E is a set of edges connecting agents with particular relations, and A is a symmetric, weighted adjacency
matrix. If there is an edge connecting agents k and (cid:96), then [A]k(cid:96) = ak(cid:96) > 0 reflects the strength of the relation
between k and (cid:96); otherwise, [A]k(cid:96) = 0. We introduce the graph Laplacian L, which is a differential operator
(cid:96)=1 ak(cid:96). Since
L is symmetric positive semi-definite, it possesses a complete set of orthonormal eigenvectors. We denote them
by {v1, . . . , vN}. For convenience, we order the set of real, non-negative eigenvalues of L as 0 = λ1 < λ2 ≤
. . . ≤ λN = λmax(L), where, since the network is connected, there is only one zero eigenvalue with corresponding
eigenvector v1 = 1√
1N [26]. Thus, the Laplacian can be decomposed as:
N
where Λ = diag{λ1, . . . , λN} and V = [v1, . . . , vN ].
L = V ΛV (cid:62),
(1)
4
Fig. 1. Agents linked by an edge can share information. The weight ak(cid:96) over an edge reflects the strength of the relation between wo
node k and wo
(cid:96) at node (cid:96).
k at
Let wk ∈ RM denote some parameter vector at agent k and let W = col{w1, . . . , wN} denote the collection of
parameter vectors from across the network. We associate with each agent k a risk function Jk(wk) : RM → R
assumed to be strongly convex. In most learning and adaptation problems, the risk function is expressed as the
expectation of a loss function Qk(·) and is written as Jk(wk) = E Qk(wk; xk), where xk denotes the random data.
The expectation is computed over the distribution of this data. We denote the unique minimizer of Jk(wk) by wo
k.
We introduce a common assumption on the risks {Jk(wk)}. This condition is applicable to many situations of
interest (see, e.g., [7], [10]).
Assumption 1. (Strong convexity) It is assumed that the individual costs Jk(wk) are each twice differentiable and
strongly convex such that the Hessian matrix function Hk(wk) = ∇2
Jk(wk) is uniformly bounded from below
and above, say, as:
wk
0 < λk,minIM ≤ Hk(wk) ≤ λk,maxIM ,
(2)
where λk,min > 0 for k = 1, . . . , N.
In many situations, there is prior information available about Wo = col{wo
N}. In the current work, the
prior belief we want to enforce is that the target signal Wo is smooth with respect to the underlying weighted graph.
References [13] -- [15] provide variations for such problems for the special case of mean-square-error costs. Here
we treat general convex costs. Let L = L ⊗ IM . The smoothness of W can be measured in terms of a quadratic
form of the graph Laplacian [24], [25], [27]:
1, . . . , wo
N(cid:88)
(cid:88)
S(W) = W(cid:62)LW =
1
2
k=1
(cid:96)∈Nk
where Nk is the set of neighbors of k, i.e., the set of nodes connected to agent k by an edge. Figure 1 provides
an illustration. The smaller S(W) is, the smoother the signal W on the graph is. Intuitively, given that the weights
ak(cid:96)(cid:107)wk − w(cid:96)(cid:107)2,
(3)
are non-negative, S(W) shows that W is considered to be smooth if nodes with a large ak(cid:96) on the edge connecting
them have similar weight values {wk, w(cid:96)}. Our objective is to devise and study a strategy that solves the following
regularized problem:
Wo
η = arg minW J glob(W) =
Jk(wk) +
W(cid:62)LW,
η
2
(4)
N(cid:88)
k=1
N,η}.
in a distributed manner where each agent is interested in estimating the k-th sub-vector of Wo
The tuning parameter η ≥ 0 controls the trade-off between the two components of the objective function. Refer-
ence [1] provides a theoretical motivation for the optimization framework where it is shown that, under a Gaussian
η = col{wo
1,η, . . . , wo
Markov random field assumption, solving problem (4) is equivalent to finding a maximum a posteriori (MAP)
estimate for W. We are particularly interested in solving the problem in the stochastic setting when the distribution
of the data xk in Jk(wk) = E Qk(wk; xk) is generally unknown. This means that the risks Jk(wk) and their gradients
∇wkJk(wk) are unknown. As such, approximate gradient vectors need to be employed. A common construction in
stochastic approximation theory is to employ the following approximation at iteration i:
(cid:92)∇wkJk(wk) = ∇wkQk(wk; xk,i),
(5)
where xk,i represents the data observed at iteration i. The difference between the true gradient and its approximation
is called the gradient noise sk,i(·):
sk,i(w) (cid:44) ∇wkJk(w) − (cid:92)∇wkJk(w).
Each agent can employ a stochastic gradient descent update to estimate wo
k,η:
wk,i = wk,i−1 − µ (cid:92)∇wkJk(wk,i−1) − µη
ak(cid:96)(wk,i−1 − w(cid:96),i−1),
(cid:88)
(cid:96)∈Nk
(cid:88)
(cid:96)∈Nk
¯Jk,i−1(wk) (cid:44) Jk(wk) +
η
2
ak(cid:96)(cid:107)wk − w(cid:96),i−1(cid:107)2.
ak(cid:96)(ψk,i − ψ(cid:96),i),
(cid:96)∈Nk
(cid:88)
wk,i = ψk,i − µη
ψk,i = wk,i−1 − µ (cid:92)∇wkJk(wk,i−1)
(cid:88)
ak(cid:96), k = (cid:96)
1 − µη
µηak(cid:96),
ck(cid:96) =
(cid:96)∈Nk
0,
(cid:96) ∈ Nk \ {k}
(cid:96) /∈ Nk
5
(6)
(7)
(8)
(9)
(10)
where µ > 0 is a small step-size parameter. In this implementation, each agent k collects from its neighbors the
estimates w(cid:96),i−1, and performs a stochastic-gradient descent update on:
By introducing an auxiliary variable ψk,i, strategy (7) can be implemented in an incremental manner:
where we replaced (wk,i−1 − w(cid:96),i−1) in the second step by the difference (ψk,i − ψ(cid:96),i) since we expect ψk,i to
be an improved estimate compared to wk,i−1. Note that if we introduce the coefficients:
ψk,i = wk,i−1 − µ (cid:92)∇wkJk(wk,i−1)
(cid:88)
(cid:96)∈Nk
wk,i =
ck(cid:96)ψ(cid:96),i,
(11)
6
then recursion (9) can be written in the diffusion form [6] -- [10]:
where the second step is a combination step. If we collect the scalars {ck(cid:96)} into the matrix C = [ck(cid:96)], then the
entries of C are non-negative for small enough µ and its columns and rows add up to one, i.e., C is a doubly-
stochastic matrix. We shall continue with form (9) because the second step in (9) makes the dependence on η
explicit. We will show later that by varying the value of η we can make the algorithm behave in different ways
from fully non-cooperative to fully single-task with many other modes in between.
B. Summary of main results
Before delving into the study of the learning capabilities of (9) and its performance limits, we summarize in this
section, for the benefit of the reader, the main conclusions of this Part I, and its accompanying Part II [23]. One key
insight that will follow from the detailed analysis in this Part I is that the smoothing parameter η can be regarded
as an effective tuning parameter that controls the nature of the learning process. The value of η can vary from
η = 0 to η → ∞. We will show that at one end, when η = 0, the learning algorithm reduces to a non-cooperative
mode of operation where each agent acts individually and estimates its own local model, wo
k. On the other hand,
when η → ∞, the learning algorithm moves to a single-mode of operation where all agents cooperate to estimate
a single parameter (namely, the Pareto solution of the aggregate cost function). For any values of η in the range
0 < η < ∞, the network behaves in a multitask mode where agents seek their individual models while at the same
time ensuring that these models satisfy certain smoothness and closeness conditions dictated by the value of η. We
are not only interested in a qualitative description of the network behavior. Instead, we would like to characterize
these models in a quantitative manner by deriving expressions that allow us to predict performance as a function
of η and, therefore, fine tune the network to operate in different scenarios.
To begin with, recall that the objective of the multitask strategy (9) is to exploit similarities among neighboring
agents in an attempt to improve the overall network performance in approaching the collection of individual
minimizer Wo by means of local communications. In light of the fact that algorithm (9) has been derived as
an (incremental) gradient descent recursion for the regularized cost (4), whose minimizer Wo
η is in general different
from Wo, the limiting point of algorithm (9) will therefore be generally different from Wo, the actual objective of
the multitask learning problem. This mismatch is the "cost" of enforcing smoothness. The analysis in the paper will
reveal that the mismatch is a function of the similarity between the individual minimizers {wo
k}, of second-order
properties of the individual costs, of the network topology captured by L, and of the regularization strength η.
In particular, future expression (31) will allow us to understand the interplay between these quantities which is
important for the design of effective multitask strategies. The key conclusion will be that, while the bias (difference
η and Wo) will in general increase as the regularization strength η increases, the size of this increase is
between Wo
7
determined by the smoothness of Wo which is in turn function of the network topology captured by L. The more
similar the tasks at neighboring agents are, the smaller the bias will be. This result, while intuitive, is reassuring,
as it implies that as long as Wo is sufficiently smooth, the bias induced by regularization will remain small, even
for moderate regularization strengths η.
The analysis also quantifies the benefit of cooperation, namely, the objective of improving the mean-square
deviation around the limiting point of the algorithm. This analysis is challenging due to coupling among agents,
and the multi-task nature of the learning process (where agents have individual targets but need to meet certain
smoothness and closeness conditions with their neighbors). Section III in this Part I and Sections III and IV in
Part II [23], and the supporting appendices, are devoted to carrying out this analysis in depth leading, for example
to Theorem 1 in Part II [23]. This theorem gives expressions for the mean-square-deviation (MSD) relative to
Wo
η. The expressions reveal the effect of the step-size parameter µ, regularization strength η, network topology,
and data characteristics (captured by the smoothness profile, second-order properties of the costs, and second-
order moments of the gradient noise) on the size of the steady-state mean-square-error performance. The results
established in Theorem 1 and expression (82) in Part II [23] provide tools for characterizing the performance of
multitask strategies in some great detail.
To illustrate the power of these results, consider a connected network where each agent is subjected to streaming
data. The goal at each agent is to estimate a local parameter vector wo
k from the observed data by minimizing
a cost of the form Jk(wk) = E Qk(wk; xk), where xk denotes the random data. Consider network applications
where the minimizers at neighboring agents tend to be similar [13], [25]. Although each agent is interested in
k, cooperating neighboring agents can still benefit from their interactions because of this
closeness. Given the graph Laplacian and data characteristics, one problem of interest would be to determine the
estimating its own task wo
optimal cooperation rule, i.e., the value of η that minimizes the network mean-square-error performance. Future
expression (82) in Part II [23] can be used to solve this problem since it allows us to predict the network MSD
relative to Wo. By using expression (82) in Part II [23], for example, we will be able to construct curves of the form
shown in Fig. 2, which illustrate how performance is dependent on the smoothness parameter η and how the nature
of the limiting solution varies as a function of this parameter. As it can be seen from this figure, η = 4 gives the
best network steady-state mean-square performance. Note that η = 0 corresponds to the non-cooperative scenario
and that a large η induces a large bias in the estimation. In the sequel we will show that as η varies from η = 0 to
η → ∞, the network behavior moves from the non-cooperative mode of operation (where agents act independently)
to the single-task mode of operation (where all agents focus on estimating a single parameter). For values of η
in between, the network can operate in any of a multitude of multitask modes (where agents estimate their own
local parameters under smoothness conditions to allow for some similarity between adjacent nodes). These limits
are indicated in Fig. 2.
Finally, we would like to mention that one of the main tools used in the analysis in this work, and its accompanying
Part II [23], is the linear transformation relative to the eigenspace of the graph Laplacian L from [6], which is also
8
Fig. 2. Network steady-state MSD relative to a smooth signal Wo as a function of the regularization strength η ∈ [0, 350] at µ = 0.005.
known as the graph Fourier transform [25], [28], [29]. Under some conditions on the data and costs profile, we
show in Section VI-A in Part II [23] how the diffusion type algorithm (9) exhibits a low-pass graph filter behavior.
Such filters are commonly used to reduce the network noise profile when the signal to be estimated is smooth with
respect to the underlying topology [25], [30] -- [32]. Interestingly, the theoretical results established in this Part I,
and its accompanying Part II [23], reveal the reasons for performance improvements under localized cooperation.
C. Network limit point and regularization strength
Before examining the behavior and performance of strategy (9) with respect to the limiting point Wo
η in (4), we
η = Wo and strategy (9) reduces to the
discuss the influence of η on Wo
single-agent mode of operation or the non-cooperative solution where each agent minimizes Jk(wk) locally without
cooperation. When η → ∞, we have from (4) that Wo
η. When η = 0, we have from (4) that Wo
η = 1N ⊗ w(cid:63) where
N(cid:88)
k=1
w(cid:63) (cid:44) arg min
w
Jk(w),
(12)
and we are in the single-task mode of operation where all agents seek to estimate a common parameter vector w(cid:63)
corresponding to the minimizer of the aggregate sum of individual costs [6] -- [10]. In order to study more closely
the influence of (finite) η > 0 on the network output Wo
η, we examine the influence of η on the transformed vector:
(cid:44) (V (cid:62) ⊗ IM )Wo
η = col(cid:8)wo
(cid:9)N
(13)
Wo
η
,
m,η
m=1
with the m-th sub-vector wo
η at the m-th eigenvalue λm of the Laplacian:
From (1), the quadratic regularization term S(W) in (3) can be written as:
m,η denoting the spectral content of Wo
m ⊗ IM )Wo
η.
m,η = (v(cid:62)
wo
N(cid:88)
N(cid:88)
S(W) = W(cid:62)LW =
λm(cid:107)wm(cid:107)2 =
λm(cid:107)wm(cid:107)2,
m=1
m=2
(14)
(15)
210-310-210-1100101102103MSDindB-26.5-26-25.5-25-24.5-24-23.5-23-22.5-22X: 0.0035Y: -23.42X: 4.001Y: -25.27X: 350Y: -22.66Non-cooperationSingle-tasklearning 9
m⊗IM )W and where we used the fact that λ1 = 0. Intuitively, given that λm > 0 for m = 2, . . . , N,
where wm = (v(cid:62)
the above expression shows that W is considered to be smooth if (cid:107)wm(cid:107)2 corresponding to large λm is small. As a
result, for a fixed λm > 0, and as the regularization strength η > 0 in (4) increases, one would expect (cid:107)wo
m,η(cid:107)2 to
decrease. Similarly, for a fixed η ≥ 0, and as λm > 0 increases, one would expect (cid:107)wo
m,η(cid:107)2 to decrease as well.
However, as we will see in the sequel, this behavior does not always hold. We show in Section VI-A in Part II [23]
that this is valid when the Hessian matrix function Hk(wk) (cid:44) ∇2
Jk(wk) is independent of wk, i.e., the cost
Jk(wk) is quadratic in wk and is uniform across the network. For more general scenarios, this is not necessarily
the case. What is useful to note, however, is that as η moves from 0 towards ∞, a variety of solution points Wo
can occur ranging from the non-cooperative to the single-task solution at both extremes.
wk
η
From the optimality condition of (4), we have:
col(cid:8)∇wkJk(wo
k,η)(cid:9)N
= −ηLWo
η,
k=1
Using the mean value theorem [33, pp. 24], we can write:
(cid:123)(cid:122)
k,η) − ∇wkJk(wo
k)
(cid:90) 1
∇wkJk(wo
where
(cid:125)
(cid:124)
=0
= H o
k,η(wo
k,η − wo
k),
(cid:110)
(cid:111)N
k=1
Let Ho
η
(cid:44) diag
H o
k,η
(cid:44)
H o
k,η
∇2
wk
Jk(wo
k + t(wo
k,η − wo
k))dt.
0
. Relation (16) can then be rewritten more compactly as:
η =(cid:0)Ho
η + ηL(cid:1)−1 Ho
Wo
ηWo.
(16)
(17)
(18)
(19)
(20)
(21)
(22)
(23)
(24)
Note that the inverse in (19) exists for all η ≥ 0 since the matrix L is positive semi-definite and, under Assumption 1,
the matrix Ho
η is positive definite. Pre-multiplying both sides of the above relation by (V ⊗ IM )(cid:62) gives:
η = (Ho
Wo
η + ηJ )−1Ho
ηWo,
where Wo
η is defined in (13), Wo (cid:44) V(cid:62)Wo, V (cid:44) V ⊗ IM , and
J (cid:44) Λ ⊗ IM ,
ηV.
Ho
(cid:44) V(cid:62)Ho
η
Since L has a single eigenvalue at zero, Λ and V can be partitioned as follows:
Λ = diag{0, Λo}, V = [v1, VR], and V (cid:62) = col{v(cid:62)
1 , V (cid:62)
R }, .
Lemma 1. (Limiting point) Under Assumption 1, it can be shown that Wo
η given by (20) satisfies:
IM Q−1
0
11 Q12
(cid:0)IM (N−1) − K(cid:1)
K
wo
1
[Wo]2:N
,
Wo
η =
10
where wo
1 = (v(cid:62)
1 ⊗ IM )Wo, [Wo]2:N = (V (cid:62)
Q11 (cid:44) (v(cid:62)
R ⊗ IM )Wo and
1 ⊗ IM )Ho
η(v1 ⊗ IM ) =
N(cid:88)
k=1
1
N
H o
k,η,
Q12 (cid:44) (v(cid:62)
Q22 (cid:44) (V (cid:62)
1 ⊗ IM )Ho
R ⊗ IM )Ho
12Q−1
η(VR ⊗ IM ),
η(VR ⊗ IM ) + ηΛo ⊗ IM ,
G (cid:44) (Q22 − Q(cid:62)
11 Q12)−1,
K (cid:44) IM (N−1) − η G (Λo ⊗ IM )
Proof. See Appendix A where we also show that:
(cid:18)
(cid:107)K(cid:107) ≤
(cid:19)(cid:18)
(cid:19)−1
max
1≤k≤N
λk,max
ηλ2(L) + min
1≤k≤N
λk,min
=
O(1)
(O(1) + O(η))
.
(25)
(26)
(27)
(28)
(29)
(30)
(31)
Consider the difference between Wo
To see this, let us subtract Wo from both sides of equation (24). We obtain:
η and Wo. It turns out that the smoother Wo is, the smaller (cid:107)Wo − Wo
η(cid:107) will be.
Q−1
11 Q12
(cid:0)IM (N−1) − K(cid:1)
K − IM (N−1)
[Wo]2:N .
η − Wo =
Wo
The difference Wo
smaller (cid:107)Wo
η − Wo(cid:107) = (cid:107)Wo
η − Wo(cid:107) will be.
η − Wo depends on [Wo]2:N . Thus, from (15) and (31), we conclude that the smoother Wo is, the
Lemma 1 will be useful in the sequel to establish Theorem 1 and to provide a low-pass graph filter interpretation
for the uniform Hessian matrices scenario considered in Section VI-A in Part II [23].
III. NETWORK STABILITY
We examine the behavior of algorithm (9) under Assumption 2 on the gradient noise processes {sk,i(·)} defined
in (6). As explained in [7], [10], these conditions are automatically satisfied in many situations of interest in
learning and adaptation. Condition (32) essentially states that the gradient vector approximation should be unbiased
conditioned on the past data, which is a reasonable condition to require. Condition (33) states that the second-order
moment of the gradient noise process should get smaller for better estimates, since it is bounded by the squared
norm of the iterate. Condition (34) states that the gradient noises across the agents are uncorrelated.
Assumption 2. (Gradient noise process) The gradient noise process defined in (6) satisfies for any w ∈ F i−1 and
for all k, (cid:96) = 1, 2, . . . , N:
E[sk,i(w)F i−1] = 0,
E[(cid:107)sk,i(w)(cid:107)2F i−1] ≤ β2
(cid:96),i(w)F i−1] = 0,
E[sk,i(w)s(cid:62)
k(cid:107)w(cid:107)2 + σ2
s,k,
k (cid:54)= (cid:96),
(32)
(33)
(34)
11
k ≥ 0, σ2
for some β2
all (cid:96) = 1, . . . , N and j ≤ i − 1.
s,k ≥ 0, and where F i−1 denotes the filtration generated by the random processes {w(cid:96),j} for
In this section, we analyze how well the multitask strategy (9) approaches the optimal solution Wo
η of the
regularized cost (4). We examine this performance in terms of the mean-square-error measure, E(cid:107)wo
k,η − wk,i(cid:107)2,
the fourth-order moment, E(cid:107)wo
k,η − wk,i). To establish mean-square
error stability, we extend the energy analysis framework of [6] to handle multitask distributed optimization. Then,
k,η − wk,i(cid:107)4, and the mean-error process, E(wo
following a similar line of reasoning as in [7, Chapter 9], we establish the stability of the first and fourth-order
moments, which is necessary to arrive at an expression for the steady-state performance in Part II [23].
Let us introduce the network block vector Wi = col{w1,i, . . . , wN,i}. At each iteration, we can view (9) as a
mapping from Wi−1 to Wi:
Wi = (IM N − µηL)
(cid:18)
(cid:110) (cid:92)∇wkJk(wk,i−1)
(cid:111)N
k=1
Wi−1 − µ col
(cid:19)
(35)
We introduce the following condition on the combination matrix (IM N − µηL), which is necessary for studying
the performance of (9). It can be easily verified that this requirement is always met by selecting µ and η to satisfy
the bounds (36) -- (37).
Assumption 3. (Combination matrix) The symmetric combination matrix (IM N − µηL) has nonnegative entries
and its spectral radius is equal to one. Since L has an eigenvalue at zero, these conditions are satisfied when the
step-size µ > 0 and the regularization strength η ≥ 0 satisfy:
0 ≤ µη ≤
2
λmax(L)
0 ≤ µη ≤ min
1≤k≤N
,
(cid:40)
1(cid:80)N
(cid:96)=1 ak(cid:96)
(cid:41)
,
(36)
(37)
where condition (36) ensures stability and condition (37) ensures non-negative entries.
A. Stability of Second-Order Error Moment
We first show that algorithm (9), in the absence of gradient noise, converges and has a unique fixed-point. Then,
we analyze the distance between this point and the vectors wo
k,η and wk,i in the mean-square-sense.
1) Existence and uniqueness of fixed-point: Without gradient noise, relation (35) reduces to:
Wi = (IM N − µηL)
Wi−1 − µ col{∇wkJk(wk,i−1)}N
k=1
.
(38)
Let X (cid:44) col{x1, . . . , xN} denote an N × 1 block vector, where xk is M × 1. The mapping (38) is equivalent to
the deterministic mapping X → Y defined as:
Y = (IM N − µηL)
X − µ col{∇wkJk(xk)}N
k=1
.
(39)
(cid:16)
(cid:16)
(cid:17)
(cid:17)
12
Lemma 2. (Contractive mapping) Under Assumption 1 and condition (36), the deterministic mapping defined in (39)
satisfies:
(cid:107)Y1 − Y2(cid:107) ≤ γ(cid:107)X 1 − X 2(cid:107),
with γ (cid:44) max1≤k≤N{γk} where:
γk (cid:44) max{1 − µλk,min,1 − µλk,max}.
This mapping is contractive when µ satisfies:
Proof. See Appendix B.
0 < µ < min
1≤k≤N
(cid:26) 2
λk,max
(cid:27)
.
(40)
(41)
(42)
It then follows from Banach's fixed point theorem [34, pp. 299 -- 303] that iteration (38) converges to a unique
fixed point W∞ = limi→∞ Wi = col{w1,∞, . . . , wN,∞} at an exponential rate given by γ. Observe that this fixed
point is not Wo
η. Since we wish to study lim supi→∞ E(cid:107)Wo
η − Wi(cid:107)2, which can be decomposed as:
E(cid:107)Wo
η − W∞ + W∞ − Wi(cid:107)2
lim sup
i→∞
E(cid:107)Wo
η − Wi(cid:107)2 = lim sup
i→∞
≤ 2(cid:107)Wo
η − W∞(cid:107)2 + 2 lim sup
i→∞
E(cid:107)W∞ − Wi(cid:107)2,
(43)
we shall first asses the size of (cid:107)Wo
η − W∞(cid:107)2 and then examine the quantity lim supi→∞ E(cid:107)W∞ − Wi(cid:107)2.
the step-size µ is small. We carry out the analysis in two steps. First, we derive an expression for (cid:101)W∞ (cid:44) Wo
2) Fixed point bias analysis: Now we analyze how far this fixed point W∞ is from the desired solution Wo
η when
η − W∞
and then we asses its size. Since W∞ is the fixed point of (38), we have at convergence:
Let (cid:101)wk,∞ (cid:44) wo
W∞ = (IM N − µηL)
W∞ − µ col{∇wkJk(wk,∞)}N
k=1
k,η − wk,∞. Using the mean-value theorem [33, pp. 24], [7, Appendix D], we can write:
(cid:17)
∇wkJk(wk,∞) = ∇wkJk(wo
where
Hk,∞ (cid:44)
∇2
wk
Jk(wo
Subtracting the vector (IM N − µηL)Wo
(cid:101)W∞ = (IM N − µηL)(IM N − µH∞)(cid:101)W∞ + µηLWo
k,η) − Hk,∞(cid:101)wk,∞,
k,η − t(cid:101)wk,∞)dt.
η + µ(IM N − µηL)col(cid:8)∇wkJk(wo
k,η)(cid:9)N
k=1
η from both sides of (44) and using relation (45), we obtain:
where H∞ (cid:44) diag{H1,∞, . . . , HN,∞}. From (16), recursion (47) can be written alternatively as:
(cid:101)W∞ = (IM N − µηL)(IM N − µH∞)(cid:101)W∞ + µ2η2L2Wo
(cid:101)W∞ = µ2η2 [IM N − (IM N − µηL)(IM N − µH∞)]
η,
−1 L2Wo
η
so that:
(cid:16)
(cid:90) 1
0
(44)
(45)
(46)
,
(47)
(48)
(49)
The inverse exists when (IM N − µηL)(IM N − µH∞) is stable, i.e., its spectral radius is less than one. Since the
spectral radius of a matrix is upper bounded by any of its induced norms, we have:
ρ((IM N − µηL)(IM N − µH∞)) ≤ (cid:107)IM N − µηL(cid:107)(cid:107)IM N − µH∞(cid:107),
(50)
in terms of the 2−induced norm. Under condition (36) and since λ1(L) = 0, we have (cid:107)IM N − µηL(cid:107) = 1. From
Assumption 1, we have:
13
(1 − µλk,max)IM ≤ IM − µHk,∞ ≤ (1 − µλk,min)IM ,
(51)
so that (cid:107)IM N − µH∞(cid:107)2 ≤ max1≤k≤N γk with γk given in (41). We conclude that when (36) and (42) are satisfied,
the inverse exists.
From (49), we observe that (cid:101)W∞ is zero in two cases: i) when η = 0; ii) when wo
In the second case, consider (24) and observe that wo
1 =
η = (V ⊗ IM )Wo
1 = 1N ⊗ wo and LWo
η = L(1N ⊗ wo) = 0.
Wo
steady-state bias (cid:101)W∞ = Wo
Theorem 1. (Fixed point bias size) Under Assumption 1 and for small µ satisfying conditions (36) and (42), the
N wo, [Wo]2:N = 0, and Wo
k = wo ∀k, i.e., Wo = 1N ⊗ wo.
1, 0}. Thus,
η = col{wo
η = (v1 ⊗ IM )wo
√
η − W∞ of the mapping (38) satisfies:
≤ µ
η − W∞(cid:107)
(cid:107)Wo
(cid:18) 1
(cid:19)
µ lim
µ→0
µ
O(η2)
(O(1) + O(η))2 .
Proof. See Appendix C.
(52)
(53)
(54)
(55)
(56)
(57)
(58)
3) Evolution of the stochastic recursion: We now examine how close the stochastic algorithm (9) approaches
η. First, we introduce the mean-square perturbation vector (MSP) at time i relative to W∞:
Wo
MSPi (cid:44) col(cid:8)E(cid:107)wk,∞ − wk,i(cid:107)2(cid:9)N
k=1 .
The k-th entry of MSPi characterizes how far away the estimate wk,i at agent k and time i is from wk,∞.
Theorem 2. (Network mean-square-error stability) Under Assumptions 1, 2, and 3, the MSP at time i can be
recursively bounded as:
MSPi (cid:22) (IN − µηL) G MSPi−1 + µ2(IN − µηL)b,
where:
G (cid:44) diag(cid:8)γ2
b (cid:44) col(cid:8)σ2
(cid:9)N
k + 3µ2β2
k=1 ,
k
k(cid:107)wo
k,η(cid:107)2 + 3β2
(cid:40)
(cid:40)
s,k + 3β2
k(cid:107)wo
k,η − wk,∞(cid:107)2(cid:9)N
(cid:41)(cid:41)
k=1
.
.
2λk,min
λ2
k,min + 3β2
k
,
2λk,max
λ2
k,max + 3β2
k
A sufficient condition for the stability of the above recursion is:
0 < µ < min
1≤k≤N
min
It follows that
(cid:107) lim sup
i→∞
MSPi(cid:107)∞ = O(µ),
14
and
lim sup
i→∞
E(cid:107)Wo
η − Wi(cid:107)2 = O(µ) +
O(µ2η4)
(O(1) + O(η))4 = O(µ).
(59)
Proof. See Appendix D. With regards to (59) note first that for fixed η, we have O(µ) + O(µ2) = O(µ). When η
and µ are coupled (η = µ−), we obtain:
O(µ) +
O(µ2−4)
O(1) + O(µ−4)
.
For < 0, O(1) dominates O(µ−4) in the denominator and we obtain O(µ) + O(µ2−4) = O(µ). For > 0,
O(µ−4) dominates O(1) in the denominator and we obtain O(µ) + O(µ2) = O(µ).
B. Stability of Fourth-Order Error Moment
The results so far establish that the iterates wk,i converge to a small O(µ)− neighborhood around the regularized
solution wo
k,η. We can be more precise and determine the size of this neighborhood, i.e., assess the size of the
constant multiplying µ in the O(µ)−term. To do so, we shall derive in Part II [23] an accurate first-order expression
for the mean-square error (59); the expression will be accurate to first-order in µ. This expression will be useful
because it will allow us to highlight several features of the limiting point of the network as a function of the
parameter η.
To arrive at the desired expression, we first need to introduce a long-term approximation model and assess how
close it is to the actual model. We then derive the performance for the long-term model and use this closeness
to transform this result into an accurate expression for the performance of the original learning algorithm. When
this argument is concluded we arrive at the desired performance expression, which we then use to comment on
the behavior of the algorithm in a more informed manner. To derive the long-term model, we shall follow the
approach developed in [7]. The first step is to establish the asymptotic stability of the fourth-order moment of the
η − Wi(cid:107)4. This property is needed to justify the validity of the long-term approximate model that
error vector, E(cid:107)Wo
will be introduced in Part II [23].
To establish the fourth-order stability, we replace condition (33) on the gradient noise process by the following
condition on its fourth order moment:
E(cid:2)(cid:107)sk,i(wk)(cid:107)4F i−1
(cid:3) ≤ β
4
k(cid:107)wk(cid:107)4 + σ4
s,k,
(60)
for some β
4
k ≥ 0, and σ4
s,k ≥ 0. As explained in [7], condition (60) implies (33) and, likewise, condition (60) holds
for important cases of interest.
Exploiting the convexity of the norm functions (cid:107)x(cid:107)4 and (cid:107)x(cid:107)2 and using Jensen's inequality, we can write:
E(cid:107)Wo
η − Wi(cid:107)4 ≤ 8(cid:107)Wo
η − W∞(cid:107)4 + 8E(cid:107)W∞ − Wi(cid:107)4,
(61)
and
E(cid:107)W∞ − Wi(cid:107)4 = E(cid:0)(cid:107)W∞ − Wi(cid:107)2(cid:1)2
= E
(cid:33)2
(cid:107)wk,∞ − wk,i(cid:107)2
(cid:32) N(cid:88)
k=1
(cid:33)2
(cid:107)wk,∞ − wk,i(cid:107)2
1
N
= N 2E
(cid:32) N(cid:88)
N(cid:88)
k=1
≤ N
E(cid:107)wk,∞ − wk,i(cid:107)4.
k=1
Let us introduce the mean-fourth perturbation vector at time i relative to W∞:
MFPi (cid:44) col(cid:8)E(cid:107)wk,∞ − wk,i(cid:107)4(cid:9)N
k=1 .
15
(62)
(63)
Theorem 3. (Fourth-order error moment stability) Under Assumptions 1, 2, 3, and condition (60), the MFP at time
i can be recursively bounded as:
where
MFPi (cid:22) (IN − µηL)G(cid:48)MFPi−1 + µ2(IN − µηL)BMSPi−1 + µ4(IN − µηL)b(cid:48),
γ4
k + 24µ2γ2
(cid:110)
kdiag(cid:8)σ2
(cid:110)
(cid:111)N
k + 81µ4β
kβ2
,
k(cid:107)wo
k(cid:107)wo
k,η(cid:107)2 + 3β2
s,k + 3β2
k,η(cid:107)4 + 81β
k(cid:107)wo
k(cid:107)wo
k=1
4
k
4
4
3σ4
s,k + 81β
G(cid:48) (cid:44) diag
B (cid:44) 8γ2
b(cid:48) (cid:44) col
k,η − wk,∞(cid:107)2(cid:9)N
k,η − wk,∞(cid:107)4(cid:111)N
,
.
k=1
k=1
A sufficiently small µ ensures the stability of the above recursion. It follows that
(cid:107) lim sup
i→∞
MFPi(cid:107)∞ = O(µ2),
and
Proof. See Appendix E.
lim sup
i→∞
E(cid:107)Wo
η − Wi(cid:107)4 = O(µ2) +
O(µ4η8)
(O(1) + O(η))8 = O(µ2).
(64)
(65)
(66)
(67)
(68)
(69)
C. Stability of First-order Error Moment
We next need to examine the evolution of the mean-error vector E(Wo
η − Wi). To establish the mean-stability,
we need to introduce a smoothness condition on the Hessian matrices of the individual costs. This smoothness
condition will be adopted in the next Part II [23] when we study the long term behavior of the network.
Assumption 4. (Smoothness condition on individual cost functions). It is assumed that each Jk(wk) satisfies a
smoothness condition close to wo
k,η, in that the corresponding Hessian matrix is Lipchitz continuous in the proximity
of wo
k,η with some parameter κd ≥ 0, i.e.,
(cid:107)∇2
for small perturbations (cid:107)∆wk(cid:107) ≤ .
wk
Jk(wo
k,η + ∆wk) − ∇2
wk
Jk(wo
k,η)(cid:107) ≤ κd(cid:107)∆wk(cid:107),
(70)
16
From the triangle inequality, we have:
(cid:107)E(Wo
η − Wi)(cid:107) ≤ (cid:107)Wo
η − W∞(cid:107) + (cid:107)E(W∞ − Wi)(cid:107).
Let us introduce the square-mean perturbation (SMP) vector at time i relative to W∞:
SMPi (cid:44) col(cid:8)(cid:107)E(wk,∞ − wk,i)(cid:107)2(cid:9)N
k=1 .
(71)
(72)
Theorem 4. (First-order error moment stability) Under Assumptions 1, 2, 3, and 4, the SMP at time i can be
recursively bounded as:
SMPi (cid:22) (IN−µηL)G(cid:48)(cid:48)SMPi−1+µ2(IN−µηL)(IN−G(cid:48)(cid:48))−1B(cid:48)MSPi−1+µ2 1
2
d)2(IN−µηL)(IN−G(cid:48)(cid:48))−1MFPi−1.
(κ(cid:48)
(73)
where
G(cid:48)(cid:48) (cid:44) diag{γk}N
B(cid:48) (cid:44) 2(κ(cid:48)
d)2diag(cid:8)(cid:107)wo
k=1 ,
k,η − wk,∞(cid:107)2(cid:9)N
(75)
}. Under condition (60), a sufficiently small µ ensures the stability of the above
k=1
,
(74)
d = max{κd, λk,max−λk,min
with κ(cid:48)
recursion. It follows that
(cid:107) lim sup
i→∞
SMPi(cid:107)∞ = O(µ2),
lim sup
i→∞
(cid:107)E(Wo
η − Wi)(cid:107) = O(µ) +
O(µη2)
(O(1) + O(η))2 .
(76)
(77)
and that
Proof. See Appendix F.
We have established so far the stability of the mean-error process, E(Wo
η −
η − Wi(cid:107)4. Building on these results, we will derive in Part II [23] closed
Wi(cid:107)2, and the fourth order moment E(cid:107)Wo
form expressions for the steady-state performance of algorithm (9). Section VI in Part II [23] will provide illustration
η − Wi), the mean-square-error E(cid:107)Wo
for the theoretical results in this part (Theorems 1, 2, and 4), and its accompanying Part II.
IV. SIMULATION RESULTS WITH REAL DATASET
In this section, we test algorithm (9) on a weather dataset corresponding to a collection of daily measurements
(mean temperature, mean dew point, mean visibility, mean wind speed, maximum sustained wind speed, and rain
or snow occurrence) taken from 2004 to 2017 at N = 139 weather stations located around the continental United
States [35]. We construct a representation graph G = (N ,E, A) for the stations using geographical distances
between sensors. Each sensor corresponds to a node k and is connected to Nk neighbor nodes with undirected
edges weighted according to ak(cid:96) = 1
2 (pk(cid:96) + p(cid:96)k) with [30]:
(cid:113)(cid:80)
pk(cid:96) =
km(cid:80)
e−d2
e−d2
k(cid:96)
m∈Nk,0
n∈N(cid:96),0
(cid:96) ∈ Nk,0,
,
e−d2
(cid:96)n
(78)
RAIN PREDICTION ERROR (80) IN WEATHER SENSOR NETWORKS FOR DIFFERENT VALUES OF REGULARIZATION STRENGTH η.
TABLE I
η = 0
η = 10
η = 45
η = 100
η = 1000
η = µ−1
prediction error
0.309
0.232
0.225
0.226
0.228
0.232
17
where Nk,0 is the set of 4-nearest neighbors of node k and dk(cid:96) denotes the geodesical distance between the k-th
and (cid:96)-sensors -- see Fig. 3 (left). Let hk,i ∈ RM denote the feature vector at sensor k and day i composed of
M = 5 entries corresponding to the mean temperature, mean dew point, mean visibility, mean wind speed, and
maximum sustained wind speed reported at day i at sensor k. Let γk(i) denote a binary variable associated with
the occurrence of rain (or snow) at node k and day i, i.e, γk(i) = 1 if rain (or snow) occurred and γk(i) = −1
otherwise. We would like to construct a classifier that allows us to predict whether it will rain (or snow) or not based
on the knowledge of the feature vector hk,i. In principle, each station could use an individual logistic regression
machine [7], [36], [37], that seeks a vector wo
k,iwo
k) and
k, such that (cid:98)γk(i) = sign(h(cid:62)
(cid:17)
(cid:16)
1 + e−γ k(i)h(cid:62)
E ln
k,iwk
+ ρ(cid:107)wk(cid:107)2.
wo
k
(cid:44) arg min
wk
(79)
(80)
In this application, however, it is expected that the decision rules {wo
In the experiment, the dataset is split into a training set used to learn the decision rule wo
k} at neighboring stations will be similar.
which (cid:98)γk(i) are generated for performance evaluation. The first dataset comprises daily weather data recorded at
k, and a test set from
the stations in the interval 2004 − 2012 (a total number of Da = 3288 days) and the training set contains data
recorded in the interval 2012 − 2017 (a total number of Dt = 1826 days). We set µ = 3 · 10−4 and ρ = 10−5. We
generate the first iterate wk,0 from the Gaussian distribution N (0, IM ) and we run strategy (9) over the training set
(i = 1, . . . , Da) for different values of η. For each value of η, we report in Table I the prediction error over the
test set defined as:
N(cid:88)
1
N
1
Dt
Dt=1826(cid:88)
k,i(cid:98)wk,∞) (cid:54)= γk(i)],
I[sign(h(cid:62)
k=1
where N = 139 is the number of nodes, (cid:98)wk,∞ is the average of the last 200 iterates generated by the algorithm
at agent k, and I[x] is the indicator function at x, namely, I[x] = 1 if x is true and 0 otherwise. Table I shows
that through cooperation, the agents improve performance. This is due to the fact that the non-cooperative solution
i=1
(η = 0) may suffer from a slow convergence rate [1, Section V-B] in which case some nodes may not be able to
converge in the finite dataset scenario. By increasing η, the convergence rate improves. However, a large value of
η (such as η = µ−1) yields a deterioration in the accuracy since in this case all agents converge approximately to
the same classifier. By setting η = 45, we obtain the smallest prediction error. We show in Fig. 3 (right) the results
of the prediction on July 30, 2015 across the US for η = 45.
18
(Left) Occurrence of rain reported by 139 weather stations across the US on July 30, 2015. (Right) Prediction of rain occurrence
Fig. 3.
from weather data based on logistic regression and multitask learning.
V. CONCLUSION
In this work, we considered multitask inference problems where agents in the network have individual parameter
vectors to estimate subject to a smoothness condition over the graph. Based on diffusion adaptation, we proposed
a strategy that allows the network to minimize a global cost consisting of the aggregate sum of the individual costs
regularized by a term promoting smoothness. We showed that, for small step-size parameter, the network is able
to approach the minimizer of the regularized problem to arbitrarily good accuracy levels. Furthermore, we showed
how the regularization strength can steer the convergence point of the network toward many modes starting from
the non-cooperative mode and ending with the single-task mode.
APPENDIX A
PROOF OF LEMMA 1
Consider the matrix inversion identity [38]:
(A + BCD)−1 = A−1 − A−1B(C−1 + DA−1B)−1DA−1,
which allows us to write:
(U + W )−1U = I − U−1(I + W U−1)−1W = I − (U + W )−1W,
for any invertible matrix U. Using (82), we write (20) alternatively as:
(cid:16)
Wo.
Wo
η =
Let
I − η(cid:0)Ho
η + ηJ(cid:1)−1 J(cid:17)
Q (cid:44) Ho
η + ηJ .
(81)
(82)
(83)
(84)
RainydayDrydayRainydayDryday Using the definitions (21) and (22), we can partition Q into blocks:
Q11 Q12
Q(cid:62)
12 Q22
,
Q =
19
(85)
1
N
with Q11, Q12, and Q22 defined in (25), (26), and (27), respectively. Since v1 = 1√
(cid:80)N
1N , we have Q11 =
k,η which is positive definite from Assumption 1. Observe that Q is invertible since it is similar to
Ho
η + ηL which is positive definite under Assumption 1. Now, by applying the block inversion formula to Q, we
obtain:
k=1 H o
N
Q−1 = (Ho
Q−1
η + ηJ )−1
11 + Q−1
−GQ(cid:62)
11 Q12GQ(cid:62)
12Q−1
11
12Q−1
11 −Q−1
11 Q12G
G
=
where
G (cid:44) (Q22 − Q(cid:62)
12Q−1
11 Q12)−1.
Replacing (86) into (83) and using (21), we arrive at:
ηQ−1
11 Q12G (Λo ⊗ IM )
IM (N−1) − η G (Λo ⊗ IM )
η =
Wo
0
IM
Wo.
,
(86)
(87)
(88)
Using definition (29) into (88), we conclude (24).
Now, we establish (30). Let us first introduce the matrix G(cid:48):
(89)
Using the above definition and expressions (28) and (27), we can re-write the matrix K in (29) alternatively as:
G(cid:48) (cid:44) (V (cid:62)
R ⊗ IM )Ho
η(VR ⊗ IM ) − Q(cid:62)
12Q−1
11 Q12.
K = IM (N−1) − η ((V (cid:62)
R ⊗ IM )Ho
η(VR ⊗ IM ) + ηΛo ⊗ IM − Q(cid:62)
12Q−1
11 Q12)−1 (Λo ⊗ IM )
= IM (N−1) − η(G(cid:48) + ηΛo ⊗ IM )−1(Λo ⊗ IM )
= (G(cid:48) + ηΛo ⊗ IM )−1G(cid:48).
(82)
The matrix G(cid:48) in (89) is the Schur complement of Ho
η in (22) which can be partitioned as:
Q11
Q(cid:62)
12
Ho
η =
.
Q12
R ⊗ IM )Ho
(V (cid:62)
η(VR ⊗ IM )
(90)
(91)
Thus, G(cid:48) is positive definite since it is the Schur complement of the positive definite matrix Ho
G(cid:48) is symmetric, from Weyl's inequality [40, pp. 239] we have:
η [39, pp. 651]. Since
0 < ηλ2(L) + λmin(G(cid:48)) ≤ λmin(G(cid:48) + ηΛo ⊗ IM ) ≤ ηλ2(L) + λmax(G(cid:48)).
Furthermore, since G(cid:48) is the Schur complement of the positive definite matrix Ho
η, we have [41, Theorem 5]:
λmin(G(cid:48)) ≥ λmin(Ho
η) = λmin(Ho
η) ≥ min
1≤k≤N
λk,min,
(92)
(93)
20
λmax(G(cid:48)) ≤ λmax(Ho
η) = λmax(Ho
η) ≤ max
1≤k≤N
λk,max.
Therefore, from (92) and (93), we get:
λmin(G(cid:48) + ηΛo ⊗ IM ) ≥ ηλ2(L) + min
1≤k≤N
λk,min,
and
λmax((G(cid:48) + ηΛo ⊗ IM )−1) =
Since the 2−induced norm of a positive definite matrix is equal to its maximum eigenvalue, we obtain:
λmin(G(cid:48) + ηΛo ⊗ IM )
ηλ2(L) + min
1≤k≤N
(cid:107)(G(cid:48) + ηΛo ⊗ IM )−1(cid:107) ≤
ηλ2(L) + min
1≤k≤N
λk,min
.
From the sub-multiplicative property of the 2−induced norm and from (90), (97), and (94), we obtain:
(cid:107)K(cid:107) ≤ (cid:107)(G(cid:48) + ηΛo ⊗ IM )−1(cid:107) · (cid:107)G(cid:48)(cid:107) ≤
max
1≤k≤N
λk,max
ηλ2(L) + min
1≤k≤N
λk,min
.
(cid:18)
≤
(cid:19)(cid:18)
(cid:19)−1
(cid:19)−1
λk,min
.
(cid:19)−1
1
(cid:18)
(cid:18)
Given any two input vectors X 1 and X 2 with corresponding updated vectors Y1 and Y2, we have from (39):
Y1 − Y2 = (IM N − µηL)
.
(99)
(94)
(95)
(96)
(97)
(98)
(100)
(101)
(102)
APPENDIX B
PROOF OF LEMMA 2
(cid:16)
X 1 − X 2 − µ col(cid:8)∇wkJk(x1
(cid:18)(cid:90) 1
(cid:17)
k)(cid:9)N
k=1
k) − ∇wkJk(x2
(cid:19)
From the mean-value theorem [33, pp. 24], we have:
∇wkJk(x1
k) − ∇wkJk(x2
k) =
∇2
wk
Jk(x2
k + t(x1
k − x2
k))dt
k − x2
(x1
k).
Using (100) into (99), and the sub-multiplicative property of the 2−induced norm [7], we obtain:
(cid:26)
where
We have
D (cid:44) diag
(cid:107)Y1 − Y2(cid:107) ≤ (cid:107)IM N − µηL(cid:107)(cid:107)D(cid:107)(cid:107)X 1 − X 2(cid:107),
IM − µ
∇2
wk
Jk(x2
k + t(x1
k − x2
k))dt
(cid:27)N
.
k=1
(cid:107)IM N − µηL(cid:107) = (cid:107)(IN − µηL) ⊗ IM(cid:107) = (cid:107)IN − µηL(cid:107).
(103)
Let ρ(·) denote the spectral radius of its matrix argument. Since L is symmetric, we have (cid:107)IN−µηL(cid:107) = ρ(IN−µηL).
Since L has one eigenvalue at zero, ρ(IN − µηL) is guaranteed to be equal to 1 if µη satisfies condition (36). For
the block diagonal symmetric matrix D in (102), we have:
∇2
(cid:90) 1
(104)
k − x2
k))dt
Jk(x2
k + t(x1
wk
(cid:107)D(cid:107) = max
1≤k≤N
(cid:13)(cid:13)(cid:13)(cid:13) .
0
Due to Assumption 1, we have:
0 < λk,minIM ≤
Jk(x2
k + t(x1
k − x2
k))dt ≤ λk,maxIM .
(105)
0
(cid:90) 1
0
(cid:13)(cid:13)(cid:13)(cid:13)IM − µ
(cid:90) 1
∇2
wk
0
It follows that (cid:107)D(cid:107) ≤ γ where γ (cid:44) max1≤k≤N{γk} and γk is given in (41). It holds that 0 < γk < 1 when µ is
chosen according to (42). Combining the previous results, we arrive at:
(cid:107)Y1 − Y2(cid:107)2 ≤ γ(cid:107)X 1 − X 2(cid:107)2,
(106)
for γ < 1 when (36) and (42) are satisfied and, in this case, the deterministic mapping (39) is a contraction.
21
APPENDIX C
From (49), we obtain the following expression for (cid:101)W∞:
PROOF OF THEOREM 1
Pre-multiplying both sides of (107) by V(cid:62) = V (cid:62) ⊗ IM gives:
(cid:101)W∞ = µη2[H∞ + ηL − µηLH∞]−1L2Wo
W∞ = µη2(cid:2)H∞ + ηJ − µηJ H∞(cid:3)−1 J 2 Wo
η.
η,
(cid:44) V(cid:62)Wo
η,
H∞ (cid:44) V(cid:62)H∞V,
where W∞ (cid:44) V(cid:62)(cid:101)W∞, Wo
η
and J is given by (21).
In the following we show that W∞ can be written as:
11 P12T
T
W∞ = µη2
−P−1
(Λ2
o ⊗ IM )K[Wo]2:N ,
where K is defined in (29) and:
T (cid:44) (P22 − P21P−1
P11 (cid:44) (v(cid:62)
1 ⊗ IM )H∞(v1 ⊗ IM )=
11 P12)−1,
N(cid:88)
k=1
1
N
Hk,∞,
P12 (cid:44) (v(cid:62)
1 ⊗ IM )H∞(VR ⊗ IM ),
P21 (cid:44) ((IN−1 − µηΛo) ⊗ IM )(V (cid:62)
P22 (cid:44) ηΛo ⊗ IM + ((IN−1 − µηΛo) ⊗ IM )(V (cid:62)
R ⊗ IM )H∞(v1 ⊗ IM ),
R ⊗ IM )H∞(VR ⊗ IM ),
We introduce the following matrix, which appears in (108):
P (cid:44) (IM N − µηJ )H∞ + ηJ =
(107)
(108)
(109)
(110)
(111)
(112)
(113)
(114)
(115)
(116)
where the blocks {Pij} are given by (112) -- (115). Note that, under Assumption 1, P11 in (112) is invertible since
it can be bounded as follows:
0 <
1
N
λk,min
IM ≤ P11 ≤ 1
N
λk,max
IM .
(117)
(cid:33)
(cid:32) N(cid:88)
k=1
P21 P22
P11 P12
(cid:32) N(cid:88)
,
(cid:33)
k=1
Since the Euclidean norm is continuous, we have limµ→0 (cid:107)g(µ)(cid:107) = (cid:107) limµ→0 g(µ)(cid:107). In the following we show that
From (110), we have:
(cid:107)Wo
η − W∞(cid:107)
µ
lim
µ→0
= η2 lim
µ→0
T (Λ2
o ⊗ IM )K[Wo]2:N
and
η4
P−1
11 P12T (Λ2
o ⊗ IM )K[Wo]2:N
µ→0
η4
(cid:13)(cid:13)(cid:13)(cid:13) lim
(cid:13)(cid:13)(cid:13)(cid:13) lim
µ→0
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) .
11 P12T
T
(Λ2
o ⊗ IM )K[Wo]2:N
−P−1
(cid:13)(cid:13)(cid:13)(cid:13)2 ≤ O(η4)(O(1) + O(η))−4,
(cid:13)(cid:13)(cid:13)(cid:13)2 ≤ O(η4)(O(1) + O(η))−4.
(118)
(119)
(120)
(121)
(122)
(123)
(124)
(125)
(126)
22
Applying the block inversion formula to P, we obtain:
P−1
P−1 =
11 + P−1
11 P12T P21P−1
−T P21P−1
11 −P−1
11 P12T
T
11
,
with T defined in (111). Replacing (118) into (108), and using (21) and (24), we conclude (110).
Our goal now is to show that
(cid:107)Wo
η − W∞(cid:107)
µ
lim
µ→0
= c,
for some constant c that may depend on η (the regularization strength), but not on µ (the step-size parameter).
From (120), (121), and (122), we can conclude (52).
Let us first establish (121). We have:
(cid:107)T (Λ2
o ⊗ IM )K[Wo]2:N(cid:107)2 ≤ (cid:107)T (cid:107)2(cid:107)Λ2
o(cid:107)2(cid:107)K(cid:107)2(cid:107)[Wo]2:N(cid:107)2.
From (30), we have (cid:107)K(cid:107)2 ≤ (O(1) + O(η))−2. For sufficiently small step-sizes, we have:
Following the same line of reasoning as in (89) -- (97), we can show that, when µ → 0, we have:
R ⊗ IM )H∞(v1 ⊗ IM ),
P21 = (V (cid:62)
P22 = ηΛo ⊗ IM + (V (cid:62)
lim
µ→0
lim
µ→0
(cid:18)
R ⊗ IM )H∞(VR ⊗ IM ).
(cid:19)−2
= (O(1) + O(η))−2.
(cid:107)T (cid:107)2 ≤
Thus, we conclude (121).
ηλ2(L) + min
1≤k≤N
λk,min
Now, we establish (122). From (117), we have P11 = O(1) and (cid:107)P−1
11 (cid:107)2 = O(1). Similarly, we can conclude
from (113) that (cid:107)P12(cid:107)2 ≤ O(1). Thus, using (121), we arrive at (122).
APPENDIX D
PROOF OF THEOREM 2
From (6), (35), and (44), we have:
W∞−Wi = (IM N−µηL)
(cid:16)
W∞ − Wi−1 − µ col{∇wkJk(wk,∞) − ∇wkJk(wk,i−1)}N
k=1 − µ col{sk,i(wk,i−1)}N
23
(cid:17)
.
k=1
(127)
(132)
(133)
Using the mean-value theorem (100), the above relation can be written as:
W∞ − Wi = (IM N − µηL)(cid:0)(IM N − µHi−1)(W∞ − Wi−1) − µ col{sk,i(wk,i−1)}N
k=1
where Hi−1 (cid:44) diag{H 1,i−1, . . . , H N,i−1} with:
∇2
H k,i−1 (cid:44)
(cid:90) 1
Let
0
Jk(wk,∞ − t(wk,∞ − wk,i−1))dt
wk
(cid:44) (IM N − µHi−1)(W∞ − Wi−1) − µ col{sk,i(wk,i−1)}N
k=1
φi
C (cid:44) IN − µηL.
(cid:1) ,
(128)
(129)
(130)
(131)
From the Laplacian matrix definition, it can be verified that the off-diagonal entries of the matrix C are non-negative
and that its diagonal entries are non-negative under condition (37). Furthermore, since we have L1N = 0, the entries
on each row of C will add up to one. Thus, applying Jensen's inequality [39, pp. 77] to the convex function (cid:107)·(cid:107)2,
we obtain from (128) and (130):
E(cid:107)wk,∞ − wk,i(cid:107)2 ≤ N(cid:88)
[C]k(cid:96)E(cid:107)φ(cid:96),i(cid:107)2,
where φk,i is the k-th sub-vector of φi given by:
(cid:96)=1
φk,i = (IM − µH k,i−1)(wk,∞ − wk,i−1) − µsk,i(wk,i−1).
Squaring both sides of (133), conditioning on F i−1, and taking expectations we obtain:
E[(cid:107)φk,i(cid:107)2F i−1] = (cid:107)wk,∞ − wk,i−1(cid:107)2
(134)
where Σk,i−1 (cid:44) (IM − µH k,i−1)2 and where the cross term is zero because of the zero-mean condition (32). Due
to Assumption 1, Σk,i−1 can be bounded as follows:
+ µ2E[(cid:107)sk,i(wk,i−1)(cid:107)2F i−1].
Σk,i−1
where γk is given by (41). From Assumption 2, E[(cid:107)sk,i(wk,i−1)(cid:107)2F i−1] can be bounded as follows:
0 < Σk,i−1 ≤ γ2
kIM ,
E[(cid:107)sk,i(wk,i−1)(cid:107)2F i−1] ≤ β2
= β2
≤ 3β2
k(cid:107)wk,i−1(cid:107)2 + σ2
k(cid:107)wo
k(cid:107)wo
s,k
k,η − wk,∞ + wk,∞ − wk,i−1 − wo
k,η − wk,∞(cid:107)2 + 3β2
k,η(cid:107)2 + σ2
k(cid:107)wk,∞ − wk,i−1(cid:107)2 + 3β2
s,k
k(cid:107)wo
k,η(cid:107)2 + σ2
s,k.
(135)
(136)
24
Taking expectation again in (134), and using the bounds (135) and (136), we obtain:
E(cid:107)φk,i(cid:107)2 = E(cid:107)wk,∞ − wk,i−1(cid:107)2
k)E(cid:107)wk,∞ − wk,i−1(cid:107)2 + µ2(cid:0)3β2
+ µ2E(cid:107)sk,i(wk,i−1)(cid:107)2
k(cid:107)wo
Σk,i−1
≤ (γ2
k + 3µ2β2
k,η − wk,∞(cid:107)2 + 3β2
k(cid:107)wo
k,η(cid:107)2 + σ2
s,k
Now, combining (137) and (132), we obtain (54).
Iterating (54) starting from i = 1, we get:
MSPi (cid:22) (CG)iMSP0 + µ2
i−1(cid:88)
j=0
(CG)jCb.
(cid:1) .
(137)
(138)
Under Assumption 3 and condition (57), the matrix CG can be guaranteed to be stable. To see this, we upper
bound the spectral radius as follows:
ρ(CG) ≤ (cid:107)CG(cid:107)∞ ≤ (cid:107)C(cid:107)∞(cid:107)G(cid:107)∞ = (cid:107)G(cid:107)∞ = max
1≤k≤N
k + 3µ2β2
γ2
k,
where we used the fact that, under condition (37), the matrix C is a right-stochastic matrix. We have:
k},
k,max + 3µ2β2
k, 1 − 2µλk,max + µ2λ2
k,min + 3µ2β2
γ2
k + 3µ2β2
which is guaranteed to be less than one when:
k = max{1 − 2µλk,min + µ2λ2
(cid:40)
0 < µ < min
(cid:41)
.
2λk,min
λ2
k,min + 3β2
k
,
2λk,max
λ2
k,max + 3β2
k
Then we conclude that the matrix CG is stable under condition (57). In this case, we have:
Using the submultiplicative property of the induced infinity norm, we obtain:
MSPi (cid:22) µ2
(CG)jCb.
lim sup
i→∞
(139)
(140)
(141)
(142)
(143)
(144)
(145)
(146)
j=0
∞(cid:88)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)∞
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞(cid:88)
∞(cid:88)
∞(cid:88)
j=0
j=0
j=0
(cid:107) lim sup
i→∞
MSPi(cid:107)∞ ≤ µ2
≤ µ2
≤ µ2
(CG)j
(cid:107)C(cid:107)∞(cid:107)b(cid:107)∞
(cid:107)(CG)j(cid:107)∞(cid:107)b(cid:107)∞,
(cid:107)C(cid:107)j∞(cid:107)G(cid:107)j∞(cid:107)b(cid:107)∞ =
µ2(cid:107)b(cid:107)∞
1 − (cid:107)G(cid:107)∞
,
where we used the fact that (cid:107)C(cid:107)∞ = 1 and where (cid:107)G(cid:107)∞ = max1≤k≤N γ2
k + 3µ2β2
k. From (140), we have:
γ2
k + 3µ2β2
k = 1 − µζk,
where
Thus,
ζk (cid:44) min{2λk,min − µλ2
k,min − 3µβ2
k, 2λk,max − µλ2
k,max − 3µβ2
k}.
(cid:107)G(cid:107)∞ = max
1≤k≤N
{1 − µζk} = 1 − µ min
1≤k≤N
ζk.
Substituting into (143), we obtain:
(cid:107) lim sup
i→∞
MSPi(cid:107)∞ ≤
µ(cid:107)b(cid:107)∞
min1≤k≤N ζk
.
25
(147)
For sufficiently small µ, we have from (56) and Theorem 1 that (cid:107)b(cid:107) = O(1) + O(µ2η4)(O(1) + O(η))−4. We
conclude that (cid:107) lim supi→∞ MSPi(cid:107)∞≤O(µ).
From (43), we have:
lim sup
i→∞
E(cid:107)Wo
η − Wi(cid:107)2 ≤ 2(cid:107)Wo
η − W∞(cid:107)2 + 2 lim sup
i→∞
N · MSPi
1(cid:62)
(148)
Therefore, from Theorem 1 and (58), we conclude (59).
APPENDIX E
PROOF OF THEOREM 3
Applying Jensen's inequality [39, pp. 77] to the convex function (cid:107) · (cid:107)4, we obtain from (128) and (130):
E(cid:107)wk,∞ − wk,i(cid:107)4 ≤ N(cid:88)
[C]k(cid:96)E(cid:107)φ(cid:96),i(cid:107)4,
where C and φk,i are given by (131) and (133), respectively. Using the inequality [7, pp. 523]:
(cid:96)=1
(cid:107)a + b(cid:107)4 ≤ (cid:107)a(cid:107)4 + 3(cid:107)b(cid:107)4 + 8(cid:107)a(cid:107)2(cid:107)b(cid:107)2 + 4(cid:107)a(cid:107)2(a(cid:62)b),
we obtain from (133) under Assumption 2 on the gradient noise:
E(cid:107)φk,i(cid:107)4 ≤ E(cid:107)(IM − µH k,i−1)(wk,∞ − wk,i−1)(cid:107)4 + 3µ4E(cid:107)sk,i(wk,i−1)(cid:107)4+
8µ2(cid:0)E(cid:107)(IM − µH k,i−1)(wk,∞ − wk,i−1)(cid:107)2(cid:1)(cid:0)E(cid:107)sk,i(wk,i−1)(cid:107)2(cid:1) .
From Assumption 1, the matrices (IM − µH k,i−1)2 and (IM − µH k,i−1)4 can be bounded as follows:
0 < (IM − µH k,i−1)2 ≤ γ2
kIM ,
0 < (IM − µH k,i−1)4 ≤ γ4
kIM ,
(149)
(150)
(151)
(152)
(153)
where γk is given by (41). Thus, we obtain:
E(cid:107)φk,i(cid:107)4 ≤ γ4
k
E(cid:107)wk,∞ − wk,i−1(cid:107)4 + 3µ4E(cid:107)sk,i(wk,i−1)(cid:107)4 + 8µ2γ2
k
(cid:0)E(cid:107)wk,∞ − wk,i−1(cid:107)2(cid:1)(cid:0)E(cid:107)sk,i(wk,i−1)(cid:107)2(cid:1) .
Under condition (60), we have:
E(cid:2)(cid:107)sk,i(wk,i−1)(cid:107)4F i−1
(cid:3) ≤ β
4
k(cid:107)wk,i−1(cid:107)4 + σ4
k(cid:107)wo
4
s,k
= β
≤ 27β
k,η − wk,∞ + wk,∞ − wk,i−1 − wo
k(cid:107)wo
k,η − wk,∞(cid:107)4 + 27β
4
k,η(cid:107)4 + σ4
s,k
4
k(cid:107)wk,∞ − wk,i−1(cid:107)4 + 27β
(154)
(155)
4
k(cid:107)wo
k,η(cid:107)4 + σ4
s,k,
26
where we applied Jensen's inequality to the function (cid:107) · (cid:107)4. Furthermore, from (136), the last term on the RHS
of (154) can be bounded as follows:
(cid:0)E(cid:107)wk,∞ − wk,i−1(cid:107)2(cid:1)(cid:0)E(cid:107)sk,i(wk,i−1)(cid:107)2(cid:1)
(cid:0)E(cid:107)wk,∞ − wk,i−1(cid:107)2(cid:1)2
+(cid:0)3β2
E(cid:107)wk,∞ − wk,i−1(cid:107)4 +(cid:0)3β2
≤ 3β2
≤ 3β2
k(cid:107)wo
k
k
k(cid:107)wo
k,η − wk,∞(cid:107)2 + 3β2
k,η − wk,∞(cid:107)2 + 3β2
k(cid:107)wo
k(cid:107)wo
k,η(cid:107)2 + σ2
k,η(cid:107)2 + σ2
s,k
(cid:1) E(cid:107)wk,∞ − wk,i−1(cid:107)2
(cid:1) E(cid:107)wk,∞ − wk,i−1(cid:107)2,
s,k
where we used the fact that for any random variable a, we have (Ea)2 ≤ Ea2. Replacing (155) and (156) into (154),
we obtain:
(156)
(157)
(158)
E(cid:107)φk,i(cid:107)4 ≤ (γ4
k)E(cid:107)wk,∞ − wk,i−1(cid:107)4+
8µ2γ2
kβ2
4
k + 81µ4β
k + 24µ2γ2
k,η − wk,∞(cid:107)2 + 3β2
k(cid:107)wo
k,η − wk,∞(cid:107)4 + 81µ4β
k(3β2
k(cid:107)wo
4
81µ4β
s,k)E(cid:107)wk,∞ − wk,i−1(cid:107)2+
k(cid:107)wo
k(cid:107)wo
k,η(cid:107)2 + σ2
k,η(cid:107)4 + 3µ4σ4
s,k.
4
Now, combining (157) and (149), we arrive at (64).
Iterating (64) starting from i = 1, we get:
MFPi (cid:22) (CG(cid:48))iMFP0 + µ2
i−1(cid:88)
j=0
(CG(cid:48))jCBMSPi−1−j + µ4
i−1(cid:88)
(CG(cid:48))jCb(cid:48).
j=0
Under Assumption 3 and for sufficiently small µ, the matrix CG(cid:48) can be guaranteed to be stable. To see this, we
upper bound its spectral radius as follows:
ρ(CG(cid:48)) ≤ (cid:107)CG(cid:48)(cid:107)∞ ≤ (cid:107)C(cid:107)∞(cid:107)G(cid:48)(cid:107)∞ = (cid:107)G(cid:48)(cid:107)∞,
(159)
since under condition (37), C is a right-stochastic matrix. The ∞−norm of G(cid:48) is given by:
(cid:107)G(cid:48)(cid:107)∞ = max
1≤k≤N
= max
1≤k≤N
= 1 − µ min
1≤k≤N
4
k
kβ2
k + 81µ4β
γ4
k + 24µ2γ2
(cid:110)
(cid:111)
(cid:8) max{1 − 4µλk,min + 6µ2λ2
k,min − 4µ3λ3
(cid:8) min{4λk,min − 6µλ2
k,max − 4µ3λ3
k,min + 4µ2λ3
1 − 4µλk,max + 6µ2λ2
4λk,max − 6µλ2
k,max + 4µ2λ3
k,min + µ4λ4
k,min + 24µ2γ2
kβ2
k + 81µ4β
k,max + 24µ2γ2
k,max + µ4λ4
k,min − µ3λ4
k,max − µ3λ4
k,min − 24µγ2
kβ2
k,max − 24µγ2
kβ2
4
4
k,
k}(cid:9)
k}(cid:9).
k + 81µ4β
kβ2
k − 81µ3β
4
k,
k − 81µ3β
4
(160)
A sufficiently small µ ensures (cid:107)G(cid:48)(cid:107)∞ < 1 and, thus, ensures the stability of CG(cid:48).
We have established in Theorem 2 that, for small µ, after sufficient iterations have passed, MSPj converges to a
bounded region on the order of µ. This implies that, there exists a jo large enough such that for all j ≥ jo it holds
that:
(cid:107)MSPj(cid:107)∞ ≤ smax = O(µ).
(161)
In this case, we have from (158):
MFPi (cid:22) µ4
lim sup
i→∞
= µ4
= µ4
= µ4
= µ4
j=0
j=0
∞(cid:88)
∞(cid:88)
∞(cid:88)
∞(cid:88)
∞(cid:88)
j=0
j=0
j=0
(CG(cid:48))jCb(cid:48) + µ2 lim sup
i→∞
(CG(cid:48))jCBMSPi−1−j
(CG(cid:48))jCb(cid:48) + µ2 lim sup
i→∞
(CG(cid:48))i−1−jCBMSPj
i−1(cid:88)
(CG(cid:48))jCb(cid:48) + µ2 lim sup
i→∞
(CG(cid:48))jCb(cid:48) + µ2 lim sup
i→∞
(CG(cid:48))jCb(cid:48) + µ2 lim sup
i→∞
(CG(cid:48))i−1−jCBMSPj +
(CG(cid:48))i−1−jCBMSPj
jo(cid:88)
j=jo+1
(CG(cid:48))−1−jCBMSPj +
i−1(cid:88)
j=0
j=jo+1
(CG(cid:48))i−1−jCBMSPj
(CG(cid:48))i−1−jCBMSPj.
(162)
Using the submultiplicative and sub-additive properties of the induced infinity norm, we obtain:
27
j=0
j=0
j=0
i−1(cid:88)
i−1(cid:88)
jo(cid:88)
(CG(cid:48))i
i−1(cid:88)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)∞
j=jo+1
(cid:107) lim sup
i→∞
MFPi(cid:107)∞ ≤ µ4
≤ µ4
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞(cid:88)
∞(cid:88)
j=0
j=0
(CG(cid:48))j
(cid:107)b(cid:48)(cid:107)∞ + µ2 lim sup
i→∞
(cid:107)(CG(cid:48))i−1−jCBMSPj(cid:107)∞
i−1(cid:88)
j=jo+1
i−jo−2(cid:88)
j=0
(cid:107)G(cid:48)(cid:107)j∞(cid:107)B(cid:107)∞smax
(cid:107)G(cid:48)(cid:107)j∞(cid:107)b(cid:48)(cid:107)∞ + µ2 lim sup
i→∞
(cid:107)b(cid:48)(cid:107)∞
+ µ2(cid:107)B(cid:107)∞smax
1 − (cid:107)G(cid:48)(cid:107)∞
= µ4
(163)
where in the second line we used (161) and where (cid:107)G(cid:48)(cid:107)∞ is given by (160). Since (cid:107)b(cid:48)(cid:107)∞ = O(1), (cid:107)B(cid:107)∞ = O(1),
smax = O(µ), and 1 − (cid:107)G(cid:48)(cid:107)∞ = O(µ), we conclude (68).
1 − (cid:107)G(cid:48)(cid:107)∞
From (61) and (62), we have:
lim sup
i→∞
E(cid:107)Wo
η − Wi(cid:107)4 ≤ 8(cid:107)Wo
η − W∞(cid:107)4 + 8N lim sup
i→∞
1N · MFPi.
(164)
Therefore, from Theorem 1 and (68), we conclude (69).
APPENDIX F
PROOF OF THEOREM 4
Conditioning both sides of (128), invoking the conditions on the gradient noise from Assumption 2, and computing
the conditional expectations, we obtain:
E[(W∞ − Wi)F i−1] = (IM N − µηL)(IM N − µHi−1)(W∞ − Wi−1).
Taking expectations again, we arrive at:
E(W∞ − Wi) = (IM N − µηL)E[(IM N − µHi−1)(W∞ − Wi−1)].
Applying Jensen's inequality [39, pp. 77] to the convex function (cid:107) · (cid:107)2, we obtain from the above relation:
(cid:107)E(wk,∞ − wk,i)(cid:107)2 ≤ N(cid:88)
(cid:96)=1
[C]k(cid:96) (cid:107)E [(IM − µH (cid:96),i−1)(w(cid:96),∞ − w(cid:96),i−1)](cid:107)2 ,
(165)
(166)
(167)
28
where C and H k,i−1 are given by (131) and (129), respectively. Let
(cid:102)H k,i−1 (cid:44) Hk,η − H k,i−1,
where
Then, we can write:
Hk,η = ∇2
wk
Jk(wo
k,η).
E [(IM − µH k,i−1)(wk,∞ − wk,i−1)] = (IM − µHk,η)E(wk,∞ − wk,i−1) + µck,i−1,
in terms of a deterministic perturbation sequence defined by
ck,i−1 (cid:44) E[(cid:102)H k,i−1(wk,∞ − wk,i−1)].
By applying Jensen's inequality to the convex function (cid:107) · (cid:107)2, we obtain:
(168)
(169)
(170)
(171)
(cid:13)(cid:13)(cid:13)(cid:13)t
(cid:13)(cid:13)(cid:13)(cid:13) 1
1
t
(cid:107)E(IM − µH k,i−1)(wk,∞ − wk,i−1)(cid:107)2
= (cid:107)(IM − µHk,η)E(wk,∞ − wk,i−1) + µck,i−1(cid:107)2
(IM − µHk,η)E(wk,∞ − wk,i−1) + µ(1 − t)
=
≤ t
(IM − µHk,η)E(wk,∞ − wk,i−1)
t
(cid:107)(IM − µHk,η)E(wk,∞ − wk,i−1)(cid:107)2 + µ2 1
1 − t
1
t
(cid:13)(cid:13)(cid:13)(cid:13)2
ck,i−1
(cid:13)(cid:13)(cid:13)(cid:13)2
1
1 − t
+ µ2(1 − t)
ck,i−1
(cid:13)(cid:13)(cid:13)(cid:13) 1
1 − t
(cid:107)ck,i−1(cid:107)2,
(cid:13)(cid:13)(cid:13)(cid:13)2
=
(172)
for any arbitrary positive number t ∈ (0, 1). We select t = γk where γk is given by (41), which is guaranteed to
be less than one under condition (42). From Assumption 1, we have (cid:107)IM − µHk,η(cid:107)2 ≤ γ2
1
(cid:107)E(IM − µH k,i−1)(wk,∞ − wk,i−1)(cid:107)2 ≤ γk(cid:107)E(wk,∞ − wk,i−1)(cid:107)2 + µ2
k. Thus, we obtain:
(cid:107)ck,i−1(cid:107)2.
(173)
1 − γk
As shown in [7, Appendix E], the Hessian of a twice differentiable strongly convex function Jk(wk) satisfying
Assumptions 1 and 4 is globally Lipschitz relative to wo
∀wk,
(174)
(175)
where κ(cid:48)
k,η, namely, it satisfies:
k,η)(cid:107) ≤ κ(cid:48)
k,η(cid:107),
d(cid:107)wk − wo
Jk(wk) − ∇2
}. Then, for each agent k we obtain:
Jk(wo
wk
wk
(cid:107)(cid:102)H k,i−1(cid:107) (cid:44) (cid:107)Hk,η − H k,i−1(cid:107)
(cid:107)∇2
d = max{κd, λk,max−λk,min
(cid:90) 1
(cid:90) 1
(cid:90) 1
d(cid:107)wo
κ(cid:48)
d(cid:107)wo
κ(cid:48)
k,η − wk,∞(cid:107) +
d(cid:107)wo
≤
= κ(cid:48)
(cid:13)(cid:13)∇2
≤
≤
wk
0
0
0
wk
k,η) − ∇2
Jk(wk,∞ − t(wk,∞ − wk,i−1))(cid:13)(cid:13) dt
Jk(wo
(cid:90) 1
k,η − wk,∞ + t(wk,∞ − wk,i−1)(cid:107)dt
k,η − wk,∞(cid:107)dt +
d(cid:107)wk,∞ − wk,i−1(cid:107),
κ(cid:48)
d(cid:107)t(wk,∞ − wk,i−1)(cid:107)dt
κ(cid:48)
0
1
2
i−1(cid:88)
j=0
i−1(cid:88)
j=0
and, hence,
(cid:107)ck,i−1(cid:107) = (cid:107)E[(cid:102)H k,i−1(wk,∞ − wk,i−1)](cid:107)
≤ E[(cid:107)(cid:102)H k,i−1(cid:107)(cid:107)wk,∞ − wk,i−1(cid:107)]
d(cid:107)wo
≤ κ(cid:48)
k,η − wk,∞(cid:107)E(cid:107)wk,∞ − wk,i−1(cid:107) +
where we used the stochastic version of Jensen's inequality:
29
E(cid:107)wk,∞ − wk,i−1(cid:107)2.
κ(cid:48)
d
1
2
(176)
f (Ea) ≤ E(f (a))
(177)
when f (x) ∈ R is convex. Applying Jensen's inequality to the convex function (cid:107) · (cid:107)2 and using the fact that
(Ea)2 ≤ Ea2 for any real-valued random variable a, we obtain from (176):
k,η − wk,∞(cid:107)2 (E(cid:107)wk,∞ − wk,i−1(cid:107))2 + 2
(κ(cid:48)
1
4
d)2E(cid:107)wk,∞ − wk,i−1(cid:107)4.
k,η − wk,∞(cid:107)2E(cid:107)wk,∞ − wk,i−1(cid:107)2 +
(κ(cid:48)
1
2
d)2(cid:0)E(cid:107)wk,∞ − wk,i−1(cid:107)2(cid:1)2
(cid:107)ck,i−1(cid:107)2 ≤ 2(κ(cid:48)
≤ 2(κ(cid:48)
d)2(cid:107)wo
d)2(cid:107)wo
(178)
From (167) and using the above bound in (173), we conclude (73).
Iterating (73) starting from i = 1, we obtain:
d)2
(κ(cid:48)
(CG(cid:48)(cid:48))jC(I − G(cid:48)(cid:48))−1MFPi−1−j + µ2
SMPi (cid:22) (CG(cid:48)(cid:48))iSMP0 + µ2 1
2
(CG(cid:48)(cid:48))jC(I − G(cid:48)(cid:48))−1B(cid:48)MSPi−1−j.
(179)
Under Assumption 3 and condition (42), the matrix CG(cid:48)(cid:48) is guaranteed to be stable. From (58), (68), and following
similar arguments as the ones used to establish (68) in Appendix E, we conclude that
(cid:107) lim sup
i→∞
SMPi(cid:107)∞ = O(µ2) +
O(µ3η4)
(180)
where we used the fact that (cid:107)B(cid:48)(cid:107)∞ ≤ O(µ2η4)/(O(1) + O(η))4 from Theorem 1 and (cid:107)(IN − G(cid:48)(cid:48))−1(cid:107)∞ ≤ O(µ−1).
(O(1) + O(η))4 = O(µ2),
Using (71) and since (cid:107)E(W∞ − Wi)(cid:107)2 = 1N · SMPi, we conclude (77) from Theorem 1 and (76).
REFERENCES
[1] R. Nassif, S. Vlaski, and A. H. Sayed, "Distributed inference over multitask graphs under smoothness," in Proc. IEEE International
Workshop on Signal Processing Advances in Wireless Communications, Kalamata, Greece, Jun. 2018.
[2] D. P. Bertsekas, "A new class of incremental gradient methods for least squares problems," SIAM J. Optim., vol. 7, no. 4, pp. 913 -- 926,
1997.
[3] R. Olfati-Saber, J. A. Fax, and R. M. Murray, "Consensus and cooperation in networked multi-agent systems," Proc. IEEE, vol. 95,
no. 1, pp. 215 -- 233, 2007.
[4] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione, "Gossip algorithms for distributed signal processing," Proc.
IEEE, vol. 98, no. 11, pp. 1847 -- 1864, 2010.
[5] S. S. Ram, A. Nedi´c, and V. V. Veeravalli, "Distributed stochastic subgradient projection algorithms for convex optimization," J. Optim.
Theory Appl., vol. 147, no. 3, pp. 516 -- 545, 2010.
[6] J. Chen and A. H. Sayed, "Distributed Pareto optimization via diffusion strategies," IEEE J. Sel. Topics Signal Process., vol. 7, no. 2,
pp. 205 -- 220, 2013.
30
[7] A. H. Sayed, "Adaptation, learning, and optimization over networks," Foundations and Trends in Machine Learning, vol. 7, no. 4-5,
pp. 311 -- 801, 2014.
[8] J. Chen and A. H. Sayed, "On the learning behavior of adaptive networks -- Part I: Transient analysis," IEEE Trans. Inf. Theory, vol.
61, no. 6, pp. 3487 -- 3517, Jun. 2015.
[9] J. Chen and A. H. Sayed, "On the learning behavior of adaptive networks -- Part II: Performance analysis," IEEE Trans. Inf. Theory,
vol. 61, no. 6, pp. 3518 -- 3548, Jun. 2015.
[10] A. H. Sayed, "Adaptive networks," Proc. IEEE, vol. 102, no. 4, pp. 460 -- 497, Apr. 2014.
[11] S. Vlaski, L. Vandenberghe, and A. H. Sayed, "Diffusion stochastic optimization with non-smooth regularizers," in Proc. Int. Conf.
Acoust., Speech, Signal Process., Shanghai, China, Mar. 2016, pp. 4149 -- 4153.
[12] J. Plata-Chaves, A. Bertrand, M. Moonen, S. Theodoridis, and A. M. Zoubir, "Heterogeneous and multitask wireless sensor networks
-- Algorithms, applications, and challenges," IEEE J. Sel. Topics Signal Process., vol. 11, no. 3, pp. 450 -- 465, Apr. 2017.
[13] J. Chen, C. Richard, and A. H. Sayed, "Multitask diffusion adaptation over networks," IEEE Trans. Signal Process., vol. 62, no. 16,
pp. 4129 -- 4144, 2014.
[14] R. Nassif, C. Richard, A. Ferrari, and A. H. Sayed, "Proximal multitask learning over networks with sparsity-inducing coregularization,"
IEEE Trans. Signal Process., vol. 64, no. 23, pp. 6329 -- 6344, 2016.
[15] X. Cao and K. J. R. Liu, "Decentralized sparse multitask RLS over networks," IEEE Trans. Signal Process., vol. 65, no. 23, pp.
6217 -- 6232, 2017.
[16] C. Eksin and A. Ribeiro, "Distributed network optimization with heuristic rational agents," IEEE Trans. Signal Process., vol. 60, no.
10, pp. 5396 -- 5411, Oct. 2012.
[17] D. Hallac, J. Leskovec, and S. Boyd, "Network Lasso: Clustering and optimization in large graphs," in Proc. ACM SIGKDD, Sydney,
Australia, Aug. 2015, pp. 387 -- 396.
[18] V. Kekatos and G. B. Giannakis, "Distributed robust power system state estimation," IEEE Trans. Signal Process., vol. 28, no. 2, pp.
1617 -- 1626, 2013.
[19] J. Plata-Chaves, N. Bogdanovi´c, and K. Berberidis, "Distributed diffusion-based LMS for node-specific adaptive parameter estimation,"
IEEE Trans. Signal Process., vol. 63, no. 13, pp. 3448 -- 3460, 2015.
[20] S. A. Alghunaim, K. Yuan, and A. H. Sayed,
"Decentralized exact coupled optimization,"
in Proc. Ann. Allerton Conf. on
Communication, Control, and Computing, Illinois, USA, 2017, pp. 338 -- 345.
[21] R. Nassif, C. Richard, A. Ferrari, and A. H. Sayed, "Diffusion LMS for multitask problems with local linear equality constraints,"
IEEE Trans. Signal Process., vol. 65, no. 19, pp. 4979 -- 4993, 2017.
[22] J. Chen, C. Richard, A. O. Hero, and A. H. Sayed, "Diffusion LMS for multitask problems with overlapping hypothesis subspaces,"
in Proc. IEEE Int. Workshop Mach. Learn. Signal Process., Reims, France, Sep. 2014, IEEE, pp. 1 -- 6.
[23] R. Nassif, S. Vlaski, C. Richard, and A. H. Sayed, "Learning over multitask graphs -- Part II: Performance analysis," Submitted for
publication, Nov. 2019.
[24] D. Zhou and B. Scholkopf, "A regularization framework for learning from graph data," in Proc. ICML Workshop on Statistical
Relational Learning and Its Connections to Other Fields, 2004, vol. 15, pp. 67 -- 68.
[25] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, "The emerging field of signal processing on graphs:
Extending high-dimensional data analysis to networks and other irregular domains," IEEE Signal Process. Mag., vol. 30, no. 3, pp.
83 -- 98, May 2013.
[26] F. R. K. Chung, Spectral Graph Theory, American Mathematical Society, 1997.
[27] K. Q. Weinberger, F. Sha, Q. Zhu, and L. K. Saul, "Graph Laplacian regularization for large-scale semidefinite programming," in
Advances in Neural Information Processing Systems, Vancouver, Canada, Dec. 2007, pp. 1489 -- 1496.
[28] A. Ortega, P. Frossard, J. Kovacevi´c, J. M. F. Moura, and P. Vandergheynst, "Graph signal processing: Overview, challenges, and
applications," Proc. IEEE, vol. 106, no. 5, pp. 808 -- 828, 2018.
31
[29] M. Tsitsvero, S. Barbarossa, and P. Di Lorenzo, "Signals on graphs: Uncertainty principle and sampling," IEEE Trans. Signal Process.,
vol. 64, no. 18, pp. 4845 -- 4860, 2016.
[30] A. Sandryhaila and J. M. F. Moura, "Discrete signal processing on graphs," IEEE Trans. Signal Process., vol. 61, no. 7, pp. 1644 -- 1656,
Apr. 2013.
[31] S. Chen, A. Sandryhaila, J. M. F. Moura, and J. Kovacevic, "Signal denoising on graphs via graph filtering," in Proc. IEEE Glob.
Conf. Signal Information Process.,, Atlanta, GA, USA, Dec. 2014, pp. 872 -- 876.
[32] D. I. Shuman, P. Vandergheynst, and P. Frossard, "Chebyshev polynomial approximation for distributed signal processing," in Proc.
IEEE Int. Conf. Dist. Comp. Sensor Syst., 2011, pp. 1 -- 8.
[33] B. T. Polyak, "Introduction to Optimization," Optimization Software, New York, 1987.
[34] E. Kreyszig, Introductory Functional Analysis with Applications, John Wiley & Sons, 1989.
[35] J. H. Lawrimore, M. J. Menne, B. E. Gleason, C. N. Williams, and D. B. Wuertz and, Global Historical Climatology Network -- Monthly
(GHCN-M), NOAA National Climatic Data Center. Available: ftp:/ftp.ncdc.noaa.gov/pub/data/gsod.
[36] D. W. Hosmer and S. Lemeshow, Applied Logistic Regression, Wiley, NJ, 2nd edition, 2000.
[37] S. Theodoridis and K. Koutroumbas, Pattern Recognition, Academic Press, 4th edition, 2008.
[38] T. Kailath, Linear Systems, Prentice-Hall, Englewood Cliffs, NJ, 1980.
[39] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, NY, 2004.
[40] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University Press, 2nd edition, 2012.
[41] R. L. Smith, "Some interlacing properties of the Schur complement of a Hermitian matrix," Linear Algebra and its Applications, vol.
177, pp. 137 -- 144, 1992.
|
1010.0150 | 1 | 1010 | 2010-10-01T13:03:02 | Implementing Lego Agents Using Jason | [
"cs.MA"
] | Since many of the currently available multi-agent frameworks are generally mostly intended for research, it can be difficult to built multi-agent systems using physical robots. In this report I describe a way to combine the multi-agent framework Jason, an extended version of the agent-oriented programming language AgentSpeak, with Lego robots to address this problem. By extending parts of the Jason reasoning cycle I show how Lego robots are able to complete tasks such as following lines on a floor and communicating to be able to avoid obstacles with minimal amount of coding. The final implementation is a functional extension that is able to built multi-agent systems using Lego agents, however there are some issues that have not been addressed. If the agents are highly dependent on percepts from their sensors, they are required to move quite slowly, because there currently is a high delay in the reasoning cycle, when it is combined with a robot. Overall the system is quite robust and can be used to make simple Lego robots perform tasks of an advanced agent in a multi-agent environment. | cs.MA | cs | Implementing Lego Agents Using Jason
Andreas Schmidt Jensen
1 October 2010
Abstract
Since many of the currently available multi-agent frameworks are
generally mostly intended for research, it can be difficult to built multi-
agent systems using physical robots. In this report I describe a way to
combine the multi-agent framework Jason, an extended version of the
agent-oriented programming language AgentSpeak, with Lego robots
to address this problem. By extending parts of the Jason reasoning cy-
cle I show how Lego robots are able to complete tasks such as following
lines on a floor and communicating to be able to avoid obstacles with
minimal amount of coding. The final implementation is a functional
extension that is able to built multi-agent systems using Lego agents,
however there are some issues that have not been addressed. If the
agents are highly dependent on percepts from their sensors, they are
required to move quite slowly, because there currently is a high delay
in the reasoning cycle, when it is combined with a robot. Overall the
system is quite robust and can be used to make simple Lego robots
perform tasks of an advanced agent in a multi-agent environment.
0
1
0
2
t
c
O
1
]
A
M
.
s
c
[
1
v
0
5
1
0
.
0
1
0
1
:
v
i
X
r
a
1
3
3
5
6
6
6
7
7
8
9
10
11
11
13
14
16
16
18
19
20
20
21
Contents
1 Introduction
2 Jason
2.1 Communication . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Synchronous vs. asynchronous approach . . . . . . . . . . . .
3 Lego Mindstorms NXT
3.1 Sensors
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Sampling data from the sensors . . . . . . . . . . . . .
3.2 Motors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1
4 LeJOS JVM
5 Communication
6 Case studies
6.1 Line Follower . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Communication . . . . . . . . . . . . . . . . . . . . . . . . . .
7 Future thoughts
8 Conclusion
A Extensions of the Jason syntax
A.1 Actions
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.2 Percepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
A.3 Project file
B Case studies: Source code
B.1 Line Follower . . . . . . . . . . . . . . . . . . . . . . . . . . .
B.2 Communication . . . . . . . . . . . . . . . . . . . . . . . . . .
2
1
Introduction
Many of the currently available multi-agent frameworks are purely intended
for research purposes and are as such not very suitable for implementing
agents in physical robots. This paper describes an attempt to built an
extension for one such framework that enables us to use it for implementing
physical robots.
The Jason framework provides an intuitive and easy way to implement
advanced Multi-Agent Systems. Being highly customizable, it enables the
user to easily represent models of artificial and real scenarios. However,
there is currently no link between real world agents (such as robots) and
software agents (implemented in Jason).
The Lego Mindstorms NXT is a robot built in Lego. It includes sev-
eral actuators and sensors, which potentially gives unlimited possibilities
for building advanced robots. It comes with a drag-and-drop programming
interface, which makes it impossible to exploit the potential fully, however
by installing the LeJOS Java Virtual Machine on the NXT, the robot is able
to run programs implemented in Java.
A combination of the Jason framework with Lego robots having LeJOS
installed, could result in an easy and efficient way to implement real world
multi-agent systems.
This paper describes how to implement Lego agents using the Jason
framework. While having considered porting the entire framework to the
LeJOS JVM, it seems more reasonable to extend the Jason framework to
be able to communicate with Lego robots.
To be able to distinguish between the agent in Jason and the physical
lego agent, I write AJason to specify that I am talking about the Jason part
of agent A, and similarly ALego when I am refering to the Lego agent of A.
2 Jason
Instead of changing the general structure of the Jason reasoning cycle the
idea is to extend parts of it, allowing Jason agents to send intentions to the
Lego agent and receive percepts.
The general control loop for a practical reasoning agent, which is some-
what analogous to the Jason reasoning cycle[1], is seen in figure 1. While
the reasoning cycle is a bit more detailed, the similarities are obvious and
it is straightforward to map the steps in the control loop to steps in the
reasoning cycle.
3
Figure 1: The general control loop for a BDI practical reasoning agent[1].
4
The control loop continuously perceives the environment, updates be-
liefs, decides what intention to achieve and looks for a plan to achieve it.
The plan is then executed until it is empty, has succeeded or is impossible.
It should also be noted that the plan is reconsidered often, to ensure that it
is still sound.
To adapt the control loop for real world agents in general, we must
look at what parts of the control loop these agents use to interact with the
environment. Generally, an agent will be able to perceive the environment
and to execute actions. How it perceives and act depends on the sensors
and actuators, but generally speaking, this is what a real world agent is able
to do.
In order to let Lego agents execute Jason plans, the parts of the control
loop concerning perceiving the environment and executing plans must be
extended. In the general control loop, this means that lines 4, 11 and 13
should be extended. All other lines refers to computations regarding the
believes, desires, intentions and plans, all of which are common for agents
in general, and it seems reasonable to let these computations happen within
Jason.
2.1 Communication
Being a framework for Multi-Agent Systems, Jason naturally allows agents
to communicate by sending messages to one another. The Lego agents
should also be able to send messages to one another. There are two possi-
bilities for how this can happen:
• The agents send messages directly to one another. This implies they
have knowledge on how to contact each other (the protocol, address,
etc). A received message should then be interpreted as a percept and
afterwards sent to the Jason engine.
• The agents send messages to the Jason representations of each another.
This only implies knowledge on the name of the agent to contact.
The Jason engine would then be responsible for everything regarding
sending and receiving messages.
The obvious advantage of using the second approach is that no new
functionality needs to be implemented. The agents are able to send messages
using the available internal actions in Jason. In that case, it is already a
part of the AgentSpeak language and the agents will readily be able to react
to the communication.
5
2.2 Synchronous vs. asynchronous approach
One thing to keep in mind is the fact that a noticeable amount of delay
is going to arise, since the agents will have to wait for the engine to send
actions, while the engine will have to wait for the agents to send percepts --
the entire process is synchronous. This means that even though the agents
may be able to quickly execute actions and the engine is able to quickly
perform a reasoning cycle, the overall performance could be very poor.
It may be more efficient to implement a number of queues, which holds a
number of items waiting to be examined and used. The agents holds a queue
of actions to perform, while the engine has a queue of percepts to consider.
The idea is that if an agent has no items in the queue, it must wait. However,
if the engine has not received any percepts, it should keep revising the plans,
using the knowledge it currently has available. Eventually, the agent will
send new percepts, which would let the engine update the plans to conform
with the changes in the environment.
Both approaches have advantages and disadvantages. Using the syn-
chronous approach, the engine and the agents are always up-to-date. No
matter what happens in the environment, the engine will know immediately,
potentially giving rise to a change in the plan. However, as mentioned, the
process is synchronous, meaning that it may prove to perform very poorly.
In the asynchronous approach, on the other hand, the engine contin-
uously updates the plans and lets the agents execute actions, even though
new information is not available yet. This means that the performance could
be much better. However, it gives rise to situations where the agent keeps
following an impossible plan, because the engine did not yet receive any
percepts showing changes in the environment.
3 Lego Mindstorms NXT
The Lego Mindstorms NXT is a toolkit consisting of a set of sensors and
motors, a simple computer and a set of Lego bricks for building robots.
Included is a programming language and editor called NXT-G. It allows
the user to develop simple applications by a drag-and-drop interface, using
loops, branching statements and simple calculations.
3.1 Sensors
The following sensors are included in the Lego Mindstorms NXT kit:
6
• The touch sensor is a sensor with a button. The sensor can detect
whether or not this button is pressed and can for instance be used to
detect whether something has been caught by the arms of a robot.
• The color sensor can detect colors. It is tuned to detect standard
Lego colors and is able to distinguish between these.
• The light sensor is a simpler version of the color sensor, which can
distinguish between light and dark areas and could for instance be
used to follow a black line on a white surface.
• The sound sensor can measure noise level and is also able to recognize
sound patterns and identify tone differences.
• The ultrasonic sensor can measure distances by emitting ultrasonic
sound waves. By sensing the time it takes for these sound waves to
bounce of an object and return to the sensor, it can measure distances
with a precision of 3 centimeters.
Other standard and non-standard sensors, such as older sensors can also
be used by the Lego Mindstorms NXT, but they will not be discussed here.
Refer to appendix A for a description on how to use percepts from these
sensors in Jason.
3.1.1 Sampling data from the sensors
The data provided by the sensors is sent to the extension of the Jason frame-
work, so that AJason is able to use the values. The samples from a sensor
however tend to be somewhat noisy. This means that to successive samples
may be very different, even though the environment has not changed.
The samples are therefore compared to previous values, to see whether
they seem reasonable. An easy way to achieve this is to calculate the me-
dian of the last n values. This should give a more accurate view of the
environment, since extreme values would not be chosen.
3.2 Motors
The kit includes three servo motors, each containing a rotation sensor en-
abling the user to control the movement of a robot very precisely1. It should
also be possible to add motors from the previous version of Lego Mindstorms,
but this will not be considered here.
1http://mindstorms.lego.com/eng/Overview/Interactive_Servo_Motors.aspx
7
A motor can move forward or backwards continuously at a given speed.
It can also be controlled to rotate a number of degrees or even to rotate to
a given degree. The extension to the Jason framework will include features
allowing agents to move forward, backwards and a rotate number of degrees.
This should enable the robots to move around continuously and precisely
rotate the motors to adjust to changes in the environment.
The motor also allows for other possibilities and it is also possible to
get values from the motor, such as its current speed and angle. This could
be used to let the robot increase speed until it has reached some value. I
will not go further into details with this, but only use dedicated sensors to
perceive.
Refer to appendix A for a description on how to use the motors in Jason.
4 LeJOS JVM
The LeJOS JVM is a tiny Java Virtual Machine, which has been ported to
the Lego NXT Brick [3]. Because of the limits of the hardware of an NXT
brick, the LeJOS JVM is only a subset of standard Java. It does contain
many of the standard Java types and classes (lists, input/output support
amongst others), but some classes are not included.
The really useful parts of the LeJOS JVM are the classes giving access to
the sensors and motors included with an NXT brick. The motors and each of
the standard sensors can be controlled using these classes. The functionality
of these classes is what the extension of Jason should utilize to control the
robots.
Since Jason is not going to be compiled and installed on each brick2, the
two systems should communicate in order to give orders and send percepts.
LeJOS includes an API for communicating between a computer and a brick
with LeJOS installed, using either bluetooth or USB.
Since applications are excecuted on the brick, the debugging possibilities
are quite limited. If an exception is thrown and is not caught, the infor-
mation about what caused the exception is very sparse. Basically the user
is given a class- and method number. These numbers can then be used to
track down where an exception was thrown, by looking at information given
by the compiler. LeJOS also gives the possibility of opening a connection to
a remote console. This forwards all output from the brick to the computer,
2Using standard LeJOS JVM Jason cannot even be compiled because of the lack of
many classes.
8
making it a lot easier to debug applications, since the screen of the brick is
quite small.
5 Communication
Communication is a twofold matter, since the system should be able to
handle both communication between AJason and ALego for any agent A as
well as communication between two distinct agents, A and B.
AJason ↔ ALego LeJOS includes a PC API, which enables a computer to
communicate with an NXT brick using either bluetooth or USB. While USB
is faster than bluetooth, it also highly reduce the usefulness of the robot,
since the it will be fastened to the computer by a cable. Using bluetooth,
the agent can move much more freely (within the range of the bluetooth
connection). Since the communication is mostly going to be small messages
containing actions and percepts, a bluetooth connection should not decrease
performance noticeably compared to a USB connection.
Some percepts may be more valuable than other but this may also be
highly dependent on the problem at hand. Therefore there will not be any
sorting in the information perceived, rather will it be send to the computer
running Jason when available. However, the agents will not add the beliefs
immediately to the belief base, since this does not conform with the reason-
ing cycle. Instead, by overriding the perceive() method, beliefs will only
be added at the point in time where the agent is supposed to perceive.
At one step in the reasoning cycle, the agent must perform an action.
In Jason, one can alter the behaviour of acting by overriding the method
act(). Since an actions in this context is something performed by ALego,
the method simply parses the term and sends the corresponding message
to ALego, who will then perform the action immediately. Since LeJOS is a
subset of Java, the terms are parsed and converted into a format that is
more easily parsed by LeJOS.
Note that even though the use of bluetooth should not decrease commu-
nication noticeably, overall slight delays may occur when an action should
be performed. First percepts are received and parsed, then the plan is re-
vised and a new intention may be chosen. This may result in choosing a
new plan, in which an action must be performed. This action is then parsed
and sent to the robot, who performs it.
9
A ↔ B Naturally, the agents in the system must be able to communicate
with each other. The question is whether the agents should communicate
directly using bluetooth, or using the Jason internal actions used for sending
messages to one another.
As discussed in section 2.1 there are obvious advantages of using the
Jason internal actions for communication instead of letting the agents com-
municate directly. One is that the functionality is already implemented and
that by representing each Lego agent as a Jason agent, this functionality is
readily available. Another advantage is seen when comparing the two alter-
natives in the case where one agent sends a percept to another agent. Since
percepts are used in the Jason reasoning cycle to decide upon an intention
and find a plan, percepts sent from agent A to agent B will always end up
in the percept database of BJason.
Step Jason internal actions Direct communication
ALego
AJason
Bluetooth
−−−−−−→ AJason
−−−−−→ BJason
Internal
1
2
3
ALego
BLego
(ALego
Bluetooth
Bluetooth
−−−−−−→ BLego
−−−−−−→ BJason
−−−−−−→ AJason)
Bluetooth
Table 1: Comparison of communication methods when sharing percepts.
Table 1 shows the communication that will happen between the agents
(of both kinds), when a percept is sent. The third message sent during the
direct communication is not exactly relevant for the process it self, however,
whenever an agent ALego perceives something, this is automatically sent to
the corresponding agent AJason.
The differences should be clear. By using direct communication, one
sends three messages using bluetooth, while the use of internal actions en-
ables one to limit the number of bluetooth messages to one. Sending a
message internally in Jason is assumed to be faster than sending a mes-
sage using bluetooth. It is therefore straightforward to conclude that using
internal actions is much more efficient than using bluetooth communication.
6 Case studies
I have developed a few applications entirely in Jason to show the abilities
of the Lego-Jason framework. First I discuss a classical robot example, and
then I try to let two agents cooperate in order to perform a simple task.
10
6.1 Line Follower
The line following robot is basically a robot, which given an environment
with a marked path (i.e. a line), is able to follow exactly that path.
It
requires the use of one or more sensors, which is able to perceive what
is beneath the robot. I have implemented a version using two Lego light
sensors. By putting these sensors close to the floor, the robot is able to see
whether it is moving on a line which is distinguishable from the rest of the
floor (e.g. black line on white floor).
The two sensors are placed so that if the agent is on the line, both sensors
will show bright values. Once the path is turning, one of the sensors will
begin to read darker values, something the agent should react to.
The implementation basically consists of the following three lines (plus
a few rules, which could have been explicitly written in each plan3):
+!move : light(S1, V1) & light(S2, V2)
& distinct_sensors(S1, S2) & on_line(V1, V2)
<- forward([a,b],[60,60]); !!move.
+!move : light(1, V1) & light(2, V2) & turning(V1,V2)
<- rotate([a,b],[-30,15]); !!move.
+!move : light(1, V1) & light(2, V2) & turning(V2,V1)
<- rotate([a,b],[15, -30]); !!move.
The agent has a goal !move, which it at all times will pursue. The first
line says that if it has perceived light values from two distinct sensors, and
they indicate that the robot is on the line, then it should move forward. The
second and third line are the cases where the agent is about to leave the line
and must turn to avoid this. The two cases corresponds to the line turning
left and right, respectively.
It must be noted that in order to successfully follow lines, the robot must
move very slowly forward, so that it is able to react to percepts before it
is too late (i.e. before it has moved completely away from the line). This
imples that communcation is a greater bottleneck than initially suspected.
The implementation assumes that the agent initially is placed on the
line, and the result may otherwise be unpredictable. However, given that
assumption, the entire implementation can be done in three lines.
6.2 Communication
In the second case study I consider two agents: One with a light sensor, an
ultrasonic sensor and two motors -- the "obstacle finder" and one with only
3The entire implementation is included in appendix B
11
a light sensor and two motors -- the "blind agent". The agents are placed in
a very simple environment resembling a pedestrian crossing, i.e. alternating
white and black bars, on which is placed an obstacle that the agents must
avoid when crossing the environment. Both agents know how to avoid the
obstacle, when they find it, but none of them initially know where it is.
The problem is to ensure that both agents successfully cross the path
without touching the obstacle. The obstacle finder is able to detect the
obstacle by using its ultrasonic sensor and can therefore fairly easily avoid
it. The blind agent can only detect how many black bars it has passed. The
idea is therefore to let the obstacle finder send a message containing the
number of black bars to pass, before the agent must avoid the obstacle. This
information can then be used by the blind agent to successfully complete
the task.
This problem is easily solved using Jason when noted that the agents
have a goal to avoid the obstacle and that they should react to changes
in the environment. The implementation is found in appendix B. Below is
shown code snippets that show how the agents cooperate:
Obstacle Finder
+light(_, X)[source(percept)] : goal(search) & on_bar(X) & last_color(white)
<-
-+last_color(black); ?bars_passed(N);
BarsPassed = N + 1; -+bars_passed(BarsPassed).
+light(_, X)[source(percept)]
goal(search) & not on_bar(X) & last_color(black)
-+last_color(white).
+obstacle(_, X)[source(percept)]
:
<-
:
<-
goal(search) & X < 15
-+goal(avoid); ?bars_passed(N);
.send(blindagent, tell, obstacle_after(N)); !!avoid.
Basically, the agent has a counter incrementing each time it moves over
a black bar. When an obstacle is observed near the agent, it tells the blind
agent at which bar the obstacle was found. Then it begins to pursue the
goal !avoid.
Blind Agent
+light(_, X)[source(percept)]
:
<-
goal(search) & on_bar(X) & last_color(white)
-+last_color(black); ?bars_passed(N);
BarsPassed = N + 1; -+bars_passed(BarsPassed).
+light(_, X)[source(percept)]
:
goal(search) & not on_bar(X) & last_color(black)
12
+bars_passed(N) :
<-
<-
-+last_color(white).
obstacle_after(N) & goal(search)
-+goal(avoid); !!avoid.
The agent uses the same approach to count the number of bars passed,
since it also has a light sensor.
It pursues the goal !avoid, whenever it
perceives that it is near the obstacle (i.e. has passed the same number of
bars as the obstacle finder).
The program could also be extended to include the case where the blind
agent does not know how to avoid the obstacle. This can be achieved by
letting the obstacle finder send the plan for avoiding, using the internal
action tellHow.
This case study shows that the agents are able to cooperate and that
messaging is actually working quite efficiently. It must be noted that infor-
mation in the message should be reacted to immediately after reception, so
in other examples, where fast reaction to messages is necessary, results may
not be as promising.
7 Future thoughts
As was shown in the case studies, the system is able to map AJason to ALego.
This means that given an implementation of a system in Jason, a group
of Lego robots would be able to act as the agents implemented in Jason.
However, some parts of the system are not as efficient as they could be, and
there are still plenty of possibilities for extending the system further. In the
following I will discuss some of these issues and possibilities.
Performance It was shown that performance is not optimal, since in order
to give AJason enough time to react on a percept sent by ALego, the latter
is forced to act very slowly. In order to make the system really useful, it
would be necessary to improve performance dramatically.
Navigation The multi-agent case study revealed one obvious problem
with the current solution: the agent has no idea where it is, and has there-
fore little possibility of navigating. Actually being able to have knowledge of
the environment could make it possible for the agent to know where it is and
enable it to navigate from (x1, y1) to (x2, y2) in a world using a navigation
algorithm.
Gaining knowledge of the environment is possible by using the ultrasonic
sensor, since it is able to detect obstacles, such as walls. The problem is that
13
this knowledge cannot directly be used to create a map, since the robot does
not know where it is with regards to that obstacle. It knows that if it has
detected an obstacle x centimeters away, then it is x centimeters from an
obstacle. However, it has no starting point to combine this knowledge with,
since it does not exactly know how far it has moved, before the obstacle was
detected.
A technique called odometry uses data from the movement of the actu-
ators (motors) to estimate how far the robot has moved over time. Since
LeJOS is able to give some information regarding the motors, it may be
possible to use this to generate a world model. In [4] odometry is used to
create such map, and it could be very useful to implement such feature in
Jason and Lego Mindstorms.
By sharing such map between all agents in the environment, it would be
possible for the agents to find and assist each other pursuing their goals.
Sensors The system is currently able to handle the four standard types
of sensors: Touch, light, sound and ultrasonic sensor. The LeJOS JVM
includes support for many other types of sensors, including the sensors from
the first version of Lego Mindstorms. As a future addition, it could be
relevant to add support for these sensors as well.
Actions An action is something performed by ALego. This means that
whenever a term in the implementation is not a Jason primitive or an in-
ternal action, it will be regarded as an action, and therefore be parsed and
sent to the Lego agent.
Section 3 describes the actions that the system is able to perform. The
LeJOS JVM gives a few more possibilities, that could be supported by Jason
as well, by implementing the actions in the parser. However, the big issue is
that the robot only has one type of actuator, the standard NXT motor. By
adding support for other actuators, such as motors with less precision but
with more power, new possibilities could become available.
8 Conclusion
The main goal of this project has been to implement an extension for the
Jason multi-agent system that allows one to write applications executed
by Lego robots. This would enable the programmer to write multi-agent
systems in a logic programming language and even execute these applications
on a real robot.
14
I have examined the Jason reasoning cycle to realize how to communicate
with a robot in terms of sending actions and receiving percepts. This has
resulted in an implementation, which does not alter Jason, but only adds
functionality to allow communication with Lego robots.
The extension to Jason is functional, and enables Lego robots to perform
actions sent from Jason. Furthermore, it was shown that the agent is able
to react to the environment by using sensors. It must be noted that while
the case studies showed that the agents can perform tasks, they cannot do
this fast, since delay must be taken into account when sending percepts,
choosing an intention and sending an action back.
Performance-wise, much could be improved, however the robots are eas-
ily programmed using Jason, since the language naturally expresses how
agents react to the environment and specify how they act to these percepts.
15
A Extensions of the Jason syntax
In this appendix I briefly describe the actions and percepts that can be
used in the LEGO extension of Jason. The design of a Jason-project file
project.mas2j for LEGO projects is also described.
A.1 Actions
forward(Motors, Speed)
Moves a list of motors forward at given speed.
Motors: List of motors.
Speed: List of speed, corresponding to the
motor of the same index in the other list.
Examples:
• forward([a,b],[300,300]): The motors A and B will move forward
at speed 300.
backward(Motors, Speed) Moves a list of motors backward at given
speed.
Motors: List of motors.
Speed: List of speed, corresponding to the
motor of the same index in the other list.
Examples:
• backward([a,b],[300,300]): The motors A and B will move back-
ward at speed 300.
rotate(Motors, Speed)
Rotates a list of motors backward a given
angle.
Motors: List of motors.
Speed: List of angles, corresponding to the
motor of the same index in the other list.
16
Examples:
• rotate([a,b],[300,300]): The motors A and B will rotate 300 de-
grees.
• rotate([a,b],[100,-100]): Motor A will rotate 100 degrees, and
motor B will rotate -100 degrees.
reverse(Motors)
Reverse the direction of a list of motors.
Examples:
Motors: List of motors.
• reverse([a,b]): The motors A and B will reverse their direction.
This happens individually. If A was moving forward and B was moving
backwards, A will now move backwards and B forwards.
speed(Motors,Speed)
Sets the speed for each motor in the list.
Motors: List of motors.
Speed: List of speed, corresponding to the
motor of the same index in the other list.
Examples:
• speed([a,b],[200,300]): Sets the speed of A to 200 and the speed
of B to 300.
stop(Motors)
Stops the motors in the list.
Examples:
• stop([a,b]): Stops motor A and B.
Motors: List of motors.
block(Blocking)
Sets whether the rotation command should
return control to the thread immediately.
Blocking: Boolean value specifying whether
rotation is blocking or control should be re-
turn immediately.
17
Examples:
• block(true): Blocks further actions of the robot until a rotation is
done. This has no immediate effect, but affects how rotate(Motors,
Speed) works.
Tells the (Lego) agent to shut down.
exit
Examples:
• exit
A.2 Percepts
light(Port, Value)
Percept from a light sensor.
Port: Port on which the sensor is connected.
Can be used to distinguish between several
sensors of the same kind.
Value: Light value read from the sensor.
Examples:
• light(1,360): The light sensor on port 1 has read a value of 360.
obstacle(Port, Value)
Percept from an ultrasonic sensor.
Port: Port on which the sensor is connected.
Can be used to distinguish between several
sensors of the same kind.
Value: Distance to obstacle read from the
sensor (max = 255).
Examples:
• obstacle(1,40): The ultrasonic sensor on port 1 has found an obsta-
cle 40 centimeters away.
touching(Port, Value) Percept from a touch sensor.
18
Port: Port on which the sensor is connected.
Can be used to distinguish between several
sensors of the same kind.
Value: The value is either true or false, cor-
responding to whether or not something is
touching the button of the sensor.
Examples:
• touching(1, true): The touch sensor on port 1 has detected a touch.
• touching(1, false): The touch sensor on port 1 has not detected a
touch.
sound(Port, Value)
Percept from a sound sensor.
Port: Port on which the sensor is connected.
Can be used to distinguish between several
sensors of the same kind.
Value: Sound value read from the sensor.
Examples:
• sound(1,55): The sound sensor on port 1 has read a value of 55.
A.3 Project file
I will not go into details with how a general project file looks. I will only
describe which classes an agent must use for their architecture and belief
base, and which parameters are available and must be specified, to allow
Jason to connect the the Lego robots.
An agent in the file project.mas2j must look like this:
agentname agentsource.asl
[btname="BTNAME", btaddress="BTADDRESS",
motora="t/f", motorb="t/f", motorc="t/f",
sensor1="...", sensor2="...", sensor3="...",
sensor4="...", sleep="..."]
agentArchClass arch.LEGOAgArchitecture
beliefBaseClass agent.UniqueBelsBB("light(port,_)","sound(port,_)",
"obstacle(port,_)","touching(port,_)");
The agentArchClass and beliefBaseClass must be set to the specified
classes. The agent architecture takes care of all communication between
19
AJason and ALego. The belief base makes sure that at any time, we can
have at most one percept from a single sensor. This is done to make sure
that the belief base does not increase exponentially in size. Below is shown
the parameters that can be passed to the agent. Parameters in bold are
mandatory.
Key
btname
btaddress
motor{a,b,c}
sensor{1,2,3,4}
sleep
Value
Name of the NXT brick
Bluetooth address in the format 12:34:56:78:90:AB
true or false, depending on whether the motor is con-
nected to the brick.
the name of the sensor connected or "none". The
name can be one of {touch,light,sound,ultrasonic}.
Milliseconds a sensor should sleep between sampling.
Default: 50 ms.
B Case studies: Source code
B.1 Line Follower
linefollower.asl
// Linefollower agent for the Lego Mindstorms NXT
// Uses two sensors to detect whether it is on the line.
// The agent assumes that it is initially on a line
// (i.e. both sensors are showing bright light).
// It may therefore not work if not placed on a line
/* Initial beliefs and rules */
// the sensors must be different, otherwise the light value
// from one sensor will be compared to it self
distinct_sensors(Sensor1, Sensor2) :- Sensor1 \== Sensor2.
// the robot is on the line if both sensors are showing a
// bright value (currently a value greater than 350)
on_line(Value1, Value2) :- Value1 >= 350 & Value2 >= 350.
// the robot should turn if one value is dark and the other is bright
// (i.e. the line is turning)
turning(Value1, Value2) :- Value1 < 350 & Value2 >= 350.
/* Initial goals */
!move.
/* Plans */
// move forward if both sensors are sampling bright values
20
+!move : light(S1, V1) & light(S2, V2)
& distinct_sensors(S1, S2) & on_line(V1, V2)
<- forward([a,b],[60,60]); !!move.
// turning
+!move : light(1, V1) & light(2, V2) & turning(V1,V2)
<- rotate([a,b],[-30,15]); !!move.
+!move : light(1, V1) & light(2, V2) & turning(V2,V1)
<- rotate([a,b],[15, -30]); !!move.
// if nothing has been perceived yet, wait, then reintroduce goal.
+!move : not light(_,_)
<- .wait("+light(_,_)"); !move.
// upon failure, start from scratch and reintroduce.
-!move <-.drop_all_desires; .abolish(light(_,_)); !!move.
B.2 Communication
obstaclefinder.asl
// Agent for finding obstacles and avoiding them.
// When an obstacle is found, the agent sends information
// to a blind agent about how to avoid the obstacle.
/* Initial beliefs and rules */
// How many bars have the agent currently passed
bars_passed(0).
// Assume we alternate between black and white bars
last_color(white).
// intially the agent searches for an obstacle
goal(search).
// The agent is on a bar if the sensors is reading a
// value less than 350
on_bar(Value) :- Value < 350.
/* Initial goals */
!init.
/* Plans */
+!init
<-
block(true); !!move.
// Increment bar-counter when on a black bar and last bar was white.
// Also make mental note that the agent is on a black bar now.
+light(_, X)[source(percept)]
:
goal(search) & on_bar(X) & last_color(white)
21
<-
-+last_color(black); ?bars_passed(N);
BarsPassed = N + 1; -+bars_passed(BarsPassed).
// Make mental note that now the agent is on a white bar.
+light(_, X)[source(percept)]
:
<-
goal(search) & not on_bar(X) & last_color(black)
-+last_color(white).
// when an obstacle is near -- tell the blind agent and avoid!
+obstacle(_, X)[source(percept)]
:
<-
goal(search) & X < 15
-+goal(avoid); ?bars_passed(N);
.send(blindagent, tell, obstacle_after(N));
!!avoid.
// Avoiding the obstacle. It is assumed that an obstacle
// can only be of one size.
+!avoid <- stop([a,b]); speed([a,b],[300,300]);
rotate([a,b],[-200,200]); rotate([a,b],[400,400]);
rotate([a,b],[200,-200]); rotate([a,b],[800,800]);
rotate([a,b],[200,-200]); rotate([a,b],[400,400]);
rotate([a,b],[-200,200]); !!move.
// approach the obstacle slowly.
+!move <-
forward([a,b],[60,60]).
blindagent.asl
// Blind agent that can avoid obstacles,
// when it is told where they are.
/* Initial beliefs and rules */
// How many bars have the agent currently passed
bars_passed(0).
// Assume we alternate between black and white bars
last_color(white).
// intially the agent searches for an obstacle
goal(search).
// The agent is on a bar if the sensors is reading a value less than 350
on_bar(Value) :- Value < 400.
/* Initial goals */
!init.
/* Plans */
+!init
<-
block(true); !!move.
// Increment bar-counter when on a black bar and last bar was white.
// Also make mental note that the agent is on a black bar now.
22
+light(_, X)[source(percept)]
:
<-
goal(search) & on_bar(X) & last_color(white)
-+last_color(black); ?bars_passed(N);
BarsPassed = N + 1; -+bars_passed(BarsPassed).
// Make mental note that now the agent is on a white bar.
+light(_, X)[source(percept)]
:
<-
goal(search) & not on_bar(X) & last_color(black)
-+last_color(white).
// when the amount of bars passed means the obstacle is near -- avoid!
+bars_passed(N)
:
<-
obstacle_after(N) & goal(search)
-+goal(avoid); !!avoid.
// Avoiding the obstacle. It is assumed that an obstacle can only be of one size.
+!avoid <- stop([a,b]); speed([a,b],[300,300]);
rotate([a,b],[-200,200]); rotate([a,b],[400,400]);
rotate([a,b],[200,-200]); rotate([a,b],[800,800]);
rotate([a,b],[200,-200]); rotate([a,b],[400,400]);
rotate([a,b],[-200,200]); !!move.
// approach the obstacle slowly.
+!move <-
forward([a,b],[60,60]).
23
References
[1] Rafael H. Bordini, Jomi Fred Hubner, and Michael Wooldridge. Pro-
gramming Multi-Agent Systems in AgentSpeak Using Jason. John Wiley
& Sons, 2007.
[2] Lin Padgham and Michael Winikoff. Developing Intelligent Agent Sys-
tems: A Practical Guide. John Wiley & Sons, 2007.
[3] LeJOS, Java for Lego Mindstorms. http://lejos.sourceforge.net/.
Last checked: September, 2010.
[4] Hagit Shatkay and Leslie Pack Kaelbling. Learning Topological Maps
with Weak Local Odometric Information. In Proceedings of IJCAI-97.
IJCAI, Inc, 2007.
Acknowledgement
I thank associate professor Jørgen Villadsen, DTU Informatics, Denmark.
Source code
The source code for the system is available here:
http://imm.dtu.dk/~jv/LEGO-Jason-NXT
24
|
cs/9809025 | 1 | 9809 | 1998-09-18T00:07:03 | Novelty and Social Search in the World Wide Web | [
"cs.MA",
"cs.DL"
] | The World Wide Web is fast becoming a source of information for a large part of the world's population. Because of its sheer size and complexity users often resort to recommendations from others to decide which sites to visit. We present a dynamical theory of recommendations which predicts site visits by users of the World Wide Web. We show that it leads to a universal power law for the number of users that visit given sites over periods of time, with an exponent related to the rate at which users discover new sites on their own. An extensive empirical study of user behavior in the Web that we conducted confirms the existence of this law of influence while yielding bounds on the rate of novelty encountered by users. | cs.MA | cs | Novelty and Social Search in the World Wide Web
Bernardo A. Huberman and Lada A. Adamic
Xerox Palo Alto Research Center
3333 Coyote Hill Road
Palo Alto, CA 94304
Abstract
The World Wide Web is fast becoming a source of information for a large part of the world’s
population. Because of its sheer size and complexity users often resort to recommendations from
others to decide which sites to visit. We present a dynamical theory of recommendations which
predicts site visits by users of the World Wide Web. We show that it leads to a universal power
law for the number of users that visit given sites over periods of time, with an exponent related to
the rate at which users discover new sites on their own. An extensive empirical study of user
behavior in the Web that we conducted confirms the existence of this law of influence while
yielding bounds on the rate of novelty encountered by users.
1
The World Wide Web (Web) has become in a very short period one of the most useful sources of
information for a large part of the world’s population. Its exponential growth, from a few sites in
1994 to millions today, has transformed it into an ecology of knowledge in which highly diverse
information is linked in extremely complex and arbitrary fashion(1). In addition, the rapid growth
of electronic commerce is bringing about a further increase in the amount of information
available on the Web.
A problem with this bounty is that it is often hard for users to choose which sites to visit, if only
because it is difficult to judge a priori their information value. In addition, since providers of
electronic commerce sometimes lack recognizable reputations and can offer similar services, it is
seldom possible to make optimal decisions as to which sites to access and which ones to avoid.
As with many other situations were choice is costly, people resort to a cooperative mechanism
which relies on the collective search performed by a whole community to find desirable and
useful sites. This is an extremely successful strategy, which has been shown to lead to faster
problem solving(2), increased genetic variability among certain species(3) and speedier diffusion
of innovations in many social settings(4).
In the case of the Web, social search can also solve the problem posed by its explosive growth,
for it is impossible for a single individual or the best search engines to keep track of all the newly
created sites and the informational value that they contribute. Large groups of people surfing on
their own can sample a much larger information space than single individuals, and any exchange
of relevant findings can increase the awareness of possibly interesting sites. Even though
recommendations, both personal and institutional(5-7), can be unreliable and highly idiosyncratic,
they decrease the cost of searching for optimal sources of information, while leading to the
discovery of new sites and improved ways of surfing the Web.
This social search mechanism is particularly useful when looking for useful information, and
should be contrasted with surfing, in which users follow hyperlinks as long as they find value in
them(8). Users typically use recommendations to choose a starting site, while their activity within
the site is determined by their surfing behavior. Thus one expects that the mechanism of social
search will be manifested in the statistics of site visits of the Web.
In what follows we study the dynamics of recommendations and their effect on site visits by users
of the World Wide Web. We show that this form of social search leads to a universal power law
for the number of users that visit given sites over periods of time. The exponent of the power law
is related to the rate at which users discover new sites on their own, as opposed to following
recommendations. An extensive empirical study of user behavior in the Web that we conducted
confirms the existence of this law of influence while yielding bounds on the rate of novelty
encountered by users.
Consider the familiar situation of Internet surfers having to choose among a large number of Web
sites. Given the unstructured character of the Web, users will either surf on their own or follow
recommendations from others
these
the nature of
that
to visit useful sites. Notice
recommendations can range from messages suggesting visits to particular sites, to brand
recognition and links to other sites that appear in pages that individuals are visiting. If people in
turn find value in those sites, they might recommend them to others. Recommendations will thus
lead individuals to preferentially choose those sites that have been already selected by other
individuals. This probabilistic process determines the number of users that visit given sites. In
addition, there is an important novelty factor, since users can visit sites that have not been
previously recommended by using search engines that do not explicitly rank their links so as to
bias users to particular sites. This is an important discovery mechanism for newly created sites
that have either few visitors or small number of links to them.
2
In order to understand the dynamics of site visits as a function of both site discovery and
recommendations, it is useful to formulate this problem in terms of a probabilistic urn model(9).
If one identifies sites with colors and visitors to those sites as balls of that color, a user following
a recommendation can be thought of as blindly extracting a ball from an urn, noting its color, and
replacing it in the urn along with another ball of the same color. Since in this urn model the
probability that a user will select a site is proportional to the number of previous visitors to the
site, the addition of the user to the site’s visitors makes subsequent selection of the site more
probable. In addition, the process of discovery of new sites can be thought of as the occasional
random insertion of a new color into the urn.
This urn mechanism can be formulated as a branching process(10) by stating that at a given time
step the probability that a site visited by a user has not been recommended is n, so that if s is the
n=
+fi+fi
(
)1
;1
sP
s
t
t
. On the other hand, with
total number of sites seen at time t,
probability n-1
a user will select a site previously viewed and recommended by someone else,
sn , that have visited it
with the probability of being selected proportional to the number of users,
, with (cid:229)
= S
t
=
1
s
+fi+fi
;1
t
n
t
s
previously, i.e.,
(
nP
n
)
tn
s
sn
1(
)1
-
=
-
1
.
s
The evolution of the system is described by the master equations for the probability that at time t,
sn users. In
a total of s novel sites have been discovered, and that the s site has been visited by
terms of time steps it is given by
+
=
+
+
n
),(
tsP
-+
1(
n
)
),1
t
(
sP
(
sP
,1
t
)1
(1)
(
nP
s
+
,1
t
+
)1
=
-
n
)
1(
-
1
(
),
tnPtn
s
s
+
n
(
nP
s
+
),1
t
+
n
(
sP
-
d
(),1
t
n
s
)0,
(2)
with the last term of Eq. (2) expressing the constraint that the s site is discovered at time t. Notice
that since at any given time step either a site is either discovered or visited following a
recommendation, the value of the novelty factor n is bounded away from one.
The solution to these equations gives the probability that n users visit a site, and depends on the
time of discovery of that site. Since Web sites appear at different times, the probability that over a
long period of time a site has been visited by n new users, regardless of the time of its
appearance, is a mixture of those distributions.
This mixture is obtained by summing the individual site probabilities, which are geometrically
distributed, weighted by an exponential factor reflecting the growth of the Web. As originally
shown by Yule in the context of speciation (11, 12), the resulting distribution is given by
=
)(
nP
a
G-
G
a
)()()1
(
n
a
+G
)
(
n
, and the coefficient a is given in terms of the novelty coefficient n
(3)
where
G
)(
x
=
-
1
x
-
u
e
du
u
(cid:242)¥
0
by
3
-=
n
2
n
-
1
which is shown in Fig. 1 for a range of values of n. Since for large values of n ,
aa -»
+G
G
/)(
(
)
n
n
n
we can write Eq. (3) as,
na
)(
(4)
)(
nP
G=
aa -
)
(
n
(5)
which implies that, up to a constant given by the Gamma function, the probability of finding n
users having visited a given site scales inversely in proportion to their number. Thus, the law of
influence predicts that when plotted on a log-log plot, the data shows a straight line whose slope
is determined by the novelty coefficient through Eq. (4).
2.7
2.6
2.5
2.4
2.3
2.2
2.1
a
2
10−4
10−3
10−2
10−1
n
Fig.1 The relationship between n, the novelty coefficient
and a, the power law exponent.
In order to test the theory we conducted an extensive analysis of several Web client traces
containing data about site visits for thousands of sites. The first set of data consisted of America
On-Line (AOL) traces obtained separately for the dates of December 3rd and 5th, 1997. On
December 3rd 47,399 users visited 93,838 sites, while on December 5th 23,316 users visited 87,430
sites. We assume that on a given day users already familiar with a site will revisit it with some
probability and that new users will visit it in proportion to the number of users that have already
visited the site. Thus, an analysis of trace data over a window of time should yield the relative
fraction of users frequenting each site. In the case of AOL, we extracted the number of unique
visitors a site received by removing multiple visits by the same user.
Figure 2 shows the resulting distribution on a log-log scale for the data of Dec. 3rd, 1997. As can
=a
–
. Using
be seen, it exhibits the power law predicted by Eq. (5) with an exponent
03.2
05.0
Eq. (4) this corresponds to a value of the novelty factor n < 0.1. which implies that out of ten
users visiting new sites at least nine were following recommendations. Regression analysis yields
-
6
<P
goodness of fit parameters R2 = 0.998,
10
, confirming the ability of the recommendation
model to account for user statistics. Similarly, a log-log plot of the data for Dec. 5th, yields a
straight line, with a similar exponent, which implies the same bound on n.
4
105
104
103
102
s
e
t
i
s
f
o
r
e
b
m
u
n
101
100
101
number of unique visitors
Fig.2 Distribution of unique AOL visitors per site on Dec. 3, 1997.
102
In order to test the universality of the theory we also analyzed an older set of trace data collected
by Cunha et al. at Boston University (13). It contains traces of 558 students who accessed 2,225
sites during the months of January and February of 1995. Once again, as shown in Figure 3, one
=a
–
08.2
08.0
, corresponding to a
obtains power law behavior, with a value of the exponent
novelty rate < 0.2.
s
e
t
i
s
f
o
r
e
b
m
u
n
103
102
101
100
100
101
number of unique visitors
Fig. 3 Distribution of unique visitors per site from BU traces Jan.-Feb. 1995.
The difference in value between the exponent corresponding to the Boston University and AOL
data can be attributed to two different factors. The first one has to do with the fact that the novelty
rates among the sets of users in the two cases are different, whereas the second relates to the
5
length of time sampled by the data. Consider two sites that have the same number of unique
visitors but are visited at different frequencies-one might be a news site users access daily, the
other an online travel agent accessed only once a month. If one studies the usage logs for the two
sites during a single day, one will appear to have more visitors than the other. Thus, data covering
a shorter period of time would have the effect of broadening the distribution given by Eq. (5) and
therefore yield a smaller exponent value. Since the AOL data covers only a single day of Web
usage, it has a smaller value of athan the Boston University data, which was collected over two
months and hence is less sensitive to the frequency of access to sites.
Another interesting result is obtained if the site visit data is plotted as a function of the site rank,
instead of number of unique visitors. As shown in Figure 4, a Zipf-like law (14) is observed, with
an exponent whose value is 0.9. This confirms ealier findings by Glassman (15) and Aida et al.
(16). We point out however, that Zipf-like regularities do not provide useful information about
the particular process that generates them. This is because the process of ranking random
variables stemming from any broad distribution always produces a narrow and monotically
decreasing power law of the type originally discussed by Zipf (17).
105
104
103
102
101
s
r
o
t
i
s
i
v
e
u
q
i
n
u
f
o
#
100
100
101
102
site rank
103
104
105
Fig.4 Ranked order distribution of the number of unique
visitors per site.
These results show that a theory of social search in the context of the World Wide Web can
account for the strong statistical regularities observed in site visits. A universal power law, whose
exponent relates to the rate at which individuals discover new Web sites, describes these
regularities. We conclude by pointing out that the sheer size of the Web and its highly interactive
nature makes it provides a fertile ground for the discovery and testing of laws of social behavior
that are harder to verify in other settings.
Acknowledgements
We thank Jim Pitkow for providing the data for our analysis and his comments on the paper, and
Rajan Lukose for many useful discussions. This work was partly supported by NSF Grant IRI-
961511.
6
References
Special issue on the Internet, Scientific American 276 (1997).
S. H. Clearwater, B. A. Huberman and T. Hogg, Cooperative Solution of Constraint
Satisfaction Problems, Science 254, 1181-1183 (1991).
M. J.Wade and S.G. Pruett-Jones, Female Copying Increases the Variance in Male
Mating Success, Proc. Natl. Acad. Sci. USA, 87, 5749-5753 (1990).
E. M .Rogers. Diffusion of Innovations. Free Press (1983).
http://www.alexa.com
W. C. Hill, L. Stead, M. Rosenstein and G. Furnas, Recommending and Evaluating
Choices in a Virtual Community of Use, in Proceedings of CHI’95, 194-201 ACM Press
(Denver, CO), 1995
N. Glance, D. Arregui and M. Dardenne, Knowledge Pump: Supporting the Flow and
Use of Knowledge, in Information Technology for Knowledge Management. Eds. U.
Borghoff and R. Pareschi, Springer-Verlag, 1998
B. A. Huberman, P. Pirolli, J. Pitkow and R. M. Lukose, Strong Regularities in World
Wide Web Surfing, Science 280, 95-97 (1998).
N. L. Johnson and S. Kotz, Urn Models and their Application (John Wiley), (1977).
K. B. Athreya and P. E. Ney, Branching Processes, (Springer-Verlag, 1972).
G. U. Yule, A Mathematical Theory of Evolution Based on the Conclusions of Dr. J. C.
Willis, Philosophical Transactions of the Royal Society of London, Series B, 213, 21-87
(1924).
N. L. Johnson, S. Kotz, A.W. Kemp, Univariate Discrete Distributions (John Wiley &
Sons) (1992).
C. R. Cunha, A. Bestavros and M. E. Crovella, Characteristics of WWW Client-based
Traces, Dept. of Computer Science, Boston University Report (1995).
G. K. Zipf, Human behavior and the principle of least effort (Addison-Wesley,
Cambridge, MA, 1949
S. Glassman, Computer Networks and ISDN Systems 27, 165-173 (1994).
M. Aida and N. Takahashi, Evaluation of the Number of Destination Hosts for Data
Networking, Proceedings of
the Sixth International Conference on Computer
Communications and Networks, IEEE 342-349 (1997).
R.Gunther, L. Levitin, B. Schapiro and P. Wagner, Zipf’s Law and the Effect of Ranking
on Probability Distributions, International Journal of Theoretical Physics, 35, 395-417
(1996).
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
.
7
|
1508.05328 | 2 | 1508 | 2016-03-31T15:44:26 | Multi-agent Reinforcement Learning with Sparse Interactions by Negotiation and Knowledge Transfer | [
"cs.MA",
"cs.AI"
] | Reinforcement learning has significant applications for multi-agent systems, especially in unknown dynamic environments. However, most multi-agent reinforcement learning (MARL) algorithms suffer from such problems as exponential computation complexity in the joint state-action space, which makes it difficult to scale up to realistic multi-agent problems. In this paper, a novel algorithm named negotiation-based MARL with sparse interactions (NegoSI) is presented. In contrast to traditional sparse-interaction based MARL algorithms, NegoSI adopts the equilibrium concept and makes it possible for agents to select the non-strict Equilibrium Dominating Strategy Profile (non-strict EDSP) or Meta equilibrium for their joint actions. The presented NegoSI algorithm consists of four parts: the equilibrium-based framework for sparse interactions, the negotiation for the equilibrium set, the minimum variance method for selecting one joint action and the knowledge transfer of local Q-values. In this integrated algorithm, three techniques, i.e., unshared value functions, equilibrium solutions and sparse interactions are adopted to achieve privacy protection, better coordination and lower computational complexity, respectively. To evaluate the performance of the presented NegoSI algorithm, two groups of experiments are carried out regarding three criteria: steps of each episode (SEE), rewards of each episode (REE) and average runtime (AR). The first group of experiments is conducted using six grid world games and shows fast convergence and high scalability of the presented algorithm. Then in the second group of experiments NegoSI is applied to an intelligent warehouse problem and simulated results demonstrate the effectiveness of the presented NegoSI algorithm compared with other state-of-the-art MARL algorithms. | cs.MA | cs | Multi-agent Reinforcement Learning with Sparse Interactions by
Negotiation and Knowledge Transfer
Luowei Zhou, Pei Yang, Chunlin Chen, Member, IEEE, Yang Gao, Member, IEEE
1
6
1
0
2
r
a
M
1
3
]
A
M
.
s
c
[
2
v
8
2
3
5
0
.
8
0
5
1
:
v
i
X
r
a
Abstract-Reinforcement learning has significant applications
for multi-agent systems, especially in unknown dynamic en-
vironments. However, most multi-agent reinforcement learning
(MARL) algorithms suffer from such problems as exponential
computation complexity in the joint state-action space, which
makes it difficult to scale up to realistic multi-agent problems.
In this paper, a novel algorithm named negotiation-based MARL
with sparse interactions (NegoSI) is presented. In contrast to
traditional sparse-interaction based MARL algorithms, NegoSI
adopts the equilibrium concept and makes it possible for agents
to select the non-strict Equilibrium Dominating Strategy Profile
(non-strict EDSP) or Meta equilibrium for their joint actions.
The presented NegoSI algorithm consists of four parts: the
equilibrium-based framework for sparse interactions, the nego-
tiation for the equilibrium set, the minimum variance method
for selecting one joint action and the knowledge transfer of
local Q-values. In this integrated algorithm, three techniques,
i.e., unshared value functions, equilibrium solutions and sparse
interactions are adopted to achieve privacy protection, better
coordination and lower computational complexity, respectively.
To evaluate the performance of the presented NegoSI algo-
rithm, two groups of experiments are carried out regarding
three criteria: steps of each episode (SEE), rewards of each
episode (REE) and average runtime (AR). The first group of
experiments is conducted using six grid world games and shows
fast convergence and high scalability of the presented algorithm.
Then in the second group of experiments NegoSI is applied to an
intelligent warehouse problem and simulated results demonstrate
the effectiveness of the presented NegoSI algorithm compared
with other state-of-the-art MARL algorithms.
Index Terms-Knowledge transfer, multi-agent reinforcement
learning, negotiation, sparse interactions.
I. INTRODUCTION
Multi-agent learning is drawing more and more interests
from scientists and engineers in multi-agent systems (MAS)
and machine learning communities [1]-[4]. One key technique
for multi-agent learning is multi-agent reinforcement learning
(MARL), which is an extension of reinforcement learning in
multi-agent domain [5]. Several mathematical models have
This work was supported by the National Natural Science Foundation of
China (Nos.61273327 and 61432008).
L. Zhou is with the Department of Control and Systems Engineering, School
of Management and Engineering, Nanjing University, Nanjing 210093, China
and with the Robotics Institute, University of Michigan, Ann Arbor, MI 48109,
USA (e-mail: [email protected]).
P. Yang is with the Department of Control and Systems Engineering, School
of Management and Engineering, Nanjing University, Nanjing 210093, China
and with the State Key Laboratory for Novel Software Technology, Nanjing
University, Nanjing 210093, China (e-mail: [email protected]).
C. Chen is with the Department of Control and Systems Engineer-
ing, School of Management and Engineering, Nanjing University, Nanjing
210093, China and with the Research Center for Novel Technology of
Intelligent Equipments, Nanjing University, Nanjing 210093, China (e-mail:
[email protected]).
Y. Gao is with the State Key Laboratory for Novel Software Technology,
Department of Computer Science, Nanjing University, Nanjing 210023, China
(e-mail: [email protected]).
been built as frameworks of MARL, such as Markov games
(MG) [6] and decentralized sparse-interaction Markov decision
processes (Dec-SIMDP) [7]. Markov games are based on the
assumption of full observability of all agents in the entire
joint state-action space. Several well-known equilibrium-based
MARL algorithms [6]-[12] are derived from this model. Dec-
SIMDP based algorithms rely on agents' local observation,
i.e., the individual state and action. Agents in Dec-SIMDP
are modeled with single-agent MDP when they are outside
of the interaction areas, while the multi-agent model such as
MG is used when they are inside. Typical Dec-SIMDP based
algorithms include LoC [13] and CQ-learning [14]. Besides,
other models such as learning automata [2] [15] are also
valuable tools for designing MARL algorithms.
In spite of the rapid development of MARL theories and
algorithms, more efforts are needed for practical applications
of MARL when compared with other MAS techniques [16]-
[18] due to some limitations of the existing MARL methods.
The equilibrium-based MARL relies on the tightly coupled
learning process which hinders their applications in practice.
Calculating the equilibrium (e.g., Nash equilibrium [19]) for
each time step and all joint states are computationally ex-
pensive [8], even for relatively small scale environments with
two or three agents. In addition, sharing individual states,
individual actions or even value functions all the time with
other agents is unrealistic in some distributed domains (e.g.,
streaming processing systems [20], sensor networks [21])
given the agents' privacy protections and huge real-time com-
munication costs [1]. As for MARL with sparse interactions,
agents in this setting have no concept of equilibrium policy
and they tend to act aggressively towards their goals, which
results in a high probability of collisions.
Therefore, in this paper we focus on how the equilibrium
mechanism can be used in sparse-interaction based algo-
rithms and a negotiation-based MARL algorithm with sparse
interactions (NegoSI) is proposed for multi-agent systems
in unknown dynamic environments. The NegoSI algorithm
consists of four parts: the equilibrium-based framework for
sparse interactions, the negotiation for the equilibrium set,
the minimum variance method for selecting one joint action
and the knowledge transfer of local Q-values. Firstly, we start
with the proposed algorithm based on the MDP model and the
assumption that the agents have already obtained the single-
agent optimal policy before learning in multi-agent settings.
Then, the agents negotiate for pure strategy profiles as their
potential set for the joint action at the "coordination state" and
they use the minimum variance method to select the relatively
good one. After that, the agents move to the next joint state
and receive immediate rewards. Finally, the agents' Q-values
are updated by their rewards and equilibrium utilities. When
initializing the Q-value for the expanded "coordination state",
the agents with NegoSI utilize both of the environmental infor-
mation and the prior-trained coordinated Q-values. To test the
effectiveness of the proposed algorithm, several benchmarks
are adopted to demonstrate the performances in terms of
the convergence, scalability, fairness and coordination ability.
In addition, aiming at solving realistic MARL problems,
the presented NegoSI algorithm is also further tested on an
intelligent warehouse simulation platform.
The rest of this paper is organized as follows. Section II
introduces the basic MARL theory and the MAS with sparse
interactions. In Section III, the negotiation-based MARL al-
gorithm with sparse interactions (NegoSI) is presented in
details and related issues are discussed. Experimental results
are shown in Section IV. Conclusions are given in Section V.
II. BACKGROUND
In this section, some important concepts in multi-agent
reinforcement learning and typical sparse-interaction based
MARL algorithms are introduced.
A. MDP and Markov Games
We start with reviewing two standard decision-making
models that are relevant to our work, i.e., Markov decision
process (MDP) and Markov Games, respectively. MDP is the
foundation of Markov Games, while Markov Games adopt
game theory in multi-agent MDPs. MDP describes a sequential
decision problem as follows [22]:
Definition 1: (Markov Decision Process, MDP) An MDP
is a tuple hS,A,R,T i, where S is the state space, A is the
action space of the agent, R : S × A → R is the reward function
mapping state-action pairs to rewards, T : S × A × S → [0,1] is
the transition function.
An agent in an MDP is required to find an optimal policy
which maximizes some reward-based optimization criteria,
such as expected discounted sum of rewards:
V ∗(s) = maxp Ep {
k=0
g krt+kst = s},
(1)
: S ×A → [0,1] stands for the policy of an agent, Ep
where V ∗(s) stands for the value of a state s under the optimal
policy, p
is
the expectation under policy p , t is any time step, k represents
a future time step, rt+k denotes the reward at the time step
(t + k) and g ∈ [0,1] is a parameter called the discount factor.
This goal can also be equivalently described using the Q-value
for a state-action pair:
Q∗(s,a) = r(s,a) + g (cid:229)
T (s,a,s′)max
a′
s′
Q∗(s′,a′),
(2)
where Q∗(s,a) stands for the value of a state-action pair (s,a)
under the optimal policy, s′ is the next state and r(s,a) is
the immediate reward when agent adopts the action a at the
state s, T (s,a,s′) is the transition possibility for the agent to
transit from s to s′ given action a. One classic RL algorithm
for estimating Q∗(s,a) is Q-learning [23], whose one-step
updating rule is as follows:
2
Q(s,a) ← (1 − a )Q(s,a) + a [r(s,a) + g max
a′
Q(s′,a′)],
(3)
where Q(s,a) denotes the state-action value function at a state-
action pair (s,a) and a ∈ [0,1] is a parameter called the
learning rate. Provided that all state-action pairs are visited
infinite times with a reasonable learning rate, the estimated
Q-value Q(s,a) converges to Q∗(s,a) [24].
Markov games are widely adopted as a framework for multi-
agent reinforcement learning (MARL) [6] [10]. It is regarded
as multiple MDPs in which the transition probabilities and
rewards depend on the joint state-action pairs of all agents. In a
certain state, agents' individual action sets generate a repeated
game that could be solved in a game-theoretic way. Therefore,
Markov game is a richer framework which generalizes both
of the MDP and the repeated game [25]-[27].
Definition 2: (Markov game) An n-agent (n ≥ 2) Markov
game is a tuple hn, {Si}i=1,...,n, {Ai}i=1,...,n, {Ri}i=1,...,n,T i,
where n is the number of agents in the system, Si is the set
of the state space for ith agent, S = {Si}i=1,...,n is the set of
state spaces for all agents, Ai is the set of the action space
for ith agent, A = {Ai}i=1,...,n is the set of action spaces for
all agents, Ri : S × A → R is the reward function of agent i,
T : S × A × S → [0,1] is the transition function.
Denote the individual policy of agent i by p i = S × Ai →
[0,1] and the joint policy of all agents by p = (p 1, . . . ,p n).
The Q-value of the join state-action pair for agent i under the
joint policy p can be formulated by
p
i (~s,~a) = Ep {
Q
k=0
g krt+k
i
~st =~s,~at = ~a},
(4)
where ~s ∈ S stands for a joint state, ~a ∈ A for a joint action and
rt+k
is the reward received at the time step (t + k). Unlike the
i
optimization goal in an MDP, the objective of a Markov game
is to find an equilibrium joint policy p
rather than an optimal
joint policy for all agents. Here, the equilibrium policy concept
is usually transferred to finding the equilibrium solution for the
one-shot game played in each joint state of a Markov game
[8]. Several equilibrium-based MARL algorithms in existing
literatures such as NashQ [10] [28] and NegoQ [8] have been
proposed, so that the joint state-action pair Q-value can be
updated according to the equilibrium:
Qi(~s,~a) ← (1 − a )Qi(~s,~a) + a (ri(~s,~a) + gf
(5)
where f i(~s′) denotes the expected value of the equilibrium in
the next joint state ~s′ for agent i and can be calculated in the
one-shot game at that joint state.
i(~s′))),
B. MAS with Sparse Interactions
The definition of Markov game reveals that all agents need
to learn their policies in the full joint state-action space and
they are coupled with each other all the time [10] [29]. How-
ever, this assumption does not hold in practice. The truth is
that the learning agents in many practical multi-agent systems
are loosely coupled with some limited interactions in some
¥
(cid:229)
¥
(cid:229)
1
2
3
4
5
6
7
1
2
3
4
5
6
7
R1
G2
R3
G3
G1
R2
Fig. 1: A part of an intelligent warehouse with three robots, where
Ri (i = 1,2,3) represents for robot i with its corresponding goal Gi
and the shaded grids are storage shelves.
particular areas [5] [13] [29]. Meanwhile, the interactions
between agents may not always involve all the agents. These
facts lead to a new mechanism of sparse interactions for
MARL research.
Sparse-interaction based algorithms [13] [14] have recently
found wide applications in MAS research. An example of
MAS with sparse interactions is the intelligent warehouse
systems, where autonomous robots only consider other robots
when they are close enough to each other [25] (see Fig. 1),
i.e., when they meet around the crossroad. Otherwise, they can
act independently.
Earlier works such as the coordinated reinforcement learn-
ing [30] [31] and sparse cooperative Q-learning [32] used
coordination graphs (CGs) to learn interdependencies between
agents and decomposed the joint value function to local value
functions. However, these algorithms cannot learn CGs online
and only focus on finding specific states where coordination is
necessary instead of learning for coordination [13]. Melo and
Veloso [13] extended the Q-learning to a two-layer algorithm
and made it possible for agents to use an additional COORDI-
NATE action to determine the state when the coordination was
necessary. Hauwere and Vrancx proposed Coordinating Q-
learning (CQ-learning) [14] and FCQ-learning [33] that helped
agents learn from statistical information of rewards and Q-
values where an agent should take other agents' states into
account. However, all these algorithms allow agents to play
greedy strategies at a certain joint state rather than equilibrium
strategies, which might cause conflicts.
More recently, Hu et al
[8] proposed an efficient
equilibrium-based MARL method, called Negotiation-based
Q-learning, by which agents can learn in a Markov game with
unshared value functions and unshared joint state-actions. In
later work, they applied this method for sparse interactions
by knowledge transfer and game abstraction [29], and demon-
strated the effectiveness of the equilibrium-based MARL in
solving sparse-interaction problems. Nevertheless, as opposed
to single Q-learning based approaches like CQ-learning, Hu's
equilibrium-based methods for sparse interactions require a
the joint states
great deal of real-time information about
3
and joint actions of all the agents, which results in huge
amount of communication costs. In this paper, we focus
on solving learning problems in complex systems. Tightly
coupled equilibrium-based MARL methods discussed above
are impractical
in these situations while sparse-interaction
based algorithms tend to cause many collisions. To this end,
we adopt the sparse-interaction based learning framework and
each agent selects equilibrium joint actions when they are in
coordination areas.
III. NEGOTIATION-BASED MARL WITH SPARSE
INTERACTIONS
When people work in restricted environments with possible
conflicts, they usually learn how to finish their individual
tasks first and then learn how to coordinate with others. We
apply this commonsense to our sparse-interaction method and
decompose the learning process into two distinct sub-processes
[5]. First, each agent learns an optimal single-agent policy by
itself in the static environment and ignores the existences of
other agents. Second, each agent learns when to coordinate
with others according to their immediate reward changes, and
then learns how to coordinate with others in a game-theoretic
way. In this section, the negotiation-based framework for MAS
with sparse interactions is first introduced and then related
techniques and specific algorithms are described in details.
A. Negotiation-based Framework for Sparse Interactions
We assume that agents have learnt their optimal policies and
reward models when acting individually in the environment.
Two situations might occur when agents are working in a
multi-agent setting. If the received immediate rewards for
state-action pairs are the same as what they learned by reward
models, the agents act independently. Otherwise, they need to
expand their individual state-action pairs to the joint ones by
adding other agents' state-action pairs for better coordination.
This negotiation-based framework for sparse interactions is
given as shown in Algorithm 1, while the expansion and
negotiation process is explained as follows:
1) Broadcast joint state. Agents select an action at a certain
state, and they detect a change in the immediate rewards.
This state-action pair is marked as "dangerous" and
these agents are called "coordinating agents". Then,
"coordinating agents" broadcast their state-action infor-
mation to all other agents and receive corresponding
state-action information from others. These state-action
pairs with reward changes form a joint state-action and
is marked as a "coordination pair". Also, these states
form a joint state called a "coordination state", which is
included in the state space of each "coordinating agent".
2) Negotiation for equilibrium policy. When agents select a
"dangerous" state-action pair, they broadcast their state-
action information to each other to determine whether
they are staying at a "coordination pair". If so, the
agents need to find an equilibrium policy rather than
their inherent greedy policies to avoid collisions. We
propose a negotiation mechanism similar to the work in
[8] to find this equilibrium policy. Each "coordinating
4
agent" broadcasts the set of strategy profiles that are
potential Non-strict Equilibrium-Dominating Strategy
Profile (non-strict EDSP) according to its own utilities
(See Algorithm 2). If no non-strict EDSP is found, the
agents search for a Meta equilibrium set (See Algorithm
3) instead, which is always nonempty [8]. Then, if there
are several candidates in the equilibrium set, the agents
use the minimum variance method to find the relatively
good one (See Algorithm 4).
3) When an agent selects an action at a certain state, if
neither a change in the immediate reward nor a "dan-
gerous" state-action pair is detected, the agent continue
to select its action independently.
Remark 1: If the number of agents n ≥ 2, the agents may
have several different expansions for different "coordination
states" in one state. Each expanded joint-state is assigned an
index and has a corresponding local Q-value. Once the agents
form a "coordination state", they search for the joint state in
their expanded joint-state pool and obtain the corresponding
local Q-value. An example is shown as in Fig. 2. In this
example, agent 1-4 have twelve states, four states, five states
and three states, respectively. For convenience, we assume
here when agent 1 is expanding joint states, it adds all states
of relevant agents in its joint states even though in most of
the cases it only needs to add a small part of the states.
Specifically, the fourth state of agent 1 is expanded to two
joint states, the first one is with agent 4 and the second one
is with agent 2 and agent 4. Similarly, the fifth state of agent
1 is expanded to three joint states, the first one is with agent
2, the second one is with agent 2 and agent 3, and the third
one is with agent 3.
Remark 2: Owing to the broadcast mechanism, an agent
will add the states of all the "coordinating agents" to its state
space, even though it does not need to coordinate with some
of these agents. The agents cannot distinguish the agent they
need to coordinate with from other agents, which brings un-
necessary calculations. This problem becomes more significant
with the growth of the complexity of the environment and
the number of the agents. Our tentative solution is setting
the communication range of the agents so that they can only
add neighbouring agents' states to their state space and ignore
those that do not need to coordinate with.
Remark 3: It is worth noting that if the agent detects
a "dangerous" state-action pair, it should observe a reward
change, that is, if we replace the former concept with the latter
one, the algorithm can work equally. Defining dangerous state-
action pairs mainly help us to better explain our thoughts and
describe the algorithm.
Algorithm 1 Negotiation-based Framework for Sparse Interactions
Input: The agent i, state space Si, action space Ai, learning rate a , discount rate g , exploration factor e
for the e − Greedy
exploration policy.
Initialize: Global Q-value Qi with optimal single-agent policy.
1: for each episode do
Initialize state si;
2:
while true do
3:
4:
5:
6:
Select ai ∈ Ai from Qi with e − Greedy;
if (si,ai) is "dangerous" then
Broadcasts (si,ai) and receives (s−i,a−i), form (~s,~a);
/* See Section III-D for the definitions of s−i and a−i*/
if (~s,~a) is not "coordination pair" then
Mark (~s,~a) as a "coordination pair" and ~s as a "coordination state", initialize Local Q-value QJ
knowledge (See Equation 9);
i at ~s with transfer
end if
Negotiate for the equilibrium joint action with Algorithm 2 - Algorithm 4. Select new ~a with e − Greedy;
else {detected an immediate reward change}
Mark (si,ai) as "dangerous", broadcasts (si,ai) and receives (s−i,a−i), form (~s,~a);
Mark (~s,~a) as "coordination pair" and ~s as "coordination state", initialize QJ
Equation 9);
Negotiate for the equilibrium joint action with Algorithm 2 - Algorithm 4. Select new ~a with e − Greedy;
i at ~s with transfer knowledge (See
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
end if
Move to the next state s′
if ~s exists then
Update QJ
i (~s,~a) = (1 − a )QJ
i and receive the reward ri;
i (~s,~a) + a (ri + g max
a′
i
end if
Update Qi(si,ai) = (1 − a )Qi(si,ai) + a (ri + g max
a′
i
si ← s′
i;
Qi(s′
i,a′
i));
Qi(s′
i,a′
i));
21:
22:
23: end for
end while until si is a terminal state.
5
Expanded state space of agent 2
5-1 5-2 5-3 5-4
Agent 3
Agent 2
5-1-1 5-1-2 5-1-3 5-1-4 5-1-5
5-2-1 5-2-2 5-2-3 5-2-4 5-2-5
1
5-3-1 5-3-2 5-3-3 5-3-4 5-3-5
2
5-4-1 5-4-2 5-4-3 5-4-4 5-4-5
3
Expanded state space of
agent 2 and agent 3
5-1 5-2 5-3 5-4 5-5
Expanded state space of agent 3
9 10 11 12
5 6 7 8
1 2 3 4
State space of agent 1
Expanded state space of agent 4
4-1 4-2 4-3
1
2
Agent 4
4-1-1 4-1-2 4-1-3
4-2-1 4-2-2 4-2-3
4-3-1 4-3-2 4-3-3
Agent 2
4-4-1 4-4-2 4-4-3
Expanded state space of
agent 2 and agent 4
Fig. 2: An example of the expansion of an agent's state space.
B. Negotiation for Equilibrium Set
One contribution of this paper is to apply the equilibrium
solution concept to traditional Q-learning based MARL algo-
rithms. Different from previous work like CQ-learning and
Learning of Coordination [14], our approach aims at finding
the equilibrium solution for the one-shot game played in
each "coordination state". As a result, we focus on two pure
strategy profiles [8], i.e., Non-strict Equilibrium-Dominating
Strategy Profile (non-strict EDSP) and Meta Equilibrium set.
The definition of Non-strict EDSP is as follows.
Definition 3: (Non-strict EDSP): In an n-agent (n ≥ 2)
normal-form game G
, ~ei ∈ A(i = 1,2, . . . ,m) are pure strategy
Nash equilibriums. A joint action ~a ∈ A is a non-strict EDSP
if ∀ j ≤ n,
U j(~a) ≥ min
i
U j(~ei)
(6)
The negotiation process of finding the set of Non-strict
EDSP is shown as in Algorithm 2, which generally consists of
two steps: 1) the agents find the set of strategy profiles that are
potentially Non-strict EDSP according to their individual util-
ities; 2) the agents solve the intersection of all the potentially
strategy sets and get the Non-strict EDSP set.
However, given the fact that the pure strategy Nash equi-
librium can be non-existent in some cases, the set of Non-
strict EDSP might also be empty. On this occasion, we replace
this strategy profile with Meta Equilibrium, which is always
nonempty. The sufficient and necessary condition of Meta
Equilibrium is defined as follows.
Definition 4: (Sufficient and Necessary Condition of
Meta Equilibrium) [8]: In an n-agent (n ≥ 2) normal-form
game G
, a joint action ~a is called a Meta equilibrium from
a metagame k1k2 . . . krG
if and only if for any i there holds
Ui(~a) ≥ min
~aPi
max
ai
min
~aSi
Ui( ~aPi ,ai, ~aSi),
(7)
For example, in a three agents metagame 213G
where Pi is the set of agents listed before sign i in the prefix
k1k2 . . .kr, Si is the set of agents listed after sign i in the prefix.
, we have
P1 = {2},S1 = {3};P2 = /0,S2 = {1,3};P3 = {2,1},S3 = /0. A
Meta equilibrium ~a meets the following constraints:
U1(~a) ≥ min
a2
U2(~a) ≥ max
a2
U3(~a) ≥ min
a2
max
a1
min
a1
min
a1
min
a3
min
a3
max
a3
U1(a1,a2,a3),
U2(a1,a2,a3),
U3(a1,a2,a3).
(8)
Hu et al [8] used a negotiation-based method to find Meta
Equilibrium set and we simplified the method as shown in
Algorithm 3. It is also pointed out that both of these two
strategy profiles belong to the set of symmetric meta equilib-
rium, which to some degree strengthens the convergence of
the algorithm.
Algorithm 2 Negotiation for Non-strict EDSPs Set
Input: A normal-form game hn, {Ai}i=1,...,n, {Ui}i=1,...,ni.
/* To be noted, "coordinating agent" i only has the
knowledge of n, {Ai}i=1,...,n and Ui*/
Initialize: The set of non-strict EDSP candidates for "coor-
dinating agents" i: Ji
Minimum utility value of pure strategy Nash equilibrium
(PNE) candidates for "coordinating agents" i: MinU i
PNE ←
NS ← /0;
;
The set of non-strict EDSPs: JNS ← /0.
1: for each ~a−i ∈ A−i do
2:
if max
Ui(a′
i, ~a−i) < MinU i
PNE then
a′
i
MinU i
3:
PNE = max
a′
i
Ui(a′
i, ~a−i);
end if
4:
5: end for
6: for each ~a ∈ A do
if Ui(~a) ≥ MinU i
7:
8:
end if
9:
10: end for
Ji
NS ← Ji
NS ∪ {~a};
PNE then
/* Broadcast Ji
NS and corresponding utilities*/
i=1 Ji
NS
11: JNS ← Tn
¥
C. Minimum Variance Method
In Section III-B, we presented the process for all "coordi-
nating agents" to negotiate for an equilibrium joint-action set.
However, the obtained set usually contains many strategy pro-
files and it is difficult for the agents to choose an appropriate
one. In this paper a Minimum Variance method is proposed
to help the "coordinating agents" choose the joint action with
relatively high total utility and minimum utilities variance to
guarantee the cooperation and fairness of the learning process.
In addition, if the equilibrium set is nonempty, the best solution
defined in the minimum variance method always exists. The
minimum variance method is described in Algorithm 4.
After negotiating for one equilibrium, the agents update
their states according to the joint action and receive immediate
rewards, which are used to update local Q-values as well
as global Q-values. Unlike other sparse-interaction algorithms
(e.g., CQ-learning), we update the global optimal Q-values to
avoid possible misleading information. In fact, in some cases,
the selected policies in multi-agent setting are totally opposite
to the agents' individual optimal policies due to dynamic
coordinating processes. The whole negotiation-based learning
algorithm has already been given in Algorithm 1.
D. Local Q-value transfer
At the beginning of the learning process, we use the transfer
knowledge of agents' optimal single agent policy to accelerate
the learning process. Furthermore, it is possible to improve
the algorithm performance by the local Q-value transfer. In
most previous literatures [13] [14], the initial local Q-values
of the newly expanded joint states are zeros. Recently, Vrancx
et al proposed a transfer learning method [34] to initialize
these Q-values with prior trained Q-value from the source
task, which is reasonable in the real world. When people
meet with others on the way to their individual destinations,
Algorithm 3 Negotiation for Meta Equilibrium Set for 3-agent
games
Input: A normal-form game < n, {Ai}i=1,...,n, {Ui}i=1,...,n >.
/* To be noted, "coordinating agents" i only has the
knowledge of n, {Ai}i=1,...,n and Ui*/
Initialize: The set of Meta Equilibrium candidates for "co-
MetaE ← /0;
ordinating agents" i: Ji
Minimum utility value of Meta Equilibrium candidates for
"coordinating agents" i: MinU i
;
The set of Meta Equilibrium: JMetaE ← /0;
1: Randomly initialize the prefix s1s2s3
from the set
MetaE ← ¥
{123,132,213,231,312,321}.
MetaE according to Equation 7;
MetaE then
MetaE ∪ {~a};
2: Calculate MinU i
3: for each ~a ∈ A do
if Ui(~a) ≥ MinU i
4:
5:
end if
6:
7: end for
Ji
MetaE ← Ji
/* Broadcast Ji
i=1 Ji
8: JMetaE ← Tn
MetaE and corresponding utilities*/
MetaE
6
Algorithm 4 Minimum Variance Method
Input: The
ns, ~a2
equilibrium set with m elements
ns} and corresponding utilities {U ns
Initialize: threshold value for total utility t ;
ns, . . . , ~am
{ ~a1
JNS =
i }i=1,...,n.
/* We set the threshold value to the mean value of the sum
i=1 Uns
i (
utilities of different joint-action profiles
m
Minimum variance of these equilibriums MinV ← ¥
Best equilibrium of the non-strict EDSPs set JBestNS ← /0.
(cid:229) m
j=1 (cid:229) n
~a j
ns)
*/
;
1: for each ~a j
ns ∈ JNS do
ns) < t then
~a j
i=1 U ns
i (
2:
JNS ← JNS\{
~a j
ns};
if (cid:229) n
3:
end if
4:
5: end for
/* Minimize the joint action's utility variance*/
if q 1
6: for each ~a j
ns ∈ JNS do
~a j
(cid:229) n
ns) − 1
k=1[U ns
k (
7:
n
MinV = q 1
(cid:229) n
k=1[U ns
k (
n
~a j
ns};
JBestNS = {
8:
n
9:
end if
10:
11: end for
12: Output JBestNS as the adopted joint action.
(cid:229) n
i=1 U ns
i (
~a j
ns) − 1
n
~a j
ns)]2 < MinV then
(cid:229) n
i=1 U ns
i (
~a j
ns)]2;
they usually have prior knowledge of how to avoid collisions.
Based on this prior commonsense and the knowledge of how
to finish their individual tasks, the agents learn to negotiate
with others and obtain fixed coordination strategies suitable
for certain environments. However, Vrancx et al only used
coordination knowledge to initialize the local Q-values and
overlooked the environmental information, which was proved
to be less effective in our experiments. In our approach, we
initialize the local Q-values of newly detected "coordination"
to hybrid knowledge as follows:
QJ
i ((si, ~s−i), (ai, ~a−i)) ← Qi(si,ai) + QCT
i
(~s,~a),
(9)
for
i (~s,~a), QCT
i
is the joint-state set except
where ~s−i
the state si
and ~a−i
is the joint-action set except for the action ai,
i ((si, ~s−i), (ai, ~a−i)) is equal with QJ
QJ
(~s,~a) is the
transferred Q-value from a blank source task at joint state
~s. In this blank source task, joint action learners (JAL) [35]
learn to coordinate with others disregarding the environmental
information. They are given a fixed number of learning steps to
learn stable Q-value QCT
(~s,~a). Similar to [34], the joint state
(~s,~a) is presented as the relative position (D x,D y),
~s in QCT
horizontally and vertically. When agents attempt to move into
the same grid location or their previous locations (as shown in
Fig. 3), these agents receive a penalty of -10. In other cases,
the reward is zero.
i
i
Take a two-agent source task for example (as shown in Fig.
4(a)). The locations of agent 1 and agent 2 are (4,3) and (3,4),
respectively. Then we have the joint state ~s = (x1 − x2,y1 −
y2) = (1, −1) and the joint action for this joint state
1
2 3 4 5
7
~a = (cid:0)(a1,a2)(cid:1) =
(↑, ↑)
(↓, ↑)
(←, ↑)
(→, ↑)
(↑, ↓)
(↓, ↓)
(←, ↓)
(→, ↓)
(↑, ←)
(↓, ←)
(←, ←)
(→, ←)
(↑, →)
(↓, →)
(←, →)
(→, →)
,
where {↑, ↓, ←, →} denotes the action set of {up, down, left,
right} for each agent. After sufficient learning steps, the agents
learn their Q-value matrices at joint state ~s as
QCT
1 (~s, ···) = QCT
2 (~s, ···) =
0
0
0
0
0
0
0 −10
−10 0
0
0
0
0
0
0
,
that is because when agent 1 and agent 2 select the action
pair (↑, ←) or (→, ↓), a collision will occur, which leads to
a -10 punishment for each agent. Otherwise, the reward is
0. Suppose that the agents are in the "coordination state"
as shown in Fig. 4(b). When acting independently in the
environment, the agents learn their single agent optimal Q-
value vectors at state s1 or s2 as Q1(s1, ···) = (−1, −10, −5, −1),
Q2(s2, ···) = (−10, −1, −1, −5). Then the local Q-value matri-
ces of this "coordination state" need to be initialized as
1(~s, ···) = Q1(s1, ···)T × (1,1,1,1) + QCT
QJ
1 (~s, ···)
−1 −1 −11 −1
−10 −10 −10 −10
−5 −5 −5 −5
−1 −11 −1 −1
=
,
QJ
2(~s, ···) = (1,1,1,1)T × Q2(s2, ···) + QCT
2 (~s, ···)
−10 −1 −11 −5
−10 −1 −1 −5
−10 −1 −1 −5
−10 −11 −1 −5
=
.
So the pure strategy Nash equilibria for this joint state are
(up, down) and (right, left). If we initialize the local Q-value
with the way used in [34] or just initialize them to zeros, the
learning process would be much longer.
For the "coordination state" with three agents, the Q-value
QCT
(~s,~a) can be calculated in the same way, except for the
i
relative positions (D x1,D y1,D x2,D y2) = (x1 − x2,y1 − y2,x2 −
x3,y2 − y3) and the cubic Q-value for each joint state.
Collision scenario I:
1
2
1
2
3
4
5
2
1
(a) A 5 × 5 blank source task with two agents.
2
G1
G2
1
(b) The detected "coordination states".
Fig. 4: An example of the local Q-value transfer in a two-agent
system.
IV. EXPERIMENTS
To test the presented NegoSI algorithm, several groups of
simulated experiments are implemented and the results are
compared with those of three other state-of-the-art MARL
algorithms, namely, CQ-learning [14], NegoQ with value
function transfer (NegoQ-VFT) [29] and independent learners
with value function transfer (IL-VFT). In next two subsections,
the presented NegoSI algorithm is applied to six grid world
games and an intelligent warehouse problem, which shows
that NegoSI is an effective approach for MARL problems
compared with other existing MARL algorithms.
For all these experiments, each agent has four actions: up,
down, left, right. The reward settings are as follows:
1) When an agent reaches its goal/goals,
it receives a
reward of 100. Its final goal is an absorbing state. One
episode is over when all agents reach their goals.
2) A negative reward of -10 is received if a collision
happens or an agent steps out of the border. In these
cases, agents will bounce to their previous states.
3) Otherwise, the reward is set to -1 as the power consump-
tion.
The settings of other parameters are the same for all algo-
rithms: learning rate a = 0.1, discount rate g = 0.9, exploration
factor e = 0.01 for e -greedy strategy. All algorithms run 2000
iterations for the grid world games and 8000 iterations for the
intelligent warehouse problem. We use three typical criteria to
evaluate these MARL algorithms, i.e., steps of each episode
(SEE), rewards of each episode (REE) and average runtime
(AR). All the results are averaged over 50 runs.
move right
move left
A. Tests on grid world games
Collision scenario II:
1
2
move right
move left
Fig. 3: Two scenarios of possible collisions.
The proposed NegoSI algorithm is evaluated in the grid
world games presented by Melo and Veloso [13], which are
shown in Fig. 5. The first four benchmarks, i.e., ISR, SUNY,
MIT and PENTAGON (shown as Fig. 5(a)-(d), respectively),
are two-agent games, where the cross symbols denote the
initial locations of each agent and the goals of the other agent.
In addition, we design two highly competitive games to further
test the algorithms, namely, GW nju (Fig. 5(e)) and GWa3
(Fig. 5(f)). The game GW nju has two agents and the game
GWa3 has three agents. The initial location and the goal of
agent i are represented by the number i and Gi, respectively.
We first examine the performances of the tested algorithms
regarding the SEE (steps of each episode) for each benchmark
map (See Fig. 6). The state-of-the-art value function transfer
mechanism helps NegoQ and ILs to converge fast in all games
except for SUNY. NegoSI also has good convergence char-
acteristics. CQ-learning, however, has a less stable learning
curve especially in PENTAGON, GW nju and GWa3. This
is reasonable since in these highly competitive games, the
agents' prior-learned optimal policies always have conflicts
and CQ-learning cannot find the equilibrium joint action to
coordinate them. When more collisions occur, the agents are
frequently bounced to previous states and forced to take more
steps before reaching their goals. The learning steps of the
final episode for each algorithm in each benchmark map are
shown in Table I. In ISR, NegoSI converges to 7.48, which
is the closest one to the value obtained by the optimal policy.
In other cases, NegoSI achieves the learning step of the final
episode between 105.1% and 120.4% to the best solution.
e
d
o
s
p
e
i
h
c
a
e
f
o
s
p
e
t
s
e
d
o
s
p
e
i
h
c
a
e
f
o
s
p
e
t
s
80
60
40
20
0
0
80
60
40
20
0
0
e
d
o
s
p
e
i
h
c
a
e
f
o
s
p
e
t
s
30
25
20
15
10
0
CQ
ILVFT
NegoQVFT
NegoSI
5
15
episodes (· 100)
10
(a) ISR
CQ
ILVFT
NegoQVFT
NegoSI
10
5
15
episodes (· 100)
(c) MIT
CQ
ILVFT
NegoQVFT
NegoSI
10
5
15
episodes (· 100)
(e) GW nju
20
20
20
e
d
o
s
p
e
i
h
c
a
e
f
o
s
p
e
t
s
e
d
o
s
p
e
i
h
c
a
e
f
o
s
p
e
t
s
e
d
o
s
p
e
i
h
c
a
e
f
o
s
p
e
t
s
25
20
15
10
0
60
40
20
0
0
80
60
40
20
0
0
8
CQ
ILVFT
NegoQVFT
NegoSI
10
5
15
episodes (· 100)
(b) SUNY
CQ
ILVFT
NegoQVFT
NegoSI
5
15
episodes (· 100)
10
(d) PENTAGON
CQ
ILVFT
NegoQVFT
NegoSI
10
5
15
episodes (· 100)
(f) GWa3
20
20
20
Fig. 6: SEE (steps of each episode) for each tested benchmark map.
Then we analyze the REE (rewards of each episode) cri-
terion of these tested algorithms (See Fig. 7-12). The results
vary for each map. In ISR, NegoSI generally achieves the
highest REE through the whole learning process, which shows
fewer collisions and more cooperations between the learning
agents. Also, the difference between the two agents' reward
values is small, which reflects the fairness of the learning
process of NegoSI. The agents are more independent
in
SUNY. Each of them has three candidates for the single-agent
optimal policy. In this setting, ILVFT has good performance.
NegoSI shows its fairness and high REE value compared
with NegoVFT and CQ-learning. The agents in MIT have
more choices of collision avoidances. All the tested algorithms
obtain the final reward of around 80. However, the learning
curves of CQ-learning are relatively unstable. In PENTAGON,
NegoSI proves its fairness and achieves as good performance
as algorithms do with the value function transfer.
Other than these above benchmarks, we give two highly
competitive games to test the proposed algorithm. In both
games, the agents' single-agent optimal policies conflict with
(a) ISR
(b) SUNY
(c) MIT
(d) PENTAGON
G2
1
2
G1
(e) GW nju
3
1,G2
2,G1
G3
(f) GWa3
Fig. 5: Grid world games.
TABLE I: The learning steps of the final episode of each tested benchmark map.
CQ − learning
ILV FT
NegoQV FT
NegoSI
T he optimal policy
ISR
8.91
13.11
8.36
7.48
6
SUNY MIT
PENTAGON GW n ju GWa3
10.70
10.38
12.98
10.92
10
19.81
18.67
19.81
21.29
18
15.32
14.18
8.55
10.30
8
12.65
10.94
10.95
12.11
10
8.65
11.40
8.31
8.87
8
each other and need appropriate coordination. In GW nju, the
learning curves of NegoSI finally converge to 91.39 and 90.09,
which are very close to the optimal final REE values as 93
and 91, respectively. Similar to NegoSI, NegoVFT achieves the
final REE values of 91.27 and 91.25. However, NegoSI can
negotiate for a fixed and safer policy that allows one agent to
always move greedily and the other to avoid collision (while
NegoVFT cannot), which shows the better coordination ability
of NegoSI. For the three-agent grid world GWa3, even though
NegoSI does not have the lowest SEE, it achieves the highest
REE through the whole learning process. For one thing, this
result demonstrates the scalability of NegoSI in the three-agent
setting. For another, it shows that the agents using NegoSI have
the ability to avoid collisions to obtain more rewards while
the agents using traditional MARL methods have less desire
to coordinate and therefore lose their rewards. Actually, even
though we developed our method with non-cooperative multi-
agent model, it does not necessarily mean that the agents are
egoistic. Thanks to the negotiation mechanism, agents learn
to benefit themselves while doing little harm to others, which
shows an evidence of cooperation.
The results regarding AR (average runtime) are shown in
Table II. ILVFT has the fastest learning speed, which is only
five to ten times slower than Q-learning to learn single policy.
CQ-learning only considers joint states in "coordination state"
and it also has a relatively small computational complexity.
The AR of NegoVFT is five to ten times more than that
of ILVFT. This is reasonable since NegoVFT learns in the
whole joint state-action space and computes the equilibrium
for each joint state. The learning speed of NegoSI is slower
than CQ-learning but faster than NegoVFT. Even if NegoSI
adopts the sparse-interaction based learning framework and
has computational complexity similar to CQ-learning, it needs
to search for the equilibrium joint action in the "coordination
state", which slows down the learning process.
B. A real-world application: intelligent warehouse systems
MARL has been widely used in such simulated domains
as grid worlds [1], but few applications have been found for
Picking staƟon
Storage shelves
Robot
Fig. 13: Simulation platform of the intelligent warehouse system.
9
realistic problems comparing to single-agent RL algorithms
[36]-[39] or MAS algorithms [40]-[42]. In this paper, we apply
the proposed NegoSI algorithm to an intelligent warehouse
problem. Previously, single agent path planning methods have
been successfully used in complex systems [43] [44], however,
the intelligent warehouse employs a team of mobile robots
to transport objects and single-agent path planning methods
frequently cause collisions [45]. So we solve the multi-agent
coordination problem in a learning way.
The intelligent warehouse is made up of three parts: picking
stations, robots and storage shelves (shown as in Fig. 13).
There are generally four steps of the order fulfillment process
for intelligent warehouse systems:
1) Input and decomposition: input and decompose orders
into separated tasks;
2) Task allocation: the central controller allocates the tasks
to corresponding robots using task allocation algorithms
(e.g., the Auction method);
3) Path planning: robots plan their transportation paths with
a single-agent path planning algorithm;
4) Transportation and collision avoidance: robots transport
their target storage shelves to the picking station and
then bring them back to their initial positions. During
the transportation process, robots use sensors to detect
shelves and other robots to avoid collisions.
We focus on solving the multi-robot path planning and
collision avoidance in a MARL way and ignore the first two
steps of the order fulfillment process by requiring each robot
to finish a certain number of random tasks. The simulation
platform of an intelligent warehouse system is shown as in
Fig. 13, which is a 16 × 21 grid environment and each grid
is of size 0.5m × 0.5m in real world. Shaded grids are storage
shelves which cannot be passed through. The state space of
each robot is made up of two parts: the location and the task
number. Each robot has 4 actions, namely, "up", "down", "left"
and "right".
We only compare the NegoSI algorithm with CQ-learning
for the intelligent warehouse problem. In fact, ILVFT has no
guarantee of the convergence characteristics and it is difficult
to converge in the intelligent warehouse setting in practice.
NegoVFT is also infeasible since its internal memory cost in
MATLAB is estimated to be 208 × 208 × 208 × 10 × 10 × 10 ×
4 × 4 × 4 × 8B = 4291GB, while the costs of other algorithms
are about 2 MB. Like NegoVFT, other MG-based MARL
algorithms cannot solve the intelligent warehouse problem
either. Experiments were performed in 2-agent and 3-agent
settings, respectively, and the results are shown as in Fig. 14
and Fig. 15.
In the 2-agent setting, the initial position and final goal of
robot 1 are (1,1) and those of robot 2 are (1,16). Each robot
needs to finish 30 randomly assigned tasks. The task set is the
same for all algorithms. Robots with NegoSI achieve lower
SEE (steps of each episode) than robots with CQ-learning
throughout the whole learning process (as shown in Fig.
14). NegoSI finally converges to 449.9 steps and CQ-learning
converges to 456.9 steps. In addition, robots with NegoSI have
higher and more stable REE (rewards of each episode). Finally,
the average runtime for completing all the tasks is 2227s for
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
50
0
−50
−100
0
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
90
80
70
60
0
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
80
60
40
20
0
0
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
80
60
40
20
0
0
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
95
90
85
80
75
70
0
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
90
80
70
60
0
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
80
60
40
20
0
0
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
80
60
40
20
0
0
10
Agent 1
Agent 2
10
5
15
episodes (· 100)
(d) NegoSI
20
Agent 1
Agent 2
20
Agent 1
Agent 2
20
Agent 1
Agent 2
20
Agent 1
Agent 2
20
Agent 1
Agent 2
Agent 3
20
10
5
15
episodes (· 100)
(d) NegoSI
10
5
15
episodes (· 100)
(d) NegoSI
10
5
15
episodes (· 100)
(d) NegoSI
10
5
15
episodes (· 100)
(d) NegoSI
10
5
15
episodes (· 100)
(d) NegoSI
Agent 1
Agent 2
10
5
15
episodes (· 100)
(a) CQ-learning
Agent 1
Agent 2
10
5
15
episodes (· 100)
(a) CQ-learning
20
20
Agent 1
Agent 2
20
5
15
episodes (· 100)
10
(a) CQ-learning
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
50
0
−50
−100
0
20
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
50
0
−50
−100
0
Agent 1
Agent 2
10
5
15
episodes (· 100)
(b) ILVFT
Agent 1
Agent 2
10
5
15
episodes (· 100)
(c) NegoQVFT
20
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
50
0
−50
−100
0
Fig. 7: REE (rewards of each episode) for each tested algorithm in ISR.
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
90
80
70
60
0
Agent 1
Agent 2
20
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
90
80
70
60
0
Agent 1
Agent 2
20
10
5
15
episodes (· 100)
(c) NegoQVFT
10
5
15
episodes (· 100)
(b) ILVFT
Fig. 8: REE (rewards of each episode) for each tested algorithm in SUNY.
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
80
60
40
20
0
0
Agent 1
Agent 2
20
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
80
60
40
20
0
0
Agent 1
Agent 2
20
10
5
15
episodes (· 100)
(c) NegoQVFT
10
5
15
episodes (· 100)
(b) ILVFT
Fig. 9: REE (rewards of each episode) for each tested algorithm in MIT.
20
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
80
60
40
20
0
0
Agent 1
Agent 2
20
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
80
60
40
20
0
0
Agent 1
Agent 2
20
10
5
15
episodes (· 100)
(c) NegoQVFT
10
5
15
episodes (· 100)
(b) ILVFT
Agent 1
Agent 2
10
5
15
episodes (· 100)
(a) CQ-learning
Fig. 10: REE (rewards of each episode) for each tested algorithm in PENTAGON.
Agent 1
Agent 2
20
5
15
episodes (· 100)
10
(a) CQ-learning
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
95
90
85
80
75
70
0
Agent 1
Agent 2
20
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
95
90
85
80
75
70
0
10
5
15
episodes (· 100)
(b) ILVFT
Agent 1
Agent 2
20
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
95
90
85
80
75
70
0
10
5
15
episodes (· 100)
(c) NegoQVFT
Fig. 11: REE (rewards of each episode) for each tested algorithm in GW nju.
i
e
d
o
s
p
e
h
c
a
e
f
o
s
d
r
a
w
e
r
100
80
60
40
20
0
0
Agent 1
Agent 2
Agent 3
10
5
15
episodes (· 100)
(a) CQ-learning
20
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
100
80
60
40
20
0
0
i
e
d
o
s
p
e
h
c
a
e
Agent 1
Agent 2
Agent 3
20
f
o
s
d
r
a
w
e
r
10
5
15
episodes (· 100)
(b) ILVFT
100
80
60
40
20
0
0
i
e
d
o
s
p
e
h
c
a
e
Agent 1
Agent 2
Agent 3
20
f
o
s
d
r
a
w
e
r
10
5
15
episodes (· 100)
(c) NegoQVFT
100
80
60
40
20
0
0
Fig. 12: REE (rewards of each episode) for each tested algorithm in GWa3.
11
TABLE II: Average runtime for each tested map.
CQ − learning
ILV FT
NegoQV FT
NegoSI
QL f or Single Policy
ISR
8.54
4.91
13.92
16.74
0.51
SUNY MIT
PENTAGON GW n ju GWa3
5.91
4.14
20.18
7.33
0.45
13.68
9.78
36.84
19.58
1.29
8.52
6.56
18.45
16.18
0.05
5.07
2.53
21.89
7.08
0.19
5.95
7.21
50.48
16.41
0.07
NegoSI and is 3606s for CQ-learning. Robots used 38% less
time to complete all the tasks with NegoSI than that with CQ-
learning.
The performances of NegoSI are also better in the 3-agent
setting than that of CQ-learning. The initial position and the
final goal of different robots are (1,1), (1,8) and (1,16). Each
robot needs to finish 10 randomly assigned tasks. The task
set is the same for different algorithms. SEE (steps of each
episode) for robots with NegoSI finally converges to 168.7
steps and that for robots with CQ-learning converges to 177.3
steps (as shown in Fig. 15). In addition, the learning curves
of NegoSI are more stable. Robots with NegoSI have higher
and more stable REE (rewards of each episode). Finally, the
average runtime for completing all the tasks is 1352s for
NegoSI and 2814s for CQ-learning. The robots use 52% less
time to complete all the tasks with NegoSI than that with CQ-
learning.
Remark 4: The agents with CQ-learning algorithm learn
faster than the agents with NegoSI for the benchmark maps
in Section IV-A, but slower than those with NegoSI for the
intelligent warehouse problem in Section IV-B. The reason is
that for agents with CQ-learning, the number of "coordination
state" is several times higher than that with NegoSI. This
difference becomes significant with the increase of the task
number, the environment scale and the number of agents. Thus,
the time used to search for one specific "coordination state" in
the "coordination state" pool increases faster in CQ-learning
than in NegoSI, which results in the increase of the whole
learning time. According to all these experimental results,
the presented NegoSI algorithm maintains better performances
regarding such characteristics as coordination ability, conver-
gence, scalability and computational complexity, especially for
practical problems.
V. CONCLUSIONS
In this paper a negotiation-based MARL algorithm with
sparse interactions (NegoSI) is proposed for the learning
and coordination problems in multi-agent systems. In this
e
d
o
s
p
e
i
h
c
a
e
f
o
s
p
e
t
s
650
600
550
500
450
400
0
NegoSI
CQlearning
20
40
episodes (· 100)
60
80
(a) SEE
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
2600
2550
2500
2450
2400
2350
0
NegoSI agent1
NegoSI agent2
CQlearning agent1
CQlearning agent2
60
80
20
40
episodes (· 100)
(b) REE
Fig. 14: SEE (steps of each episode) and REE (rewards of each episode) for NegoSI and CQ-learning in the 2-agent intelligent warehouse.
e
d
o
s
p
e
i
h
c
a
e
f
o
s
p
e
t
s
300
250
200
150
0
NegoSI
CQlearning
20
40
episodes (· 100)
60
80
(a) SEE
e
d
o
s
p
e
i
h
c
a
e
f
o
s
d
r
a
w
e
r
900
850
800
750
700
650
0
NegoSI agent1
NegoSI agent2
NegoSI agent3
CQlearning agent1
CQlearning agent2
CQlearning agent3
20
40
episodes (· 100)
60
80
(b) REE
Fig. 15: SEE (steps of each episode) and REE (rewards of each episode) for NegoSI and CQ-learning in the 3-agent intelligent warehouse.
integrated algorithm the knowledge transfer mechanism is
also adopted to improve agent's learning speed and coordina-
tion ability. In contrast to traditional sparse-interaction based
MARL algorithms, NegoSI adopts the equilibrium concept
and makes it possible for agents to select non-strict EDSP
or Meta equilibrium for their joint actions, which makes it
easy to find near optimal (or even optimal) policy and to
avoid collisions as well. The experimental results demonstrate
the effectiveness of the presented NegoSI algorithm regarding
such characteristics as fast convergence, low computational
complexity and high scalability in comparison to the state-of-
the-art MARL algorithms, especially for practical problems.
Our future work focuses on further comparison of NegoSI
with other existing MARL algorithms and more applications
of MARL algorithms to general and realistic problems. In
addition, multi-objective reinforcement learning (MORL) [46]-
[48] will also be considered to further combine environment
information and coordination knowledge for local learning.
ACKNOWLEDGEMENT
The authors would like to thank Dr. Yujing Hu and Dr.
Jiajia Dou for helpful discussion. The authors would also like
to thank the Associate Editor and the anonymous reviewers
for their constructive comments that have greatly improved
the original manuscript.
REFERENCES
[1] L. Bus¸oniu, R. Babuska and B. D. Schutter, "A comprehensive survey
of multi-agent reinforcement learning," IEEE Transactions on System,
Man, and Cybernetics, Part C: Applications and Reviews, vol. 38, no.
2, pp. 156-172, 2008.
[2] P. Vrancx, K. Verbeeck and A. Now´e, "Decentralized learning in markov
games," IEEE Transactions on Systems, Man, and Cybernetics, Part B:
Cybernetics, vol. 38, no. 4, pp. 976-981, 2008.
[3] R. Bianchi, M. F. Martins, C. H. Ribeiro and A. H. Costa, "Heuristically-
accelerated multiagent reinforcement learning," IEEE Transactions on
Cybernetics, vol. 44, no. 2, pp. 252-265, 2014.
[4] K. Hwang, S. Tan, M. Hsiao and C. Wu, "Cooperative multiagent
congestion control for high-speed networks," IEEE Transactions on
Systems, Man, and Cybernetics, Part B: Cybernetics, vol. 35, no. 2,
pp. 255-268, 2005.
[5] C. Yu, M. Zhang, F. Ren and G. Tan, "Multiagent learning of coor-
dination in loosely coupled multiagent systems," IEEE Transaction on
Cybernetics, vol. 45, no. 12, pp. 2853-2867, 2015.
[6] M. L. Littman, "Markov games as a framework for multi-agent rein-
forcement learning," in Proceedings of the International Conference on
Machine Learning (ICML-94), pp.157-163, New Brunswick, NJ, July
10-13, 1994.
[7] F. S. Melo and M. Veloso, "Decentralized MDPs with sparse interac-
tions," Artificial Intelligence, vol. 175, no. 11, pp. 1757-1789, 2011.
[8] Y. Hu, Y. Gao and B. An, "Multiagent reinforcement learning with
unshared value functions," IEEE Transaction on Cybernetics, vol. 45,
no. 4, pp. 647-662, 2015.
[9] Y. Hu, Y. Gao and B. An, "Accelerating multiagent reinforcement
learning by equilibrium transfer," IEEE Transaction on Cybernetics,
vol.45, no. 7, pp. 1289 - 1302, 2015.
[10] J. Hu and M. P. Wellman, "Nash Q-learning for general-sum stochas-
tic games," The Journal of Machine Learning Research, vol. 4, pp.
1039C1069, 2003.
[11] M. L. Littman, "Friend-or-foe Q-learning in general-sum games," in
Proceedings of
the International Conference on Machine Learning
(ICML-01), pp. 322-328, Williams College, Williamstown, MA, USA,
2001.
[12] A. Greenwald, K. Hall and R. Serrano, "Correlated Q-learning," in
the International Conference on Machine Learning
Proceedings of
(ICML-03), pp. 84-89, Washington, DC, USA, 2003.
12
[13] F. S. Melo and M. Veloso, "Learning of coordination: Exploiting sparse
interactions in multiagent systems," in Proceedings of the International
Conference on Autonomous Agents and Multiagent Systems (AAMAS),
pp. 773-780, 2009.
[14] Y. D. Hauwere, P. Vrancx and A. Now´e, "Learning multi-agent state
space representations," in Proceedings of the International Conference
on Autonomous Agents and Multiagent Systems (AAMAS), vol. 1, no. 1,
pp. 715-722, 2010.
[15] Y. D. Hauwere, P. Vrancx and A. Now´e, "Generalized learning automata
for multi-agent reinforcement learning," AI Communications, vol. 23, no.
4, pp. 311-324, 2010.
[16] W. Wang and Y. Jiang, "Community-aware task allocation for social
networked multiagent systems," IEEE Transactions on Cybernetics, vol.
44, no. 9, pp. 1529-1543, 2014.
[17] J. Enright and P. R. Wurman, "Optimization and coordinated auton-
omy in mobile fulfillment systems," Automated Action Planning for
Autonomous Mobile Robots, pp. 33-38, 2011.
[18] P. Stone and M. Veloso, "Multiagent systems: a survey from the machine
learning perspective," Autonomous Robots, vol. 8, no. 3, pp. 345-383,
2000.
[19] J. F. Nash, "Equilibrium points in n-person games," in Proceedings of
the National Academy of Sciences of the United States of America, vol.
36, no. 1, pp. 48-49, 1950.
[20] B. An, F. Douglis and F. Ye, "Heuristics for negotiation schedules in
multi-plan optimization," in Proceedings of the International Conference
on Autonomous Agents and Multiagent Aystems (AAMAS), vol. 2, pp.
551-558, 2008.
[21] B. An, V. Lesser, D. Westbrook and M. Zink, "Agent-mediated multi-
step optimization for resource allocation in distributed sensor networks,"
in Proceedings of the International Conference on Autonomous Agents
and Multiagent Systems (AAMAS), vol. 2, pp. 609-616, 2011.
[22] R. S. Sutton and A. G. Barto, "Reinforcement learning: an introduction,"
MIT press, 1998.
[23] C. Watkins, "Learning from delayed rewards," PhD thesis, University of
Cambridge, 1989.
[24] J. Tsitsiklis, "Asynchronous stochastic approximation and Q-learning,"
Machine Learning, vol. 16, no. 3, pp. 185-202, 1994.
[25] A. Now´e, P. Vrancx, and Y. D. Hauwere, "Game theory and multi-
agent reinforcement learning," Reinforcement Learning, Springer Berlin
Heidelberg, pp. 441-470, 2012.
[26] A. Burkov and B. Chaib-Draa, "Repeated games for multiagent systems:
a survey," The Knowledge Engineering Review, vol. 29, no. 1, pp. 1-30,
2014.
[27] A. Burkov and B. Chaib-Draa, "An approximate subgame-perfect equi-
librium computation technique for repeated games," Proceedings of the
Twenty-Fourth AAAI Conference on Artificial Intelligence, pp. 729-736,
2010.
[28] R. Porter, E. Nudelman and Y. Shoham, "Simple search methods for
finding a Nash equilibirum," Games and Economic Behavior, vol. 63,
no. 2, pp. 642-662, 2008.
[29] Y. Hu, Y. Gao and B. An, "Learning in Multi-agent Systems with
Sparse Interactions by Knowledge Transfer and Game Abstraction," in
Proceedings of the International Conference on Autonomous Agents and
Multiagent Systems (AAMAS), pp. 753-761, Istanbul, Turkey, 4-8 May,
2015.
[30] C. Guestrin, M. Lagoudakis and R. Parr, "Coordinated reinforcement
learning," in Proceedings of the International Conference on Machine
Learning (ICML-02), vol. 2, pp. 227-234. 2002.
[31] C. Guestrin, S. Venkataraman and D. Koller, "Context-specific multia-
gent coordination and planning with factored MDPs," in Proceedings of
the National Conference on Artificial Intelligence, pp. 253-259. 2002.
[32] J. R. Kok and N. A. Vlassis, "Sparse cooperative Q-learning," in
the International Conference on Machine Learning
Proceedings of
(ICML-04), pp. 61-68, Banff, Alberta, Canada, 2004.
[33] Y. D. Hauwere, P. Vrancx, and A. Now´e, "Solving sparse delayed
coordination problems in multi-agent reinforcement learning," Adaptive
and Learning Agents, Springer Berlin Heidelberg, pp. 114-133, 2012.
[34] P. Vrancx, Y. D. Hauwere and A. Now´e, "Transfer learning for multi-
agent coordination," in Proceedings of the International Conference on
Agents and Artificial Intelligence (ICAART), pp. 263-272, 2011.
[35] C. Claus and C. Boutilier, "The dynamics of reinforcement learning
in cooperative multiagent systems," in Proceedings of the fifteenth na-
tional/tenth conference on Artificial intelligence/Innovative applications
of artificial intelligence, pp. 746-752, July 26C30, 1998.
[36] D. Liu, X. Yang, D. Wang and Q. Wei, "Reinforcement-learning-based
robust controller design for continuous-time uncertain nonlinear systems
13
subject to input constraints," IEEE Transactions on Cybernetics, vol. 45,
no. 7, pp. 1372-1385, 2015.
[37] S. Haeri and L. Trajkovic, "Intelligent Deflection Routing in Buffer-
Less Networks," IEEE Transactions on Cybernetics, vol. 45, no. 2, pp.
316-327, 2015.
[38] D. Ernst, M. Glavic, F. Capitanescu and L. Wehenkel, "Reinforcement
learning versus model predictive control: a comparison on a power
system problem," IEEE Transactions on Systems, Man, and Cybernetics,
Part B: Cybernetics, vol. 39, no. 2, pp. 517-529, 2009.
[39] J. Valasek, J. Doebbler, M. D. Tandale and A. J. Meade, "Improved
adaptive-reinforcement
learning control for morphing unmanned air
vehicles," IEEE Transactions on Systems, Man, and Cybernetics, Part
B: Cybernetics, vol. 38, no. 4, pp. 1014-1020, 2008.
[40] L. Sabattini, C. Secchi and N. Chopra, "Decentralized estimation and
control for preserving the strong connectivity of directed graphs," IEEE
Transactions on Cybernetics, vol. 45, no. 10, pp. 2273-2286, 2015.
[41] S. Su, Z. Lin and A. Garcia, "Distributed synchronization control of
multiagent systems with unknown nonlinearities," IEEE Transactions
on Cybernetics, vol. 46, no. 1, pp. 325-338, 2016.
[42] H. Zhang, T. Feng, G. Yang and H. Liang, "Distributed cooperative
optimal control for multiagent systems on directed graphs: an inverse
optimal approach," IEEE Transactions on Cybernetics, vol. 45, no. 7,
pp. 1315-1326, 2015.
[43] L. Zhou, Y. Shi, J. Wang and P. Yang, "A balanced heuristic mechanism
for multirobot task allocation of intelligent warehouses," Mathematical
Problems in Engineering, vol. 2014, article ID 380480, 2014.
[44] H. H. Viet, P. H. Kyaw and T. Chung. "Simulation-based evaluations
of reinforcement learning algorithms for autonomous mobile robot path
planning," IT Convergence and Services, Springer Netherlands, pp. 467-
476, 2011.
[45] P. R. Wurman, R. D'Andrea and M. Mountz, "Coordinating hundreds
of cooperative, autonomous vehicles in warehouses," AI Magazine, vol.
29, no. 1, pp. 9-20, 2008.
[46] P. Vamplew, R. Dazeley, A. Berry, R. Issabekov and E. Dekker,
"Empirical evaluation methods for multiobjective reinforcement learning
algorithms," Machine Learning, vol. 84, no. 1-2, pp. 51-80, 2011.
[47] C. Lin and I. Chung, "A reinforcement neuro-fuzzy combiner for
multiobjective control," IEEE Transactions on Systems, Man, and Cy-
bernetics, Part B: Cybernetics, vol. 29, no. 6, pp. 726-744, 1999.
[48] C. Liu, X. Xu and D. Hu, "Multiobjective reinforcement learning: A
comprehensive overview," IEEE Transactions on Systems, Man, and
Cybernetics: Systems, vol. 45, no. 3, pp. 385-398, 2015.
|
1009.1132 | 2 | 1009 | 2010-09-28T17:17:54 | Efficient Collaborative Application Monitoring Scheme for Mobile Networks | [
"cs.MA",
"cs.CR",
"cs.DC"
] | New operating systems for mobile devices allow their users to download millions of applications created by various individual programmers, some of which may be malicious or flawed. In order to detect that an application is malicious, monitoring its operation in a real environment for a significant period of time is often required. Mobile devices have limited computation and power resources and thus are limited in their monitoring capabilities. In this paper we propose an efficient collaborative monitoring scheme that harnesses the collective resources of many mobile devices, "vaccinating" them against potentially unsafe applications. We suggest a new local information flooding algorithm called "TTL Probabilistic Propagation" (TPP). The algorithm periodically monitors one or more application and reports its conclusions to a small number of other mobile devices, who then propagate this information onwards. The algorithm is analyzed, and is shown to outperform existing state of the art information propagation algorithms, in terms of convergence time as well as network overhead. The maximal "load" of the algorithm (the fastest arrival rate of new suspicious applications, that can still guarantee complete monitoring), is analytically calculated and shown to be significantly superior compared to any non-collaborative approach. Finally, we show both analytically and experimentally using real world network data that implementing the proposed algorithm significantly reduces the number of infected mobile devices. In addition, we analytically prove that the algorithm is tolerant to several types of Byzantine attacks where some adversarial agents may generate false information, or abuse the algorithm in other ways. | cs.MA | cs |
Efficient Collaborative Application Monitoring Scheme for Mobile Networks
New operating systems for mobile devices allow their users to download millions of applications
created by various individual programmers, some of which may be malicious or flawed. In order to
detect that an application is malicious, monitoring its operation in a real environment for a significant
period of time is often required. Mobile devices have limited computation and power resources and
thus are limited in their monitoring capabilities. In this paper we propose an efficient collaborative
monitoring scheme that harnesses the collective resources of many mobile devices, "vaccinating"
them against potentially unsafe applications. We suggest a new local information flooding algorithm
called TTL Probabilistic Propagation (TPP). The algorithm is implemented in any mobile device,
periodically monitors one or more application and reports its conclusions to a small number of other
mobile devices, who then propagate this information onwards, whereas each message has a predefined
TTL. The algorithm is analyzed, and is shown to outperform existing state of the art information
propagation algorithms, in terms of convergence time as well as network overhead. The maximal
load of the algorithm (the fastest arrival rate of new suspicious applications, that can still guarantee
complete monitoring), is analytically calculated and shown to be significantly superior compared to
any non-collaborative approach. Finally, we show both analytically and experimentally using real
world network data received among others from the Reality Mining Project, that implementing the
proposed algorithm significantly reduces the number of infected mobile devices.
In addition, we
analytically prove that the algorithm is tolerant to several types of Byzantine attacks where some
adversarial agents may generate false information, or abuse the algorithm in other way.
I.
INTRODUCTION
The market share of Smart-phones is rapidly increasing and is expected to increase even faster with the introduction
of 4th generation mobile networks, reaching from 350 million in 2009 to one billion by 2012 [9]. Companies that are
distributing new mobile devices operating systems had created marketplaces that motivate individuals and other
companies to introduce new applications (such as Apple's App Store Google's Android Market, Nokia's Ovi Store
and others). The main assumption behind these marketplaces is that users will prefer a mobile device based on
an operating system with larger marketplace offerings. It is expected that in the future, various communities will
develop additional alternative marketplaces that will not be regulated. These marketplaces will allow mobile users to
download from a variety of millions of new applications. An example for such a marketplace is GetJar, offering 60,000
downloadable applications for over 2,000 models of mobile devices, counting a total of over 900 million downloads by
August 2010 [25]. The content of most marketplaces is currently not verified by their operators and thus some of the
applications they contain may be malicious or contain faulty code segments. Downloading a malicious application from
the marketplace is not the only way that a mobile device may be infected by malicious code. This may also happen
as a result of a malicious code that manages to exploit a vulnerability in the operating systems and applications or
through one of the mobile phone communication channels such as Bluetooth, Wi-Fi, etc' [26, 29]. McAfee's Mobile
Security Report for 2008 states that nearly 14% of global mobile users have been directly infected or have known
someone who was infected by a mobile virus (this number had even increased in the following year) [35, 36].
In
many cases, in order to detect that an application is malicious, monitoring its operation in a real environment for
a significant period of time is required. The monitored data is being processed using advanced machine learning
algorithms in order to assess the maliciousness of the application. For a variety of techniques for local monitoring of
mobile phone applications see [28, 37 -- 39].
In recent years most of the prominent security and privacy threats for communication networks had relied on the use
of collaborative methods (e.g., Botnets). The danger stemming from such threats is expected to significantly increase
in the near future, as argued in [29, 49] and others. The amount of resources a single unit may allocate in order to
defend from such threats without interfering with its routine work is very limited. Therefore, the development of an
efficient collaborative defense infrastructure for mobile users is strongly required [51].
In this work we propose such a collaborative application monitoring infrastructure, that is capable of dramatically
decreasing the susceptibility of mobile devices to malicious applications. Indeed, for every new malicious application
that is introduced to the network, there will always be a few users that will encounter it for the first time, and
subsequently may be exposed to its negative effects. However, by wisely assimilating this information throughout the
network, based on their (poor) experience the vast majority of the mobile devices will be notified on the properties
of the new malicious application, well ahead before encountering it. This behavior resembles mammals vaccination
system, as a new threat have some (small) probability of damaging the organism (or network), but once defeated, all
the cells of the organism are familiar with this threat, and will be able to easily overcome it (or avoid installing it)
should they encounter it in the future. As the portion of "new" applications (namely, applications that no network
user had ever experienced with) is rather small, the mobile devices are kept protected at all times against the vast
majority of potentially harmful marketplace applications (this property of the proposed scheme is illustrated in Figure
1).
This paper presents and analyzes an algorithm that efficiently implements this proposed scheme. The algorithm is
shown to provide fast convergence, low network overhead, and is capable of handling a significantly higher number of
potential malicious applications concurrently, than any other collaborative or non collaborative approach. In addition,
the algorithm is fault tolerance for several known adversarial and Byzantine attacks. These features of the algorithm
are shown both analytically and empirically, using simulation of various random networks, as well as using real life
network that are based on MIT's Reality Mining project[17].
2
FIG. 1: Each month mobile users download and install various kinds of mobile applications. These applications are comprised
of "new applications" (namely, applications that were only introduced to the marketplace recently), and "old applications"
(applications that were released to the marketplace in previous months). Whereas new applications may contain a few ma-
licious applications, the vast majority of malicious applications are already in the marketplace for some time. Using the
proposed scheme, users of mobile devices become "vaccinated" against all of the potentially harmful applications among the
old applications. As a result, the number of infected mobile devices decreases dramatically.
In this paper we present a collaborative monitoring algorithm, called TPP -- TTL Probability Propagation. The
algorithm uses a time to live counter for each alerting message, that is logarithmic in the number of the mobile
devices, n. Using these short lived messages that are sent between the network's devices the algorithm is analytically
shown to provide a prior knowledge of every potential malicious application to the vast majority of the network users.
The fastest arrival rate of new applications for which this property can still be maintained (e.g. the maximal load
of the algorithm) is analytically calculated (see Theorem 5), shown to be far superior than any non-collaborative
host-based monitoring approach (Corollary 1) or collaborative approach. Furthermore, we show that the local cost of
the algorithm (namely, the amount of local efforts required from each participating user) monotonically decreases as
the network's size increases.
Using real-world numbers, implemented as a service executed by 1,000,000 mobile devices, assuming that each device
can locally monitor only a single application per week, and that the requested vaccination efficiency is 99% (namely,
we allow only 1% of the network's users to be susceptible to attacks), the system is guaranteed to collaboratively
monitor 25,566 new applications each month (achieved using an average of 2.763 messages per device per day). A
single non-collaborative device under the same requirements would be able to monitor only 5 new applications each
month. A similar result was also demonstrated on a real world network, using the data collected in the Reality Mining
Project [17] (see Section IX).
In addition, the algorithm is shown to be partially fault tolerant to the presence of Byzantine devices, that are
capable of distributing messages containing false information, or selectively refuse to forward some of the messages
they receive[52] (see Theorem 6 in Section VI discussing devices who distribute false messages and Corollaries 3, 4
and 5 in Section VII concerning devices who block passing messages).
Furthermore, we show that the cost of defending against adversarial abuse of the proposed algorithm grows asymp-
totically slow whereas the cost of increasing the strength of such attacks grows approximately linearly. This property
of the algorithm makes it economically scalable, and desirable to network operators (see more details in Section VI
and specifically an illustration that appears in Figure 4).
The efficiency of the algorithm stems from the generation of an implicit collaboration between a group of random
walking agents, released from different sources in the network (and at different times). This technique is novel, as
most related works do not take into consideration that the agents might be released collaboratively from different
3
sources [15, 16]. Hence, analysis of such systems was limited to the probabilistic analysis of the agents' movements.
The rest of this work is organized a follows : Related work and comparison to state-of-the-art techniques are
presented in Section II. The threat model is formally defined in Section III. The TPP collaborative algorithm is
presented in IV, and its performance is analyzed. The algorithm's robustness with regards to attacks by adversaries is
discussed in Sections VI and VII. An alternative model which does not require a random network overlay assumption
is discussed in Section VIII. A wealth of experimental results are presented in Section IX, where conclusions and
suggestions for future research appears in Section X.
In addition, this paper contains an appendix, discussing in
details several issues that are mentioned during the paper.
II. RELATED WORK
Flooding a network with messages intended for a large number of nodes is arguably the simplest form of information
dissemination in communication networks (specifically when previous knowledge about the network topology is limited
or unavailable). Since the problem of finding the minimum energy transmission scheme for broadcasting a set of
messages in a given network is known to be NP-Complete [8], flooding optimization often relies on approximation
algorithms. For example, in [43] messages are forwarded according to a set of predefined probabilistic rules, whereas
[42] discusses a deterministic algorithm, approximating the connected dominating set of each node.
In this work we apply a different approach -- instead of a probabilistic forwarding of messages, we assign a TTL
value for each message, guiding the flooding process. The analysis of the algorithm is done by modeling the messages
as agents practicing random walk in a random graph overlay of the network[53]. The optimal value of this TTL is
shown, guaranteeing a fast completion of the task, while minimizing the overall number of messages sent. The use of
a TTL dependent algorithm was previously discussed in [1] (demonstrating O(log2 n) propagation completion time)
and [34] (where no analytic bound over the completion time was demonstrated).
Flooding Algorithms. The simplest information propagation technique is of course the flooding algorithm. It is well
known that the basic flooding algorithm, assuming a single source of information, guarantees propagation completion
in a worse case cost of O(n2) messages (O(E) in expanders) and time equals to the graph's diameter, which in
log(n·p) ) ≈ O(log n) [12, 18]. We later show that the TPP algorithm
the case of a random graph G(n, p) equals O( log n
discussed in this work achieves in many cases the same completion time, with a lower cost of O(n log n) messages
(Theorems 3 and 4).
Other flooding algorithms variants use various methods to improve the efficiency of the basic algorithm. One
example is a probabilistic forwarding of the messages in order to reduce the overall cost of the propagation process,
optionally combined with a mechanism for recognizing and reducing repetitions of packets transitions by the same
device [41]. Other methods may include area based methods [41] or neighborhood knowledge methods [44]. In many of
these works, completion time is traded for a reduced overall cost, which results in a similar cost as the TPP algorithm
proposed in this work (namely, O(n log n)), but with a significantly higher completion time. Additional details on
various flooding algorithms can be found in [50].
An extremely efficient flooding algorithm in terms of completion time is the network coded flooding algorithm,
discussed in [13]. In this work, dedicated to G(n, p) random graphs, a message is forwarded by any receiving vertex
k
d(v) times, while k is a parameter which depends on the network's topology [21]. Using this method, the algorithm
achieves a completion time of approximately O( n3
E2 ). This algorithm, however, is still outperformed by the TPP
algorithm.
Epidemics Algorithms. An alternative approach to be mentioned in this scope is the use of epidemic algorithms [5,
47]. There exist a variety of epidemic algorithms, starting with the basic epidemic protocol [14],through neighborhood
epidemics [22] and up to hierarchical epidemics [46].
In general, all the various epidemic variants has a trade-off
between number of messages sent, completion time, and previous knowledge required for the protocols (concerning
the structure of the network). However, the most efficient results of such approaches are still outperformed by the
TPP algorithm.
Distributed Coverage Algorithms. A different approach for a collaborative assimilation of an important piece
of information throughout the network is the use of cooperative exploration algorithms (in either known or unknown
environments), guaranteeing that all (or a large enough portion) of the graph is being "explored" by agents carrying the
alerting messages. Planar networks can be sampled into R2, and then be collaboratively explored by a decentralized
group of myopic agents (see swarm coverage examples in [2, 31]).
In [45] a swarm of ant-like robots is used for repeatedly covering an unknown area, using a real time search method
called node counting. Using this method, the agents are analytically shown to cover the network graph efficiently.
Another algorithm to be mentioned in this scope is the LRTA* algorithm [32], that was shown in [30] to guarantee
cover time of O(n2) in degree bounded undirected planar graphs. Interestingly, in such graphs, the random walk
4
algorithm is also known to require at most O(n2) time (and at least Ω(n(log n)2)) [27].
In [48] a collaborative algorithm that guarantees coverage of regions in the Z2 grid, in O(n1.5) time, using extremely
simple and myopic "ant like" mobile agents and using no direct communication by the agents is discussed. In [3]
this algorithm was later shown to guarantee (under some limitations) coverage of dynamically expanding domains in
O(n2.5 log n) time.
Summary. The maximal arrival rate of new applications for which a successful decentralized monitoring can still be
guaranteed (namely, the maximal load ), of the TPP algorithm is analyzed in Section V.
As previously discussed, the efficiency of the TPP algorithm is derived from the fact that participating devices form
a collaborative infrastructure, using alerting messages that can be sent between each two users -- which implicitly
assumes the existence of an appropriate network overlay. The algorithm can still be implemented for any given
network topology, assuming no network overlays. This variant of the algorithm is discussed in Section VIII. It should
also be mentioned that as the TPP algorithm uses random elements, it requires approximately O(ln2 n) random bits
by each device for its proper execution.
III. THE COLLABORATIVE APPLICATION MONITORING PROBLEM
Given a mobile network of n devices, let us denote the network's devices by V = {v1, v2, . . . , vn}. Note that
the network's topology may be dynamic[54]. Each device may occasionally visit the marketplace, having access to
N new downloadable applications every month. We assume that downloading of applications is done independently,
namely -- that the probability that a user downloads application a1 and the probability that the same user downloads
application a2 are uncorrelated. Let us denote the average number of applications each user downloads each month
by η, out of which ηn are applications that were introduced to the marketplace this month ("new applications")
and ηo are applications that were introduced to the marketplace in previous months ("old applications"), such that
ηo + ηn = η.
For some malicious application ai, let pai denote the application's penetration probability -- the probability that
given some arbitrary device v, it is unaware of the maliciousness of ai. The penetration probability of every new
malicious application is 1. Our goal is to verify that at the end of the month, the penetration probability of all malicious
applications released during this month are lower than a penetration threshold pM AX , resulting in a "vaccination" of
the network with regards to these applications. This way, although a new malicious application may infect a handful
of devices -- the first one it encounters -- the rest of the network would quickly become immune to it. Formally, for
some small we require that :
∀Malicious application ai P rob (pai > pM AX ) <
The rational behind the use of the threshold pM AX is increasing the efficiency of the collaborative system, defending
against dominant malicious applications. Given additional available resources, the parameter pM AX can be decreased,
resulting in a tighter defense grid (traded for decreased load and increased messages overhead).
We assume that any device v can send a message of some short content to any other device u. In addition, we
assume that at the initialization phase of the algorithm each device is given a list containing the addresses of some X
random network members. This can be implemented either by the network operator, or by distributively constructing
and maintaining a random network overlay.
A revised model that does not require this assumption is analyzed in Section VIII.
We assume that each device can locally monitor applications that are installed on it (as discussed for example in
[4, 38]). However, this process is assumed to be expensive (in terms of the device's battery), and should therefore be
executed as fewer times as possible. The result of an application monitoring process is a non-deterministic boolean
value : {true, f alse}.
False-positive and false-negative error rates are denoted as :
P (Monitoring(ai) = trueAi is not malicious) = E+
P (Monitoring(ai) = falseAi is malicious) = E−
We assume that the monitoring algorithm is calibrated in such way that E+ ≈ 0.
As we rely on the propagation of information concerning the maliciousness of applications, our system might be
abused by injection of inaccurate data. This may be the result of a deliberate attack, aimed for "framing" a benign
application (either as a direct attack against a competitive application, or as a more general attempt for undermining
the system's reliability), or simply as a consequence of a false-positive result of the monitoring function. Therefore,
in order for a device v to classify an application ai as malicious, one of the following must hold :
• Device v had monitored ai and found it to be malicious.
• Device v had received at least ρ alerts concerning ai from different sources (for some decision threshold ρ).
In addition, note that the information passed between the devices concerns only applications discovered to be
malicious. Namely, when an application is found to be benign, no message concerning this is generated. This is
important not only for preserving a low messages overhead of the algorithm but also to prevent malicious applications
from displaying a normative behavior for a given period of time, after which they initiate their malicious properties.
In such cases, soon after an application exits its "dormant" stage, it will be detected and subsequently reported,
generating a "vaccination reaction" throughout the network.
5
IV. ALGORITHM, CORRECTNESS & ANALYSIS
We shall now present the TPP algorithm. The algorithm is executed by each device separately and asynchronously,
where no supervised or hierarchical allocation of tasks, as well as no shared memory are required. Following is a list
of the main notations used in the presentation and analysis of the proposed algorithm.
n The number of devices participating in the TPP algorithm
η Average number of applications each user downloads each month
ηn Average number of "new applications" users download each month
ηo Average number of "old applications" users download each month
X The number of alert messages generated and sent upon the discovery of a malicious application
pN The ratio X
n
pM AX Penetration threshold -- the maximal percentage of non-vaccinated devices allowed by the network operator
E− False negative error rate of the local monitoring mechanism
T Delay time between each two consecutive local monitoring processes
N Number of new applications introduced to the network each month
pai The penetration probability of application ai
1 − Confidence level of the correctness of the convergence time estimation
α The polynomial confidence level lnn −1
λA Applications arrival rate (number of new applications per time unit)
λA Maximal applications arrival rate that a non-collaborative device can monitor
λM Monitoring rate (number of applications monitored per time unit)
λT Number of algorithm's time steps per time unit
timeout The Time-To-Live counter assigned to alert messages
ρ Decision threshold -- the number of alerts a device must receive in order to classify an application as malicious
CS Cost of sending a single message
CM Cost of locally monitoring a single application
6
A. TTL Probabilistic Propagation (TPP) -- a Collaborative Monitoring Algorithm
The TPP algorithm conceptually relies on the fact that in order to "vaccinate" a network with regards to malicious
applications, it is enough that only a small number of devices will monitor this application. This way, although the
application monitoring process is relatively expensive (in terms of battery and CPU resources), the amortized cost
of monitoring each malicious application is kept to a minimum. A detailed implementation of the TPP algorithm
appears in Algorithm 1.
At its initialization (lines 1 through 6), all the applications installed on the device are added to a list of suspected
applications. In addition, an empty list of known malicious applications is created. Once an application is determined
as malicious, it is added to the known malicious applications list. In case this application was also in the suspected
applications list (namely, it is installed on the device, but has not been monitored yet), it is deleted from that list.
Once a new application is encountered it is compared to the known malicious applications list, and if found, an alert is
sent to the user (alternatively, the application can be chosen to be uninstalled automatically). This feature resembles
the long-term memory of the immune system in living organisms.
If the new application is not yet known to be
malicious, the application is added to the suspected applications list.
Once executed, a periodic selection of an arbitrary application from the list of suspected applications is done, once
every T steps (lines 23 through 31). The selected application is then monitored for a given period of time, in order
to discover whether it is of malicious properties (see details about such monitoring in Section III). If the application
is found to be malicious, it is removed from the list of suspected applications and added to the known malicious
applications list (lines 28 and 27). In addition, an appropriate alert is produced and later sent to X random devices.
The alert message is also assigned a specific TTL value (lines 29 and 30). Once a network device receives such an
alert message it automatically forwards it to one arbitrary device, while decreasing the value of TTL by 1. Once TTL
reaches zero, the forwarding process of this message stops (lines 21 and 22). In case a monitored application displays
no malicious properties, it is still kept in the list of suspicious applications, for future arbitrary inspections.
A device may also classify an application as malicious as a result of receiving an alert message concerning this
application (lines 14 through 20). In order to protect begins applications from being "framed" (reported as being
malicious by adversaries abusing the vaccination system), a device classifies an application as malicious only after it
receives at least ρ messages concerning it, from different sources (for a pre-defined decision threshold ρ). Note that
when a device v receives an alert message concerning application ai, it still forwards this message (assuming that
T T L > 0), even when v has not yet classified ai as malicious (for example, when the number of alert messages received
is still smaller than ρ). When the ρ-th alerting message concerning an application is received, the device adds the
application to its known malicious applications list and removes it from the suspected applications list if needed. The
selection of the optimal value of ρ is discussed in Section VI. The values of T , ρ and T T L, as well as the number of
generated alert messages can be determined by the network operators, or be changed from time to time according to
the (known or estimated) values of n and N , and the penetration threshold pM AX . Selecting an optimal value for
T T L is discussed in Section IV.
B. Optimal Parameters for Guaranteeing Successful Monitoring
Outline of Analysis. In order to analyze the algorithm's behavior, we shall model the movements of the notification
messages between the network's devices as random walking agents, traveling in a random graph G(n, p). Taking into
account the fact that the messages have limited lifespan (namely, T T L), a relation between the size of the graph
and the lifespan of the agents is produced. Having the value of T T L that guarantees a coverage of the graph, the
algorithm's completion time, as well as the overall number of messages sent, can then be calculated.
While analyzing the performance of the TPP algorithm we imagine a directed Erdos-Renyi random graph G(V, E) ∼
G(n, pN ), where pN = X
n . The graph's vertices V denote the network's devices, and the graph's edges E represent
messages forwarding connections. Notice that as G is a random graph, it can be used for the analysis of the performance
of the TPP algorithm, although the message forwarding connections of the TPP algorithm are dynamic (notice again
that an alternative model which does not require this randomness assumption is discussed in Section VIII).
In
addition, although the identity of the "neighbors" of a vertex v in the real network overlay may change from time to
time (as the overlay graph can be dynamic), it can still be modeled by static selection of X random neighbors of v.
Observing some malicious application ai, every once in a while some device which ai is installed on randomly selects
it for monitoring. With probability (1− E−) the device discovers that ai is malicious and issues alerts to X network's
members. We look at these reports as the system's "agents", and are interested in finding :
• The time it takes the graph to be explored by the agents. Namely, the time after which every device was visited
by at least ρ agents (and is now immune to ai).
7
Initialize A(v) ← A(v)
1: Initialization :
2:
3:
4:
5:
6:
Let A(v) be the list of installed applications
Let A(v) be the list of suspected applications
Let (cid:101)A(v) be the list containing known malicious applications
Initialize (cid:101)A(v) ← ∅
If ai ∈ (cid:101)A(v) then
7: Interrupt upon encountering a new application ai :
8:
9:
10:
11:
12:
A(v) ← A(v) \ {ai}
Issue an alert to the user concerning ai
A(v) ← A(v) ∪ {ai}
End if
If j ≥ ρ then
(cid:101)A(v) ← (cid:101)A(v) ∪ {ai}
A(v) ← A(v) \ {ai}
If ai ∈ A(v) then
Issue an alert to the user concerning ai
13: Interrupt receives malicious application ai notice, for the j-th time :
14:
15:
16:
17:
18:
19:
20:
21:
22:
End if
Decrease T T L of report by 1
Forward report to a random network member
End if
Select a random application ai from A(v) for monitoring
If ai is found to be malicious then
(cid:101)A(v) ← (cid:101)A(v) ∪ {ai}
23: Execute every T time-steps :
24:
25:
26:
27:
28:
29:
30:
31:
Issue an alert to the user concerning ai
A(v) ← A(v) \ {ai}
Report ai to X random network members
Set T T L = timeout
End if
Algorithm 1: TTL Probabilistic Propagation
• Total number of messages sent during this process.
• The minimal TTL which guarantees a successful vaccination of the network.
Note that the agents have a limited lifespan, equals to timeout. As the graph is a random graph, the location of the
devices in which ai is installed is also random. Therefore, as they are the sources of the agents, we can assume that
the initial locations of the agents are uniformly and randomly distributed along the vertices of V . In compliance with
the instruction of the TPP algorithm, the movement of the agents is done according to the random walk algorithm.
Application ai is installed on n · pai devices, each of which monitors a new application every T time steps. Upon
selecting a new application to monitor, the probability that such a device will select ai is 1
N . The probability that
upon monitoring ai the device will find it to be malicious is (1 − E−), in which case it will generate n · pN alerting
messages. The expected number of new agents created at time t, denoted as k(t), therefore equals :
(cid:80)
k(i).
k(t) =
n2 · pai · pN
T · N
(1 − E−)
and the accumulated number of agents which have been generated in a period of t time-steps is therefore kt =
i≤t
The value of timeout (the assigned TTL) is selected in such a way that the complete coverage of the graph, and
therefore its vaccination against ai, is guaranteed (in probability greater than 1 − ). We now artificially divide the
mission to two phases, the first containing the generation of agents and the second discussing the coverage of the
graph. Note that this division ignores the activity of the agents created in the second phase. Note also that the fact
that the agents are working in different times (and in fact, some agents may already vanish while others have not even
been generated yet) is immaterial. The purpose of this technique is to ease the analysis of the flow of the vaccination
process.
Denoting the completion time by TV ac we therefore have :
TV ac ≤ TGeneration + TP ropagation
It is easy to see that essentially TP ropagation (cid:44) timeout. We now artificially set :
timeout = λ · (TGeneration + timeout)
8
From this we can see that :
TGeneration = (1 − λ) · (TGeneration + timeout)
TGeneration =
(1 − λ)
λ
· timeout
We later demonstrate an upper bound for timeout, based on the number of agents created in t ≤ TGeneration
(ignoring the activity of the agents created between t = TGeneration and t = TGeneration + timeout).
In this case, we can now write: TV ac ≤ 2 · timeout.
We now examine the case of λ = 0.5 (which we show to be the optimal selection for λ, in the paper's Appendix).
Let us denote the number of agents created in the first TGeneration time-steps by k = kTGeneration. We now find the
time it takes those k agents to completely cover the graph G, and from this, derive the value of timeout.
Let us denote by ρ-coverage of a graph the process the result of which is that every vertex in the graph was visited
by some agent at least ρ times.
Theorem 1. The time it takes k random walkers to complete a ρ-coverage of G in probability greater than 1 −
(denoted as T(n)) can be bounded as follows :
2(cid:0)ρ − ln
(cid:1)
−
n
3k
2n(1− 1
ln n
)
≤ T (n) ≤ 2(cid:0)ρ − ln
n
(cid:1)
1 − e− k
2n
Proof. See Appendix for complete proof.
1 − e
We now show how to select a value of timeout that guarantees a successful vaccination process :
Theorem 2. For every values of timeout that satisfy the following expression, the TPP algorithm is guaranteed to
achieve successful vaccination for any penetration threshold pM AX , in probability greater than 1 − :
2(cid:0)ρ − ln
(cid:1)
n
(cid:88)
timeout(1 − e−timeout· n·pM AX ·pN
2T ·N
(1−E−))
= 1
E[k] =
i≤TGeneration
n2 · pai · pN
T · N
(1 − E−)
Proof. Recalling the expected number of agents generated at each time step, the expected number of agents k that
appears in Theorem 1 equals :
A successful termination of TPP means that the penetration probability of (any) malicious application is decreased
below the threshold pM AX . Until this is achieved, we can therefore assume that this probability never decreases below
pM AX :
Therefore, we can lower bound the number of agents as follows :
∀t < TGeneration : pai ≥ pM AX
k ≥ timeout · n2 · pM AX · pN
(1 − E−)
T · N
2(cid:0)ρ − ln
(cid:1)
n
− n·pM AX ·pN
2T ·N·timeout−1 (1−E−)
Assigning timeout = m into Theorem 1, successful vaccination is guaranteed for :
timeout =
2(ρ − ln
n )
1 − e− k
2n
≤
1 − e
and the rest is implied
C. Number of Messages and Time Required for Vaccination
From the value of timeout stated in Theorem 2, the vaccination time TV ac as well as the overall cost of the TPP
algorithm can now be extracted. The cost of the algorithm is measured as a combination of the overall number of
messages sent during its execution and the total number of monitoring activities performed. Let us denote the cost
of sending a single message by CS and the cost of executing a single local monitoring process by CM .
Observation 1. For any timeout = τ which satisfies Theorem 2, the time and cost of the TPP algorithm can be
expressed as :
9
TV ac = O (τ )
(cid:18) pM AX · pN
n−2T · N
O
, M = O(k · τ · CS +
k
X
CM ) =
· (1 − E−) ·
τ 2 · CS +
τ
n · pN
· CM
(cid:19)(cid:19)
(cid:18)
(cid:18)
Theorem 3. The time it takes a network that implements the TPP algorithm to guarantee vaccination against all
the malicious applications from an N available applications is upper bounded as follows :
(cid:115)
TV ac ≤ 4
T · N (ρ + (α + 1) ln n)
n · pM AX · pN · (1 − E−)
= O
ρ + ln n +
T · N
pM AX · (1 − E−) · ln n
(cid:19)
Proof. Let us assume that is polynomial in 1
Using the bound (1 − x) < e−x for x < 1 we can see that when assuming [55] that :
n , namely : = n−α s.t. α ∈ Z+.
timeout · n · pM AX · pN
2T · N
(1 − E−) < 1
Theorem 2 can be written as :
and therefore :
ρ + (α + 1) ln n ≥ timeout2 · n · pM AX · pN · (1 − E−)
4T · N
(cid:115)
timeout ≤
4T · N (ρ + (α + 1) ln n)
n · pM AX · pN · (1 − E−)
Assigning this approximation of timeout into the assumption above, yields :
(cid:115)
2T · N
n · pM AX · pN (1 − E−)
· 2 (ρ + (α + 1) ln n) <
2T · N
n · pM AX · pN (1 − E−)
which is satisfied for the following values of pN :
pN <
T · N
n · pM AX · (ρ + (α + 1) ln n)(1 − E−)
For constant (or smaller) values of pM AX and number of applications larger than O(ln n) we can safely assign pN = ln n
n .
As the vaccination time equals twice the value of timeout, the rest is implied.
Assigning the upper bound for pN into Theorem 3 immediately yields O(ρ + ln n). When assigning ln n
n as a lower
bound for pN which guarantees connectivity [12], the following expression is received :
(cid:115)
TV ac ≤ 4
T · N · (ρ + (α + 1) ln n)
ln n · pM AX · (1 − E−)
However, using the upper bound for pN derived above, we see that :
ln n
n
<
T · N
n · pM AX · (ρ + (α + 1) ln n)(1 − E−)
which in turn implies that :
Combining the two yields :
(cid:16)
(cid:17)
Note that although O
into the upper bound for pN we can see that :
pM AX (1−E−)
T·N
10
ρ + (α + 1) ln n <
pM AX · (1 − E−) · ln n
T · N
(cid:18)
TV ac = O
(cid:19)
T · N
pM AX · (1 − E−) · ln n
is allegedly independent of n, assigning the connectivity lower bound PN > ln n
n
T · N
pM AX (1 − E−)
= Ω(ρ ln n + ln2 n)
Theorem 4. The overall cost (messages sent and monitoring) of a network that implements the TPP algorithm for
guaranteeing vaccination against all the malicious applications from an N available applications is upper bounded as
follows :
M ≤ k · timeout · CS +
· CM ≤
k
X
≤ 4n (ρ + (α + 1) ln n) CS + 2CM
n (ρ + (α + 1) ln n) · pM AX · (1 − E−)
=
pN · T · N
(cid:115)
(cid:18) n
ln n
(cid:18)
O
(nρ + n ln n)CS +
+ n(ρ + ln n)
pM AX (1 − E−)
T · N
(cid:19)
(cid:19)
CM
Proof. When the vaccination process is completed, no new messages concerning the malicious application are sent.
The above is received by assigning the approximated value of timeout into Observation 1.
In networks of E− < 1 − o(1), provided that[56] ρ = O(ln n), and remembering that in this case
T·N
pM AX (1−E−) =
Ω(ln2 n) Theorem 4 is dominated by :
M = O
(cid:16)
(cid:17)
n ln nCS +
n
ln n
CM
V. MAXIMAL LOAD OF THE TPP ALGORITHM
We shall now calculate the maximal load of the TPP algorithm, namely, the maximal rate at which new applications
can be introduced into the system, while guaranteeing that each of them would still be monitored by enough devices.
Definition 1. Let λA denote the rate at which new applications are introduced to the system (units are #applications
per time units).
Definition 2. Let λM denote the rate at which each device monitors an arbitrary new application (units are #appli-
cations per time units).
Definition 3. Let λT denote the rate at which alert messages are processed throughout the system (namely, how many
"time steps" pass per time unit).
We can now derive the maximal load of the system, as follows :
Theorem 5. An upper bound for the maximal applications load λA is :
11
λA ≤ n · pM AX · pN · (1 − E−)
16 (ρ + (α + 1) ln n)
· λ2(cid:101)T
· λM
where:
λ(cid:101)T
(cid:44) min
(cid:40)
(cid:41)
λT ,
2ρ + 2(α + 1) ln n
(1− 1
ln n
pN
)(1−pN − 1
n
)
− 3
2
1 − e
T
Proof. Interpreting λ−1
assign N = λA and T = λ−1
1. λ(cid:101)T denotes the maximal number of steps that may be required by the algorithm to complete the proliferation
as the number of months it takes to process a single step of the algorithm, we can now
M into Theorem 3, requesting that the upper bound over TV ac shall be kept smaller than
of the alerting messages produced in the last monitoring of each month, calculated using the lower bound part of
Theorem 1.[57]
Observation 2. For ρ = O(ln n) and constant values of pM AX and E−, then for pN > ln n
monotonically increasing with n.
n the maximal load is
We can now calculate the algorithm's benefit factor, comparing the system's maximal load to the maximal load of
any non collaborative monitoring algorithm :
Corollary 1. Let us denote by λA the maximal load of a single device. The ratio between the maximal load of
applications a single device can non-collaboratively monitor and the maximal load of the TPP algorithm is :
= λ2(cid:101)T
λA
λA
· (1 − pM AX ) · pM AX ·
n · pN
16 (ρ + (α + 1) ln n)
Proof. We can easily see that :
Dividing λA by λA the rest is implied.
λA = (1 − E−) ·
1
1 − pM AX
· λM
Observation 3. For ρ = O(ln n) and constant values of pM AX and E−, then for pN > ln n
n and regardless of the
monitoring rate λM the benefit factor of participating in the collaborative TPP monitoring algorithm is asymptotically
greater than 1, and is monotonically increasing with n.
Numeric Example I (Large Network ) :
assume a network of n = 1, 000, 000 devices who participate in the
collaborative effort. We shall require that pM AX = 0.01 (namely, that 99% of the network must be fully vaccinated)
and that the reliability of the analysis should be 99.9% (as n = 1, 000, 000 it means that α = 0.5). Disregarding
adversarial attacks, we can use ρ = 1, and in addition, we shall assume that the error probability is very small. We
also assume that the time that is required for a message to be processed and forwarded is approximately 5 seconds
(namely, λT = 720, assuming time units of hours). Upon discovery of a malicious application, we shall assume that a
device sends this information to 400 network members (namely, pN = 1
2500 ). We shall also assume that each device
24·7 ). Using Theorem 5 we see that λA ≤ 35.51163 which
can monitor a single application per week (namely, λM = 1
equals to 25,566 new applications per month![58]
As to the number of messages each participating device is required to send, using Theorem 4 we can see that each
device sends on average of at most 6· ln n messages during each month, namely -- 2.763 messages on average per day.
Comparing this to the performance of a single device relying solely on a non collaborative host-based monitoring,
such a device could cope with monitoring less then 5 new applications each month, as λA ≤ 0.006). This reflects a
benefit factor of more than 5,000 in the maximal load of suspicious applications.
Numeric Example II (Reality Mining ) :
In order to assess the performance of the algorithm for real-world
networks, we have used the data generated by the Reality Mining project [17]. This project analyzed 330,000 hours of
continuous behavioral data logged by the mobile phones of 94 subjects, forming a mobile based social network. Using
this network, we can estimate the efficiency of the TPP algorithm, as follows. Based on the Reality Mining data we
can calculate that n = 94, α = 1.52 (in order to guarantee a 99.9% confidence level) and pN = 0.092123. In addition,
as in the previous example, we assume that each device can monitor a single application per week (λM = 1
24·7 )
and that the time required for a message to be processed equals 5 seconds (λT = 720 and λ(cid:101)T = 139) and were no
adversaries are present (ρ = 1). Then the maximal load of the system λA as a function of the penetration threshold
pM AX equals (using Theorem 5) :
The maximal load of the algorithm for this network appears in Figure 2.
λA ≤ 4.997 · pM AX
12
FIG. 2: Maximal monthly load of the TPP algorithm implemented over the Reality Mining network. Note that the maximal
load increases as the vaccination mechanism loosens (namely, guarantees vaccination for a slightly smaller portion of the
network). For example, where for 98% vaccination the algorithm guarantees monitoring of 72 new applications per month,
when we require that only 90% of the network's devices are vaccinated, the number of applications that can be monitored each
month increases to 360.
It is important to notice that although the number of applications the average user downloads each month is far
smaller than 5,000, the importance of the collaborative monitoring scheme is still high, as the main merit of the
proposed scheme lays in its predictive ability. Namely, whereas a user downloading 5 applications each month may
locally monitor them successfully, should one of them be malicious, the user will fall victim to its malicious behavior.
However, should users that participate in the proposed collaborative monitoring scheme try to download a malicious
application, they will be immediately notified about the maliciousness of the application (in high probability), before
they are exposed to its dangers. Indeed, the first users that encounter a new malicious application may fall victim to
its perils -- an unavoidable property of any host-based monitoring system. However, using the proposed collaborative
scheme this is kept to a minimum.
13
VI. AVOIDING BENIGN APPLICATIONS "FRAME" THROUGH ADVERSARIAL USE OF THE TPP
VACCINATION ALGORITHM
As mentioned in previous sections, the TPP algorithm is fault tolerant to the presence of adversarial devices which
try to compromise the system's integrity by causing a large enough portion of the network to consider (one or more)
benign applications as malicious. This aspect of the algorithm is crucial, as it is a collaborative algorithm which relies
on the participation of as many mobile devices as possible -- devices who should be guaranteed that the efficient
collaborative monitoring will not be accompanied with erroneous results. In the TPP algorithm, this fault tolerance
is being provided by the introduction of the ρ "decision threshold" into the decision system. Namely, the fact that
an application is being marked as malicious only when at least ρ relevant messages (from different origins) are being
received. Relying on the common agreement of several devices as a tool for overcoming system noise (which can be
either coincidental or intentional) is often used in swarm based systems. For example, a similar mechanism called
"threshold cryptography" is used for enabling the collaborative use of cryptographic tasks (see for example [24, 40].
Definition 4. Let us denote by PAttack(T T L, ρ, k
n , ε) the probability that a "framing attack" done by a group of k
organized adversaries will successfully convince at least ε · n of the network's devices that some benign application ai
is malicious.
A trivial example is the use of very large values for T T L, which allow a group of k adversaries to convince the
entire network that any given application is malicious, provided that k > ρ, namely :
∀k ≥ 1
,
∀ρ ≤ k ,
∀ε < 1
,
lim
T T L→∞ PAttack(T T L, ρ,
k
n
, ε) = 1
Theorem 6. The probability that k attackers will be able to make at least an portion of the network's devices treat
some benign application ai as malicious, using the TPP algorithm with a decision threshold ρ is :
(cid:18)
(cid:19)
k
n
PAttack
T T L, ρ,
, ε
≤ 1 − Φ
n ·
√
ε − P(cid:113)
P (1 − P )
(cid:33)ρ
where :
P = e(ρ−T T L·(1−e
− k·pN
2
)) ·
(cid:32)
T T L · (1 − e− k·pN
2
)
ρ
and where Φ(x) is the cumulative normal distribution function, defined as :
Φ(x) =
e− 1
2 t2
dt
(cid:90) x
−∞
1√
2π
(cid:16)
(cid:17)
and also provided that :
ρ > T T L
1 − e− k·pN
2
Proof. We use Lemma 2 to calculate the probability that a device v ∈ V will be reported of some malicious application
ai by a message sent by one of the k adversaries at the next time-step. This is yet again a Bernoulli trial with :
ps = 1 − e− (k·n·pN )
2n = 1 − e− k·pN
2
Denoting as Xv(t) the number of times a notification message had arrived to v after t steps, using Chernoff bound :
P [Xv(t) > (1 + δ)t · ps] <
in which we set δ = ρ
t·ps
− 1. We can therefore see that :
P (cid:44) PAttack(T T L, ρ,
k
n
, n−1) = P [Xv(T T L) > ρ] < e(ρ−T T L·ps) ·
(cid:18) T T L · ps
(cid:19)ρ
ρ
(cid:18)
eδ
(1 + δ)(1+δ)
(cid:19)t·ps
It is important to note that the Chernoff bounds requires that δ > 0. This is immediately translated to the
following requirement :
ρ > T T L(1 − e− k·pN
2
)
14
As we want to bound the probability that at least εn of the devices are deceived, we shall use the above as a success
probability of a second Bernoulli trial. As n is large, the number of deceived devices can be approximated using
normal distribution :
(cid:18)
(cid:19)
k
n
, ε
≤ 1 − Φ
ε · n − n · P
(cid:113)
n · P (1 − P )
PAttack
T T L, ρ,
and the rest is implied.
Theorem 6 is illustrated in Figures 3 and 4. Figure 3 presents the execution of the vaccination algorithm in a
network of 1,000,000 devices, where each device is connected to 50 neighbors. In this example we require that the
number of devices that may be deceived by adversarial attacker would be at most 100. With the decision threshold
ρ properly calibrated according to Theorem 6, we show that as long as the number of adversaries is below 480 they
cannot launch a successful attack. However, as the number of adversaries increases, such an attack quickly becomes
a feasible option. Specifically, by increasing the number of adversaries by 3% (from 480 to 494) the probability of a
successful attack rises from 0 to 1. In order to compensate the increase in adversaries number, devices operating the
TPP algorithm have to simply increase the value of ρ by 1. Figure 4 shows how each small increase in the value of the
decision threshold ρ requires the adversaries to increase their numbers by approximately 7%. Interestingly, the effect
of such changes in the value of ρ on the algorithm's load and overhead is very small, as can be observed in Figure 5.
Note that in the proof of Theorem 6 we assumed that the adversarial devices may decide to send a false message
concerning an application's maliciousness, but they must do so using the definitions of the TPP algorithm. Namely,
each device may send at most pN · n messages, and send these messages to random destinations.
If adversarial
devices had been allowed flexibility in those constraints as well, a small group of adversarial devices could have sent
an infinitely large number of false messages, that would have been propagated throughout the network, resulting in
a successful attack (similarly to the case where T T L → ∞). Alternatively, had adversarial devices been allowed to
chose the destination of the pN · n messages they send, they could have all send them to the same destination v, thus
guaranteeing that v would be deceived. More generically, a group of k = ı · ρ of adversarial devices could have send
their messages to i · pN · n different devices, guaranteeing their deception.
Definition 5. Let us denote by PAttack−Destination
devices are allowed to control the destination of the messages they produce.
n , ε(cid:1) the attack success probability when adversarial
(cid:0)T T L, ρ, k
The following Corollary can be drawn :
Corollary 2. The probability that k attackers that can control the destination of the pN · n messages they produce
will be able to make at least an portion of the network's devices treat some benign application ai as malicious, using
the TPP algorithm with a decision threshold ρ is :
(cid:18)
(cid:19)
(cid:18)
(cid:19)
PAttack−Destination
T T L, ρ,
k
n
, ε
≤ PAttack
T T L − 1, ρ,
k
n
, ε − k
ρ
· pN
VII. FAULT TOLERANCE TO "LEECHING" AND "MUTING ATTACKS"
Collaborative systems, by their nature, are based on the fact that the participants are expected to allocate some
of their resources (computational, energy, storage, etc') to be used by some collaborative algorithm. However, what
happens when users decide to benefit from the system's advantages without providing the contribution that is expected
from them ? This behavior, known as leeching is a known artifact in many Peer-to-Peer data distribution systems,
in which users often utilize the system for data download, without allocating enough upload bandwidth in return. A
similar behavior can also be the result of a deliberate attack on the system -- a muting attack. In this attack, one or
more participants of the system block all the messages that are sent through them (namely -- automatically decrease
the TTL of those messages to zero). In addition, no original messages are being generate by these participants. The
purpose of this attack is to compromise the correctness of the TPP vaccination algorithm, which relies on the paths
messages of a given TTL value are expected to perform.
15
FIG. 3: An illustration of Theorem 6 -- an upper bound for the success probability of a collaborative "framing attack" as
a function of the number of adversarial devices. In this example, a changing number of adversaries are required to deceive
at least 100 devices to think that some benign application is in fact malicious. Notice the phase transition point around 485
adversarial device.
FIG. 4: An illustration of Theorem 6 -- the influence of the value of decision threshold ρ on the number of adversarial devices
required for a successful attack. Note how as the adversaries increase their numbers, all that has to be done in order to prevent
them from launching successful attacks is simply to slightly increase the value of the decision threshold ρ. This makes the TPP
algorithm economically favorable for networks operators. From experimental observations it can be seen that increasing the
value of ρ results only in slight increase in the vaccination time and network overhead, which converge asymptotically, whereas
the resources required in order to launch stronger attacks grow approximately linearly. See more details in Figures 5
Either performed as a way of evading the need to allocate resources for the collective use of the network, or
maliciously as an adversarial abuse of the system, this behavior might have a significant negative influence on the
performance of the TPP algorithm. The operators of the network therefore need to have a way of calibrating the
system so that it will overcome disturbances caused by any given number of participants who select to adopt this
behavior. In this section we show that in terms of completion time the TPP algorithm is fault tolerant to the presence
of blocking devices up to a certain limit.
Definition 6. Let pmute denote the probability that a given arbitrary network device stops generating vaccination
messages or starts blocking some or all of the messages that are sent through it.
16
We now show (Corollary 3) that the expected vaccination time remains unchanged as long as : pmute (cid:28) 1
ρ ln n . For
higher values of pmute Theorem 7 presents an analytic upper bound for the algorithm's expected completion time.
For extremely high values of pmute the completion time of the algorithm is shown in Corollary 4 to be upper bounded
by
1
4 pmute
1−pmute
As to the overall cost of the TPP algorithm, it is shown in Corollary 5 to remain completely unaffected by the
presence of blocking devices, regardless of their number. Due to space considerations, some of the proofs were omitted
and can be found in the Appendix.
V ac (denoting by TV ac the vaccination time with the presence of no blocking devices).
· T 2
Definition 7. Let us denote by T (n, pmute) the vaccination time of a network of n devices, with a probability of pmute
to block messages.
Theorem 7. The vaccination completion time of the TPP algorithm for some critical penetration pM AX in probability
greater than 1 − , while at most n · pmute devices may block messages forwarding and generation, is :
T (n, pmute) =
2(ρ − ln
n )
− 1−pmute−e−timeout·pmute
pmute
1 − e
· n·pM AX ·pN
2T·N·(1−E− )−1
while for the calculation of timeout we can use the expressions that appear in Theorem 2 or Theorem 3.
Proof. See Appendix for complete proof.
timeout. We shall note three complementary cases : rmute (cid:28) 1 or rmute (cid:29) 1 or rmute ≈ 1.
We shall now observe the behavior of the expression above for various values of the ratio rmute defined as pmute ·
It is easy to see that when rmute (cid:28) 1, the decay of the number of messages is negligible, namely :
1 − pmute − e−timeout·pmute
pmute
≈ timeout
As a result, Theorem 7 can be approximated by Theorem 2. Subsequently, the Theorems and Corollaries that are
derived from Theorem 2 would hold, according to which we can see that : pmute (cid:28) 1
O(ρ log n) .
Based on the above, we can now state the fault tolerance of the TPP algorithm, with respect to the muting attack :
(cid:17)
(cid:16) n
ρ ln n
mute mobile devices
Corollary 3. The TPP algorithm is fault tolerant with respect to the presence of c · O
(for some c (cid:28) 1). Namely :
T (n, pmute) ≈ T (n, 0)
Let us now observe the case where rmute (cid:29) 1. In this case, most of the messages are likely to be blocked before
completing their TTL-long path. This results in an increased vaccination time, as shown in the following Corollary :
(for some c (cid:29) 1), the completion
Corollary 4. When the number of blocking devices is greater than c · O
time of the TPP algorithm is affected as follows :
(cid:16) n
(cid:17)
ρ ln n
Proof. See Appendix for complete proof.
T (n, pmute) ≤ 1
4 pmute
1 − pmute
· T (n, 0)2
17
Note that for every other value of rmute which revolves around 1, the expected vaccination time would move
· T (n, 0)2 (see more details concerning the monotonic nature of T (n, pmute) in the proof
1
4 pmute
1−pmute
between T (n, 0) and
of Corollary 5).
Let us now examine the affect blocking devices may have on the number of messages sent throughout the execution
of the algorithm. As shown in the following Corollary, the overall cost of the TPP algorithm remains unaffected by
the presence of any given number of blocking devices.
Definition 8. Let us denote by M (n, pmute) the overall cost of the TPP algorithm (messages sent + monitoring) for
a network of n devices, with a probability of pmute to block messages.
Corollary 5. The overall cost of the TPP algorithm is unaffected by the presence of blocking devices. Namely :
∀pmute < 1
, M (n, pmute) = M (n, 0)
Proof. See Appendix for complete proof.
VIII. VACCINATION IN GENERAL GRAPHS USING NO OVERLAYS
An interesting question is raised when the forwarding of notification messages between the vertices is not assumed
to be done using a random network overlay, but rather -- using only the edges of the network graph. This is also
motivated by works such as [11] and [23] which show how network topology may play an important role in the
spreading of virus / malware (and subsequently, also of the vaccination messages).
In addition, we shall assume that we have no knowledge concerning the structure of the network graph, or its type.
We now show that even in this case, the TPP algorithm can still be used, albeit with a much larger completion time.
Theorem 8. Using the TPP algorithm in an unknown graph with no network overlay, to guarantee a successful
vaccination process for some critical penetration pM AX , the following completion time is obtained :
Proof. We shall first observe the following upper bound concerning the exploration time of a general graph using a
decentralized group of k random walkers [7] :
(cid:32)(cid:18)
(cid:19) 2
TV ac = O
T · N
pM AX · (1 − E−)
3 · n− 2
3 · log
4
3 n
(cid:18)E2log3n
(cid:19)
k2
E(exG) = O
(cid:33)
(cid:33)
Note that the coverage time of random walkers in graphs is also upper bounded by 4
27 n3 + o(n3) [20]. However, as
we assume that pN < O(n−(0.5+)) (for some > 0), using the bound of [7] gives a tighter bound. Using our lower
bound over the number of agents for k, and as E = n · pN we get :
(cid:32)
E(exG) = O
n2p2
timeout2 · n4·p2
N log3 n
M AX·p2
T 2·N 2
N
(1 − E−)2
Multiplying the exploration time by ρ (in order to guarantee that each device will get at least ρ messages) and
replacing E(exG) with timeout we see that :
TV ac = O
(cid:118)(cid:117)(cid:117)(cid:116) ρ ·(cid:16)
3
(cid:17)2
· log(n)
T·N
pM AX·(1−E−)
n2
Assuming also that ρ = O(ln n), and the rest is implied.
The overall cost of the algorithm in graphs with no overlays is given in Theorem 9.
Theorem 9. Using the TPP algorithm in an unknown graph with no network overlay, to guarantee a successful
vaccination for some penetration threshold pM AX , the overall cost of the algorithm is upper bounded by the following
expression :
18
(cid:18)
O
T · N
pM AX · (1 − E−)
3 · n
2
3 · log
8
3 n · CS +
E ·(cid:16)
4
3 · log
n 7
T·N
3 n
pM AX·(1−E−)
(cid:17) 1
3
· CM
(cid:19) 1
(cid:18)
Proof. As to the overall cost of the algorithm, we know that :
M = O
k · timeout · CS +
· CM
k
X
(cid:19)
Using the expressions for timeout, k and pN , the rest is implied.
IX. EXPERIMENTAL RESULTS
We have implemented the TPP algorithm in a simulated environment and tested its performance in various sce-
narios. Due to space considerations, we present in this Section only a sample of the results we have obtained. We
have tested a network of n = 1000 devices, having access to N = 100 applications, one of which was malicious[59].
We assume that each device downloads 30 random applications, monitors 1 application every week, and can send
notification messages to 10 random network members (namely, pN = 0.01). We require that upon completion, at
least 990 network members must be aware of the malicious application (namely, pM AX = 0.01), with = 0.001. In
addition, we assumed that among the network members there are 100 adversaries, trying to deceive at least 50 mobile
devices to believe that some benign application is malicious.
Figure 5 shows the time (in days) and messages required in order to complete this monitoring, as a function of
the decision threshold ρ. We can see that whereas the adversaries succeed in probability 1 for ρ < 3, they fail in
probability 1 for any ρ ≥ 3. Note the extremely efficient performance of the algorithm, with completion time of ∼ 260
days using only 5 messages and at most 30 monitored applications per user. The same scenario would have resulted
in 100 messages per user using the conventional flooding algorithm, or alternatively, in 700 days and 100 monitored
applications per user using a non-collaborative scheme. Notice how the increase in the time and network overhead as
a result of an increase in ρ (in order to compensate an increase in the number of adversaries) converge asymptotically.
This makes the algorithm economically scalable, as discussed in Section VI.
Figure 6 demonstrates the decrease in completion time and network overhead as a result of increasing the pene-
tration threshold pM AX . Figure 7 demonstrates the evolution in the malicious application's penetration probability
throughout the vaccination process. An interesting phenomenon is demonstrated in Figure 8, where the number of
adversarial devices is gradually increased, and their success in deceiving 5% of the network's members is studied. As
the deception rate increases linearly, the success to generate a successful attack displays a phase transition -- growing
rapidly from "a very low attack success probability" to "very high attack success probability" with the increase of
only 20% in the number of adversaries.
As stated in previous sections, the main contribution that stems from implementing the TPP algorithm is the
fact that the vast majority of the mobile devices become immune to any new malicious applications that may be
introduced to the network. We have implemented a network of 70,000 units based on the aggregated calls-graph that
was received from an actual mobile network. Our simulation assumed that each user downloads an average of 50
applications per month, out of which ηn = 5 new applications. To this model we have periodically injected a small
number of malicious applications (approximately 0.5% of the applications). In this simulation we used the well known
SIR epidemic model, with a slight variation -- as the network is dynamic, we assumed that there is a constant flow
of users leaving the network, which is compensated by a flow of users joining the network. This allows the network
to avoid the "everyone dies eventually" syndrome, by constantly trading a portion of the recovered devices for new
susceptible ones. Notice that a device may become infected either as a result of downloading a malicious application
or as a result to an exposure to an infected device. In both cases, a mandatory condition for the infection is that the
device was not vaccination against this application already. Figure 9 presents the dynamics of the network, comparing
the number of infected and damaged devices when the collaborative vaccination mechanism is present to the number
of infected and damaged mobile devices without it. Notice how the proposed algorithm decreases the number of
infectious devices at the steady state of the network, as well as the accumulated number of malicious incidents, by
approximately 75%.
19
FIG. 5: An experimental result of a network of n = 1000 members, with N = 100 applications, pM AX = 0.01, pN = 0.01 and
100 adversaries that try to mislead at least 5% of the network into believing that some benign application is malicious. Notice
how changes in ρ dramatically effect the adversaries' success probability, with almost no effect on the completion time.
FIG. 6: The effect of decreasing the penetration threshold pM AX on the algorithm's completion time and number of messages
(ρ = 1).
X. CONCLUSIONS AND FUTURE WORK
In this work we have presented the TPP TTL-based propagation algorithm, capable of guaranteeing the collab-
orative vaccination of mobile network users against malicious applications. The performance of the algorithm was
analyzed and shown to be superior compared to state of the art in this field, guaranteeing higher number of suspicious
applications that can be monitored concurrently, and do so using a lower network overhead.
The algorithm was also shown to be capable of overcoming the presence of adversarial devices who try to inject
messages of false information into the network. The challenge of filtering out false information which is injected into a
collaborative information propagation networks resembles the known "faulty processors" problem. For example, the
following work discusses the challenge of synchronizing the clock of a communication network of size n when n
3 faulty
processors are present [6]. Another interesting work in this scope is the work of [10] which discusses a collaborative
fault tolerant "acknowledgement propagation" algorithm, for reporting on the receipt (or lack of) of sent messages.
It should be noted that during the analysis of the fault tolerance of the algorithm, we assumed that although an
attacker can send reports of benign applications, or alternatively -- refuse to forward messages passed through it,
the basic parameters of the algorithm are still preserved. Namely, adversarial devices are not allowed to send more
than X messages or generate messages with T T L values higher than the value allowed by the network operator. In
addition, we assumed that while an adversary can "kill" messages sent to it by refusing to forward them, it cannot
however interfere with the content of the message itself (e.g. by changing the identity of the message's original sender,
its current T T L, or the identity of the application it reports of). The exact implementation details of a cryptographic
protocol of these properties, however, is out of the scope of this work,
Future versions of the work should investigate the generalization of the propagation mechanism, allowing the number
of alert messages generated to be dynamically calibrated in order to further decrease the algorithm's overhead (for
example, as a function of the number of alerts already received concerning the application). Alternatively, upon
20
FIG. 7: The penetration probability of the malicious application, as a function of the time, with ρ = 1 (on the left) and ρ = 20
(on the right).
FIG. 8: An illustration of the phase transition that is displayed when observing the influence of the number of adversaries over
their probability to successfully generate an attack. Note how an increase of 20% in the number of adversaries increases the
probability to deceive a large enough portion of the network from less than 0.2 to approximately 0.8.
With No Collaborative Monitoring
With TPP Vaccination Mechanism
FIG. 9: An experimental demonstration of the benefits of using the proposed collaborative monitoring algorithm. Implemented
on a network that is based on real-life mobile network of 70,000 devices, the TPP algorithm significantly decreases both the
number of infectious devices as well as the accumulated number of malicious incidents at the network (i.e. damages caused by
activations of malicious applications that were installed on mobile devices).
forwarding a message, devices might be requested to send more than a single copy of the message they received.
Another interesting topic to investigate is the use of the mechanism proposed in this work as an infrastructure for
other security related problems. One example can be the problem of collaboratively coping with malicious beacons
in hostile wireless environment, as discussed in [33]. Most of the existing localization protocols for sensor networks
are vulnerable in hostile environments, requiring for the enhancement of the security of location discovery. The work
of [33] presents voting-based methods to tolerate malicious attacks against range-based location discovery in sensor
networks. This problem seems to benefit from the use a mechanism which is fault tolerant to the injection of false
information, such as the algorithm we propose in this work.
21
Appendix A: Proof of Theorem 1
THEOREM 1 The time it takes k random walkers to complete a ρ-coverage of G in probability greater than 1−
(denoted as T(n)) can be bounded as follows :
2(cid:0)ρ − ln
(cid:1)
−
n
3k
2n(1− 1
ln n
)
1 − e
≤ T (n) ≤ 2(cid:0)ρ − ln
n
(cid:1)
1 − e− k
2n
Proof. Since our bounds are probabilistic, we can state that the following "bad events" occur with very low probability
(e.g. 2−ω(n)). Event Elow degree, defined as the existence of a vertex v ∈ V with deg(v) < n·pN
. Event Ehigh degree,
defined as the existence of a vertex v ∈ V with deg(v) > 3·n·pN
n · pN
. Using the Chernoff bound on G we get :
] < e− n·pN
prob[deg(v) <
2
2
8
2
Applying union bound on all vertices we get :
Similarly :
P rob[Elow degree] < n · e− n·pN
8 < 2−ω(n)
P rob[Ehigh degree] < 2−ω(n)
From now on we assume that Elow degree and Ehigh degree do not occur, and condition all probabilities over this
assumption. In the private case of ∀v ∈ V , deg(v) = pN · n, every analysis that is based on the expected number of
neighbors shall hold.
In order to continue analyzing the execution process of the TPP algorithm we note that as the initial placement
of the agents is random, their movement is random and the graph G is random, we can see that the placement of the
agents after every-step is purely random over the nodes. Using these observation, the number of agents residing in
adjacent vertices from some vertex v can be produced :
Lemma 1. Let v ∈ V be an arbitrary vertex of G. Let N1(v, t) be the number of agents which reside on one of
N eighbor(v) (adjacent vertices to v) after step t :
∀t ≥ 0 :
3 · pN · k
2
≥ E[N1(v, t)] ≥ pN · k
2
In other words, the expected number of agents who reside in distance 1 from v after every step is at least pN·k
and
2
at most 3·pN·k
.
2
Proof. Upon our assumption, in G(n, pN ) the number of incoming neighbors for some vertex v is at least 1
at most 3
that :
2 pN · n. In addition, for every u ∈ V (G), P rob[some agent resides on u] = k
2 pN · n and
n . Combining the two we see
∀t ≥ 0 : E[N1(v, t)] ≥ pN · n · k
≥ 1
2
pN · k
2n
∀t ≥ 0 : E[N1(v, t)] ≤ 3pN · n · k
2n
≥ 3
2
pN · k
2n and at most 1 − e
22
Lemma 2. For any vertex v ∈ V , the probability of v being notified at the next time-step that ai is malicious is at
least 1 − e− k
Proof. The probability that an agent located on a vertex u such that (u, v) ∈ E will move to v at the next time-step
pN·n . The number of agents that are located in adjacent vertices to v is between k
2 pN . Therefore, the
is
probability that v will not be reported about ai at the next time-step is between (1− 1
2 pN·k.
Using the well known inequality (1 − x) < e−x for x < 1, we can bound this probability from above by :
2 pN and 3k
pN·n ) 1
2 pN·k and (1− 1
pN·n ) 3
2n(1− 1
ln n
−
) .
3k
1
− 1
pN ·n )
(e
1
2 pN·k ≤ e− k
2n
Using the inequality (1 − x) > e− x
1−x for x < 1, it will be bounded from below by :
(cid:16)
e
pN ·n )−1(cid:17) 3
−(pN·n)−1(1− 1
2 pN·k ≥ e
−
3pN k
2pN (n− 1
pN
−
) ≥ e
3k
2n(1− 1
ln n
)
In order to guarantee that the graph is connected we know that PN ≥ ln n
n [19]. Therefore, the probability that v will
−
2n and at most 1 − e
be notified on the next time-step is at least 1 − e− k
2n(1− 1
ln n
) .
3k
Interestingly, this fact holds for any positive pN (the density parameter of G).
Lemma 2 states the probability that some vertex v ∈ V will be reported of ai at the next time-step. This is in fact
a Bernoulli trial with probability psuccess. Now we bound the probability of failing this trial (not notifying vertex v
enough times) after m steps. Let Xv(m) denote the number of times that a notification message had arrived to v
after m steps, and let Fv(m) denote the event that v was not notified enough times after m steps (i.e. Xv(m) < ρ).
We additionally denote by F (m) the event that one of the vertices of G where not notified enough times after m steps
(i.e. (cid:83)
v∈V (G) Fv(m)). We use the Chernoff bound :
P [Xv(m) < (1 − δ)psuccessm] < e−δ2 mpsuccess
2
in which we set δ = 1 −
ρ
mpsuccess
. We can then see that :
P [Xv(m) < ρ] < e
−(1−
ρ
mpsuccess
)2 mpsuccess
2
namely : P [Fv(m)] < eρ− mpsuccess
2
. Applying the union bound we get :
P [e1 ∪ e2 ∪ . . . ∪ en] ≤ P [e1] + P [e2] + . . . + P [en]
on all n vertices of G. Therefore we can bound the probability of failure on any vertex v (using Lemma 2) as follows :
P r[F (m)] ≤ neρ− mpsuccess
2
≤ neρ− m(1−e
− k
2
2n )
≤
Assigning psuccess = 1 − e− k
2n (worse case) and psuccess = 1 − e
−
3k
2n(1− 1
ln n
) (best case), the rest is implied.
Appendix B: Leeching and Muting Attacks
In this section we give the complete proofs for the Theorems and Corollaries that appear in Section VII of the main
body of the paper.
Detailed Analysis -- Completion Time. Recall we denote by pmute the probability that a given network device
may decide to stop generating vaccination messages and block some or all of the messages that are received by it. In
addition, recall that T (n, pmute) denotes the vaccination time of a network of n devices, with a probability of pmute
to block messages. Theorem 2 can now be revised in the following way :
probability greater than 1 − , while at most n · pmute devices may block messages forwarding and generation, is :
THEOREM 7 : The vaccination completion time of the TPP algorithm for some critical penetration pM AX in
T (n, pmute) =
2(ρ − ln
n )
− 1−pmute−e−timeout·pmute
pmute
1 − e
· n·pM AX ·pN
2T·N·(1−E− )−1
while for the calculation of timeout we can use the expressions that appear in Theorem 2 or Theorem 3.
Proof. The expected number of new messages created at time t, k(t) is expected to be :
k(t) = (1 − pmute)
n2 · pai · pN
T · N
(1 − E−)
Given a message created with a TTL of timeout, at every time step it has a probability of pmute to be sent to a
network node which will not forward it onwards. Therefore, given a group of m messages, created at time t. Then at
time t + i (for every i < timeout), only (1− pmute)i · m messages would remain alive. The average number of messages
at any time step between t and t + timeout is therefore :
23
As we assume that ∀t < TV accination pai ≥ pM AX , the number of agents k would be at least :
1 − (1 − pmute)timeout
timeout · pmute
m
1 − pmute − (1 − pmute)timeout+1
pmute
· n2 · pM AX · pN
T · N
(1 − E−)
Using again the fact that ∀x < 1 (1 − x) < e−x, we can see that :
k >
1 − pmute − e−timeout·pmute
pmute
· n2 · pM AX · pN
T · N
(1 − E−)
Recalling Theorem 1 :
T (n, 0) =
2(ρ − ln
n )
1 − e− k
2n
and assigning the revised expression for k, the rest is implied.
COROLLARY 4 : When the number of blocking devices is greater than c · O
completion time of the TPP algorithm is affected as follows :
(cid:17)
(cid:16) n
ρ ln n
(for some c (cid:29) 1), the
T (n, pmute) ≤ 1
Proof. When rmute (cid:29) 1 Theorem 7 converges as follows :
Observation 4. When pmute (cid:29) timeout−1, the vaccination time of the TPP algorithm is :
4 pmute
1 − pmute
· T (n, 0)2
Assuming again that = n−α
we see that :
2(ρ − ln
n )
2T ·N
· n·pM AX ·pN
(1−E−)
T (n, pmute) =
1 − e
− 1−pmute
pmute
s.t. α ∈ Z+ then provided that :
1 − pmute
pmute
· n · pM AX · pN
2T · N
(1 − E−) < 1
T (n, pmute) =
4N · T · pmute(ρ + (α + 1) ln n)
(1 − pmute) · n · pM AX · pN (1 − E−)
Using Theorem 3, we shall calculate ∆pmute , denoting the increase in the vaccination time as a result of the presence
of the blocking nodes :
∆pmute =
T (n, pmute)
T (n, 0)
=
4N·T·pmute(ρ+(α+1) ln n)
(cid:113) T·N (ρ+(α+1) ln n)
(1−pmute)·n·pM AX·pN (1−E−)
n·pM AX·pN·(1−E−)
4
and after some arithmetics we can see that :
∆pmute =
T (n, pmute)
T (n, 0)
=
1
4 pmute
1 − pmute
· T (n, 0)
24
Detailed Analysis -- Cost. Recall we denote by M (n, pmute) the overall cost of the TPP algorithm (messages sent
+ monitoring) for a network of n devices, with a probability of pmute to block messages. As shown in the following
Corollary, the overall cost of the TPP algorithm remains unaffected by the presence of any given number of blocking
devices.
COROLLARY 5 : The overall cost of the TPP algorithm is unaffected by the presence of blocking devices.
Namely :
Proof. We have already shown before (Corollary 3) that when the number of blocking devices is smaller than c·O( n
(for some c (cid:28) 1), the system is approximately unaffected by the blocking devices. Therefore, we can expect that :
ρ ln n )
∀pmute < 1
, M (n, pmute) = M (n, 0)
M (n, pmute) ≈ M (n, 0)
Interestingly, this is also the case when the number of blocking devices is far greater. Recalling Observation 1 the
cost of the algorithm with no blocking devices equals :
(cid:18)
(cid:19)
M (n, 0) = O
k · T (n, 0) +
k
n · pN
· CM
Denoting by k(n,pmute) the expected number of active agents at each time step, we have already shown in the proof
of Theorem 7 that for large values of pmute :
The overall cost of the algorithm assuming large number of blocking devices equals :
k(n,pmute) ≈ 1 − pmute
pmute
· T (n, 0)−1 · k(n,0)
(cid:19)
k(n,pmute)
n · pN
CM
(cid:18)
k(n,pmute) · T (n, pmute) +
(cid:32)
M (n, pmute) = O
Assigning the values of k(n,pmute) and T (n, pmute) we can now see that :
· k(n,0)
1−pmute
pmute
T (n, 0) · n · pN
k(n,0) · T (n, 0) +
M (n, pmute) = O
As we assumed that pmute (cid:29) 1
ρ ln n we know that :
1 − pmute
pmute
(cid:28) 1 − 1
ρ ln n
1
ρ ln n
(cid:28) ρ ln n − 1
(cid:33)
CM
(cid:19)
CM
The expression for the overall cost of the algorithm can now be rewritten as :
ρ ln n · k(n,0)
T (n, 0) · n · pN
k(n,0) · T (n, 0) +
M (n, pmute) = O
and as T (n, 0) = Ω(ln n), we get the requested result of :
M (n, pmute) = O
k(n,0) · T (n, 0) +
(cid:19)
CM
= M (n, 0)
k(n,0)
n · pN
(cid:18)
(cid:18)
We have shown that M (n, pmute) = M (n, 0) for pmute (cid:28) timeout−1 and well as for pmute (cid:29) timeout−1. We shall
now upper bound the overall cost of the algorithm for values of pmute which are close to timeout−1. For this, we shall
find the value of pmute which maximizes M (n, pmute) :
∂M (n, p)
∂p
=
∂k(n,p)
∂p
T (n, p) +
∂T (n, p)
∂p
k(n,p) +
∂k(n,p)
∂p
CM
n · pN
Recalling (from the proof of Theorem 7) that :
25
we see that :
where:
Therefore :
k >
1 − pmute − e−timeout·pmute
pmute
· n2 · pM AX · pN
T · N
(1 − E−)
∂k(n,p)
∂p
=
(p · timeout + 1) · e−timeout·p − 1
p2
· α
n2 · pM AX · pN
T · N
α =
(1 − E−)
∀p > 0
,
∂k(n,p)
∂p
< 0
and the number of agents is monotonously decreasing.
Examining the behavior of the cleaning time, we can see that :
∂T (n, p)
∂p
= − ∂k(n,p)
∂p
·
2e
n(1 − e
− 1−p−e−timeout·p
·β
− 1−p−e−timeout·p
·β)2
p
p
where :
β =
n · pM AX · pN
2T · N · (1 − E−)−1
Using our observation concerning ∂k(n,p)
∂p
, we see that :
∀p > 0
,
∂T (n, p)
∂p
> 0
and the cleaning time is monotonously increasing.
Returning to the derivative of the overall price function, let us divide it to two components. The first component
representing the number of messages sent during the algorithm while the second representing the monitoring activities
of the devices :
where :
and :
∂M (n, p)
∂p
= M1(p) + M2(p)
M1(p) =
∂k(n,p)
∂p
T (n, p) +
∂T (n, p)
∂p
k(n,p)
M2(p) =
∂k(n,p)
∂p
1
n · pN
· CM
As k(n,p) is monotonously decreasing, the cost of the algorithm due to monitoring activities, represented by M2(p),
would be maximized when pmute = 0 (for which we have already shown that M (n, p) = M (n, 0)).
T (n, p) −
e−β·xp
(1 − e−β·xp )2 · xp · β
(cid:19)
As to M1(p), it can be written as :
M1(p) =
∂k(n,p)
∂p
(cid:18)
where :
β =
n · pM AX · pN
2T · N · (1 − E−)−1
26
and :
xp =
1 − p − e−timeout·p
p
Studying this function reveals that it has a single minimum point, while M1(0) = 0 and M1(1) → ∞. This means
that the maximal values of M (n, pmute) are received either at pmute = 0 or at pmute = 1. As we have already shown
that at these points the overall cost of the algorithm remains unchanged, we can conclude that the overall cost of the
algorithm at any point between these values of pmute is also bounded by M (n, 0).
Appendix C: λ Optimization
Recall that while analyzing the performance of the TPP algorithm we artificially divided the vaccination process
into two phases. At the first phase, the vaccinating agents are generated, while the proliferation of the agents is done
in the second phase. The ratio between the two phases was assumed to be 1, namely -- we assumed that the two
phases are identical in length. Following is the analysis of the selection of this ratio, which demonstrates that this is
indeed the optimal division.
Let us recall that :
and that :
timeout = λ · (TGeneration + timeout)
TGeneration =
1 − λ
λ
· timeout
As we assumed in the previous section that λ = 0.5, it is interesting to know whether another value of λ (perhaps,
a value that depends of the properties of the network) may yield a lower completion time.
Concerning TV ac, it can now be seen that :
TV ac ≤ 1
λ
· timeout
Similar to the calculation of the number of agents k in Theorem2, we know that :
(cid:88)
k =
i≤TGeneration
n2 · pai · pN eighbor
T · N
(1 − E−) ≥
By selecting timeout = m we verify that the information propagation process will be completed successfully, and
so we can now write :
timeout =
2(ρ − ln(
1 − e− k
2n
≤
≥ 1 − λ
λ
(1 − E−)
T · N
· timeout · n2 · pM AX · pN eighbor
2(cid:0)ρ − ln(
n )(cid:1)
n )(cid:1)
1 − e− 1−λ
λ ·timeout· n·pM AX ·pN eighbor
2T ·N
(1−E−)
We can therefore write :
Using a similar approximation as in the previous section, we shall assume that the requested is polynomial in 1
n .
2(ρ + (α + 1) ln(n)(cid:1) = timeout(1 − e− 1−λ
λ ·timeout· n·pM AX ·pN eighbor
2T ·N
(1−E−))
Using the bound (1 − x) < e−x for x < 1 we can see that when assuming that :
1 − λ
λ
· timeout · n · pM AX · pN eighbor
2T · N
(1 − E−) < 1
we can write the previous expression as :
2ρ + (2α + 2) ln(n) ≥ 1 − λ
λ
· timeout2 · n · pM AX · pN eighbor
2T · N
(1 − E−)
27
4T · N (ρ + (α + 1) ln(n))
n · pM AX · pN eighbor · (1 − E−) 1−λ
λ
and therefore :
which means that :
(cid:115)
timeout ≤
(cid:115)
TV ac ≤ 1
λ
4T · N (ρ + (α + 1) ln(n))
n · pM AX · pN eighbor · (1 − E−) 1−λ
λ
(cid:115)
In order to find the optimal value of λ we now calculate ∂TV ac
:
∂λ
∂TV ac
∂λ
= − 1 − 2λ
2(λ − λ2)1.5
It can now be easily seen that :
4T · N (ρ + (α + 1) ln(n))
n · pM AX · pN eighbor · (1 − E−)
and therefore the value of λ = 0.5 minimizes the upper bound over the algorithm's completion time.
λ =
1
2
−→ ∂TV ac
∂λ
= 0 ,
∂2TV ac
∂λ2 > 0
[1] Adamic, L. A.; Lukose, R. M.; Puniyani, A. R.; and Huberman, B. A. 2001. Search in power-law networks. Phys. Rev. E
64(4):046135.
[2] Altshuler, Y.; Yanovsky, V.; Bruckstein, A.; and Wagner, I. 2008. Efficient cooperative search of smart targets using uav
swarms. ROBOTICA 26:551 -- 557.
[3] Altshuler, Y.; Wagner, I.; and Bruckstein, A. 2009. Collaborative exploration in grid domains. In Sixth International
Conference on Informatics in Control, Automation and Robotics (ICINCO).
[4] Apap, F.; Honig, A.; Hershkop, S.; Eskin, E.; and Stolfo, S. 2002. Recent Advances in Intrusion Detection. Springer Berlin
Heidelberg. chapter Detecting Malicious Software by Monitoring Anomalous Windows Registry Accesses, 36 -- 53.
[5] Bailey, N. 1975. The Mathematical Theory of Infectious Diseases and its Applications (second edition). Hafner Press.
[6] Barak, B.; Halevi, S.; Herzberg, A.; and Naor, D. 2000. Clock synchronization with faults and recoveries (extended
abstract). In PODC '00: Proceedings of the nineteenth annual ACM symposium on Principles of distributed computing,
133 -- 142. New York, NY, USA: ACM.
[7] Broder, A.; Karlin, A.; Raghavan, P.; and Upfal, E. 1989. Trading space for time in undirected s− t connectivity. In ACM
Symposium on Theory of Computing (STOC), 543 -- 549.
[8] Cagalj, M.; Hubaux, J.; and Enz, C. 2002. Minimum-energy broadcast in all-wireless networks: Np-completness and
distribution issues. In MOBICOM.
[9] Canalys. 2009. Canalys estimates, Canalys.
[10] Castelluccia, C.; Jarecki, S.; Kim, J.; and Tsudik, G. 2006. Secure acknowledgement aggregation and multisignatures with
limited robustness. Computer Networks 50(10):1639 -- 1652.
[11] Chakrabarti, D.; Wang, Y.; Wang, C.; Leskovec, J.; and Faloutsos, C. 2008. Epidemic thresholds in real networks. ACM
Trans. Inf. Syst. Secur. 10(4):1 -- 26.
[12] Chung, F., and Lu, L. 2001. The diameter of sparse random graphs. Advances in Applied Mathematics 26:257 -- 279.
[13] Crisostomo, S.; Barros, J.; and Bettstetter, C. 2008. Flooding the network: Multipoint relays versus network coding. In
4th IEEE Intl. Conference on Circuits and Systems for Communications (ICCSC), 119 -- 124.
[14] Demers, A.; Greene, D.; Hauser, C.; Irish, W.; Larson, J.; Shenker, S.; Sturgis, H.; Swinehart, D.; and Terry, D. 1987.
Epidemic algorithms for replicated database maintenance. In In Proc. of the Sixth ACM Symp. on Principles of Distributed
Computing, 1 -- 12.
[15] Dolev, S., and Tzachar, N. 2010. Spanders: Distributed spanning expanders. In Proc. of the 25th ACM Symposium on
Applied Computing (SAC-SCS).
[16] Dolev, S.; Schiller, E.; and Welch, J. 2006. Random walk for self-stabilizing group communication in ad hoc networks.
IEEE Transactions on Mobile Computing 5:893 -- 905.
[17] Eagle, N.; Pentland, A.; and Lazer, D. 2009. Inferring social network structure using mobile phone data. Proceedings of
the National Academy of Sciences (PNAS) 106:15274 -- 15278.
[18] Erdos, P., and Renyi, A. 1959. On random graphs. Publ. Math. Debrecen 6:290 -- 291.
[19] Erdos, P., and Renyi, A. 1960. On the evolution of random graphs. Publications of the Mathematical Institute of the
Hungarian Academy of Sciences 5:17 -- 61.
28
[20] Feige, U. 1995. A tight upper bound on the cover time for random walks on graphs. Random Struct. Algorithms 6(1):51 -- 54.
[21] Fragouli, C.; Widmer, J.; and Boudec, J. L. 2006. A network coding approach to energy efficient broadcasting:
from
theory to practice. In The 25th IEEE International Conference on Computer Communications (INFOCOM2006), 1 -- 11.
[22] Ganesa, D.; Krishnamachari, B.; Woo, A.; Culler, D.; Estrin, D.; and Wicker, S. 2002. An empirical study of epidemic
algorithms in large scale multihop wireless networks -- technical report ucla/csd-tr 02-0013. Technical report, UCLA
Computer Science.
[23] Garetto, M.; Gong, W.; and Towsley, D. 2003. Modeling malware spreading dynamics.
In In Proceedings of IEEE
INFOCOM, 1869 -- 1879.
[24] Gemmel, P. 1997. An introduction to threshold cryptography. CryptoBytes 7 -- 12.
[25] GetJar. 2010. Getjar statistics.
[26] Hyponnen, M. 2006. Malware goes mobile. Sci. American 295:70.
[27] Jonasson, J., and Schramm, O. 2000. On the cover time of planar graphs. Electron. Comm. Probab. 5:85 -- 90 (electronic).
[28] Kim, H.; Smith, J.; and Shin, K. G. 2008. Detecting energy-greedy anomalies and mobile malware variants. In MobiSys
'08: Proceeding of the 6th international conference on Mobile systems, applications, and services, 239 -- 252. New York, NY,
USA: ACM.
[29] Kleinberg, J. 2007. The wireless epidemic. Nature 449:287 -- 288.
[30] Koenig, S.; Szymanski, B.; and Liu, Y. 2001. Efficient and inefficient ant coverage methods. Annals of Mathematics and
Artificial Intelligence 31:41 -- 76.
[31] Kong, C.; Peng, N.; and Rekleitis, I. 2006. Distributed coverage with multi-robot system. In IEEE International Conference
on Robotics and Automation.
[32] Korf, R. 1990. Real-time heuristic search. Artificial Intelligence 42:189 -- 211.
[33] Liu, D.; Ning, P.; Liu, A.; Wang, C.; and Du, W. K. 2008. Attack-resistant location estimation in wireless sensor networks.
ACM Trans. Inf. Syst. Secur. 11(4):1 -- 39.
[34] Lv, Q.; Cao, P.; Cohen, E.; Li, K.; and Shenker, S. 2002. Search and replication in unstructured peer-to-peer networks.
In ICS '02: Proceedings of the 16th international conference on Supercomputing, 84 -- 95. New York, NY, USA: ACM.
[35] 2008. Mcafee mobile security report 2008. Technical report.
[36] 2009. Mcafee mobile security report 2009. Technical report.
[37] Moskovitch, R.; Gus, I.; Pluderman, S.; Stopel, D.; Glezer, C.; Shahar, Y.; and Elovici, Y. 2007a. Detection of un-
known computer worms activity based on computer behavior using data mining. In CISDA 2007. IEEE Symposium on
Computational Intelligence in Security and Defense Applications, 169 -- 177.
[38] Moskovitch, R.; Pluderman, S.; Gus, I.; Stopel, D.; Feher, C.; Parmet, Y.; Shahar, Y.; and Elovici, Y. 2007b. Host based
intrusion detection using machine learning. In 2007 IEEE Intelligence and Security Informatics, 107 -- 114.
[39] Mutz, D.; Valeur, F.; Vigna, G.; and Kruegel, C. 2006. Anomalous system call detection. ACM Trans. Inf. Syst. Secur.
9(1):61 -- 93.
[40] Narasimha, M.; Tsudik, G.; and Yi, J. H. 2003. On the utility of distributed cryptography in p2p and manets: the case
of membership control. In Proceedings of the 11th IEEE International Conference on Network Protocols, 336 -- 345.
[41] Ni, S.; Tseng, Y.; Chen, Y.; and Sheu, J. 1999. The broadcast storm problem in a mobile ad hoc network. In In Proceedings
of the ACM/IEEE International Conference on Mobile Computing and Networking (MOBICOM), 151 -- 162.
[42] Qayyum, L., and Laouiti, A. 2002. Multipoint relaying for flooding broadcast messages in mobile wireless networks. In
Proceedings of HICSS.
[43] Sasson, Y.; Cavin, D.; and Schiper, A. 2003. Probabilistic broadcas for flooding in wireless mobile ad-hoc networks. In
Proceedings of IEEE Wireless communication and networks (WCNC).
[44] Stojmenovic, I.; Seddigh, M.; and Zunic, J. 2002. Dominating sets and neighbor elimination-based broadcasting algorithms
in wireless networks. IEEE Transactions on Parallel and Distributed Systems 13(1):14 -- 25.
[45] Svennebring, J., and Koenig, S. 2004. Building terrain-covering ant robots: A feasibility study. Autonomous Robots
16(3):313 -- 332.
[46] van Renesse, R., and Birman, K. 2002. Scalable management and data mining using astrolabe. In In Proc. of the First
International Workshop on Peer-to-Peer Systems (IPTPS02).
[47] Vogels, W.; van Renesse, R.; and Birman, K. 2003. The power of epidemics: robust communication for large-scale
distributed systems. SIGCOMM Comput. Commun. Rev. 33(1):131 -- 135.
[48] Wagner, I.; Altshuler, Y.; Yanovski, V.; and Bruckstein, A. 2008. Cooperative cleaners: A study in ant robotics. The
International Journal of Robotics Research (IJRR) 27(1):127 -- 151.
[49] Wang, P.; Gonzalez, M.; Hidalgo, C.; and Barabasi, A. 2009. Understanding the spreading patterns of mobile phone
viruses. Science 324:1071 -- 1075.
[50] Williams, B., and Camp, T. 2002. Comparison of broadcasting techniques for mobile ad hoc networks. In MOBIHOC,
9 -- 11.
[51] Zhou, C. V.; Leckie, C.; and Karunasekera, S. 2010. A survey of coordinated attacks and collaborative intrusion detection.
Computers and Security 29:124 -- 140.
[52] We assume that interference in the messages' content, or generation of messages using false identity are impossible, due
to, say, the use of cryptographic means.
[53] An alternative that does not require such an overlay assumption is presented in Section VIII.
[54] This will later come into effect when messages will be sent between the network's members, at which case the selection
of "an arbitrary network member" can be assumed to be purely random. Notice that assuming pure randomness among
devices selection is not mandatory -- see Section VIII for analysis of the TPP algorithm assuming no such network overlays.
29
[55] We aspire that the number of messages each device is asked to send upon discovering a new malicious application is
kept to a minimum. As the value of PN must be greater than ln n
n in order to guarantee connectivity [19], it is safe to
assume PN = O( ln n
n ) [15]. We lower
bound pN as we are interested in demonstrating the result for any random graph G(n, p). In addition, we later show that
timeout ≈ O(log n). It is also safe to assume that N ≈ Ω(ln n) and that PM AX ≈ O(cid:0) 1
n ). Note that a connected pseudo-random graph can also be generated with pN = O( 1
(cid:1). This assumption in later
ln n
discussed in great details.
[56] See Section VI for more details.
[57] The number of nodes to be covered that is assigned to the expression of Theorem 1 is n − n · pN − 1 as the n · pN alerting
messages are known to be sent to different nodes, at the first step of the algorithm. In addition, the node that generates
these messages is naturally aware of their content.
[58] Notice that by using days (or weeks) instead of hours, the maximal monthly load of the algorithm could have further been
increased. However, this would have also increased the time between the initial discovery of a malicious application and
time at which the vaccination process is completed. Selecting a short time unit (i.e. an hour) ensures that this time is kept
to a minimum.
[59] Note that the number of malicious applications does not influence the completion time of algorithm, as monitoring and
notification is done in parallel. The number of message, however, grows linearly with the number of malicious applications.
|
0904.2320 | 1 | 0904 | 2009-04-15T13:49:42 | Why Global Performance is a Poor Metric for Verifying Convergence of Multi-agent Learning | [
"cs.MA",
"cs.LG"
] | Experimental verification has been the method of choice for verifying the stability of a multi-agent reinforcement learning (MARL) algorithm as the number of agents grows and theoretical analysis becomes prohibitively complex. For cooperative agents, where the ultimate goal is to optimize some global metric, the stability is usually verified by observing the evolution of the global performance metric over time. If the global metric improves and eventually stabilizes, it is considered a reasonable verification of the system's stability.
The main contribution of this note is establishing the need for better experimental frameworks and measures to assess the stability of large-scale adaptive cooperative systems. We show an experimental case study where the stability of the global performance metric can be rather deceiving, hiding an underlying instability in the system that later leads to a significant drop in performance. We then propose an alternative metric that relies on agents' local policies and show, experimentally, that our proposed metric is more effective (than the traditional global performance metric) in exposing the instability of MARL algorithms. | cs.MA | cs |
Why Global Performance is a Poor Metric for Verifying
Convergence of Multi-agent Learning
Sherief Abdallah∗,a,b
aBritish University in Dubai, United Aran Emirates
bUniversity of Edinburgh, United Kingdom
Abstract
Experimental verification has been the method of choice for verifying the
stability of a multi-agent reinforcement learning (MARL) algorithm as the
number of agents grows and theoretical analysis becomes prohibitively com-
plex. For cooperative agents, where the ultimate goal is to optimize some
global metric, the stability is usually verified by observing the evolution of
the global performance metric over time. If the global metric improves and
eventually stabilizes, it is considered a reasonable verification of the system's
stability.
The main contribution of this note is establishing the need for better
experimental frameworks and measures to assess the stability of large-scale
adaptive cooperative systems. We show an experimental case study where
the stability of the global performance metric can be rather deceiving, hiding
an underlying instability in the system that later leads to a significant drop
in performance. We then propose an alternative metric that relies on agents'
local policies and show, experimentally, that our proposed metric is more
effective (than the traditional global performance metric) in exposing the
instability of MARL algorithms.
Key words: Multi-agent Systems, Multi-agent Learning, Experimental
Verification
∗Email: [email protected]
Preprint submitted to ...
September 25, 2018
1. Introduction
The term convergence, in reinforcement learning context, refers to the sta-
bility of the learning process (and the underlying model) over time. Similar
to single agent reinforcement learning algorithms (such as Q-learning (1)),
the convergence of a multi-agent reinforcement learning (MARL) algorithm is
an important property that received considerable attention (2; 3; 4; 5). How-
ever, proving the convergence of a MARL algorithm via theoretical analysis
is significantly more challenging than proving the convergence in the single
agent case. The presence of other agents that are also learning deem the en-
vironment non-stationary, therefore violating a foundational assumption in
single agent learning. In fact, proving the convergence of MARL algorithm
even in 2-player-2-action single-stage games (arguably the simplest class of
multi-agent systems domains) has been challenging (2; 4; 5).
As a consequence, experimental verification is usually the method of
choice as the number of agents grows and theoretical analysis becomes pro-
hibitively complex. For cooperative agents, researchers typically verified the
stability of a MARL algorithm by observing the evolution of some global
performance metric overtime (6; 7; 8; 9; 5). This is not surprising since the
ultimate goal of a cooperative system is to optimize some global metric. Ex-
amples of global performance metrics include the percentage of total number
of delivered packets in routing problems (10), the average turn around time
of tasks in task allocation problems (5), or the average reward (received by
agents) in general (8).
If the global metric improves over time and eventually appears to stabilize,
it is usually considered a reasonable verification of convergence (6; 7; 8; 9; 5).
Even if the underlying agent policies are not stable, one could argue that at
the end, global performance is all that matters in a cooperative system.
This paper challenges the above (widely-used) practice and establishes
the need for better experimental frameworks and measures for assessing the
stability of large-scale cooperative systems. We show an experimental case
study where the stability of the global performance metric can hide an un-
derlying instability in the system. This hidden instability later leads to a
significant drop in the global performance metric itself. We propose an al-
ternative measure that relies on agents' local policies: the policy entropy.
We experimentally show that the proposed metric is more effective than the
traditional global performance metric in exposing the instability of MARL
algorithms in large-scale multi-agent systems.
2
The paper is organized as follows. Section 2 describes the case study we
will be using throughout the paper. Section 3 reviews MARL algorithms
(with particular focus on WPL and GIGA-WoLF, the two algorithms we use
in our experimental evaluation). Section 4 presents our initial experimental
results, where the global performance metric leads to a (misleading) con-
clusion that a MARL algorithm converges. Section 5 presents our proposed
measure and illustrates how it is used to expose the hidden instability of a
MARL algorithm. We conclude in Section 6.
2. Case Study: Distributed Task Allocation Problem (DTAP)
We use a simplified version of the distributed task allocation domain
(DTAP) (5), where the goal of the system is to assign tasks to agents such
that the service time of each task is minimized. For illustration, consider the
example scenario depicted in Figure 1. Agent A0 receives task T1, which
can be executed by any of the agents A0, A1, A2, A3, and A4. All agents
other than agent A4 are overloaded, and therefore the best option for agent
A0 is to forward task T1 to agent A2 which in turn forwards the task to its
left neighbor (A5) until task T1 reaches agent A4. Although agent A0 does
not know that A4 is under-loaded (because agent A0 interacts only with its
immediate neighbors), agent A0 should eventually learn (through experience
and interaction with its neighbors) that sending task T1 to agent A2 is the
best action without even knowing that agent A4 exists.
A1
A3
A0
T1
A5
A2
A4
Figure 1: Task allocation using a network of agents.
The DTAP domain has an essential property that appears in many real
world problems yet not captured by most of the domains that were used to
3
analyze MARL algorithms experimentally: communication delay. The effect
of an action does not appear immediately because it is communicated via
messages and messages take time to route. Not only is the reward delayed
but so is any change in the system's state. A consequence of communication
delay is partial observability: an agent can not observe the full system state
(the queues at every other agent, messages on links and in queues, etc.).
Each time unit, agents make decisions regarding all task requests received
during this time unit. For each task, the agent can either execute the task
locally or send the task to a neighboring agent. If an agent decides to execute
the task locally, the agent adds the task to its local queue, where tasks are
executed on a first come first serve basis, with unlimited queue length.
Each agent has a physical location. Communication delay between two
agents is proportional to the Euclidean distance between them, one time unit
per distance unit. Agents interact via two types of messages. A REQUEST
message hi, j, T i indicates a request sent from agent i to agent j requesting the
execution of task T . An UPDATE message hi, j, T, Ri indicates a feedback
(reward signal) from agent i to agent j that task T took R time steps to
complete (the time steps are computed from the time agent i received T 's
request).
T ST (T )
T τ
PT ∈T τ
The main goal of DTAP is to reduce the total service time, averaged over
tasks, AT ST =
, where T τ is the set of task requests received
during a time period τ and T ST (T ) is the total time a task T spends in the
system. The T ST (T ) time consists of the time for routing a task request
through the network, the time the task request spends in the local queue,
and the time of actually executing the task.
Although the underlying simulator has different underlying states, we
deliberately made agents oblivious to these states. The only feedback an
agent gets is its own reward. This simplifies the agent's decision problem and
re-emphasises partial observability: agents collectively learn a joint policy
that makes a good compromise over the different unobserved states (because
the agents can not distinguish between these states).
4
3. Multiagent Reinforcement Learning
The experimental results in the following section focus on two gradient-
ascent MARL algorithms: GIGA-WoLF (3) and WPL (5).1 We chose these
two algorithms because they allow agents to learn a stochastic policy based on
the expected reward gradient. Both algorithms were also shown to converge
in benchmark two-player-two-action games as well as some larger games.
The specifics of WPL and GIGA-WoLF (such as their update equations, the
underlying intuition, their differences and similarities) are neither relevant
to the purpose of this paper nor needed to follow our analysis in Section
4. Nevertheless, and for completeness, we mention below (very briefly) the
equations for updating the policy for the two algorithms. Further details
regarding the two algorithms can be found elsewhere (3; 5).
An agent i using WPL updates its policy πi according to the following
equations:
∀j ∈ neighbors(i) : ∆πt+1
i
(j) ←
∂V t
∂πt
i (π)
i (j)
if ∂V t
∂πt
i (π)
i (j) < 0
i (j) otherwise
· η ·( πt
i (j)
1 − πt
i + ∆πt+1
)
i
πt+1
i ← projection(πt
where η is a small learning constant and Vi(πi) is the expected reward
agent i would get if it interacts with its neighbors according to policy πi.
The projection function ensures that after adding the gradient ∆πi to the
policy, the resulting policy is still valid.
An agent i using GIGA-WoLF updates its policy πi according to the
following equations:
1A large number of MARL algorithms have been proposed that vary in their in their
underlying assumptions and target domains (11). MARL algorithms that can only learn a
deterministic policy (such as Q-learning (1)) are not suitable for the DTAP domain. For
example, even if two neighbors have practically the same load, Q-learning will assign all
incoming requests to one of the neighbors until a feedback is received later indicating a
change in the load. On the other hand, an agent using a gradient ascent MARL algorithm
has the ability to adjust its policy to a non-deterministic (or stochastic) distribution (5).
Q-learning was successfully used in the packet routing domain (6; 12), where load balancing
is not the main concern (the main objective is routing a packet from a particular source
to a particular destination).
5
i (π)t)
i (π)/3)
πt+1
i + ηV t
i = projection(πt
zt+1
i + ηV t
i = projection(πt
zt+1
i − zt
i
zt+1
i − πt
(zt+1
i = min(cid:18)1,
δt+1
πt+1
i = πt+1
i + δt+1
i (cid:19)
i − πt+1
i
)
i
4. Stability Under the Global Metric Results
We have evaluated the performance of WPL and GIGA WoLF using the
following setting.2 100 agents are organized in a 10x10 grid. Communication
delay between two adjacent agents is two time units. Tasks arrive at the 4x4
sub-grid at the center at rate 0.5 tasks/time unit. All agents can execute a
task with a rate of 0.1 task/time unit (both task arrival and service durations
follow an exponential distribution). Figure 2 illustrates the setting.
Figure 3 plots the global performance (measured in terms of ATST) of the
two multi-agent learning algorithms in the DTAP domain. Just by looking at
ATST plot, it is relatively safe to conclude that WPL converges quickly while
GIGA-WoLF converges after about 75,000 time steps. The following section
presents the measure we have used in order to discover that the stability of
GIGA-WoLF is actually spurious.
5. Verifying Stability Using Policy Entropy
Ideally, we would want to visualize and analyze the evolution of all learn-
ing parameters, including action values and the policy of all agents. However,
going to such detail is only possible for small number of agents. As the num-
ber of agents increases, one needs aggregated measures that summarize the
system's behavior yet can reflect the stability of the system's dynamics.
We propose using a simple measure that summarizes an agent's policy
into a single number: the policy entropy, H(πx), for a particular agent x:
H(πx) = − Xy∈neighbors(x)
πx(y)lgπx(y)
2The simulator is available online at http://www.cs.umass.edu/~shario/dtap.html.
6
a9
a8
a7
a6
a5
a4
a3
a2
a1
a0
a19
a29
a39
a49
a59
a69
a79
a89
a99
a18
a28
a38
a48
a58
a68
a78
a88
a98
a17
a27
a37
a47
a57
a67
a77
a87
a97
a16
a26
a36
a46
a56
a66
a76
a86
a96
a15
a25
a35
a45
a55
a65
a75
a85
a95
a14
a24
a34
a44
a54
a64
a74
a84
a94
a13
a23
a33
a43
a53
a63
a73
a83
a93
a12
a22
a32
a42
a52
a62
a72
a82
a92
a11
a21
a31
a41
a51
a61
a71
a81
a91
a10
a20
a30
a40
a50
a60
a70
a80
a90
Figure 2: The simulation setting for the DTAP domain. Only the 16 nodes at the
center receive tasks from the environment. A node's diameter reflects its local
queue length.
1800
1600
1400
1200
1000
800
600
400
200
0
WPL
GIGA−WoLF
0
50000
100000
150000
200000
Figure 3: Comparing the average total service time for 200,000 time steps of the
DTAP problem for WPL and GIGA-WoLF.
7
The function πx is the policy of agent x (πx(y) is the probability that
agent x interacts with its neighbor y). In case of gradient ascent, the pol-
icy is explicitely learned. For deterministic learners (such as Q-learning),
the effective policy can be estimated by counting the number of times each
neighbor is chosen.
Figure 4 plots the average policy entropy and the associated standard
deviation, over the 100 agents, against time. Agent policies under WPL do
converge but the policies under GIGA-WoLF have not converged yet. The
policy entropy is still decreasing, which suggests that GIGA-WoLF is still
adapting.
GIGA-WoLF
WPL
y
p
o
r
t
n
E
y
c
i
l
o
P
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
0
0 0
1
0
0
0 0
3
0
0
0 0
5
0
0
0 0
7
0
0
0 0
9
0
0
0
0
0
0
0
0
0
7 0
1
5 0
1
3 0
1
1 0
1
0
0
0
9 0
1
0
0
0
Time
Figure 4: The policy entropy of WPL and GIGA-WoLF for 200,000 time steps.
This intrigued us to rerun the simulator, this time allowing the simulator
to run for 600,000 times steps instead of just 200,000 time steps. To our
surprise, the global performance metric (the ATST in this case) of GIGA-
WoLF starts slowly to diverge after 250,000 time steps and the corresponding
policy entropy continue to decrease. WPL's policy entropy remains stable,
as well as the global performance metric.
More in-depth analysis is needed in order to fully understand the dy-
namics of GIGA-WoLF and WPL in the DTAP domain and in large-scale
systems in general. However, this is beyond the scope of this research note
8
3500
3000
2500
2000
1500
1000
500
0
WPL
GIGA−WoLF
0
100000
200000
300000
400000
500000
600000
Figure 5: Comparing the average total service time for 600,000 time steps of the
DTAP problem for WPL and GIGA-WoLF.
y
p
o
r
t
n
E
y
c
i
l
o
P
GIGA-WoLF
WPL
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0 0
1
0
0
0 0
6
0
0
1
0 0
1
0
1 0
0 0
2
0
6 0
0 0
2
0 0
3
0
6 0
0
1 0
0 0
3
0 0
4
0
6 0
0 0
4
0
1 0
0 0
5
0
6 0
0 0
5
6 0
0
1 0
0
1 0
0 0
0
Time
Figure 6: The average policy entropy of WPL and GIGA-WoLF for 600,000 time
steps.
9
and should not distract from the main point we are trying to make: the
common practice of using a global performance metric to verify the stability
of a MARL algorithm is not reliable and can be misleading.
6. Conclusion and Future work
The main contribution of this paper is showing that using a global perfor-
mance metric for verifying the stability (or even the usefulness) of a MARL
algorithm is not a reliable methodology. In particular, we present a case study
of 100 agents where the global performance metric can hide an underlying
instability in the system that later leads to a significant drop in performance.
We propose a measure that successfully exposes such instability.
One of the issues indirectly raised by this paper is for how long shall a
performance metric be stable in order to conclude the stability of the un-
derlying MARL algorithm? Currently, no theoretical framework addresses
this question, which we believe to be an essential requirement for adopting
MARL in practical large-scale applications.
References
[1] R. Sutton, A. Barto, Reinforcement Learning: An Introduction, MIT
Press, 1999.
[2] M. Bowling, M. Veloso, Multiagent learning using a variable learning
rate, Artificial Intelligence 136 (2) (2002) 215 -- 250.
[3] M. Bowling, Convergence and no-regret in multiagent learning, in: Pro-
ceedings of the Annual Conference on Advances in Neural Information
Processing Systems, 2005, pp. 209 -- 216.
[4] V. Conitzer, T. Sandholm, AWESOME: A general multiagent learning
algorithm that converges in self-play and learns a best response against
stationary opponents, Machine Learning 67 (1-2) (2007) 23 -- 43.
[5] S. Abdallah, V. Lesser, A multiagent reinforcement learning algorithm
with non-linear dynamics, Journal of Artificial Intelligence Research 33
(2008) 521 -- 549.
10
[6] J. A. Boyan, M. L. Littman, Packet routing in dynamically changing
networks: A reinforcement learning approach, in: Proceedings of the
Annual Conference on Advances in Neural Information Processing Sys-
tems, 1994, pp. 671 -- 678.
[7] L. Peshkin, V. Savova, Reinforcement learning for adaptive routing, in:
Proceedings of the International Joint Conference on Neural Networks,
2002, pp. 1825 -- 1830.
[8] M. Ghavamzadeh, S. Mahadevan, R. Makar, Hierarchical multi-agent
reinforcement learning, Autonomous Agents and Multi-Agent Systems
13 (2) (2006) 197 -- 229.
[9] S. Abdallah, V. Lesser, Multiagent reinforcement learning and self-
organization in a network of agents, in: Proceedings of the International
Joint Conference on Autonomous Agents and Multiagent Systems, 2007,
pp. 1 -- 8.
[10] Y.-H. Chang, T. Ho, Mobilized ad-hoc networks: A reinforcement learn-
ing approach, in: ICAC '04: Proceedings of the First International Con-
ference on Autonomic Computing, IEEE Computer Society, Washing-
ton, DC, USA, 2004, pp. 240 -- 247.
[11] L. Panait, S. Luke, Cooperative multi-agent learning: The state of the
art, Autonomous Agents and Multi-Agent Systems 11 (3) (2005) 387 --
434.
[12] P. S. Dutta, N. R. Jennings, L. Moreau, Cooperative information shar-
ing to improve distributed learning in multi-agent systems, Journal of
Artificial Intelligence Research 24 (2005) 407 -- 463.
11
|
1101.0820 | 1 | 1101 | 2010-12-27T15:18:31 | Emotionally Colorful Reflexive Games | [
"cs.MA"
] | This study addresses the matter of reflexive control of the emotional states by means of Reflexive Game Theory (RGT). It is shown how to build a bridge between RGT and emotions. For this purpose the Pleasure-Arousal-Dominance (PAD) model is adopted. The major advantages of RGT are its ability to predict human behavior and unfold the entire spectra of reflexion in the human mind. On the other hand, PAD provides ultimate approach to model emotions. It is illustrated that emotions are reflexive processes and, consequently, RGT fused with PAD model is natural solution to model emotional interactions between people. The fusion of RGT and PAD, called Emotional Reflexive Games (ERG), inherits the key features of both components. Using ERG, we show how reflexive control can be successfully applied to model human emotional states. Up to date, EGR is a unique methodology capable of modeling human reflexive processes and emotional aspects simultaneously. | cs.MA | cs |
Emotionally Colorful Reflexive Games
Sergey Tarasenko
Deparment of Intelligence Science and Technology, Graduate School of Informatics
Kyoto University, Yoshida honmachi, Kyoto 606-8501, Japan
[email protected]
Abstract. This study addresses the matter of reflexive control of the
emotional states by means of Reflexive Game Theory (RGT). It is shown
how to build a bridge between RGT and emotions. For this purpose the
Pleasure-Arousal-Dominance (PAD) model is adopted. The major ad-
vantages of RGT are its ability to predict human behavior and unfold
the entire spectra of reflexion in the human mind. On the other hand,
PAD provides ultimate approach to model emotions. It is illustrated that
emotions are reflexive processes and, consequently, RGT fused with PAD
model is natural solution to model emotional interactions between peo-
ple. The fusion of RGT and PAD, called Emotional Reflexive Games
(ERG), inherits the key features of both components. Using ERG, we
show how reflexive control can be successfully applied to model human
emotional states. Up to date, EGR is a unique methodology capable of
modeling human reflexive processes and emotional aspects simultane-
ously.
Key words: Reflexive Game Theory, Bipolar Choices, Theory of Bipo-
larity and Reflexivity, the Golden Ratio, Implicit Primordial Knowledge
(IPK), Inner Computer, Algebraic Model of Ethical Cognition
1 Introduction
The theory of personal constructs proposed by Kelly [10] in 1955 initiated the
strong debates. This theory suggests that each person has one's own unique sys-
tem of dichotomous (bipolar) constructs, which serve as special axes for "pro-
jecting" self and other persons. Most of the constructs may be mapped onto the
scale of "good-bad". Kelly [10] developed the experimental approach for essem-
beling the list of personal constructs and a list of one's images of other people.
The poles, the constructs are represented with, are of negative ("bad") and pos-
itive ("good") types. Hereafter, the theory proposed by Kelly will be referred
as Theory of Bipolar Constructs. According to Kelly, these poles should be cho-
sen with equal probability. However the subsequent psychological experiments
proved this suggestion wrong.
In earlier experiments, it was shown by Benjafield and Adams-Webber[5],
Adams-Webber[1], Benjafield and Green [6], Shalit [28], Osgood and Richards
[25] and Osgood [24] that under great variety of experimental conditions the
positive pole is chosen with the frequency 0.62.
2
Sergey Tarasenko
The persistent appearance of the frequency 0.62 resulted in two ways. First,
the reasonable question "What is the true value of the probability this frequency
represents?". Second, it was suggested by Lefebvre [14](p. 291) that
"... human cognition has a special mechanism for modeling self and
others. It works as a universal 'inner computer' and creates the core
of the images that later are 'dressed' and 'colored' with nuances. The
constant 0.62 is a characteritic of this 'computer'."
It was determined by Lefebvre that the true value of the frequency 0.62 is
probability 0.618..., which is inverted Golden Ratio φ (φ = 1.618...) [14, 13].
Therefore, the Golden Ratio, as being a characteristic of the 'inner computer',
is related to the Implicit Primordial Knowledge (IPK) [31].
Now we return to the Theory of Bipolar Constructs. It had strong impact on
subsequently conducted psychological researches. The research about the emo-
tions was not an exception.
The semantic differential approach originally proposed by Osgood et al. [26]
considers three dimensions to characterize the person's personality. These di-
mensions are Evalution, Activity and Potency.
This approach was further tested from the point of view of emotions. Russel
and Mehrabian [30] proved in their study that the entire spectra of emotions
can be described by the 3-dimensional space spanned by Pleasure (Evaluation),
Arousal (Activity) and Domination (Potency) axes. For every dimension, the
lower and upper bounds (ends) are recognized as negative and positive poles,
respectively. Consequently, the negative pole can be described by negative ad-
jectives, and positive one by positive adjectives.
It was shown by Russel and Mehrabian [30] that these three dimensions are
not only necessary dimensions for an adequate description of emotions, but they
are also sufficient to define all the various emotional states. In other words,
the Emotional state can be considered as a function of P leasure, Arousal and
Dominance. On the basis of this study, Mehrabian [22] proposed Pleasure-
Arousal-Dominance (PAD) model of Emotional Scales. The emotional states
defined as combinations of ends from various dimensions are presented in Fig. 1.
In this study, we discuss the matter of how the PAD model can be used
in Reflexive Game Theory (RGT) [16, 17] to emotionally color the interactions
between people and humans and robots.
2 Building the Bridge between PAD Model and RGT
2.1 Representation of PAD model in the space of binary 3D vectors
The most important issue to fuse RGT inference and PAD Emotional Scale is
to build the bridge between two approaches.
By definition, PAD model is spanned by three dimensions. The value of
each component continuously ranges from -1 to 1. The notation in the PAD
model space presented in Fig. 1 are as follows: 1) pair +P vs −P corresponds
Emotionally Colorful Reflexive Games
3
to P leasure (positive pole: value 1) vs Displeasure (negative pole: value -1); 2)
pair +A vs −A corresponds to Arousal (positive pole: value 1) vs N on-arousal
(negative pole: value -1); and 3) pair +D vs −D corresponds to Dominance
(positive pole: value 1) vs Submissiveness (negative pole: value -1).
Accodring to Mehrabian [22], "pleasure vs displeasure" distinguishes the pos-
itive vs negative emotional states, "arousal vs non-arousal" refers to combination
of physical activity and mental alertness, and "dominance vs submissiveness" is
defined in terms of control vs lack of control [22].
i.e.,
Though, PAD model operates with continous values,
for example,
curious is coded as (0.22, 0.62,−0.01). Mehrabian [22] defines 8 basic states,
which are all possible combinations of high vs low pleasure (+P vs −P ), high
vs low arousal (+A vs −A), and high vs low dominance (+D vs −D). In other
words, there 8 basic states are all possiblecombinations of the poles. In total,
there are 6 poles - two for each of three scales. Therefore, there are 23 = 8
possible combinations of poles. These states (combinations) are considered as
extreme states and can be refered as approximations of the intermediate states.
For instance, intermediate state "angry" (-0.51, 0.59, 0.25) can be approximated
by extreme state "hostile" (−P, +A, +D). Therefore, only eight extreme states
can be used to describe the entire variety of intermediate emotional states.
Fig. 1. The Pleasure-Arousal-Dominance (PAD) model's space.
The combinations of poles in the PAD space can be presented as 3-dimensional
binary vectors, where 1 corresponds to positive pole (1 in PAD notation) and 0
refers to negative pole (-1 in PAD notation). Therefore, emotional states, which
are combinations of different poles of various dimensions, can be represented as
follows: for instance, emotional state Hostile (-P,+A,+D) or (-1,1,1) can be sub-
EmotionallyColorfulReflexiveGames3toPleasure(positivepole:value1)vsDispleasure(negativepole:value-1);2)pair+Avs−AcorrespondstoArousal(positivepole:value1)vsNon-arousal(negativepole:value-1);and3)pair+Dvs−DcorrespondstoDominance(positivepole:value1)vsSubmissiveness(negativepole:value-1).AccodringtoMehrabian[22],"pleasurevsdispleasure"distinguishesthepos-itivevsnegativeemotionalstates,"arousalvsnon-arousal"referstocombinationofphysicalactivityandmentalalertness,and"dominancevssubmissiveness"isdefinedintermsofcontrolvslackofcontrol[22].Though,PADmodeloperateswithcontinousvalues,i.e.,forexample,curiousiscodedas(0.22,0.62,−0.01).Mehrabian[22]defines8basicstates,whichareallpossiblecombinationsofhighvslowpleasure(+Pvs−P),highvslowarousal(+Avs−A),andhighvslowdominance(+Dvs−D).Inotherwords,there8basicstatesareallpossiblecombinationsofthepoles.Intotal,thereare6poles-twoforeachofthreescales.Therefore,thereare23=8possiblecombinationsofpoles.Thesestates(combinations)areconsideredasextremestatesandcanbereferedasapproximationsoftheintermediatestates.Forinstance,intermediatestate"angry"(-0.51,0.59,0.25)canbeapproximatedbyextremestate"hostile"(−P,+A,+D).Therefore,onlyeightextremestatescanbeusedtodescribetheentirevarietyofintermediateemotionalstates.Bored(−P,−A,−D)Anxious(−P,+A,−D)Disdainful(−P,−A,+D)Hostile(−P,+A,+D)Docile(+P,−A,−D)Dependent(+P,+A,−D)Relaxed(+P,−A,+D)Exuberant(+P,+A,+D)P(Pleasure)D(Dominance)A(Arousal)Fig.1.PADmodelThecombinationsofpolesinthePADspacecanbepresentedas3-dimensionalbinaryvectors,where1correspondstopositivepole(1inPADnotation)and0referstonegativepole(-1inPADnotation).Therefore,emotionalstates,whicharecombinationsofdifferentpolesofvariousdimensions,canberepresentedas 4
Sergey Tarasenko
Table 1. Disjunction operation
x y x ∪ y
0 0
0
1 0
0 1
1 1
1
1
1
Table 2. Conjunction operation
x y x ∩ y
0 0
0
1 0
0 1
1 1
0
0
1
The complete set of emotional states represented as binary vectors is
stituted for {0, 1, 1}, while Relaxed (+P,-A,+D) or (1,-1,1) can be represented
as {1, 0, 1} in the rescaled coordinates.
Docile (+P,−A,−D) is coded as {1, 0, 0};
Anxious (−P, +A,−D) is coded as {0, 1, 0};
Disdainf ul (−P,−A, +D) is coded as {0, 0, 1};
Hostile (−P, +A, +D) is coded as {0, 1, 1};
Dependent (+P, +A,−D) is coded as {1,1,0};
Relaxed (+P,−A, +D) is coded as {1, 0, 1};
Exuberant (+P, +A, +D) is coded as {1, 1, 1};
Bored (−P,−A,−D) is coded as {0, 0, 0}.
Among eight basic states, there are three special emotional states Docile
(+P,-A,-D) ≡ {1, 0, 0}), Anxious (−P, +A,−D) ≡ {0, 1, 0}) and Disdainf ul
(−P,−A, +D) ≡ {0, 0, 1}). These three states are the basis of the 3d binary
space. Thus, any other five emotional states can be considered as disjunction
(notations OR/∪/+) of these three basis vectors (emotional states). The binary
disjunction operation is presented in Table 1.
For example, emotional state Dependent (+P, +A,−D) is disjunction of
basis states Docile (+P,−A,−D) and Anxious (−P, +A,−D): Docile OR
Anxious = (+P,−A,−D) ∪ (−P, +A,−D) = {1, 0, 0} ∪ {0, 1, 0} = {1, 1, 0}
= (+P, +A,−D) = Dependent.
Emotionally Colorful Reflexive Games
5
2.2 Formalism of the Reflexive Game Theory
In this section, we present the basics of the Reflexive Game Theory (RGT). In the
RGT, it is considered that behavior of each subject in the group is determined
by reflexive function Φ, which is function of subjects' mutual influences [16, 17,
18, 13]:
(1)
where n is a total number of subjects in the group; aj, for any j = 1, ..., n; j (cid:54)= i,
influence of subject aj on subject ai.
Φ = Φ(a1, ..., ai, ..., an),
Then the choice of subject ai is defined by decision equation [16, 17]:
ai = Φ(a1, ..., ai, ..., an),
(2)
The RGT uses Boolean algebra of alternatives for calculus. There three op-
erations defined in Boolean algebra. These are unary negation operation and
binary conjunction and disjunction operations. The disjuction and conjunction
operations are defined in Tables 1 and 2, respectively.
In this study, we employ Boolean algebra containing eight 3d binary vectors:
{1, 0, 0}, {0, 1, 0}, {0, 0, 1}, {1, 1, 0}, {1, 0, 1}, {0, 1, 1}, {0, 0, 0}, and {1, 1, 1}. The
element of Boolean algebra code particular influences. Each element of Boolean
algebra is referred as alternative.
The RGT deals with groups of individuals. Any group is represented with
relationship graph. Any group, whose graph is decomposable1, can be represented
as polynomial P .
The relationships in the group are indicated with ribs. The solid-line ribs
represent alliance relationship. The dashed-line ribs correspond to conflict rela-
tionship. In algebraic form, conjunction or multiplication operations represent
alliance, while disjunction or summation operations correspond to conflict rela-
tionship. The idea behind is as follows. Two subjects being in alliance can find
the compromise or a common influence, therefore their interaction can be char-
acterized as conjunction of their influences. At the same time, two conflicting
subjects will never reach a compromise and their interaction should include both
influences. Therefore disjunction operation corresponds to the conflict relation-
ship.
Considerring the disjunction and conjunction operations, the group, in which
subjects a and b are in conflict with each other and in alliance with subject c, is
represented as (a + b)c polynomial. Relationship graph is presented in Fig. 2.
Fig. 2. Relationship graph for three subject a, b and c.
1 See [16, 17, 18] for detailed discussion about graph decomposition.
6SergeyTarasenkoacbFig.2.Relationshipgraph[(a+b)c][a+b][a][b][c]·+Fig.3.Polynomialstratificationtreeforpolynomialab+c.ThereflexivefunctionΦisdefinedbythecorrespondingdiagonalform.There-fore,reflexivefunctionforsubjecta,bandcis[a]+[b][a+b]·[c]Φ(a,b,c)=[(a+b)c].Therefore,therootofthediagonalformispolynomialP.WedefinetheremainingpartofthediagonalformasW.Therefore,thediagonalformcanbepresentedasPW.AccrodingtothedefinitiongivenbyLefebvre[13],theexponentialoperationPWisdefinedbyformulaPW=P+W,(3)whereWstandsfornegationofW[16,17,18].Thisoperationisusedtofoldthediagonalform.Therefore,ingeneralcasereflexivefunctionΦisdefinedasΦ(a1,...,ai,...,an)=P+W,(4)Nextwefoldthediagonalformforthreesubjectsa,bandc:([a+b]+[a]+[b])[c][(a+b)c]=1[c][(a+b)c]=(a+b)c+c.ConsiderreflexivefunctionΦforthesubjecta,bandcdiscussedabove:Φ(a,b,c)=(a+b)c+c,(5) 6
Sergey Tarasenko
Then on the basis of polynomial P , the polynomial stratification tree is con-
structed (Fig. 3).
Fig. 3. Polynomial stratification tree for polynomial (a + b)c.
By omiting the branches of the polynomial stratification tree, the diagonal
form [16, 17, 13, 15, 32] is obtained straight forwardly. Consider the diagonal
form corresponding to the PST presented in Fig. 3:
[a] + [b]
[a + b]
·[c]
.
[(a + b)c]
The reflexive function Φ is defined by the corresponding diagonal form. Thus,
reflexive function for subjects a, b and c is
Φ(a, b, c) = [(a + b)c]
[a] + [b]
[a + b]
·[c]
.
Therefore, the root of the diagonal form is polynomial P . We define the
remaining part of the diagonal form as W . Thus, the diagonal form can be
presented as P W .
Accroding to the definition given by Lefebvre [13], the exponential operation
P W is defined by formula
P W = P + W ,
(3)
where W stands for negation of W [16, 17, 18].
This operation is used to fold the diagonal form. Therefore, in general case
reflexive function Φ is defined as
Φ(a1, ..., ai, ..., an) = P + W ,
(4)
Next we fold the diagonal form for three subjects a, b and c:
([a + b] + [a] + [b])[c]
[(a + b)c]
=
1[c]
[(a + b)c]
= (a + b)c + c .
6SergeyTarasenkoacbFig.2.Relationshipgraph[(a+b)c][a+b][a][b][c]·+Fig.3.Polynomialstratificationtreeforpolynomialab+c.ThereflexivefunctionΦisdefinedbythecorrespondingdiagonalform.There-fore,reflexivefunctionforsubjecta,bandcis[a]+[b][a+b]·[c]Φ(a,b,c)=[(a+b)c].Therefore,therootofthediagonalformispolynomialP.WedefinetheremainingpartofthediagonalformasW.Therefore,thediagonalformcanbepresentedasPW.AccrodingtothedefinitiongivenbyLefebvre[13],theexponentialoperationPWisdefinedbyformulaPW=P+W,(3)whereWstandsfornegationofW[16,17,18].Thisoperationisusedtofoldthediagonalform.Therefore,ingeneralcasereflexivefunctionΦisdefinedasΦ(a1,...,ai,...,an)=P+W,(4)Nextwefoldthediagonalformforthreesubjectsa,bandc:([a+b]+[a]+[b])[c][(a+b)c]=1[c][(a+b)c]=(a+b)c+c.ConsiderreflexivefunctionΦforthesubjecta,bandcdiscussedabove:Φ(a,b,c)=(a+b)c+c,(5) Emotionally Colorful Reflexive Games
7
The reflexive function Φ for the subjects a, b and c discussed above is then:
Φ(a, b, c) = (a + b)c + c ,
The general form of decision equation for each subject is then
x = (a + b)c + c ,
(5)
(6)
where x is any subject variable a, b or c.
Finally, we briefly present the solution method of the decision equation. Con-
sider the decision equaition in a form:
x = Ax + Bx ,
(7)
where A and B are some sets.
This form of the decision equation is called canonical form [16, 17, 32, 33]
of the decision equation. If sets A and B are such that A ⊇ B, then decision
equation in canonical form has at least one solution from the interval A ⊇ x ⊇ B.
2.3 Merging the RGT and PAD model
Summarizing the facts about the RGT and PAD model, we highlight that RGT
has been proven to predict human choices in the groups of people and allows to
control human behavior by means of particular influences on the target individ-
uals. Next we note that PAD model provides description of how the emotional
states of humans can be modelled, meaning that a certain emotional state of a
particular person can be changed to the desired one. Furthermore, it is straight-
forward to see that the coding of the PAD emotional states and alternatives of
Boolean algebra are identical.
Therefore, it is possible to change the emotional states of the subjects in the
groups by making influences as elements of the Boolean algebra. In such a case,
vector {1, 0, 0}, for example, plays a role of influence towards emotional state
Docile.
Besides, we have distinguished three basis emotional states Docile ({1, 0, 0}),
Anxious ({0, 1, 0}) and Disdainf ul ({0, 0, 1}). The interactions (as defined by
disjunction and conjunction operations) of these basic emotional states can result
in derivative emotional states such as Dependent, Relaxed, etc. Before, consid-
ering the example of PAD application in RGT, we note that reflexive function
Φ defines state, which subject is going to switch to. This process goes uncon-
sciously. We have discussed above the reasons the conjunction and disjunction
represent alliance and conflict relatioships, respectively.
Another important issue is that often people directly express their emotions
in actions. Therefore in such a case, a particular emotional state of a subject can
be considered as the influence he is making on other subjects.
Example 1. Subjects a and b are in alliance relationship. Subject a makes in-
fluence Dependent {1, 1, 0}. Subject b makes influence Relaxed {1, 0, 1}. Their
8
Sergey Tarasenko
resultant influence will be (a · b) = {1, 1, 0}{1, 0, 1} = {1, 0, 0} or Docile. Con-
sequently, the influence of the group, including subjects in alliance with each
other, on a given subject is considered as conjunction (defining compromise of
all the subjects in alliance) of the influences of all the subjects' influences.
Example 2. Subjects a and b are in conflict relationship. Subject a makes
influence Docile {1, 0, 0}. Subject b makes influence Disdainf ul {0, 0, 1}. Their
resultant influence will be (a + b) = {1, 0, 0}{0, 0, 1} = {1, 0, 1} or Relaxed.
Therefore, the influence of the group, including subjects in conflict with each
other, on a given subject is considered as disjunction (defining overall influence
since compromise is impossible) of the influences of all the subjects' influences.
Next, we consider an example of reflexive interactions controlling emotional
states.
3 Emotionally Colored Reflexive Games: Sample Situation
Consider a group of four subjects - the director d and his advisors a, b and c. Let
advisors a,b and c are in alliance with each other and in conflict with director
d. The graph of such group is presented in Fig. 4. This groups is described by
polynomial abc + d.
Fig. 4. Relationship graph of four subjects a, b, c and d.
The canonical form of decision equation for each subject are:
a = (bc + d)a + da
b = (ac + d)b + db
c = (ab + d)c + dc
d = d + abcd
The corresponding solution intervals are
(bc + d) ⊇ a ⊇ d
(ac + d) ⊇ b ⊇ d
(ab + d) ⊇ c ⊇ d
1 ⊇ d ⊇ abc
(8)
(9)
(10)
(11)
(12)
(13)
(14)
(15)
8SergeyTarasenkoExample2.Subjectsaandbareinconflictrelationship.SubjectamakesinfluenceDocile{1,0,0}.SubjectbmakesinfluenceDisdainful{0,0,1}.Theirresultantinfluencewillbe(a+b)={1,0,0}{0,0,1}={1,0,1}orRelaxed.Therefore,theinfluenceofthegroup,includingsubjectsinconflictwitheachother,onagivensubjectisconsideredasdisjunction(definingoverallinfluencesincecompromiseisimpossible)oftheinfluencesofallthesubjects'influences.Next,weconsideranexampleofreflexiveinteractionscontrollingemotionalstates.3EmotionallyColoredReflexiveGames:SampleSituationabcdFig.4.Relationshipgraphoffoursubjectsa,b,candd.Consideragroupoffoursubjects-thedirectordandhisadvisorsa,bandc.Letadvisorsa,bandcareinalliancewitheachotherandinconflictwithdirectord.ThegraphofsuchgroupispresentedinFig.4.Thisgroupsisdescribedbypolynomialabc+d.Thecanonicalformofdecisionequationforeachsubjectare:a=(bc+d)a+da(8)b=(ac+d)b+db(9)c=(ab+d)c+dc(10)d=d+abcd(11)Thecorrespondingsolutionintervalsare(bc+d)⊇a⊇d(12)(ac+d)⊇b⊇d(13)(ab+d)⊇c⊇d(14)1⊇d⊇abc(15)Considerthateachsubjectisinaparticularuniqueemotionalstate.LetdirectorwillbeinExuberantemotionalstate.Theadvisorsa,bandcareinRelaxed({1,0,1}),Docile({1,0,0})andAnxious({0,1,0})emotionalstates,respectevely.Thisvarietyinemotionalstatesrefrainsdirectorandhisadvisors Emotionally Colorful Reflexive Games
9
It is assumed that each subject is in a particular unique emotional state. Let
director will be in Exuberant emotional state. The advisors a, b and c are in
Relaxed ({1, 0, 1}), Docile ({1, 0, 0}) and Anxious ({0, 1, 0}) emotional states,
respectevely. This variety in emotional states refrains director and his advisors
from reaching a fruitful decision. Understanding this emotional situation, direc-
tor decides to apply reflexive control on emotional level. Let adivors' influence
on all the other subjects coincides with their emotional state, while directly is in
complete control and can decide, which emotional influence to on each particular
subject.
Using RGT, we can predict the emotional states of each subject in the group
after the reflexive emotional interaction:
for subject a: ({1, 0, 0}{0, 1, 0} + d) ⊇ a ⊇ d ⇒ a = d;
for subject b: ({1, 0, 1}{0, 1, 0} + d) ⊇ b ⊇ d ⇒ a = d;
for subject c: ({1, 0, 1}{1, 0, 0} + d) ⊇ c ⊇⇒ {1, 0, 0} + d ⊇ c ⊇ d;
for subject d: 1 ⊇ d ⊇ {1, 0, 1}{1, 0, 0}{0, 1, 0} ⇒ 1 ⊇ d ⊇ 0 ⇒ d = d.
of advisors a and b is completely defined by the director's influence.
Therefore, under conditions of such group structure and influences, decision
Fig. 5. Interpretation of the Diagonal form levels.
The entire diagonal form (P + W ) represents the state of the subject. In
Fig. 5, the diagonal form is marked as Subject. The term W is called a general
image of the self. On the next level, there are two images of Image 1 ([abc][a][b][c])
and Images 2 ([d]). The Images 1 and 2 are images of the self, which general
image of the self W has. Finally, there are the images [a], [b] and [c] are the
images of the self, which Image 1 has.
Following this interpretation of the diagonal form [16, 17], we can calculate
each emotional state in each image. We analyze the structure of reflexion for
advisor c. His state is ({1, 0, 0} + d) or (Docile(+P,−A,−D) plus director's
influence). The emotional state in the Image 1 is Exuberant (1 = {1, 1, 1}),
because [abc][abc] = [abc] + [abc] = 1. The Image 2 is [d] that means it is en-
tirely defined by director's influence. Finally, the general images of the self W is
Exuberant (1 = {1, 1, 1}).
[abc+d][d][abc]+[a][b][c][abc][a][b][c][a][b][c][d]SubjectFirst Image of Self (Image 1)Second Image of Self (Image 2)Images of Self , which Image 1 hasGeneral Image of Self (W)+ 10
Sergey Tarasenko
In this simple example, we have illustated how the emotional states can
appear to be the subject for the reflexive control. We call such reflexive control
to be reflexive emotional control. We have shown how the reflexive emotional
control can be sucessfully implemented by means of the Reflexive Game Theory.
Besides, the RGT allows to unfold the entire sequence of reflexion in the
human mind including its emotional aspects.
4 Discussion
In the previous sections, we have introduced the bridge between the PAD model
of emotional states and RGT calculus by coding PAD's emotoinal states in terms
of RGT's Boolean algebra of alternatives. We also illustrated how the emotional
states can be the subject of reflexive control and can be successfully managed.
This is possible, because the variables in the PAD, namely, Pleasure, Arousal,
Dominance, are proved to be necessary and sufficient variables to control human
emotions [30]. In other words, necessity means that if emotional state is changed
the value of all three variables is changed accordingly. On the other hand, if the
values of the variables are changed, the emotional state is also changed. We have
illustrated how to merge the RGT calculus with proper coding of PAD emotional
states. Finally, we have provided a simple explanatory example of how reflexive
emotional control can be applied in action.
Furthermore, we have not only illustated how to apply RGT to the control
of subject's emotions, but uncovered the entire cascade of human reflexion as a
sequence of subconscious reflexion, which allows to trace emotional reflection of
each reflexive image. This provides us with unique ability to unfold the sophis-
ticated structure of reflexive decision making process, involving emotions. Up to
date, there has been no approach, capable of doing such thing, reported.
The emotional research based on PAD model is transparent and clear. The
models of robots exhibiting human like emotional behavior using only PAD has
been successfully illustrated in recent book by Nishida et al. [23]. The ability
to influence the factors (variables) is the major justification for such approaches
and for the application of the RGT. Yet, it is just a "mechanical" part of the
highly sophisticated matter of human emotions.
The present study introduces the brand new approach to modeling of human
emotional behavior. We call this new breed of RGT application to be emotional
Reflexive Games (eRG). At present RGT fused with PAD model is the unique
approach allowing to explore the entire diversity of human emotional reflexion
and model reflexive interaction taming emotions. This fusion is automatically
formalized in algorithms and can be easily applied for further developing emo-
tional robots.
Since the proposed mechanism has no heavy negative impact on human psy-
chologycal state, robots should be enabled to deal with such approach in order
to provide human subjects with stress free psychological friendly environments
for decision making.
Emotionally Colorful Reflexive Games
11
Next, we discuss more complex question regarding emotion and reveal the
facts of why the Reflexive Game Theory is the natural approach to model emo-
tions.
We start from trying to answer the question - what are the emotions them-
selves?
According to Ekman [7], emotions shoud be regarded as highly automat-
ics processing algorithms responsible for our everyday life survival. Being entire
unconscious (automatic) regulators of the human body's physiology, emotions
bring the instantenous solutions of how to act in front of rapidly approaching
possible threat. This instantenous activity is possible due to unconscious pro-
cessing algorithms, which have been characterized by Lewicki et al. [20, 21] as
highly non-linear by nature and extremely fast by performance.
Therefore, emotions are fast processors of the information coming apart from
environment. Usually, the emotions are characterized by some physiological pat-
terns of body activity on the one hand, and external expression by face mimic or
gestures on the other hand. From this point of view, the fact that the reproduc-
tion by the self of physical part of emotional pattern, i.e., just making an angry
face can elicit the anger as emotional state itself [9, 19], is completely scientif-
ically unexpected phenomenon. Yet this seems to be widely known by writers.
For instance, Edgar Allan Poe, in his story "The Purloined Letter", describes
how one character is attemping to understand the intensity of emotional expe-
rience of another character by self-mimicing (or imitating) facial expressions of
another one.
This imitation is interesting from the point of view of reflexive processes.
Vladimir Lefebvre in his book "Structure of Awareness" [12] writes:
"Reflexion in traditional phylosophical psychological sense is ability
to place one's mind into external position of 'observer', 'researcher' or
'controler' of one's own body, actions and thoughts. We extend such
understanding of reflexion and will consider that reflexion is ability to
place one's mind into external position to another 'character', his actions
and thoughts.".
From this point of view, the self-mimicing is reflexive process. Thus, it is pos-
sible to elicit and understand the emotions of others by ref lexion. Ekman et al.
[9](p. 1210) suggested that this is possible due to "... direct connections between
motor cortex and hypothalamus that traslates between emotion-prototypic ex-
pression in the face and the emoition-specific pattering in the autonomic nervous
system".
On the other hand, the reflexive processes are at the core of the RGT. There-
fore RGT is natural approach to model emotions.
Continuing the line of logic that emotions can be elicited by self reproducing
patterns of motor activity accompanying expression of emotions, we suggest the
not only facial expressions, but also other motor patterns such as gestures can
elicit the corresponding emotion.
Consequently wide variaty of motor actions can cause some emotional states
to emerge. On the other hands, other actions, not related to the physical activity,
12
Sergey Tarasenko
can result in particular emotions. Therefore, it is possible to control human
emotions not directly, but making people to perform particular actions. From
this perspetive, PAD model can serve as a measuring device of how actions
influence emotions.
However, since the emotions are automatic processing algorithms, which take
their input apart from environment, the environment itself is natural medium to
influence on emotions. How this can be applied in interior of shops to provide a
customer with delightful emotional state is described by Soriano et al. [29].
Therefore, the continously on-going decision making process of selecting and
implementing action etc. is always accomapanied by emotional process, which
characterize each action from the emotional perspective.
We consider that the material presented in this study has essential influence
on the brand new Reflexive Game Theory. Furthermore, the results presented
extend widely recognized field of Affective Compiting [27] with new powerful
computational paradigm, which is ready to be applied in robots.
References
1. Adams-Webber, F.: A further test of golden section hypothesis. British Journal of
Psychology, 69, 439-442 (1978).
2. Adams-Webber, F.: Personal Construct Theory. New York; Wiley (1979).
3. Adams-Webber, F.: Assimilation and contrast in personal judgement: the di-
chotomy corollary. J. C. Mancuso and F. Adams-Webber (Eds.) In The construct-
ing person. New York: Praeger (1982).
4. Adams-Webber, F. and Benjafield, F.: The relation between lexical marking and
rating extremity in international judgement. Canadian Journal of Behavioral Sci-
ence, 5, 234-241 (1973).
5. Benjafield, F., and Adams-Webber, F.: The golden section hypothesis. British Jour-
nal of Psychology, 67, 11-15 (1976).
6. Benjafield, F. and Green, T.R.G.: Golden section relation in interpersonal judge-
ment. British Journal of Psychology, 69, 25-35 (1978).
7. Ekman, P.: Emotions Revealed: Recognizing Faces and Feelings to Improve Com-
munication and Emotional Life. Holt Paperbacks (2007).
8. Ekman, P., Levenson, R. W., and Friesen, W. V.: Autonomic nervous system ac-
tivity distinguishes between emotions. Science, 221, 1208-1211 (1983).
9. Ekman, P., and Davidson, R. J.: Voluntary changes regional brain activity. Psy-
chological Science, 4, 5, 342-345 (1993).
10. Kelly, G.A.: The psychology of personal constructs. New York: Norton (1955).
11. Le Gros Clark, W. E., and Meyer, M.: Anantomical relationship between the cere-
bral cortex and the hypothalamus. British Medical Bulletin, 6, 4, 341-344 (1950).
12. Lefebvre V.A.: The Structure of Awareness: Toward a Symbolic Language of Hu-
man Awareness. Beverly Hills, Calif.: Sage (1977).
13. Lefebvre, V.A.: Algebra of Conscience. D. Reidel, Holland (1982).
14. Lefebvre, V.A.: The Golden section and an algebraic model of ethical cognition.
Journal of Mathematical Psychology, 29, 289-310 (1985).
15. Lefebvre, V.A.: Algebra of Conscience. 2nd Edition. Holland: Kluwer (2001).
16. Lefebvre, V.A.: Lectures on Reflexive Game Theory. Leaf & Oaks, Los Angeles
(2010).
Emotionally Colorful Reflexive Games
13
17. Lefebvre, V.A.: Lectures on Reflexive Game Theory. Cogito-Center, Moscow (2009)
(in Russian).
18. Lefebvre, V.A.: Reflexive analysis of groups. In: Argamon, S. and Howard, N.
(eds.) Computational methods for counterterrorism. 173 -- 210. Springer, Heidelberg
(2009).
19. Levenson, R. W., Ekman, P., and Friesen, W. V.: Voluntary facial action generates
emotion specific autonomic nervous system activity. Psychophysiology, 27 (4), 363-
384 (1990).
20. Lewicki, P., Hill, T, and Bizot, E.: Acquisition of procedural knowledge about
a pattern of stimuli that cannot be articulated. Cognitive Psychology, 20, 24-37
(1988).
21. Lewicki, P., Hill, T, and Czyzewska, M.: Nonconscious Acquisition of Information.
American Psychologist, 47, 6, 796-801 (1992).
22. Mehrabian, A.: Pleasure-arousal-dominance: a general framework for describing
and measuring individual differences in temperament. Current Psychology: Devel-
opmental, Learning, Personality, Social 14, 4, 261-292 (1996).
23. Nishida, T., Jain, C. L., and Faucher, C.: Modelling Machine Emotions for Realiz-
ing Intelligence. Foundations and Application. Springer-Verlag Berlin Heidelberg
(2010).
24. Osgood, C.E.: From Yang and Yin to AND and BUT in cross cultural perspective.
International Journal of Psychology, 14, 1-35 (1979).
25. Osgood, C.E., and Richards, M.M.: From Yang and Yin to AND and BUT. Lan-
guage, 49, 1, 380-412 (1973).
26. Osgood, C.E., Suci, G., and Tannenbaum, P.: The measurement of meaning. Ur-
bana, IL: University of Illinois Press (1957).
27. Picard, R.W.: Affective Computing, MIT Technical Report 321 (1995).
28. Shalit, B.: The golden section relation in the evaluation of environment factors.
British Journal of Psychology, 71, 39-42 (1980).
29. Yani-de-Soriano, M.M., and Foxall, G.R.: The emotional power of place: The fall
and rise of dominance in retail research. Journal of Retailing and Consumer Ser-
vices, 13, 403-416 (2006).
30. Russell, J.A., and Mehrabian, A.: Evidence for a three-factor theory of emotions.
Journal of Research in Personality 11, 273-294 (1977).
31. Tarasenko, S. S., Inui, T., and Abdikeev, N.M.: The Golden Ratio-based Blind
Choice Performance. In Proceedings for 29th Annual Conference of the Cognitive
Science Society, 1864 (2007).
32. Tarasenko, S.: Modeling mixed groups of humans and robots with Reflexive Game
Theory. In Lamers, M.H., and Verbeek, F.J. (eds.): HRPR 2010, LINCST 59, 108-
117 (2011).
33. Tarasenko, S. The Inverse Task of the Reflexive Game Theory: Theoretical Mat-
ters, Practical Applications and Relationship with Other Issues (2010, submitted;
arXiv:1011.3397v1).
|
1505.01668 | 5 | 1505 | 2018-12-02T14:30:12 | Multi-Target Tracking in Distributed Sensor Networks using Particle PHD Filters | [
"cs.MA",
"eess.SY",
"stat.AP"
] | Multi-target tracking is an important problem in civilian and military applications. This paper investigates multi-target tracking in distributed sensor networks. Data association, which arises particularly in multi-object scenarios, can be tackled by various solutions. We consider sequential Monte Carlo implementations of the Probability Hypothesis Density (PHD) filter based on random finite sets. This approach circumvents the data association issue by jointly estimating all targets in the region of interest. To this end, we develop the Diffusion Particle PHD Filter (D-PPHDF) as well as a centralized version, called the Multi-Sensor Particle PHD Filter (MS-PPHDF). Their performance is evaluated in terms of the Optimal Subpattern Assignment (OSPA) metric, benchmarked against a distributed extension of the Posterior Cram\'er-Rao Lower Bound (PCRLB), and compared to the performance of an existing distributed PHD Particle Filter. Furthermore, the robustness of the proposed tracking algorithms against outliers and their performance with respect to different amounts of clutter is investigated. | cs.MA | cs | SUBMITTED FOR PUBLICATION
1
Multi-Target Tracking in Distributed Sensor
Networks using Particle PHD Filters
Mark R. Leonard, Student Member, EURASIP, and Abdelhak M. Zoubir, Member, EURASIP,
8
1
0
2
c
e
D
2
]
A
M
.
s
c
[
5
v
8
6
6
1
0
.
5
0
5
1
:
v
i
X
r
a
Abstract
Multi-target tracking is an important problem in civilian and military applications. This paper investigates multi-target tracking
in distributed sensor networks. Data association, which arises particularly in multi-object scenarios, can be tackled by various
solutions. We consider sequential Monte Carlo implementations of the Probability Hypothesis Density (PHD) filter based on
random finite sets. This approach circumvents the data association issue by jointly estimating all targets in the region of interest.
To this end, we develop the Diffusion Particle PHD Filter (D-PPHDF) as well as a centralized version, called the Multi-Sensor
Particle PHD Filter (MS-PPHDF). Their performance is evaluated in terms of the Optimal Subpattern Assignment (OSPA) metric,
benchmarked against a distributed extension of the Posterior Cram´er-Rao Lower Bound (PCRLB), and compared to the performance
of an existing distributed PHD Particle Filter. Furthermore, the robustness of the proposed tracking algorithms against outliers
and their performance with respect to different amounts of clutter is investigated.
Multi-target tracking, distributed target tracking, Particle Filter, PHD Filter, robustness, Posterior Cram´er-Rao Lower Bound
Index Terms
I. INTRODUCTION
T HE problem of multi-target tracking (MTT) is becoming increasingly important in many military and civilian applications
such as air and ground traffic control, harbor surveillance, maritime traffic control, or video communication and
surveillance [1] -- [3]. Distributed sensor networks offer a desirable platform for MTT applications due to their low cost and
ease of deployment, their lack of a single point of failure, as well as their inherent redundancy and fault-tolerance [4]. A
comprehensive overview of the state-of-the-art of distributed single-target tracking (STT) is given in [5]. Distributed versions
of the Kalman Filter [5], [6] and its nonlinear, non-Gaussian counterpart, the Particle Filter (PF) [7], have been well-studied.
However, they cannot be applied directly to MTT as they do not account for the problem of data association. Although there
are methods such as the Joint Probabilistic Data Association Filter (JPDAF) [8] or the Multiple Hypothesis Tracker (MHT)
[9] that address this problem in STT algorithms, the resource constraints in sensor networks might pose a challenge on finding
suitable distributed implementations [10]. The Probability Hypothesis Density (PHD) filter [11], [12], in contrast, resorts to
the concept of random finite sets (RFSs) to circumvent the problem of data association altogether.
The authors are with the Signal Processing Group, Institute of Telecommunications, Technische Universitat Darmstadt, Darmstadt 64283, Germany (e-mail:
[email protected], [email protected])
Manuscript submitted November 27, 2018
SUBMITTED FOR PUBLICATION
2
PSfrag replacements
25
20
15
10
5
0
-5
-10
-15
-20
-25
-25
]
m
[
y
30
-30
PSfrag replacements
Target 1
Target 2
Target 3
Target 1
Target 2
Target 3
]
m
[
y
30
-30
-20
-15
-10
-5
0
5
10
15
20
25
x[m]
(a)
25
20
15
10
5
0
-5
-10
-15
-20
-25
-25
-20
-15
-10
-5
0
5
10
15
20
25
x[m]
(b)
Fig. 1: (a) Distributed sensor network with 1-coverage of the region of interest and 3 exemplary target tracks. (b) Example
of tracking 3 targets with the Diffusion Particle PHD Filter (D-PPHDF). The small colored dots represent the target location
estimates obtained by the respective node with the same color. The light gray dots show the collective measurements obtained
by all nodes in the network.
In this work, we investigate distributed MTT in a sensor network with 1-coverage of the region of interest (ROI), i.e., the
sensor nodes have non- or barely overlapping fields of view (FOVs) and are distributed such that maximum area coverage is
attained [13]. An exemplary network layout with these properties is depicted in Figure 1a). Autonomous distribution algorithms
for realizing such a topology have been studied in our previous work [14]. The nodes in the network communicate with their
neighbors in order to collaboratively detect and track targets in the ROI. In addition, all of the sensors are equipped with a
signal processing unit, allowing them to form decisions without a fusion center. That way, the network can autonomously react
to events such as the detection of an intruder without relying on a network operator. For the sake of simplicity, the network
is considered to be static. However, the consideration of mobile sensor nodes would enable reactions such as target pursuit or
escape.
Since the FOV and communication radius of each node are limited, a target is only seen by a subset of the network, which
changes as the target moves through the ROI. Hence, at each time instant, there is an active and an inactive part of the network.
The goal, thus, is to detect and observe the target in a distributed and collaborative fashion as it travels across the ROI, rather
than reaching a network-wide consensus on its state and have the estimate available at each node.
In the sequel, we develop a distributed Particle PHD filter called Diffusion Particle PHD Filter (D-PPHDF), which uses
neighborhood communication to collaboratively estimate and track a single-sensor PHD at each node in the active subnetwork.
In addition, we formulate the Multi-Sensor Particle PHD Filter (MS-PPHDF), a centralized extension of the D-PPHDF. The
performance of both algorithms is evaluated in terms of the Optimal Subpattern Assignment (OSPA) metric [15], which is
calculated for the joint set of target state estimates of the active subnetwork. Furthermore, a distributed version of the Posterior
Cram´er-Rao Lower Bound (PCRLB) [16] -- [18] -- again averaged over the active subnetwork -- is introduced and used as a
SUBMITTED FOR PUBLICATION
3
benchmark. Moreover, we investigate the robustness of the proposed tracking algorithms against outliers and examine their
performance under different amounts of clutter.
Other distributed solutions for MTT in a multi-sensor setup using the PHD filter have been studied, e.g., in [19], [20], [21].
Contrary to our approach, they either assume overlapping FOVs or employ a pairwise communication scheme. The common
idea, however, is to extend the single-sensor PHD filter to the multi-sensor case through communication between multiple
nodes, or nodes and a fusion center. A more rigorous approach for MTT with multiple sensors is to use a multi-sensor PHD
filter [22], [23], which seeks to estimate and track a single multi-sensor PHD instead of multiple single-sensor PHDs. In this
work, we compare our methods to the approach in [20] (adapted to our scenario), which is also based on single-sensor PHDs.
The consideration of methods based on a multi-sensor PHD will be the focus of future work.
The paper is organized as follows: Section 2 presents the considered state-space model and recapitulates the theory of
RFSs as well as the PHD and the PHD filter. The problem of distributed MTT is addressed in Section 3. Here, we will first
detail our modification of Adaptive Target Birth (ATB) before formulating the D-PPHDF and investigate its computational
complexity and communication load. In Section 4, the MS-PPHDF is developed and analyzed in terms of computational
complexity and communication load. Section 5 is dedicated to simulations. First, the Distributed Posterior Cram´er-Rao Lower
Bound (DPCRLB) is introduced. Then, we present the simulation setup and discuss our results. Finally, a conclusion is given
in Section 6.
A. State-Space and Measurement Model
II. MODELS AND THEORY
A linear state-space model is considered for each target at time instant i ≥ 0. The target state vector stgt(i) = [xtgt(i), xtgt(i)]⊤
contains the target location vector xtgt as well as the velocity vector xtgt. For the sake of simplicity, we restrict ourselves to a
2D-environment. The target state evolves according to the state equation [24]:
stgt(i) = F (i)stgt(i − 1) + G(i)ntgt(i).
(1)
The matrices F and G as well as the vector ntgt will be explained shortly. Node k obtains a measurement zk of the target
location as given by the measurement equation [24]:
zk(i) = H k(i)stgt(i) + ν tgt
k (i),
k ∈ M
(2)
with M = {m ∈ {1, . . . , N } kxm(i) − xtgt(i)k2 ≤ Rsen} denoting the set of all nodes m that are located such that the
Euclidean distance kxm(i) − xtgt(i)k2 between their location xm and the target location xtgt is not greater than their sensing
radius Rsen. Note that N is the total number of nodes in the network. Furthermore, ntgt(i) ∼ N (02,1, Q(i)) and νtgt
N (02,1, Rk(i)) denote the state and measurement noise processes, respectively, with the zero-mean vector 02,1 = [0, 0]⊤
k (i) ∼
. Both
noise processes are spatially and temporally white, as well as uncorrelated with the initial target state stgt(0) and each other
SUBMITTED FOR PUBLICATION
for all i. For the sake of simplicity, we choose a time-invariant measurement noise covariance matrix
Rk(i) = Rk = σ2
r I 2,
where σ2
r is the variance of each component of the measurement noise and I n denotes the identity matrix of size n.
In target tracking, the model matrices are usually chosen to be time-invariant and given by [24]
4
(3)
(4)
F =
I 2 ∆iI2
02,2
I2
,
G =
∆i2
2 I 2
∆iI2
,
Q = σ2
q I 2,
where 02,2 is the 2 × 2 zero matrix. Furthermore, ∆i is the time step interval in seconds with which the state-space model
progresses. In addition, σ2
q denotes the variance of a state noise component. We assume that the sensor nodes only obtain
information on the location of a target. One common set of measurements that is often found in applications at sea is the
combination of distance and bearing measurements from which an estimate of the target location can easily be calculated. Since
we are not interested in the exact nature of the measured location information but rather in how this information is processed
by different tracking algorithms, we formulate our measurement model based on the local target location estimates at each
node. This gives us a general model that is applicable to a wide variety of application irrespective of the exact measurement
quantities. Thus, we obtain a general measurement matrix H k of the form
H k =(cid:20)I 2 02,2(cid:21) .
(5)
B. Random Finite Sets (RFSs)
A random finite set (RFS) is an unordered finite set that is random in the number of its elements as well as in their values
[25] -- [27]. Therefore, RFSs are a natural choice for representing the multi-target states and measurements in MTT: the state
and measurement vectors of all targets are collected in corresponding RFSs [28], [29]. Given the realization Ξi−1 of the RFS
Ξi−1 at time instant i − 1, the multi-target state of our tracking problem can be described by the RFS Ξi according to
Ξi = Si(Ξi−1) ∪ Bi,
(6)
where the survival set S i(Ξi−1) denotes the RFS of targets that already existed at time step i − 1 and have not exited the ROI,
i.e., the region covered by the sensor network, in the transition to time step i. In addition, the birth set Bi is the RFS of new
targets that spontaneously appear at the border of the ROI at time instant i [1], [12], [29]. Note that the statistical behavior of
Ξi can be described by the conditional probability fii−1(ΞiΞi−1).
The multi-target measurement model is given by the RFS Σi as
Σi = Θi(Ξi) ∪ Ci(Ξi),
(7)
where Θi(Ξi) is the RFS of measurements generated by Ξi. In addition, the RFS Ci(Ξi) represents clutter or false alarms.
Given a realization Σi of Σi, the statistical behavior of the RFS Σi is described by the conditional probability fi(ΣiΞi).
SUBMITTED FOR PUBLICATION
5
C. The Probability Hypothesis Density (PHD)
In analogy to the single-target case, the optimal Bayesian filter for MTT recursively propagates the multi-target posterior
fii(ΞiΣ0:i) over time, according to
fii(ΞiΣ0:i) =
fi(ΣiΞi)fii−1(ΞiΣ0:i−1)
R fi(ΣiΞ)fii−1(ΞΣ0:i−1)µs(dΞ)
fii−1(ΞiΣ0:i−1) =Z fii−1(ΞiΞ)fi−1i−1(ΞΣ0:i−1)µs(dΞ),
(8)
(9)
where µs is a dominating measure as described in [29]. This approach requires the evaluation of multiple integrals, which
makes it even more computationally challenging than its single-target counterpart. A common solution is to find a set of
statistics, e.g., the moments of first or second order, which yield a good approximation of the posterior, and propagate them
instead [1].
The Probability Hypothesis Density (PHD) Dii(s(i)Σ0:i) is an indirect first-order moment of fii(ΞiΣ0:i) [30]. It is given
by the following integral [12], [31]:
Dii(s(i)Σ0:i) = Xs
n (i)∈ΞiZ δ(s(i) − stgt
tgt
n (i))fii(s(i)Σ0:i)ds(i),
where R f (Y )δY denotes a set integral.
The PHD has the following two properties [30]:
1) The expected number of targets Ntgt(i) at time step i is obtained by integrating the PHD according to
Ntgt(i) =Z Dii(s(i)Σ0:i)ds(i).
(10)
(11)
This is in contrast to probability density functions (PDFs), which always integrate to 1.
2) Estimates of the individual target states can be found by searching for the j Ntgtm highest peaks of the PHD, where ⌊·⌉
denotes rounding to the nearest integer.
Because of these two properties, the number of targets as well as their states can be estimated independently at each time step
without any knowledge of their identities. That way, the data association issue is avoided. However, this also means that PHD
Filters cannot deliver the continuous track of a specific target. If continuous tracks are required, an additional association step
has to be performed. Two possible association algorithms for track continuity can be found in [11].
D. The PHD Filter
The PHD Filter is an approach for recursively propagating the PHD Dii(s(i)Σ0:i) at time step i given measurements up
to time step i over time. If the RFS Ξ is Poisson-distributed, then its PHD is equal to its intensity function and is, hence, a
SUBMITTED FOR PUBLICATION
sufficient statistic [12]. In this case, the PHD recursion is given by the following prediction and update equations [12]:
Dii−1(s(i)Σ0:i−1) = bi(s(i)) +Z pS(s(i − 1))fii−1(s(i)s(i − 1))Di−1i−1(s(i − 1)Σ0:i−1)ds(i − 1)
6
(12)
(13)
Dii(s(i)Σ0:i) ="1 − pD + Xz∈Σi
pDfi(zs(i))
λFAcFA(z) + pDR fi(zs(i))Dii−1(s(i)Σ0:i−1)ds(i)#Dii−1(s(i)Σ0:i−1)
Note that bi(s(i)) is the PHD of the birth set Bi of new targets appearing at time step i. In addition, pS(s(i − 1)) denotes the
probability that a target survives the transition from time step i − 1 to i. The probability of survival depends on the previous
state s(i − 1) because a target that is close to the border of the ROI and has a velocity vector pointing away from it is
unlikely to be present at time step i. Furthermore, fii−1(s(i)s(i − 1)) and fi(zs(i)) denote the transition probability and the
likelihood, respectively. The probability of detection pD is constant over time and the tracker's field of view (FOV) since it is
assumed that all targets can be detected if the ROI is covered. The term λFAcFA(z) represents Poisson-distributed false-alarms
due to clutter, where λFA is the false alarm parameter, which is distributed according to its spatial distribution cFA(z).
III. DISTRIBUTED MULTI-TARGET TARGET TRACKING
In this section we introduce the Diffusion Particle PHD Filter (D-PPHDF), a distributed Particle Filter implementation of
the PHD Filter for performing MTT in a sensor network without a fusion center. Before diving into the algorithm, we briefly
review the concept of Adaptive Target Birth (ATB) and discuss the modification we applied in the D-PPHDF.
A. Adaptive Target Birth (ATB)
Standard formulations of the PHD Filter consider the PHD bi(s(i)) of the birth set Bi to be known a priori [32]. For typical
tracking applications such as air surveillance, this is a reasonable assumption since new targets should appear at the border of
the ROI given continuous observation. An alternative is to make the target birth process adaptive and measurement-driven as
suggested in [32], [33]. To this end, the PHD -- and consequently the set of particles and weights approximating it in a Particle
Filter implementation -- is split into two densities corresponding to persistent objects, which have survived the transition from
time step i − 1 to i, and newborn objects, respectively.
In [32], [33], the PHD of newborn objects is approximated by randomly placing NP new particles around each target
measurement, with NP denoting the number of particles per target. We improve upon this approach by only considering
measurements with no noticeable impact on any persistent particle weight, as these may indicate the appearance of a new
target. That way, the number of newborn particles is further reduced and a possible overlap between persistent and newborn
PHD is avoided. With the transition to time step i + 1, the newborn particles become persistent. Furtherore, we perform the
ATB step towards the end of each iteration of the algorithm and only consider the particles representing the persistent PHD
in the prediction, weighting, and resampling steps. Hence, the update equation (13) does not have to be modified as in [32],
[33].
While ATB delays the tracking algorithm by one time step, it is much more efficient as it only places new particles in
regions in which a target is likely to be found. In addition, there is no need for an explicit initialization step since the first
incoming target will trigger the deployment of a newborn particle cloud around its corresponding measurement.
SUBMITTED FOR PUBLICATION
7
B. The Diffusion Particle PHD Filter (D-PPHDF)
The proposed Diffusion Particle PHD Filter (D-PPHDF) is an extension of the single-sensor Particle PHD Filter (PPHDF)
[11], [28], [34] for the multi-sensor case. Furthermore, it relies on ATB for a more efficient target detection. The communication
scheme we employ to exchange measurements and estimates between nodes is inspired by the two-step communication used
in the context of Diffusion Adaptation [35]. However, the algorithm does not rely on least-mean-squares or any other kind of
adaptive filter. First, each node k in the active part of the network obtains an intermediate estimate of the states of the targets
present, i.e., of the PHD of persistent targets -- represented by the set nsp
k,pers(i), wp
k,pers(i)oNk,pers(i)
p=1
corresponding weights -- based on neighborhood measurements. In other words, every active node runs a separate PPHDF with
of persistent particles with
access to measurements from its neighborhood Nk, defined as
where Rcom denotes the communication radius. Second, each active node combines the intermediate estimates from its
Nk = {l ∈ {1, . . . , N }(cid:12)(cid:12) kxl − xkk2 ≤ Rcom}, k = 1, . . . , N.,
(14)
neighborhood to a final, collaborative estimate. To this end, the persistent particle sets of all neighbors are merged into a
of persistent neighborhood particles and corresponding weights before the clustering
collective setnsp
k,coll(i), wp
k,coll(i)oNk,coll(i)
p=1
step, with Nk,coll(i) denoting the number of collective persistent neighborhood particles. In the sequel, we will look at the
individual steps of the D-PPHDF in more detail:
• Merging: The sets nsp
k,coll(i − 1), wp
k,coll(i − 1)oNk,coll(i−1)
p=1
and nsp
k,new(i − 1), wp
k,new(i − 1)oNk,new(i−1)
k,coll(i − 1) and sp
p=1
k,new(i − 1), respectively,
consist of the
collective persistent neighborhood particles and newborn particles of node k, sp
at time step i − 1 with their respective weights wp
k,coll(i − 1) and wp
k,new(i − 1). These sets are merged to become the
total set nsp
k,tot(i), wp
k,tot(i)oNk,tot(i)
p=1
of particles of node k at time step i, which is given by
of particles and weights of node k at time step i. Here, Nk,tot(i) is the total number
Nk,tot(i) = Nk,coll(i − 1) + Nk,new(i − 1),
(15)
with Nk,coll(i − 1) and Nk,new(i − 1) denoting the respective number of persistent neighborhood and newborn particles at
the previous time step. Note that since the sets of particles and weights represent PHDs, merging the sets corresponds to
summing these PHDs.
• Predicting: Each particle is propagated through the system model to become a persistent particle. The system model is
assumed to be the same for each target and given by Equation (1). Since the process noise is captured by the spread of
the particle cloud, the respective term can be removed from the equation, yielding
sp
k,pers(i) = F sp
k,tot(i),
p = 1, . . . , Nk,tot(i).
(16)
The corresponding weights are multiplied with the probability of survival pS, which is assumed to be constant for the
SUBMITTED FOR PUBLICATION
sake of simplicity1, according to
wp
k,pers(ii − 1) = pSwp
k,tot(i),
p = 1, . . . , Npers(i).
8
(17)
The prediction of particles and weights corresponds to the second term in Equation (12).
• Measuring & Broadcasting (1): The sensor nodes obtain measurements of the targets and forward them to their neighbors.
• Weighting: The persistent particle weights of node k are updated by applying a weighting step corresponding to Equation
(13) iteratively for each neighbor. Using the product operator, this weighting step can be compactly denoted as
wp
k,pers(i) = Yl∈Nk"1 − pD + Xzj ∈Σl
i
wp
k,j,update(i)#wp
k,pers(ii − 1),
with
wp
k,j,update(i) =
pDfi(zj xp(i))
λFAcFA(zj) + L(zj)
,
where Σl
i is the set of measurements obtained by node l and L(zj) is calculated as
L(zj) =
N p
k,pers(i)
Xq=1
pDfi(zj xq(i))wq
k,pers(ii − 1).
(18)
(19)
(20)
Note that fi(zj xp(i)) is the likelihood and xp(i) is the location vector of particle p.
Afterwards, each node k obtains the set Σk
i,cand of candidate measurements, i.e., measurements that are not responsible for
the highest weighting of any persistent particle, to be used in the ATB step later on. The set Σk
i,cand is found according to
Σk
i,cand = Σk
i \nzmp (cid:12)(cid:12)(cid:12)
mp = arg max
j
wp
k,j,update(i), p = 1, ..., Nk,pers(i)o.
(21)
• Resampling: Each node k calculates its own expected number of targets Nk,tgt(i) from its total persistent particle mass
according to
Consequently, the number of persistent particles of node k is updated as
Nk,tgt(i) =
Nk,tot(i)
Xp=1
wp
k,pers(i)
Nk,pers(i) = Nk,tgt(i)NP .
.
(22)
(23)
Furthermore, the set of persistent particles of node k has to be resampled by drawing Nkpers(i) particles with replacement
from it. Note that the probability of drawing particle p is given by
wp
k,pers(i)
Nk,tgt(i)
since the weights do not sum to unity. Then,
1A constant probability of survival pS is a reasonable assumption if the targets move relatively slowly with respect to the observation time and the size of
the ROI.
SUBMITTED FOR PUBLICATION
the weights are reset to equal values as
wp
k,pers(i) =
Nk,tgt(i)
Nk,pers(i)
,
p = 1, . . . , Nk,pers(i).
9
(24)
• Broadcasting (2): Every node k transmits its set of resampled persistent particles and weightsnsp
k,pers(i), wp
to its neighbors.
k,pers(i)oNk,pers(i)
p=1
• Clustering: Each node k forms a collective set of persistent neighborhood particles sp
k,coll(i) and corresponding weights
wp
k,Nh(i) according to
with
k,coll(i), wp
nsp
k,coll(i)oNk,coll(i)
p=1
= [l∈Nknsp
l,pers(i), wp
l,pers(i)oNl,pers(i)
p=1
Nk,coll(i) = Xl∈Nk
Nl,pers(i)
,
(25)
(26)
denoting the number of collective persistent neighborhood particles of node k. As in the merging step, this corresponds
to summing the corresponding PHDs to obtain an updated single-sensor PHD with a probability distribution reflecting the
information of the entire neighborhood of node k. Note that the PHDs might not be independent if a target is detected by
more than one neighbor. However, this is not a problem since merging the particle sets simply results in the respective
target being represented by more particles. Hence, node k will be able to estimate the corresponding location more
accurately.
The estimated target states are found by clustering the collective persistent particles. Since the expected number of targets
Nl,tgt(i), l ∈ Nk might be different for each neighbor, we resort to hierarchical clustering of the single-linkage type [36].
Here, the sum of the expected number of targets over the neighborhood can serve as an upper bound for the number of
clusters. Note, however, that if two targets are close to each other, clustering algorithms might not be able to resolve both
targets correctly.
• Roughening: A roughening step is performed to counter sample impoverishment [37]. To this end, an independent jitter
sj(i) is added to every resampled particle. Each component sj
c(i), c = 1, . . . , d of the jitter with dimensionality d is
sampled from the Gaussian distribution N (0, (σj
c (i))2). The component-wise standard deviation of the jitter is given by
c (i) = KEcNk,coll(i)−1/d,
σj
(27)
where Ec is the interval length between the maximum and minimum samples of the respective component. To avoid
evaluating Ec separately for each particle cluster, it is assigned an empirically found constant value.2 Note that d = 4
since the dimensionality of the jitter vector sj(i) and the particle state vector sp(i) have to coincide. In addition, K is a
tuning constant, which controls the spread of the particle cloud.
• Adaptive Target Birth: NP new particles are placed randomly around each candidate measurement zj ∈ Σk
i,cand leading
2Since the noise variances as well as the network topology are fixed, the true value of Ec will not change significantly over time and between clusters, so
this is a valid simplification.
SUBMITTED FOR PUBLICATION
10
to a total number of Nk,new(i) = NP · Σk
i,cand newborn particles for node k. Every newborn particle is associated with a
weight that is chosen according to
wp
k,new(i) =
pB
k,new(i)
N p
,
p = 1, . . . , Nk,new(i),
(28)
where pB is the probability of birth. Depending on the application, pB can depend on time as well as on the location
of the respective particle. For simplicity, the probability that a new target enters the ROI is assumed to be equal for all
locations in the birth region over time. The target birth process corresponds to the first term in Equation (12).
Figure 1b) shows an example of tracking three targets, which move along the deterministic tracks depicted in Figure 1a),
using the D-PPHDF. Note that each small colored dot corresponds to a target location estimate obtained by the respective node
with the same color while the light grey dots represent the collective measurements from all nodes. From this illustration, the
following properties of the D-PPHDF are apparent: First, the algorithm only delivers separate location estimates -- represented
by the small colored dots -- for each time instant rather than continuous tracks, which -- as mentioned before -- is a common
property of PHD filters. Second, the network as a whole would be able to correctly track all three targets, while a single node
only obtains the locally relevant subtracks of the targets in its vicinity. Third, the employed two-step communication scheme
is able to extend the vicinity of a node far beyond its own sensing radius of Rsen = 6 m. This can, for instance, be seen from
the fact that the lime-green node located at [−14, −23] is able to obtain location estimates of target 2, which enters the ROI
from the south. Finally, Figure 1b) also illustrates the resolution problem of clustering. When targets 1 and 2, which enter the
ROI from the north and the south, respectively, cross paths, the nodes in their vicinity see them as just one target. This leads
to an aggregation of target location estimates around [9, 0].
The pseudo-code of the D-PPHDF is given in Table I.
C. Computational Complexity and Communication Load
In this section we take a look at the computational complexity and the communication load the D-PPHDF imposes on each
node in the active subnetwork. The following steps are performed at every time instant i but time dependency is omitted for
simplicity. Note that each of the steps scales with the number of active nodes when considering the computational complexity
of the network as a whole.
• Prediction: The prediction step described by Equations (16) and (17) is performed for each particle at every active node.
Hence, it scales with the number of particles Nk,tot and the dimensionality d of the particle vectors. In order to obtain a
tractable expression for the computational complexity, we assume each node to have the same number of particles Ntot.
⇒ O(Ntotd)
• Weighting: Each particle is updated in the weighting step given by Equations (18)-(21). The weight update as well as the
designation of candidate measurements for ATB depends on the neighborhood size Nk of node k and the number of
measurements(cid:12)(cid:12)Σl(cid:12)(cid:12) of each of its neighbors l. For tractability reasons, we assume each node to have the same number of
neighbors Nnb and to obtain the same number of measurements Nmeas.
⇒ O(NtotNnbNmeas)
SUBMITTED FOR PUBLICATION
1: input: d, Ec, K, n, N, NP , pB, pS, λFA, cFA
k,coll(0), wp
2: initialize: nsp
3: while i ≤ n do
4:
for k = 1, . . . , N do
k,coll(0)oNk,coll(0)
p=1
=nsp
k,new(0), wp
k,new(0)oNk,new(0)
p=1
= ∅.
5:
Merge the sets of collective persistent and newborn particles with corresponding weights:
k,tot(i), wp
nsp
k,tot(i)oNk,tot(i)
p=1
=nsp
k,coll(i − 1), wp
k,coll(i − 1)oNk,coll(i−1)
p=1
∪nsp
k,new(i − 1), wp
k,new(i − 1)oNk,new(i−1)
p=1
6:
7:
for p = 1, . . . , Nk,tot(i) do
Predict the new state of each particle and update the weight with the probability of survival pS:
k,pers(i) = F sp
sp
k,pers(ii − 1) = pSwp
wp
k,tot(i)
k,tot(i).
8:
Update the weights using neighborhood measurements:
11
.
wp
k,pers(i) = Yl∈Nk"1 − pD + Xzj ∈Σl
i
wp
k,j,update(i)#wp
k,pers(ii − 1),
wp
k,j,update(i) =
pDfi(zj xp(i))
λFAcFA(zj) + L(zj)
Nk,pers(i)
,
L(zj) =
Xq=1
pDfi(zj xq(i))wq
k,pers(ii − 1).
end for
Form the set of candidate measurements for ATB:
Σk
i,cand = Σk
mp = arg max
j
wp
k,j,update(i),
p = 1, ..., Nk,pers(i),
i \nzmp (cid:12)(cid:12)(cid:12)
j = 1, ...,(cid:12)(cid:12)Σl
i(cid:12)(cid:12) ∀l ∈ Nko.
9:
10:
11:
12:
14:
15:
16:
17:
end for
i ← i + 1
18:
19: end while
20: return
Calculate the estimated number of targets:
Resample Nk,pers(i) = Nk,tgt(i)NP particles and reset the weights:
Nk,tgt(i) =
Nk,tot(i)
Xp=1
wp
.
k,pers(i)
wp
k,pers(i) =
Nk,tgt(i)
Nk,pers(i)
,
p = 1, . . . , Nk,pers(i).
13:
Merge the sets of persistent neighborhood particles and weights:
k,coll(i), wp
nsp
k,coll(i)oNk,coll(i)
p=1
= [l∈Nknsp
l,pers(i), wp
l,pers(i)oNl,pers(i)
p=1
.
Use single-linkage clustering to identify Ntgt(i) clusters and find the set of estimated target states nsl
by calculating the centroids.
Add an independent jitter to each particle using a component-wise standard deviation of:
k(i)o
Ntgt(i)
l=1
Place NP new particles randomly around each candidate measurement zj ∈ Σk
i,cand. Set the weights as:
c (i) = KEcNk,coll(i)−1/d.
σj
wp
k,new(i) =
pB
Nk,new(i)
,
p = 1, . . . , Nk,new(i).
SUBMITTED FOR PUBLICATION
12
• Resampling: The estimation of the number of targets and the resampling step in Equations (22)-(24) are linear in the
number of particles used for the calculation [38]. For the sake of simplicity, we assume each active node to have the same
estimate of the number of targets Ntgt.
⇒ O(Ntot + NactiveNtgtNP )
• Clustering: The complexity of single-linkage clustering is cubic in the number of particles, i.e., in the number of neighbors
Nnb of each node, the estimated number of targets Ntgt, the number of particles per target NP , and the dimensionality d
of the particles [39].
⇒ O((NnbNtgtNP d)3)
• Roughening: Roughening (Equation (27)) is performed for every collective particle and is linear in the dimensionality of
the particles.
⇒ O(NnbNtgtNP d)
• Adaptive Target Birth: The birth process depends on the number of particles per target NP as well as the number of
candidate measurements Ncand, which is assumed equal for each active node to ensure tractibility.
⇒ O(NP Ncand)
As far as the communication load is concerned, the D-PPHDF requires the broadcasting of measurements, i.e., 2 scalars per
measurement, over the neighborhood in the first broadcasting step. In the second step, the sets of particles and weights, i.e., 5
scalars per particle, are transmitted. Clearly, the communication load strongly depends on the number of nodes in the network,
or more precisely the number of active nodes and the size of their respective neighborhood. As an extension of the D-PPHDF,
one could think of changing the second broadcasting step and transmit Gaussian Mixture Model representations -- instead of
the actual particles and weights -- that will be resampled at the receiver node (see e.g., [40]). That way, communication load
could be reduced to transmitting only a few scalars in the second broadcasting step at the cost of estimation accuracy and
additional computational complexity. However, a thorough treatment of this extension is beyond the scope of this work.
Having presented the D-PPHDF as a distributed solution for MTT in a sensor network, we propose the centralized counterpart
IV. CENTRALIZED MULTI-TARGET TRACKING
to our approach in the sequel.
A. The Multi-Sensor Particle PHD Filter (MS-PPHDF)
The proposed Multi-Sensor Particle PHD Filter (MS-PPHDF) is a centralized, multi-sensor PPHDF that relies on a fusion
center with access to the measurements of all nodes in the network. It is based on the formulation of the single-sensor PPHDF
in [11], [28], [34] but with an extended measurement set comprising the measurements of the entire network. Hence, one
might obtain more than one measurement per target -- a change to the typical assumption in target tracking that each target
produces at most one measurement [8]. To account for this change, we add a pre-clustering step before the weighting step and
normalize the weight update accordingly. A similar partitioning of the measurement set is used in extended target tracking,
where a sensor can receive multiple target reflections due to the target's physical extent [41], [42].
SUBMITTED FOR PUBLICATION
13
In the following, we will look at the individual steps of the algorithm in more detail:
• Merging: The sets {sp
pers(i − 1), wp
pers(i − 1)}Npers(i−1)
p=1
and {sp
new(i − 1), wp
new(i − 1)}Nnew(i−1)
p=1
consist of the persistent and
newborn particles, sp
pers(i − 1) and sp
new(i − 1), respectively, at time step i − 1 with their respective weights wp
pers(i − 1)
and wp
new(i − 1). These sets are merged to become the total set {sp
tot(i), wp
tot(i)}Ntot(i)
p=1
of particles and weights at time step
i. Here, Ntot(i) is the total number of particles at time step i, which is given by
Ntot(i) = Npers(i − 1) + Nnew(i − 1),
(29)
with Npers(i − 1) and Nnew(i − 1) denoting the respective number of persistent and newborn particles at the previous time
step.
• Predicting: As in the D-PPHDF, each particle is propagated through the system model according to
pers(i) = F sp
sp
tot(i),
p = 1, . . . , Npers(i) = Ntot(i)
to become a persistent particle. The corresponding weights are multiplied with the probability of survival pS as
wp
pers(ii − 1) = pSwp
tot(i),
p = 1, . . . , Npers(i).
(30)
(31)
• Measuring: The sensor nodes obtain measurements of the targets.
• Pre-Clustering: Since there might be more than one measurement per target, the measurements of the entire network are
pre-clustered before the weighting step and each measurement is assigned a label C(z) that reflects the cardinality of its
own cluster. This can be done, for instance, using single-linkage clustering [36]. The clustering is based on the distance
between measurements, i.e., spatially close measurements are assumed to stem from the same target. Hence, when two
or more targets are too close to each other, cardinality errors may occur.
• Weighting: All available target measurements, which comprise the set Σi, are used to update the persistent particle weights
according to
with
and
wp
pers(i) ="1 − pD + Xzj ∈Σi
wp
j,update(i)#wp
pers(ii − 1),
wp
j,update(i) =
pDfi(zj xp(i))
(λFAcFA(zj) + L(zj)) C(zj)
,
L(zj) =
Npers(i)
Xq=1
pDfi(zj xq(i))wq
pers(ii − 1).
(32)
(33)
(34)
Note that -- in contrast to the D-PPHDF -- the weighting step is applied only once using the entire set of measurements.
Therefore -- and since there might be more than one measurement per target -- we have to ensure that the weight update
terms wp
j,update -- and consequently the particle weights -- still sum to the number of targets present. This is done by
SUBMITTED FOR PUBLICATION
14
normalizing Equation (33) with C(zj), i.e., the cardinality of the cluster to which the current measurement zj belongs.
Afterwards, we form the set Σi,cand of candidate measurements for the ATB step according to
• Resampling: The expected number of targets Ntgt(i) is calculated from the total persistent particle mass as
Σi,cand = Σi\nzmp (cid:12)(cid:12)(cid:12)
mp = arg max
j
wp
j,update(i), p = 1, ..., Npers(i)o.
Consequently, the number of persistent particles is updated according to
Ntgt(i) =
Ntot(i)
Xp=1
wp
.
pers(i)
Npers(i) = Ntgt(i)NP .
(35)
(36)
(37)
Furthermore, the set of persistent particles is resampled by drawing Npers(i) particles with probability
wp
pers(i)
Ntgt(i)
. Then, the
weights are reset to equal values as
wp
pers(i) =
Ntgt(i)
Npers(i)
,
p = 1, . . . , Npers(i).
(38)
• Clustering: In contrast to the D-PPHDF, there is only one estimate of the expected number of targets. Hence, we can use
k-means clustering [43] to find the estimated target states by grouping the resampled particles into Ntgt(i) clusters and
calculating the centroid of each cluster.
• Roughening: Roughening is performed analogously to the D-PPHDF.
• Adaptive Target Birth: NP new particles are placed randomly around each candidate measurement zj ∈ Σi,cand yielding
a total number of Nnew(i) = NP · Σi,cand newborn particles. The corresponding weights are chosen according to
wp
new(i) =
pB
Nnew(i)
,
p = 1, . . . , Nnew(i),
(39)
where pB is the probability of birth.
The pseudo-code of the MS-PPHDF is given in Table II.
B. Computational Complexity and Communication Load
In this section we analyze the computational complexity and the communication load of the MS-PPHDF. The following
steps are performed at every time instant i but time dependency is omitted for simplicity:
• Prediction: The prediction step described by Equations (30) and (31) is performed for each of the Ntot particles and is
linear in the dimensionality d.
⇒ O(Ntotd)
• Pre-Clustering: The pre-clustering step relies on single-linkage clustering. The complexity is therefore cubic in the total
number of measurements Nmeas. [39]
⇒ O(N 3
meas)
SUBMITTED FOR PUBLICATION
15
1: input: d, Ec, K, n, N, NP , pB, pS, λFA, cFA
2: initialize: {sp
p=1 = {sp
3: while i ≤ n do
4: Merge the sets of persistent and newborn particles with corresponding weights:
coll(0)}Ncoll(0)
new(0)}Nnew(0)
p=1 = ∅.
coll(0), wp
new(0), wp
{sp
tot(i), wp
tot(i)}Ntot(i)
pers(i − 1), wp
∪ {sp
new(i − 1), wp
new(i − 1)}Nnew(i−1)
p=1
.
p=1 =(cid:8)sp
pers(i − 1)(cid:9)Npers(i−1)
p=1
5:
6:
7:
8:
for p = 1, . . . , Ntot(i) do
Predict the new state of each particle and update the weight with the probability of survival pS:
pers(i) = F sp
sp
pers(ii − 1) = pSwp
tot(i)
tot(i).
wp
Cluster the measurements using single-linkage clustering and assign each measurement z a label C(z) reflecting
the cardinality of its cluster.
Update the weights using the measurements of the entire network:
wp
pers(i) ="1 − pD + Xzj ∈Σi
wp
j,update(i)#wp
pers(ii − 1)
wp
j,update(i) =
pDfi(zj xp(i))
(λFAcFA(zj) + L(zj)) C(zj)
Npers(i)
,
9:
10:
end for
Form the set of candidate measurements for ATB:
L(zj) =
Xq=1
pDfi(zj xq(i))wq
pers(ii − 1).
Σi,cand = Σi\nzmp (cid:12)(cid:12)(cid:12)
Calculate the estimated number of targets:
mp = arg max
j
wp
j,update(i),
p = 1, ..., Npers(i),
j = 1, ..., Σio.
Resample Npers(i) = Ntgt(i)NP particles and reset the weights:
Ntgt(i) =
Ntot(i)
Xp=1
wp
.
pers(i)
wp
pers(i) =
Ntgt(i)
Npers(i)
,
p = 1, . . . , Npers(i).
Find the set of estimated target states nsl(i)o
each cluster.
Add an independent jitter to each particle using a component-wise standard deviation of:
Ntgt(i)
l=1
by using k-means clustering and calculating the centroid of
11:
12:
13:
14:
15:
Place NP new particles randomly around each candidate measurement zj ∈ Σi,cand. Set the weights as:
c (i) = KEcNpers(i)−1/d.
σj
wp
new(i) =
pB
Nnew(i)
,
p = 1, . . . , Nnew(i).
i ← i + 1
16:
17: end while
18: return
TABLE II: The Multi-Sensor Particle PHD Filter.
SUBMITTED FOR PUBLICATION
16
• Weighting: Each particle is updated in the weighting step given by Equations (32)-(34). The weight update as well as the
designation of candidate measurements for ATB depends on the number of measurements Nmeas = Σ.
⇒ O(NtotNmeas)
• Resampling: The estimation of the number of targets and the resampling step in Equations (36)-(38) are linear in the
number of particles used for the calculation [38].
⇒ O(Ntot + NtgtNP )
• Clustering: In contrast to the D-PPHDF we can use k-means clustering. The complexity of Lloyd's implementation is
given by [44]
⇒ O((NtgtNP )dNtgt+1 log(NtgtNP )).
• Roughening: Roughening is linear in the dimensionality of the particles and their number.
⇒ O(NtgtNP d)
• Adaptive Target Birth: The birth process depends on the number of particles per target NP as well as the number of
candidate measurements Ncand.
⇒ O(NP Ncand)
In summary, the computational complexity of the MS-PPHDF is largely comparable to that of the D-PPHDF. The only
exception is the pre-custering step, which scales cubicly with the total number of measurements and adds additional complexity
to the algorithm. As a tradeoff the communication load of the MS-PPHDF clearly is lower compared to the D-PPHDF because
there is only the initial transmission of measurements from the nodes to the fusion center. However, considering a setup
with relatively small communication radii, this initial communication step requires a lot of relaying and leads to high traffic
density in the vicinity of the fusion center. Furthermore, this communication structure exhibits a single point of failure while
a distributed sensor network is inherently redundant.
V. SIMULATIONS
In this section, we evaluate the performance of the D-PPHDF as well as the MS-PPHDF for tracking multiple targets
in a sensor network with 1-coverage. To this end, we consider Gaussian measurement noise of different variance as well
as ε-contaminated noise with different contamination ratios to investigate the robustness of the algorithms. In addition, the
performance for different amounts of clutter is analyzed. We compare the proposed algorithms to the alternative distributed
PPHDF from [20], which will be referred to as Distributed Data Fusion Particle PHD Filter (DDF-PPHDF). Here, each node
runs its own PPHDF using only its own measurements. In a subsequent step, the particles are distributed over the neighborhood
and reweighted by fusing their corresponding Exponential Mixture Densities.
Furthermore, we formulate the DPCRLB as a lower bound for evaluating the performance of the three algorithms in terms
of the OSPA [15] metric. In our simulations, we compute the OSPA metric with respect to the joint set of target state estimates
of the entire active network. The latter is found by clustering the target state estimates of all active nodes. Furthermore, we
p (cid:17)2
consider the squared OSPA metric scaled by the number of targets, i.e., Ntgt ·(cid:16) ¯d(c)
DPCRLB, which will be introduced in the following, as a benchmark.
, as in [45]. That way, we can use the
SUBMITTED FOR PUBLICATION
17
A. The Distributed Posterior Cram´er-Rao Lower Bound (DPCRLB)
Rather than evaluating the performance of the different MTT algorithms based on an error metric, it makes more sense to
derive a minimum variance bound on the estimation error, which enables an absolute performance evaluation. For time-invariant
statistical models, the most commonly used bound is the Cram´er-Rao Lower Bound (CRLB), which is given by the inverse
of Fisher's information matrix [46]. In [45] and [47], the CRLB is used in the context of multi-sensor MTT of an unknown
number of unlabeled targets in order to evaluate the performance, as well as prove the asymptotic efficiency of the PHD as
the number of nodes goes to infinity. Since we are more interested in the tracking behavior of a fixed network over time, we
resort to the Posterior Cram´er-Rao Lower Bound (PCRLB), which is an extension of the CRLB for the time-variant case [16].
This bound can be calculated sequentially with the help of a Riccati-like recursion derived in [48]. Furthermore, in [17] and
[49], the PCRLB is adapted for an MTT scenario in which the tracker can obtain more than one measurement per target.
Let πm
i , m = 1, ..., M denote the probability that any measurement is associated with target m at time instant i as defined in
[18]. With the corresponding stochastic process Πm
i , the new stochastic process of association probabilities and target states to
can now be formed as described
be estimated becomes Φi =(cid:16)Π1:M
i
, Ξ1:M
i
(cid:17). Fisher's information matrix JΦi =
JΠi
J Πi
Ξi
J Ξi
Πi
JΞi
in [49] and [17]. However, as the number of targets varies over time, i.e., targets might enter or exit the ROI, JΦi has to be
expanded or shrunk in the inverse matrix domain as described in [50]. The PCRLB Bi at time instant i can be obtained as
the trace of the inverted submatrix JΞi according to [49]
Bi = trace(cid:26)hJΞi − J Πi
Ξi
J −1
Πi
J Ξi
Πii−1(cid:27) .
(40)
Note that, in a distributed MTT scenario, Bi corresponds to a lower bound on the estimation error of a central processing
unit with access to all measurements. Since we are interested in completely distributed MTT with in-network processing, we
extend the PCRLB to the Distributed Posterior Cram´er-Rao Lower Bound (DPCRLB). To this end, each node k computes its
own PCRLB Bk
i considering only the measurements of its two-hop neighborhood, which is given by
N (2)
k = [l∈Nk
Nl,
(41)
i.e., the neighbors of node k and their neighbors. Furthermore, only the targets within the sensing range of Nk are taken into
account. Clearly, only nodes with a neighborhood in the vicinity of at least one target will be able to calculate a PCRLB. The
DPCRLB Bi,dist at time instant i is then obtained by averaging over these values according to
where M is the set of all nodes that are able to compute a PCRLB.
Bi,dist =
1
M Xk∈M
Bk
i ,
(42)
SUBMITTED FOR PUBLICATION
18
VARIABLE VALUE
DESCRIPTION
∆i
N
σ2
r
σ2
q
ε
Rcom
Rsen
Ec
K
NP
pB
pD
pS
λFA
cFA(z)
c
p
1 s
30
time step of the tracking algorithm
number of nodes
0.1, 0.3 m2
0.01 m2
0.1, 0.3
componentwise power of meas. noise
componentwise power of state noise
contamination ratio
2Rsen
6 m
6
0.2
500
0.8
0.95
0.98
0.1, 0.3
1
πR2
sen
2
2
communication radius
sensing radius
empirical interval length for jitter
tuning constant for roughening
number of particles per target
probability of birth
probability of detection
probability of survival
average no. of false alarms / clutter
PDF of false alarms / clutter (uniform)
cut-off value (OSPA)
order of the OSPA metric
TABLE III: Simulation parameters
B. Simulation Setup
In the following simulations, a static sensor network as depicted in Figure 1a) is used to perform MTT. The network is
centered around the point of origin [0, 0]⊤
and distributed such that 1-coverage of the ROI is guaranteed. It covers an area of
approximately 2500 m2. Clutter is assumed Poisson and uniformly distributed over the sensing range of each node with an
average rate of λFA = 0.1 and 0.3. Moreover, we consider Gaussian measurement noise with variance σ2
r = 0.1 and 0.3 as
well as ε-contamination noise with a ten-times higher variance and a contamination rate of ε = 0.1 and 0.3. For the sake of
simplicity, collisions between targets and sensor nodes are neglected.
An overview of all simulation parameters is given in Table III. Since the purpose of this work is to introduce the MS-PPHDF
as well as the D-PPHDF, verify their functionality, and compare them to alternative approaches, we consider a rather simple
scenario with a high probability of detection and relatively low clutter levels. In our future work, we will study more sophisticated
scenarios to define possible breakdown points of our algorithms.
We use the MS-PPHDF, the D-PPHDF, as well as the DDF-PPHDF to track three targets for i = 0, ..., 30. The targets enter
the ROI at time steps i = 0, 9, 14 from the north, south, and west, respectively. A Monte Carlo simulation with NMC = 1000
runs is performed to evaluate the performance of the tracking algorithms in terms of the Optimal Subpattern Assignment
(OSPA) metric. Note that the target trajectories as shown in Figure 1a) are deterministic, as is often the case in target tracking
simulations [50] in order to guarantee the comparability of the different Monte Carlo runs regarding, for instance, the number
of targets present.
SUBMITTED FOR PUBLICATION
MS-PPHDF
D-PPHDF
DDF-PPHDF
DPCRLB
9
8
7
6
0.2
PSfrag replacements
2
0.4
(cid:17)
5
t
s
i
d
,
i
35
B
d
n
a
4
)
2
0.6
2
(
¯d
(cid:16)
0.8
·
3
Ntgt ·(cid:16) ¯d
2 (cid:17)2
3.5
3
35
2.5
19
0.2
PSfrag replacements
0.4
2
t
s
e
0.6
N
0.8
1.5
t
g
t
N
1.2
true number
1.4
1.6
1.8
t
s
i
d
,
i
35
B
d
n
a
0.2
2
1
0
0
DPCRLB
1.2
1.4
and Bi,dist
1.6
1.8
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
5
10
15
20
25
30
Time steps i (seconds)
(a) Squared and scaled OSPA compared to DPCRLB, σ2
r
= 0.1
MS-PPHDF
D-PPHDF
DDF-PPHDF
DPCRLB
9
8
7
6
5
35
0.2
0.4
2
0.4
(cid:17)
4
)
2
0.6
2
(
¯d
(cid:16)
0.8
·
3
t
g
t
N
1.2
1.4
1.6
1.8
2
1
0
0
5
10
15
20
25
30
Time steps i (seconds)
(c) Squared and scaled OSPA compared to DPCRLB, σ2
r
= 0.3
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
5
10
15
20
25
30
Time steps i (seconds)
(b) Estimated number of targets, σ2
r
= 0.1
1
0.5
0
0
3.5
3
2.5
2
t
s
e
0.6
N
0.8
1.5
1.2
1.4
1.6
1.8
1
0.5
0
0
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
5
10
15
20
25
30
Time steps i (seconds)
(d) Estimated number of targets, σ2
r
= 0.3
Fig. 2: Simulation I: Results for Gaussian noise and clutter rate λFA = 0.1. The left part of the figure shows the squared and
scaled Optimal Subpattern Assignment (OSPA) metric for each algorithm compared to the Distributed Posterior Cram´er-Rao
Lower Bound (DPCRLB), while the right part compares the estimated to the true number of targets.
C. Simulation I: Results
In the first simulation, we compare the performance of the MS-PPHDF and the D-PPHDF to the alternative DDF-PPHDF
and the DPCRLB, which serves as a benchmark. Measurement noise is zero-mean Gaussian with variance σ2
r = 0.1, 0.3 and
the average number of clutter is 0.1.
The simulation results are depicted in Figure 2. While the top part considers zero-mean Gaussian measurement noise with a
per-component variance of σ2
r = 0.1, the bottom part shows the results for σ2
r = 0.3. In addition to evaluating the performance
of the MS-PPHDF, the D-PPHDF, and the DDF-PPHDF in terms of the squared and scaled OSPA metric over time and
comparing it to the DPCRLB as can be seen in the left part of the figure, we also look at the estimated number of targets,
which is depicted in the right column. Since the OSPA metric contains a penalty for an erroneous estimate of the number of
targets, this side-by-side comparison facilitates the interpretation of the tracking results.
Let us start by considering Figures 2a) and 2b), i.e., the case of σ2
r = 0.1. First of all, we observe that neither tracking
SUBMITTED FOR PUBLICATION
2.5
2
1.5
t
s
i
d
,
i
35
B
d
0.2
n
a
0.4
2
(cid:17)
)
0.6
2
2
(
¯d
1
(cid:16)
0.8
·
t
g
t
N
1.2
1.4
1.6
1.8
0.5
0
0
MS-PPHDF
D-PPHDF
DDF-PPHDF
DPCRLB
5
10
15
20
25
30
Time steps i (s)
2.5
2
1.5
t
s
i
d
,
i
35
B
d
0.2
n
a
0.4
2
(cid:17)
)
0.6
2
2
(
¯d
1
(cid:16)
0.8
·
t
g
t
N
1.2
1.4
1.6
1.8
0.5
0
0
20
MS-PPHDF
D-PPHDF
DDF-PPHDF
DPCRLB
5
10
15
20
25
30
Time steps i (s)
(a) Squared and scaled OSPA compared to DPCRLB, σ2
r
= 0.1
(b) Squared and scaled OSPA compared to DPCRLB , σ2
r
= 0.3
Fig. 3: Simulation I: Results for Gaussian noise and clutter rate λFA = 0.3 (zoomed in). The squared and scaled Optimal
Subpattern Assignment (OSPA) metrics of using the Multi-Sensor Particle PHD Filter (MS-PPHDF), the Diffusion Particle
PHD Filter (D-PPHDF), and the Distributed Data Fusion Particle PHD Filter (DDF-PPHDF), are compared to the Distributed
Posterior Cram´er-Rao Lower Bound (DPCRLB).
algorithm provides an OSPA value or an estimate of the number of targets for i = 0. This is expected and due to ATB, which
initializes new particle clouds based on the measurements from the previous time step. Thus, target birth is delayed by one
time step and tracking can only be performed for i > 0. The same effect can be witnessed at i = 9 and i = 14, respectively,
which mark the time instants at which targets 2 and 3 enter the ROI. Here, the OSPA curves of all trackers exhibit a spike,
which is due to the fact that the newborn particles are not yet considered in the tracker and, hence, the number of estimated
targets is too low, as can be seen in Figure 2b).
Another sudden rise of all the OSPA curves can be observed in the time interval 20 ≤ i ≤ 24 with a valley at i = 22.
Looking at the estimated number of targets, we can attribute this phenomenon to the fact that only two of the three targets
are recognized by the tracking algorithms. Since the target trajectories are deterministic, we know that in the given time
interval targets 2 and 3 cross paths. Due to the inability of the clustering algorithm to separate strongly overlapping sets of
measurements, the two targets merge into one as long as they are close to each other. When the two targets occupy almost
exactly the same position, i.e., at i = 22, the OSPA metric decreases due to the decrease in measurement variance. As the
targets drift apart, the variance and with it the OSPA metric increases up to the point where the two targets can be recognized
as separate again and the corresponding penalty is switched off.
Looking at the overall picture in Figure 2a), which shows the case of σ2
r = 0.1, it is evident that the centralized MS-PPHDF
and the distributed D-PPHDF achieve approximately the same performance with OSPA values closely approaching the DPCRLB
when the number of targets stays constant. Furthermore, both algorithms deliver very accurate estimates of the number of targets,
given they are separable by clustering, as can be seen in Figure 2b). The DDF-PPHDF, however, continuously exhibits a worse
performance than the D-PPHDF, both in terms of the OSPA metric as well as the estimated number of targets. This is where
the additional communication in the proposed D-PPHDF shows its strength in reducing uncertainty due to measurement noise
and clutter. Apart from achieving worse tracking results, the DDF-PPHDF also has more difficulty in separating targets 1 and
SUBMITTED FOR PUBLICATION
21
2 when they cross paths, resulting in an earlier rise and a later fall of the OSPA metric, compared to our approach.
In the case of σ2
r = 0.3, the overall performance of the different tracking algorithms is very similar to the case of σ2
r = 0.1.
In order to make a statement on how the different tracking algorithms compare, let us neglect the penalty due to an erroneous
estimate of the number of targets and take a look at Figures 3a) and 3b), which are zoomed-in versions of Figures 2a) and
2c), respectively.
In Figures 3a) and 3b) the DPCRLB is given as a benchmark for tracking performance. One can observe that its value
is always smaller or equal to the respective measurement variance. As stated before, the centralized MS-PPHDF and the
distributed D-PPHDF exhibit very similar performance and deliver better tracking results than the DDF-PPHDF. While the
MS-PPHDF achieves lower OSPA values than the D-PPHDF when the number of targets stays constant, i.e., for 3 ≤ i ≤ 8
and 24 ≤ i ≤ 30, the D-PPHDF performs better directly after a new target appears, i.e., for 1 ≤ i ≤ 2, 10 ≤ i ≤ 13, and
15 ≤ i ≤ 18. This is likely due to the fact that the two-step communication scheme employed in the D-PPHDF is able to
reduce the impact of measurement noise and clutter faster than the centralized MS-PPHDF can.
Looking at the case of σ2
r = 0.3 in Figure 3b), we observe that the higher measurement noise affects the performance
of all algorithms, resulting in higher OSPA curves. While the OSPA curves of the MS-PPHDF and the DDF-PPHDF are
proportionally shifted upward by approximately the same value, i.e., they are equally impacted by the higher noise level, the
D-PPHDF seems to be slightly more affected by the change. However it still outperforms the DDF-PPHDF at all time instants.
All in all, the proposed D-PPHDF yields better performance than the existing DDF-PPHDF in estimating the number of
targets and tracking them, irrespective of the amount of measurement noise. In addition, it is also a bit faster in delivering
correct state estimates of new targets than the centralized MS-PPHDF and performs only slightly worse once the number of
targets stays constant. In our future work, we will look at ways to further improve the performance of the MS-PPHDF and
the D-PPHDF in order to approach the DPCRLB even more closely.
D. Simulation II: Results
In the second simulation, we evaluate the performance of the MS-PPHDF, the D-PPHDF and the DDF-PPHDF under a
higher clutter rate of 0.3. The remaining parameters are chosen as in the previous simulation. The simulation results are shown
in Figure 4. The top part considers zero-mean Gaussian measurement noise with a per-component variance of σ2
r = 0.1 while
the bottom part shows the results for σ2
r = 0.3.
While the higher clutter rate causes an increase in the OSPA value of all algorithms, the MS-PPHDF is still able to correctly
estimate the number of targets (except for the crossing period 20 ≤ i ≤ 24) in both cases. When taking the next lower integer
of the estimate, the D-PPHDF also yields acceptable results for σ2
r = 0.1. For σ2
r = 0.3 the number of targets is overestimated
by 1 for 1 ≤ i ≤ 15, causing a stronger degradation of the scaled and squared OPSA value in this interval.
Apparently, the DDF-PPHDF is not able to cope with a clutter rate of 0.3 as the number of targets is largely overestimated.
Hence, no accurate target tracking is possible.
Ntgt ·(cid:16) ¯d
2 (cid:17)2
SUBMITTED FOR PUBLICATION
22
PSfrag replacements
35
MS-PPHDF
D-PPHDF
DDF-PPHDF
18
16
14
12
t
s
i
d
,
i
B
d
n
a
10
0.2
DPCRLB
0.4
true number
0.6
and Bi,dist
0.8
2
(cid:17)
)
2
2
(
¯d
8
(cid:16)
6
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
·
t
g
t
N
t
s
e
N
1
1.2
1.4
1.6
1.8
35
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
5
10
20
Time steps i (seconds)
15
25
30
MS-PPHDF
D-PPHDF
DDF-PPHDF
4
2
0
0
18
16
14
12
10
8
6
4
2
0
0
5
20
10
Time steps i (seconds)
15
25
30
Nest
PSfrag replacements
35
0.2
DPCRLB
0.4
0.6
and Bi,dist
0.8
1.2
1.4
1.6
1.8
35
0.2
0.4
0.6
0.8
1.2
1.4
1.6
1.8
7
6
5
t
s
e
N
4
3
2
1
0
7
6
5
t
s
e
N
4
3
2
1
0
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
5
10
20
Time steps i (seconds)
15
25
30
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
5
10
20
Time steps i (seconds)
15
25
30
Fig. 4: Simulation II: Results for Gaussian noise and clutter rate λFA = 0.3. The left part of the figure shows the squared and
scaled Optimal Subpattern Assignment (OSPA) metric for each algorithm compared to the Distributed Posterior Cram´er-Rao
Lower Bound (DPCRLB), while the right part compares the estimated to the true number of targets.
E. Simulation III: Results
In the third simulation, we evaluate the robustness of the MS-PPHDF, the D-PPHDF and the DDF-PPHDF in the face of
ε-contaminated noise and different clutter rates. We consider a per-component variance of the measurement noise of σ2
r = 0.1
and 0.3, an average number of clutter of λFA = 0.1 and 0.3, as well as a contamination of 10% and 30%. The remaining
parameters are chosen as before. The simulation results for clutter rates λFA = 0.1 and 0.3 are given in Figures ?? and 6,
respectively. The top half of each figure considers σ2
r = 0.1 while the bottom half pertains to σ2
r = 0.3. Rows 1 and 3 deal
with a noise contamination of 10 %, rows 2 and 4 show the results for 30 %.
Let us look at the case of λFA = 0.1 first. We observe that the centralized MS-PPHDF is still the best performing algorithm,
being largely unaffected by higher noise variance and outliers. The D-PPHDF is a close second, being primarily affected by the
higher clutter rate and the higher noise variance. It shows only a slight additional performance degradation when increasing the
noise contamination to 30 %. Hence, it can be said that both algorithms are robust against outliers and can handle a fraction
of at least 10 % in the given scenario. The MS-PPHDF can also cope with λFA = 0.3 while the target position estimates of
the D-PPHDF might be too imprecise in this case, depending on the problem at hand. The DDF-PPHDF, in contrast, is more
severely affected by outliers. Both the OSPA value and the estimated number of targets increase with the introduction of noise
SUBMITTED FOR PUBLICATION
23
contamination. When the number of targets remains constant and no target crossing takes place, i.e. for i < 9 and i > 24, the
number of targets is only slightly overestimated. However, when targets two and three enter the scene, i.e. for 10 ≤ i < 20,
the estimate is inaccurate, which imposes a penalty on the scaled and squared OPSA metric. Hence, the DDF-PPHDF is not
a robust algorithm for the considered tracking scenario.
In the case of λFA = 0.3, the DDF-PPHDF, again, breaks down completely. The MS-PPHDF, however, is still able to
give accurate results with only slight deviations from the true number of targets and a small OSPA value. Unfortunately, the
combination of noise contamination and more clutter is too much for the D-PPHDF to handle. It overestimates the number of
targets by one to two, causing the OSPA value to rise as well.
In summary, the proposed MS-PPHDF and D-PPHDF are -- to a certain extent -- robust against outliers of the ε-contamination
kind. This property is due to the employed two-way communication scheme, which vets measurements as well as intermediate
target position estimates against the entire network or the neighborhood of each node. The alternative DDF-PPHDF, however,
breaks down in the face of outliers.
In [51] -- [53], we successfully proposed to use robust estimators to robustify sequential detectors for distributed sensor
networks. We applied the same concept to the distributed D-PPHDF. However, no further performance improvement could be
gained here since the twofold neighborhood averaging already exhausted the power of neighborhood communication.
VI. CONCLUSION
In this work, we developed a distributed as well as a centralized PPHDF for MTT in sensor networks. We, furthermore,
came up with a distributed version of the PCRLB that served as a benchmark in the performance evaluation. Our simulation
results showed that the distributed D-PPHDF is faster in correctly tracking new targets than the centralized MS-PPHDF and
performs only slightly worse when the number of targets stays constant. In addition, it delivers accurate tracking results as long
as the targets are far enough apart so that their corresponding measurement clouds are separable. Our approach outperforms
the existing DDF-PPHDF at the cost of additional communication between sensor nodes. Moreover, the proposed trackers
are inherently robust against outliers and the centralized MS-PPHDF is even able to handle higher clutter rates. The existing
DDF-PPHDF, in contrast, is neither robust nor able to cope with more than 10 % clutter.
The authors would like to thank Dr. Paolo Braca from the NATO Science & Technology Organization, Centre for Maritime
Research and Experimentation in La Spezia, Italy, for his valuable comments.
ACKNOWLEDGMENT
REFERENCES
[1] S. Challa, R. Evans, M. Morelande, and D. Musicki, Fundamentals of object tracking. Cambridge University Press, 2011.
[2] S. Maresca, P. Braca, J. Horstmann, and R. Grasso, "Maritime surveillance using multiple high-frequency surface-wave radars," Geoscience and Remote
Sensing, IEEE Transactions on, vol. 52, no. 8, pp. 5056 -- 5071, Aug 2014.
[3] J. Rambach, M. F. Huber, M. R. Balthasar, and A. M. Zoubir, "Collaborative multi-camera face recognition and tracking," in Proceedings of the 12th
IEEE International Conference on Advanced Video- and Signal-based Surveillance (AVSS2015), August 2015.
SUBMITTED FOR PUBLICATION
24
[4] R. Olfati-Saber, J. A. Fax, and R. M. Murray, "Consensus and cooperation in networked multi-agent systems," Proceedings of the IEEE, vol. 95, no. 1,
pp. 215 -- 233, Jan 2007.
[5] O. Hlinka, F. Hlawatsch, and P. M. Djuric, "Distributed particle filtering in agent networks: A survey, classification, and comparison," IEEE Signal
Processing Magazine, vol. 30, no. 1, pp. 61 -- 81, 2013.
[6] F. S. Cattivelli, C. G. Lopes, and A. H. Sayed, "Diffusion strategies for distributed Kalman filtering: Formulation and performance analysis," Proceedings
of the 1st IAPR Workshop on Cognitive Information Processing (CIP), pp. 36 -- 41, 2008.
[7] M. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, "A tutorial on particle filters for online nonlinear/non-Gaussian Bayesian tracking," IEEE
Transactions on Signal Processing, vol. 50, no. 2, pp. 174 -- 188, 2002.
[8] Y. Bar-Shalom, P. K. Willett, and X. Tian, Tracking and Data Fusion: A Handbook of Algorithms. Storrs, CT: YBS Publishing, 2011.
[9] D. Reid, "An algorithm for tracking multiple targets," IEEE Transactions on Automatic Control, vol. 24, no. 6, pp. 843 -- 854, 1979.
[10] S. Oh, L. Schenato, P. Chen, and S. Sastry, "Tracking and coordination of multiple agents using sensor networks: System design, algorithms and
experiments," Proceedings of the IEEE, vol. 95, no. 1, pp. 234 -- 254, 2007.
[11] D. E. Clark, "Multiple target tracking with the probability hypothesis density filter," Ph.D. dissertation, Department of Electrical, Electronic and Computer
Engineering, Heriot-Watt University, UK, October 2006.
[12] R. Mahler, "Multitarget Bayes filtering via first-order multitarget moments," IEEE Transactions on Aerospace and Electronic Systems, vol. 39, no. 4,
pp. 1152 -- 1178, 2003.
[13] X. Wang, G. Xing, Y. Zhang, C. Lu, R. Pless, and C. Gill, "Integrated coverage and connectivity configuration in wireless sensor networks," in Proceedings
of the 1st international conference on Embedded networked sensor systems. ACM, 2003, pp. 28 -- 39.
[14] M. R. Balthasar, S. Al-Sayed, S. Leier, and A. M. Zoubir, "Optimal area coverage in autonomous sensor networks," in Proceedings of the 2nd International
Conference and Exhibition on Underwater Acoustics (UA2014), June 2014.
[15] D. Schuhmacher, B.-T. Vo, and B.-N. Vo, "A consistent metric for performance evaluation of multi-object filters," Signal Processing, IEEE Transactions
on, vol. 56, no. 8, pp. 3447 -- 3457, Aug 2008.
[16] H. L. Van Trees, Detection, estimation, and modulation theory.
John Wiley & Sons, 2004.
[17] C. Hue, J.-P. Le Cadre, and P. Perez, "Performance analysis of two sequential Monte Carlo methods and posterior Cramer-Rao bounds for multi-target
tracking," in Proceedings of the Fifth International Conference on Information Fusion, vol. 1, 2002, pp. 464 -- 473.
[18] -- -- , "Sequential Monte Carlo methods for multiple target tracking and data fusion," IEEE Transactions on Signal Processing, vol. 50, no. 2, pp.
309 -- 325, 2002.
[19] M. Uney, D. Clark, and S. Julier, "Distributed fusion of PHD filters via exponential mixture densities," IEEE Journal of Selected Topics in Signal
Processing, vol. 7, no. 3, pp. 521 -- 531, June 2013.
[20] M. Uney, S. Julier, D. Clark, and B. Ristic, "Monte Carlo realisation of a distributed multi-object fusion algorithm," in Sensor Signal Processing for
Defence (SSPD 2010), Sept 2010, pp. 1 -- 5.
[21] G. Battistelli, L. Chisci, C. Fantacci, A. Farina, and A. Graziano, "Consensus CPHD filter for distributed multitarget tracking," Selected Topics in Signal
Processing, IEEE Journal of, vol. 7, no. 3, pp. 508 -- 520, 2013.
[22] R. Mahler, "The multisensor PHD filter: I. General solution via multitarget calculus," in Proc. SPIE Signal Processing, Sensor Fusion, and Target
Recognition XVIII, vol. 7336, May 2009.
[23] -- -- , "Approximate multisensor CPHD and PHD filters," in Proceesings of the 13th Conference on Information Fusion (FUSION), July 2010, pp. 1 -- 8.
[24] F. Gustafsson, F. Gunnarsson, N. Bergman, U. Forssell, J. Jansson, R. Karlsson, and P.-J. Nordlund, "Particle filters for positioning, navigation, and
tracking," IEEE Transactions on Signal Processing, vol. 50, no. 2, pp. 425 -- 437, 2002.
[25] B. T. Vo, "Random finite sets in multi-object filtering," Ph.D. dissertation, University of Western Australia, 2008.
[26] R. P. Mahler, Statistical multisource-multitarget information fusion. Artech House, Inc., 2007.
[27] -- -- , Advances in statistical multisource-multitarget information fusion. Artech House, 2014.
[28] B.-N. Vo, S. Singh, and A. Doucet, "Sequential Monte Carlo implementation of the PHD filter for multi-target tracking," in Proceedings of the 6th
International Conference on Information Fusion, 2003, pp. 792 -- 799.
[29] -- -- , "Sequential Monte Carlo methods for multitarget filtering with random finite sets," IEEE Transactions on Aerospace and Electronic Systems,
vol. 41, no. 4, pp. 1224 -- 1245, 2005.
[30] R. Mahler, "Multitarget moments and their application to multitarget tracking," Technical Report, DTIC Document, 2001.
SUBMITTED FOR PUBLICATION
25
[31] -- -- , "'Statistics 102' for multisource-multitarget detection and tracking," IEEE Journal of Selected Topics in Signal Processing, vol. 7, no. 3, pp.
376 -- 389, 2013.
[32] B. Ristic, D. Clark, B.-N. Vo, and B.-T. Vo, "Adaptive target birth intensity for PHD and CPHD filters," IEEE Transactions on Aerospace and Electronic
Systems, vol. 48, no. 2, pp. 1656 -- 1668, April 2012.
[33] B. Ristic, D. Clark, and B.-N. Vo, "Improved SMC implementation of the PHD filter," in Proceesings of the 13th Conference on Information Fusion
(FUSION), July 2010, pp. 1 -- 8.
[34] S. Hong, L. Wang, Z.-G. Shi, and K. S. Chen, "Simplified particle PHD filter for multiple-target tracking: Algorithm and architecture," Progress In
Electromagnetics Research, vol. 120, pp. 481 -- 498, 2011.
[35] A. H. Sayed, "Diffusion adaptation over networks," in E-Reference Signal Processing, R. Chellappa and S. Theodoridis, Eds. New York: Elsevier,
2013.
[36] B. Everitt, S. Landau, M. Leese, and D. Stahl, Cluster Analysis.
John Wiley & Sons, 2011.
[37] N. J. Gordon, D. J. Salmond, and A. F. M. Smith, "Novel approach to nonlinear/non-Gaussian Bayesian state estimation," in IEE Proceedings F -- Radar
and Signal Processing, vol. 140, no. 2, 1993, pp. 107 -- 113.
[38] F. Gustafsson, "Particle filter theory and practice with positioning applications," IEEE Aerospace and Electronic Systems Magazine, vol. 25, no. 7, pp.
53 -- 82, 2010.
[39] F. Murtagh, "A survey of recent advances in hierarchical clustering algorithms," The Computer Journal, vol. 26, no. 4, pp. 354 -- 359, 1983.
[40] B.-N. Vo and W.-K. Ma, "The Gaussian mixture probability hypothesis density filter," IEEE Transactions on signal processing, vol. 54, no. 11, pp.
4091 -- 4104, 2006.
[41] K. Granstrom and U. Orguner, "A PHD filter for tracking multiple extended targets using random matrices," IEEE Transactions on Signal Processing,
vol. 60, no. 11, pp. 5657 -- 5671, Nov 2012.
[42] K. Granstrom, A. Natale, P. Braca, G. Ludeno, and F. Serafino, "PHD extended target tracking using an incoherent X-band radar: Preliminary real-world
experimental results," in 17th International Conference on Information Fusion (FUSION), July 2014, pp. 1 -- 8.
[43] J. MacQueen et al., "Some methods for classification and analysis of multivariate observations," in Proceedings of the 5th Berkeley Symposium on
Mathematical Statistics and Probability, vol. 1, no. 281-297. California, USA, 1967, p. 14.
[44] M. Inaba, N. Katoh, and H. Imai, "Applications of weighted voronoi diagrams and randomization to variance-based k-clustering: (extended abstract),"
in Proceedings of the 10th Annual Symposium on Computational Geometry, ser. SCG '94. New York, NY, USA: ACM, 1994, pp. 332 -- 339. [Online].
Available: http://doi.acm.org/10.1145/177424.178042
[45] P. Braca, S. Marano, V. Matta, and P. Willett, "Asymptotic efficiency of the PHD in multitarget/multisensor estimation," IEEE Journal of Selected Topics
in Signal Processing, vol. 7, no. 3, pp. 553 -- 564, June 2013.
[46] S. M. Kay, Fundamentals of statistical signal processing: Estimation theory. Prentice-Hall, Inc., 1993.
[47] P. Braca, S. Marano, V. Matta, and P. Willett, "Multitarget-multisensor ML and PHD: Some asymptotics," in Proceedings of the 15th International
Conference on Information Fusion (FUSION), 2012, pp. 2347 -- 2353.
[48] P. Tichavsky, C. Muravchik, and A. Nehorai, "Posterior Cramer-Rao bounds for discrete-time nonlinear filtering," IEEE Transactions on Signal Processing,
vol. 46, no. 5, pp. 1386 -- 1396, 1998.
[49] C. Hue, J.-P. Le Cadre, and P. P´erez, "Performance analysis of two sequential Monte Carlo methods and posterior Cram´er-Rao bounds for multi-target
tracking," IRISA, Tech. Rep., 2002.
[50] A. Bessell, B. Ristic, A. Farina, X. Wang, and M. Arulampalam, "Error performance bounds for tracking a manoeuvring target," in Proceedings of the
6th International Conference of Information Fusion (FUSION), vol. 2, 2003, pp. 903 -- 910.
[51] W. Hou, M. R. Leonard, and A. M. Zoubir, "Robust distributed sequential detection via robust estimation," Proceedings of the 25th European Signal
Processing Conference (EUSIPCO), Aug 2017.
[52] M. R. Leonard and A. M. Zoubir, "Robust sequential detection in distributed sensor networks," IEEE Transactions on Signal Processing, Feb 2018,
submitted. [Online]. Available: https://arxiv.org/abs/1802.00263
[53] M. R. Leonard, M. Stiefel, M. Fauss, and A. M. Zoubir, "Robust sequential testing of multiple hypotheses in distributed sensor networks," in Proceedings
of the 43rd IEEE International Conference on Accoustics, Speech and Signal Processing (ICASSP), April 2018.
SUBMITTED FOR PUBLICATION
PSfrag replacements
PSfrag replacements
26
MS-PPHDF
D-PPHDF
DDF-PPHDF
6
5
4
t
s
i
d
,
i
B
d
n
a
2
(cid:17)
3
35
DPCRLB
0.2
true number
0.4
and Bi,dist
σ2
r = 0.1
0.6
ε = 0.1
0.8
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
)
2
2
(
¯d
(cid:16)
2
·
t
g
t
N
1.2
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
35
DPCRLB
0.2
0.4
and Bi,dist
0.6
0.8
1.2
1.4
PSfrag replacements
1.6
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
4
3.5
3
t
s
e
N
2.5
2
1.5
1.4
PSfrag replacements
1.6
1.8
35
DPCRLB
0.2
true number
0.4
and Bi,dist
0.6
0.8
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
σ2
r = 0.1
ε = 0.3
1.2
1.4
PSfrag replacements
1.6
1.8
1
0
0
7
6
5
4
t
s
i
d
,
i
B
d
n
a
2
(cid:17)
)
2
2
(
¯d
3
(cid:16)
·
t
g
t
N
Nest
35
DPCRLB
0.2
true number
0.4
and Bi,dist
0.6
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
0.8
1.2
1.4
t
s
i
d
,
i
B
d
n
a
σ2
r = 0.3
ε = 0.1
1.6
1.8
2
3
(cid:17)
)
2
2
(
¯d
2.5
(cid:16)
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
·
t
g
t
N
Nest
35
and Bi,dist
0.2
Nest
0.4
0.6
0.8
1.2
1.4
1.6
1.8
t
s
i
d
,
i
B
d
n
a
2
1.5
1
0.5
0
8
7
6
5
2
1
0
0
5.5
5
4.5
4
3.5
Ntgt ·(cid:16) ¯d
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
2 (cid:17)2
(2)
σ2
r = 0.3
ε = 0.3
and Bi,dist
Nest
2
(cid:17)
)
2
2
(
¯d
4
(cid:16)
Nest
·
t
g
t
N
and Bi,dist
Nest
3
2
1
0
5
5
10
20
Time steps i (seconds)
15
25
30
1.8
1
0
5
10
20
Time steps i (seconds)
15
25
30
MS-PPHDF
D-PPHDF
DDF-PPHDF
5
MS-PPHDF
D-PPHDF
DDF-PPHDF
5
MS-PPHDF
D-PPHDF
DDF-PPHDF
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
35
DPCRLB
0.2
0.4
and Bi,dist
0.6
0.8
1.2
1.4
PSfrag replacements
1.6
1.8
Nest
35
and Bi,dist
DPCRLB
30
0.2
0.4
and Bi,dist
0.6
0.8
1.2
1.4
1.6
1.8
Nest
35
and Bi,dist
0.2
Nest
0.4
and Bi,dist
0.6
30
0.8
1.2
1.4
1.6
1.8
Nest
and Bi,dist
Nest
and Bi,dist
Nest
and Bi,dist
30
10
20
Time steps i (seconds)
15
Ntgt ·(cid:16) ¯d
Ntgt ·(cid:16) ¯d
25
(2)
2 (cid:17)2
2 (cid:17)2
(2)
Ntgt ·(cid:16) ¯d
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
2 (cid:17)2
(2)
25
10
20
Time steps i (seconds)
15
(2)
(2)
Ntgt ·(cid:16) ¯d
Ntgt ·(cid:16) ¯d
Ntgt ·(cid:16) ¯d
2 (cid:17)2
2 (cid:17)2
2 (cid:17)2
(2)
25
10
20
Time steps i (seconds)
15
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
4.5
4
3.5
3
2.5
2
1.5
t
s
e
N
1
0
5
20
10
Time steps i (seconds)
15
25
30
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
4
3.5
3
t
s
e
N
2.5
2
1.5
1
0
5
20
10
Time steps i (seconds)
15
25
30
5
4.5
4
3.5
t
s
e
N
3
2.5
2
1.5
1
0
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
5
20
10
Time steps i (seconds)
15
25
30
Fig. 5: Simulation III: Results for ε-contaminated noise and clutter rate λFA = 0.1. The two upper rows consider σ2
the lower ones show the results for σ2
r = 0.3. Rows 1 and 3 consider ε = 0.1, rows 2 and 4 show the results for ε = 0.3.
r = 0.1,
SUBMITTED FOR PUBLICATION
PSfrag replacements
25
20
15
t
s
i
d
,
i
B
d
n
a
2
(cid:17)
)
2
2
(
¯d
10
(cid:16)
35
DPCRLB
0.2
true number
0.4
and Bi,dist
σ2
r = 0.1
0.6
ε = 0.1
0.8
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
1
1.2
·
t
g
t
N
1.4
PSfrag replacements
1.6
1.8
2
PSfrag replacements
MS-PPHDF
D-PPHDF
DDF-PPHDF
35
DPCRLB
0.2
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
0.4
and Bi,dist
0.6
t
s
e
N
0.8
1.2
1.4
PSfrag replacements
1.6
27
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
8
7
6
5
4
3
2
5
10
20
Time steps i (seconds)
15
25
30
1.8
1
0
5
10
20
Time steps i (seconds)
15
25
30
5
0
0
30
25
20
5
0
0
30
25
20
35
DPCRLB
0.2
true number
0.4
and Bi,dist
0.6
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
σ2
r = 0.1
ε = 0.3
1.4
PSfrag replacements
1.6
t
s
i
d
,
i
B
d
n
a
2
(cid:17)
15
(cid:16)
10
)
2
2
(
¯d
·
t
g
t
N
0.8
1
1.2
1.8
2
Nest
35
DPCRLB
0.2
true number
0.4
and Bi,dist
0.6
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
0.8
1
1.2
1.4
t
s
i
d
,
i
B
d
n
a
σ2
r = 0.3
ε = 0.1
1.6
1.8
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
2
Nest
and Bi,dist
0.2
Nest
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Nest
σ2
r = 0.3
ε = 0.3
and Bi,dist
2
15
(cid:17)
)
2
2
(
¯d
(cid:16)
10
·
t
g
t
N
t
s
i
d
,
i
B
d
n
a
5
0
0
35
30
25
20
2
(cid:17)
)
2
2
(
¯d
15
(cid:16)
Ntgt ·(cid:16) ¯d
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
2 (cid:17)2
(2)
Nest
·
t
g
t
N
10
and Bi,dist
Nest
5
0
0
5
5
5
(2)
Ntgt ·(cid:16) ¯d
MS-PPHDF
D-PPHDF
DDF-PPHDF
2 (cid:17)2
35
DPCRLB
0.2
0.4
and Bi,dist
0.6
0.8
1.2
1.4
PSfrag replacements
1.6
1.8
Nest
35
and Bi,dist
DPCRLB
30
0.2
0.4
and Bi,dist
0.6
Ntgt ·(cid:16) ¯d
Ntgt ·(cid:16) ¯d
25
(2)
2 (cid:17)2
2 (cid:17)2
(2)
MS-PPHDF
D-PPHDF
DDF-PPHDF
10
20
Time steps i (seconds)
15
Ntgt ·(cid:16) ¯d
Ntgt ·(cid:16) ¯d
(2)
2 (cid:17)2
2 (cid:17)2
(2)
25
10
20
Time steps i (seconds)
15
Nest
35
and Bi,dist
0.2
Nest
0.4
and Bi,dist
0.6
30
MS-PPHDF
D-PPHDF
DDF-PPHDF
(2)
(2)
Ntgt ·(cid:16) ¯d
Ntgt ·(cid:16) ¯d
Ntgt ·(cid:16) ¯d
2 (cid:17)2
2 (cid:17)2
2 (cid:17)2
(2)
25
20
10
Time steps i (seconds)
15
Nest
and Bi,dist
Nest
and Bi,dist
Nest
and Bi,dist
30
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
5
20
10
Time steps i (seconds)
15
25
30
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
5
20
10
Time steps i (seconds)
15
25
30
MS-PPHDF
D-PPHDF
DDF-PPHDF
true number
9
8
7
6
t
s
e
N
5
4
3
2
1
0
9
8
7
6
t
s
e
N
5
4
3
2
1
0
11
10
9
8
7
6
5
4
3
2
t
s
e
N
1
0
5
20
10
Time steps i (seconds)
15
25
30
0.8
1.2
1.4
1.6
1.8
0.8
1.2
1.4
1.6
1.8
Fig. 6: Simulation III: Results for ε-contaminated noise and clutter rate λFA = 0.3. The two upper rows consider σ2
the lower ones show the results for σ2
r = 0.3. Rows 1 and 3 consider ε = 0.1, rows 2 and 4 show the results for ε = 0.3.
r = 0.1,
|
1710.03189 | 2 | 1710 | 2017-11-18T18:37:59 | Towards Agent-Based Model Specification in Smart Grid: A Cognitive Agent-based Computing Approach | [
"cs.MA",
"cs.AI",
"physics.soc-ph"
] | A smart grid can be considered as a complex network where each node represents a generation unit or a consumer. Whereas links can be used to represent transmission lines. One way to study complex systems is by using the agent-based modeling (ABM) paradigm. An ABM is a way of representing a complex system of autonomous agents interacting with each other. Previously, a number of studies have been presented in the smart grid domain making use of the ABM paradigm. However, to the best of our knowledge, none of these studies have focused on the specification aspect of ABM. An ABM specification is important not only for understanding but also for replication of the model. In this study, we focus on development as well as specification of ABM for smart grid. We propose an ABM by using a combination of agent-based and complex network-based approaches. For ABM specification, we use ODD and DREAM specification approaches. We analyze these two specification approaches qualitatively as well as quantitatively. Extensive experiments demonstrate that DREAM is a most useful approach as compared with ODD for modeling as well as for replication of models for smart grid. | cs.MA | cs |
Towards Agent-Based Model Specification of Smart Grid: A
Cognitive Agent-Based Computing Approach
Waseem Akrama, Muaz A. Niazi∗a, Laszlo Barna Iantovicsb
aComputer Science Department,
COMSATS Institute of Information Technology,
Islamabad, Pakistan
bCorresponding author
Abstract
A smart grid can be considered as a complex network where each node represents a generation
unit or a consumer. Whereas links can be used to represent transmission lines. One way to study
complex systems is by using the agent-based modeling (ABM) paradigm. An ABM is a way of
representing a complex system of autonomous agents interacting with each other. Previously, a
number of studies have been presented in the smart grid domain making use of the ABM paradigm.
However, to the best of our knowledge, none of these studies have focused on the specification
aspect of ABM. An ABM specification is important not only for understanding but also for repli-
cation of the model. In this study, we focus on development as well as specification of ABM for
smart grid. We propose an ABM by using a combination of agent-based and complex network-
based approaches. For ABM specification, we use ODD and DREAM specification approaches.
We analyze these two specification approaches qualitatively as well as quantitatively. Extensive
experiments demonstrate that DREAM is a most useful approach as compared with ODD for
modeling as well as for replication of models for smart grid.
Keywords: Agent-based modeling, Complex networks, Smart Grid
1. Introduction
A smart grid focuses on the complex interactions between utility service providers and con-
sumers. It involves the non-linear dialogue of power and information data between utility service
providers and consumers [1]. The complex interactions in the form of repeated auction, fluctuating
supply and demand add complexity to the nature of a smart grid. Because of this complex nature,
it can be considered as a complex system.
The study and understanding of any complex system are associated with the modeling of the
system. Modeling complex system allows better understanding and analyzing the emergent be-
havior of each entity involved in the system [2]. Being a complex system, a smart grid can also
essentially be modeled in the form of either agent-based or complex network-based models [3, 4].
These models can well represent the smart grid in term of its various components, their behavior,
and communication among them for the energy distribution and management.
Preprint submitted to Elsevier
November 21, 2017
A particular way of modeling smart grid as a complex network is by including its various
components such as generating units, consumers, distributors and other components as nodes and
communication as edges. Chasing et al. have developed the complex network model for the US
power grid by considering the number of nodes as power sources and consumers, while the number
of edges as communication lines [5]. After developing different complex networks, we able to use
the mathematical tool for computing centrality measures and metrics on such networks. Then
these measurements will allow studying the global behavior of each component in large-scale
power system network.
Regarding complex systems, Batool and Niazi [6] have proposed a hybrid modeling approach
for the Internet of things (IoT) domain. In this work, a combination of agent-based and complex
network-based approaches are used for modeling IoT-based complex scenarios. This work demon-
strates the use of standard complex networks such as small-world, scale-free and random network.
However, this previous work is not extended and applied to the smart grid domain. We think this
previous modeling methodology is also useful for the smart grid-based complex scenarios.
In scientific literature, agent-based modeling (ABM) and the multi-agent system (MAS) are
effectively used in a smart grid domain. Some of these works have been discussed in the later
section of the paper (See Discussion section). However, these works lack in any ABM specification
approach for documenting ABM. An ABM specification is most important for understanding as
well as replicating ABM. So there is a need for an easily understandable methodology to describe
an ABM, specifically in the smart grid domain.
The purpose of this study is twofold. First, we propose an ABM for the smart grid using a
combination of agent-based and complex networks-based approaches. In our work, we will use
the previously developed standard complex networks such as small-world, scale-free and random
network. For validation, we also present a hybrid centrality-based routing algorithm. This will
allow the end to end delivery between consumers and utility service providers. Second, for ABM
specification, we follow two specification approaches, the one is ODD (short for Overview, De-
sign concept, and Details) [7] and the second is a DREAM (short for Descriptive Agent-based
modeling) [3]. Then we present a comparative analysis of both specification techniques.
Our main contributions can be listed as follows:
1. A proposed approach for modeling and simulation of smart grid using complex network and
agent-based modeling approaches.
2. A hybrid centrality-based routing algorithm.
3. The ODD specification approach used for ABM model of smart grid.
4. The DREAM specification approach used for ABM model of smart grid.
5. A comparative analysis of ODD and DREAM specification techniques.
The rest of the paper is structured as follows: Section 2 presents basic background and con-
cepts, in Section 3 a model development is presented, Section 4 is dedicated for results and dis-
cussions, the paper ends with conclusions formulated in Section 5.
2. Background
In this section, we present basic concept and understanding of cognitive agent-based comput-
ing approach, DREAM, and ODD specification approaches.
2
2.1. Cognitive agent-based computing approach
Niazi and Hussain in [8] presented a unified cognitive agent-based computing approach for
agent-based and complex network-based modeling. This approach involves the process of taking
any complex system from the real world and to convert it into a suitable simple model by using
specific modeling level such as exploratory or descriptive agent-based approach. The exploratory
approach involves the use of agents to explore the complex system, identifies which agent-based
model is feasible for the specific problem, then develops the proof concept and also explains what
kind of data is required for validation and verification of the model. The Descriptive agent-based
model level is the process of presenting ABM specification in the form of pseudo-code, a complex
network of the model and social network analysis. Other modeling levels involve complex network
and validation/ verification of the model.
2.2. Descriptive agent-based modeling
Descriptive agent-based modeling (DREAM) is a cognitive agent-based computing approach
developed by Niazi in [3]. DREAM offers an ABM specification technique which comprises of
developing a complex network of the ABM, pseudo-code specification model and social network
analysis of the network model. It offers a detailed description of ABM as well as visual based
analysis. It provides an easy translation of network model into pseudo-code followed by ABM
development. In figure 1, Dream methodology has been shown.
Figure 1: DREAM methodology for ABM specification adopted from [3]. We can see that it starts from a basic
understanding of the model, then followed by developing a complex network of the model. There are two steps after
network formation, one is to present pseudocode specification, this step allows for the actual code translation, second
after network formation is by applying social network analysis tool to compute centrality measures of the network.
The final step involves the results analysis.
2.3. Overview, Design concept and Details
Overview, Design concept and Details (ODD) is originally developed by Grimm et al. in [7]
and then an updated version is presented in [9]. It is a textually based specification technique for
3
documenting ABM. It provides a checklist which covers key features of the model. It comprises of
three main sections which are Overview, Design concept, and Details. These sections are further
divided into subsections. In figure 2, ODD specification methodology is presented. A detailed
description can be found in [9].
Figure 2: ODD methodology for ABM specification adopted from [9]. We can see that ODD is divided into three main
sections that are named: overview, design concept, and details. Then each section is further divided into subsections.
These sections cover key features of the model.
3. Model development
In this section, we present a case scenario of the smart grid, followed by ODD and DREAM
specification of the model.
Figure 3 shows our research methodology. First, we develop ABM for the smart grid, then fol-
lowed by two specification approaches ODD and DREAM. After this, we perform a comparative
analysis of these specification approaches. In the final step, we present results.
4
Figure 3: Our research method. We can see that the first step comprises of ABM development followed by the
specification of the ABM. Two specification methods are adopted (ODD and DREAM). After specification, the next
step is to compare and analyze both specifications approaches qualitatively as well as quantitatively. The final step
shows results of the comparative analysis.
3.1. Scenario of smart grid
To model smart grid, let us consider a large set of network, in which different consumers
and power generation units are connected with each other through different configuration. To
model different possibility of different configurations of the large-scale power system, we use
standard complex networks such as small-world [10], scale-free [11] and random network [12].
For validation, we apply routing techniques such as random walk and centrality-based routing.
The routing process involves the selection of path from the source toward the destination.
Routing strategy in complex networks can be categorized into two types, i.e. Local and Global
Routing. The local routing strategy needs local information about neighbor nodes. These are in-
cluding local static routing, local dynamic routing, and local pheromone routing [13]. On the other
side, global routing strategy needs global information like topological structure, characteristics of
each node and real-time information. These are including shortest path routing, efficient routing,
and global dynamic routing [14].
For large-scale complex networks, global routing remains problematic. It is difficult to have
the characteristic of each node and to have real time information. Another difficulty consists in
the increases of computational time. While on the other hand, local routing remains promising
for large scale real world complex networks. It offers less computational time as well as easy
implementation.
In a smart grid environment, two types of routing occur. The one is energy demand from
consumers side to generation unit, while the generation unit responses by providing energy to
consumers. The second is data and information routing about demand profile from consumers
side and energy cost from grid unit [15].
5
3.2. Model specification according to ODD
3.2.1. Overview
1. The purpose of the model:
To understand how a combination of agent-based and complex network-based modeling ap-
proaches can be used to simulate large-scale power system. Further, how routing techniques
can be used to validate the model.
2. The involved entities:
The model consists of three types of agents that have the names; consumers, producers, and
walkers that are represented by nodes in a complex network. The model allows that produc-
ers and consumers are generated randomly in a network. Producers generate power and can
transmit to the consumers through communication lines that we called links. Consumers
demand and finally use energy power. The environment is set as a complex network where
nodes represent producers and consumers and links represent transmission lines. State vari-
ables visited? and consumer? are used to mark once a node being visited and to check is
any available consumer? node in the neighbor list.
3. Routing purpose:
For routing purpose, the concept of walkers is deployed. Initially, the walkers are located at
the producer's nodes. They search for the neighbor nodes and move to one of the neighbor's
node. Once a node is visited, it is marked as visited?. The walkers also check for the
consumer's node. The simulation time is kept as continues. At each time step, plots are
generated in order to measure visited nodes and visiting consumers.
3.2.2. Design concept
1. Basic principles:
The basic hypothesis of our model is that a cognitive agent-based computing approach is
better for modeling and simulation of the large-scale power system. In our approach, we
used a combination of agent-based and complex network based modeling approaches. We
used previously developed complex network models such as small-world, scale-free and
random network to simulate a smart grid based environment.
2. Emergence:
The "emergence" feature shows information about "What kind of outputs of the model are
modeled". In other words, we can say that what are the expected results from the model. In
case of our approach, the routing techniques (random walk and centrality routing) are used
for transmission from producers towards consumers. The key results are the computation of
end to end delivery from producers towards consumers.
3. Adaptation:
Adaptive feature of the model shows decision making capability for the agents against the
changing environment. Decisions are taken by using well-defined constraints to adapt the
variation in the environment accordingly. There are two rules applied for making a decision.
When using random walk, the walker search for neighbor's node and select one of them,
while by using centrality routing, the walkers search for a neighbor node with maximum
value and select that one.
6
4. Objectives:
In an adaptive environment, individual agents also receive effects or rewards from the envi-
ronment for their adaptive behavior to achieve one's own objective. In our model, the main
objective is to measure how much time is taken while moving from one node to the other.
5. Sensing:
In the decision-making process among agents, there are some specific features related to
each agent which allow communicating neighbors to make their decision according to values
of those features. In our case scenario, the walkers use the sensing property, if a neighbor
node is already been visited then they avoid rerouting. They also sense for consumers, if
any visited node is a consumer, then they deliver packets or energy.
6. Interaction:
Producers and consumers can communicate with each other for power transmission.
7. Stochasticity:
The routing process is modeled as random.
8. Observation:
When the simulation is running, at each time step the following data are collected.
(a) Number of nodes.
(b) Number of producers.
(c) Number of consumer.
(d) Number of walkers.
(e) Number of nodes with visited.
(f) Number of consumers with visited.
3.2.3. Details
The details section of the ODD protocol covers features of the model about what is the initial
state of the model, what kind of data is used, and what types of parameters and parameter values
are set in the model.
1. Initialization:
The model is implemented using NetLogo an agent-based modeling tool. The model en-
vironment is initialized by calling "draw-network" method. This method is used to draw
any selected network. Then consumers and producers are generated randomly by specifying
their number. After this, the walkers are placed at the producer's location.
2. Input data:
The standard complex networks are generated and kept as external source files. These net-
work files are used as input for the model.
3. Submodels:
The model parameters and parameter values are given in table 1.
7
Table 1: Evaluation metrics: These parameters and parameter values are used for model simulation. The region shows
the simulation environment which is kept as 100 by 100. The number of nodes in the network is considered as 500,
800, 1000. The number of power sources and consumers are used as 5, 10, 50, 100, 150, 200. Three different standard
complex networks are used. For routing purpose, random walk and centrality-based routing algorithm are applied.
The Performance of the model is measured in term of average delivery rate (computation time from sources to the
consumers). A series of experiments were carried out during simulation model.
Parameter
Region
No. of Nodes
Power sources
Power consumer
Network
Routing
Performance measure Average delivery rate
No. of runs
Value
100X100
500, 800, 1000
5, 10, 50, 100
50, 100, 150, 200
small-world, scale-free, random network
random-walk, centrality-based routing
4(1, 10, 20, 30runs)
3.3. Model specification according to DREAM
In this section, we present our ABM documentation according to DREAM specification ap-
proach. We describe our model using Pseudo-code specification part of the DREAM approach.
3.3.1. Agent design
There are two types of agents which are used in the simulation model. Agents by node type,
in our simulation model, we used complex networks. These complex networks consist of nodes
which are connected through communication lines called links. These node agents represent pro-
ducers and consumers in the network. Agents by walker type, for routing purpose the Walker
concept is deployed. These walkers initially placed on producers node. They have the ability to
move around the network.
Algorithm 1 Breed Node: This node agent is used to represent producers and consumers in the
network
Internal Variables: < source?, target?, visited? >
1: source?: All nodes that represent sources(used for producers)
2: target?: All nodes that represent target (used for consumers)
3: visited?: Used to check the status of a Node agent
In the Node specification model, first, we described the Breed Node agent. The "Breed" is
a global keyword in NetLogo(Agent-based modeling toolkit) describing a set of similar-behavior
agents. As we noted first that nodes are used in the network to represent producers and consumers.
After this, we specified the internal variables for Node agent. There are three internal variables
used for Node agent. The source? this variable is used to represent producers or generating unit in
power system. The target? variable is used to represent consumers in a power system environment.
8
The last one visited? is used to check the status of the node agent where it is visited or not. Next,
we define a specification for Walker agent.
Algorithm 2 Breed Walker: This agent is used for walker that can move around the network
Internal Variables: < location, is- f inish?, location-list? >
1: location: Keep current location information of the walker
2: is−finish?: Check the finish goal
3: location−list: Keep all visited locations record
The breed Walker represents Walker agents. These walkers are deployed for routing purpose
that can move around the network. Next, we specified three internal variables for Walker agent.
The first one is location. The location variable is used to keep information about the current
location of a walker. The is-finish? is a Boolean variable that returns true if the finish condition is
met. The location-list is a list variable that used as a memory with a walker. This variable keeps
information of all visited locations. After describing agent specification model, next, we are going
to present global variables specification model.
3.3.2. Global
For setup simulation, the key variables are five input global variables.
Algorithm 3 Input Globals: < get-network-type, num-node, num-walker, num-source, num-
target >
Chooser:
Slider:
get−network−type: Used for selecting network type from the giving list
num−node: Used for specifying number of nodes in a network
num−walker: Used for specifying number of walker in a network
num−source: Used for creating number of producer nodes
num−target: Used for creating number of consumer nodes
Here, five input global variables are used. The get-network-type is input provided by Chooser
(GUI element of the NetLogo simulation toolkit). This is used for selecting network type from
the available list. The network list comprises of the small-world, scale-free and random network.
The other four input variables are provided by Slider (GUI element of NetLogo). The num-node is
used to specify the number of nodes in the network. The num-walker is used to specify the number
of Walker in the network. The num-source and num-target are used for specification of a number
of producer and consumer respectfully.
3.3.3. Setup procedure
Here, we present the main setup procedure for model initialization.
9
Algorithm 4 Procedure setup: Creating simulation environment
Input: All global parameters from the user interface
Output: Setup simulation environment
begin
1: clear screen
2: draw-selected-network
3: add-sources
4: add-targets
5: calculate-centrality
End
The setup procedure is the global simulation setup specification model. This is used to create
a simulation model. The input parameters are provided by a user interface. The procedure starts
with the calling of clear-all function. It will clear all the previous work. Next, all four individual
procedures are called. The first one is "draw-selected-network". Next, we describe these individual
procedures.
Algorithm 5 Procedure draw-selected-network: Creating the desired network
Input: network-file, list of available network
Output: Setup the selected network
begin
1: if network-type = "small-world" then
2:
3: else
4:
5:
6:
7:
End
elsenetwork-type = "random-network"
[draw-random-network]
[draw-small-world]
if network-type = "scale-free" then
[draw-scale-free]
"draw-selected-network" is a procedure which is used to setup the selected network from the
giving options (a list of networks). The procedure checks the input type and called the appropriate
function of it. Here three individual procedures are used. Next, we present the specification model
of these individual procedures which are called by draw-selected-network.
3.3.4. Small-world network
In this network topology, any individual is linked to any other individual by a maximum of six
edges. In social network terminology, this is called "six degrees of separation". It has fewer nodes
with more links. It is satisfied by: G(V, E), L ∝ log(N).
10
Algorithm 6 Procedure draw-small-world: Creating small-world network
Input: N
Output: Setup small-world network
create-nodes N;
for all Nodes do;
set all-wired? false;
while all-wired? ! = true do
wired-them [connect-nodes];
set all-wired=Do-calculation [clustering-coefficient];
Findclustering-coefficient;
If no Node leftthen stop;
end
create link with one node to other
calculate shortest path [set node i=1;
count distance of node i to every other node;
i++;
set shortest path=min distance]
set layout circle;
End
3.3.5. Scale-free network
In this network, the link of an individual with another individual is uneven. There are some
nodes which have dense connections, while some others have fewer connections. The dense con-
nections are also called hubs. These hubs have the tendency to join with other new nodes. This
network follows the power law of the degree distribution. The probability of joining new nodes
with the existing hub can be defined by: (cid:89)
In Eq.1, Ki represents degree of hub i.
(
Ki(cid:80)
j K j
ki) =
(1)
11
Algorithm 7 Procedure draw-scale-free: Creating scale-free network
Input: N
Output: Setup scale-free network
set shapes circles;
create node 1;
set node=nobody;
create node 1;
[ set node = new-node]
if old-node = nobody then
[create link with old-node];
while count nodes < total-nodes do
[add nodes];
set layout spring;
set layout circle;
End
3.3.6. Random network
In this network, each individual node is formed randomly; there is no specific structure to be
followed. This network can be formed by joining vertex with other arbitrary. Formally a random
network GR(N, P) is framed with edges associated with likelihood, p given that 0 < P < 1 . The
connectivity of nodes with other does not depend on the degree of nodes.
Algorithm 8 Procedure draw-random-network: Creating random network
Input: N
Output: Setup random network
create nodes N;
for all Nodes do
set shapes circles;
set location random location;
create-links with nodes with p;
check p > 0 and p < 1;
End
Next, we present all other procedures that describe different processes which are called by the
main setup procedure.
3.3.7. Setup source and target nodes procedures
"add-source" procedure is used to setup source nodes (producers) on the network. It takes
network type, nodes, links, and a number of the source from the user interface. Next create source
nodes of a specified number from user randomly on the network.
12
Algorithm 9 Procedure add-source: Setup sources on the network
Input: network-type, num-source
Output: Setup source nodes on the network
begin
random nodes[;
set source true;
End
"add-target" procedure is used to setup target nodes (consumers) on the network.
It takes
network type, nodes, links, and a number of the target from the user interface. Next create target
nodes of a specified number from user randomly on the network.
Algorithm 10 Procedure add-target: Setup target on the network
Input: network-type, num-target
Output: Setup target nodes on the network
begin
Select random nodes[;
set target true;
End
After setting source and target nodes on the network. Next, the main setup procedure calls
"calculate-centrality" procedure. Next, we present calculate-centrality procedure.
3.3.8. Centrality measure
The complex network provides centrality measure feature. The centrality measure is the pro-
cess to find out which vertex is the most central which is important. It is widely used for measuring
the relative importance of nodes within a network. It is a numerical number assigned to each node
for pair-wise comparison within a whole network. It has four types.
1. Degree of nodes:
It measures the total number of connections that a particular node has in a network. A
node with a higher degree has most importance as compared with those which have a lower
degree. If a node with a higher degree is removed, then it can disrupt the structure as well
as the flow of the network.
Algorithm 11 Procedure calculate-degree: Calculates degree centrality on the network
Input: selected network
Output: calculates degree centrality for all nodes
for all Nodes do
set degree count all connected links
set label degree
End
13
2. Closeness of nodes:
It is used to find out how much data from a particular node moves to every other node in a
network. To calculate closeness of a particular node i from all other nodes t. Mathematical
it can be written as:
(cid:88) 1
dist(i, t)
Ccloseness(s) =
(2)
Algorithm 12 Procedure calculate-closeness: Calculates closeness centrality on the network
Input: selected network
Output: calculates closeness centrality for all nodes
for all Nodes do
set closeness Ccloseness(i) =(cid:80) 1
dist(i,t)
set label closeness
End
3. Betweenness of nodes:
Betweenness centrality is the process of counting the number of times a specific vertex
comes in the shortest path between any two vertexes in a network. It has the capability to
observe the network transmission. Mathematically can be written as:
(cid:88) σst(i)
σst
Cbetweenness(i) =
(3)
Where σst(i) denotes the number of shortest paths between nodes s and t passing through
the node i. While σst is the total number of shortest paths exist between node s and t.
Algorithm 13 Procedure calculate-betweenness: Calculates closeness centrality on the network
Input: selected network
Output: calculates betweenness centrality for all nodes
for all Nodes do
set betweenness Cbetweenness(i) =(cid:80) σst(i)
σst
set label betweenness
End
4. Eigen-vector of nodes:
It measures the impact of a particular node in a network. It defines which node is connected
to the most important node in a network. It depends on neighbors in term of connections
that neighbors have with other nodes in a network.
14
Algorithm 14 Procedure calculate-eigenvector: Calculates eigenvector centrality on the network
Input: selected network
Output: calculates eigenvector centrality for all nodes
for all Nodes do
(cid:80)
t∈g Aad j−matrix ∗ Xt OR Ax = λx
set eigenvector Ceigenvector(i) = 1
set label eigenvector
λ
End
3.3.9. Procedure Go
To validate our model, we apply routing techniques on our developed model. There are two
routing approaches that we used in our work. The first one is random walk and the second one
is centrality-based routing. Next, we present the details and specification models of these two
routing approaches described as procedures.
1. Procedure random-walk:
In case of a random walk, the walkers are set initially on the source nodes. They search their
neighbors and select one of them randomly. This process goes repeatedly until all the target
nodes have been visited. Let us consider an undirected graph G(V, E) , a random walk is a
stochastic process that starts from a given vertex, then select one of its neighbors randomly
to visit next. It has no memory that keeps information of previous moves. It stops when the
termination condition meets.
Algorithm 15 Procedure random-walk: Used for routing on network
Input: All input parameters provided by user interface
Output: End to end delivery from sources to the destinations
Execution: Called repeatedly on simulation execution
begin
1: start from sources
2: while is − f inish (cid:44) true do
3:
4:
5:
End
search neighbors
select any of neighbor
move to the selected
2. Procedure centrality-rw:
The procedure "centrality-rw" is another approach which is used in our work for routing
purpose. The working of this technique is as follows. The Walker search for neighbor
nodes. They take a decision on the basis of centrality. They search for that neighbor node
which has maximum centrality value. Then move to the selected node. If they found no
node with maximum centrality value, then they select randomly a node from neighbors. The
working of centrality-base routing is shown in figure 4.
15
Figure 4: Flowchart for centrality-based routing algorithm. We can see that the algorithm starts with making a list
of neighbor nodes of the current location. In the next step, it checks the content of the list. If the list is empty, then
the simulation stops, in case no, the algorithm search for the node having maximum centrality value. In case, yes, it
selects that node, in case no, it selects any node from the neighbor list. The next step is to move to the selected node.
After this, it checks the finish goal, in case yes, the simulation stops and in case no the control goes to the first step.
Algorithm 16 Procedure centrality-rw: Used for routing on network
Input:All input parameters provided by user interface
Output: End to end delivery from sources to the destinations
Execution: Called repeatedly on simulation execution
begin
1: start from sources
2: while is- f inish (cid:44) true do
get-list-of-neighbors
3:
if is-list-empty=true then
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
End
Else
Search node with max-centrality in the list
if node with max-value exist= true then
die
stop
select node with max-value
Else
select any of node in the list
move to the selected
16
3. Generate plot:
"do-plot" procedure is used to plot on the execution of the simulation of the model. Here,
this procedure plots two types of information. The first, at each time step, it counts the
number of nodes that have been visited. The second, it counts the number of consumers that
are visited. Next, we present model specification and detail for the experiment in our model.
Algorithm 17 Procedure generate-graph: Generate graphs of the current simulation state
Input: No input required
Output: Generate plot on simulation execution
Execution: Called by Go procedure
begin
Plot the number of nodes that are visited
Plot the number of targets visited
End
3.3.10. Performed experiment
Two types of experiments are performed in our model simulation. The first one is used for
random walk algorithm; the second is used for centrality-based routing algorithm.
Algorithm 18 Experiment: Demonstrates the effect of random walk and centrality routing on
different networks
Input: < num − nodes, num − source, num − target, num − walker, network − type >
Setup Procedure: < setup >
Go Procedure: < random − walk, centrality − rw >
Repetition: 10
Input:
num-nodes: 500
num-source: 5, 10, 50, 100
num-targets: 50, 100, 150, 200
num-walker: 5, 10, 50, 100
networ-type: small-world, scale-free, random-network
Reporter:
measure-nodes: count number of nodes that have been visited
measure-targets: count number of targets that have been visited
Stop condition:
If all targets are visited, Stop
3.4. Centrality-based routing algorithm: Time complexity analysis
In this section, we present the centrality-based routing algorithm time complexity analysis.
The time complexity of our proposed centrality-based routing is a linear function of n that is O(n).
17
In algorithm analysis, we analyze the cost and number of times that each step takes for execution.
All steps take a constant time, except step 2 and 8 which takes n time for execution. In step 2, a
while loop executes for n times. In step 8, algorithm searches in the list item of nodes and then
selects the node with the largest value, so it takes n time. The total running time is the sum of
running time and cost of each step in the algorithm.
get-list-of-neighbors
if is-list-empty=true then
Procedure centrality-rw: Used for routing on network
Input:All input parameters provided by user interface
Output: End to end delivery from sources to the destinations
Execution: Called repeatedly on simulation execution
begin
1: start from sources
2: while is − f inish (cid:44) true do
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
End
Else
Search node with max-centrality in the list
if node with max-value exist= true then
die
stop
select node with max-value
Else
select any of node in the list
move to the selected
Cost
C1
C2
C3
C4
C5
C6
C7
C8
C9
C10
C11
time
1
n
1
1
1
1
n
1
1
1
1
T(n) = C1(1) +C2(n) +C3(1) +C4(1) +C5(1) +C6(1) +C7(n) +C8(1) +C9(1) +C10(1) +C11(1) (4)
c = (C1 + C3 + C4 + C5 + C6 + C8 + C9 + C10 + C11)(1)
C2(n) + C7(n) = (C2 + C7)n
a = C2 + C7
By putting Eq.5 and Eq.7 values in Eq.4, we get:
T(n) = a(n) + c
T(n) = O(n)
(5)
(6)
(7)
(8)
(9)
4. Results and Discussion
In this section, we present results obtained from DREAM methodology, then we compare
ODD vs DREAM followed by empirical analysis and related work.
18
4.1. Complex network of the model
Figure 5 shows the complex network of our proposed ABM. This network presents our ABM
in a visualized form. We develop the network model of our proposed ABM using Gephi (a network
toolkit).
First, we start from the root node "ABM". This node is expanded into leaf nodes "global
variables, agent, procedures, and expts". The global variables are further expanded into global
output and input. These are input provided by the user interface to the model. "Agent" node
represents the involved entity in the model which is further expanded into two types of agents
that are named as node and walker. After defining global inputs, agents, our next focus is on
"procedure" node. The "procedure" node is the root for all procedures used in the model. This
node has the highest node degree in the network model. The node "expts" is a parent for all
sub-nodes that represent different experiments carried out in model simulation.
Figure 5: A network model of our proposed ABM for the smart grid. We can see that in the network the root node is
"ABM" which is connected to the four nodes named as "globals, agents, procedure, and expts". The "globals" node
is connected to the globals output and input parameters. The "agent" node is connected with the involved entities in
the model. The "procedure" node is connected with all other processes and functions in the model. The "expts" node
is connected with all experiments carried out in the model simulation.
4.2. Social network analysis
In this section, we present the results obtained by applying social network analysis(SNA) tool
on the network model. The SNA provides quantitative measures to give network topological de-
tails. Using these quantitative measures we can perform a comparison of different network models.
In figure 8 degree centrality of the network is plotted. It shows that the procedure node has
highest centrality. Next the ABM node and third is global-input.
19
In figure 9 betweenness centrality is presented. It demonstrates that ABM node has highest
betweenness centrality and the second highest betweennes centrality demonstrated by procedure
node.
Figure 10 shows closeness centrality of the network.
It demonstrates that ABM node has
highest closeness centrality and procedure node is on second number.
Figure 11 shows eigenvector centrality of the network. It shows that procedure node has high-
est eigenvector centrality followed by ABM node.
4.3. Comparison of ODD and DREAM
In this section, we provide a qualitative as well as quantitative comparison between ODD and
DREAM specification techniques.
The ODD specification allows a textual based description of ABM with the purpose to make
model readable and promotes the rigorous formulation of models.
It comprises checklist that
covers key features through which one can describe an ABM. The ODD specification has some
limitations that are described in the following.
The ODD specification only provides a textual based description of ABM. Sometimes for
large ABM, it has less description, which is insufficient to cover all the features of the ABM. It
has no quantitative assessment of the ABM on the base of which one can perform a comparison
between different ABM. Reviewing and comparison of different ABM is difficult. For comparison
and classification purpose, the only possible way is to make a table and put ODD checklists of
different ABM. Then search for similarities and differences.
According to the Grimm et al. [9], a survey was conducted from 2006 to 2009 of those pub-
lications in which ODD was used. According to this survey, only 75% publications used ODD
correctly, while 25% publications used ODD incorrectly and some parts of the protocol were
compromised. Grimm et al. [9] formulates the conclusion that it is difficult to write an ABM
specification by following ODD protocol.
Another issue which is identified in the ODD specification is redundancy. Some parts of the
specification like purpose section is also included in the introduction section of the document. the
design concept section is also repeated in the sub model section of the method. Sub-model section
is repeated in the process and scheduling section.
Sometimes there may be different publications with a different version of the same ABM. Then
these publications have the same ODD with little modification in entities and process section.
Another limitation of ODD is that the textual based description is too specific which is not useful
for replication of ABM; there exist some ambiguities and misunderstanding about ABM.
On the other side, DREAM allows a detailed specification of ABM. It comprises of making a
complex network of the model, pseudo-code specification, and network analysis steps.
This method allows for inter-disciplinary comparative study and communication among differ-
ent scientific domains. So, if a model is developed in a social science, it can be compared visually
and quantitatively with a model developed in biological science and vice versa. For an instance, a
social model for aid spreading presented in [16] and biological model developed for the emergence
of snake-like structure in [17]. By developing complex networks of both ABM models, we can
easily compare and analyze both networks in the same manner. DREAM specification approach
can be applied to any ABM of any research domain.
20
DREAM provides to present ABM in the complex network model. This allows reading and
understanding ABM visually without going to the exact code specification. By performing net-
work analysis on this network of ABM, it also gives a quantitative measurement of the ABM.
These quantitative measures are the digital footprint of ABM and can be used to compare different
ABM.
DREAM further allows pseudo-code specification and details of the ABM. This specification
helps in understanding ABM completely without regards discipline. This specification then offers
a translation to the code and developing ABM. It concludes that by using DREAM, any ABM can
be replicated easily.
Next, we present an empirical analysis of the ODD and DREAM methodologies. We evaluated
these methodologies based on 13 features as shown in table 2. We used H=2, M=1 and L=0 for
evaluation purpose that demonstrates which methodology offers which feature to what extent.
Then we computed the rank of each methodology by averaging results shown in table 3.
Table 2: Selected features for empirical analysis of ODD and DREAM
Feature
Social and technical
process
Adaptability
Reusability
Redundancy
Stepwise
Documentation
Network model
Pseudo-code
Network analysis
Transition
Communication
Confusion
User satisfaction
Description
Considers the level of social and technical aspects in
the methodology
Referring how much the methodology is flexible to be
adopted in different domains
It refers the extent of the methodology to be used for
model replication
Concerns with the number of repeated modules in the
methodology
Measures how much methodology is based on se-
quential steps
It involves the process of documenting the model
It is the developing of network for the model
Concerns with the presenting Pseudo-code specifica-
tion for the model
Concerns to applying social network analysis tool on
network
Refers the extent of conversion to model development
Refers to the level of communication among multiple
disciplines
It shows the degree of complexity in the methodology
Refers to the level of convenience
21
Table 3: Empirical analysis of ODD vs DREAM
Feature ODD DREAM
F(1)
F(2)
F(3)
F(4)
F(5)
F(6)
F(7)
F(8)
F(9)
F(10)
F(11)
F(12)
F(13)
Rank
H
H
H
L
H
M
H
H
H
H
H
H
H
1.76
H
L
L
M
H
H
H
L
L
L
L
L
L
0.69
4.4. Simulation results
4.4.1. Simulation setup
To simulate a smart grid based complex scenario, we developed small-world, scale-free and
random complex network using Agent-Based Modeling approach. In order to validate our work,
we applied routing techniques such as random walk and centrality-based routing on large-scale
complex networks, specifically in the smart grid domain. For comparison, we applied random and
centrality-based routing on these networks and analyzed their behavior on these networks.
4.4.2. Evaluation metrics
For performance evaluation purpose, we used average delivery rate parameter. The average
delivery rate is defined as the number of packets sent by sources and successfully received by
consumers. Mathematically can be written as following:
n(cid:88)
1
Ds
Dc
100
(10)
Where Ds represents data packets sent by the source and Dc represents data packets received by
consumers. The experiment was performed for different case studies such as a different number of
consumers and generation units. Then the simulation results were averaged over 30 executions. To
see the behavior of routing techniques, when going from source locations towards the destinations
through different paths at each time step. We used different parameters and observed for which
combination it takes less convergence time. The simulation environment is set according to the
parameters as shown in table 1.
22
4.4.3. Results of Random walk
We applied random walk routing technique on different complex networks. The simulation
results demonstrate that random walk routing technique showed less iteration in the case of small-
world topology as compared to others network topologies.
For small world network, figure 12a shows the simulation result for different number of sources
and consumers. This shows convergence rate on different case studies. The results show conver-
gence rate lies between 180-280 iterations.
For scale-free network, figure 12b shows simulation results for different numbers of sources
and consumers. Random walk shows large convergence rate as compared to small world network.
For random network, figure 13a shows simulation results for different numbers of sources and
consumers. In this network topology, random walk demonstrate very large convergence rate as
compared to both scale-free and small world. This is due to the network topology.
Figure 13b shows the performance of random walk on different network topologies. The small-
world topology demonstrates less iteration while random network has very large convergence rate.
4.4.4. Issues identified in Random walk
1. Agents can move randomly on the network, they select a random node from their neighbor
list.
2. Agents can move to previously visited nodes.
3. Agents do not maintain records when traversing nodes of the network.
4. Sometimes, agents get stuck on the network, this increases computational time.
4.4.5. Results of Centrality-based routing
Figure 14a shows centrality routing(CR) on small-world network with different numbers of
consumers and generating units. The simulation results show that on average, each CR has equal
convergence rate. When it is compared with other complex networks, it is found through simula-
tion results that CR on small-world has less convergence rate compared with other networks. In
this case, CR on small-world has convergence rate between 80 to 105 iterations.
Figure 14b shows simulation results of the CR on scale-free network with different numbers
of consumers and generating units. The simulation results demonstrate that on scale-free network,
Degree-CR has high convergence rate as compared to other approaches.
Figure 15a shows simulation results of Degree-CR, Closeness-CR, Betweenness-CR and Eigenvector-
CR applied on random network using different numbers of consumers and generating unit. Then
the obtained results were plotted according to the figure 15a. This shows that Degree-CR has less
convergence time as compared to others centrality routing techniques.
4.4.6. Random-walk vs Centrality-based routing
Figure 15b shows simulation results of different routing techniques on random network. It
demonstrates that random walk has large iteration as compared to other routing techniques.
Figure 16a shows simulation results of different routing techniques on scale-free network. It
demonstrates that Degree-CR and random walk have large convergence time.
23
Figure 16b shows the simulation results of different routing techniques on small-world net-
work. It demonstrates that CR techniques have a similar convergence rate while random walk has
large convergence rate.
The total time reduction of centrality-based routing compared with the random-walk algorithm
on the random network, scale-free network, and small-world network are shown in tables 4,5,6
respectively.
4.5. Comparison with previous work
In this section, we present some related work. Previously, the agent-based and complex
network-based approaches have successfully been used in the smart grid. However, there is no
such model with the conjunction of specification aspect of the model. Next, we present some of
the previous studies from an agent-based and complex network perspective.
In [18], a conceptual model is developed for energy domain. This model is integrated with
the ODD methodology for documenting ABM. In this model, some other concept was added like
layers, objects, actor and working point to bridge between the social and technical system in the
energy domain. However, this conceptual model was not validated on ABM.
In [19], a check-in based routing approach is proposed for network traffic model. In this work,
betweenness centrality used to assign node as the check-in node between source and destination.
The proposed routing strategy was implemented on the scale-free network. However, the opti-
mization of routing remained as the open problem of this work.
In [20], Niazi and Hussain have presented agent-based tools for modeling and simulation of
self-organizing in wireless sensor network. They demonstrated the usability of agent-based tool
NetLogo, and developed different experiments that show how to model different scenario in the
sensor network domain.
Batool and Niazi in [6], presented a novel hybrid approach for modeling internet of things
(IoT)-based complex scenario using standard complex networks such as small-world, scale-free,
random and lattice network. In [21], novel agent-based and complex network-based modeling
approaches are used in the wireless sensor network domain. In this work, different standard com-
plex network such as small-world, scale-free, and random network is used. Further, random walk
routing strategy is implemented for communication between different sensors in the network.
In [22], another routing technique proposed for large-scale sensor network-based environment.
In this work, local and global update strategy is introduced for maintenance and efficient routing
in the network. This approach monitors any changes in network and update routing path according
to the situation. Results demonstrate the effectiveness of the techniques and reduced end to end
delivery rate as compared to the previous techniques.
In [23], research work is carried out on frequency synchronization in the power grid system.
In this work, the network theory concept was used to monitor, control and exploit the frequency
variation of the power system.
In [24], research work is carried out on security analysis using complex network approach in
power system. In this work, the power adjacency matrix approach is proposed for analysis and
measurement of the power flow and activities of each node and links on the network.
In [25], proposed a novel routing strategy based on betweenness centrality in a complex net-
work. In this work scale-free network is used and routing was performed based on expanding
24
betweenness of each node. This method shifts the load from the node with higher betweenness to
the lower.
Liu et al. in [26], proposed an efficient probability routing strategy using scale-free complex
network topology. This method utilizes the probability concept for redistributing load from critical
nodes to the non-critical nodes. Results showed that routing path is reduced 30% as compared with
previously probability routing technique.
In [27], research work is carried out on routing strategy in wireless sensor network and pro-
posed sink betweenness distributed routing algorithm. In this approach, betweenness of each node
is calculated in which sink node exists as a terminal node. This work was implemented on the
random network.
Ansari et al. in [28], proposed a multi-agent system for reactive power control system in a smart
grid. This work reduced power losses and provided exploitation of available power resources.
Zhang et al. in [29], presented a voltage variation control strategy. This approach controls voltage
profile in the specified range of the studied system, which results in reducing system loss and
improving system reliability. Authors in [30, 31] have worked on fault location and restoration
in smart grid by using the complex network approach. These studies demonstrate the modeling
of fault location and restoration process in a distributed power network. Weng et al.
in [32]
also worked on fault location and proposed the use of particle swarm optimization technique for
locating voltage disturbance sources in a distributed power grid.
Regarding communication management, different studies also have been presented in smart
grid. Wang et al. in [33], proposed an adaptive strategy for energy trading between utility grid
and consumers. In this work, each agent can communicate with each other for sharing information
about energy usage and cost. Tsai et al. in[34] have worked on distributed large-scale consumers
load with the conjunction of renewable energy resources. In this work, a neighbor communication
strategy is applied. This results in low communication cost. E. Kremers et al. in [35] presented
a bottom-up approach for the smart grid modeling. It consists of two layers; physical layer for
electrical power transmission and logical layer for communication. This model has the ability
to integrate new devices in the smart grid environment. It provides dynamic load management,
power, communication control and monitoring.
Regarding power scheduling, In [36], Chao and Hsiung proposed Fair Energy Resource Allo-
cation (FERA) algorithm for electricity trading in smart grid. This technique prevents starvation
situation and fatal problem, it also reduces power cost. In [37] Li et al. proposed a look-ahead
scheduling model for flexible loads in a smart grid environment. This model provides flexible
strategies to handle the large flexible load.
In [38] have proposed a multiple microgrid based
model, in which each micro grid trades energy to the market and randomly selects their strategy
for maximizing their revenue. Samadi et al. in [39] worked on load scheduling and power trading
with a combination of large scale renewable energy resources. This model allows users to con-
sume local energy. In [40] an agent-based model is proposed for large scale virtual power plant.
The proposed model constitutes of complex, heterogeneous distributed energy resources with a
combination of local and multi-objective function. In [41], Shirzeh et al. worked on management
of renewable energy resources and storage devices in a smart grid environment. They proposed a
multi-agent system based on a plug-and-play technique for management and controlling storage
resources in the smart grid. In [42], Kuznetsova et al. presented two step-ahead reinforcement
25
learning algorithm for battery scheduling within microgrid architecture. It is composed of local
consumers, generator and storage devices connected to the external grid. This technique predicts
and forecasts power demand of consumers.
5. Conclusions
In this paper, we proposed modeling and simulation of the smart grid by using a combination
of agent-based and complex network-based approaches. To further validate our work, we applied
routing techniques such as random walk and centrality-based routing. We used different standard
complex networks such as small-world, scale-free and random network. The simulation results
demonstrate that centrality-based routing gives better results on the small-world network as com-
pared to other networks. For ABM specification, we followed two approaches, the one is ODD and
the other is DREAM methodology. We presented qualitative as well as a quantitative comparison
of both ODD and DREAM specification techniques. The comparative study of ODD and DREAM
proved that DREAM methodology is the more useful approach for documenting an ABM not only
in terms of modeling but also for replication of the models, specifically in the smart grid domain.
26
Figure 6: Screenshot of the developed ABM of smart grid. The image shows user interface of NetLogo simulation
tool. It consists of sliders, chooser, monitors, buttons, and world ( a simulation environment).
27
(a) Small-world network
(b) Scale-free network
(c) Random network
Figure 7: Developed smart grid scenarios based on standard complex networks: Part(a) shows small-world consti-
tutes of 500 nodes, number of consumers and sources are selected randomly. Part(b) shows scale-free network with
500nodes, a number of consumers and sources are selected randomly. Part(c) demonstrates random network consists
of 500nodes, a number of consumers and sources are selected randomly.
28
Figure 8: Degree centrality of the network: It shows Prodecdure node has the highest degree in the network.
Figure 9: Betweenness centrality of the network: ABM node has the highest betweeness centrality value in the net-
work.
29
Figure 10: Closeness centrality of the network: ABM and Procedure node has the highest closeness centrality value.
Figure 11: Eigenvector centrality of the network: It shows Procedure node is on the top list in the network model.
30
(a) Random Walk on Small-world network
(b) Random Walk on scale-free network
Figure 12: The simulation results of Random walk on different networks: Part(a) shows results of a random walk on
the small-world network. The x-axis shows a number of consumers and the y-axis shows average end to end delivery
rate against a different number of sources. Part(b) shows results of random walk for different number of consumers
and sources on scale-free network of five hundred nodes.
(a) Random Walk on random network
(b) Random Walk on overall networks
Figure 13: The simulation results of random walk: Part(a) shows random walk results on a random network with five
hundred nodes. Part(b) shows random walk results on overall complex networks. Results of all three networks are
compared with(number of consumers: 50, 100, 150, 200, number of sources: 10, 50, 100, 150).
31
(a) Centrality-based routing on small-world network (b) Centrality-based routing on scale-free network
Figure 14: The simulation results of Centrality-based routing on complex networks. Results for different types of
centrality routing based on Closeness, Betweenness, Eigenvector, and Degree are plotted. The experiments were
performed for consumers: 50, 100, 150, 200 and sources: 10, 50, 100, 150: Part(a) shows centrality-based routing on
the small-world network. Part(b) shows centrality-based routing results on scale-free network.
(a) Centrality-based routing on random network
(b) RW vs CR on random network
Figure 15: Part(a) shows simulation results of different centrality-based routing algorithm on a random complex net-
work comprises of five hundred nodes. The number of consumers: 50,100,150,200 and number of sources: 10,50,100
are used. In part(b), the simulation results of random-walk and centrality-based routing algorithms are compared for
different number consumers and sources.
32
(a) RW vs CR on scale-free network
(b) RW vs CR on small-world network
Figure 16: Comparison of random-walk and centrality-based routing algorithms on different networks. Experiments
were performed for consumers:50,100,150,200 and source:50,100,150,200. Part(a) shows comparative results of
random-walk and centrality-based routing on scale-free network. Part(b) shows comparative results of random-walk
and centrality-based routing on small-world network. The simulation results of centrality routing based on closeness,
betweenness, eigenvector and degree are same for small-world network.
Table 4: Comparison on Random network
Technique
RW vs Eigenvector
RW vs Betweenness
RW vs Closeness
RW vs Degree
Time reduction (%)
55
35
54
56
Table 5: Comparison on Scale-free network
Technique
RW vs Eigenvector
RW vs Betweenness
RW vs Closeness
RW vs Degree
Time reduction (%)
40.25
24
46.8
-82.5
33
Table 6: Comparison on small-world network
Technique
RW vs Eigenvector
RW vs Betweenness
RW vs Closeness
RW vs Degree
Time reduction (%)
60
60
59.9
59.8
Table 7: A comparative analysis of previous studies in the smart grid. We analyzed the previous study based on ABM,
complex network, specification techniques ODD and DREAM. The comparative study confirms that there is no such
specification study for ABM in the smart grid.
Author
Ansari
Objective
MAS for reactive power man-
agement
ABM CN
ODD DREAM
Zhang et al Adaptive strategy for energy
Ref.
[28]
[33]
[39]
[34]
[35]
[37]
[38]
[42]
[18]
[23]
[24]
[30, 31] Erikson et
al.
Samadi
al.
Tsai et al.
et
Kremers et
al.
in
trading
Fault location and restoration
in power system network
Appliances
scheduling
smart home
Communication
among
large-scale distributed con-
sumers load
Simple power system mod-
eling with consumers and
power generators
Scheduling of flexible loads
Li et al.
Wang et al. Modeling multiple micro-
grids
Kuznetsova Battery storage scheduling
Hinker
Wang
Jia
A conceptual model for smart
grid
Frequency synchronization in
power system
Security analysis in power
system
34
Category
Overview
Design concept
Detail
Table 8: Model specification follows ODD
Sub-category
Purpose
Entities
Process
Basic principle
Emergence
Adaptation
Objective
Sensing
Interaction
Stochasticity
Obervation
Initialization
Input data
Sub-model
Our-model
Modeling smart grid using agent-based and complex
network-based approach
Producers, consumers, walkers
a hybrid centrality-based routing algorithm for end to
end delivery from producers to consumers
a cognitive agent-based computing approach is better
for modeling and simulation of the large scale power
system.
Computation time of end to end delivery from pro-
ducers towards consumers
Based on connected neighbors
To measure how much time is taken while moving
from one node to the other
Check the state of the neighbor nodes
Local communication
Random process
Collect data about number of consumers, number of
producers, number of nodes visited
Complex network setup
External network setup files
Parameters:
• number of nodes
• number of sources
• centrality-based routing
• average delivery rate calculation
35
References
References
[1] O. Ellabban, H. Abu-Rub, Smart grid customers' acceptance and engagement: An overview, Renewable and
Sustainable Energy Reviews 65 (2016) 1285–1298.
[2] J. M. Epstein, Why model?, Journal of Artificial Societies and Social Simulation 11 (4) (2008) 12.
[3] M. A. Niazi, Towards a novel unified framework for developing formal, network and validated agent-based
simulation models of complex adaptive systems, Ph.D. thesis, University of Stirling (2011).
[4] M. A. Niazi, Complex adaptive systems modeling: a multidisciplinary roadmap, Complex Adaptive Systems
Modeling 1 (1) (2013) 1.
[5] D. P. Chassin, C. Posse, Evaluating north american electric grid reliability using the barab´asi–albert network
model, Physica A: Statistical Mechanics and its Applications 355 (2) (2005) 667–677.
[6] K. Batool, M. A. Niazi, Modeling the internet of things: a hybrid modeling approach using complex networks
and agent-based models, Complex Adaptive Systems Modeling 5 (1) (2017) 4.
[7] V. Grimm, U. Berger, F. Bastiansen, S. Eliassen, V. Ginot, J. Giske, J. Goss-Custard, T. Grand, S. K. Heinz,
G. Huse, et al., A standard protocol for describing individual-based and agent-based models, Ecological mod-
elling 198 (1) (2006) 115–126.
[8] M. A. Niazi, A. Hussain, Cognitive agent-based computing-I: a unified framework for modeling complex adap-
tive systems using agent-based & complex network-based methods, Springer Science & Business Media, 2012.
[9] V. Grimm, U. Berger, D. L. DeAngelis, J. G. Polhill, J. Giske, S. F. Railsback, The odd protocol: a review and
first update, Ecological modelling 221 (23) (2010) 2760–2768.
[10] D. J. Watts, S. H. Strogatz, Collective dynamics of'small-world'networks, nature 393 (6684) (1998) 440.
[11] A.-L. Barab´asi, R. Albert, H. Jeong, Mean-field theory for scale-free random networks, Physica A: Statistical
Mechanics and its Applications 272 (1) (1999) 173–187.
[12] G. Barrenechea, B. Beferull-Lozano, M. Vetterli, Lattice sensor networks: Capacity limits, optimal routing and
robustness to failures, in: Information Processing in Sensor Networks, 2004. IPSN 2004. Third International
Symposium on, IEEE, 2004, pp. 186–195.
[13] K.-K. Kleineberg, D. Helbing, Collective navigation of complex networks: Participatory greedy routing, Scien-
tific Reports 7.
[14] B. Lin, B. Chen, Y. Gao, K. T. Chi, C. Dong, L. Miao, B. Wang, Advanced algorithms for local routing strategy
on complex networks, PloS one 11 (7) (2016) e0156756.
[15] M. Rekik, Z. Chtourou, N. Mitton, A. Atieh, Geographic routing protocol for the deployment of virtual power
plant within the smart grid, Sustainable Cities and Society 25 (2016) 39–48.
[16] M. A. Niazi, A. Siddiqa, G. Fortino, Modeling aids spread in social networks, in: German Conference on
Multiagent System Technologies, Springer, 2013, pp. 361–371.
[17] M. A. Niazi, Emergence of a snake-like structure in mobile distributed agents: an exploratory agent-based
modeling approach, The Scientific World Journal 2014.
[18] J. Hinker, C. Hemkendreis, E. Drewing, S. Marz, D. I. H. Rodr´ıguez, J. M. Myrzik, A novel conceptual model
facilitating the derivation of agent-based models for analyzing socio-technical optimality gaps in the energy
domain, Energy.
[19] Z.-Y. Jiang, J.-F. Ma, Y.-L. Shen, Check-in based routing strategy in scale-free networks, Physica A: Statistical
Mechanics and its Applications 468 (2017) 205–211.
[20] M. Niazi, A. Hussain, Agent-based tools for modeling and simulation of self-organization in peer-to-peer, ad
hoc, and other complex networks, IEEE Communications Magazine 47 (3).
[21] K. Batool, M. A. Niazi, S. Sadik, A. R. R. Shakil, Towards modeling complex wireless sensor networks using
agents and networks: a systematic approach, in: TENCON 2014-2014 IEEE Region 10 Conference, IEEE, 2014,
pp. 1–6.
[22] L. Pradittasnee, S. Camtepe, Y.-C. Tian, Efficient route update and maintenance for reliable routing in large-scale
sensor networks, IEEE Transactions on Industrial Informatics 13 (1) (2017) 144–156.
[23] C. Wang, Synchronisation in complex networks with applications to power grids, Ph.D. thesis, University of
Aberdeen (2017).
36
[24] Y. Jia, Z. Xu, S.-L. Ho, Z. X. Feng, L. L. Lai, Security analysis of smart grids-a complex network perspective.
[25] Z.-H. Guan, L. Chen, T.-H. Qian, Routing in scale-free networks based on expanding betweenness centrality,
Physica A: Statistical mechanics and its applications 390 (6) (2011) 1131–1138.
[26] W. yan Liu, X. Li, J. Li, B. Liu, An efficient probability routing algorithm for scale-free networks, Chinese
[27] E. M. Oliveira, H. S. Ramos, A. A. Loureiro, Centrality-based routing for wireless sensor networks, in: Wireless
Journal of Physics 55 (3) (2017) 667–673.
Days (WD), 2010 IFIP, IEEE, 2010, pp. 1–5.
[28] J. Ansari, A. Gholami, A. Kazemi, Multi-agent systems for reactive power control in smart grids, International
Journal of Electrical Power & Energy Systems 83 (2016) 411–425.
[29] X. Zhang, A. J. Flueck, C. P. Nguyen, Agent-based distributed volt/var control with distributed power flow solver
in smart grid, IEEE Transactions on Smart Grid 7 (2) (2016) 600–607.
[30] M. Eriksson, M. Armendariz, O. O. Vasilenko, A. Saleem, L. Nordstrom, Multiagent-based distribution automa-
tion solution for self-healing grids, IEEE Transactions on Industrial Electronics 62 (4) (2015) 2620–2628.
[31] M. J. Ghorbani, M. A. Choudhry, A. Feliachi, A multiagent design for power distribution systems automation,
IEEE Transactions on Smart Grid 7 (1) (2016) 329–339.
[32] G. Weng, F. Huang, Y. Tang, J. Yan, Y. Nan, H. He, Fault-tolerant location of transient voltage disturbance
source for dg integrated smart grid, Electric Power Systems Research 144 (2017) 13–22.
[33] Z. Wang, L. Wang, Adaptive negotiation agent for facilitating bi-directional energy trading between smart build-
ing and utility grid, IEEE Transactions on Smart Grid 4 (2) (2013) 702–710.
[34] S.-C. Tsai, Y.-H. Tseng, T.-H. Chang, Communication-efficient distributed demand response: A randomized
admm approach, IEEE Transactions on Smart Grid 8 (3) (2017) 1085–1095.
[35] E. Kremers, J. G. de Durana, O. Barambones, Multi-agent modeling for the simulation of a simple smart micro-
grid, Energy Conversion and Management 75 (2013) 643–650.
[36] H.-L. Chao, P.-A. Hsiung, A fair energy resource allocation strategy for micro grid, Microprocessors and Mi-
crosystems 42 (2016) 235–244.
[37] Y. Li, T. Yong, J. Cao, P. Ju, J. Yao, S. Yang, A consensus control strategy for dynamic power system look-ahead
scheduling, Neurocomputing 168 (2015) 1085–1093.
[38] H. Wang, T. Huang, X. Liao, H. Abu-Rub, G. Chen, Reinforcement learning in energy trading game among
smart microgrids, IEEE Transactions on Industrial Electronics 63 (8) (2016) 5109–5119.
[39] P. Samadi, V. W. Wong, R. Schober, Load scheduling and power trading in systems with high penetration of
renewable energy resources, IEEE Transactions on Smart Grid 7 (4) (2016) 1802–1812.
[40] A. Clausen, A. Umair, Y. Demazeau, B. N. Jørgensen, Agent-based integration of complex and heterogeneous
distributed energy resources in virtual power plants, in: International Conference on Practical Applications of
Agents and Multi-Agent Systems, Springer, 2017, pp. 43–55.
[41] H. Shirzeh, F. Naghdy, P. Ciufo, M. Ros, Balancing energy in the smart grid using distributed value function
(dvf), IEEE Transactions on Smart Grid 6 (2) (2015) 808–818.
[42] E. Kuznetsova, Y.-F. Li, C. Ruiz, E. Zio, G. Ault, K. Bell, Reinforcement learning for microgrid energy man-
agement, Energy 59 (2013) 133–146.
37
|
1810.13195 | 1 | 1810 | 2018-10-31T10:17:33 | A multi-agent system for managing the product lifecycle sustainability | [
"cs.MA"
] | The international competitive market causes the increasing of shorten product life cycle and product development process with the improvement in term of time, cost and quality while increasing the waste generation. Product life cycle sustainability can reduce waste, conserve resources, use recycling materials, design product for easy disassembly and avoid using hazardous material. This paper proposes a knowledge management architecture, based on a multi-agent system, which focuses on the "sustainability" in order to manage knowledge in each stage of the product lifecycle, and particularly in the recovery process. The aim of this research work is to make the link between a decision-making system based on the agent's knowledge about the sustainability (environmental norms, rules...) and a PLM (Product Lifecycle Management) system. The software Agents will help the decision makers in each stage of the lifecycle and make them take into account the environmental impact of their decisions. | cs.MA | cs | A multi-agent system for managing the product lifecycle sustainability
T.MANAKITSIRISUTHI, Y.OUZROUT, A. BOURAS
Université Lumière Lyon 2,
LISEP Laboratory
160 Bd de l'université
69676 Bron Cedex
Abstract. The international competitive market causes the increasing of shorten product life
cycle and product development process with the improvement in term of time, cost and quality
while increasing the waste generation. Product life cycle sustainability can reduce waste,
conserve resources, use recycling materials, design product for easy disassembly and avoid using
hazardous material. This paper proposes a knowledge management architecture, based on a
multi-agent system, which focuses on the "sustainability" in order to manage knowledge in each
stage of the product lifecycle, and particularly in the recovery process.
The aim of this research work is to make the link between a decision-making system based on
the agent's knowledge about the sustainability (environmental norms, rules…) and a PLM
(Product Lifecycle Management) system. The software Agents will help the decision makers in
each stage of the lifecycle and make them take into account the environmental impact of their
decisions.
The proposed architecture will be illustrated in a case study.
Keywords: Recovery process, Product
Knowledge management.
lifecycle sustainability, Multi-Agents System,
1.
INTRODUCTION
Today, increasing international competition and growing markets and the advanced technology
cause the shorten product life and product development which result on the improvement of its
performance in terms of time, cost and quality at the same time increasing the waste generation.
The huge amounts with different kind of products have been introduced into the market to satisfy
several of customers' requirements thus, the environment can be affected inevitability. The
degree of environmental impacts is determined by materials and energy used in the production
process and outputs generated at each stage of product's lifecycle. Managing the returned
product can help in reducing the damage of environment. Environmental laws and regulations
have been established to require the organization to take responsibility of return products when
they come to end of their life or end of usage. Some organizations have found that returned
products or used products can be the additional source of revenues by recycling materials or
reusing product's components or parts after disassembling in the manufacturing process. This
brings the attention of organization in managing the returned products.
In order to produce the products that are sustainable and less harmful the environment,
knowledge that occurred during the production process, recovery process, and the related
activities of organization including the knowledge about environmental performance should be
captured, evaluated and stored.
Consideration of environmental issues and regulations during the design process, product
development and particularly in recovery process can help user for decision making in each stage
of lifecycle, consequently, minimizing the waste generation of product and improving the
environmental performance.
This research purposes a knowledge management architecture based on multi-agents system in
order to manage knowledge about the environmental management and the system will facilitate
in decision making by making the link of agent's knowledge about sustainability and PLM
system. The system will help decision makers in each stage of the lifecycle and make them take
into account the environmental impacts of their decisions.
2. LITERATURE REVIEW
In the past, organizations have been asked to develop products and services which are better in
quality, less expensive within timely manner. Environmental effects were overlooked during the
design stage for new products and processes. Hazardous wastes were discarded inattention in the
most convenient ways which damage to our environment. Concerning the environmental impact,
customers are increasing their expectation that organizations are be able to handle and willing to
take responsibility of their returned products (relationship between trust and reverse logistics
performance) [1] and to be able to reduce the environmental impact from their activities and
products.
Organizations can create their profitability from repairing, remanufacturing and recycling of
returned products [2]. Since customers are conscious to environmental impact, organizations are
increasing the improvement of customer satisfaction by minimizing raw materials used in the
production process at the same time reducing costs of production and wastes generated into
environment [3]. Therefore, reverse logistics are increasingly utilized by organizations as
competitive strategy [4] [5]. Environmental legislations have been raised in order to improve the
environmental performance of organization [6].
There are legislations concerned about environmental performance to encourage organizations
towards sustainable product by increasing the reuse, recycle, and recovery of returned product.
Many organizations have considered environmental issue as a key driver influencing to
environmental performance [7] [8]. The objective of Environmental management system is to
minimize the damage on the environment which cause by the organization's activities and
products [9].
The influence of environmental regulations, customer awareness and social responsibility, the
decision making during the design stage plays an important role to environmental performance.
Thus, organization needs to design product with the improvement of effective the use of
materials and energy and for better quality with less effect to environment. Moreover, in
recovery process, the effective of reusing, recycling and remanufacturing of returned product can
help decreasing environmental impacts and conserving resources and energy. The theoretical
model of eco-oriented design [10] has presented in order for suggestion the design alternatives to
designers and decision makers for balancing the use of nature resources and industry. Another
research [11] shown the concept of design for sustainability, the concept will encourage
organization to design products and services in dimensions of environmental, social and
economic performances.
Figure 1 shows an overview of product's activities in both forward logistics and reverse logistics.
Used product will be returned back to the organization with varieties reasons of returns. The
inspection stage will need a lot of knowledge intensive to consider where the returned product
should go to recycle, repair, resale or reuse.
Figure1. Product lifecycle stage in forward and reverse logistics
Recovery process can help organization consume less of materials by reusing the second life of
materials or parts and components from the return products which results in reduction the use of
virgin of raw material and increases productivity while minimizes the environmental impacts.
Increasing awareness of need for sustainable products will result in the integration of
environmental aspect into product design and product development. The proposed knowledge
management architecture based on multi agent system model in order to link agents' knowledge
and PLM system and support user decision making in each stage of product lifecycle related to
environmental impacts.
3. KNOWLEDGE MANAGEMENT ARCHITECTURE
There are many reasons of using agents for developing knowledge management system. First,
they are proactive due to user in recovery process or production process does not know or does
not have time to search on product information or knowledge related to environmental
performance. Therefore, agents can help user to collect and to manage those knowledge.
Second, agents can make the system become more efficient because agents can learn from their
own previous experiences either failures or successes [12]. Third, each agent in the system can
use different technique to solve the problem such as Inspect agent uses CBR (Case-Based
Reasoning) to help inspector solving problems. And the last one is agents can manage its own
knowledge and interact to other agents ask for the solutions which does not exist in their
knowledge base.
The architecture is composed of Inspect agent, Redesign agent, Recover agent and Disposal
agent (see Figure2).
Figure2. Knowledge management system architecture based on multi-agent system
The Inspect agent is an agent to check the quality of returned product from the previous
experience in their knowledge base. It also sends message asking the other agents if there any
solutions exist (see Figure3). Then Inspect agent will determine the destination of returned
product (repair, reuse, resale, recycling or redesign). For example, in case of the returned product
need to be fixed or replace the damaged parts/components, Inspect agent will set the status of
returned product to repair.
Figure3. Inspect agent sends a request to for information
The Recover agent, after receives the message from Inspect agent that returned product needs to
be repaired, Recover agent will search all the solutions how to recover product. The solutions in
Recover agent's knowledge base are the solutions concerned to the environmental performance.
For example, Recover agent may send information about parts or materials which can be reused
in the inspected product or inform the substitute materials.
Once an Inspect agent analyzes and finds out that the returned product is needed to redesign in
order to respond customer satisfactions. Redesign agent will do their work by search the solution
in knowledge base along with the product information in PLM system. The given solutions will
be presented and correspondence with the concept of product lifecycle sustainability such as
redesign the user manual to reduce the amount of paper or to extend the life of product usage in
case of reason of return is customers do not know how to operate the product. In term of
conserving resources and energies, Redesign agent may inform designer to increase the use of
recycled materials in product, reduce weight and size of components and products.
In order to reduce environmental impact as much as possible, Disposal agent manages
information related to reuse materials or components from the damaged product. Disposal agent
can send information about the recyclable design product or inform to label material type in
order to reduce the time and cost of recycling.
The aim of this research is to encourage user of making their decision in each stage of product
lifecycle mainly concentrate on environmental impact of product. To develop the prototype and
testing the proposed architecture based on multi agent concept, JADE (Java Agent Development
Framework) platform was chosen. JADE is an open source, implemented in java language and
compliance with FIPA specifications.
4. CONCLUSION AND FUTURE WORK
This paper presents knowledge management architecture based on multi agent system to help
user in each stage of product lifecycle for decision making by taking to account on
environmental impacts.
The prototype is in the process of developing for testing some scenarios for product lifecycle
sustainability.
REFERENCE
1. Daugherty, P.J., Richey, R.G., Hudgens, B.J. and Autry, C.W., "Reverse logistics in the
automobile aftermarket industry", International Journal of Logistics Management, 2003, Vol.
14, No. 1, pp. 49-61.
2. Giuntini, Ron and Andel, Tom," Master the six R's of reverse logistics" Transportation and
Distribution, March 1995.
3. Ayres, R.U., Ferrer, G., Van Leynseele, T., 1997. "Eco-efficiency, asset recovery and
remanufacturing". Eur. Manag. J. 15 (5), 557 -- 574.
4. Edward J. Marien, "Reverse Logistics as competitive strategy", Supply Chain Management
Review, 1998.
5. Stock, J.R., Speh, T.W., Shear, H.W., 2002. Many happy (product) returns. Harvard Business
Review 80 (7), 16-17
6. Bloemhof J, van Nunen J. "Integration of environmental management and SCM". ERIM,
Report Series Research in Management 2005; ERS-2005-030-LIS.
7. H.-J. Bullinger, J. von Steinaecker, A. Weller "Concepts and methods for a production
integrated environmental protection", International Journal of Product Economics, 60-61
(1999) 35-42
8. Shane J. Schvaneveldt, "Environmental performance of products: Benchmarks and tools for
measuring improvement", Benchmarking: An International Journal, 2003, vol.10, issue 2, p
137-152.
9. http://www.iso.org/iso/iso_catalogue/management_standards/iso_9000_iso_14000/iso_14000
_essentials.htm
10. Hernane Borges de Barros Pereira and Paulo Fernando de Almeida Souza, "Design for
sustainability: Method in search for a better harmony between industry and nature".
11. Marcel Crul, Jan Carel Diehl, "Design for Sustainability (D4S): Manual and Tools for
Developing Countries", 2008, 7th Annual ASEE Global Colloquium on Engineering
Education.
12. P. Maes, "Agents that reduce work and information overload", Communications of the ACM,
37(7), (1994), p. 31-40.
|
1609.03996 | 1 | 1609 | 2016-09-13T19:45:59 | SEAL's operating manual: a Spatially-bounded Economic Agent-based Lab | [
"cs.MA"
] | This text reports in detail how SEAL, a modeling framework for the economy based on individual agents and firms, works. Thus, it aims to be an usage manual for those wishing to use SEAL or SEAL's results. As a reference work, theoretical and research studies are only cited. SEAL is thought as a Lab that enables the simulation of the economy with spatially bounded microeconomic-based computational agents. Part of the novelty of SEAL comes from the possibility of simulating the economy in space and the instantiation of different public offices, i.e. government institutions, with embedded markets and actual data. SEAL is designed for Public Policy analysis, specifically those related to Public Finance, Taxes and Real Estate. | cs.MA | cs |
RESEARCH REPORT
1
SEAL's operating manual:
a Spatially-bounded Economic Agent-based Lab
Bernardo Alves Furtado, Isaque Daniel Rocha Eberhardt, and Alexandre Messa
Abstract -- This text reports in detail how SEAL, a modeling framework for the economy based on individual agents and
firms, works. Thus, it aims to be an usage manual for those wishing to use SEAL or SEAL's results. As a reference work,
theoretical and research studies are only cited. SEAL is thought as a Lab that enables the simulation of the economy with
spatially bounded microeconomic-based computational agents. Part of the novelty of SEAL comes from the possibility
of simulating the economy in space and the instantiation of different public offices, i.e. government institutions, with
embedded markets and actual data. SEAL is designed for Public Policy analysis, specifically those related to Public
Finance, Taxes and Real Estate.
Index Terms -- Agent-based modeling, public policy, Python, computational economics, municipalities, public finance.
!
1 INTRODUCTION
text
T HIS
reports in detail how SEAL
works.1 Thus, it aims to be a a reference
for those wishing to use SEAL as a model, use
SEAL's outputs, study its details, or work as a
SEAL's collaborator. As a manual, the focus of
the text is not on a specific research question
or a literature review. Thus, the accompanying
literature is not discussed in the report. We
direct the interested reader to other sources [1] --
[6] Rather, here we describe the framework,
the Lab, that we hope may be used for a
number of applications. Henceforth, we have
applied it to [7], [8]. This report follows the
ODD protocol recommendations of Grimm and
colleagues [9]. However, we aim at being more
specific, describing in detail the components of
the model.
We refer to SEAL as a Lab or a framework
that enables the simulation of the economy, its
markets, considering spatially bounded com-
putational agents. SEAL simulates the econ-
omy in space,
its markets and government
institutions. SEAL is specifically designed to
• B. Furtado, I. Eberhardt and A. Silva are with the Institute
for Applied Economic Research, Government of Brazil, Bras´ılia,
Brazil. B. Furtado is also a researcher at CNPq
E-mail: [email protected]
1. This applies to SEAL version 2.0 as of September 2016.
study Public Policy, with an emphasis on Public
Finance, Taxes and Real Estate.
This claim of usability derives directly from
the available modeling elements present in the
model. The Lab comprises Citizens, organized
as families, living and moving among house-
holds, traveling to firms both of which: house-
holds and firms, are spatially geocoded within
actual municipalities' boundaries. Citizens are
born, age and die and they are initiated in the
model according to official statistics. Munic-
ipalities' government is also present. Finally,
interaction amongst agents happens in three
markets: labor, real estate and goods. All those
agents are operated in an agent-based model
simulated in Python 3.4.4.
Besides this introduction, this report includes
a simple statement of purpose (section 2),
a description of
the process overview and
scheduling (section 3), a full description of the
agents,
its methods, the markets (section 4)
and the different possibilities of simulations
to run (section 5). Section 6 describes all the
necessary data input, parameters setting and
demographic and spatial
information neces-
sary to run SEAL for different configurations
(section 6). Section 7 delineates the outputs of
SEAL (section 7). This report concludes (sec-
tion 8) with the description of the design con-
cepts, following [9], and the possibilities we see
RESEARCH REPORT
for this framework (section 9).
2 PURPOSE
SEAL was designed specifically to tackle pub-
lic policy concerns. A main justification can
be found in [10]. As such, SEAL provides an
economic-based framework to answer a num-
ber of questions. A proposed list is available
in [7], section 4. Illustratively, we could cite:
general evaluation of tax impacts; economic
impacts on mobility; general impacts of agents'
decisions; impacts of changes in government
spatial outreach, to name a few.
3 PROCESS OVERVIEW AND SCHEDULE
SEAL is setup so that the modeler can run a
single test or multiple runs, in order to make
sure the results are consistent (see section 5).
When running a single run, the modeler should
simply type: python main.py in a python
3.X interpreter (Figure 1). However, it would
be wise to set a number of parameters first
(see section 6) and make sure all the necessary
libraries are installed.
3.1 Time_iteration module
SEAL is organized around a discrete time
frame. As such, time is divided into days,
months, quarters and years. It is possible to
insert periodical activities, for any of the agents
in the model, at any one of these periods.
The first function that runs before the actual
time_iteration is have_a_go.
Have_a_go initiates my_simulation as an
instance of time_iteration.TimeControl
class. Then it runs up to the total number of
days setup for the full simulation. Typically,
since we only consider working days, months
have 21 days, thus, we run for 5.040 days
(which is equivalent to 20 years).
3.1.1 Daily activities
Daily activities are restricted to production up-
dates. On Day 0 of the simulation, however,
some setup is put in place before actual sim-
ulation begins. This includes the single time a
2
product is created 2 and a first round of labor
market matching up, so that firms may have
employees and be able to produce during the
first month of simulation.
3.1.2 Monthly activities
In the beginning of every month, the demo-
graphics dynamic is updated. Given the citi-
zen's month of birth, the probabilities of her
death and of her giving birth (see details at
section 4.7) are calculated. Then, firms pay
salaries (workers receive their salaries before
they can go shopping).
Families distribute their money equally
among their members. Then, each member
acquires his consumption goods, and imme-
diately consumes them (see details at goods
market (see section 4.6.1). Governments collect
taxes from consumption. In turn, firms update
or maintain their prices according to their in-
ventory levels (see section 4.5). Finally, each
government municipality spends the collected
taxes on public goods. The region quality of
life (QLI) is improved by the same amount as
the government spending per capita. However,
a TREASURE_INTO_SERVICES parameter ad-
justs the numbers so that QLIs remain close to
a 0 to 1 indicator.
Before the end of the month, the labor market
(see section 4.6.1) is processes, followed by one
run of the real estate market (see section 4.6.3).
Finally, statistics and outputs are calculated
and saved (section 7). Presently, quarters and
years are used exclusively for statistics compu-
tation.
Four
located at
time_iteration for
importing or usage
reasons. However, they pertain to the labor
market and the equalization of
resources
among family members.
functions
other
are
In short, activities run every month are sub-
sequently:
1) Record month
2) Process demographics
deaths).
3) Firms make payments.
(aging, births,
2. The structure of the Lab is compatible with the idea that
companies may create more products with different character-
istics during simulation. The current configuration does not
differentiate among products, which are homogeneous.
RESEARCH REPORT
3
Fig. 1. SEAL flowchart for the major processes and python modules. The grey objects are
definitions or data to start running a simulation. Population data come from Brazilian demographics
census; number and size of firms from RAIS. The orange objects represent decisions to be made
while running the model. Rectangles are parameters definitions and the lozenges represent the
kind of simulation chosen. The gold rectangles represent the most important modules, which
control the model and run the processes. Green parallelograms represent output data (graphs,
maps, txt and csv files).
(cid:1)In parameters.py set the OUTPUT_PATHand in actual_parameters.pyset parametersvalues (cid:1)In actual_parameters.py CHOOSE THE SIMULATION TO RUNthrough True/False parameters. Set a number of other run parameters.Single TestMulti-RunSimulationSensitivity AnalysisAuto AdjustPreparationEntry data (provided):BoundariesPopulation census dataFirms' RAIS data Control_Autoadjust_Model.pyControl_Multi_runs_simulation.pymain.pyControl_Sensitivity_analysis.py1st Run "main.py" or any "control.py" generator.py2nd"main.py" checks need to call "generator.py"have_a_go.pyAgents (Pop. and Firms)Create and save agents3th"main.py" calls "have_a_go.py"4th "have_a_go.py" calls "time_iteration.py"time_iteration.pyRun:DaysMonthsQuartersYearsActivities:Production,ConsumptionHousing, Labor, Real Estate markets,DemographicsCompute statistics:GeneralRegionalAgents FirmsFamilies5th"main.py" calls "plotting.py"ENDs simulationOutput data *.txt:GeneralRegionalFirmsFamiliesAgentsGenerate final graphs and tables RESEARCH REPORT
4
4) Families
redistribute their own cash
within members.
5) Family members consume (Goods mar-
ket).
6) Governments collect taxes.
7) Governments spend the collected taxes on
life quality improvement (QLI).
8) Firms calculate profits and update prices.
9) Labor market is processed.
10) Real estate market is processed.
11) Statistics and output are processed.
4 ACTUAL ENTITIES AND PROCESSES
This section aims at detailing each model agent
variables and methods. All five following agent
types are treated as classes and instantiated
once (see section 6.3) before running the simu-
lation.
4.1 Citizens
Agents have the following variables:
• fixed: id, gender, month of birth,
qualification, family_id.
any
given moment),
(an
• variable: age, money (amount owned
at
savings,
indicator
firm_id, utility
of
consumed),
address (osgeo, ogr), distance,
region_id.
cumulative
amount
When instantiated, agents are given an id
(in successive order of generation), a gender,
an initial age, a fixed qualification and some
initial money. The number of agents,
their
gender and their age come from official data
(see section 6.3). Qualification is drawn from
a gamma distribution with α = β = 3. Thus,
it is independent of age and fixed throughout
the model. Money is drawn from a uniform
distribution in (20, 40).
Agents have a number of methods. The
usual get() methods are used to return vari-
ables values, to check conditions related to age
(such as is_minor and is_retired), and
update variables (such as update_age and
update_money given a certain amount).
A number of methods are related to fam-
ily business. set_family, which then never
changes, get_family_id and a check to see if
the agent belongs_to_family. Address and
consequent region membership change accord-
ing to the changes the family does. Every time
a family changes addresses, values are updated
at the agent level.
Methods related to the workplace include
setting and getting workplace and a check to
see if the agent is_employed. When firm_id
is None, then either agent is unemployed or
is out of working age. The variable distance
always refers to the distance between current
household where the agent resides and current
firm, when employed. Otherwise, it is set to
None.
Each agent also has a method that calculates
current house
distance from his or her
location to prospective and current place
of employment. Information that is printed
and outputted in TXT format
individually
includes: id, gender,
for
qualification, age, money, firm_id,
utility, address, and region_id.
agent
each
4.2 Citizens' families
Families have a single id variable. They also
hold values
that are constantly updated
for
balance, savings, household_id
and
None),
region_id
house_price, address and house. House
is a variable that actually contains the full
instance of a household. It is useful to access
the house price, its address and region. This
scheme is necessary because families move
from house to house, given the endogenous
changes of the model as a whole.
(initiated
at
Finally, an important variable of families is
an inventory members (actually, a Python dic-
tionary) that contains all the instances of the
families' members. Related proper methods are
is_member and num_members, which report
the family size.
The methods in Families include typical get
and set ones; an add_agent, which brings
the agent
instance into members and sets
the family_id inside the agent instance. A
counterpart method is that of death_member,
which removes the member from the fam-
ily (at the death event). Families also have
access to sum_balance, sum_savings and
RESEARCH REPORT
5
update_balance, which is a method that di-
vides a received amount equally among family
members.
Other
which
assign_house
two important pair of methods
updates
are
household_id, region_id,
house
and empty_house (which makes
itself,
into None, and also
all
class).
calls empty() within the house
and
address
set_address
region_id for all members within the
family.
those variables
updates
the
A relevant function consume() is hosted at
the level of the family, but it will be described
in the Goods and Services Market, section 4.6.1.
Finally, families calculate internally the pro-
portion of employed members as well as aver-
age utility.
all
include:
4.3 Households
Households are simple, fixed class objects that
contain mostly get and set methods. Fixed
variables
house_id, address,
size, region_id, quality,
of
which are set at
the beginning via the
generator module (section 6.3). Among those
methods, we
can list: get_family_id,
update_ownership, is_occupied,
add_family, and empty which changes
family_id into None.
variables
price,
so that both
family_id and ownership
the current family living on a given household
as well as its owner can change dynamically.
Non-fixed
include
These processes of changing ownership and
residence both happen at the real estate market
(section 4.6.3). Price is updated through the
relation size times quality times a region
Quality of Life Index (the latter is updated
at the Government module, according to the
amount of taxes collected).
class
represents
4.4 Government
individual
This
the
administrative-political unit
municipalities.
is generated based on
shapefile objects (that contain the geometric
features of
the boundaries), using OSGEO
objects.
basically,
the
It
It is the REGION where taxes are applied
and the REGION where taxes collected are
spent in order to increase the quality of life. The
class also contains self-calculated information
including firms and citizens within its territory.
In detail.
The generation of the Govern class object
needs to be read as ogr from OSGEO. This
implies that an address_envelope, which
is read from geometry().GetEnvelope() is
possible as an __init__ method. Addresses
also comes from geometry().
Apart from the geometric features, Govern
has a name which is extracted from the avail-
able fields associated in the shapefile using
GetFieldAsString(0) and an ID which
also comes from a field GetField(0).
Other variables include index, which is the
current value of the Quality of Life Index
3, treasure,
the current amount of avail-
able cash, the region_gdp, which is set at
statistics (section 7.1), its population (pop),
and a calculated total_commute.
The method to calculate the population re-
ceives all the families as input and for those
families in which region_id is the same of
the current region, the members of such family
are added up.
spending (that
The government
the
amount of collected taxes)
is directly con-
verted in increase in the quality of life in-
dex. In order to keep values nearly compat-
ible with empirical data, treasure is nor-
malized dividing it by population, and mul-
tiplying the result by a reducing value of
TREASURE_INTO_SERVICES = 0.0005.
is,
4.5 Firms
Firms are generated at
the
simulation and remain fixed afterwards. Thus,
are
firm_id, address and region_id
fixed variables. Besides, at the beginning of
the simulation, each firm has a small amount
of cash that is stored at total_balance.
the start of
However, the number of employed workers,
their quality, and the price of the firms' product
vary through time. Variables used to follow the
3. For simulation purposes, this index is initiated as the
value of the municipal Human Development Index in 2000
RESEARCH REPORT
6
firms dynamics include: last_qtr_balance,
updated every three months; profit 4
whether current total_balance is above
last_qtr_balance − amount_sold that
saves the cumulative number of products sold;
and amount_produced. The usual get and
set methods allow access to the values of the
firms' variables.
The firms also contain a dictionary of
employees with all of its employees at any
given time. As an object oriented programming
paradigm, the dictionary carries the full in-
stances of the employed agents, thus providing
access to all of the agents' information.
Prices can go up, down or remain the same.
Prices remain unchanged when either the num-
ber of employees or the inventory are zero.
above
When
firms
given
inventory
parameter
(Q)
(K)
is
then the
(QUANTITY_TO_CHANGE_PRICES),
price of its product is reduced in proportion to
the MARKUP parameter. The opposite happens
when quantity goes below that same level,
and they increase proportional to the MARKUP
value.
a
a
(cid:26)pricet−1 ∗ (1 + markup),
pricet =
if Q < K
pricet−1 ∗ (1 − markup), otherwise
(1)
Sale is the method that operationalizes the
goods market. It is detailed in section 4.6.1.
Other methods pertaining the functioning of
the firm are related to employees' business.
Those include add_employee, which not only
includes the agent in the employee dictionary,
but also passes the firm_id to the agent
(who can then store it and thus change his
status to employed). The obit method deletes
the employee from the firms' dictionary when
he or she passes (see demographics module).
When fire is called, the employee is also
removed from the dictionary of employees of
the firm and his firm_id is set back to None.
A relevant role of the firm is to control staff
at all times.
A method to add_employee makes sure
to register the agent within a firm's dictio-
nary employees. It also adds firm_id to the
4. Profit starts at 1 in order to guarantee an opening
positive value.
agent using method set_workplace. When
an employee dies and is withdrawn from the
simulation, he is deleted from the firm's team
with obit. When called, fire chooses a (sin-
gle) random employee and deletes him or her
from the staff and also sets workplace to None.
Other methods related to employment include
a check, is_worker and num_employees
(which returns the size of current staff).
Payment of employees is such that when the
firm is having profits and cashflow is positive,
profits are distributed as salary/bonuses us-
ing individual productivity criteria. Otherwise,
salary is based on productivity times a unitary
price.
Payment of employees is calculated in a two-
stage process. When the firm is distributing
profits, a get_wage_base is defined, such that
the base salary is the unit plus the percentage of
profits over current available cash. So that if the
company has a current cashflow of 1,000 and
profit was 200, then base salary is incremented
by a 20% margin at 1.20. Otherwise, base salary
remains at unit. Then, payment is made ex-
actly according to production. That is, each
employee produces the quantity equivalent to
qualification exponentiated by a productivity
parameter α.
Thus, salary is the unit (or unit plus profit
distribution margin) times days of production
(21) times qualification exponentiated by α.
salaryt =
prof itt−1
cashf lowt−1
∗ qualif icationα
(2)
4.5.1 Products
Another dictionary is inventory which con-
tains the list of the firm's products. At the
current implementation, there is only a single
product. However, code is implemented so that
other products could be easily added. In order
to save and control the IDs of possible prod-
ucts, there is also a product_index variable,
which starts at 0.
A special method, create_product,
is
used only once, before the simulation actu-
ally begins. Products contain information on
price, originally set at 1, and quantity, obvi-
ously, initially set at 0. They are also produced
as a class, stored at products module.
RESEARCH REPORT
7
J(cid:88)
update_product_quantity is actually
the method that implements the production
function. Production is based on the quantity
and quality of the employees. Given that any
employee has a given level of qualification, the
quantity produced is each employee's qualifi-
cation raised to an alpha parameter that indi-
cates the productivity level. Thus:
Qit =
jα
i
(3)
j=1
where Qi represents the quantity produced
by firm i, ji, each employee qualification, and
α, the productivity parameter. Production only
occurs if the firm has a positive number of
employees, positive monetary funds and at
least one registered product at inventory.
4.6 Markets
There are three markets at the current imple-
mentation of the model. The operation of goods
and services is distributed among methods
within the agents and the firms, and is monthly
called for every agent at time_iteration.
In turn, the labor market has a module of
its own, named communications indeed, that
market was thought out as an announcement
board where companies and potential employ-
ees meet. Finally, the real estate market also has
a module of its own, called housing_market.
4.6.1 Goods and services
The function consume is located within each
family with the following steps. First, the fam-
ily needs to have a positive amount of cash
in order to enter the market. If that amount
is below the unit, a random number between 0
and the available amount is chosen to be spent.
If, on the other hand, that amount is above
the unit, the chosen consumption is a percent-
age of the available amount of money (M).
Such percentage is given by a number drawn
from a beta distribution, so that the average is
given by the modeler chosen β parameter, i.e.
an average propensity to consume
(cid:26)U (0, 1) ∗ M,
Beta(1, 1−β
to spend ∼
if M < 1
β ) ∗ M, otherwise (4)
Once the decision of the amount to be spent
is made, the agent checks another parameter
called SIZE_MARKET. This parameter deter-
mines the number of firms the agent makes
contact before doing the consumption decision.
Third, among the contacted firms, the agent
chooses to buy either from the cheapest firm,
or the closest one. This decision is random and
equally weighted with probability 0.5.
Once those decisions have been made, con-
sumption is straightforward, given that prod-
ucts are homogeneous.
If the firm can sell the full agent demand,
according to his full amount to spend, the sale
occurs. Otherwise, the firm sells as much quan-
tity as it can provide, and returns monetary
change for the quantity it cannot provide.
At each month, the cash that the agent de-
cides not to spend goes to a savings account.
This account enables the family to enter the real
estate market.
Sales are located within the firm agents.
A simple check is made to see whether the
amount available from the consumer is posi-
tive. Given the original structure that allowed
for more than one product, the amount to
spend is distributed equally among the prod-
ucts available. However, for this implementa-
tion, there is a single product available and the
entire amount is directed to such a product.
The first operation within the firm is to trans-
form the cash amount into product quantity,
given current price. The resulting quantity is
deduced from the firm's inventory and taxes
are deduced from the amount paid. The firm's
balance is increased by the amount bought, and
the treasure of the region where the firm is
located is increased by the corresponding tax
collected (that is, given the tax rate and the
amount spent). The firm registers the amount
sold for statistics purposes.
4.6.2 Labor
At the end of every month, the labor market
is initiated from time_iteration that calls
hire_fire and look_for_jobs:
the first
function fills in a list with the firms that are
hiring; the second one compiles a list of able
and currently unemployed workers.
RESEARCH REPORT
8
The entrance of the firm in the labor mar-
ket is determined by a parameter (set by the
modeler), at the parameters module, called
LABOUR_MARKET. The larger this parameter,
less often the firm enters the labor market. If it
is set, for instance, at 0.75, it means that, at each
month, the firm has a 25% chance of entering
the market on average, firms would then enter
the labor market once every four months.
At the labor market, the firm decides to hire
if its profits are positive (profits are calculated
based on cash flow, having last quarter as refer-
ence). Otherwise, the firm fires one employee.
That is, if the firm has a negative result since
last quarter and the strategy is to enter the
market, then the firm makes one employee
redundant. Decision on the employee is made
randomly. That process guarantees a dynamic
labor market.
On the side of likely work seekers, citizens
enter the market every month when they fulfill
all of the three following conditions: (a) being
currently unemployed, (b) being of age 16 or
older, (c) being active in the market (up to 69
years old).
The module communications, that makes
the matching among firms and employees,
has a class called Posting. There are two
lists, one of candidates and one of available
postings. They contain all the firms and cit-
izens. The method that makes the match is
assign_post.
First, both lists are sorted: the firms offering
posts are sorted from those paying the high-
est base salaries; and the agents looking for
jobs are sorted by years of qualification (and,
implicitly, productivity). Then, the firm paying
the highest salary chooses first: the match is
made randomly with a 0.5 probability for the
candidate living the closest to the firm or the
most qualified in the list. While there are firms
offering opened positions and candidates inter-
ested, the match goes on. Every month both
lists are emptied and the process restarts.
4.6.3 Real estate
The houses prices are updated according to
current Quality of Life Indexes (QLI) of their
respective regions (municipalities). Thus, ev-
ery time the municipality collects taxes and
spends in public goods, the offering prices of
households increase accordingly. In order to
make its value comparable to actual Human
Development Index for Municipalities, the QLI
is normalized by multiplying it by a parameter
(k) TREASURE_INTO_SERVICE (typically set at
0.0005). Further, QLI is weighted by the change
in population (N).
QLIt =
QLIt−1 ∗ Nt−1
Nt
treasure ∗ k
Nt
+
(5)
Once it is empty, the house enters the real
estate market. In turn, families entering the
market are chosen according to the parameter
PERCENTAGE_CHECK_NEW_LOCATIONS,
which is a percentage of
families.
the universe of
The price of each house is then updated,
according to its size and quality and region.
P riceit = size ∗ quality ∗ QLIt
(6)
The lists of houses available and interested
families are sorted according, respectively, to
house prices and the amount of families' sav-
ings. The first step is for the family to acquire a
new ownership. Families that own more than
one house decides, at the end of the real estate
market procedure, whether to move or not.
Starting from the family with the largest
amount, the family will buy the most expensive
house it can afford, if any.
Once the matching family-house has been
made, the savings of the buying family is de-
duced of the price paid and the family who
previously owned the house receives the same
amount. The actual transfer of registering of
house ownership is then made, followed by the
removal of the house from the sales list.
Once the family has successfully accom-
plished a purchase, it verifies if it is worth
moving or not by checking the quality of the
houses and the proportion of employed in the
house. Thus, there are three options:
1) if the family lives in the best house, but
all members of the house are unemployed
(or out of the market, such as children
and retired members), then the family
moves to the second best house;
RESEARCH REPORT
9
2) if the family does not live in the house
with the best quality but at least one
person is employed, the family moves
into the best house;
3) all other options maintain the family in
the same house.
Once the decision to move has been made,
the family empties the previous house and
register the new house, the new address and,
eventually, the new region.
4.7 Dmographics
Demographics is the module that controls
aging, dying and giving births of agents.
It
is implemented mainly as one function
check_demographics.
The first step is to check the current year
of the simulation (by construction, the first
year is 2000). All of the citizens age. Then,
according to their gender and month of birth,
a death probability, derived from official IBGE
data estimates, is evaluated and saved.
Females then evaluate a probability to give
birth if they are between 15 and 49 years old.
If so, a new agent, child with age zero, is gen-
erated and it is set to live within the mother's
family.
Then, a number is drawn to evaluate the
death of agents. When death occurs, agents
are withdrawn from the agents list, from their
families and from their firms. For statistical
purposes, they are added to a new list called
my_graveyard.
4.8 Spatial boundness
In the present model, distance is used as one
of the criterions for two decisions: when firms
choose candidates for employment; and when
citizens choose a firm in order to consume its
product.
Distance is also used to compute a regional
commuting distance, defined as the sum of the
commuting distances of all employed citizens
living in a given region.
the beginning of
Furthermore, space plays a role in the model
when, at
the simulation,
households, firms and citizens are spatially
allocated. This allocation is made considering
urban and rural proportion for each munici-
pality (see details at section 6.3).
5 RUNNING DIFFERENT SIMULATIONS
The model can run in four different ways:
5.1 Simple; 5.2 Sensitivity analysis; 5.3 Multi-
run; and, 5.4 Auto adjustment analysis. The
definition of the process of each run is based on
the actual_parameters module, where each
selection is made by means of a combination of
variables as logical values (True or False).
are
variables
These
sensitivity_choice,
multi_run_simulation, and
auto_adjust_sensitivity_test.
Despite all these possibilities, the model's
core is the same, but called to run in different
ways.
All destinations of outputs from SEAL model
are identified by the four different simulation
possibilities. Below is the scheme of logical
variables and the resulted simulation that is
run:
• All logical variables are False: Single run
• Only sensitivity_choice is True:
Sensitivity_analysis
• Only multi_run_simulation is True:
Multi_Run_Simulation
• Only auto_adjust_sensitivity_test
is True: Adjust_test
variables
5.1 Simple run or Test
When the idea is to run a single simulation (a
simple test), it is necessary to set all control
logical
(sensitivity_choice,
multi_run_simulation, and
auto_adjust_sensitivity_test)
the main module.
as False,
Automatically,
run
a simulation over the selected region and save
results accordingly.
The model will
save results based on
the saving option in the
the selection of
actual_parameters module, so, check care-
fully the output saving options before running
a simulation.
the SEAL model will
and call
5.2 Sensitivity analysis
In order to run a sensitivity_choice,
select
it as True on actual_parameters
from the module
call
and
run
the
RESEARCH REPORT
10
control_sensitivity_analysis.
This
module controls the multiple runs of the main
module. As such, the model keeps working as
a simple model, but called many times.
to
the
The
linked
varying
analysis
parameters
the model
sensitivity
all
is performed
the
over
economics of
(ALPHA, BETA,
QUANTITY_TO_CHANGE_PRICES, MARKUP,
LABOUR_MARKET, SIZE_MARKET,
CONSUMPTION_SATISFACTION,
PERCENTAGE_CHECK_NEW_LOCATION,
TAX_CONSUMPTION),
values
in between the limits of each one these
parameters. For example, for ALPHA [0.01, 1].
Thus, ALPHA will be divided by the
number of intervals defined by the user in
the
control_sensitivity_analysis
module,
parameter
number_of_test_parameter_values
variable as an integer value. Using the
same example of ALPHA, if it the parameter
is
number_of_test_parameter_values
defined as 6, the values for ALPHA will be:
0.01; 0.208; 0.406; 0.604; 0.802; 1.
the
set
on
However, note that, because of the number
of parameters (9) and the example of 6 possible
values for each one, a combinatory analysis of
10,077,696 possible combinations is generated.
This would take a long time to run. Therefore,
we chose to adopt a ceteris paribus in-
stance in which each parameter new value is
tested while the others remain fixed as default.
Therefore, when the user chooses the num-
the state True for sensi-
ber of
tivity analysis and runs the simulation call-
ing control_sensitivity_analysis.py,
the simulation will run all possible runs, as
described, and produce a dataset of results,
statistics and plots for each parameter. Anal-
ysis of the results may indicate how much
the results vary based on each different set of
parameter values.
intervals,
5.3 Multi-run
The Multi-run simulation is focused on iden-
tifying the consistence of the model, varying
the random numbers involved on setting the
decision-making processes. The idea is to repli-
cate the same model, varying only the random
numbers, so that we can isolate average con-
sistent results, from outlier runs. We aim at
reproducing a typical run along with pseudo
variance information on its results.
In
order
to
run
it
the
nec-
in
True
the
to
set
multi_run_simulation,
essary
actual_parameters module
and
control_multi_runs_simulations.py.
simulation
variable
calling
logical
run
the
the
to
is
In this module,
it
the number of
fine
number_of_runs.
is necessary to de-
simulations
to run
5.4 Auto-adjustment analysis
The option to use the Auto adjustment anal-
ysis aims at verifying the best combination of
parameters for the model to produce optimal
results. The process is based on three condi-
tions:
The subdivision of each parameter interval
based on a number of possibilities defined
by the user on interval_for_values and
the control_autoadjust_model module.
The pattern defined as a reference for the
model to run the auto adjustment process. The
parameters values that produce the maximum
GDP and the minimum GINI index at same
time in the last month of the simulation are
set as default.
The number of times the model will run
in order to approximate the best values of
parameters that
lead to the optimal result
on control_autoadjust_model,
is
specifically
variable
times_test_aproximations, as an integer
number.
setting
the
set
This structure was defined based on the
possible number of combinations for the four
most important economic parameters: ALPHA,
BETA, MARKUP, and TAX_CONSUMPTION.
In order to restrict a huge number of pos-
sibilities, a subdivision method for each pa-
rameter was adopted. Thus, in the beginning,
the model divides the interval in a number
of pieces. For example, for ALPHA and the
interval_for_values set as 5, the values
to be tested will be: 0.01, 0.2575, 0.505, 0.7525,
RESEARCH REPORT
11
1 (the other three parameters will remain con-
stant). Such a combination would produce a list
of 255 possibilities of parameters.
After the initial approximation, given the
respective values of GDP and GINI, the best
parameter interval is chosen and a subsequent
division of the parameter values is applied. For
example, if ALPHA was selected with values
0.505 and 0.7525 for the best combination of
GDP and GINI, then the model will produce
a new list of five values that lies within the
previously set boundaries: 0.505, 0.57, 0.63,
0.69, 0.75; and so on until the number defined
of times_test_aproximations. Using this
approach, the total combinations of model pro-
cess will be of 1,020 tests.
5.5 Public Policy Test application: example
A final alternative run is to set alternative0
in the parameters module as False. In such a
case, the model will run considering, for taxes
purposes, all regions within each chosen Popu-
lation Concentration Area (ACP) [11] as a sin-
gle region. It is possible to run alternative0
and multi_run simulations together. In such
case, the runs alternate between True and
False alternative0.
newer
module
A
control_multi_ACPs_alternate_test
runs the alternative0 test automatically
and plots the differences and median of some
indicators for both cases, municipalities as a
single entity and municipalities as they are
presently.
6 DATA INPUT AND REQUIREMENTS
All dataset used to run the model are available
on the internet and come from official Brazil-
ian Agencies websites, especially The Brazilian
Institute of Geography and Statistics IBGE.
The data used to run the model is fundamen-
tally from the demographics Census of 2000
and 2010. The data used on the process are
divided in Agents, Firms and Government.
6.1 Agents data from IBGE
Agents data come from Demographics Census
of 2000 (as a start point):
• Tables of mortality probability by year,
age group and gender. The data was
interpolated
to
generalize it into every year data. Available
ftp://ftp.ibge.gov.br/Projecao da
on:
Populacao/Projecao da Populacao 2013/
tabuas de mortalidade xls.zip.
values
using
equal
• Tables
of
fecundity
probability
are
(between 2000 and
available by year
and
(between 15
2030), group age
and state. Available
49 years
ftp://ftp.ibge.gov.br/Projecao da
on:
Populacao/Projecao da Populacao 2013/
projecoes 2013 indicadores xls.zip.
old),
2010).
• IDHM (Human Development Index) by
IPEA/IBGE, Census
municipalities
This measure
(2000,
years
conditions
gives
the
for
on:
https://docs.google.com/gview?url=http:
//ivs.ipea.gov.br/ivs/data/rawData/
atlasivs dadosbrutos pt.xlsx
each municipality. Available
development
• Qualification by municipality and divided
in years of study. Available on: http:
//www.sidra.ibge.gov.br/bda/tabela/
listabl.asp?z=cd&o=32&i=P&c=2986
6.2 Firms data from DATAVIVA
• Dataset for number of firms by municipal-
ities, year (from 2002 to 2013), and produc-
tive sector. Provided by Ministry of Labor
and Employment (MTE) and available on:
http://dataviva.info/pt/data/.
6.3 Input spatial data
• The spatial dataset
uses
maps,
is also from IBGE.
the municipalities
on:
The model
boundaries
ftp://geoftp.ibge.gov.br/organizacao
do territorio/malhas territoriais/malhas
municipais/municipio 2010//.
available
(ACPs),
defined
each metropolitan
• The urban area boundary definition uses
Population Concentration
the High
IBGE
Areas
for
of
Brazil. These data can be found in:
ftp://geoftp.ibge.gov.br/cartas e mapas/
mapas do brasil/sociedade e economia/
areas urbanizadas//areas urbanizadas
by
region
RESEARCH REPORT
12
• The
do Brasil 2005 shapes.zip. These maps
represent the spatial delimitation of urban
areas (occurrence of populated areas).
last demographics data are
the
urban and rural populations at each
municipality, available on IBGE census
200 on IBGE SIDRA
website
database
2000
and 2010): http:
//www.sidra.ibge.gov.br/bda/tabela/
listabl.asp?z=cd&o=27&i=P&c=200.
(table
for
6.4 Instances generator
The generator module is the one responsible
for creating instances of all agents, families,
regions, houses and firms. It follows official
data and allocates them according to urban
and rural proportion, and municipalities pop-
ulation numbers.
Firstly, information on the proportion of ur-
ban inhabitants and number of firms by munic-
ipality are loaded. Then, information on 2000s
Municipal Human Development Index (HDM-
I) is loaded.
In order to create regions, the only necessary
data is their HDM-I and their respective shape-
file, read from 6.3. Such information is then
passed on to class Govern in Government.
All instances are saved in lists that are passed
from time_iteration to each process. Thus,
they are dynamic lists with changing instances.
For time saving purposes, agents for a given
spatial configuration and percentage of popu-
lation are saved in files that can later be loaded,
configuring its persistence.
Using the newly created regions, the main
function is create_all, which creates, within
each region, all other mentioned instances.
The first step is to select, from population
data for 2000, the number of inhabitants by
age group and gender for each municipality.
This procedure enables the successive creation
of agents for each age group and, alternately,
for each gender.
Additionally, each agent gets: a qualification
level assigned probabilistically, given munic-
ipal data of 2000; a random age that falls
within his group age; a small random amount
of money; and unique IDs.
The number of families created in each re-
gion is proportional to the number of inhab-
itants with an average given by parameter
MEMBERS_PER_FAMILY (typically set at 2.5 cit-
izens per family).
Houses are above the number of families,
given by parameter HOUSE_VACANCY (typi-
cally set at 10%).
The necessary additional
The number of firms, as stated before, are
given by reading of actual data by municipality.
are
functions
create_family, create_household
and create_firm; allocate_to_family
and allocate_to_households; and
get_random_point_in_polygon.
are
long
then
For
filled
as
Create_family provides empty family
using
vessels, which
allocate_to_family.
as
there are agents to be allocated into families
for a given municipality, one family and one
agent are drawn from the available list and
the match is made. There is one check to see
if the agent does not belong to any family
before allocation. That procedure ensures that
the number of agents per family is random
and the defined proportion is only valid for
the total numbers per municipality.
By construction, the number of families is
always smaller than the number of houses
available. Hence, all
families that have a
positive number of members are allocated
randomly to an available household using
allocate_to_households. In such a pro-
cess which of course happens only before the
simulation actually begins the family gets the
ownership of that first house they are allocated
to. The remaining houses that are empty are
then subsequently distributed among all the
families for a given municipality.
That implies that, on average, 10% of the
families own their own house plus one other
empty house that they may make available at
the real estate market. Nothing withholds the
possibility that luck will provide one family
with two or three empty houses.
The create_household function is a bit
more complex as it needs to abide to some
spatial restrictions. Given the urban or rural
proportion of the municipality, some of the
houses will be set on rural areas (according
RESEARCH REPORT
13
to official urban municipal legislation) or on
urban areas. That proportion is probabilistic
and we do not guarantee that a specific fixed
proportion will be urban or rural.
into
of
guarantees
that
falls
areas
function
address
rural
that
the
correct
the map is
The
actual
urban or
get_random_point_in_polygon.
uses
This
ogr.Geometry,
and
geometry().Contains
AddPoint
within a given shapefile previously divided
into urban and rural areas so that the allocation
is correctly determined.
function
Overall, besides address, the household is
also given a random size drawn from the in-
terval (20, 120), a fixed quality, from one to
four and an endogenous price value given by
Equation 6. Throughout the simulation, size
and quality will remain the same, whereas I
will be regularly updated, depending on the
economic dynamics of the municipality.
end,
In
four
my_families and my_firms. 5.
create_all returns
my_agents, my_houses,
the
lists:
6.5 Parameters and settings
Parameters module contains not only all the
parameters of the simulation, but also a lot of
settings and decisions. We will describe them
below, in order of appearance.
The
first
parameters
is:
PERCENTAGE_ACTUAL_POP and it establishes
the percentage of the population that is going
to be used in the simulation. A typical value is
0.01. However, the higher the proportion, the
longest the time to complete the simulation.
Simple tests can be run with 0.0001.
The
second parameter
a decision:
SIMPLIFY_POP_EVOLUTION and it accepts
Boolean values
(True or False). When
True, the generation of agents is made using
age groups. When False, the actual number
of inhabitants by each specific age is used
in the simulation. However, given smaller
municipalities and smaller populations,
it
becomes inadequate to run a simulation of 1%
of the population of 54-year-olds, for example,
is
5. These objects are then saved using pickles so that in a
subsequent run they can be uploaded and save running time
when there are only 27 of those. That way, it
is easier to reduce the population in bundles
of age groups for small municipalities.
LIST_NEW_AGE_GROUPS insures that
the
age groups are malleable. The modeler may
define the superior limit of the groups. So far,
we have been using [6, 12, 17, 25, 35, 45, 65,
100].
In the sequence,
Next comes the decision on region of study.
processing_states can be any one of
Brazilian 27 states or any combination of them,
using the two capital letters of its name, such
as 'MG' or 'SP'. To select them all, choose 'BR'.
is possible to choose
which Population Concentration Areas (ACPs),
established by IBGE in 2005 within each state.
ACPs approximate real conurbated metropoli-
tan areas [11]. Thus, they exclude rural, de-
tached municipalities that may be officially part
of metropolitan areas but that are not strongly
integrated.
it
Possible ACPs include:
AM, Manaus; PA Bel´em; AP Macap´a; MA
Sao Lu´ıs, Teresina; PI Teresina; CE Fortaleza,
Juazeiro do Norte/Crato/ Barbalha; RN Natal;
PB Joao Pessoa, Campina Grande; PE Recife,
Petrolina/Juazeiro; AL Macei´o; SE Aracaju;
BA Salvador, Feira de Santana, Ilh´eus/Itabuna,
Petrolina/Juazeiro; MG Belo Horizonte, Juiz
de Fora, Ipatinga, Uberlandia; ES Vit´oria; RJ
Volta Redonda/Barra Mansa, Rio de Janeiro,
Campos dos Goytacazes; SP Sao Paulo, Camp-
inas, Sorocaba, Sao Jos do Rio Preto, San-
Jundia, Sao Jos dos Campos, Ribeirao
tos,
Preto; PR Curitiba, Londrina, Maring´a; SC
Joinville, Florian´ıpolis; RS Porto Alegre, Novo
Hamburgo/Sao Leopoldo, Caxias do Sul,
Pelotas/Rio Grande; MS Campo Grande; MT
Cuiab´a; GO Goiania, Bras´ılia; DF Bras´ılia
According to available data, a starting year
has to be selected. For the case of Brazil, the
starting data, based on Census data, is the year
2000.
YEAR_TO_START
The next selection is the folder where to
storage output. It is essential for the running
of the program and it has to be set in any new
machine.
OUTPUT_PATH
Then, a number of settings related to saving,
RESEARCH REPORT
14
is
output, plotting and type of simulation have to
be completed with either False or True.
choice
Another
important
run with
merged/changed
actual municipalities
is whether
to
or
with
municipalities.
indicates
that
alternative0
boundaries are kept as is. When False,
municipalities
and
collecting of taxes thus benefits all the citizens
within a metropolitan region equally.
are merged together
True
sensitivity_choice refers to the sensi-
tivity analysis (section 5.2); the same is true
for multi_run_simulation (section 5.3) and
auto_ajdust_sensitivity_test (section
5.4). As said before, a regular run should have
all these parameters set to False.
keep_random_seed is necessary given the
fact that random module is used in a number
of different modules. In order to guarantee
that all random numbers are drawn from the
same sequence, when needed, a fixed_seed
= random.Random() is in place.
Some of the settings/control are related to
following the simulation as it unfolds, such
as: print_statistics_and_results and
show_plots_of_each_simulation.
option
Another
is
time_to_be_eliminated, which is related
to the amount of data that will enter the plots.
The percentage determines the portion to be
left out. For example, 0.2 will leave the first
20% of the data out of the plots.
Finally,
a
savings
number
save_plots_figures,
of
options:
save_agents_data,
save_agents_data_monthly,
save_agents_data_quarterly,
save_agents_data_annually, and
create_csv_files.
The module then initiates the actual sim-
ulation running parameters. Those include:
TOTAL_DAYS, measured in 21 working days a
month, meaning that 5.040 days amount to 20
years.
ALPHA, which is the productivity factor-
decaying exponent and can vary from 0 to 1,
being 1 the most productive worker.
BETA, which is the consumption exponent
and determines the amount the family decides
to consume/save. The larger the exponent,
more consumption and less savings.
QUANTITY_TO_CHANGE_PRICES, which is
the number threshold that firms check against
their stocks in order to decide whether demand
is high (below the quantity) and thus prices
should rise.
MARKUP, which is the amount the firm ap-
plies when varying prices.
LABOUR_MARKET, which is the frequency
that the firm makes decisions about entering
the market. The higher the parameter, less of-
ten.
SIZE_MARKET, the number of firms that the
consumer check for prices and distances when
deciding where to shop.
CONSUMPTION_SATISFACTION, this is just
a parameter that scales down consumption in
order to keep track of accumulated satisfaction
given by consumption.
PERCENTAGE_CHECK_NEW_LOCATION,
the
percentage of total families that enter the real
estate market each month.
TAX_CONSUMPTION, the general consump-
tion tax applied to all sales and collected by
municipal government.
The process of rearrange population data is
performed after the definition of parameters
values. Can be chose to perform a simplifi-
cation of population when the percentage of
population chosen are too small. In the end the
parameters module save the files for control
process and generate the output *.txt data.
6.6 Necessary Python libraries
We run the model in Python 3.4.4.
The following Python libraries are neces-
sary:
from numpy import median
from operator
import methodcaller, attrgetter
from osgeo import ogr
from pandas import read_csv
from timeit import default_timer
ggplot
glob
itertools
matplotlib.pyplot
numpy
os
RESEARCH REPORT
pandas
pickle
random
subprocess
sys
A simpler way to install the necessary envi-
ronment is:
Install
Anaconda3-2.3.0-Windows-x86_64
Then type in Terminal
conda install python=3.4.4
conda install pandas=0.18.0
conda install numpy=1.10.4
conda install -c ioos
geopandas=0.2.0.dev0
pip install ggplot==0.6.8
7 OUTPUTS
Output module is called at the end of the
month and it is divided into general statis-
tics call_statistics and regional statis-
tics call_regional_stats. The former calls
my_statistics and writes TXT outputs. The
information it contains include:
actual_month
price_index
GDP_index
unemployment
average_workers
families_wealth
firms_wealth
firms_profit
GINI_index
average_utility
The regional statistics include:
actual_month
region_id
commuting
get_pop
get_gdp
regional_gini
regional_unemployment
index
GDP_region_capita
15
Some of
the methods are detailed in
the next subsection. Finally, output mod-
ule also includes save_agents_data and
sum_region_gdp functions.
7.1 Statistics
The statistics class is just a bundle of func-
tions together without a permanent instance
of data. Thus, every time Statistics() is
called, it is initiated anew. The functions in-
clude average price of the firms, regional GDP,
based on FIRMS' revenues, GDP per capita, un-
employment, families' wealth, GINI, regional
GINI and commuting information.
average_price method checks for every
firm in my_firms the quantity and price of
each product (currently just one). Then it cal-
culates and returns the average.
calculate_region_GDP goes through the
firms in a given region and adds up the cumu-
lative sold value, inclusive of taxes.
update_GDP_capita also goes through the
firms in the region, adding up cumulative
value sold and then dividing it by current
population.
calculate_unemployment goes through
all agents checking whether they are minor
or retired and then classifying those left as
employed or unemployed. Then it updates un-
employment rate as the ration between unem-
ployed and the total workforce (employed plus
unemployed).
calculate_families_wealth
returns
the sum of all families wealth.
calculate_GINI
calculate_utility returns the average
of families' average utility. That is, first, the
average within each family is calculated and
then we calculate the average of such values.
given
the
calculated upon
typical GINI
is,
this very families average utility. That
the inequality present
the GINI
among
consumption.
calculate_regional_GINI is made exactly
the same way but considering only the families
residing in a given region (municipality).
is
coefficient
cumulative
is in fact
families
by
update_commuting is calculated for every
member of each family in each region that
is_employed.
RESEARCH REPORT
16
In
7.2 Plotting and analysis
The model produces some data and graphs by
default in the plotting module and others in
accordance to users' choices. The plotting of the
model may use different modules.
the
all models,
plotting module
produces the same set of plots from the
temp_general_%s.txt data. This output
data represents the aggregated measures of
all regions in the simulation. These measures
are: Price index, GDP, Unemployment,
Average of number of workers by
firms, Families wealth, Firms
wealth, Firms profit, GINI index,
these
Average utility. For each one of
is produced to allow
measures one plot
for a general understanding of the patterns
observed in the running model.
represent
The same module produces
the firms'
plots using data from temp_firm%s.txt.
They
Amount produced,
Price, Mean and Median of Number
employees. They are used to understand the
firms' general behavior.
the
(one
value
for
each municipality).
the plotting module produces
stats
each
in
The
are: Commuting, Population,
Finally,
regional
measure
measures
GDP of region, GINI by region,
Unemployment by municipality,
Quality life index by municipality,
GDP per capita by municipality.
When the model is in a Multi Run simula-
tion mode, the module plot_multi_run.py
use the values of temp_general_%s.txt and
stack up each measure (i.e. Unemployment)
of all simulations in the same graph. The
same process is observed for regional data and
agents' data. They are used to produce plots
considering each municipality.
7.3 Output data
Output data represents the minimal data
necessary to analyze
the
model. Therefore, all values of simulation
for general and regional data are saved by
default. These data contain the economics and
social measures
selected
and
region
the whole
temp_general_%s.txt
results of
the
for
in
regional measures
the
the
of
temp_regional_stats%s.txt.
social
(by municipality)
economic measures
and
The
choose
the
the
and the
can
data
user
agents'
scale
files:
of
temp_firm%s.txt,
temp_agent%s.txt
time
and temp_house%s.txt,
such data: monthly:
frequency to
quarterly,
save_agents_data_monthly,
or
save_agents_data_quarterly
save_agents_data_annually.
annually,
Such
produce
aggregated or simple plots (in accordance
to the type of model running).
information
used
save
to
is
7.3.1 Details of TXTs files
When the option to save agents data is set to
True, all data is saved and 5 *.TXT files are
created.
The files are: 'agent, firm, general, house' and
'regional'. Family information is also available
in the house file. They contain no headers,
delimiter = , and decimal = .
AGEN T contain the following columns:
month region_id gender long lat id
age qualification firm_id family_id
utility distance
F IRM contain the
following columns:
month firm_id region_id long lat
total_balance$ number_employees
total_quantity_in_stock
amount_produced price
the
contain
GEN ERAL
following
columns: month price_index gdp_index
unemployment average_workers
families_wealth families_savings
firms_wealth firms_profit
gini_index average_utility
the
HOU SE
contain
following
month house_id long lat
columns:
house_size house_price family_id
family_savings region_id
the
REGION AL
contain
following
month region_id commuting
columns:
pop gdp_region regional_gini
regional_unemployment qli_index
gdp_percapta treasure
The file's names always include the follow-
ing parameters: "None", alternative0,
PERIODICITY_SAVE_DATA, TOTAL_DAYS,
RESEARCH REPORT
17
total_pop, SIZE_MARKET, ALPHA,
BETA, QUANTITY_TO_CHANGE_PRICES,
MARKUP, LABOUR_MARKET,
CONSUMPTION_SATISFACTION,
PERCENTAGE_CHECK_NEW_LOCATION,
TAX_CONSUMPTION.
and
8 DESIGN CONCEPTS
There are two main drives of the agent-based
model here presented. Firstly, it is a restricted
adherence to investigate public policies within
a spatial continuum. In order to achieve that,
there are locational Cartesian attributes for
houses and for firms with agents commut-
ing the distance. Further, region space
or-
ganized in municipalities contain both their
actual geodesic boundaries, which determines
their action space, but also comprehend the
urban rural divide, which enables urban con-
centration and rural sparseness. Thus, distance
explicitly is considered within interaction in
the labor market and consumption market.
Further, there is a real estate market that is
very much dependent on how well the firms
in the region are performing. Stronger firms
with higher sales increase taxes, which are
directly applied to improving estates quality
and directly, prices.
Secondly, the model is proposed in a rather
flexible and generic way so that it can be
constituted as a framework for later public
policy analysis. Agents, firms, markets, space,
activities are constructed so that newer research
questions can be easily adapted into the model,
which would provide rapid answering. As it
is, for instance, results can be provide for an
alteration in a specific tax or a change in a
firms strategy to enter the labor market, or a
municipality fusion or, yet, in the influence of
those changes in mobility patterns, for instance.
Apart from those two specific design con-
cepts, obviously, the model follows the typ-
ical agent-based specifications, that is: agent
heterogeneity, decision-making based on local
variables, and emergence.
Specifically, we can say that
the agents
in this model adapt in the sense that they
change behavior giving observed variables at
their environment. For example, firms change
decision-making (and salaries paid) depending
on whether they have available balance and
profits. Agents decide whether to move into a
better quality house depending on the propor-
tion of employed adults in the family.
The model is not intended as a forecasting
tool at this stage. The focus is more on setting
a framework,
learning about complex inter-
actions among different factors across a large
spectrum and gaining magnitude of changes
insights.
Interactions are plenty. Specifically, they hap-
pen in the three proposed markets and within
the families (co-responsible for consumption,
savings and utility). Indirectly, there is inter-
action among those agents within the same
region.
Stochasticity is present in the model with the
use of drawing and random decision-making
in a number of procedures. However, a control
variable is present that enables the reproduc-
tion of the same result. Typically, results should
be presented as summaries of a number of
simulated runs.
A collective entity that is relevant for the
model is that of the family (see section 4.2).
9 FINAL CONSIDERATIONS
This is still a model in progress. However, at
this stage, we believe we have most of the
framework as intended.
Thus, this research project reveals from a
code perspective nearly all aspects of
the
model. That includes a description of purpose
and concept, model scheduling, detailed infor-
mation and methods on all agents and classes
used, the market interactions, data generated
and output produced. It also includes informa-
tion on running the model for multiple runs
and explicitly using it as a policy example
(section 5.5).
However, we have yet to publish major re-
search studies using such framework, except
for [7], [8]. This is exactly the next step: apply
the framework to public policy research ques-
tions.
Finally, we would like to add that the current
working team is short and we would be willing
to cooperate with other fellow scientists.
RESEARCH REPORT
18
ACKNOWLEDGMENT
The authors would like to thank Institute for
Applied Economic Research and especially the
Department of Innovation and Infrastructure
(DISET) for their continuing support.
[3]
[4]
[5]
[2] U. Wilensky
REFERENCES
[1] C. M. Macal, "Everything you need to know about
agent-based modelling and simulation," Journal
of
Simulation, vol. 10, no. 2, pp. 144 -- 156, May 2016.
[Online]. Available: http://www.palgrave-journals.com/
doifinder/10.1057/jos.2016.7
Introduction
and W.
to Agent-Based Modeling.
2015.
[Online]. Available: https://mitpress.mit.edu/books/
introduction-agent-based-modeling
J. M. Epstein, Generative Social Science: Studies in Agent-
Based Computational Modeling. Princeton University Press,
Dec. 2011.
J. H. Miller and S. E. Page, Complex adaptive systems.
Princeton University Press, Mar. 2007.
J. M. Epstein and R. Axtell, Growing artificial societies:
social science from the bottom up. Cambridge, MA: Brook-
ings/MIT Press, 1996.
Rand, An
The MIT Press,
[6] L. Tesfatsion, "Agent-Based Computational Economics: A
Constructive Approach to Economic Theory," in Handbook
of Computational Economics. Elsevier, 2006, vol. 2, pp. 831 --
880.
[7] B. A. Furtado and I. D. R. Eberhardt, "A simple
agent-based spatial model of the economy:
tools for
policy," Journal of Artificial Societies & Social Simulation,
2016, arXiv: 1510.04967. [Online]. Available: http://arxiv.
org/abs/1510.04967
[8] -- -- , "Da mobilidade metropolitana vinculada a econo-
mia : analise a partir de um modelo baseado em agentes,"
RADAR, vol. 43, pp. 9 -- 20, Feb. 2016. [Online]. Available:
http://repositorio.ipea.gov.br/handle/11058/6050
[9] V. Grimm, U. Berger, D. L. DeAngelis, J. G. Polhill,
J. Giske, and S. F. Railsback, "The ODD protocol:
a review and first update," Ecological Modelling, vol.
221, no. 23, pp. 2760 -- 2768, 2010.
[Online]. Avail-
able: http://www.sciencedirect.com/science/article/pii/
S030438001000414X
[10] B. A. Furtado, P. A. M. Sakowski, and M. H. Tovolli,
Modeling complex systems for public policies. Brasilia: IPEA,
Instituto de Pesquisa Economica Aplicada, 2015.
[11] IBGE. Ministerio do Planejamento, Arranjos populacionais
e Concentracoes urbanas do Brasil. Rio de Janeiro: IBGE,
2015.
|
0807.2028 | 4 | 0807 | 2009-03-12T21:43:21 | On Krause's multi-agent consensus model with state-dependent connectivity (Extended version) | [
"cs.MA"
] | We study a model of opinion dynamics introduced by Krause: each agent has an opinion represented by a real number, and updates its opinion by averaging all agent opinions that differ from its own by less than 1. We give a new proof of convergence into clusters of agents, with all agents in the same cluster holding the same opinion. We then introduce a particular notion of equilibrium stability and provide lower bounds on the inter-cluster distances at a stable equilibrium. To better understand the behavior of the system when the number of agents is large, we also introduce and study a variant involving a continuum of agents, obtaining partial convergence results and lower bounds on inter-cluster distances, under some mild assumptions. | cs.MA | cs | On Krause’s multi-agent consensus model with
state-dependent connectivity (Extended version)
Vincent D. Blondel, Julien M. Hendrickx and John N. Tsitsiklis
9
0
0
2
r
a
M
2
1
]
A
M
.
s
c
[
4
v
8
2
0
2
.
7
0
8
0
:
v
i
X
r
a
Abstract—We study a model of opinion dynamics introduced
by Krause: each agent has an opinion represented by a real
number, and updates its opinion by averaging all agent opinions
that differ from its own by less than 1. We give a new proof of
convergence into clusters of agents, with all agents in the same
cluster holding the same opinion. We then introduce a particular
notion of equilibrium stability and provide lower bounds on
the inter-cluster distances at a stable equilibrium. To better
understand the behavior of the system when the number of
agents is large, we also introduce and study a variant involving
a continuum of agents, obtaining partial convergence results
and lower bounds on inter-cluster distances, under some mild
assumptions.
Keywords: Multi-agent system, consensus, opinion dynam-
ics, decentralized control.
I. INTRODUCTION
There has been an increasing interest in recent years in the
study of multi-agent systems where agents interact according
to simple local rules, resulting in a possibly coordinated global
behavior. In a prominent paradigm dating back to [10] and
[28], each agent maintains a value which it updates by taking a
linear, and usually convex combination of other agents’ values;
see e.g., [5], [16], [17], [25], [28], and [26], [27] for surveys.
The interactions between agents are generally not all-to-all,
but are described by an interconnection topology. In some
applications, this topology is fixed, but several studies consider
the more intriguing case of changing topologies. For example,
in Vicsek’s swarming model [30], animals are modeled as
agents that move on the two-dimensional plane. All agents
have the same speed but possibly different headings, and
at each time-step they update their headings by averaging
the headings of those agents that are sufficiently close to
them. When the topology depends on the combination of the
agent states, as in Vicsek’s model, an analysis that takes this
dependence into account can be difficult. For this reason, the
sequence of topologies is often treated as exogenous (see e.g.
[4], [17], [25]), with a few notable exceptions [7], [8], [18].
This research was supported by the National Science Foundation under
grant ECCS-0701623, by the Concerted Research Action (ARC) “Large
Graphs and Networks” of the French Community of Belgium and by the
Belgian Programme on Interuniversity Attraction Poles initiated by the Bel-
gian Federal Science Policy Office. The scientific responsibility rests with
its authors. Julien Hendrickx holds postdoctoral fellowships from the F.R.S.-
FNRS (Belgian Fund for Scientific Research) and the B.A.E.F. (Belgian
American Education Foundation); a part of this research was conducted when
he was with the Universit´e catholique de Louvain.
V. D. Blondel is with Department of Mathematical Engineering, Universit´e
catholique de Louvain, Avenue Georges Lemaitre 4, B-1348 Louvain-la-
Neuve, Belgium; [email protected]. J. M. Hen-
drickx and J. N. Tsitsiklis are with the Laboratory for Information and
Decision Systems, Massachusetts Institute of Technology, Cambridge, MA
02139, USA; [email protected], [email protected].
xi(t + 1) = Pj:xi(t)−xj (t)<1 xj(t)
Pj:xi(t)−xj (t)<1 1
.
(1)
For instance, the authors of [7] consider a variation of the
model studied in [17], in which communications are all-to-
all, but with the relative importance given by one agent to
another weighted by the distance separating the agents. They
provide conditions under which the agent headings converge
to a common value and the distance between any two agents
converges to a constant. The same authors relax the all-to-
all assumption in [8], and study communications restricted to
arbitrarily changing but connected topologies.
We consider here a simple discrete-time system involving
endogenously changing topologies, and analyze it while taking
explicitly into account the dependence of the topology on the
system state. The discrete-agent model is as follows. There are
n agents, and every agent i (i = 1, . . . , n), maintains a real
value xi. These values are synchronously updated according
to
Two agents i, j for which xi(t) − xj(t) < 1 are said to
be neighbors or connected (at time t). Note that with this
definition, an agent is always its own neighbor. Thus, in this
model, each agent updates its value by computing the average
of the values of its neighbors. In the sequel, we usually refer to
the agent values as “opinions,” and sometimes as “positions.”
The model (1) was introduced by Krause [19] to capture
the dynamics of opinion formation. Values represent opinions
on some subject, and an agent considers another agent as
“reasonable” if their opinions differ by less than 11. Each
agent thus updates its opinion by computing the average of the
opinions it finds “reasonable”. This system is also sometimes
referred to as the Hegselmann-Krause model, following [14].
It has been abundantly studied in the literature [19], [20], [22],
[23], and displays some peculiar properties that have remained
unexplained. For example, it has been experimentally observed
that opinions initially uniformly distributed on an interval tend
to converge to clusters of opinions separated by a distance
slightly larger than 2, as shown in Figure 1. In contrast,
presently available results can only prove convergence to
clusters separated by at least 1. An explanation of the inter-
cluster distances observed for this system, or a proof of a
nontrivial lower bound is not available.
Inter-cluster distances larger than the interaction radius
(which in our case was set to 1) have also been observed
by Deffuant et al. [9] for a related stochastic model, often
referred to as the Deffuant-Weisbuch model. In that model,
1In Krause’s initial formulation, all opinions belong to [0, 1], and an agent
considers another one as reasonable if their opinions differ by less than a
pre-defined parameter ǫ.
x
10
9
8
7
6
5
4
3
2
1
0
0
5
t
(a)
10
15
x
10
9
8
7
6
5
4
3
2
1
0
6
8
10
0
2
4
t
(b)
Figure 1. Time evolution of 1000 agent opinions, according to the model (1).
Initial opinions are either uniformly spaced (case (a)) or chosen at random
(case (b)), on an interval of length 10. In both cases, opinions converge to
limiting values (“clusters”) that are separated from each other by much more
than the interaction radius, which was set to 1.
two randomly selected agents update their opinions at any
given time step. If their opinions differ by more than a
certain threshold, their opinions remain unchanged; otherwise,
each agent moves to a new opinion which is a weighted
average of its previous opinion and that of the other agent.
Thus, the Krause and Deffuant-Weisbuch models rely on the
same idea of bounded confidence, but differ because one is
stochastic while the other is deterministic. Besides, Krause’s
model involves simultaneous interactions between potentially
all agents, while the interactions in the Deffuant-Weisbuch
model are pairwise. Despite these differences, the behavior of
these two systems is similar, including inter-cluster distances
significantly larger than the interaction radius. The behavior
of the Deffuant-Wesibuch model — and in particular the
final positions of the clusters — has also been studied by
considering a continuous density approximating the discrete
distribution of agents, and examining the partial differential
equation describing the evolution of this density [2], [3].
Other models, involving either discrete or continuous time,
and finitely or infinitely many agents, have also been proposed
[1], [12], [29]. For a survey, see for example [24].
The model that we consider also has similarities with certain
rendezvous algorithms (see, e.g., [21]) in which the objective is
to have all agents meet at a single point. Agents are considered
neighbors if their positions are within a given radius R. The
update rules satisfy two conditions. First, when an agent
moves, its new position is a convex combination of its previous
position and the positions of its neighbors. Second, if two
agents are neighbors, they remain neighbors after updating
their positions. This ensures that an initially connected set of
agents is never split into smaller groups, so that all agents can
indeed converge to the same point.
In this paper, we start with a simple convergence proof
for the model (1). We then introduce a particular notion of
equilibrium stability, involving a robustness requirement when
an equilibrium is perturbed by introducing an additional agent,
and prove that an equilibrium is stable if and only if all
inter-cluster distances are above a certain nontrivial lower
bound. We observe experimentally that
the probability of
converging to a stable equilibrium increases with the number
of agents. To better understand the case of a large numbers
of agents, we introduce and study a variation of the model,
which involves a continuum of agents (the “continuous-agent”
2
introduced in [6],
Our continuous-agent model, first
model). We give partial convergence results and provide a
lower bound on the inter-cluster distances at equilibrium,
under some regularity assumptions. We also show that for a
large number of discrete agents, the behavior of the discrete-
agent model indeed approximates the continuous-agent model.
is
obtained by indexing the agents by a real number instead of an
integer. It is equivalent to the so-called “discrete-time density
based Hegselmann-Krause model” proposed independently in
[24], which is in turn similar to a model presented in [12] in
a continuous-time setup. Furthermore, our model can also be
viewed as the limit, as the number of discrete opinions tends
to infinity, of the “interactive Markov chain model” introduced
by Lorenz [23]; in the latter model, there is a continuous
distribution of agents, but the opinions take values in a discrete
set.
We provide an analysis of the discrete-agent model (1) in
Section II. We then consider the continuous-agent model in
Section III. We study the relation between these two models
in Section IV, and we end with concluding remarks and open
questions, in Section V.
II. THE DISCRETE-AGENT MODEL
A. Basic properties and convergence
We begin with a presentation of certain basic properties of
the discrete-agent model (1), most of which have already been
proved in [14], [20], [22].
Proposition 1 (Lemma 2 in [20]): Let
(x(t)) be a se-
quence of vectors in ℜn evolving according to (1). The order
of opinions is preserved: if xi(0) ≤ xj(0), then xi(t) ≤ xj(t)
for all t.
Proof: We use induction. Suppose that xi(t) ≤ xj(t). Let
Ni(t) be the set of agents connected to i and not to j, Nj(t)
the set of agents connected to j and not to i, and Nij(t) the
set of agents connected to both i and j, at time t. We assume
here that these sets are nonempty, but our argument can easily
be adapted if some of them are empty. For any k1 ∈ Ni(t),
k2 ∈ Nij(t), and k3 ∈ Nj(t), we have xk1 (t) ≤ xk2 (t) ≤
xk3 (t). Therefore, ¯xNi ≤ ¯xNij ≤ ¯xNj , where ¯xNi , ¯xNij , ¯xNj ,
respectively, is the average of xk(t) for k in the corresponding
set.
It follows from (1) that
and
xi(t + 1) =
Nij ¯xNij + Ni ¯xNi
Nij + Ni
≤ ¯xNij ,
xj(t + 1) =
Nij ¯xNij + Nj ¯xNj
Nij + Nj
≥ ¯xNij ,
where we use A to denote the cardinality of a set A.
In light of this result, we will assume in the sequel, without
loss of generality, that the initial opinions are sorted: if i < j
then xi(t) ≤ xj(t). The next Proposition follows immediately
from the definition of the model.
Proposition 2: Let (x(t)) be a sequence of vectors in ℜn
evolving according to (1), and such that x(0) is sorted,
i.e., if i < j, then xi(0) ≤ xj (0). The smallest opinion
x1 is nondecreasing with time, and the largest opinion xn
is nonincreasing with time. Moreover, if at some time the
distance between two consecutive agent opinions xi(t) and
xi+1(t) is larger than or equal to 1 it remains so for all
subsequent times t′ ≥ t, so that the system can then be
decomposed into two independent subsystems containing the
agents 1, . . . , i, and i + 1, . . . , n, respectively.
Note that unlike other related models as the Deffuant-
Weisbusch model [9] or the continuous-time model in [15], the
average of the opinions is not necessarily preserved, and the
“variance” (sum of squared differences from the average) may
occasionally increase. See [15] for examples with three and
eight agents respectively. The convergence of (1) has already
been established in the literature (see [11], [22]), and is also
easily deduced from the convergence results for the case of
exogenously determined connectivity sequences (see e.g., [5],
[16], [22], [25]), an approach that extends to the case of higher-
dimensional opinions. We present here a simple alternative
proof, which exploits the particular dynamics we are dealing
with.
Theorem 1: If x(t) evolves according to (1), then for every
in finite time. Moreover, for
i, xi(t) converges to a limit x∗
i
any i, j, we have either x∗
i = x∗
Proof: Since x(0) is assumed to be sorted, the opinion x1
is nondecreasing and bounded above by xn(0). As a result, it
1. Let p be the highest index for which
converges to a value x∗
xp converges to x∗
1.
We claim that if p < n, there is a time t such that xp+1(t)−
xp(t) ≥ 1. Suppose, to obtain a contradiction, that the claim
does not hold, i.e., that xp+1(t) − xp(t) is always smaller than
1. Fix some ǫ > 0 and a time after which the distance of xi
from x∗
1, for i = 1, . . . , p, is less than ǫ. Since xp+1 does not
converge to x∗
1, there is a further time at which xp+1 is larger
1 + δ for some δ > 0. For such a time t, xp(t + 1) is
than x∗
at least
i − x∗
j or (cid:12)(cid:12)x∗
j(cid:12)(cid:12) ≥ 1.
1
p + 1 p+1
Xi=1
xi(t)! ≥
1
p + 1
(p(x∗
1 − ǫ) + (x∗
1 + δ)) ,
which is larger than x∗
1 + ǫ if ǫ is chosen sufficiently
small. This however contradicts the requirement that xp remain
within ǫ from x∗
1. This contradiction shows that there exists
a time t at which xp+1(t) − xp(t) ≥ 1. Subsequent to that
time, using also Proposition 2, xp cannot increase and xp+1
cannot decrease, so that the inequality xp+1−xp ≥ 1 continues
to hold forever. In particular, agents 1, . . . , p will no more
interact with the remaining agents. Thus, if p < n, there
will be some finite time after which the agents p + 1, . . . , n
behave as an independent system, to which we can apply
the same argument. Continuing recursively, this establishes
the convergence of all opinions to limiting values that are
separated by at least 1.
It remains to prove that convergence takes place in finite
time. Consider the set of agents converging to a particular
limiting value. It follows from the argument above that there
is a time after which none of them is connected to any agent
outside that set. Moreover, since they converge to a common
value, they eventually get sufficiently close so that they are all
connected to each other. When this happens, they all compute
the same average, reach the same opinion at the next time
step, and keep this opinion for all subsequent times. Thus,
3
they converge in finite time. Finite time convergence for the
entire systems follows because the number of agents is finite.
We will refer to the limiting values to which opinions
converge as clusters. With some abuse of terminology, we
will also refer to a set of agents whose opinions converge
to a common value as a cluster.
It can be shown that the convergence time is bounded above
by some constant c(n) that depends only on n. On the other
hand, an upper bound that is independent of n is not possible,
even if all agent opinions lie in the interval [0, L] for a fixed L.
To see this, consider n agents, with n odd, one agent initially
placed at 1, and (n − 1)/2 agents initially placed at 0.1 and
1.9. All agents will converge to a single cluster at 1, but the
convergence time increases to infinity as n grows.
We note that the convergence result in Theorem 1 does
not hold if we consider the same model but with a countable
number of agents. Indeed, consider a countably infinite number
of agents, all with positive initial opinions. Let m(y) be the
number of agents having an initial opinion y. Suppose that α ∈
(1/2, 1), and consider an initial condition for which m(0) = 0,
m(α) = 1, m(α(k + 1)) = m(αk) + 3m(α(k − 1)) for every
integer k > 1, and m(y) = 0 for every other value of y. Then,
the update rule (1) implies that xi(t + 1) = xi(t) + α/2, for
every agent i and time t, and convergence fails to hold. A
countable number of agents also admits equilibria where the
limiting values are separated by less than 1. An example of
such an equilibrium is obtained by considering one agent at
every integer multiple of 1/2.
We also note that equilibria in which clusters are separated
by less than 1 become possible when opinions are elements
of a manifold, instead of the real line. For example, suppose
that opinions belong to [0, 2π) (identified with elements of the
unit circle), and that two agents are neighbors if and only if
xi − xj (mod 2π) < 1. If every agent updates its angle
by moving to the average of its neighbors’ angles, it can
be seen that an initial configuration with n agents located at
angles 2πk/n, k = 0, . . . , n − 1, is an equilibrium. Moreover,
more complex equilibria also exist. Convergence has been
experimentally observed for models of this type, but no proof
is available.
B. Experimental observations
Theorem 1 states that opinions converge to clusters sepa-
rated by at least 1. Since the smallest and largest opinions
are nondecreasing and nonincreasing, respectively, it follows
that opinions initially confined to an interval of length L can
converge to at most ⌈L⌉ + 1 clusters. It has however been
observed in the literature that the distances between clusters
are usually significantly larger than 1 (see [20], [23], and
Figure 1), resulting in a number of clusters that is significantly
smaller than the upper bound of ⌈L⌉ + 1. To further study this
phenomenon, we analyze below different experimental results,
similar to those in [23].
Figure 2 shows the dependence on L of the cluster number
and positions, for the case of a large number of agents
and initial opinions that are uniformly spaced on an interval
/
2
L
−
x
25
20
15
10
5
0
−5
−10
−15
−20
−25
0
5
10
15
20
25
L
30
35
40
45
50
Figure 2. Locations of the different clusters at equilibrium, as a function of
L, for 5000L agents whose initial opinions are uniformly spaced on [0, L],
represented in terms of their distance from L/2. The dashed lines correspond
to the endpoints 0 and L of the initial opinion distribution. Similar results
are obtained if the initial opinions are chosen at random, with a uniform
distribution.
of length L. Such incremental analyses also appear in the
literature for various similar systems [2], [13], [23], [24].
We see that the cluster positions tend to change with L in
a piecewise continuous (and sometimes linear) manner. The
discontinuities correspond to the emergence of new clusters, or
to the splitting of a cluster into two smaller ones. The number
of clusters tends to increase linearly with L, with a coefficient
slightly smaller than 1/2, corresponding to an inter-cluster
distance slightly larger than 2. Note however that this evolution
is more complex than it may appear: Irregularities in the
distance between clusters and in their weights can be observed
for growing L, as already noted in [23]. Besides, for larger
scale simulations (L = 1000, n = 106), a small proportion
of clusters take much larger or much smaller weights than
the others, and some inter-cluster distances are as large as
4 or as small as 1.5. These irregularities could be inherent
to the model, but may also be the result of the particular
discretization chosen or of the accumulation of numerical
errors in a discontinuous system.
Because no nontrivial lower bound is available to explain
the observed inter-cluster distances in Krause’s model, we start
with three observations that can lead to some partial under-
standing. In fact, the last observation will lead us to a formal
stability analysis, to be developed in the next subsection.
(a) We observe from Figure 2 that the minimal value of
L that leads to multiple clusters is approximately 5.1, while
Theorem 1 only requires that this value be at least 1. This mo-
tivates us to address the question of whether a more accurate
bound can be derived analytically. Suppose that there is an odd
number of agents whose initial opinions are uniformly spaced
on [0, L]. An explicit calculation shows that all opinions
2 − O( 1
belong to an interval [ 1
n )] after one
iteration, and to an interval [ 11
12 + O( 1
n )]
after two iterations. Furthermore, by Proposition 2, all opinions
must subsequently remain inside these intervals. On the other
hand, note that with an odd number of agents,
there is
one agent that always stays at L/2. Thus, if all opinions
eventually enter the interval (L/2 − 1, L/2 + 1), then there
2 + O( 1
n ), L − 11
n ), L − 1
12 − O( 1
4
x
8
7
6
5
4
3
2
1
0
0
5
t
10
15
Figure 3.
Time evolution when the initial opinions are uniformly spaced
on a semi-infinite interval, with a density of 100 per unit length. Groups of
agents become separated from the remaining agents, and converge to clusters
separated by approximately 2.2.
6 − O( 1
12 + O( 1
can only be a single cluster. This implies that there will
be a single cluster if L − 11
n ) < L/2 + 1 , that is, if
L < 23
n ) ≃ 3.833. This bound is smaller than the
experimentally observed value of about 5.1. It can be further
improved by carrying out explicit calculations of the smallest
position after a further number of iterations. Also, as long as
the number of agents is sufficiently large, a similar analysis
is possible if the number of agents is even, or in the presence
of random initial opinions.
(b) When L is sufficiently large, Figure 2 shows that
the position of the leftmost clusters becomes independent of
L. This can be explained by analyzing the propagation of
information: at each iteration, an agent is only influenced by
those opinions within distance 1 of its own, and its opinion is
modified by less than 1. So, information is propagated by at
most a distance 2 at every iteration. For the case of uniformly
spaced initial opinions on [0, L], with L large, the agents
with initial opinions close to 0 behave, at least in the first
iterations, as if opinions were initially distributed uniformly
on [0, +∞). Moreover, once a group of opinions is separated
from other opinions by more than 1, this group becomes
decoupled. Therefore, if the agents with initial opinions close
to 0 become separated from the remaining agents in finite time,
their evolution under a uniform initial distribution on [0, L] for
a sufficiently large L is the same as in the case of a uniform
initial distribution on [0, +∞).
We performed simulations with initial opinions uniformly
spaced on [0, ∞), as in [23]. We found that every agent
eventually becomes connected with a finite number of agents
and disconnected from the remaining agents. The groups
formed then behave independently and converge to clusters.
As shown in Figure 3, the distances between two consecutive
clusters are close to 2.2. These distances partially explain the
evolution of the number of clusters (as a function of L) shown
in Figure 2. However, a proof of these observed properties
is not available, and it
is unclear whether the successive
inter-cluster distances possess some regularity or convergence
properties.
(c) A last observation that leads to a better understanding of
x
4
3.5
3
2.5
2
1.5
1
0.5
0
0
5
t
10
15
Figure 4. Example of a temporary “meta-stable” state. Initially, two groups
are formed that do not interact with each other, but they both interact with a
small number of agents lying in between. As a result, the distance separating
the two groups decreases slowly and eventually becomes smaller than 1. At
that point, the groups attract each other directly and merge into a single cluster.
the size of the inter-cluster distances is the following. Suppose
that L is just below the value at which two clusters are
formed, and note the special nature of the resulting evolution,
shown in Figure 4. The system first converges to a “meta-
stable state” in which there are two groups, separated by a
distance slightly larger than 1, and which therefore do not
interact directly with each other. The two groups are however
slowly attracted by some isolated agents located in between;
furthermore, these isolated agents are being pulled by both
of these groups and remain at the weighted average of the
opinions in the two groups. Eventually, the distance between
the two groups becomes smaller than 1, the two groups start
attracting each other directly, and merge into a single cluster.
(This corresponds to one of the slow convergence phenomena
observed in [23].) The initial convergence towards a two-
cluster equilibrium is thus made impossible by the presence
of a few agents in between. Moreover, the number of these
isolated agents required to destabilize a meta-stable state can
be arbitrarily small compared to the number of agents in the
two groups. On the other hand, this phenomenon will not
arise if the two clusters are separated by a sufficiently large
distance. For example, if the distance between the two groups
is more than 2, no agent can be simultaneously connected
to both groups. This suggests that, depending on the distance
between clusters, some equilibria are stable with respect to the
presence of a small number of additional agents, while some
are not.
C. Stability with respect to a perturbing agent
In this section, we introduce a notion of equilibrium sta-
bility, motivated by the last observation in the preceding
subsection. We first generalize the model (1), so that each
agent i has an associated weight wi and updates its opinion
according to the weighted discrete-agent model
xi(t + 1) = Pj:xi(t)−xj (t)<1 wjxj(t)
Pj:xi(t)−xj(t)<1 wj
.
(2)
5
It can be verified that the convergence results in Theorem 1
and the properties proved in Propositions 1 and 2 continue to
hold. We will use the term weight of a cluster to refer to the
sum of the weights of all agents in the cluster. Observe that if
a number w of agents in system (1) have the same position,
they behave as a single agent with weight w in the model (2).
This correspondence can also be reversed, so that (2) can be
viewed as a special case of (1), whenever the weights wi are
integer, or more generally, rational numbers.
is again an equilibrium. We define ∆x0,δ = Pi wi ¯xi − ¯x′
Let ¯x be a vector of agent opinions at equilibrium. Suppose
that we add a perturbing agent indexed by 0, with weight δ
and initial opinion x0, that we let the system evolve again,
until it converges to a new, perturbed equilibrium, and then
remove the perturbing agent. The opinion vector ¯x′ so obtained
i,
which is a measure of the distance between the original and
perturbed equilibria. We say that ¯x is stable if supx0 ∆x0,δ, the
supremum of distances between initial and perturbed equilibria
caused by a perturbing agent of given weight δ, converges to
zero as δ vanishes. Equivalently, an equilibrium is unstable if
a substantial change in the equilibrium can be induced by a
perturbing agent of arbitrarily small weight.
Theorem 2: An equilibrium is stable if and only if for any
two clusters A and B with weights WA and WB, respectively,
the following holds: either WA = WB and the inter-cluster
distance is greater than or equal to 2; or WA 6= WB and the
inter-cluster distance is strictly greater than 1 + min(WA,WB )
max(WA,WB ) .
(Note that the two cases are consistent, except that the second
involves a strict inequality.)
Proof: We start with an interpretation of the strict in-
equality in the statement of the theorem. Consider two clusters
A and B, at positions xA and xB, and let m = (WAxA +
WBxB)/(WA + WB), which is their center of mass. Then, an
easy calculation shows that
xA − xB > 1 + min(WA,WB )
max(WA,WB )
if and only if
max{m − xA, m − xB} > 1
(3)
Suppose that an equilibrium ¯x0 satisfies the conditions in the
theorem. We will show that ¯x0 is stable. Let us insert a perturb-
ing agent of weight δ. Note that since ¯x0 is an equilibrium, and
therefore the clusters are at least 1 apart, the perturbing agent is
connected to at most two clusters. If this agent is disconnected
from all clusters, it has no influence, and ∆x0,δ = 0. If it
is connected to exactly one cluster A, with position xA and
weight WA, the system reaches a new equilibrium after one
time step, where both the perturbing agent and the cluster have
an opinion (x0δ + xAWA)/(δ + WA). Then,
∆x0,δ = x0 − xA ·
δ
δ + WA
≤
δ
δ + WA
,
which converges to 0 as δ → 0. Suppose finally that the
perturbing agent
is connected to two clusters A, B. This
implies that the distance between these two clusters is less
than 2, and since x0 satisfies the conditions in the theorem, it
must be greater than 1 + min(WA,WB )
max(WA,WB ) . Therefore, using (3),
the distance of one these clusters from their center of mass m
is greater than 1. The opinion of the perturbed agent after one
iteration is within O(δ) from m, while the two clusters only
move by an O(δ) amount. Since the original distance between
one of the two clusters and m is greater than 1, it follows
that after one iteration, and when δ is sufficiently small, the
distance of the perturbing agent from one of the clusters is
greater than 1, which brings us back to the case considered
earlier, and again implies that ∆x0,δ converges to zero as δ
decreases.
To prove the converse, we now suppose that the distance
between two clusters A and B, at positions xA and xB, is
less than 2, and also less than 1 + min(WA,WB )
max(WA,WB ) . Assuming
without loss of generality that xA < xB, their center of mass
m is in the interval (xB − 1, xA + 1). Let us fix an ǫ > 0 such
that (m − ǫ, m + ǫ) ⊆ (xB − 1, xA + 1). Suppose that at some
time t after the introduction of the perturbing agent we have
x0(t) ∈ (m(t) − ǫ, m(t) + ǫ) ⊆ (xB(t) − 1, xA(t) + 1) ,
(4)
with xB(t) − xA(t) ≥ 1, where x0(t), xA(t), xB(t), and m(t)
represent the positions at time t of the perturbing agent, of the
clusters A and B, and of their center of mass, respectively. One
can easily verify that xA(t + 1) = xA(t) + Θ(δ) > xA(t),
and xB(t + 1) = xB(t) − Θ(δ), so that xB(t + 1) − xA(t +
1) < xB(t) − xA(t), and (m(t + 1) − ǫ, m(t + 1) + ǫ) ⊆
(xB(t + 1) − 1, xA(t + 1) + 1) .
Moreover, observe that if δ were 0, we would have x0(t +
1) = m(t). For δ 6= 0, x0(t + 1) is close to m(t), and we have
x0(t + 1) = m(t) + O(δ). Since
m(t + 1) =
WAxA(t + 1) + WBxB(t + 1)
WA + WB
= m(t) + O(δ),
we obtain m(t + 1) − m(t) = O(δ), and therefore x0(t +
1) ∈ (m(t + 1) − ǫ, m(t + 1) + ǫ), as long as δ is sufficiently
small with respect to ǫ.
We have shown that if x0(0) = x0 is chosen so that the
condition (4) is satisfied for t = 0, and if δ is sufficiently small,
the condition (4) remains satisfied as long as xB(t) − xA(t) ≥
1. The perturbing agent remains thus close to the center of
mass, attracting both clusters, until at some time t∗ we have
xB(t∗) − xA(t∗) < 1. The two clusters then merge at the
next time step. The result of this process is independent of
the weight δ of the perturbing agent, which proves that ¯x is
not stable. Finally, a similar but slightly more complicated
argument shows that ¯x is not stable when xA − xB = 1 +
min(WA,WB )
max(WA,WB ) , and xA − xB < 2.
Theorem 2 characterizes the stable equilibria in terms of a
lower bound on the inter-cluster distances. It allows for inter-
cluster distances at a stable equilibrium that are smaller than
2, provided that the clusters have different weights. This is
consistent with experimental observations for certain initial
opinion distributions, as shown in Figure 5. On the other
hand, for the frequently observed case of clusters with equal
weights, stability requires the inter-cluster distances to be at
least 2. Thus, this result comes close to a full explanation of
the observed inter-cluster distances of about 2.2.
In general, there is no guarantee that the system (1) will
converge to a stable equilibrium. (A trivial example is obtained
by initializing the system at an unstable equilibrium, such as
6
3
2.5
2
x
1.5
1
0.5
0
0
1
2
t
3
4
5
Figure 5. Example of convergence to a stable equilibrium where the clusters
are separated by less than 2. The initial distribution of opinions is obtained by
taking 251 uniformly spaced opinions on [0, 2.5] and 500 uniformly opinions
on [2.5, 3]. Opinions converge to two clusters with 153 and 598 agents,
respectively, that are separated by a distance 1.6138 > 1.2559 = 1 + 153
598 .
Similar results are obtained when larger number of agents are used, provided
that the initial opinions are distributed in the same way, i.e, with a density on
[2.5, 3] which is ten times larger than the density on [0, 2.5].
2 for half of the agents and xi(0) = 1
xi(0) = − 1
2 for the other
half). On the other hand, we have observed that for a given
smooth distribution of initial opinions, and as the number of
agents increases, we almost always obtain convergence to a
stable equilibrium. This leads us to the following conjecture.
Conjecture 1: Suppose that the initial opinions are chosen
randomly and independently according to a particular con-
tinuous and bounded probability density function (PDF) with
connected support. Then, the probability of convergence to a
stable equilibrium tends to 1, as the number of agents increases
to infinity.
Besides the extensive numerical evidence (see e.g., Figure
6), this conjecture is supported by the intuitive idea that if the
number of agents is sufficiently large, whenever two groups
of agents start forming two clusters,
there will still be a
small number agents in between, whose presence will preclude
convergence to an unstable equilibrium. The conjecture is also
supported by Theorem 7 in Section III, which deals with a
continuum of agents, together with the results in Section IV
that provide a link between the discrete-agent and continuous-
agent models.
III. THE CONTINUOUS-AGENT MODEL
The discussion in the previous section indicates that much
insight can be gained by focusing on the case of a large number
of agents. This motivates us to consider a model involving a
continuum of agents. We use the interval I = [0, 1] to index
the agents, and we consider opinions that are nonnegative and
bounded above by a positive constant L. We denote by xt(α)
the opinion of agent α ∈ I at time t. We use X to denote the
set of measurable functions x : I → ℜ, and XL ⊂ X the set
of measurable functions x : I → [0, L]. The evolution of the
opinions is described by
xt+1(α) = Rβ:(α,β)∈Cxt
Rβ:(α,β)∈Cxt
xt(β) dβ
dβ
,
(5)
3
2.5
2
x
1.5
1
0.5
0
0
6
8
2
4
t
(a)
3
2.5
2
x
1.5
1
0.5
0
0
20
25
30
5
10
15
t
(b)
Figure 6. Time evolution of agent opinions, when initial opinions are drawn
from a common PDF which is larger on the interval (2.5,3) than on the
interval (0,2.5). In (a), we have 501 agents and they converge to an unstable
the clusters have respective weights 152 and 349, and their
equilibrium:
distance is 1.399 < 1 + 152
349 ≃ 1.436. In (b), we have 5001 agents and they
converge to a stable equilibrium: we see two clusters being formed originally,
but they are eventually drawn together by a small number of agents in between.
where Cx ⊆ I 2 is defined for any x ∈ X by
Cx := {(α, β) ∈ I 2 : x(α) − x(β) < 1}.
If the denominator in (5) is zero, we use the convention
xt+1(α) = xt(α). However, since the set of agents α for which
this convention applies has zero measure, we can ignore such
agents in the sequel. We assume that x0 ∈ XL. We then see
that for every t > 0, we have xt ∈ XL, so that the dynamics
are well-defined. In the sequel, we denote by χx the indicator
function of Cx, that is, χx(α, β) = 1 if (α, β) ∈ Cx, and
χx(α, β) = 0 otherwise.
We note that for the same reasons as in the discrete-agent
model, if for some α and β we have the relation xt(α) ≤ xt(β)
or xt(α) = xt(β) at some t, then the same relation continues
to hold at all subsequent times. Furthermore, if x0 only takes a
finite number of values, the continuous-agent model coincides
with the weighted discrete-agent model (2), with the same
range of initial opinions, and where each discrete agent’s
weight is set equal to the measure of the set of indices α
for which x0(α) takes the corresponding value.
In the remainder of this section, we will study the con-
vergence properties of the continuous-agent model, and the
inter-cluster distances at suitably defined stable equilibria.
A. Operator formalism
To analyze the continuous-agent model (5), it is convenient
to introduce a few concepts, extending well known matrix and
graph theoretic tools to the continuous case. By analogy with
interaction graphs in discrete multi-agent systems, we define
for x ∈ X the adjacency operator Ax, which maps the set X
of measurable functions on I into itself, by letting
(Axy) (α) =Z χx(α, β)y(β) dβ.
Applying this operator can be viewed as multiplying y by the
“continuous adjacency matrix” χx, and using an extension of
the matrix product to the continuous case. We also define the
degree function dx : I → ℜ+, representing the measure of the
set of agents to which a particular agent is connected, by
dx(α) =Z χx(α, β)dβ = (Ax1)(α),
7
where 1 : I → {1} is the constant function that takes the
value 1 for every α ∈ I. Multiplying a function by the degree
function can be viewed as applying an operator Dx : X → X
defined by
(Dxy) (α) = dx(α)y(α) =Z χx(α, β)y(α) dβ.
When dx is positive everywhere, we can also define the
operator D−1
x , which multiplies a function by 1/dx. Finally,
we define the Laplacian operator Lx = Dx − Ax. It follows
directly from these definitions that Lx1 = 0, similar to what
is known for the Laplacian matrix. In the sequel, we also use
the scalar product hx, yi =R x(α)y(α) dα. We now introduce
two lemmas to ease the manipulation of these operators.
Lemma 1: The operators defined above are symmetric with
respect to the scalar product: for any x, y, z ∈ X, we have
hz, Axyi = hAxz, yi, hz, Dxyi = hDxz, yi, and hz, Lxyi =
hLxz, yi.
Proof: The result is trivial for Dx. For Ax, we have
hz, Axyi = R z(α)(cid:0)R χx(α, β)y(β) dβ(cid:1) dα
= R y(β)(cid:0)R χx(α, β)z(α) dα(cid:1) dβ.
Since χx(α, β) = χx(β, α) for all α, β,
this implies
hz, Axyi = hAxz, yi. By linearity, the result also holds for
Lx and any other linear combination of those operators.
Lemma 2: For any x, y ∈ X, we have
hy, (Dx ± Ax)yi =
1
2Z χx(α, β) (y(α) ± y(β))2 dα dβ.
In particular, Lx = Dx − Ax is positive semi-definite.
Proof: From the definition of the operators, we have
hy, (Dx ± Ax)yi =Z χx(α, β)y(α) (y(α) ± y(β)) dα dβ.
The right-hand side of this equality can be rewritten as
The symmetry of χx then implies that hy, (Dx ± Ax)yi equals
1
+ 1
2(cid:0)R χx(α, β)y(α) (y(α) ± y(β)) dα dβ(cid:1)
2(cid:0)R χx(β, α)y(β) (y(β) ± y(α)) dα dβ(cid:1) .
2Z χx(α, β)(cid:0)y(α)2 ± 2y(α)y(β) + y(β)2(cid:1) dα dβ,
1
from which the results follows directly.
The update equation (5) can be rewritten, more compactly,
in the form
∆xt := xt+1 − xt = −D−1
x Lxtxt, or Dxt∆xt = −Lxtxt,
(6)
where the second notation is formally more general as it also
holds on the possibly nonempty zero-measure set on which
dx = 0. We say that xt ∈ XL is a fixed point of the system
if ∆xt = 0 holds almost everywhere (a.e., for short), that is,
except possibly on a zero-measure set. It follows from (6) that
the set of fixed points is characterized by the equality Lxx = 0,
a.e. One can easily see that the set of fixed points contains the
set F := {x ∈ XL : x(α) 6= x(β) ⇒ x(α) − x(β) ≥ 1} of
opinion functions taking a discrete number of values that are
at least one apart. Let ¯F be the set of functions x ∈ XL for
which there exists s ∈ F such that s = x, a.e. We prove later
that ¯F is exactly the set of solutions to Lxx = 0, a.e., and
thus the set of fixed points of (6).
B. Convergence
In this section we present some partial convergence results.
In particular, we show that the change ∆xt of the opinion
function decays to 0, and that xt tends to the set of fixed
points. We begin by proving the decay of a quantity related
to ∆xt.
Theorem 3: For any initial condition of the system (6), we
have
∞
Xt=0Z χxt (α, β) (∆xt(α) + ∆xt(β))2 dα dβ < ∞.
Proof: We consider the nonnegative potential function V :
X → ℜ+ defined by
V (x) =
1
2Z min(cid:16)1, (x(α) − x(β))2(cid:17) dα dβ ≥ 0,
(7)
and show that
V (xt+1) − V (xt) ≤ − h∆xt, (Axt + Dxt)∆xti ,
which by Lemma 2 implies the desired result.
We
observe
that
for
every x, y
∈ X,
min(cid:16)1, (y(α) − y(β))2(cid:17) is
smaller
both 1 and (y(α) − y(β))2, there holds
than or
equal
since
to
2RCx
= hy, Lxyi + 1
V (y) ≤ 1
(y(α) − y(β))2 dα dβ + 1
1 dα dβ
2(cid:12)(cid:12)I 2 \ Cx(cid:12)(cid:12) ,
(8)
where Lemma 2 was used to obtain the last equability. For
y = x, it follows from the definition of Cx that the above
inequality is tight. In particular, the following two relations
hold for any s and t:
2RI 2\Cx
V (xt) = hxt, Lxtxti + 1
V (xs) ≤ hxs, Lxtxsi + 1
Taking s = t + 1, we obtain
2(cid:12)(cid:12)I 2 \ Cxt(cid:12)(cid:12)
2(cid:12)(cid:12)I 2 \ Cxt(cid:12)(cid:12) .
V (xt+1) − V (xt) ≤ hxt+1, Lxtxt+1i − hxt, Lxtxti
= 2 h∆xt, Lxtxti + h∆xt, Lxt∆xti ,
where we have used the symmetry of Lxt. It follows from (6)
that Lxtxt = −Dxt∆xt, so that
V (xt+1) − V (xt) ≤ −2 h∆xt, Dxtxti + h∆xt, Lxtxti
= − h∆xt, (Axt + Dxt )∆xti ,
since Lx = Dx − Ax.
As will be seen below, this result implies the convergence
of ∆xt to 0 in a suitable topology. We now show that Lxx is
small only if x is close to F , the set of functions taking discrete
values separated by at least 1. As a corollary, we then obtain
the result that ¯F is exactly the set of fixed points, as also shown
in [24]. The intuition behind the proof of these results parallels
our proof of Theorem 1, and is as follows. Consider an agent α
with one of the smallest opinions x(α). If the change in x(α)
is small, its attraction by agents with larger opinions must be
small, because almost no agents have an opinion smaller than
x(α). Therefore, there must be very few agents with an opinion
significantly larger than x(α) that interact with α, while there
might be many of them who have an opinion close to x(α).
In other words, possibly many agents have approximately the
8
same opinion x(α), and very few agents have an opinion in
the interval [x(α)+ǫ, x(α)+1), so that x is close to a function
in F in that zone. Take now an agent α′ with an opinion larger
than x(α)+1+ǫ, and such that very few agents have an opinion
in (x(α) + 1 + ǫ, x(α′)). This agent interacts with very few
agents having an opinion smaller than its own. Thus, if the
change in such an agent’s opinion is small, this implies that
its attraction by agents having larger opinions is also small,
and we can repeat the previous reasoning.
In order to provide a precise statement of the result, we
associate an opinion function x with a measure that describes
the distribution of opinions, and use a measure-theoretic
formalism. For a measurable function x : I → [0, L] (i.e.,
x ∈ XL), and a measurable set S ⊆ [0, L], we let µx(S) be the
Lebesgue measure of the set {α : x(α) ∈ S}. By convention,
we let µ(S) = 0 if S ⊆ ℜ \ [0, L]. To avoid confusion with
µ, we use S to denote the standard Lebesgue measure of
a set S. We also introduce a suitable topology on the set of
opinion functions. We write x ≤µ ǫ if {α : x(α) > ǫ} ≤ ǫ.
Similarly, x <µ ǫ if {α : x(α) ≥ ǫ} < ǫ, and x =µ 0
if {α : x(α) 6= 0} = 0. We define the “ball” Bµ(x, ǫ) as
the set {y ∈ XL : x − y <µ ǫ}. This allows us to define
a corresponding notion of limit. We say that xt →µ y if for
all ǫ > 0, there is a t′ such that for all t > t′ we have
xt ∈ Bµ(y, ǫ). We write xt →µ S for a set S if for all ǫ > 0,
there is a t′ such that for all t > t′, there is a y ∈ S for which
xt ∈ Bµ(y, ǫ).
The result below, proved in Appendix A, states that the
distance between x ∈ XL and F (the subset of XL consisting
of functions taking discrete values separated by at least 1)
decreases to 0 (in a certain uniform sense) when Lxx →µ 0.
Theorem 4: For any ǫ > 0, there exists a δ > 0 such that if
Lxx <µ δ, then there exists some s ∈ F with x − s <µ ǫ.
In particular, if Lxx =µ 0, then x ∈ ¯F .
The next theorem compiles our convergence results.
Theorem 5: Let (xt) be a sequence of functions in XL
evolving according to the model (5), and let F be the set of
functions taking discrete values separated by at least 1. Then
(xt+1 −xt) →µ 0 and xt →µ F . (In particular, periodic trajec-
tories, other than fixed points, are not possible.) Furthermore,
x is a fixed point of (5) if and only if x ∈ ¯F .
Proof: We begin by proving the convergence of ∆xt.
Suppose that ∆xt = (xt+1 − xt) →µ 0 does not hold. Then,
there is an ǫ > 0 such that for arbitrarily large t, there is a
set of measure at least ǫ such that ∆xt(α) > ǫ for every α
in that set. Consider such a time t. Without loss of generality,
assume that there is a set S ⊆ I of measure at least ǫ/2 on
which ∆xt(α) > ǫ. (Otherwise, we can use a similar argument
for the set on which ∆xt(α) < −ǫ.) Fix some L′ > L.
For i ∈ {1, . . . , 2⌈L′⌉}, let Ai ⊂ I be the set on which
xt ∈ [(i − 1)/2, i/2]. For any i and for any α, β ∈ Ai, there
holds xt(α) − xt(β) < 1 and thus (α, β) ∈ Cxt. Therefore,
i ⊆ Cxt for all i. Moreover, the sets Ai cover [0, 1], so that
A2
Ai ∩ S ≥ S ≥ ǫ/2. Thus, there exists some i∗ such
i=1
P2⌈L′⌉
that Ai∗ ∩ S ≥ ǫ/(4⌈L′⌉). We then have
RCxt
(∆xt(α) + ∆xt(β))2 dα dβ
≥ R(Ai∗ ∩S)2 (∆xt(α) + ∆xt(β))2 dα dβ
≥ 4ǫ2Ai∗ ∩ S2 ≥ ǫ4/4⌈L′⌉2.
if ∆xt
does
→µ
not
0
Thus,
R(α,β)∈Cxt
then
(∆xt(α) + ∆xt(β))2 does not decay to 0, which
contradicts Theorem 3. We conclude that ∆xt →µ 0. Using
also (6) and the fact dxt(α) ≤ 1, we obtain Lxtxt →µ 0.
Theorem 4 then implies that xt →µ F .
hold,
If x ∈ ¯F , it is immediate that x is a fixed point. Conversely,
if x0 = x is a fixed point, then xt = x0, a.e., for all t. Then,
the fact xt →µ F implies that x ∈ ¯F .
We note that the fact xt →µ F means that the measure µx
associated with any limit point x of xt is a discrete measure
whose support consists of values separated by at least 1.
Furthermore, it can be shown that at least one such limit point
exists, because of the semi-compactness of the set of measures
under the weak topology.
Theorem 5 states that xt tends to the set F , but does not
guarantee convergence to an element of this set. We make the
following conjecture, which is currently unresolved.
Conjecture 2: Let (xt) be a sequence of functions in XL,
evolving according to the model (5). Then, there is a function
x∗ ∈ F such that xt →µ x∗.
C. Inter-cluster distances and stability of equilibria
We have found that x is a fixed point of (5) if and only if it
belongs to ¯F , that is, with the exception of a zero-measure set,
the range of x is a discrete set of values that are separated by
at least one. As before, we will refer to these discrete values as
clusters. In this section, we consider the stability of equilibria,
and show that a condition on the inter-cluster distances similar
to the one in Theorem 2 is necessary for stability. Furthermore,
we show that under a certain smoothness assumption, the
system cannot converge to a fixed point that does not satisfy
this condition.
In contrast to the discrete case, we can study the continuous-
agent model using the classical definition of stability. We say
that s ∈ F is stable if for any ǫ > 0, there is a δ > 0 such
that for any x0 ∈ Bµ(s, δ), we have xt ∈ Bµ(s, ǫ) for all
t. It can be shown that this notion encompasses the stability
with respect to the addition of a perturbing agent used in
Section II-C. More precisely, if we view the discrete-agent
system as a special case of the continuum model, stability
under the current definition implies stability with respect to the
notion used in Section II-C. The introduction of a perturbing
agent with opinion x0 can indeed be simulated by taking
x0(α) = s(α) everywhere except on an appropriate set of
measure less than δ, and x0(α) = x0 on this set. (However, the
converse implication turns out to not hold in some pathological
cases. Indeed, consider two agents separated by exactly 2.
They are stable with respect to the definition of Section II-C,
but not under the current definition. This is because if we
introduce a small measure set of additional agents that are
uniformly spread between the two original agents, we will
obtain convergence to a single cluster.) Moreover, it can be
9
verified that the notion of stability used here is equivalent to
both L1 and L2 stability. In the sequel, and to simplify the
presentation, we will neglect any zero measure sets on which
∆xt(α) 6= 0, and will give the proof for a fixed point in
F . The extension to fixed points in ¯F is straightforward. The
proof of the following result is similar to that of its discrete
counterpart, the necessary part of Theorem 2, and is presented
in the Appendix B.
Theorem 6: Let s ∈ F be a fixed point of (5), and let a, b
two values taken by s. If s is stable, then
b − a ≥ 1 +
min (µs(a), µs(b))
max (µs(a), µs(b))
.
(9)
With a little extra work, focused on the case where the
distance a − b between the two clusters is exactly equal to
2, we can show that the strict inequality version of condition
(9) is necessary for stability. We conjecture that this strict
inequality version is also sufficient.
We will now proceed to show that under an additional
smoothness assumption on the initial opinion function, we can
never have convergence to a fixed point that violates condition
(9). We start by introducing the notion of a regular opinion
function. We say that a function x ∈ XL is regular if there
exist M ≥ m > 0 such that any interval J ⊆ [inf α x, supα x]
satisfies m J ≤ µx(J) ≤ M J. Intuitively, a function is
regular if the set of opinions is connected, and if the density
of agents on any interval of opinions is bounded from above
and from below by positive constants. (In particular, no single
value is taken by a positive measure set of agents.) For
example, any piecewise differentiable x ∈ XL with positive
upper and lower bounds on its derivative is regular.
We will show that if x0 is regular and if (xt) converges,
then xt converges to an equilibrium satisfying the condition
(9) on the minimal distance between opinions, provided that
supα xt − inf α xt remains always larger than 2. For conve-
nience, we introduce a nonlinear update operator U on XL,
x Axx, so that the
defined by U (x) = x − D−1
recurrence (5) can be written as xt+1 = U (xt). The proof of
the following proposition is presented in Appendix C.
x Lxx = D−1
Proposition 3: Let x ∈ XL be a regular function such that
supα x − inf α x > 2. Then U (x) is regular.
We note that the assumption supα x−inf α x > 2 in Proposi-
tion 3 is necessary for the result to hold. Indeed, if the opinion
values are confined to a set [a, b], with b − a = 2 − δ < 2,
then all agents with opinions in the set [a + 1 − δ, a + 1] are
connected with every other agent, and their next opinions will
be the same, resulting in a non-regular opinion function.
As a consequence of Proposition 3, together with Theorem
5, if x0 is regular, then there are two main possibilities: (i)
There exists some time t at which supα xt − inf α xt < 2.
In this case, the measure µxt will have point masses shortly
thereafter, and will eventually converge to the set of fixed
points with at most two clusters. (ii) Alternatively, in the
“regular” case, we have supα xt − inf α xt>2 for all times.
Then, every xt is regular, and convergence cannot take place
in finite time. Furthermore, as we now proceed to show,
convergence to a fixed point that violates the stability condition
(9) is impossible. Let us note however that tight conditions
for a sequence of regular functions to maintain the property
supα xt−inf α xt>2 at all times appear to be difficult to obtain.
Theorem 7: Let (xt) be a sequence of functions in XL that
evolve according to (5). We assume that x0 is regular and
that supα xt − inf α xt > 2 for all t. If (xt) converges, then it
converges to a function s ∈ F such that
b − a ≥ 1 +
min (µs(a), µs(b))
max (µs(a), µs(b))
,
for any two distinct values a, b, with µs(a), µs(b) > 0. In
particular, if µs(a) = µs(b), then b − a ≥ 2.
Proof: Suppose that (xt) converges to some s. It follows
from Theorem 5 that s ∈ F , and from Proposition 3 that all
xt are regular. Suppose now that s violates the condition in
the theorem, for some a, b, with a < b. Then, b − a < 2,
and we must have µs ((a, b)) = 0 because all discrete values
taken by s (with positive measure) must differ by at least 1.
We claim that there exists a positive length interval J ⊆ (a, b)
such that µxt+1 (J) ≥ µxt(J) whenever xt ∈ Bµ(s, ǫ), for a
sufficiently small ǫ > 0. Since xt converges to s, this will
imply that there exists a finite time t∗ after which µxt (J) is
nondecreasing, and lim inf t→∞ µxt(J) ≥ µxt∗ (J) > 0. On
the other hand, since µs((a, b)) = 0, µxt(J) must converge to
zero. This is a contradiction and establishes the desired result.
µs(a)+µs(b) be
the weighted average of a and b. The fact that the condition
in the theorem is violated implies (cf. (3)) that c − a < 1
and c − b < 1. Let δ > 0 be such that c − δ + 1 > b and
c + δ − 1 < a, and consider the interval J = [c − δ, c + δ]. For
any x ∈ Bµ(s, ǫ), we have
We now establish the above claim. Let c = µs(a)a+µs(b)b
µx([a − ǫ, a + ǫ]) ∈ [µs(a) − ǫ, µs(a) + ǫ],
µx([b − ǫ, b + ǫ]) ∈ [µs(b) − ǫ, µs(b) + ǫ],
µx ((a − 1, b + 1) \ ([a − ǫ, a + ǫ] ∪ [b − ǫ, b + ǫ])) ≤ ǫ,
where we have used the fact that the values taken by s are
separated by at least 1. Suppose now that ǫ is sufficiently
small so that c − δ + 1 > b + ǫ and c + δ − 1 < a − ǫ.
This implies that for every γ such that x(γ) ∈ J, we have
(a − ǫ, b + ǫ) ⊆ (x(γ) − 1, x(γ) + 1). If ǫ were equal to zero,
we would have ux(d) = c. When ǫ is small, the location of the
masses at a and b moves by an O(ǫ) amount, and an additional
O(ǫ) mass is introduced. The overall effect is easily shown
to be O(ǫ) (the detailed calculation can be found in [15]).
Thus, (U (x))(γ) − c is of order O(ǫ). When ǫ is chosen
sufficiently small, we obtain c − δ ≤ (U (x))(γ) ≤ c + δ, i.e.,
(U (x))(γ) ∈ J for all γ such that x(γ) ∈ J. This implies that
µU(x)(J) ≥ µx(J), and completes the proof.
IV. RELATION BETWEEN THE DISCRETE AND THE
CONTINUOUS-AGENT MODELS
We now analyze the extent to which the continuous-agent
model (5) can be viewed as a limiting case of the discrete-
agent model (1), when the number of agents tends to infinity.
As already explained in Section III,
the continuous-agent
model can simulate exactly the discrete-agent model. In this
section, we are interested in the converse; namely, the extent
to which a discrete-agent model can describe, with arbitrarily
10
good precision, the continuous-agent model. We will rely on
the following result on the continuity of the update operator.
Proposition 4: Let x ∈ XL be a regular function. Then, the
update operator U is continuous at x with respect to the norm
k · k∞. More precisely, for any ǫ > 0 there exists some δ > 0
such that if ky − xk∞ ≤ δ then kU (y) − U (x)k∞ ≤ ǫ.
Proof: Consider a regular function x ∈ XL, and an
arbitrary ǫ > 0. Let δ be smaller than mǫ/25M, where m and
M (with m ≤ M ) are the bounds in the definition of regular
opinion functions applied to x. We will show that if a function
y ∈ XL satisfies kx − yk∞ ≤ δ, then kU (y) − U (x)k∞ ≤ ǫ.
Fix some α ∈ I, and let Sx, Sy ⊆ I be the set of agents
connected to α according to the interconnection topologies Cx
and Cy defined by x and y, respectively. We let Sxy = Sx∩Sy,
Sx\y = Sx \ Sxy and Sy\x = Sy \ Sxy. Since kx − yk∞ ≤
δ, the values x(α) − x(β) and y(α) − y(β) differ by at
most 2δ, for any β ∈ I. As a consequence, if β ∈ Sy, then
x(α) − x(β) ≤ y(α) − y(β) + 2δ. Similarly, if β 6∈ Sy,
then x(α) − x(β) ≥ y(α) − y(β) − 2δ. Combining these
two inequalities with the definitions of Sxy, Sx\y, and Sy\x,
we obtain
[x(α) − 1 + 2δ, x(α) + 1 − 2δ] ⊆ x(Sxy) ⊆ [x(α) − 1, x(α) + 1],
x(Sx\y) ⊆ [x(α) − 1, x(α) − 1 + 2δ] ∪ [x(α) + 1 − 2δ, x(α) + 1],
x(Sy\x) ⊆ [x(α) − 1 − 2δ, x(α) − 1] ∪ [x(α) + 1, x(α) + 1 + 2δ].
Since x is regular, we have Sxy ≥ m(2 − 4δ) ≥ m and
(cid:12)(cid:12)Sx\y(cid:12)(cid:12) ,(cid:12)(cid:12)Sy\x(cid:12)(cid:12) ≤ M 4δ. Let now ¯xxy and ¯xx\y be the average
value of x on Sxy and Sx\y, respectively. Similarly, let ¯yxy,
and ¯yy\x be the average value of y on Sxy and Sy\x. Since
kx − yk∞ ≤ δ, ¯xxy and ¯yxy differ by at most δ. It follows
from the definition of the model (5) that
(U (x))(α) = ¯xxy +
(U (y))(α) = ¯yxy +
Sx\y
Sxy+Sx\y
Sy\x
Sxy+Sy\x
(¯xx\y − ¯xxy),
(¯yy\x − ¯yxy).
from which we obtain that (U (y))(α) − (U (x))(α) is upper
It can be seen that (cid:12)(cid:12)¯xx\y − ¯xxy(cid:12)(cid:12) ≤ 3 and (cid:12)(cid:12)¯yy\x − ¯yxy(cid:12)(cid:12) ≤ 3,
¯xxy − ¯yxy + 3(cid:12)(cid:12)Sy\x(cid:12)(cid:12)
where we have used the fact that ¯xxy − ¯yxy ≤ δ. Since
the above is true for any α ∈ I, we conclude that
kU (y) − U (x)k∞ ≤ ǫ.
+ 3(cid:12)(cid:12)Sx\y(cid:12)(cid:12)
Sxy
4M δ
m
Sxy
≤ δ + 6
≤ ǫ.
Let U t : XL → XL be the composition of the update oper-
ator, defined by U t(x) = U(cid:0)U t−1(x)(cid:1), so that U t(x0) = xt.
Proposition 4 is readily extended to a continuity result for U t.
Corollary 1: Let x0 ∈ XL be a regular function such that
supα U t(x) − inf α U t(x) > 2 for every t ≥ 0. Then for any
finite t, U t is continuous at x with respect to the norm k · k∞.
Proof: Since x is regular and since supα U t(x) −
inf α U t(x) > 2 for all t, Proposition 3 implies that all U t(x)
are regular. Proposition 4 then implies that for all t, U is
continuous at U t(x), and therefore the composition U t is
continuous at x.
Corollary 1 allows us to prove that, in the regular case,
and for any given finite time horizon, the continuous-agent
model is the limit of the discrete-agent model, as the number
of agents grows. To this effect, for any given partition of I =
[0, 1] into n disjoint sets J1, . . . , Jn, we define an operator
G : ℜn → X that translates the opinions in an n-agent system
to an opinion function in the continuous-agent model. More
precisely, for a vector x ∈ ℜn and any α ∈ Ji, we let (G x)(α)
be equal to the ith component of x.
Theorem 8: Let x0 ∈ XL be a regular function and assume
that supα xt − inf α xt > 2 for t ≤ t∗. Then, the sequence
(xt), t = 1, . . . , t∗, can be approximated arbitrarily well by
a sequence (xt) of opinion vectors evolving according to (1),
in the following sense. For any ǫ > 0, there exists some n,
a partition of I into n disjoint sets J1, . . . , Jn, and a vector
x0 ∈ [0, L]n such that the sequence of vectors xt generated by
the discrete-agent model (1), starting from x0, satisfies kxt −
G xtk∞ ≤ ǫ, for t = 1, . . . , t∗.
Proof: Fix ǫ > 0. Since all U t are continuous at x0, there
is some δ > 0 such that if ky − x0k∞ ≤ δ, then kU t(y) −
xtk∞ ≤ ǫ, for t ≤ t∗. Since x0 is regular, we can divide I
into subsets J1, J2, . . . , Jn, so that Ji = 1/n for all i, and
x0(α) − x0(β) ≤ δ for all α, β in the same set Ji. (This is
done by letting ci be such that µx0([0, ci]) = i/n, and defining
Ji = {α : ci−1 ≤ x0(α) ≤ ci}, where n is sufficiently large.)
We define x0 ∈ [0, L]n by letting its ith component be equal
to ci. We then have kx0 − G x0k∞ ≤ δ. This implies that
kxt − U t(G x0)k∞ ≤ ǫ, for t ≤ t∗. Since the continuous-
agent model, initialized with a discrete distribution, simulates
the discrete-agent model, we have U t(G x0) = G xt, and the
desired result follows.
Theorem 8 supports the intuition that for large values of
n, the continuous-agent model behaves approximatively as
the discrete-agent model, over any finite horizon. In view of
Theorem 6, this suggests that the discrete-agent system should
always converge to a stable equilibrium (in the sense defined in
Section II) when n is sufficiently large, as stated in Conjecture
1, and observed in many examples (see, e.g., Figure 6). Indeed,
Theorem 6 states that under the regularity assumption, the
continuum system cannot converge to an equilibrium that
does not satisfy condition (9) on the inter-cluster distances.
However, this argument does not translate to a proof of the
conjecture because the approximation property in Theorem 8
only holds over a finite time horizon, and does not necessarily
provide information on the limiting behavior.
V. CONCLUSIONS AND OPEN QUESTIONS
We have analyzed the model of opinion dynamics (1)
introduced by Krause, from several angles. Our motivation was
to provide an analysis of a simple multi-agent system with an
endogenously changing interconnection topology while taking
explicitly advantage of the topology dynamics, something that
is rarely done in the related literature.
We focused our attention on an intriguing phenomenon,
the fact that equilibrium inter-cluster distances are usually
significantly larger than 1, and typically close to 2. We
proposed an explanation of this phenomenon based on a notion
of stability with respect to the addition of a perturbing agent.
We showed that such stability translates to a certain lower
bound on the inter-cluster distances, with the bound equal to
2 when the clusters have identical weights. We also discussed
the conjecture that when the number of agents is sufficiently
11
large, the system converges to a stable equilibrium for “most”
initial conditions.
To avoid granularity problems linked with the presence or
absence of an agent in a particular region, we introduced a
new opinion dynamics model that allows for a continuum of
agents. For this model we proved that under some regularity
assumptions, there is always a finite density of agents between
any two clusters during the convergence process. As a result,
we could prove that such systems never converge to an
unstable equilibrium. We also proved that the continuous-
agent model is indeed the limit of a discrete model, over any
given finite time horizon, as the number of agents grows to
infinity. These results provide some additional support for the
conjectured, but not yet established, generic convergence to
stable equilibria.
We originally introduced the continuous-agent model as
a tool for the study of the discrete-agent model, but it is
also of independent interest and raises some challenging open
questions. An important one is the question of whether the
continuous-agent model is always guaranteed to converge. (We
only succeeded in establishing convergence to the set of fixed
points, not to a single fixed point.)
Finally, the study of the continuous-agent model suggests
some broader questions. In the same way that the convergence
of the discrete-agent model can be viewed as a special
case of convergence of inhomogeneous products of stochastic
matrices, it may be fruitful to view the convergence of the
continuous-agent model as a special case of convergence of
inhomogeneous compositions of stochastic operators, and to
develop results for the latter problem.
The model (1) can of course be extended to higher di-
mensional spaces, as is often done in the opinion dynamics
literature (see [24] for a survey). Numerical experiments
again show the emergence of clusters that are separated by
distances significantly larger than 1. The notion of stability
with respect to the addition of an agent can also be extended
to higher dimensions. However, stability conditions become
more complicated, and in particular cannot be expressed as
a conjunction of independent conditions, one for each pair
of clusters. For example, it turns out that adding a cluster
to an unstable equilibrium may render it stable [15]. In
addition, a formal analysis appears difficult because in ℜn,
with n > 1,
the support of the opinion distribution can
be connected without being convex, and convexity is not
necessarily preserved by our systems. For this reason, even
under “regularity” assumptions, the presence of perturbing
agents between clusters is not guaranteed.
REFERENCES
[1] E. Ben-Naim. Rise and fall of political parties. Europhysics Letters,
69(5):671–676, 2005.
[2] E. Ben-Naim, P.L. Krapivsky, and S. Redner. Bifurcations and patterns
in compromise processes. Physica D, 183(3):190–204, 2003.
[3] E. Ben-Naim, P.L. Krapivsky, F. Vasquez, and S. Redner. Unity and
discord in opinion dynamics. Physica A, 330:99–106, 2003.
[4] D. P. Bertsekas and J. N. Tsitsiklis. Parallel and Distributed Computa-
tion. Prentice-Hall, 1989.
[5] V.D. Blondel, J.M. Hendrickx, A. Olshevsky, and J.N. Tsitsiklis. Con-
vergence in multiagent coordination, consensus, and flocking.
In
Proceedings of the 44th IEEE Conference on Decision and Control
(CDC’2005), pages 2996–3000, Seville, Spain, December 2005.
[6] V.D. Blondel, J.M. Hendrickx, and J.N. Tsitsiklis. On the 2R conjecture
for multi-agent systems.
the European Control
Conference 2007 (ECC’2007), pages 874–881, Kos, Greece, July 2007.
[7] F. Cucker and S. Smale. Emergent behavior in flocks. IEEE Transactions
In Proceedings of
on Automatic Control, 52:852–862, May 2007.
[8] F. Cucker and S. Smale. On the mathematics of emergence. Japanese
Journal of Mathematics, 2:119–227, March 2007.
[9] G. Deffuant, D. Neau, F. Amblard, and G. Weisbuch. Mixing beliefs
among interacting agents. Advances in Complex Systems, 3:87–98, 2000.
Journal of the American
[10] DeGroot, M. H. Reaching a consensus.
Statistical Association, 69:118–121, 1974.
[11] J. C. Dittmer. Consensus formation under bounded confidence. Nonlin-
ear Analysis, (47):4615–4621, 2001.
[12] S. Fortunato, V. Latora, A. Pluchino, and Rapisarda R. Vector opinion
International
dynamics in a bounded confidence consenus model.
Journal of Modern Physics C, 16:1535–1551, October 2005.
[13] R. Hegselmann. Opinion dynamics: Insights by radically simplifying
models. In D. Gillies, editor, Laws and Models in Science, pages 1–29,
London, 2004.
[14] R. Hegselmann and U. Krause. Opinion dynamics and bounded
confidence models, analysis, and simulations. Journal of Artificial Soci-
eties and Social Simulation, 5(3):http://jasss.soc.surrey.ac.uk/9/1/8.html,
2002.
[15] J.M. Hendrickx.
Graphs and Networks for the Analysis of Au-
tonomous Agent Systems. PhD thesis, Universit´e catholique de Louvain,
http://www.inma.ucl.ac.be/hendrickx/availablepublications/
Thesis Julien Hendrickx.pdf, 2008.
[16] J.M. Hendrickx and V.D. Blondel. Convergence of different linear and
non-linear Vicsek models. In Proceedings of the 17th International Sym-
posium on Mathematical Theory of Networks and Systems (MTNS’2006),
pages 1229–1240, Kyoto, Japan, July 2006.
[17] A. Jadbabaie, J. Lin, and A. S. Morse. Coordination of groups of mobile
autonomous agents using nearest neighbor rules. IEEE Transactions on
Automatic Control, 48(6):988–1001, 2003.
[18] E.W. Justh and P. S. Krishnaprasad. Equilibria and steering laws for
planar formations. Systems and Control Letters, 52(1):25–38, 2004.
[19] U. Krause. Soziale Dynamiken mit vielen Interakteuren. Eine Prob-
lemskizze. In Modellierung und Simulation von Dynamiken mit vielen
interagierenden Akteuren, pages 37–51. 1997.
[20] U. Krause. A discrete nonlinear and non-autonomous model of consen-
sus formation. Communications in Difference Equations, pages 227–236,
2000.
[21] J. Lin, A.S. Morse, and B.D.O. Anderson. The multi-agent rendezvous
problem. In Proceedings of the 42th IEEE Conference on Decision and
Control (CDC’2003), pages 1508– 1513, Hawaii (HA), USA, December
2003.
[22] J. Lorenz. A stabilization theorem for continuous opinion dynamics.
Physica A, 355(1):217–223, 2005.
Consensus
[23] J. Lorenz.
model of
dence.
9(1):http://jasss.soc.surrey.ac.uk/9/1/8.html, 2006.
strikes back in the Hegselmann-Krause
continuous opinion dynamics under bounded confi-
Journal of Artificial Societies and Social Simulation,
[24] J. Lorenz. Continuous opinion dynamics under bounded confidence: A
survey. International Journal of Modern Physics C, 18(12):1819–1838,
2007.
[25] L. Moreau. Stability of multiagent systems with time-dependent commu-
nication links. IEEE Transactions on Automatic Control, 50(2):169–182,
2005.
[26] R. Olfati-Saber, J.A. Fax, and R.M. Murray. Consensus and cooperation
in networked multi-agent systems. Proceedings of the IEEE, 95(1):215–
233, January 2007.
[27] W. Ren, R.W. Beard, and E.M. Atkins.
IEEE Control and Systems
Magazine, 27(2):71–82, April 2007.
[28] J.N. Tsitsiklis. Problems in decentralized decision making and computa-
tion. PhD thesis, Dept. of Electrical Engineering and Computer Science,
Massachusetts Institute of Technology, http://web.mit.edu/jnt/www/PhD-
84-jnt.pdf, 1984.
[29] D. Urbig.
Attitude dynamics with limited verbalisation capa-
Journal of Artificial Societies and Social Simulation,
bilities.
6(1):http://jasss.soc.surrey.ac.uk/6/1/2.html, 2003.
[30] T. Vicsek, A. Czirok, I. Ben Jacob, I. Cohen, and O. Schochet. Novel
type of phase transitions in a system of self-driven particles. Physical
Review Letters, 75:1226–1229, 1995.
12
Vincent D. Blondel received the M.Sc. degree in
mathematics from Imperial College, London, U.K.,
in 1990 and the Ph.D. degree in applied mathematics
from the Universit´e catholique de Louvain, Louvain-
la-Neuve, Belgium, in 1992.
He was a Visiting Researcher at the Royal In-
stitute of Technology, Stockholm, Sweden, and at
the Institut National de Recherche en Informatique
et en Automatique (INRIA), Rocquencourt, France.
During 2005-2006, he was an Invited Professor and
a Fulbright Scholar at Massachusetts Institute of
Technology, Cambridge. He is currently a Professor and Department Head
at the Universit´e catholique de Louvain, Louvain-la-Neuve, Belgium.
Dr. Blondel was the recipient of the Prize Wetrems of the Belgian Royal
Academy of Science, the Society for Industrial and Applied Mathematics
(SIAM) Prize on Control and Systems Theory, and the Ruberti Prize in
Systems and Control of the IEEE in 2006.
Julien M. Hendrickx received an engineering de-
gree in applied mathematics and a PhD in mathemat-
ical engineering from the Universit´e catholique de
Louvain, Belgium, in 2004 and 2008, respectively.
He has been a visiting researcher at the University
of Illinois at Urbana Champaign in 2003-2004, at the
National ICT Australia in 2005 and 2006, and at the
Massachusetts Institute of Technology in 2006 and
2008. He is currently a postdoctoral fellow at the
Laboratory for Information and Decision systems of
the Massachusetts Institute of Technology, and holds
postdoctoral fellowships of the F.R.S.-FNRS (Fund for Scientific Research)
and of Belgian American Education Foundation.
Doctor Hendrickx was the recipient of the 2008 EECI award for the best
PhD thesis in Europe in the field of Embedded and Networked Control.
John N. Tsitsiklis (F’99) received the B.S. degree in
mathematics and the B.S., M.S., and Ph.D. degrees
in electrical engineering from the Massachusetts
Institute of Technology (MIT), Cambridge, in 1980,
1980, 1981, and 1984, respectively. He is currently
a Clarence J. Lebel Professor with the Department
of Electrical Engineering, MIT. He has served as a
Codirector of the MIT Operations Research Center
from 2002 to 2005, and in the National Council on
Research and Technology in Greece (2005-2007).
His research interests are in systems, optimization,
communications, control, and operations research. He has coauthored four
books and more than a hundred journal papers in these areas.
Prof. Tsitsiklis was a recipient of an Outstanding Paper Award from
the IEEE Control Systems Society (1986),
the M.I.T. Edgerton Faculty
Achievement Award (1989), the Bodossakis Foundation Prize (1995), and
the INFORMS/CSTS Prize (1997). He is a member of the National Academy
of Engineering. Finally, in 2008, he was conferred the title of Doctor honoris
causa, from the Universit´e catholique de Louvain
A. Proof of Theorem 4
APPENDIX
Before proceeding to the main part of the proof, we start
with an elementary lemma.
Lemma 3: For any real numbers ǫ > 0, M > 1, and any
positive integer N , there exists a ∆1 > 0 and a sequence
K1, K2, . . . KN , such that:
a) Ki > M , for i = 1, . . . , N;
b) the sequence (∆i) defined by ∆i+1 = 3Ki∆i + 1
Ki
∆iKi < ǫ, for i = 1, . . . , N;
satisfies
Proof: We use induction. The result is obviously valid
for N = 1. We now assume that it holds for some N , and
prove that it also holds for N + 1. Choose some KN +1 such
that KN +1 > M . Using the induction hypothesis, choose ∆1
and a sequence K1, . . . , KN so that for i = 1, . . . N, we have
Ki∆i < ǫ
on Ki are satisfied for i = 1, . . . , N + 1, and so are those on
Ki∆i for i = 1, . . . , N. The result then follows from
and Ki > max(cid:16) 2KN +1
, M(cid:17). The conditions
6KN +1
ǫ
KN +1∆N +1 = KN +1(cid:16)3KN ∆N + 1
KN(cid:17)
2KN +1(cid:17) = ǫ.
< KN +1(cid:16)3
+ ǫ
6KN +1
ǫ
To simplify the presentation of the proof, we introduce some
new notation. For a given measure µ, we define the function
Lµ by
Lµ (y) =Z(y−1,y+1)
(y − z) dµ(z).
Thus, for any α ∈ I, we have (Lxx)(α) = Lµx (x(α)), where
µx is the measure associated to x, defined by letting µx(S)
be the Lebesgue measure of the set {α : x(α) ∈ S} for any
measurable set S. Since no ambiguity is possible here as we
only use one such measure, we will refer to µx as µ in the
sequel. We also define the nonnegative functions
and
L+
µ (y) =Z(y,y+1)
µ (y) =Z(y−1,y)
L−
(z − y) dµ(z) ≥ 0,
(y − z) dµ(z) ≥ 0,
so that Lµ = L+
<µ, we observe that if Lxx <µ δ, then the set
µ . Using the definition of the relation
µ − L−
µ (y) − L−
L+
≥ δo
µ (y)(cid:12)(cid:12)(cid:12)
S =ny ∈ [0, L] :(cid:12)(cid:12)(cid:12)
satisfies µ(S) < δ. As a consequence,
if Lxx <µ δ, then
for any z ∈ [0, L] at least one of the following must be true:
(i) there exists some y ∈ [z, L] such that L+
µ (y) + δ
and µ ([z, y)) ≤ δ; or,
(ii) we have µ([z, L]) < δ. We will make use of this
observation repeatedly in the proof of Theorem 4, which is
given below.
µ (y) < L−
Proof: Without loss of generality, we assume that L is
integer. (This is because the model with L not integer can be
viewed as a special case of a model in which opinions are
distributed on [0, ⌈L⌉].) We fix some ǫ > 0, and without loss
13
of generality we assume that ǫ < 1/2. Using Lemma 3, we
form two sequences K1, . . . , KL+1 and ∆1, . . . , ∆L+1 that
satisfy: (i) ∆i+1 = 3Ki∆i + 1
, for i = 1, . . . L; (ii) Ki >
Ki
(L + 1)/ǫ and Ki∆i < ǫ, for i = 1, . . . L+1. In particular,
∆i < ǫ2/(L + 1). We then choose some δ smaller than ∆i/3,
for all i. We will prove the following claim. If Lxx <µ δ,
then there exists some N ≤ L + 1, and two nondecreasing
finite sequences, (xi) and (yi), that satisfy
−1 = y0 < 0 ≤ x1 ≤ y1 ≤ · · · ≤ xN ≤ yN ,
the termination condition µ ((yN , L]) < ǫ2/(L + 1), and the
following additional conditions, for i = 1, . . . , N:
(a) L+
µ (xi) < ∆i;
(b) xi ≥ yi−1 + 1;
(c) µ ([yi−1, xi)) ≤ ∆i − δ;
(d) 0 ≤ yi − xi ≤ Ki∆i < ǫ.
[xi, yi]
implies
is less than ǫ. Furthermore,
The above claim, once established,
that µ is
“close” to a discrete measure whose support consists of
values that are separated by at
least 1, and provides a
proof of the theorem. To see this, note that the length of
each interval
the set
i=1[xi, yi] is covered by disjoint intervals, of the
form [0, xi), (yi−1, xi), or (yN , L]. Since intervals (yi−1, xi)
have at least unit length, the overall number of such intervals
is at most L + 1. For intervals of the form [0, xi), (yi−1, xi),
condition (c) implies that their measure is bounded above
by ∆i − δ < ∆i < ǫ2/(L + 1). Recall also the termination
condition µ ((yN , L]) < ǫ2/(L + 1). It follows that
the
[0, L] \ S N
measure of the set [0, L] \ Si[xi, yi] is at most ǫ2, hence
smaller than ǫ. Let s ∈ F be a function which for every α
takes a value xi which is closest to x(α). Since x can differ
from all xi by more than ǫ only on a set of measure smaller
than ǫ, it follows that x − s<µǫ. Finally, if Lxx =µ 0, then
Lxx <µ δ for all positive δ. As a consequence, the distance
between x and F is smaller than any positive ǫ and is thus
0. Because ¯F is the closure of F , it follows then that x ∈ ¯F .
Thus, it will suffice to provide a proof of the claim.
We will now use a recursive construction to prove the claim.
We initialize the construction as follows. Since Lxx <µ δ,
there exists some x1≥ 0 such that µ ([0, x1)) ≤ δ and
L+
µ (x1) < L−
µ (x1) + δ. Since y0 = −1, x1 satisfies condition
(b). Since δ < ∆1/3, we have µ([y0, x1)) = µ([0, x1)) ≤ δ ≤
2∆1/3 − δ < ∆1 − δ, and condition (c) is satisfied. Moreover,
L−
µ (x1) = R(x1−1,x1)(x1 − z) dµ(z)
≤ R(x1−1,x1) dµ(z)
≤ µ([0, x1)) ≤ δ.
Thus, L+
is also satisfied.
µ (x1) < L−
µ (x1) + δ ≤ 2δ < ∆1, and condition (a)
We now assume that we have chosen nonnegative x1, . . . , xi
and y1, . . . , yi−1, so that x1, . . . , xi−1 satisfy the four condi-
tions (a)-(d), and xi satisfies conditions (a)-(c). We will first
show that we can choose yi to satisfy condition (d). Then, if
µ ((yi, L]) < ǫ2/(L + 1), we will set N = i, and terminate
the construction. Otherwise, we will show that we can choose
xi+1 to satisfy conditions (a)-(c), and continue similarly. Note
that if yL has been thus constructed and the process has
not yet terminated, then the property yi+1 ≥ xi ≥ yi + 1
(from conditions (b) and (d)) implies that yL+1 ≥ L, so that
µ ((yL+1, L]) = 0, which satisfies the termination condition.
This shows that indeed N ≤ L + 1, as desired. Because all
the required conditions will be enforced, this construction will
indeed verify our claim.
The argument considers separately two different cases. For
the first case, we assume that µ ([xi, xi + 1)) ≤ δ + 1
, which
Ki
means that very few agents have opinions between xi and
xi+1. The construction described below is illustrated in Figure
7(a). We let yi = xi, so that condition (d) is trivially satisfied
by yi. If µ ((yi, L]) < ǫ2/(L+1), we let i = N and terminate.
Suppose therefore that this is not the case. Then, yi < L, and
µ ([yi, L + 1]) = µ ([yi, L]) ≥
ǫ2
L + 1
> ∆i+1.
d(cid:32)
We then have
µ ([yi + 1, L + 1]) = µ ([yi, L + 1]) − µ ([yi, yi + 1))
14
d(cid:32)
(cid:32) · (cid:38) + 1
Ki
(cid:32) · (cid:38)
xi = yi
yi + 1
xi+1
(a)
(cid:32) · (cid:38)
(cid:32) < 1
Ki
xi
yi xi + Ki(cid:6)i
(cid:32) · (cid:38)
yi + 1
xi+1
> ∆i+1 − δ − 1
Ki
> δ,
(b)
where the last inequality follows from the recurrence ∆i+1 =
3Ki∆i+ 1
(cf. Lemma 3), and the fact that 2δ < ∆i < Ki∆i
Ki
for all i. In particular, we must have yi + 1 ≤ L. Using the
assumption Lxx <µ δ, and an earlier observation, we can
find some xi+1≥ yi + 1 such that L+
µ (xi+1) + δ
and µ([yi + 1, xi+1))≤δ. Condition (b) then trivially holds for
xi+1. Remembering that yi = xi, we also have
µ (xi+1) < L−
µ ([yi, xi+1)) = µ ([xi, xi + 1)) + µ ([yi + 1, xi+1))
≤ δ + 1
Ki
+ δ ≤ ∆i+1 − δ,
where the last inequality follows, as before, from the recur-
rence ∆i+1 = 3Ki∆i + 1
, and from δ ≤ Ki∆i. As a
Ki
result, xi+1 satisfies condition (c). To prove that xi+1 satisfies
condition (a), observe that
L−
µ (xi+1) = R(xi+1−1,xi+1)(xi+1 − z) dµ(z)
≤ µ ([xi+1 − 1, xi+1))
≤ µ ([yi, xi+1)) .
where the last inequality follows from condition (b) for xi+1.
Because xi+1 satisfies condition (c), we have L−
µ (xi+1) ≤
∆i+1 − δ. Then, condition (a) for xi+1 follows from the fact
that xi+1 has been chosen so that L+
µ (xi+1) + δ.
We now consider the second case, where µ ([xi, xi + 1)) >
. Our construction is illustrated in Figure 7(b). We claim
. This is because otherwise
δ+ 1
Ki
that µ ([xi + Ki∆i, xi + 1)) ≤ 1
Ki
we would have
µ (xi+1) < L−
L+
µ (xi) = R(xi,xi+1)(z − xi) dµ(z)
≥ R(xi+Ki∆i,xi+1)(z − xi) dµ(z)
≥ Ki∆iµ ([xi + Ki∆i, xi + 1))
> ∆i,
contradicting condition (a) for xi. This implies that µ([xi, xi +
K∆i]) > δ. It follows that we can choose yi ≥ xi
so that yi ≤ xi + Ki∆i, µ ([yi, xi + Ki∆i)) ≤ δ, and
L+
µ (yi) ≤ L−
µ (yi) + δ. Then, condition (d) is satisfied by
yi. If µ ([yi, L]) < ǫ2/(L + 1), we let i = N and terminate.
Illustration of the iterative construction in the proof of Theorem 4,
Figure 7.
when (a) µ ([xi, xi + 1)) ≤ δ + 1
. For
Ki
the sake of clarity, the figure shows the density dµ as if it were continuous,
but the proof does not use the continuity, or even existence, of a density.
, and (b) µ ([xi, xi + 1)) > δ + 1
Ki
this is not the case. By the same
Suppose therefore that
argument as for the previous case, we can then choose xi+1
so that yi + 1 ≤ xi+1 ≤ L, L+
µ (xi+1) + δ, and
µ ([yi + 1, xi+1)) < δ. Thus xi+1 satisfies condition (b).
µ (xi+1) ≤ L−
To prove that xi+1 satisfies the remaining two conditions,
(c) and (a), we need an upper bound on µ ([xi + 1, yi + 1)).
Observe first that
L+
µ (yi) = R[yi,yi+1)(z − yi) dµ(z)
≥ R[xi+1,yi+1)(z − yi) dµ(z)
≥ (1 + xi − yi) µ ([xi + 1, yi + 1))
≥ 1
2 µ ([xi + 1, yi + 1)) ,
(10)
where the last inequality follows from the fact 1 + xi − yi ≥
1 − ǫ (condition (d)), and the fact that ǫ was assumed smaller
than 1
2 . Thus, to derive upper bound on µ ([xi + 1, yi + 1))
it will suffice to derive an upper bound on L+
µ (yi). We start
with the inequality, L+
µ (yi) + δ, and also make use
of the fact yi − 1 ≥ xi − 1 ≥ yi−1, which is a consequence
of conditions (b) and (d) for xi. We obtain
L−
µ (yi) ≤ R[yi−1,xi)(yi − z) dµ(z) +R[xi,yi)(yi − z) dµ(z)
≤ µ ([yi−1, xi)) + µ ([xi, yi)) (yi − xi)
≤ ∆i − δ + Ki∆i,
µ (yi) ≤ L−
where the last
inequality follows from conditions (c) and
(d) for xi, and the fact that µ ([xi, yi)) ≤ µ ([0, L]) = 1.
Combining this with the lower bound (10) leads to
µ ([xi + 1, yi + 1)) ≤ 2(Ki + 1)∆i,
(11)
which is the desired upper bound on µ ([xi + 1, yi + 1)).
We will now use the above upper bound to prove conditions
(a) and (c) for xi+1. Observe that µ ([yi, xi+1)) can be
15
expressed as
µ ([yi, xi + Ki∆i)) + µ ([xi + Ki∆i, xi + 1))
+ µ ([xi + 1, yi + 1)) + µ ([yi + 1, xi+1)) .
Recall that yi has been chosen so that µ ([yi, xi + Ki∆i)) ≤
δ, and xi+1 so that µ ([yi + 1, xi+1)) ≤ δ. Moreover,
µ ([xi + Ki∆i, xi + 1)) has been shown to be no greater than
1/Ki. It then follows from (11) that
µ ([yi, xi+1)) ≤ 2δ +
1
Ki
+ 2(Ki + 1)∆i ≤ 3Ki∆i +
1
Ki
− δ,
where we have used the facts that 3δ ≤ ∆i and Ki ≥ 3.
Condition (c) for xi+1 follows from the property ∆i+1 =
3Ki∆i+ 1
in the definition of the sequence (∆i) (see Lemma
Ki
3). To prove condition (a), we observe that
amounts to noting that the numerator and denominator in the
definition of ux(a) are both Lipschitz continuous functions of
a, and that the denominator is bounded below by m. The proof
of the lower bound is essentially a strengthening of the proof of
Proposition 1, which only established that ux(b)−ux(a) ≥ 0.)
With our choice of δ, we have either a ≥ inf α x + 1 or b ≤
supα x − 1. We only consider the second case, so that (a, b +
1) ⊆ [inf α x, supα x]; the first case can be treated similarly.
Let ¯µa\b = µx ((a − 1, b − 1]), ¯µab = µx ((b − 1, a + 1)), and
¯µb\a = µx ([a + 1, b + 1)). Let also ¯xa\b, ¯xab, and ¯xb\a be the
center of mass of the opinions of those agents whose opinions
lie in the set (a − 1, b − 1], (b − 1, a + 1), and [a + 1, b + 1),
respectively. (In case ¯µa\b = 0, we use the convention xa\b =
b − 1.)
From the definition of ux, we have
L−
µ (xi+1) = R(xi+1−1,xi+1)(xi+1 − z) dµ(z)
≤ µ ((xi+1 − 1, xi+1)) ,
ux(a) =
¯µab ¯xab + ¯µa\b ¯xa\b
¯µab + ¯µa\b
= ¯xab −
¯µa\b(¯xab − ¯xa\b)
¯µab + ¯µa\b
and then use conditions (b) and (c) for xi+1 to obtain
and
ux(b) =
¯µab ¯xab + ¯µb\a ¯xb\a
¯µab + ¯µb\a
= ¯xab +
¯µb\a(¯xb\a − ¯xab)
¯µab + ¯µb\a
,
.
L−
µ (xi+1) ≤ µ ([yi, xi+1)) ≤ ∆i − δ.
This completes the induction and the proof of Theorem 4.
B. Proof of Theorem 6
Proof: Suppose s does not satisfy the condition of this
theorem, and that a < b. Since s ∈ F , we have µs ((a, b)) = 0.
Let Sa, Sb ⊂ I be two sets on which s takes the values a and
b, respectively. We choose these sets so that the Lebesgue
measure of Sa ∪ Sb is δ, and so that the ratio Sa/Sb of
their measures is equal to µs(a)/µs(b). Let x0(α) = s(α) for
α /∈ Sa ∪ Sb, and x0(α) = µs(a)a+µs(b)b
for α ∈ Sa ∪ Sb.
µs(a)+µs(b)
Observe that µx0 (a)
µx0 (b) = µs(a)
µs(b) .
As already discussed, when x0 takes discrete values, the
evolution of xt is entirely characterized by the evolution of
a corresponding weighted discrete-agents system of the form
(2). We can then apply the reasoning in the proof of Theorem
2 to show that the two clusters initially at a and b converge
to a single cluster. Since this can be done for any, arbitrarily
small δ > 0, s is unstable.
C. Proof of Proposition 3
To ease the reading of the proof, we introduce a new
notation. For any x ∈ XL, we let ux : [0, L] → [0, L] be
a function defined so that ux(a) is the updated opinion of an
agent that held opinion a, namely
ux(a) = R a+1
a−1 z dµx(z)
µx ((a − 1, a + 1))
.
As a consequence, (U (x))(α) = ux(x(α)) and xt+1(α) =
uxt (xt(α)), for any α ∈ I.
Proof: Since x is regular, there exist m and M , with 0 <
m ≤ M , such that for any [a, b] ⊆ [inf α x, supα x] we have
m(b − a) ≤ µx([a, b]) ≤ M (b − a). Let δ = min{ 1
2 , supα x −
inf α x − 2}. We first prove the existence of M ′, m′ > 0 such
that if [a, b] ⊆ [inf α x, supα x] and b−a < δ, then m′(b−a) ≤
ux(b) − ux(a) ≤ M ′(b − a). (The proof of the upper bound
Note that a − 1 ≤¯xa\b ≤ ¯xab ≤ ¯xb\a≤ b + 1, so that ¯xb\a −
¯xab ≤ 2 + (b − a) ≤ 3, and similarly, ¯xab − ¯xa\b ≤ 3. From
the regularity assumption, we also have ¯µb\a ≤ M (b − a) and
¯µab = µx ((b − 1, a + 1)) ≥ µx ((a, a + 1)) ≥ m. Thus,
ux(b) − ux(a) ≤ 3
¯µb\a
¯µab
+ 3
¯µa\b
¯µab
≤
3M (b − a)
m
,
which proves the claimed upper bound with M ′ = 3 M
m .
For the lower bound, an elementary calculation shows that
if we have a density function on the interval [0, 1], which is
bounded above and below by M and m, respectively, then its
center of mass is at least m/2M . By applying this fact to the
interval (b − 1, a + 1) (which has length larger than 1), we
conclude that its center of mass, ¯xab is at least m/2M below
the right end-point a + 1. Since also ¯xb\a ≥ a + 1, we have
¯xb\a − ¯xab ≥ m/2M , and
¯µb\a
m(b − a)
3M
·
,
≥ ux(a) +
·
m
2M
¯µab + ¯µb\a
ux(b) ≥ ¯xab +
[a, b] ⊆ [inf α x, supα x]
m
2M
where the last inequality made use of the facts µb\a ≥ m(b−a)
and ¯µab + ¯µb\a ≤ 3M . This establishes the claimed lower
bound, with m′ = m2/6M 2.
By splitting an interval
into
subintervals of length bounded by δ, we see that the result
m′(b − a) ≤ ux(b) − ux(a) ≤ M ′(b − a) also holds for
general such intervals. Consider now an interval [a′, b′] ∈
[inf α U (x), supα U (x)], and let a = inf{z ∈ [0, L] : ux(z) ∈
[a′, b′]} and b = sup{z ∈ [0, L] : ux(z) ∈ [a′, b′]}. As a conse-
quence of the order preservation property, ux ((a, b)) ⊆ [a′, b′],
and [a′, b′] ⊆ [ux(a), ux(b)]. Since x is regular, we have
µx(a) = µx(b) = 0, which implies that µU(x)([a′, b′]) =
µx([a, b]) ∈ [m(b − a), M (b − a)]. Using the bounds on
ux(b)−ux(a)
, we finally obtain the desired result
b−a
mm′(b′ − a′) ≤ µU(x)([a′, b′]) ≤ M M ′(b′ − a′).
|
1110.2767 | 1 | 1110 | 2011-10-12T19:23:10 | Resource Allocation Among Agents with MDP-Induced Preferences | [
"cs.MA",
"cs.AI"
] | Allocating scarce resources among agents to maximize global utility is, in general, computationally challenging. We focus on problems where resources enable agents to execute actions in stochastic environments, modeled as Markov decision processes (MDPs), such that the value of a resource bundle is defined as the expected value of the optimal MDP policy realizable given these resources. We present an algorithm that simultaneously solves the resource-allocation and the policy-optimization problems. This allows us to avoid explicitly representing utilities over exponentially many resource bundles, leading to drastic (often exponential) reductions in computational complexity. We then use this algorithm in the context of self-interested agents to design a combinatorial auction for allocating resources. We empirically demonstrate the effectiveness of our approach by showing that it can, in minutes, optimally solve problems for which a straightforward combinatorial resource-allocation technique would require the agents to enumerate up to 2^100 resource bundles and the auctioneer to solve an NP-complete problem with an input of that size. | cs.MA | cs | Journal of Artificial Intelligence Research 27 (2006) 505–549
Submitted 06/06; published 12/06
Resource Allocation Among Agents with MDP-Induced
Preferences
Dmitri A. Dolgov
Technical Research Department (AI & Robotics Group)
Toyota Technical Center
2350 Green Road
Ann Arbor, MI 48105, USA
Edmund H. Durfee
Electrical Engineering and Computer Science
University of Michigan
2260 Hayward St.
Ann Arbor, MI 48109, USA
[email protected]
[email protected]
Abstract
Allocating scarce resources among agents to maximize global utility is, in general, com-
putationally challenging. We focus on problems where resources enable agents to execute
actions in stochastic environments, modeled as Markov decision processes (MDPs), such
that the value of a resource bundle is defined as the expected value of the optimal MDP
policy realizable given these resources. We present an algorithm that simultaneously solves
the resource-allocation and the policy-optimization problems. This allows us to avoid ex-
plicitly representing utilities over exponentially many resource bundles, leading to drastic
(often exponential) reductions in computational complexity. We then use this algorithm
in the context of self-interested agents to design a combinatorial auction for allocating re-
sources. We empirically demonstrate the effectiveness of our approach by showing that
it can, in minutes, optimally solve problems for which a straightforward combinatorial
resource-allocation technique would require the agents to enumerate up to 2100 resource
bundles and the auctioneer to solve an NP-complete problem with an input of that size.
1. Introduction
The problem of resource allocation is ubiquitous in many diverse research fields such as
economics, operations research, and computer science, with applications ranging from de-
centralized scheduling (e.g., Wellman, Walsh, Wurman, & MacKie-Mason, 2001) and net-
work routing (e.g., Feldmann, Gairing, Lucking, Monien, & Rode, 2003) to transportation
logistics (e.g., Sheffi, 2004; Song & Regan, 2002) and bandwidth allocation (e.g., McMillan,
1994; McAfee & McMillan, 1996), just to name a few. The core question in resource alloca-
tion is how to distribute a set of scarce resources among a set of agents (either cooperative or
self-interested) in a way that maximizes some measure of global utility, with social welfare
(sum of agents’ utilities) being one of the most popular criteria.
In many domains, an agent’s utility for obtaining a set of resources is defined by what
the agent can accomplish using these resources. For example, the value of a vehicle to a
delivery agent is defined by the additional revenue that the agent can obtain by using the
vehicle. However, to figure out how to best utilize a resource (or set of resources), an agent
c(cid:13)2006 AI Access Foundation. All rights reserved.
Dolgov & Durfee
Figure 1: Dependency Cycle: To formulate their planning problems, the agents need to
know what resources they will get, but their utility functions, which define the
input to the resource-allocation problem, depend on the solutions to the planning
problems.
often must solve a non-trivial planning problem because actions might have long-term, non-
deterministic effects. Therefore, an agent’s value for a set of resources is defined by a solution
to a planning problem, but to formulate its planning problem the agent needs to know what
resources it will obtain. This leads to cyclic dependencies (depicted in Figure 1), wherein the
input to the resource allocation problem depends on the solution to the planning problem,
and vice versa. Unfortunately, for anything but the simplest domains, neither the resource-
allocation nor the planning problem can be solved in closed form, making it impossible to
obtain parameterized solutions.
Our focus in this paper is on solving the interdependent problems of resource allocation
and stochastic planning. The main question we consider is how to allocate the resources in
a way that maximizes the social welfare of the agents when the utility function of each agent
is defined by a Markov decision process (Puterman, 1994) whose action set is parameterized
by the resources. In this paper, we specifically focus on non-consumable resources (such
as vehicles) that enable actions, but are not themselves consumed during action execution.
We briefly mention the case of consumable resources in Section 6, and refer to the work by
Dolgov (2006) for a more detailed treatment.
We assume that the agents’ MDPs are weakly-coupled, meaning the agents only interact
through the resources, and once the resources are allocated, the transition and reward
functions of their MDPs are independent. Our model of weakly-coupled MDPs connected
via shared resources is similar to that of Meuleau, Hauskrecht, Kim, Peshkin, Kaelbling,
Dean, and Boutilier (1998) and Benazera, Brafman, Meuleau, and Hansen (2005), but
differs in that we further assume that resources are only allocated once, prior to any actions
being taken. While this “one-shot” allocation assumption limits our approach somewhat,
it also allows our approach to apply more broadly to non-cooperative settings (without
this assumption, the game-theoretic analysis of agents’ interactions is significantly more
complex). More importantly, it allows us to avoid the state space explosion (due to including
resource information in the MDP states), which limits that other work to finding only
approximately optimal solutions for non-trivial problems.
The main result presented in this paper is thus a new algorithm that, under the above
conditions, optimal ly solves the resource-al location and the policy-optimization problems
simultaneously. By considering the two problems together, it sidesteps the dependency cycle
506
Resource Allocation Among Agents with MDP-Induced Preferences
mentioned above, which allows us to avoid an explicit representation of utility functions
on resource bundles, leading to an exponential reduction in complexity over combinatorial
resource allocation with flat utility functions. We empirically demonstrate that the resulting
algorithm scales well to finding optimal solutions for problems involving numerous agents
and resources.
Our algorithm can be viewed as contributing a new approach for dealing with the com-
putational complexity of resource allocation in domains with complex utility functions that
are not linearly decomposable in the resources (due to the effects of substitutability and
complementarity).
In such combinatorial al location problems, finding an optimal alloca-
tion is NP-complete in the (often exponentially large) space of resource bundles (Rothkopf,
Pekec, & Harstad, 1998). Previous approaches for addressing the complexity have included
determining classes of utility functions that lead to tractable problems (as surveyed by
de Vries & Vohra, 2003), iterative algorithms for resource allocation and preference elici-
tation (as surveyed by Sandholm & Boutilier, 2006), and concise languages for expressing
agents’ preferences (Sandholm, 1999; Nisan, 2000; Boutilier & Hoos, 2001; Boutilier, 2002).
The novelty of our approach with respect to these is that it explicitly embraces the under-
lying processes that define the agents’ utility functions, for cases where these processes can
be modeled as resource-parameterized MDPs. By doing so, not only does our approach use
such MDP-based models as a concise language for agents’ utility functions, but more impor-
tantly, it directly exploits the structure in these models to drastically reduce computational
complexity by simultaneously solving the planning and resource-allocation problems.
In the context of cooperative agents, our approach can be viewed as a way of solving
weakly-coupled multiagent MDPs, where agents’ transition and reward functions are inde-
pendent, but the space of joint actions is constrained, as, for example, in the models used
by Singh and Cohn (1998) or Meuleau et al. (1998). From that perspective, the concept of
resources can be viewed as a compact way of representing the interactions between agents,
similarly to the model used by Bererton, Gordon, and Thrun (2003); however, our work
differs in a number of assumptions. Moreover, our algorithms can be easily modified to work
with models where the constraints on the joint actions are modeled directly (for example,
via SAT formulas).
For non-cooperative agents, we apply our resource-allocation algorithm to the mechanism-
design problem (e.g., Mas-Colell, Whinston, & Green, 1995), where the goal is to allocate
resources among the agents in a way that maximizes the social welfare, given that each
participating agent is selfishly maximizing its own utility. For domains with self-interested
agents with complex preferences that exhibit combinatorial effects between the resources,
combinatorial auctions (e.g., de Vries & Vohra, 2003) are often used for resource-allocation.
The Generalized Vickrey Auction (GVA) (MacKie-Mason & Varian, 1994), which is an ex-
tension of Vickrey-Clarke-Groves (VCG) mechanisms (Vickrey, 1961; Clarke, 1971; Groves,
1973) to combinatorial auctions, is particularly attractive because of its nice analytical
properties (as described in Section 4.1). We develop a variant of a VCG auction, where
agents submit their resource-parameterized MDPs as bids, and the auctioneer simultane-
ously solves the resource-allocation and the policy-optimization problems, thus retaining the
compact representation of agents’ preferences throughout the process. We describe exten-
sions to the mechanism for distributing the computation and for encoding MDP information
to reduce the revelation of private information.
507
Dolgov & Durfee
The remainder of this paper proceeds as follows. After a brief review of MDPs in
Section 2, we present (in Section 3) our model of a decision-making agent: the resource-
parameterized MDP with capacity constraints. We analyze the problem of optimal policy
formulation in such a resource-parameterized capacity-constrained MDP, study its prop-
erties, and present a solution algorithm, based on the formulation of this (NP-complete)
problem as a mixed integer program.
With these building blocks, we move to the multiagent setting and present our main
result, the algorithm for simultaneously allocating resources and planning across agents
(Section 4). Based on that algorithm, we then design a combinatorial auction for allocating
resources among self-interested agents. We describe a distributed implementation of the
mechanism, and discuss techniques for preserving information privacy.
In Section 5, we
analyze the computational efficiency of our approach, empirically demonstrating exponen-
tial reductions in computational complexity, compared to a straightforward combinatorial
resource-allocation algorithm with flat utility functions. Finally, in Section 6, we conclude
with a discussion of possible generalizations and extensions to our approach. For conciseness
and better readability, all proofs and some generalizations are deferred to appendices.
2. Markov Decision Processes
We base our model of agents’ decision problems on infinite-horizon fully-observable MDPs
with the total expected discounted reward optimization criterion (although our results are
also applicable to other classes of MDPs, such as MDPs with the average per-step rewards).
This section introduces our notation and assumptions, and serves as a brief overview of the
basic MDP results (see, for example, the text by Puterman (1994) for a detailed discussion
of the material in this section).
A classical single-agent, unconstrained, stationary, fully-observable MDP can be defined
as a 4-tuple hS , A, p, ri, where:
• S is a finite set of states the agent can be in.
• A is a finite set of actions the agent can execute.
• p : S × A × S 7→ [0, 1] defines the transition function. The probability that the agent
goes to state σ ∈ S upon execution of action a ∈ A in state s ∈ S is p(σ s, a).
We assume that, for any action, the corresponding transition matrix is stochastic:
Pσ p(σ s, a) = 1 ∀s ∈ S , a ∈ A.
• r : S × A 7→ R defines the reward function. The agent obtains a reward of r(s, a) if it
executes action a ∈ A in state s ∈ S . We assume the rewards are bounded.
In a discrete-time fully-observable MDP, at each time step, the agent observes the current
state of the system and chooses an action according to its policy. A policy is said to be
Markovian (or history-independent ) if the choice of action does not depend on the history
of states and actions encountered in the past, but rather only on the current state and
time. If, in addition to that, the policy does not depend on time, it is called stationary. By
definition, a stationary policy is always Markovian. A deterministic policy always prescribes
the execution of the same action in a state, while a randomized policy chooses actions
according to a probability distribution.
508
Resource Allocation Among Agents with MDP-Induced Preferences
Following the standard notation (Puterman, 1994), we refer to different classes of policies
as Πxy , where x = {H, M , S } specifies whether a policy is History-dependent, Markovian,
or Stationary, and y = {R, D} specifies whether the policy is Randomized or Deterministic
(e.g., the class of stationary deterministic policies is labeled ΠSD ). Obviously, ΠHy ⊃ ΠMy ⊃
ΠSy and ΠxR ⊃ ΠxD , with history-dependent randomized policies ΠHR and stationary
deterministic policies ΠSD being the most and the least general, respectively.
A stationary randomized policy is thus a mapping of states to probability distributions
over actions: π : S × A 7→ [0, 1], where π(s, a) defines the probability that action a is
executed in state s. A stationary deterministic policy can be viewed as a degenerate case
of a randomized policy for which there is only one action for each state that has a nonzero
probability of being executed.
In an unconstrained discounted MDP, the goal is to find a policy that maximizes the
total expected discounted reward over an infinite time horizon:1
(γ )t rt(π , α)i,
Uγ (π , α) = Eh ∞Xt=0
where γ ∈ [0, 1) is the discount factor (a unit reward at time t + 1 is worth the same to the
agent as a reward of γ at time t), rt is the (random) reward the agent receives at time t,
whose distribution depends on the policy π and the initial distribution over the state space
α : S 7→ [0, 1].
One of the most important results in the theory of MDPs states that, for an uncon-
strained discounted MDP with the total expected reward optimization criterion, there al-
ways exists an optimal policy that is stationary, deterministic, and uniformly optimal, where
the latter term means that the policy is optimal for all distributions over the starting state.2
There are several commonly-used ways of finding the optimal policy, and central to all of
them is the concept of a value function of a policy, vπ : S 7→ R, where vπ (s) is the expected
cumulative value of the reward the agent would receive if it started in state s and behaved
according to policy π . For a given policy π , the value of every state is the unique solution
to the following system of S linear equations:
r(s, a)π(s, a) + γ Xσ
vπ (s) = Xa
To find the optimal policy, it is handy to consider the optimal value function v∗ : S 7→ R,
where v∗ (s) represents the value of state s, given that the agent behaves optimally. The
optimal value function satisfies the following system of S nonlinear equations:
p(σ s, a)v∗ (σ)i,
a hr(s, a) + γ Xσ
v∗ (s) = max
Given the optimal value function v∗ , an optimal policy is to simply act greedily with respect
to v∗ (with any method of tie-breaking in case of multiple optimal actions):
π(s, a) = (1 if a ∈ arg maxa hr(s, a) + γ Pσ p(σ s, a)v∗ (σ)i,
0 otherwise.
1. Notation: here and below (x)y is an exponent, while xy is a superscript.
2. Uniform optimality of policies is the reason why α is not included as a component of a textbook MDP.
p(σ s, a)vπ (σ),
∀s ∈ S .
∀s ∈ S .
(2)
(1)
(3)
(4)
509
Dolgov & Durfee
α(s)v(s)
r(s, a)x(s, a)
p(σ s, a)v(σ),
∀s ∈ S , a ∈ A,
One of the possible ways of solving for the optimal value function is to formulate the
nonlinear system (3) as a linear program (LP) with S optimization variables v(s) and
S A constraints:
min Xs
sub ject to:
v(s) ≥ r(s, a) + γ Xσ
where α is an arbitrary constant vector with S positive components (α(s) > 0 ∀s ∈ S ).3
In many problems (including the ones that are the focus of this paper), it is very use-
ful to consider the equivalent dual LP with S A optimization variables x(s, a) and S
constraints:4
x Xs Xa
max
sub ject to:
Xa
x(σ, a) − γ Xs Xa
x(s, a) ≥ 0
x(s, a)p(σ s, a) = α(σ),
∀s ∈ S , a ∈ A.
∀σ ∈ S ;
(6)
(5)
The optimization variables x(s, a) are often called the visitation frequencies or the occupa-
tion measure of a policy. If we think of α as the initial probability distribution, then x(s, a)
can be interpreted as the total expected number of times action a is executed in state s.
Then, x(s) = Pa x(s, a) gives the total expected flow through state s, and the constraints
in the above LP can be interpreted as the conservation of flow through each of the states.
An optimal policy can be computed from a solution to the dual LP as:
x(s, a)
Pa x(s, a)
,
(7)
π(s, a) =
where non-negativity of α guarantees that Pa x(s, a) > 0 ∀s ∈ S . In general, this appears to
lead to randomized policies. However, a bounded LP with n constraints always has a basic
feasible solution (e.g., Bertsimas & Tsitsiklis, 1997), which by definition has no more than
n non-zero components. If α is strictly positive, a basic feasible solution to the LP (7) will
have precisely S nonzero components (one for each state), which guarantees the existence
of an optimal deterministic policy. Such a policy can be easily obtained by most LP solvers
(e.g., simplex will always produce solutions that map to deterministic policies).
Furthermore, as mentioned above, for unconstrained discounted MDPs, there always
exist policies that are uniformly optimal (optimal for all initial distributions). The above
dual LP (6) yields uniformly optimal policies if a strictly positive α is used. However, the
3. The overloading of α as the ob jective function coefficients here and the initial probability distribution
of an MDP earlier is intentional and is explained shortly.
4. Note that some authors (e.g., Altman, 1996) prefer the opposite convention, where (6) is called the dual,
and (5), the primal.
510
Resource Allocation Among Agents with MDP-Induced Preferences
solution (x) to the dual LP retains its interpretation as the expected number of times state s
is visited and action a is executed only for the initial probability distribution α that was
used in the LP.
The main benefit of the dual LP (6) is manifested in constrained MDPs (Altman, 1999;
Kallenberg, 1983; Heyman & Sobel, 1984), where each action, in addition to producing a
reward r(s, a), also incurs a vector of costs ηk (s, a) : S × A 7→ R ∀k ∈ [1..K ]. The prob-
lem then is to maximize the expected reward, sub ject to constraints on the expected costs.
Constrained models of this type arise in many domains, such as telecommunication appli-
cations (e.g., Ross & Chen, 1988; Ross & Varadara jan, 1989), where it is often desirable to
maximize expected throughput, sub ject to conditions on the average delay. Such problems,
where constraints are imposed on the expected costs, can be solved in polynomial time
using linear programming by simply augmenting the dual LP (6) with the following linear
constraints:
Xs Xa
ηk (s, a)x(s, a) ≤ bηk ,
where bηk is the upper bound on the expected cost of type k . The resulting constrained MDP
differs from the standard unconstrained MDP: in particular, deterministic policies are no
longer optimal, and uniformly optimal policies do not, in general, exist for such problems
(Kallenberg, 1983).
For the same reason of being easily augmentable with constraints, the dual LP (6) also
forms the basis for our approach. However, the constraints that arise in resource-allocation
problems that are our focus in this paper are very different from the linear constraints in (8),
leading to different optimization problems with different properties and requiring different
solution techniques (as described in more detail in Section 3).
We conclude the background section by introducing a simple unconstrained MDP that
will serve as the basis for a running example, to which we will refer throughout the rest of
this paper.
∀k ∈ [1..K ],
(8)
Example 1 Consider a simple delivery domain, depicted in Figure 2, where the agent can
obtain rewards for delivering furniture (action a1 ) or delivering appliances (action a2 ).
Delivering appliances produces higher rewards (as shown on the diagram), but it does more
damage to the delivery vehicle. If the agent only delivers furniture, the damage to the vehicle
is negligible, whereas if the agent delivers appliances, the vehicle is guaranteed to function
reliably for the first year (state s1 ), but after that (state s2 ) has a 10% probability of failure,
per year. The vehicle can be serviced (action a3 ), resetting its condition, but at the expense
of lowering profits. If the truck does break (state s3 ), it can be repaired (action a4 ), but with
a more significant negative impact on profits. We wil l assume a discount factor γ = 0.9.
The optimal value function is v∗ (s1 ) ≈ 95.3, v∗ (s2 ) ≈ 94.7, v∗ (s3 ) ≈ 86.7, and the
corresponding optimal occupation measure (assuming a uniform α) is the fol lowing (listing
only the non-zero elements): x(s1 , a2 ) ≈ 4.9, x(s2 , a3 ) ≈ 4.8, x(s3 , a4 ) ≈ 0.3. This maps to
the optimal policy that dictates that the agent is to start by delivering appliances (action a2
in state s1 ), then service the vehicle after the first year (action a3 in state s2 ), and fix the
vehicle if it ever gets broken (a4 in s3 ) (the latter has zero probability of happening under
this policy if the agent starts in state s1 or s2 ).
(cid:3)
511
Dolgov & Durfee
Figure 2: Unconstrained MDP example for a delivery domain. Transition probabilities (p)
and rewards (r) are shown on the diagram. Actions not shown result in transition
to same state with no reward. There is also a noop action a0 that corresponds to
doing nothing; it does not change state and produces zero reward.
3. Agent Model: Resource-Parameterized MDP
In this section, we introduce our model of the decision-making agent, and describe the
single-agent stochastic policy-optimization problem that defines the agent’s preferences over
resources. We show that, for this single-agent problem, formulated as an MDP whose
action set is parameterized on the resources available to the agent, stationary deterministic
policies are optimal, but uniformly optimal policies do not, in general, exist. We also show
that the problem of finding optimal policies is NP-complete. Finally, we present a policy-
optimization algorithm, based on a formulation of the problem as a mixed integer linear
program (MILP).
We model the agent’s resource-parameterized MDP as follows. The agent has a set of
actions that are potentially executable, and each action requires a certain combination of
resources. To capture local constraints on the sets of resources an agent can use, we use
the concept of capacities: each resource has capacity costs associated with it, and each
agent has capacity constraints. For example, a delivery company needs vehicles and loading
equipment (resources to be allocated) to make its deliveries (execute actions). However,
all equipment costs money and requires manpower to operate it (the agent’s local capacity
costs). Therefore, the amount of equipment the agent can acquire and successfully utilize
is constrained by factors such as its budget and limited manpower (agent’s local capacity
bounds). This two-layer model with capacities and resources represented separately might
seem unnecessarily complex (why not fold them together or impose constraints directly
on resources?), but the separation becomes evident and useful in the multiagent model
discussed in Section 4. We emphasize the difference here: resources are the items being
allocated among the agents, while capacities define the inherent limitations of an individual
agent on what combinations of resources it can usefully possess.
The agent’s optimization problem is to choose a subset of the available resources that
does not violate its capacity constraints, such that the best policy feasible under that bundle
512
Resource Allocation Among Agents with MDP-Induced Preferences
of resources yields the highest utility. In other words, the single-agent problem analyzed in
this section has no constraints on the total resource amounts (they are introduced in the
multiagent problem in the next section), and the constraints are only due to the agent’s
capacity limits. Adding limited resource amounts to the single-agent model would be a very
simple matter, since such constraints can be handled with a simple pruning of the agent’s
action space. Further, note that without capacity constraints, the single-agent problem
would be trivial, as it would always be optimal for the agent to simply take all resources
that are of potential use. However, in the presence of capacity constraints, we face a problem
that is similar to the cyclic dependency in Figure 1: the resource-selection problem requires
knowing the values of all resource bundles, which are defined by the planning problem, and
the planning problem is ill-defined until a resource bundle is chosen. In this section, we
focus on the single-agent problem of selecting an optimal subset of resources that satisfies
the agent’s capacity constraints and assume the agent has no value for acquiring additional
resources that exceed its capacity bounds.
The resources in the model outlined above are non-consumable, i.e., actions require
resources but do not consume them during execution. As mentioned in the Introduction,
in this work we focus only on non-consumable resources, and only briefly outline the case
of consumable resources in Section 6.
We can model the agent’s optimization problem as an n-tuple hS , A, p, r, O , ρ, C , κ, bκ , αi:
• hS , A, p, ri are the standard components of an MDP, as defined earlier in Section 2.
• O is the set of resources (e.g., O = {production equipment, vehicle, . . .}). We will use
o ∈ O to refer to a resource type.
• ρ : A × O 7→ R is a function that specifies the resource requirements of all actions;
ρ(a, o) defines how much of resource o ∈ O action a ∈ A needs to be executable (e.g.,
ρ(a, vehicle) = 1 means that action a requires one vehicle).
• C is the set of capacities of our agent (e.g., C = {space, money, manpower, . . .}). We
will use c ∈ C to refer to a capacity type.
• κ : O × C 7→ R is a function that specifies the capacity costs of resources; κ(o, c) defines
how much of capacity c ∈ C a unit of resource o ∈ O requires (e.g., κ(vehicle, money) =
$50000 defines the monetary cost of a vehicle, while κ(vehicle , manpower) = 2 means
that two people are required to operate the vehicle).
• bκ : C 7→ R specifies the upper bound on the capacities; bκ(c) gives the upper bound
on capacity c ∈ C (e.g., bκ(money) = $1,000,000 defines the budget constraint, and
bκ(manpower) = 7 specifies the size of the workforce).
• α : S 7→ R is the initial probability distribution; α(s) is the probability that the agent
starts in state s.
Our goal is to find a policy π that yields the highest expected reward, under the con-
ditions that the resource requirements of that policy do not exceed the capacity bounds of
513
Dolgov & Durfee
the agent. In other words, we have to solve the following mathematical program:5
(9)
∀c ∈ C ,
max
Uγ (π , α)
π
sub ject to:
π(s, a)(cid:1)o ≤ bκ(c),
a nρ(a, o)H (cid:0) Xs
Xo
κ(o, c) max
where H is the Heaviside “step” function of a nonnegative argument, defined as:
H (z ) = ( 0 z = 0,
1 z > 0.
The constraint in (9) can be interpreted as follows. The argument of H is nonzero if
the policy π assigns a nonzero probability to using action a in at least one state. Thus,
H (Ps π(s, a)) serves as an indicator function that tells us whether the agent plans to use
action a in its policy, and max (cid:8)ρ(a, o)H (Ps π(s, a))(cid:9) tells us how much of resource o the
agent needs for its policy. We take a max with respect to a, because the same resource o can
be used by different actions. Therefore, when summed over all resources o, the left-hand
side gives us the total requirements of policy π in terms of capacity c, which has to be no
greater than the bound bκ(c).
The following example illustrates the single-agent model.
Example 2 Let us augment Example 1 as fol lows. Suppose the agents needs to obtain a
truck to perform its delivery actions (a1 and a2 ). The truck is also required by the service
and repair actions (a3 and a4 ). Further, to deliver appliances, the agent needs to acquire
a forklift, and it needs to hire a mechanic to be able to repair the vehicle (a4 ). The noop
action a0 requires no resources. This maps to a model with three resources (truck, forklift,
and mechanic): O = {ot , of , om }, and the fol lowing action resource costs (listing only the
non-zero ones):
ρ(a1 , ot ) = 1, ρ(a2 , ot ) = 1, ρ(a2 , of ) = 1, ρ(a3 , ot ) = 1, ρ(a4 , ot ) = 1, ρ(a4 , om ) = 1.
Moreover, suppose that the resources (truck ot , forklift of , or mechanic om ) have the
fol lowing capacity costs (there is only one capacity type, money: C = {c1 })
κ(ot , c1 ) = 2, κ(of , c1 ) = 3, κ(om , c1 ) = 4,
and the agent has a limited budget of bκ = 8. It can, therefore acquire no more than two of
the three resources, which means that the optimal solution to the unconstrained problem as
in Example 1 is no longer feasible.
(cid:3)
Let us observe that the MDP-based model of agents’ preferences presented above is
fully general for discrete indivisible resources, i.e., any non-decreasing utility function over
resource bundles can be represented via the resource-constrained MDP model described
above.
5. This formulation assumes a stationary policy, which is supported by the argument in Section 3.1.
514
Resource Allocation Among Agents with MDP-Induced Preferences
Theorem 1 Consider a finite set of n indivisible resources O = {oi } (i ∈ [1, n]), with
m ∈ N available units of each resource. Then, for any non-decreasing utility function
defined over resource bund les f : [0, m]n 7→ R, there exists a resource-constrained MDP
hS , A, p, r, O , ρ, C , κ, bκ , αi (with the same resource set O) whose induced utility function over
the resource bund les is the same as f . In other words, for every resource bund le z ∈ [0, m]n ,
the value of the optimal policy among those whose resource requirements do not exceed z
(cal l this set Π(z)) is the same as f (z):
∀f : [0, m]n 7→ R, ∃ hS , A, p, r, O , ρ, C , κ, bκ , αi :
∀ z ∈ [0, m]n , Π(z) = nπ (cid:12)(cid:12)(cid:12) max
π(s, a)(cid:1)i ≤ zio =⇒ max
a hρ(a, oi )H (cid:0) Xs
π∈Π(z)
Proof: See Appendix A.1.
(cid:3)
Let us comment that while Theorem 1 establishes the generality of the MDP-based pref-
erence model introduced in this section, the construction used in the proof is of little practi-
cal interest, as it requires an MDP with an exponentially large state or action space. Indeed,
we do not advocate mapping arbitrary unstructured utility functions to exponentially-large
MDPs as a general solution technique. Rather, our contention is that our techniques ap-
ply to domains where utility functions are induced by a stochastic decision-making process
(modeled as an MDP), thus resulting in well-structured preferences over resources that
can be exploited to drastically lower the computational complexity of resource-allocation
algorithms.
Uγ (π , α) = f (z).
3.1 Properties of the Single-Agent Constrained MDP
In this section, we analyze the constrained policy-optimization problem (9). Namely, we
show that stationary deterministic policies are optimal for this problem, meaning that it
is not necessary to consider randomized, or history-dependent policies. However, solutions
to problem (9) are not, in general, uniformly optimal (optimal for any initial distribution).
Furthermore, we show that (9) is NP-hard, unlike the unconstrained MDPs, which can be
solved in polynomial time (Littman, Dean, & Kaelbling, 1995).
In the
We begin by showing optimality of stationary deterministic policies for (9).
following, we use ΠHR to refer to the class of history-dependent randomized policies (the
most general policies), and ΠSD ⊂ ΠHR to refer to the class of stationary deterministic
policies.
Theorem 2 Given an MDP M = hS , A, p, r, O , ρ, C , κ, bκ , αi with resource and capacity
constraints, if there exists a policy π ∈ ΠHR that is a feasible solution for M , there exists a
stationary deterministic policy πSD ∈ ΠSD that is also feasible, and the expected total reward
of πSD is no less than that of π :
∀ π ∈ ΠHR , ∃ πSD ∈ ΠSD : Uγ (πSD , α) ≥ Uγ (π , α)
Proof: See Appendix A.2.
(cid:3)
intuitively, stationary deterministic
The result of Theorem 2 is not at all surprising:
policies are optimal, because history dependence does not increase the utility of the policy,
515
Dolgov & Durfee
and using randomization can only increase resource costs. The latter is true because in-
cluding an action in a policy incurs the same costs in terms of resources regardless of the
probability of executing that action (or the expected number of times the action will be
executed). This is true because we are dealing with non-consumable resources; this prop-
erty does not hold for MDPs with consumable resources (as we discuss in more detail in
Section 6).
We now show that uniformly optimal policies do not always exist for our constrained
problem. This result is well known for another class of constrained MDPs, where constraints
are imposed on the total expected costs that are proportional to the expected number of
times the corresponding actions are executed (discussed earlier in Section 2). MDPs with
such constraints arise, for example, when bounds are imposed on the expected usage of
consumable resources, and as mentioned in Section 2, these problems can be solved using
linear programming by augmenting the dual LP (6) with linear constraints on expected
costs (8). Below, we establish the same result for problems with non-consumable resources
and capacity constraints.
Observation 1 There do not always exist uniformly optimal solutions to (9).
In other
words, there exist two constrained MDPs that differ only in their initial conditions: M =
hS , A, p, r, O , ρ, C , κ, bκ , αi and M ′ = hS , A, p, r, O , ρ, C , κ, bκ , α′ i, such that there is no policy
that is optimal for both problems simultaneously, i.e., for any two policies π and π ′ that are
optimal solutions to M and M ′ , respectively, the fol lowing holds:
Uγ (π , α) > Uγ (π ′ , α), Uγ (π , α′ ) < Uγ (π ′ , α′ )
(10)
We demonstrate this observation by example.
Example 3 Consider the resource-constrained problem as in Example 2. It is easy to see
that if the initial conditions are α = [1, 0, 0] (the agent starts in state s1 with certainty),
the optimal policy for states s1 and s2 is the same as in Example 1 (s1 → a2 and s2 → a3 ),
which, given the initial conditions, results in zero probability of reaching state s3 (to which
the noop a0 is assigned). This policy requires the truck and the forklift. However, if the
agent starts in state s3 (α = [0, 0, 1]), the optimal policy is to fix the truck (execute a4 in
s3 ), and to resort to furniture delivery (do a1 in s1 and assign the noop ao to s2 , which is
then never visited). This policy requires the mechanic and the truck. These two policies are
uniquely optimal for the corresponding initial conditions, and are suboptimal for other initial
conditions, which demonstrates that no uniformly optimal policy exists for this example. (cid:3)
The intuition behind the fact that uniformly optimal policies do not, in general, exist
for constrained MDPs is that since the resource information is not a part of the MDP state
space, and there are constraints imposed on resource usage, the principle of Bellman opti-
mality does not hold (optimal actions for different states cannot be chosen independently).
Given a constrained MDP, it is possible to construct an equivalent unconstrained MDP with
the standard properties of optimal solutions (by folding the resource information into the
state space, and modeling resource constraints via the transition function), but the resulting
state space will be exponential in the number of resources.
We now analyze the computational complexity of the optimization problem (9).
516
Resource Allocation Among Agents with MDP-Induced Preferences
Theorem 3 The fol lowing decision problem is NP-complete. Given an instance of an MDP
hS , A, p, r, O , ρ, C , κ, bκ , αi with resources and capacity constraints, and a rational number Y ,
does there exist a feasible policy π , whose expected total reward, given α, is no less than Y ?
Proof: See Appendix A.3.
(cid:3)
Note that the above complexity result stems from the limited capacities of the agents
and the fact that we define the resource requirements of a policy as the set of all resources
needed to carry out all actions that have a nonzero probability of being executed.
If,
however, we defined constraints on the expected resource requirements, then actions with
low probability of being executed would have lower resource requirements, optimal policies
would be randomized, and the problem would be equivalent to a knapsack with continuously
divisible items, which is solvable in polynomial time via the LP formulation of MDPs with
linear constraints (6,8).
3.2 MILP Solution
Now that we have analyzed the properties of the optimization problem (9), we present a
formulation of (9) as a mixed integer linear program (MILP). Given that we have established
NP-completeness of (9) in the previous section, MILP (also NP-complete) is a reasonable
formulation that allows us to reap the benefits of a vast selection of efficient algorithms and
tools (see, for example, the text by Wolsey, 1998 and references therein).
In this section and in the rest of the paper we will assume that the resource requirements
of actions are binary, i.e., ρ(a, o) = {0, 1}. We make this assumption to simplify the
discussion, and it does not limit the generality of our results. We briefly describe the case
of non-binary resource costs in Appendix B for completeness, but refer to the work by
Dolgov (2006) for a more detailed discussion and examples.
Let us rewrite (9) in the occupation measure coordinates x by adding the constraints
from (9) to the standard LP in occupancy coordinates (6). Noticing that (for states with
nonzero probability of being visited) π(s, a) and x(s, a) are either zero or nonzero simulta-
neously:
π(s, a)(cid:1) = H (cid:0) Xs
H (cid:0) Xs
x(s, a)(cid:1),
and that, when ρ(a, o) = {0, 1}, the total resource requirements of a policy can be simplified
as follows:
x(s, a)(cid:17),
x(s, a)(cid:1)o = H (cid:16) Xa
a nρ(a, o)H (cid:0) Xs
ρ(a, o) Xs
max
we get the following program in x:
x Xs Xa
x(s, a)r(s, a)
max
sub ject to:
Xa
x(σ, a) − γ Xs Xa
x(s, a)p(σ s, a) = α(σ),
x(s, a)(cid:17) ≤ bκ(c),
κ(o, c)H (cid:16) Xa
Xo
ρ(a, o) Xs
x(s, a) ≥ 0,
517
∀s ∈ S , a ∈ A.
∀a ∈ A,
∀o ∈ O ,
(11)
∀σ ∈ S ;
∀c ∈ C ;
(12)
Dolgov & Durfee
∀c ∈ C ,
The challenge in solving this mathematical program is that the constraints are nonlinear
due to the Heaviside function H .
To linearize the Heaviside function, we augment the original optimization variables x
with a set of O binary variables δ(o) ∈ {0, 1}, where δ(o) = H (cid:16) Pa ρ(a, o) Ps x(s, a)(cid:17). In
other words, δ(o) is an indicator variable that shows whether the policy requires resource o.
Using δ(o), we can rewrite the resource constraints in (12) as:
Xo
κ(o, c)δ(o) ≤ bκ(c),
which are linear in δ . We can then synchronize δ and x via the following linear inequalities:
1/X Xa
ρ(a, o) Xs
x(s, a) ≤ δ(o),
∀o ∈ O ,
(14)
where X ≥ maxo Pa ρ(a, o) Ps x(s, a) is the normalization constant, for which any up-
per bound on the argument of H () can be used. The bound X is guaranteed to exist
for discounted problems. For example, we can use X = (1 − γ )−1 maxo Pa ρ(a, o), since
Ps,a x(s, a) = (1 − γ )−1 for any x that is a valid occupation measure for an MDP with
discount factor γ .6
Putting it all together, the problem (9) of finding optimal policies under resource con-
straints can be formulated as the following MILP:
x,δ Xs Xa
max
sub ject to:
x(σ, a) − γ Xs Xa
Xa
Xo
κ(o, c)δ(o) ≤ bκ(c),
1/X Xa
ρ(a, o) Xs
x(s, a) ≤ δ(o),
x(s, a) ≥ 0,
δ(o) ∈ {0, 1},
x(s, a)p(σ s, a) = α(σ),
∀s ∈ S , a ∈ A;
x(s, a)r(s, a)
∀σ ∈ S ;
∀c ∈ C ;
∀o ∈ O ;
∀o ∈ O .
(13)
(15)
We illustrate the MILP construction with an example.
Example 4 Let us formulate the MILP for the constrained problem from Example 3. Re-
cal l that in that problem there are three resources O = {ot , of , om } (truck, forklift, and
mechanic), one capacity type C = {c1 } (money), and actions have the fol lowing resource
requirements (again, listing only the nonzero ones):
ρ(a1 , ot ) = 1, ρ(a2 , ot ) = 1, ρ(a2 , of ) = 1, ρ(a3 , ot ) = 1, ρ(a4 , ot ) = 1, ρ(a4 , om ) = 1
6. Instead of using a single X for all resources, a different X (o) ≥ Pa ρ(a, o) Ps x(s, a) can be used for
every resource, leading to more uniform normalization and potentially better numerical stability of the
MILP solver.
518
Resource Allocation Among Agents with MDP-Induced Preferences
The resources have the fol lowing capacity costs:
κ(ot , c1 ) = 2,
κ(of , c1 ) = 3,
κ(om , c1 ) = 4,
and the agent has a limited budget, i.e., a capacity bound, bκ(c1 ) = 8.
To compute the optimal policy for an arbitrary α, we can formulate the problem as
an MILP as described above. Using binary variables δ(o) = {δ(ot ), δ(of ), δ(om )}, we can
express the constraint on capacity cost as the fol lowing inequality:
2δ(ot ) + 3δ(of ) + 4δ(om ) ≤ 8,
For the constraints that synchronize the occupation measure x and the binary indicators δ(o),
we can set X = (1 − γ )−1 maxo Pa ρ(a, o) = 4(1 − γ )−1 . Combining this with other
constraints from (15), we get an MILP with 12 continuous and 4 binary variables, and
S + C + O = 3 + 3 + 1 = 7 constraints (not counting the last two sets of range con-
straints).
(cid:3)
As mentioned earlier, even though solving such programs is, in general, an NP-complete
task, there is a wide variety of very efficient algorithms and tools for doing so. Therefore,
one of the benefits of formulating the optimization problem (9) as an MILP is that it allows
us to make use of the highly efficient existing tools.
4. Multiagent Resource Allocation
We now consider the multiagent problem of resource allocation between several agents,
where the resource preferences of the agents are defined by the constrained MDP model
described in the previous section. We reiterate our main assumptions about the problem:
1. Weak coupling. We assume that agents are weakly-coupled (Meuleau et al., 1998),
i.e., they only interact through the shared resources, and once the resources are allo-
cated, the agents’ transitions and rewards are independent. This assumption is critical
to our results.7
2. One-shot resource allocation. The resources are distributed once before the agents
start executing their MDPs. There is no reallocation of resources during the MDP
phase. This assumption is critical to our results; allowing reallocation of resources
would violate the weak-coupling assumption.
3. Initial central control over the resources. We assume that at the beginning of
the resource-allocation phase, the resources are controlled by a single authority. This
is the standard sell-auction setting. For problems where the resources are distributed
7. If agents are cooperative, the assumption about weak coupling can be relaxed (at the expense of an
increase in complexity), and the same MILP-based algorithm for simultaneously performing policy op-
timization and resource allocation can be applied if we consider the joint state spaces of the interacting
agents. For self-interested agents, a violation of the weakly-coupled assumption would mean that the
agents would be playing a Markov game (Shapley, 1953) once the resources are allocated, which would
significantly complicate the strategic analysis of the agents’ bidding strategies during the initial resource
allocation.
519
Dolgov & Durfee
among the agents to begin with, we face the problem of designing a computationally-
efficient combinatorial exchange (Parkes, Kalagnanam, & Eso, 2001), which is a more
complicated problem that is outside the scope of this work. However, many of the
ideas presented in this paper could potentially be applicable to that domain as well.
4. Binary resource costs. As before, we assume that agents’ resource costs are binary.
This assumption is not limiting. The case of non-binary resources is discussed in
Appendix B.
Formally, the input to the resource-allocation problem consists of the following:
• M is the set of agents; we will use m ∈ M to refer to an agent.
• {hS , A, pm , rm , αm , ρm , bκm i} is the collection of weakly-coupled single-agent MDPs,
as defined in the single-agent model in Section 3. For simplicity, but without loss of
generality, we assume that all agents have the same state and action spaces S and A,
but each has its own transition and reward functions pm and rm , initial conditions αm ,
as well as its own resource requirements ρm : A × O 7→ {0, 1} and capacity bounds
bκm : C 7→ R. We also assume that all agents have the same discount factor γ , but this
assumption can be trivially relaxed.
• O and C are the sets of resources and capacities, defined exactly as in the single-agent
model in Section 3.
• κ : O × C 7→ R specifies the capacity costs of the resources, defined exactly as in the
single-agent model in Section 3.
• bρ : O 7→ R specifies the upper bound on the amounts of the shared resources (this
defines the additional bound for the multiagent problem).
Given the above, our goal is to design a mechanism for allocating the resources to the
agents in an economically efficient way, i.e., in a way that maximizes the social welfare of
the agents (one of the most often-used criteria in mechanism design). We would also like
the mechanism to be efficient from the computational standpoint.
Example 5 Suppose that there are two delivery agents. The MDP and capacity constraints
of the first agent are exactly as they were defined previously in Examples 1 and 2. The MDP
of the second agent is almost the same as that of the first agent, with the only difference
that it gets a slightly higher reward for delivering appliances: r2 (s1 , a2 ) = 12 (whereas
r1 (s1 , a2 ) = 10 for the first agent). Suppose there are two trucks, one forklift, and one
mechanic that are shared by the two agents. These bounds are specified as fol lows:
bρ(om ) = 1.
bρ(of ) = 1,
bρ(ot ) = 2,
Then the problem is to decide which agent should get the forklift, and which should get the
mechanic (trucks are plentiful in this example).
(cid:3)
520
Resource Allocation Among Agents with MDP-Induced Preferences
4.1 Combinatorial Auctions
As previously mentioned, the problem of finding an optimal resource allocation among
self-interested agents that have complex valuations over combinations of resources arises
in many different domains (e.g., Ferguson, Nikolaou, Sairamesh, & Yemini, 1996; Wellman
et al., 2001) and is often called a combinatorial al location problem. A natural and widely
used mechanism for solving such problems is a combinatorial auction (CA) (e.g., de Vries
& Vohra, 2003).
In a CA, each agent submits a set of bids for resource bundles to the
auctioneer, who then decides what resources each agent will get and at what price.
Consider the problem of allocating among a set of agents M a set of indivisible re-
sources O , where the total quantity of resource o ∈ O is bounded by bρ(o). Our earlier
simplifying assumption that actions’ resource requirements are binary implies that agents
will only be interested in bundles that contain no more than one unit of a particular resource.
In a combinatorial auction, each agent m ∈ M submits a bid bm
w (specifying how much
the agent is willing to pay) for every bundle w ∈ W m that has some value um
w > 0. In some
cases, it is possible to express such bids without enumerating all bundles (for example, using
an XOR bidding language (Sandholm, 1999) where it is necessary to only consider bundles
with strictly positive value, such that no subset of a bundle has the same value). Such
techniques often reduce the complexity of the resource-allocation problem, but do not, in
general, avoid the exponential blow up in the number of bids. Therefore, below we describe
the simplest combinatorial auction with flat bids, but it should be noted that many concise
bidding languages exist that in special cases can reduce the number of explicit bids.
After collecting all the bids, the auctioneer solves the winner-determination problem
(WDP), a solution to which prescribes how the resources should be distributed among the
w − qm
w , its utility is um
agents and at what prices. If agent m wins bundle w at price qm
w
(we are assuming risk-neutral agents with quasi-linear utility functions). Thus, the optimal
bidding strategy of an agent depends on how the auctioneer allocates the resources and sets
prices.
Vickrey-Clarke-Groves (VCG) mechanisms (Vickrey, 1961; Clarke, 1971; Groves, 1973)
are a widely used family of mechanisms that have certain very attractive properties (dis-
cussed in more detail below). An instantiation of a VCG mechanism in the context of
combinatorial auctions is the Generalized Vickrey Auction (GVA) (MacKie-Mason & Var-
ian, 1994), which allocates resources and sets prices as follows. Given the bids bm
w of all
agents, the auctioneer chooses an allocation that maximizes the sum of agents’ bids. This
problem is NP-complete (Rothkopf et al., 1998) and can be expressed as the following in-
teger program, where the optimization variables zm
w = {0, 1} are indicator variables that
show whether bundle w is assigned to agent m, and nwo = {0, 1} specifies whether bundle w
contains o:8
8. There are other related algorithms for solving the WDP (e.g., Sandholm, 2002), but we will use the
integer program (16) as a representative formulation for the class of algorithms that perform a search in
the space of binary decisions on resource bundles.
521
Dolgov & Durfee
w bm
zm
w
(16)
∀o ∈ O .
∀m ∈ M;
z Xm∈M Xw∈W m
max
sub ject to:
Xw∈W m
zm
w ≤ 1,
Xm∈M Xw∈W m
zm
w nwo ≤ bρ(o),
The first constraint in (16) says that no agent can receive more than one bundle, while the
second constraint ensures that the total amount of resource o assigned to the agents does
not exceed the total amount available. Notice that MILP (16) performs the summation over
exponentially large sets of bundles w ∈ W m . As outlined above, in an auction with XOR
bidding, these sets would typically be smaller, but, in general, still exponentially large.
A GVA assigns resources according to the optimal solution ez to (16) and sets the payment
for agent m to:
−m − Xm′ 6=m ezm′
w bm′
w = V ∗
qm
w ,
(17)
where V ∗
−m is the value of (16) if m were to not participate in the auction (the optimal
value if m does not submit any bids), and the second term is the sum of other agents’ bids
in the solution ez to the WDP with m participating.
A GVA has a number of nice properties. It is strategy-proof, meaning that the dominant
strategy of every agent is to bid its true value for every bundle: bm
w = um
w . The auction
is economical ly efficient, meaning that it allocates the resources to maximize the social
welfare of the agents (because, when agents bid their true values, the ob jective function
of (16) becomes the social welfare). Finally, a GVA satisfies the participation constraint,
meaning that no agent decreases its utility by participating in the auction.
A straightforward way to implement a GVA for our MDP-based problem is the following.
Let each agent m ∈ M enumerate all resource bundles W m that satisfy its local capacity
constraints defined by bκm (c) (this is sufficient because our MDP model implies free disposal
of resources for the agents, and we make the same assumption about the auctioneer). For
each bundle w ∈ W m , agent m would determine the feasible action set A(w) and formulate
an MDP Λm (w) = hS , A(w), pm (w), rm (w), αm i, where pm (w) and rm (w) are the transition
and reward functions defined on the pruned action space A(w). Every agent would then
solve each Λm (w) corresponding to a feasible bundle to find the optimal policy eπm (w), whose
γ (eπm (w), αm ).
expected discounted reward would define the value of bundle w: um
w = U m
This mechanism suffers from two ma jor complexity problems. First, the agents have
to enumerate an exponential number of resource bundles and compute the value of each
by solving the corresponding (possibly large) MDP. Second, the auctioneer has to solve
an NP-complete winner-determination problem on exponentially large input. The following
sections are devoted to tackling these complexity problems.
Example 6 Consider the two-agent problem described in Example 5, where two trucks, one
forklift, and the services of one mechanic are being auctioned off. Using the straightforward
version of the combinatorial auction outlined above, each agent would have to consider
522
Resource Allocation Among Agents with MDP-Induced Preferences
2O = 23 = 8 possible resource bund les (since resource requirements of both agents are
binary, neither agent is going to bid on a bund le that contains two trucks). For every
resource bund le, each agent wil l have to formulate and solve the corresponding MDP to
compute the utility of the bund le.
For example, if we assume that both agents start in state s1 (different initial conditions
would result in different expected rewards, and thus different utility functions), the value of
the nul l resource bund le to both agents would be 0 (since the only action they would be able
to execute is the noop a0 ). On the other hand, the value of bund le [ot , of , om ] = [1, 1, 1] that
contains al l the resources would be 95.3 to the first agent and 112.4 to the second one. The
value of bund le [1, 1, 0] to each agent would be the same as the value of [1, 1, 1] (since their
optimal policies for the initial conditions that put them in s1 do not require the mechanic).
Once the agents submit their bids to the auctioneer, it wil l have to solve the WDP via
the integer program (16) with M2O = 2(2)3 = 16 binary variables. Given the above, the
optimal way to al locate the resources would be to assign a truck to each of the agents, the
forklift to the second agent, and the mechanic to either (or neither) of the two. Thus, the
agents would receive bund les [1, 0, 0] and [1, 1, 0], respectively, resulting in social welfare of
50 + 112.4 = 162.4. However, if at least one of the agents had a non-zero probability of
starting in state s3 , the value of the resource bund les involving the mechanic would change
drastical ly, as would the optimal resource al location and its social value.
(cid:3)
4.2 Avoiding Bundle Enumeration
To avoid enumerating all resource bundles that have non-zero value to an agent, two things
are required:
i) the mechanism has to support a concise bidding language that allows
the agent to express its preferences to the auctioneer in a compact manner, and ii) the
agents have to be able to find a good representation of their preferences in that language. A
simple way to achieve both in our model is to create an auction where the agents submit the
specifications of their resource-parameterized MDPs to the auctioneer as bids: the language
is compact and, given our assumption that each agent can formulate its planning problem
as an MDP, this does not require additional computation for the agents. However, this
only changes the communication protocol between the agents and the auctioneer, similarly
to other concise bidding languages (Sandholm, 1999; Nisan, 2000; Boutilier & Hoos, 2001;
Boutilier, 2002). As such, it simply moves the burden of solving the valuation problem from
the agents to the auctioneer, which by itself does not lead to any gains in computational
efficiency. Such a mechanism also has implications on information privacy issues, because
the agents have to reveal their local MDPs to the auctioneer (which they might not want
to do). Nevertheless, we can build on this idea to increase the efficiency of solving both the
valuation and winner-determination problems and do so while keeping most of the agents’
MDP information private. We address ways of maintaining information privacy in the next
section, and for the moment focus on improving the computational complexity of the agents’
valuation and the auctioneer’s winner-determination problems.
The question we pose in this section is as follows. Given that the bid of each agent con-
sists of its MDP, its resource information and its capacity bounds hS , A, pm , rm , αm , ρm , bκm i,
can the auctioneer formulate and solve the winner-determination problem more efficiently
523
Dolgov & Durfee
∀o ∈ O .
∀c ∈ C , m ∈ M;
(18)
than by simply enumerating each agent’s resource bundles and solving the standard integer
program (16) with an exponential number of binary variables?
Therefore, the goal of the auctioneer is to find a joint policy (a collection of single-agent
policies under our weak-coupling assumption) that maximizes the sum of the expected total
discounted rewards for all agents, under the conditions that: i) no agent m is assigned a set
of resources that violates its capacity bound bκm (i.e., no agent is assigned more resources
than it can carry), and ii) the total amounts of resources assigned to all agents do not exceed
the global resource bounds bρ(o) (i.e., we cannot allocate to the agents more resources than
are available). This problem can be expressed as the following mathematical program:
π Xm
U m
γ (πm , αm )
max
sub ject to:
κ(o, c)H (cid:0)ρm (a, o) Xs
Xo
πm (s, a)(cid:1) ≤ bκm (c),
ρm (a, o)H (cid:0) Xs
Xm
πm (s, a)(cid:1) ≤ bρ(o),
Obviously, the decision version of this problem is NP-complete, as it subsumes the single-
agent MDP with capacity constraints, NP-completeness of which was shown in Section 3.1.
Moreover, the problem remains NP-complete even in the absence of single-agent capacity
constraints. Indeed, the global constraint on the amounts of the shared resources is sufficient
to make the problem NP-complete, which can be shown with a straightforward reduction
from KNAPSACK, similar to the one used in the single-agent case in Section 3.1.
We can linearize (18) similarly to the single-agent problem from Section 3.2, yielding
the following MILP, which is simply a multiagent version of (15) (recall the assumption of
this section that resource requirements are binary):
x,δ Xm Xs Xa
max
sub ject to:
xm (σ, a) − γ Xs Xa
Xa
xm (s, a)pm (σ s, a) = αm (σ),
Xo
κ(o, c)δm (o) ≤ bκm (c),
Xm
δm (o) ≤ bρ(o),
1/X Xa
ρm (a, o) Xs
xm (s, a) ≤ δm (o),
xm (s, a) ≥ 0,
δm (o) ∈ {0, 1},
∀o ∈ O , m ∈ M,
where X ≥ maxo,m Pa ρ(a, o) Ps xm (s, a) is an upper bound on the argument of H (·),
used for normalization. As in the single-agent case, this bound is guaranteed to exist for
discounted MDPs and is easy to obtain.
∀s ∈ S , a ∈ A, m ∈ M;
xm (s, a)rm (s, a)
(19)
∀o ∈ O ;
∀σ ∈ S , m ∈ M;
∀c ∈ C , m ∈ M;
∀o ∈ O , m ∈ M;
524
Resource Allocation Among Agents with MDP-Induced Preferences
The MILP (19) allows the auctioneer to solve the WDP without having to enumerate
the possible resource bundles. As compared to the standard WDP formulation (16), which
has on the order of M2O binary variables, (19) has only MO binary variables. This
exponential reduction is attained by exploiting the knowledge of the agents’ MDP-based
valuations and simultaneously solving the policy-optimization and resource-allocation prob-
lems. Given that the worst-case solution time for MILPs is exponential in the number of
integer variables, this reduction has a significant impact on the worst-case performance of
the mechanism. The average-case running time is also reduced drastically, as demonstrated
by our experiments, presented in Section 5.
Example 7 If we apply the mechanism discussed above to our running example as an al-
ternative to the straightforward combinatorial auction presented in Example 6, the winner-
determination MILP (19) wil l look as fol lows. It wil l have MS A = (2)(3)(5) = 30 con-
tinuous occupation-measure variables xm , and MO = (2)(3) = 6 binary variables δm (o).
It wil l have MS = (2)(3) = 6 conservation-of-flow constraints that involve continuous
variables only, as wel l as MC + O + MO = (2)(1) + 3 + (2)(3) = 9 constraints that
involve binary variables.
The capacity constraints for the agents wil l be exactly as in the single-agent case described
in Example 4, and the global resource constraints wil l be:
δ1 (ot ) + δ2 (ot ) ≤ 2,
δ1 (of ) + δ2 (of ) ≤ 1,
δ1 (om ) + δ2 (om ) ≤ 1.
Notice that in this example there is one binary decision variable per resource per agent
(yielding 6 such variables for this simple problem). This is exponential ly fewer than the
number of binary variables in the straightforward CA formulation of Example 6, which
requires one binary variable per resource bund le per agent (yielding 16 such variables for
this problem). Given that MILPs are NP-complete in the number of integer variables, this
reduction from 16 to 6 variables is noticeable even in a smal l problem like this one and can
lead to drastic speedup for larger domains.
(cid:3)
The mechanism described above is an instantiation of the GVA, so by the well-known
properties of VCG mechanisms, this auction is strategy-proof (the agents have no incentive
to lie to the auctioneer about their MDPs), it attains the socially optimal resource allocation,
and no agent decreases its utility by participating in the auction.
To sum up the results of this section: by having the agents submit their MDP informa-
tion to the auctioneer instead of their valuations over resource bundles, we have essentially
removed all computational burden from the agents and at the same time significantly sim-
plified the auctioneer’s winner-determination problem (the number of integer variables in
the WDP is reduced exponentially).
4.3 Distributing the Winner-Determination Problem
Unlike the straightforward combinatorial auction implementation discussed earlier in Sec-
tion 4.1, where the agents shared some computational burden with the auctioneer, in the
mechanism from Section 4.2, the agents submit their information to the auctioneer and
then just idle while waiting for a solution. This suggests further potential improvements in
computational efficiency. Indeed, given the complexity of MILPs, it would be beneficial to
525
Dolgov & Durfee
exploit the computational power of the agents to offload some of the computation from the
auctioneer back to the agents (we assume that agents have no cost for “helping out” and
would prefer for the outcome to be computed faster).9 Thus, we would like to distribute
the computation of the winner-determination problem (19), the common ob jective in dis-
tributed algorithmic mechanism design (Feigenbaum & Shenker, 2002; Parkes & Shneidman,
2004).10
For concreteness, we will base the algorithm of this section on the branch and bound
method for solving MILPs (Wolsey, 1998), but exactly the same techniques will also work
for other MILP algorithms (e.g., cutting planes) that perform a search in the space of LP
relaxations of the MILP. In branch and bound for MILPs with binary variables, LP relax-
ations are created by choosing a binary variable and setting it to either 0 or 1, and relaxing
the integrality constraints of other binary variables. If a solution to an LP relaxation hap-
pens to be integer-valued, it provides a lower bound on the value of the global solution. A
non-integer solution provides an upper bound for the current subproblem, which (combined
with other lower bounds) is used to prune the search space.
Thus, a simple way for the auctioneer to distribute the branch and bound algorithm is
to simply farm out LP relaxations to other agents and ask them to solve the LPs. However,
it is easy to see that this mechanism is not strategy-proof. Indeed, an agent that is tasked
with performing some computation for determining the optimal resource allocation or the
associated payments could benefit from lying about the outcome of its computation to
the auctioneer. This is a common phenomenon in distributed mechanism implementations:
whenever some WDP calculations are offloaded to an agent participating in the auction, the
agent might be able to benefit from sabotaging the computation. There are several methods
to ensuring the strategy-proofness of a distributed implementation. The approach best
suited for our problem is based on the idea of redundant computation (Parkes & Shneidman,
2004),11 where multiple agents are asked to do the same task and any disagreement is
carefully punished to discourage lying. In the rest of this section, we demonstrate that this
is very easy to implement in our case.
The basic idea is very simple: let the auctioneer distribute LP relaxations to the agents,
but check their solutions and re-solve the problems if the agents return incorrect solutions
(this would make truthful computation a weakly-dominant strategy for the agents, and
a nonzero punishment can be used to achieve strong dominance). This strategy of the
auctioneer removes the incentive for the agents to lie and yields exactly the same solution
as the centralized algorithm. However, in order for this to be beneficial, the complexity of
checking a solution must be significantly lower than the complexity of solving the problem.
Fortunately, this is true for LPs.
Suppose the auctioneer has to solve the following LP, which can be written in two
equivalent ways (let us refer to the one on the left as the primal, and to the one on the right
9. As observed by Parkes and Shneidman (2004), this assumption is a bit controversial, since a desire for
efficient computation implies nonzero cost for computation, while the agents’ cost for “helping out” is
not modeled. It is, nonetheless, a common assumption in distributed mechanism implementations.
10. We describe one simple way of distributing the mechanism, others are also possible.
11. Redundant computation is discussed by Parkes and Shneidman (2004) in the context of ex post Nash
equilibria, whereas we are interested in dominant strategies, but the high-level idea is very similar.
526
Resource Allocation Among Agents with MDP-Induced Preferences
as the dual):
min α
T v
sub ject to:
AT v ≥ r.
max rT x
sub ject to:
Ax = α; x ≥ 0.
(20)
By the strong duality property, if the primal LP has a solution v∗ , then the dual also has a
solution x∗ , and αT v∗ = rT x∗ . Furthermore, given a solution to the primal LP, it is easy to
compute a solution to the dual: by complementary slackness, v∗T = rT B−1 and x∗ = B−1α,
where B is a square invertible matrix composed of columns of A that correspond to basic
variables of the solution.
These well-known properties can be used by the auctioneer to quickly check optimality
of solutions returned by the agents. Suppose that an agent returns v as a solution to the
primal LP. The auctioneer can calculate the dual solution vT = rT B−1 and check whether
rT x = αT v. Thus, the most expensive operation that the auctioneer has to perform is
the inversion of B , which can be done in sub-cubic time. As a matter of fact, from the
implementation perspective, it would be more efficient to ask the agents to return both the
primal and the dual solutions, since many popular algorithms compute both in the process
of solving LPs.
Thus, we have provided a simple method that allows us to effectively distribute the
winner-determination problem, while maintaining strategy-proofness of the mechanism and
with a negligible computation overhead for the auctioneer.
4.4 Preserving Information Privacy
The mechanism that we have discussed so far has the drawback that it requires agents
to reveal complete information about their MDPs to the auctioneer. The problem is also
exacerbated in the distributed WDP algorithm from the previous section, since not only
does each agent reveal its MDP information to the auctioneer, but that information is then
also spread to other agents via the LP relaxations of the global MILP. We now show how
to alleviate this problem.
Let us note that, in saying that agents prefer not to reveal their local information, we
are implicitly assuming that there is an external factor that affects agents’ utilities that is
not captured in the agents’ MDPs. A sensible way to measure the value of information is by
how it changes one’s decision-making process and its outcomes. Since this effect is not part
of our model (in fact, it contradicts our weak-coupling assumption), we cannot in a domain-
independent manner define what constitutes “useful” information, and how bad it is for an
agent to reveal too much about its MDP. Modeling such effects and carefully analyzing
them is an interesting research task, but it is outside the scope of this paper. Thus, for the
purposes of this section, we will be content with a mechanism that hides enough information
to make it impossible for the auctioneer or an agent to uniquely determine the transition
or reward function of any other agent (in fact, information revealed to any agent will map
to infinitely many MDPs of other agents).12 Many such transformations are possible; we
present just one to illustrate the concept.
12. A more stringent condition would require that agents’ preferences over resource bundles are not revealed
(Parkes, 2001), but we set a lower bar here.
527
Dolgov & Durfee
The main idea of our approach is to modify the previous mechanism so that the agents
submit their private information to the auctioneer in an “encrypted” form that allows the
auctioneer to solve the winner-determination problem, but does not allow it to infer the
agents’ original MDPs.
First, note that, instead of passing an MDP to the auctioneer, each agent can submit
an equivalent LP (6). So, the question becomes: can the agent transform its LP in such a
way that the auctioneer will be able to solve it, but will not be able to infer the transition
and reward functions of the originating MDP? In other words, the problem reduces to the
following. Given an LP L1 (created from an MDP Λ = hS , A, p, r, αi via (6)), we need to
find a transformation L1 → L2 such that a solution to the transformed LP L2 will uniquely
map to a solution to the original LP L1 , but L2 will not reveal the transition or the reward
functions of the original MDP (p or r). We show that a simple change of variables suffices.
Suppose agent m1 has an MDP-originated LP and is going to ask agent m2 to solve it. In
order to maintain the linearity of the problem (to keep it simple for m2 to solve), m1 should
limit itself to linear transformations. Consider a linear, invertible transformation of the
primal coordinates u = F v, and a linear, invertible transformation of the dual coordinates
y = Dx. Then, the LP from (20) will be transformed (by applying F , switching to the
dual, and then applying D) to an equivalent LP in the new coordinates y:
max rT D−1y
sub ject to:
(F −1 )T AD−1y = (F −1 )T
D−1y ≥ 0.
α;
(21)
The value of the optimal solution to (21) will be the same as the value of the optimal solution
to (20), and given an optimal solution y∗ to (21), it is easy to compute the solution to the
original: x∗ = D−1y∗ . Indeed, from the perspective of the dual, the primal transformation F
is equivalent to a linear transformation of the dual equality constraints Ax = α, which (given
that F is non-singular) has no effect on the solution or the ob jective function. Furthermore,
the dual transformation D is equivalent to a change of variables that modifies the solution
but not the value of the ob jective function.
However, a problem with the above transformations is that it gives away D−1 . Indeed,
agent m2 will be able to simply read (up to a set of multiplicative constants) the trans-
formation off the constraints D−1y ≥ 0. Therefore, only diagonal matrices with positive
coefficients (which are equivalent to stretching the coordinate system) are not trivially de-
duced by m2 , since they also map to y ≥ 0. Choosing a negative multiplier for some xi
(inverting the axis) is pointless, because that flips the non-negativity constraints to yi ≤ 0,
immediately revealing the sign to m2 .
Let us demonstrate that, given any MDP Λ and the corresponding LP L1 , we can
choose D and F such that it will be impossible for m2 to determine the coefficients of L1
(or equivalently the original transition and reward functions p and r). When agent m2
receives L2 (as in (21)), all it knows is that L2 was created from an MDP, so the columns
of the constraint matrix of the original LP L1 must sum to a constant:
Aj i = 1 − γ Xσ
Xj
528
p(σ s, a) = 1 − γ .
(22)
Resource Allocation Among Agents with MDP-Induced Preferences
(a)
(b)
Figure 3: Preserving privacy example. Two different MDPs that can lead to the same LP
constraint matrix.
This gives m2 a system of S nonlinear equations for the diagonal D and arbitrary F ,
which have a total of S A + S 2 free parameters. For everything but the most degenerate
cases (which can be easily handled by an appropriate choice of D and F ), these equations
are hugely under-constrained and will have infinitely many solutions. As a matter of fact,
by sacrificing S of the free parameters, m1 can choose D and F in such a way that the
columns of constraints in L2 will also sum to a constant γ ′ ∈ (0, 1), which would have the
effect of transforming L1 to an L2 that corresponds to another valid MDP Λ2 . Therefore,
given an L2 , there are infinitely many original MDPs Λ and transformations D and F that
map to the same LP L2 .
We also have to consider the connection of resource and capacity costs to agents’ oc-
cupation measures in the global WDP (19). There are two things that the auctioneer has
to be able to do:
i) determine the value of each agent’s policy (to be able to maximize
the social welfare), and ii) determine the resource requirements of the policies (to check
the resource constraints). So, the question is, how does our transformation affect these?
As noted earlier, the transformation does not change the ob jective function, so the first
requirement holds. On the other hand, D does change the occupation measure xm (s, a) by
arbitrary multipliers. However, a multiplicative factor of xm (s, a) has no effect on the usage
of non-consumable resources, as it only matters whether the corresponding xm (s, a) is zero
or not (step function H nullifies the scaling effect). Thus, the second condition also holds.
Example 8 Consider the two-state MDP depicted in Figure 3a that represents the decision-
making problem of a sales company, with the two states corresponding to possible market
conditions, and the two actions — to two possible sales strategies. Market conditions in state
s1 are much more favorable than in state s2 (the rewards for both actions are higher). The
transitions between the two states correspond to probabilities of market conditions changing
and the rewards reflect the expected profitability in these two states. Obtaining such numbers
in a realistic scenario would require performing costly and time-consuming research, and the
company might not want to make this information public.
Therefore, if the company were to participate in the resource-al location mechanism de-
scribed above, it would want to encrypt its MDP before submitting it to the auctioneer.
529
Dolgov & Durfee
The MDP has the fol lowing reward function
r = (1, 19.622, 0.063, 0.084)T ,
(23)
(24)
(25)
and the fol lowing transition function:
p(a1 ) = (cid:18) 1
p(a2 ) = (cid:18)0.986 0.014
0.5 0.5(cid:19) ,
0.5 (cid:19) .
0
0.5
Using γ = 0.8, this corresponds to the fol lowing conservation of flow constraint matrix:
0.6 (cid:19) .
A = (cid:18)0.2
0.212 −0.4 −0.4
0.6
0 −0.012
Before submitting its LP to the auctioneer, the agent applies the fol lowing transformations:
D = diag(1, 0.102, 47.619, 47.619), F = (cid:18) 2
−0.084 0.126(cid:19) ,
0
yielding the fol lowing new constraint matrix:
A′ = (F −1 )T AD−1 = (cid:18)0.1
0 −0.9 0.1 0.1(cid:19) .
0
0
1
However, the above constraint matrix A′ corresponds to a non-transformed conservation
of flow constraint for a different MDP (shown in Figure 3b) with γ = 0.9, the fol lowing
reward function:
(26)
(27)
r = (1, 2, 3, 4)T ,
(28)
and the fol lowing transition function:
0 1(cid:19) .
0 1(cid:19) ,
p(a2 ) = (cid:18)0 1
p(a1 ) = (cid:18)1 0
Therefore, when the auctioneer receives the constraint matrix A′ , it has no way of know-
ing whether the agent has an MDP with transition function (29) that was transformed
using (26) or the MDP with transition function (24) that was not transformed. Notice that
the dynamics of the two MDPs vary significantly: both in transition probabilities and state
connectivity. The second MDP does not reveal any information about the originating MDP
and the corresponding market dynamics.
(cid:3)
(29)
To sum up, we can, to a large extent, maintain information privacy in our mechanism by
allowing agents to apply linear transformations to their original LPs. The information that
is revealed by our mechanism consists of agents’ resource costs ρm (a, o), capacity bounds
bκm (c), and the sizes of their state and action spaces (the latter can be hidden by adding
dummy states and actions to the MDP).
The revealed information can be used to infer agents’ preferences and resource require-
ments. Further, numeric policies are revealed, but the lack of information about transition
and reward functions renders this information worthless (as just illustrated in Example 8,
there could be multiple originating MDPs with very different properties).
530
Resource Allocation Among Agents with MDP-Induced Preferences
5. Experimental Results
In this section we present an empirical analysis of the computational complexity of the
resource-allocation mechanism described in Section 4. We report results on the compu-
tational complexity of the mechanism from Section 4.2, where the agents submit their
MDPs to the auctioneer, who then simultaneously solves the resource-allocation and policy-
optimization problems. As far as the additional speedup achieved by distributing the WDP,
as described in Section 4.3, we do not report empirical results, since it is well-established
in the parallel programming literature that parallel versions of branch-and-bound MILP
solvers consistently achieve linear speedup (Eckstein, Phillips, & Hart, 2000). This is due
to the fact that branch-and-bound algorithms require very little inter-process communica-
tion.
For our experiments, we implemented a multiagent delivery problem, based on the
multiagent rover domain (Dolgov & Durfee, 2004). In this problem, agents operate in a
stochastic grid world with delivery locations randomly placed throughout the grid. Each
delivery task requires a set of resources, and there are limited quantities of the resources.
There are random delivery locations on the grid, and each location has a set of deliveries that
it accepts. Each resource has some size requirements (capacity cost), and each delivery agent
has bounded space to hold the resources (limited capacity). The agents participate in an
auction where they bid on delivery resources. In this setting, the value of a resource depends
on what other resources the agent acquires and what other deliveries it can make. Given a
bundle of resources, an agent’s policy optimization problem is to find the optimal delivery
plan. The exact parameters used in our experiments are not critical for the trends seen in
the results presented below, but for the sake of reproducibility the domain is described in
detail in Appendix C.13 All of the resource costs in the experiments presented below are
binary.
Computational complexity of constrained optimization problems can vary greatly as
constraints are tightened or relaxed. Therefore, as our first step in the analysis of empirical
computational complexity of our mechanism, we investigate how its running time depends
on the capacity constraints of each agent and on the bounds on the total amounts of
resources shared by the agents. As is common with other types of constrained optimization
or constraint-satisfaction problems, it is natural to expect that the WDP MILP will be
easy to solve when the problem is over- or under-constrained in either the capacity or the
resource bounds. To empirically verify this, we varied local capacity constraint levels from 0
(meaning agents cannot use any resources) to 1 (meaning each agent has the capacity to
use enough resources to execute its optimal unconstrained policy), as well as the global
constraint levels for which 0 meant that no resources were available to the agents, and 1
meant that there were enough resources to assign to each agent its most desired resource
bundle. In all of our experiments, the part of the MILP solver was played by CPLEX 8.1
on a Pentium-4 machine with 2GB of RAM (RAM was not a bottleneck due to the use of
sparse matrix representations). A typical running-time profile is shown in Figure 4. The
problem is very easy when it is over-constrained, becomes more difficult as the constraints
are relaxed and then abruptly becomes easy again when capacity and resource levels start
to approach utopia.
13. We also investigated other, randomly generated domains, and the results were qualitatively the same.
531
Dolgov & Durfee
c
e
s
,
t
6
4
2
0
1
0.5
Local Constraints
0
0
1
0.5
Global Constraints
l
e
v
e
L
t
n
i
a
r
t
s
n
o
C
l
a
c
o
L
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.2
0.4
0.6
Global Constraint Level
0.8
1
Figure 4: Running time for MDP-based winner-determination MILP (19) for different lev-
els of global ( bρ) and local (bκm ) constraints. The constraint levels are fractions
of utopian levels that are needed to implement optimal unconstrained policies.
Problems involved 10 agents, each operating on a 5-by-5 grid, with 10 shared
resource types. Each data point shown is an average of ten runs over randomly-
generated problems.
In all of the following experiments we aim to avoid the easy regions of constraint levels.
Therefore, given the complexity profiles, we set the constraint levels to 0.5 for both local
capacity and global resource bounds. We also set the discount factor to γ = 0.95. This
value was chosen arbitrarily, because our investigations into the effect of the value of γ on
the running time of the MILP revealed no significant trends.
We begin by comparing the performance of our MDP-based auction (Section 4.2) to the
performance of the straightforward CA with flat preferences (as described in Section 4.1).
The results are summarized in Figure 5, which compares the time it takes to solve the
standard winner-determination problem on the space of all resource bundles (16) to the
time needed to solve the combined MDP-WDP problem (19) used in our mechanism, as
the number of resources is increased (with 5 agents, on a 5-by-5 grid). Despite the fact
that both algorithms have exponential worst-case running time, the number of integer
variables in (16) is exponentially larger than in our MILP (19), the effect of which is clearly
demonstrated in Figure 5. Furthermore, this comparison gives an extremely optimistic
view of the performance of the standard CA, as it does not take into account the additional
complexity of the valuation problem, which requires formulating and solving a very large
number of MDPs (one per resource bundle). On the other hand, the latter is embedded into
the WDP of our mechanism (19), thus including the time for solving the valuation problem
in the comparison would only magnify the effect.
In fact, in our experiments, the time
required to solve the MDPs for the valuation problem was significantly greater than the
time for solving the resulting WDP MILP. However, we do not present quantitative results
to that effect here, because of the difference in implementation (iterating over resource
bundles and solving MDPs was done via a straightforward implementation in Matlab, while
532
Resource Allocation Among Agents with MDP-Induced Preferences
Figure 5: Gains in computation efficiency: MDP-based WDP versus a WDP in a straight-
forward CA implementation. The latter does not include the time for solving
the MDPs to compute resource-bundle values. Error bars show the standard
deviation over ten runs.
Figure 6: Scaling the MDP-based winner-determination MILP (19) to more agents. Agents
operated on 5-by-5 grids and shared 10 types of resources.
MILPs were solved using highly-optimized CPLEX code). No parallelization of the WDP
was performed for these experiments for either algorithm.
Below we analyze the performance of our algorithm on larger problems infeasible for
the straightforward CA. Figure 6 illustrates the scaling effect as the number of agents
participating in the auction is increased. Here and below, each point on the plot corresponds
to a single run of the experiment (with no less than ten runs performed for every value of
parameters), and the solid line is the mean. Recall that the size of the WDP scales linearly
533
Dolgov & Durfee
(a)
(b)
(c)
(d)
Figure 7: (a)–(c): scaling of the MDP-based winner-determination MILP (19) with the
number of resources on three sets of problems with different grid sizes (n) and
different numbers of agents (M); (d): a linear-scale plot of the tail of the data
in (c).
with the number of agents. The graph therefore reflects a rather standard scaling effect
for an NP-complete problem. As can be seen from the plot, problems with 25 agents
and 10 resource types are well within the reach of the method, on average taking around
30 minutes.
Next, we analyze how the method scales with the number of resource types. Figure 7
shows the solution time as a function of the number of resource types for three different
sets of problems. In these problems, the number of actions scaled linearly with the number
of resource types, but each action required a constant number of resources, i.e., the number
of nonzero ρ(a, o) per action was constant (two) regardless of the total number of resource
534
Resource Allocation Among Agents with MDP-Induced Preferences
Figure 8: Complexity of MDP-based winner-determination MILP (19) as a function of the
number of actions’ resource requirements.
types. These problems exhibit an interesting trait wherein the running time peaks for
relatively low numbers of resource types, then falls quickly, and then increases much more
slowly as the number of resource types increases (as illustrated in Figure 7d, which uses
a linear scale). This is due to the fact that when the total number of resource types is
much higher than the number of resources required by any action, there is less contention
for a particular resource among the agents and between one agent’s actions. Therefore, the
problems become relatively under-constrained and the solution time increases slowly.
To better illustrate this effect, we ran a set of experiments inverse to the ones shown in
Figure 7: we kept the total number of resource types constant and increased the number of
resource types required by each action. The results are shown in Figure 8. The running-time
profile is similar to what we observed earlier when we varied the local and global constraints:
when the total number of resources per action is low or high, the problem is under- or over-
constrained and is relatively easy to solve, but its complexity increases significantly when
the number of resources required by each resource is in the range of 50-80% of the total
number of resource types.
Based on the above, we would expect that if the actions’ resource requirements increased
with the total number of resource types, the problem would not scale as gracefully as in
Figure 7. For example, Figure 9 illustrates the running time for problems where the number
of resources required by each action scales linearly with the total number of resources. There,
the complexity does increase significantly faster. However, it is not unreasonable to assume
that in many domains the number of actions does not, in fact, increase with the total
number of resource types involved. Indeed, it is natural to assume that the total number
of resource types increases as the problem becomes more complicated and the number of
tasks the agent can perform increases. However, why should the resource requirements of an
action increase as well? If the delivery agent from our running example acquires the ability
to deliver pizza, it might need new resources to perform actions related to this new activity,
535
Dolgov & Durfee
Figure 9: Complexity when actions’ resource requirements grow proportionally to the total
number of resource types. The number of resource types needed by each action
is 10% of the total number of resource types O .
but one would not expect the resource requirements for delivering furniture or appliances
to change. Therefore, we believe that in many real applications, our method will scale up
gracefully with the total number of resource types.
The above experiments illustrate the point that for domains where agents have pref-
erences that are defined by underlying Markov decision processes, the resource-allocation
mechanism developed in this paper can lead to significant computational advantages. As
shown in Figure 7, the method can be successfully applied to very large problems that, we
argue, are well beyond the reach of combinatorial resource-allocation mechanisms with flat
preferences. As our experiments show (Figure 5), even for small problems, combinatorial
resource allocation mechanisms with flat preferences can be time-consuming, and our at-
tempts to empirically evaluate those simpler mechanisms on larger problems proved futile.
For instance, our method takes under one minute to solve a problem that, in the standard
CA, requires the agents to enumerate up to 2100 bundles and the auctioneer to solve an
NP-complete problem with an input of that size.
6. Generalizations, Extensions, and Conclusions
There are many possible extensions and generalizations of the work presented here, and we
briefly outline several below.
The treatment in this paper focused on the problem of resource allocation among
self-interested agents, but the algorithms also apply to cooperative MDPs with weakly-
interacting agents. In the cooperative setting, the concept of resources can be viewed as
a compact way to model inter-agent constraints and their inability to include some com-
binations of joint actions in their policies. Such weakly-coupled MDPs, where agents have
independent transition and reward functions, but certain combinations of joint actions are
536
Resource Allocation Among Agents with MDP-Induced Preferences
not feasible is a widely used model of agents’ interactions (e.g., Singh & Cohn, 1998). Our
model was resource-centric, but more direct models are also certainly possible. For exam-
ple, agents can use SAT formulas to describe valid combinations of joint actions. This case
can be easily handled via simple modifications to the single and multiagent MILPs (15)
and (19).
Indeed, any SAT formula can be expressed as a set of linear inequalities on
binary variables ∆(a) (or ∆m (a) in the multiagent case), which can be directly added to
the corresponding MILP (see the case of non-binary resources in Appendix B for an MILP
defined on indicators ∆(a), instead of the δ(o) used in the binary case).
As mentioned previously, our work can be extended to handle consumable resources that
are used up whenever agents execute actions. In fact, under some conditions, the problem
can be considerably simplified for domains with only these kinds of resources. The most
important change is that we have to redefine the value of a particular resource bundle to
an agent. The difficulty is that, given a policy, the total use of consumable resources is
uncertain, and the definition of the value of a resource bundle becomes ambiguous. One
possibility is to define the value of a bundle as the payoff of the best policy whose expected
resource usage does not exceed the amounts of resources in the bundle. The interpretation
of ρm (a, o) would also change to mean the amount of resource o consumed by action a
every time it is executed. This would make the constraints in (19) linear in the occupation
measure, which would tremendously simplify the WDP (making it polynomial). This is
analogous to the models used in constrained MDPs (Altman & Shwartz, 1991), as briefly
described earlier in Section 2. Information privacy can be handled similarly to the case of
non-consumable resources. However, given the transformation y = Dx, the resource cost
function ρm will also have to be scaled by D−1 (since the total consumption of consumable
resources is proportional to the occupation measure). This has the additional benefit of
hiding the resource cost functions (unlike the case of non-consumable resources where they
were revealed). A more detailed treatment of the model with consumable resources is
presented in the work by Dolgov (2006), including a discussion of risk-sensitive cases, where
the value of a resource bundle is defined as the payoff of the best policy whose probability
of exceeding the resource amounts is bounded.
In this work we exploited structure in agents’ preferences that stems from the underlying
policy-optimization problems. However, the latter were modeled using “flat” MDPs that
enumerate all possible states and actions. Such flat MDPs do not scale well due to the
curse of dimensionality (Bellman, 1961). To address this, the WDP MILP can be modified
to work with factored MDPs (Boutilier, Dearden, & Goldszmidt, 1995) by using a factored
resource-allocation algorithm (Dolgov & Durfee, 2006), which is based on the dual ALP
method for solving factored MDPs as developed by Guestrin (2003). This method allows us
to exploit both types of structure in the resource-allocation algorithms: structure in agents’
preferences induced by the underlying MDPs, as well as structure in MDPs themselves.
The resource-allocation mechanism discussed in this paper assumed a one-shot allocation
of resources and a static population of agents. An interesting extension of our work would
be to consider a system where agents and resources arrive and depart dynamically, as in the
online mechanism design work (Parkes & Singh, 2003; Parkes, Singh, & Yanovsky, 2004).
Combining the MDP-based model of utility functions with the dynamics of online problems
could be a valuable result and thus appears to be a worthwhile direction of future work. If
537
Dolgov & Durfee
the agent population is static, but a periodic re-allocation of resources is allowed, techniques
like phasing can be used to solve the resulting problem (Wu & Durfee, 2005).
To summarize the results of this paper, we presented a variant of a combinatorial auc-
tion for resource allocation among self-interested agents whose valuations of resource bundles
are defined by their weakly-coupled constrained MDPs. For such problems, our mechanism,
which exploits knowledge of the structure of agents’ MDP-based preferences, achieves an
exponential reduction in the number of integer decision variables, which in turn leads to
tremendous speedup over a straightforward implementation, as confirmed by our experimen-
tal results. Our mechanism can be implemented to achieve its reduction in computational
complexity without sacrificing any of the nice properties of a VCG mechanism (optimal out-
comes, strategy-proofness, and voluntary participation). We also discussed a distributed
implementation of the mechanism that retains strategy-proofness (using the fact that an
LP solution can be easily verified), and does not reveal agents’ private MDP information
(using a transformation of agents’ MDPs).
We believe that the models and solution algorithms described in this paper significantly
further the applicability of combinatorial resource-allocation mechanisms to practical prob-
lems, where the utility functions for resource bundles are defined by sequential stochastic
decision-making problems.
7. Acknowledgments
We thank the anonymous reviewers for their helpful comments, as well as our colleagues
Satinder Singh, Kang Shin, Michael Wellman, Demothenis Teneketsis, Jianhui Wu, and
Jeffrey Cox for the valuable discussions related to this work.
This material is based in part upon work supported by Honeywell International, and
by the DARPA IPTO COORDINATORs program and the Air Force Research Laboratory
under Contract No. FA8750–05–C–0030. The views and conclusions contained in this doc-
ument are those of the authors, and should not be interpreted as representing the official
policies, either expressed or implied, of the Defense Advanced Research Pro jects Agency or
the U.S. Government.
Appendix A. Proofs
A.1 Proof of Theorem 1
Theorem 1 Consider a finite set of n indivisible resources O = {oi } (i ∈ [1, n]), with
m ∈ N available units of each resource. Then, for any non-decreasing utility function
defined over resource bund les f : [0, m]n 7→ R, there exists a resource-constrained MDP
hS , A, p, r, O , ρ, C , κ, bκ , αi (with the same resource set O) whose induced utility function over
the resource bund les is the same as f . In other words, for every resource bund le z ∈ [0, m]n ,
the value of the optimal policy among those whose resource requirements do not exceed z
(cal l this set Π(z)) is the same as f (z):
∀f q : [0, m]n 7→ R, ∃ hS , A, p, r, O , ρ, C , κ, bκ , αi :
∀ z ∈ [0, m]n , Π(z) = nπ (cid:12)(cid:12)(cid:12) max
π(s, a)(cid:1)i ≤ zio =⇒ max
a hρ(a, oi )H (cid:0) Xs
π∈Π(z)
538
Uγ (π , α) = f (z).
Resource Allocation Among Agents with MDP-Induced Preferences
Figure 10: Creating an MDP with resources for an arbitrary non-decreasing utility function.
The case shown has three binary resources. All transitions are deterministic.
Proof: This statement can be shown via a straightforward construction of an MDP that
has an exponential number (one per resource bundle) of states or actions. Below we present
a reduction with a linear number of actions and an exponential number of states. Our choice
is due to the fact that, although the reverse mapping requiring two states and exponentially
many actions is even more straightforward, such an MDP feels somewhat unnatural.
Given an arbitrary non-decreasing utility function f , a corresponding MDP can be
constructed as follows (illustrated in Figure 10 for n = 3 and m = 1). The state space S of
the MDP consists of (m+ 1)n + 1 states – one state (sz ) for every resource bundle z ∈ [0, m]n ,
plus a sink state (s0 ).
The action space of the MDP A = a0 S{aij }, i ∈ [1, n], j ∈ [1, m] consists of mn + 1
actions: m actions per each resource oi , i ∈ [1, n], plus an additional action a0 .
The transition function p is deterministic and is defined as follows. Action a0 is appli-
cable in every state and leads to the sink state s0 . Every other action aij is applicable in
states sz , where zi = (j − 1) and leads with certainty to the states where zi = j :
p(σ s, a) =
In other words, aij only applies in states that have j − 1 units of resource i and leads to the
state where the amount of ith resource increases to j .
The reward function r is defined as follows. There are no rewards in state s0 , and
action a0 is the only action that produces rewards in other states:
r(s, a) = (f ′(z) a = ao , s = sz , ∀z ∈ [0, m]n
0
otherwise,
where f ′ is a simple transformation of f that compensates for the effects of discounting:
1 a = aij , s = sz , σ = sz′ , zi = (j − 1), z ′
i = j,
1 a = a0 , σ = s0 ,
0 otherwise.
f ′ (z) = f (z)(γ )− Pi zi .
539
Dolgov & Durfee
In other words, it takes Pi zi transitions to get to state sz , so the contribution of the above
into the total discounted reward will be exactly f (z).
The resource requirements ρ of actions are as follows: action a0 does not require any
resources, while every other action aij requires j units of resource oi .
Finally, the initial conditions are α(sz=0 ) = 1, meaning that the agent always starts
in the state that corresponds to the empty resource bundle (state s000 in Figure 10). The
capacity costs κ and limits bκ are not used, so we set C = ∅.
It is easy to see that in the MDP constructed above, given a resource bundle z, any
policy from the feasible set Π(z) has zero probability of reaching any state sz′ for which
z′ > z (for any component i). Furthermore, an optimal policy from the set Π(z) will be to
transition to state sz (since f (z) is non-decreasing) and then use action a0 , thus obtaining
a total discounted reward of f (z).
(cid:3)
A.2 Proof of Theorem 2
Theorem 2 Given an MDP M = hS , A, p, r, O , ρ, C , κ, bκ , αi with resource and capacity
constraints, if there exists a policy π ∈ ΠHR that is a feasible solution for M , there exists a
stationary deterministic policy πSD ∈ ΠSD that is also feasible, and the expected total reward
of πSD is no less than that of π :
∀ π ∈ ΠHR , ∃ πSD ∈ ΠSD : Uγ (πSD , α) ≥ Uγ (π , α)
Proof: Let us label A′ ⊆ A the set of all actions that have a non-zero probability of being
executed according to π , i.e.,
A′ = {a∃s : π(s, a) > 0}
Let us also construct an unconstrained MDP: M ′ = hS , A′ , p′ , r ′ i, where p′ and r ′ are the
restricted versions of p and r with the action domain limited to A′ :
p′ : S × A′ × S 7→ [0, 1]
r ′ : S × A′ 7→ R
p′ (σ s, a) = p(σ s, a), r ′ (s, a) = r(s, a) ∀s ∈ S , σ ∈ S , a ∈ A′
Due to a well-known property of unconstrained infinite-horizon MDPs with the total
expected discounted reward optimization criterion, M ′ is guaranteed to have an optimal
stationary deterministic solution (e.g., Theorem 6.2.10, Puterman, 1994), which we label
πSD .
Consider πSD as a potential solution to M . Clearly, πSD is a feasible solution, because
its actions come from the set A′ that includes actions that π uses with non-zero probability,
which means that the resource requirements (as in (9)) of πSD can be no greater than those
of π . Indeed:
π(s, a)(cid:1)o,
πSD (s, a)(cid:1)o ≤ max
a∈A nρ(a, o)H (cid:0) Xs
a∈A′ nρ(a, o)H (cid:0) Xs
ρ(a, o) = max
max
a∈A′
where the first inequality is due to the fact that H (z ) ≤ 1 ∀z , and the second equality
follows from the definition of A′ .
(30)
540
Resource Allocation Among Agents with MDP-Induced Preferences
Figure 11: Reduction of KNAPSACK to M-OPER-CMDP. All transitions are determinis-
tic.
Furthermore, observe that πSD yields the same total reward under M ′ and M . Addi-
tionally, since πSD is a uniformly optimal solution to M ′ , it is, in particular, optimal for
the initial conditions α of the constrained MDP M . Therefore, πSD constitutes a feasible
solution to M whose expected reward is greater than or equal to the expected reward of
any feasible policy π .
(cid:3)
A.3 Proof of Theorem 3
Theorem 3 The fol lowing decision problem is NP-complete. Given an instance of an MDP
hS , A, p, r, O , ρ, C , κ, bκ , αi with resources and capacity constraints, and a rational number Y ,
does there exist a feasible policy π , whose expected total reward, given α, is no less than Y ?
Proof: As shown in Theorem 2, there always exists an optimal policy for (9) that is
stationary and deterministic. Therefore, the presence in NP is obvious, since we can, in
polynomial time, guess a stationary deterministic policy, verify that it satisfies the resource
constraints, and calculate its expected total reward (the latter can be done by solving the
standard system of linear Markov equations (2) on the values of all states).
To show NP-completeness of the problem, we use a reduction from KNAPSACK (Garey
& Johnson, 1979). Recall that KNAPSACK in an NP-complete problem, which asks
whether, for a given set of items z ∈ Z , each of which has a cost c(z ) and a value v(z ),
there exists a subset Z ′ ⊆ Z such that the total value of all items in Z ′ is no less than
some constant bv , and the total cost of the items is no greater than another constant bc, i.e.,
Pz∈Z ′ c(z ) ≤ bc and Pz∈Z ′ v(z ) ≥ bv . Our reduction is illustrated in Figure 11 and proceeds
as follows.
Given an instance of KNAPSACK with Z = m, let us number all items as zi , i ∈
[1, m] as a notational convenience. For such an instance of KNAPSACK, we create an
MDP with m + 1 states {s1 , s2 , . . . sm+1}, m + 1 actions {a0 , . . . am}, m types of resources
O = {o1 , . . . om}, and a single capacity C = {c1 }.
The transition function on these states is defined as follows. Every state si , i ∈ [1, m]
has two transitions from it, corresponding to actions ai and a0 . Both actions lead to state
si+1 with probability 1. State sm+1 is absorbing and all transitions from it lead back to
itself.
The reward and the cost functions are defined as follows. We want action ai , i ∈ [1, m]
(which corresponds to item zi in KNAPSACK) to contribute v(zi ) to the total discounted
541
Dolgov & Durfee
reward. Hence, we set the immediate reward for every action ai to v(zi )(γ )1−i , which, given
that our transition function implies that state si is reached exactly at step i − 1, ensures
that if action ai is ever executed, its contribution to the total discounted reward will be
v(zi )(γ )1−i (γ )i−1 = v(zi ). Action a0 produces a reward of zero in all states.
The resource requirements of actions are defined as follows. Action ai , i ∈ [1, m] only
needs resource oi , i.e., ρ(ai , oj ) = 1 ⇐⇒ i = j . We set the cost of resource oi to be the
cost c(zi ) of item i in the KNAPSACK problem. The “null” action a0 requires no resources.
In order to complete the construction, we set the initial distribution α = [1, 0, . . . ] so that
the agent starts in state s1 with probability 1. We also define the decision parameter Y = bv
and the upper bound on the single capacity bκ = bc.
Essentially, this construction allows the agent to choose action ai or a0 at every state si .
Choosing action ai is equivalent to putting item zi into the knapsack, while action a0
corresponds to the choice of not including zi in the knapsack. Therefore, there exists a
policy that has the expected payoff no less than Y = bv and uses no more than bκ = bc
of the shared resource if and only if there exists a solution to the original instance of
KNAPSACK.
(cid:3)
Appendix B. Non-binary Resource Requirements
Below we describe an MILP formulation of the capacity-constrained single-agent optimiza-
tion problem (9) for arbitrary resource costs ρ : A × O 7→ R, as opposed to binary costs
that were assumed in the main parts of the paper. The corresponding multiagent winner-
determination problem (the non-binary equivalent of (19)) follows immediately from the
single-agent MILP.
For arbitrary resource costs, we obtain the following non-binary equivalent of the opti-
mization problem (12) in the occupation measure coordinates:
x Xs Xa
max
sub ject to:
Xa
x(σ, a) − γ Xs Xa
x(s, a)p(σ s, a) = α(σ),
x(s, a)(cid:1)o ≤ bκ(c),
a nρ(a, o)H (cid:0) Xs
Xo
κ(o, c) max
x(s, a) ≥ 0,
To linearize the sum of max operators in (31), let us observe that the inequality
nXi
is equivalent to the following system of Z n linear inequalities:
f (z , u1 ) + . . . + g(un ) max
z∈Z
∀s ∈ S , a ∈ A.
g(ui ) max
z∈Z
f (z , ui ) = g(u1 ) max
z∈Z
f (z , un ) ≤ a
x(s, a)r(s, a)
∀σ ∈ S ;
∀c ∈ C ;
(31)
g(u1 )f (z1 , u1 ) + g(u2 )f (z2 , u2 ) + . . . + g(un )f (zn , un ) ≤ a,
∀z1 , z2 , . . . zn ∈ Z .
542
Resource Allocation Among Agents with MDP-Induced Preferences
∀c ∈ C
(32)
Applying this to the constraints from (31), we can express the original system of C nonlinear
constraints (each of which has a max):
x(s, a)(cid:1)o ≤ bκ(c),
a nρ(a, o)H (cid:0) Xs
Xo
κ(o, c) max
as the following system of C AO constraints where the max is removed:
κ(o, c)ρ(ao , o)H (cid:0) Xs
Xo
x(s, a)(cid:1) ≤ bκ(c),
∀c ∈ C , ao1 , ao2 , . . . ∈ A.
Notice that this way of eliminating the maximization exponentially increases the number
of constraints, because the above expansion enumerates all possible actions for each resource
(i.e., it enumerates policies where each resource o is used by action a1 , where it is used by
action a2 , action a3 , etc.) However, in many problems not all resources are used by all
actions. In such cases, most of the above constraints become redundant, and the number of
constraints can be reduced from C AO to C Qo Ao , where Ao is the number of actions
that use resource o.
We can linearize the Heaviside function analogously to the case of binary resource costs
in Section 3.2: we create a binary indicator variable that corresponds to the argument of
H () and tie it to the occupation measure x via linear inequalities. The only difference is that
for non-binary resource costs, instead of using indicators on resources, we use indicators on
actions: ∆(a) ∈ {0, 1}, where ∆(a) = H (Ps x(s, a)) is an indicator that shows whether
action a is used in the policy. Using ∆ and expanding the max as above, we can represent
the optimization problem (9) as the following MILP:
x,∆ Xs Xa
max
sub ject to:
x(σ, a) − γ Xs Xa
Xa
x(s, a)p(σ s, a) = α(σ),
Xo
κ(o, c)ρ(ao , o)∆(ao ) ≤ bκ(c),
Xs
x(s, a)/X ≤ ∆(a),
x(s, a) ≥ 0,
∆(a) ∈ {0, 1},
∀a ∈ A,
where X ≥ max Ps x(s, a) is some constant finite upper bound on the expected number
of times action a is used, which exists for any discounted MDP. We can, for example, let
X = (1 − γ )−1 , since Ps,a x(s, a) = (1 − γ )−1 for any x that is a valid occupation measure
for an MDP with discount factor γ .
Example 9 Let us formulate the MILP for the constrained problem from Example 3. Recal l
that there are three resources O = {ot , of , om } (truck, forklift, and mechanic), one capacity
∀c ∈ C , ao1 , ao2 , . . . ∈ A;
x(s, a)r(s, a)
∀a ∈ A;
∀s ∈ S , a ∈ A;
∀σ ∈ S ;
(33)
543
Dolgov & Durfee
type C = {c1 } (money), and actions have the fol lowing resource requirements (listing only
the nonzero ones):
ρ(a1 , ot ) = 1, ρ(a2 , ot ) = 1, ρ(a2 , of ) = 1, ρ(a3 , ot ) = 1, ρ(a4 , ot ) = 1, ρ(a4 , om ) = 1
The resources have the fol lowing capacity costs:
κ(ot , c1 ) = 2, κ(of , c1 ) = 3, κ(om , c1 ) = 4,
and the agent has a limited budget, i.e., a capacity bound bκ(c1 ) = 8.
To compute the optimal policy for an arbitrary α, we can formulate the problem as an
MILP using the techniques described above. Using binary variables {∆(ai )} = {∆i} =
{∆1 , ∆2 , ∆3 , ∆4},14 we can express the constraint on capacity cost as the fol lowing system
of C AO = 1(4)3 = 64 linear constraints:
(2)(1)∆1 + (3)(0)∆1 + (4)(0)∆1 ≤ 8,
(2)(1)∆1 + (3)(0)∆1 + (4)(0)∆2 ≤ 8,
(2)(1)∆1 + (3)(0)∆1 + (4)(0)∆3 ≤ 8,
(2)(1)∆1 + (3)(0)∆1 + (4)(1)∆4 ≤ 8,
(2)(1)∆1 + (3)(1)∆2 + (4)(0)∆1 ≤ 8,
. . .
(2)(0)∆4 + (3)(0)∆4 + (4)(1)∆4 ≤ 8.
It is easy to see that most of these constraints are redundant, and the fact that each action
only requires a smal l subset of the resources al lows us to prune many of the constraints. In
fact, the only resource that is used by multiple actions is ot . Therefore, in accordance with
our earlier discussion, we only need Qo Ao = 1 × 4 × 1 = 4 constraints:
(2)(1)∆1 + (3)(1)∆2 + (4)(1)∆4 ≤ 8,
(2)(1)∆2 + (3)(1)∆2 + (4)(1)∆4 ≤ 8,
(2)(1)∆3 + (3)(1)∆2 + (4)(1)∆4 ≤ 8,
(2)(1)∆4 + (3)(1)∆2 + (4)(1)∆4 ≤ 8,
where each of the four constraints corresponds to a case where the first resource (ot ) is used
by a different action.
As mentioned earlier, we can set X = (1 − γ )−1 for the constraints that synchronize the
occupation measure x and the binary indicators ∆. Combining this with other constraints
from (33), we get an MILP with 12 continuous and 4 binary variables, and S + C Qo Ao +
A = 3 + 4 + 3 = 10 constraints (not counting the last two sets of range constraints). (cid:3)
Finally, let us observe that by expanding the resource and action sets, any problem
can be represented using binary resources only.
If the domain contains mostly binary
requirements, it may be more effective to expand the non-binary resource requirements ρ
by augmenting the resource set O , and then use the binary formulation of Section 3.2 rather
than directly applying the more-general formulation described above.
14. We do not create a ∆0 for the noop action a0 , as its resource costs are zero, and it drops out of all
expressions.
544
Resource Allocation Among Agents with MDP-Induced Preferences
Appendix C. Experimental Setup
This appendix details how our experimental domains were constructed. For a delivery
domain with M agents operating on an n-by-n grid and sharing O resource types, we
used the following parameters.
The resources enable agents to carry out delivery tasks. For a problem with O resource
types, there are O delivery actions, and performing action i ∈ [1, O ] requires a random
subset of resources from O (where the number of resources required by an action is an
important parameter, whose effect on complexity is discussed in Section 5). The probability
that task i ∈ [1, O ] can be carried at a location is 0.1 + 0.4(O − i)/(O − 1), i.e., uniformly
distributed between 0.1 and 0.5, as a function of the action ID (actions with lower IDs are
more rewarding, per the definition of the reward function below, but can be executed at
fewer locations).
There are n2/5 possible delivery locations randomly placed on the grid. Each delivery
location is assigned a set of delivery tasks that can be executed there (a single location can
be used for multiple delivery tasks, and a single task can be carried out at any of several
locations). The assignment of tasks to locations is done randomly.
Each agent has 4 + O actions: drive in any of the four perpendicular directions and
execute one of the delivery tasks. The drive actions result in movement in the intended
direction with probability of 0.8 and with probability of 0.2 produce no change of location.
All movement actions incur a negative reward, the amount of which depends on the size of
the agent. For a problem with M agents, the movement penalty incurred by agent m ∈
[1, M] is −1 − 9(m − 1)/(M − 1), i.e., distributed uniformly on [−1, −10] as a function
of the agent’s ID.
Execution of an action corresponding to a delivery task i ∈ [1, O ] in a location to which
the task is assigned produces a reward of 100i/O and moves the agent to a new random
location on the grid. The new location is chosen randomly at problem generation (thus
known to agent), but the transition is deterministic, which induces a topology with nearby
and remote locations. Attempting execution of a delivery task in an incorrect location does
not change state and produces zero reward.
The agents bid for delivery resources of O types. There are cg lob M units of each
resource, where cg lob is the global constraint level (set to 0.5 for most of our experiments,
as described in more detail in Section 5). There is one capacity type: size. The size
requirements for making deliveries of type i ∈ [1, O ] is i. The capacity limit of agent m
is cloc/2O (O + 1), where cloc is the local constraint level (set to 0.5 for most of our
experiments, as was described in more detail in Section 5).
The initial location of each agent is randomly selected from a uniform distribution. The
discount factor is γ = 0.95.
References
Altman, E. (1996). Constrained Markov decision processes with total cost criteria: Occu-
pation measures and primal LP. Methods and Models in Operations Research, 43 (1),
45–72.
545
Dolgov & Durfee
Altman, E., & Shwartz, A. (1991). Adaptive control of constrained Markov chains: Cri-
teria and policies. Annals of Operations Research, special issue on Markov Decision
Processes, 28, 101–134.
Altman, E. (1999). Constrained Markov Decision Processes. Chapman and HALL/CRC.
Bellman, R. (1961). Adaptive Control Processes: A Guided Tour. Princeton University
Press.
Benazera, M. E., Brafman, R. I., Meuleau, N., & Hansen, E. (2005). Planning with contin-
uous resources in stochastic domains. In Proceedings of the Nineteenth International
Joint Conference on Artificial Intel ligence (IJCAI-05), pp. 1244–1251.
Bererton, C., Gordon, G., & Thrun, S. (2003). Auction mechanism design for multi-robot
coordination. In Thrun, S., Saul, L., & Scholkopf, B. (Eds.), Proceedings of Conference
on Neural Information Processing Systems (NIPS). MIT Press.
Bertsimas, D., & Tsitsiklis, J. N. (1997).
Scientific.
Introduction to Linear Optimization. Athena
Boutilier, C. (2002). Solving concisely expressed combinatorial auction problems. In Pro-
ceedings of the Eighteenth National Conference on Artificial Intel ligence (AAAI-02),
pp. 359–366.
Boutilier, C., Dearden, R., & Goldszmidt, M. (1995). Exploiting structure in policy con-
struction. In Proceedings of the Fourteenth International Joint Conference on Artificial
Intel ligence (IJCAI-95), pp. 1104–1111.
Boutilier, C., & Hoos, H. H. (2001). Bidding languages for combinatorial auctions. In Pro-
ceedings of the Seventeenth International Joint Conference on Artificial Intel ligence
(IJCAI-01), pp. 1211–1217.
Clarke, E. H. (1971). Multipart pricing of public goods. Public Choice, 18, 19–33.
de Vries, S., & Vohra, R. V. (2003). Combinatorial auctions: A survey. INFORMS Journal
on Computing, 15 (3), 284–309.
Dolgov, D. (2006). Integrated Resource Al location and Planning in Stochastic Multiagent
Environments. Ph.D. thesis, Computer Science Department, University of Michigan.
Dolgov, D. A., & Durfee, E. H. (2004). Optimal resource allocation and policy formula-
tion in loosely-coupled Markov decision processes. In Proceedings of the Fourteenth
International Conference on Automated Planning and Scheduling (ICAPS-04), pp.
315–324.
Dolgov, D. A., & Durfee, E. H. (2006). Resource allocation among agents with preferences
induced by factored MDPs. In Proceedings of the Fifth International Joint Conference
on Autonomous Agents and Multiagent Systems (AAMAS-06), Hakodate, Japan.
Eckstein, J., Phillips, C., & Hart, W. (2000). Pico: An ob ject-oriented framework for parallel
branch and bound. In Proceedings of the Workshop on Inherently Paral lel Algorithms
in Optimization and Feasibility and their Applications.
Feigenbaum, J., & Shenker, S. (2002). Distributed algorithmic mechanism design: Recent
results and future directions.
In Proceedings of the Sixths International Workshop
546
Resource Allocation Among Agents with MDP-Induced Preferences
on Discrete Algorithms and Methods for Mobile Computing and Communications, pp.
1–13. ACM Press, New York.
Feldmann, R., Gairing, M., Lucking, T., Monien, B., & Rode, M. (2003). Selfish routing in
non-cooperative networks: A survey. In Proceedings of the Twenty-Eights International
Symposium on Mathematical Foundations of Computer Science (MFCS-03), pp. 21–
45. Springer-Verlag.
Ferguson, D., Nikolaou, C., Sairamesh, J., & Yemini, Y. (1996). Economic models for allo-
cating resources in computer systems. In Clearwater, S. (Ed.), Market-Based Control:
A Paradigm for Distributed Resource Al location, pp. 156–183, Hong Kong. World
Scientific.
Garey, M. R., & Johnson, D. S. (1979). Computers and Intractability: A Guide to the Theory
of NP-Completeness. W. H. Freeman & Co.
Groves, T. (1973). Incentives in teams. Econometrica, 41 (4), 617–631.
Guestrin, C. (2003). Planning Under Uncertainty in Complex Structured Environments.
Ph.D. thesis, Computer Science Department, Stanford University.
Heyman, D. P., & Sobel, M. J. (1984). Volume II: Stochastic Models in Operations Research.
McGraw-Hill, New York.
Kallenberg, L. (1983). Linear Programming and Finite Markovian Control Problems. Math.
Centrum, Amsterdam.
Littman, M. L., Dean, T. L., & Kaelbling, L. P. (1995). On the complexity of solving Markov
decision problems. In Proceedings of the Eleventh Annual Conference on Uncertainty
in Artificial Intel ligence (UAI–95), pp. 394–402, Montreal.
MacKie-Mason, J. K., & Varian, H. (1994). Generalized Vickrey auctions. Tech. rep.,
University of Michigan.
Mas-Colell, A., Whinston, M. D., & Green, J. R. (1995). Microeconomic Theory. Oxford
University Press, New York.
McAfee, R. P., & McMillan, J. (1996). Analyzing the airwaves auction. Journal of Economic
Perspectives, 10 (1), 159–75.
McMillan, J. (1994). Selling spectrum rights. Journal of Economic Perspectives, 8 (3),
145–62.
Meuleau, N., Hauskrecht, M., Kim, K.-E., Peshkin, L., Kaelbling, L., Dean, T., & Boutilier,
C. (1998). Solving very large weakly coupled Markov decision processes. In Proceedings
of the Fifteenth National Conference on Artificial Intel ligence (AAAI-98), pp. 165–
172.
Nisan, N. (2000). Bidding and allocation in combinatorial auctions. In Electronic Commerce.
Parkes, D. (2001). Iterative Combinatorial Auctions: Achieving Economic and Computa-
tional Efficiency. Ph.D. thesis, Department of Computer and Information Science,
University of Pennsylvania.
Parkes, D. C., Kalagnanam, J. R., & Eso, M. (2001). Achieving budget-balance with
Vickrey-based payment schemes in exchanges. In Proc. 17th International Joint Con-
ference on Artificial Intel ligence (IJCAI-01), pp. 1161–1168.
547
Dolgov & Durfee
Parkes, D. C., & Shneidman, J. (2004). Distributed implementations of Vickrey-Clarke-
Groves mechanisms. In Proceedings of the Third International Joint Conference on
Autonomous Agents and Multi Agent Systems (AAMAS-04), pp. 261–268.
Parkes, D. C., & Singh, S. (2003). An MDP-based approach to Online Mechanism Design. In
Proceedings of the Seventeenths Annual Conference on Neural Information Processing
Systems (NIPS-03).
Parkes, D. C., Singh, S., & Yanovsky, D. (2004). Approximately efficient online mechanism
design. In Proceedings of the Eighteenths Annual Conference on Neural Information
Processing Systems (NIPS-04).
Puterman, M. L. (1994). Markov Decision Processes. John Wiley & Sons, New York.
Ross, K., & Chen, B. (1988). Optimal scheduling of interactive and non-interactive traffic in
telecommunication systems. IEEE Transactions on Automatic Control, 33, 261–267.
Ross, K., & Varadara jan, R. (1989). Markov decision processes with sample path con-
straints: the communicating case. Operations Research, 37, 780–790.
Rothkopf, M. H., Pekec, A., & Harstad, R. M. (1998). Computationally manageable com-
binational auctions. Management Science, 44 (8), 1131–1147.
Sandholm, T., & Boutilier, C. (2006). Preference elicitation in combinatorial auctions. In
Cramton, Shoham, & Steinberg (Eds.), Combinatorial Auctions, chap. 10. MIT Press.
Sandholm, T. (1999). An algorithm for optimal winner determination in combinatorial
auctions. In Proceedings of the Sixteenth International Joint Conference on Artificial
Intel ligence (IJCAI-99), pp. 542–547, San Francisco, CA, USA. Morgan Kaufmann
Publishers Inc.
Sandholm, T. (2002). Algorithm for optimal winner determination in combinatorial auc-
tions. Artificial Intel ligence, 135 (1-2), 1–54.
Shapley, L. S. (1953). Stochastic games. Proceedings of National Academy of Science, USA,
39, 1095–1100.
Sheffi, Y. (2004). Combinatorial auctions in the procurement of transportation services.
Interfaces, 34 (4), 245–252.
Singh, S., & Cohn, D. (1998). How to dynamically merge Markov decision processes. In
Jordan, M. I., Kearns, M. J., & Solla, S. A. (Eds.), Advances in Neural Information
Processing Systems, Vol. 10, pp. 1057–1063. The MIT Press.
Song, J., & Regan, A. (2002). Combinatorial auctions for transportation service procure-
ment: The carrier perspective. Transportation Research Record, 1833, 40–46.
Vickrey, W. (1961). Counterspeculation, auctions and competitive sealed tenders. Journal
of Finance, 16, 8–37.
Wellman, M. P., Walsh, W. E., Wurman, P. R., & MacKie-Mason, J. K. (2001). Auction
protocols for decentralized scheduling. Games and Economic Behavior, 35, 271–303.
Wolsey, L. (1998). Integer Programming. John Wiley & Sons.
548
Resource Allocation Among Agents with MDP-Induced Preferences
Wu, J., & Durfee, E. H. (2005). Automated resource-driven mission phasing techniques for
constrained agents. In Proceedings of the Fourth International Joint Conference on
Autonomous Agents and Multiagent Systems (AAMAS-05), pp. 331–338.
549
|
1804.07183 | 1 | 1804 | 2018-04-19T14:16:50 | Industrial Symbiotic Networks as Coordinated Games | [
"cs.MA",
"cs.GT"
] | We present an approach for implementing a specific form of collaborative industrial practices-called Industrial Symbiotic Networks (ISNs)-as MC-Net cooperative games and address the so called ISN implementation problem. This is, the characteristics of ISNs may lead to inapplicability of fair and stable benefit allocation methods even if the collaboration is a collectively desired one. Inspired by realistic ISN scenarios and the literature on normative multi-agent systems, we consider regulations and normative socioeconomic policies as two elements that in combination with ISN games resolve the situation and result in the concept of coordinated ISNs. | cs.MA | cs | Industrial Symbiotic Networks as Coordinated Games
Vahid Yazdanpanah
University of Twente
Enschede, The Netherlands
[email protected]
Extended Abstract
Devrim Murat Yazan
University of Twente
Enschede, The Netherlands
[email protected]
Henk Zijm
University of Twente
Enschede, The Netherlands
[email protected]
8
1
0
2
r
p
A
9
1
]
A
M
.
s
c
[
1
v
3
8
1
7
0
.
4
0
8
1
:
v
i
X
r
a
ABSTRACT
We present an approach for implementing a specific form of collab-
orative industrial practices-called Industrial Symbiotic Networks
(ISNs)-as MC-Net cooperative games and address the so called ISN
implementation problem. This is, the characteristics of ISNs may
lead to inapplicability of fair and stable benefit allocation methods
even if the collaboration is a collectively desired one. Inspired by
realistic ISN scenarios and the literature on normative multi-agent
systems, we consider regulations and normative socioeconomic poli-
cies as two elements that in combination with ISN games resolve
the situation and result in the concept of coordinated ISNs.
KEYWORDS
Game Theory for Practical Applications; Industrial Symbiosis; MC-
Net Cooperative Games; Normative Coordination; Policy and Reg-
ulation
ACM Reference Format:
Vahid Yazdanpanah, Devrim Murat Yazan, and Henk Zijm. 2018. Industrial
Symbiotic Networks as Coordinated Games. In Proc. of the 17th International
Conference on Autonomous Agents and Multiagent Systems (AAMAS 2018),
Stockholm, Sweden, July 10–15, 2018, IFAAMAS, 3 pages.
1 INTRODUCTION
Industrial Symbiotic Networks (ISNs) are collaborative networks of
industries with the aim to reduce their materials and energy foot-
print by circulating reusable resources (e.g, physical waste material)
among the network members [5, 11, 18]. Such a symbiosis leads to
socioeconomic and environmental benefits for involved firms and
the society. One barrier against stable ISN implementations is the
lack of frameworks able to secure such networks against unfair and
unstable allocation of obtainable benefits among the involved firms.
In other words, even if economic benefits are foreseeable, lack of
stability and/or fairness may lead to non-cooperative decisions and
hence unimplementability of ISNs (ISN implementation problem).
Reviewing recent contributions in the field of industrial symbiosis
research, we encounter studies focusing on the interrelations be-
tween industrial enterprises [18] and the role of contracts in the
process of ISN implementation [1]. We believe a missed element for
shifting from theoretical ISN design to practical ISN implementation
is to model, reason about, and support ISN decisions in a dynamic
way-and not by using snapshot-based modeling frameworks.
Proc. of the 17th International Conference on Autonomous Agents and Multiagent Systems
(AAMAS 2018), M. Dastani, G. Sukthankar, E. André, S. Koenig (eds.), July 10–15, 2018,
Stockholm, Sweden. © 2018 International Foundation for Autonomous Agents and
Multiagent Systems (www.ifaamas.org). All rights reserved.
This abstract reports on extending the game-theoretic approach
of [19] with regulative rules and normative socioeconomic policies-
following the successful line of work on normative multi-agent
systems [3, 7, 17]. The extension provides a scalable solution to
the ISN implementation problem and enables enforcing desired
industrial collaborations in a fair and stable manner.
1.1 Research Questions
The following questions guide the design of a game-theoretic frame-
work and its normative coordination mechanism that jointly facili-
tate the implementation of ISNs:
(1) ISN Games: How to define a game-theoretic basis for ISNs that
both reflects their operational cost dynamics and allows the
integration of normative rules?
(2) ISN Coordination: How to uniformly represent the regulatory
dimension of ISNs using incentive rules and normative policies?
(3) Coordinated ISN Games: How to develop a framework that inte-
grates normative coordination methods into ISN games to enable
the fair and stable implementation of desirable ISNs-with re-
spect to an established policy?
Dealing with ISNs' complex industrial context [20], an ideal ISN
implementation platform would be tunable to specific industrial
settings, scalable for implementing various ISN topologies, and
would not require industries to sacrifice financially nor restrict
their freedom in the market. Below, we present the overview of an
approach for developing an ISN implementation framework with
properties close to the ideal one.
2 OVERVIEW OF THE APPROACH
As discussed in [1, 19], the total obtainable cost reduction (as an
economic benefit) and its allocation among involved firms are key
drivers behind the stability of ISNs. For any set of agents involved in
an ISN, this value-i.e., the obtainable cost reduction-characterizes
the value of the set and hence can be seen as a basis for formulating
ISNs as cooperative games. On the other hand, in realistic ISNs, the
symbiotic practice takes place in presence of economic, social, and
environmental policies and under regulations that aim to enforce
the policies by nudging the behavior of agents towards desired ones.
This is, while policies generally indicate whether an ISN is "good
(bad, or neutral)", the regulations are a set of norms that-in case of
agents' compliance-result in an acceptable spectrum of collective
behaviors. We follow this normative perspective and aim to use
normative coordination to guarantee the implementability of desir-
able ISNs-modeled as games-in a stable and fair manner. In the
following subsections, we indicate how ISN games can be modeled
and coordinated using regulatory incentive rules and normative
socioeconomic policies.
2.1 ISNs as Cooperative Games
In the game-theoretic representation of ISNs, the value of any set
of agents S is defined [19] using the difference between the total
cost that firms have to pay in case the ISN does not occur, i.e. costs
to discharge wastes and to purchase traditional primary inputs (de-
noted by T(S)), and the total cost that firms have to pay collectively
in case the ISN is realized, i.e. costs for recycling and treatment, for
transporting resources among firms, and transaction costs (denoted
by O(S)). Formally, the ISN among agents in a non-empty finite
set of agents N is a normalized superadditive cooperative game
(N , v) where for S ⊆ N , v(S) is equal to T(S) − O(S) if S > 1, and
0 otherwise.
Benefit sharing is crucial in the process of ISN implementation,
mainly because of stability and fairness concerns. Roughly speak-
ing, firms are rational agents that defect unbeneficial collaborations
(instability) and mostly tend to reject relations in which benefits
are not shared according to contributions (unfairness). Focusing
on the Core and Shapley allocations [12, 16]-as standard methods
that characterize stability and fairness-these solution concepts
appear to be applicable in a specific class of ISNs but are not gen-
erally scalable for value allocation in the implementation phase
of ISNs. In particular, relying on the balancedness of two-person
ISN games, denoted by ISNΛ, we can show that any ISNΛ is imple-
mentable in a fair and stable manner. However, in larger games-as
balancedness does not hold necessarily-the core of the game may
be empty which in turn avoids an ISN implementation that is rea-
sonable for all the involved firms. So, even if a symbiosis could
result in collective benefits, it may not last due to instable or unfair
implementations. A natural response which is in-line with realistic
practices is to employ monetary incentives as a means of normative
coordination-to guarantee the implementability of "desired" ISNs.
To allow a smooth integration with normative rules, we transform
ISN games into basic MC-Nets1 through the following steps: let
(N , v) be an arbitrary ISN game, S≥2 = {S ⊆ N : S ≥ 2} be the
set of all groups with two or more members where K = S≥2
denotes its cardinality. We start with an empty set of MC-Net
rules. Then for all groups Si ∈ S≥2, for i = 1 to K, we add a
rule {ρi : (Si , N \ Si) (cid:55)→ vi = T(Si) − O(Si)} to the MC-Net.
2.2 Normative Coordination of ISNs
Following [7, 17], we see that norms can be employed as game
transformations to bring about more desirable outcomes in ISN
games. For this account, given the economic, environmental, and
social dimensions and with respect to potential socioeconomic con-
sequences, ISNs can be partitioned in three classes by a normative
socioeconomic policy function ℘ : 2N (cid:55)→ {p+, p◦, p−}, where N is
a finite set of firms. Moreover, p+, p◦, and p− are labels-assigned
by a third-party authority-indicating whether an ISN is promoted,
permitted, or prohibited, respectively.
v(S) will be equal to
1A basic MC-Net represents a game in N as a set of rules {ρi : (Pi, Ni) (cid:55)→ vi }i∈K ,
where Pi ⊆ N , Ni ⊂ N , Pi ∩ Ni = ∅, vi ∈ R \ {0}, and K is the set of rule
indices. For a group S ⊆ N , a rule ρi is applicable if Pi ⊆ S and Ni ∩ S = ∅. Then
i∈Π(S) vi where Π(S) denote the set of rule indices that are
applicable to S. This rule-based representation allows natural integration with rule-
based coordination methods and results in relatively low complexity for computing
allocation methods such as the Shapley value [8, 10].
ι(S) (cid:66)
The rationale behind introducing policies is mainly to make sure
that the set of promoted ISNs are implementable in a fair and sta-
ble manner while prohibited ones are instable. To ensure this, in
real ISN practices, the regulatory agent introduces monetary incen-
tives, i.e., ascribes subsidies to promoted and taxes to prohibited
collaborations. We follow this practice and employ a set of rules to
ensure/avoid the implementability of desired/undesired ISNs by al-
locating incentives2. Such a set of incentive rules can be represented
by an MC-Net ℜ = {ρi : (Pi ,Ni) (cid:55)→ ιi}i∈K in which K is the set
of rule indices. Then, the incentive value for S ⊆ N , is defined as
i∈ℑ(S) ιi where ℑ(S) denotes the set of rule indices that
are applicable to S. It is provable that for any ISN game there exists
a set of incentive rules to guarantee its implementability.
2.3 Coordinated ISN Games
Having policies and regulations, we integrate them into ISN games
and introduce the concept of Coordinated ISNs (C−ISNs). Formally,
let G be an ISN and ℜ be a set of regulatory incentive rules, both
as MC-Nets among agents in N . Moreover, for each group S ⊆ N ,
let v(S) and ι(S) denote the value of S in G and the incentive value
of S in ℜ, respectively. We say the Coordinated ISN Game (C−ISN)
among agents in N is a cooperative game (N , c) where for each
group S, we have that c(S) = v(S) + ι(S).
It can be observed that employing such incentive rules is effective
for enforcing socioeconomic policies. In particular, we have that
for any promoted ISN game, under a policy ℘, there exist an imple-
mentable C−ISN game. Analogously, similar properties hold while
avoiding prohibited ISNs or allowing permitted ones. The presented
approach for incentivizing ISNs is advisable when the policy-maker
is aiming to ensure the implementability of a promoted ISN in an
ad-hoc way. In other words, an ℜ that ensures the implementability
of a promoted ISN G1 may ruin the implementability of another
promoted ISN G2. To avoid this, the set of collaborations that a pol-
icy ℘ marks as promoted should be mutually exclusive. Accordingly,
we have the desired result that the mutual exclusivity condition is
sufficient for ensuring the implementability of all the ISNs among
℘-promoted groups in a fair and stable manner.
3 CONCLUDING REMARKS
The details of the components for developing the ISN implemen-
tation framework-rooted in cooperative games and coordinated
with normative rules-consist of algorithms for generating incen-
tive rules and policy properties to ensure the implementability of
promoted ISNs. We plan to explore the possibility of having multi-
ple policies and tools for policy option analysis [13] in ISNs. Then,
possible regulation conflicts can be resolved using prioritized rule
sets (inspired by formal argumentation theory [9, 15]). We also aim
to focus on administration of ISNs by modeling them as norma-
tive multi-agent organizations [4, 20] and relying on norm-aware
frameworks [2, 6] that enable monitoring organizational behaviors.
ACKNOWLEDGMENTS
The project leading to this work has received funding from the Eu-
ropean Union's Horizon 2020 research and innovation programme
under grant agreement No. 680843.
2See [14, 21] for similar approaches on incentivizing cooperative games.
REFERENCES
[1] Vito Albino, Luca Fraccascia, and Ilaria Giannoccaro. 2016. Exploring the role
of contracts to support the emergence of self-organized industrial symbiosis
networks: an agent-based simulation study. Journal of Cleaner Production 112
(2016), 4353–4366.
[2] Huib Aldewereld, Frank Dignum, Andrés García-Camino, Pablo Noriega, Juan An-
tonio Rodríguez-Aguilar, and Carles Sierra. 2007. Operationalisation of norms
for electronic institutions. In Coordination, Organizations, Institutions, and Norms
in Agent Systems II. Springer, 163–176.
[3] Giulia Andrighetto, Guido Governatori, Pablo Noriega, and Leendert WN van der
Torre. 2013. Normative multi-agent systems. Vol. 4. Schloss Dagstuhl-Leibniz-
Zentrum fuer Informatik.
[4] Olivier Boissier and M Birna Van Riemsdijk. 2013. Organisational reasoning
agents. In Agreement Technologies. Springer, 309–320.
[5] Marian R Chertow. 2000. Industrial symbiosis: literature and taxonomy. Annual
review of energy and the environment 25, 1 (2000), 313–337.
[6] Mehdi Dastani, Leendert van der Torre, and Neil Yorke-Smith. 2016. Commit-
ments and interaction norms in organisations. Autonomous Agents and Multi-
Agent Systems (2016), 1–43.
[7] Davide Grossi, Luca Tummolini, and Paolo Turrini. 2013. Norms in game theory.
In Agreement Technologies. Springer, 191–197.
[8] Samuel Ieong and Yoav Shoham. 2005. Marginal contribution nets: a compact
representation scheme for coalitional games. In Proceedings of the 6th ACM
conference on Electronic commerce. ACM, 193–202.
[9] Souhila Kaci and Leendert van der Torre. 2008. Preference-based argumentation:
Arguments supporting multiple values. International Journal of Approximate
Reasoning 48, 3 (2008), 730–751.
[10] Julien Lesca, Patrice Perny, and Makoto Yokoo. 2017. Coalition Structure Gener-
ation and CS-core: Results on the Tractability Frontier for games represented by
MC-nets. In Proceedings of the 16th Conference on Autonomous Agents and Multi-
Agent Systems. International Foundation for Autonomous Agents and Multiagent
Systems, 308–316.
[11] D Rachel Lombardi and Peter Laybourn. 2012. Redefining industrial symbiosis.
Journal of Industrial Ecology 16, 1 (2012), 28–37.
[12] Andreu Mas-Colell, Michael Dennis Whinston, Jerry R Green, et al. 1995. Mi-
croeconomic theory. Vol. 1. Oxford university press New York.
[13] Sara Mehryar, Richard Sliuzas, Ali Sharifi, Diana Reckien, and Martin van Maarse-
veen. 2017. A structured participatory method to support policy option analysis
in a social-ecological system. Journal of Environmental Management 197 (2017),
360–372.
[14] Reshef Meir, Jeffrey S. Rosenschein, and Enrico Malizia. 2011. Subsidies, Stability,
and Restricted Cooperation in Coalitional Games. In IJCAI 2011, Proceedings of the
22nd International Joint Conference on Artificial Intelligence, Barcelona, Catalonia,
Spain, July 16-22, 2011. 301–306.
[15] Sanjay Modgil and Henry Prakken. 2013. A general account of argumentation
with preferences. Artificial Intelligence 195 (2013), 361–397.
[16] Martin J Osborne and Ariel Rubinstein. 1994. A course in game theory. MIT press.
[17] Yoav Shoham and Moshe Tennenholtz. 1995. On social laws for artificial agent
societies: off-line design. Artificial intelligence 73, 1-2 (1995), 231–252.
[18] Devrim Murat Yazan, Vincenzo Alessio Romano, and Vito Albino. 2016. The
design of industrial symbiosis: An input–output approach. Journal of cleaner
production 129 (2016), 537–547.
[19] Vahid Yazdanpanah and Devrim Murat Yazan. 2017. Industrial Symbiotic Rela-
tions as Cooperative Games. In Proceedings of the 7th International Conference on
Industrial Engineering and Systems Management (IESM-2017). 455–460.
[20] Vahid Yazdanpanah, Devrim Murat Yazan, and W Henk M Zijm. 2016. Normative
Industrial Symbiotic Networks: A Position Paper. In Multi-Agent Systems and
Agreement Technologies. Springer, 314–321.
[21] Yair Zick, Maria Polukarov, and Nicholas R. Jennings. 2013. Taxation and stability
in cooperative games. In International conference on Autonomous Agents and
Multi-Agent Systems, AAMAS '13, Saint Paul, MN, USA, May 6-10, 2013. 523–530.
|
1912.04800 | 1 | 1912 | 2019-12-10T16:34:38 | Approximate Strategy-Proofness in Large, Two-Sided Matching Markets | [
"cs.MA"
] | An approximation of strategy-proofness in large, two-sided matching markets is highly evident. Through simulations, one can observe that the percentage of agents with useful deviations decreases as the market size grows. Furthermore, there seems to be a strong connection between the length of preference order lists, the correlation of agent preferences, and the approximation of strategy proofness. Interestingly, approximate strategy proofness is reached easier with shorter length of preference orders and higher preference correlation. These findings justify the use of the deferred acceptance algorithm in large two-sided matching markets despite it not being strategy-proof. | cs.MA | cs |
Approximate Strategy-Proofness in Large, Two-Sided Matching
Markets
An empirical study of how market size, preference order length and preference
correlation impacts market participants' incentives.
Lars L. Ankile, Kjartan Krange, Yuto Yagi
CS 136 Economics and Computation Final Paper
School of Engineering and Applied Sciences
Harvard University
Cambridge, MA
December 11, 2019
Approximate Strategy-Proofness in Large,
Two-Sided Matching Markets
Lars L. Ankile, Kjartan Krange, Yuto Yagi
December 11, 2019
Abstract
An approximation of strategy-proofness in large, two-sided matching
markets is highly evident. Through simulations one can observe that the
percentage of agents with useful deviations decreases as the market size
grows. Furthermore, there seems to be a strong connection between the
length of preference order lists, the correlation of agent preferences, and
the approximation of strategy proofness. Interestingly, approximate strat-
egy proofness is reached easier with shorter length of preference orders and
higher preference correlation. These findings justify the use of deferred
acceptance algorithm in large two sided matching markets despite it not
being strategy-proof.
1
Introduction
Approximate strategy-proofness is
observed in large matching markets
where the percentage of agents who
can usefully deviate is small. A re-
search paper exploring this concept is
Roth and Elliot, 1990[1].
It presents
how the labor-market matching be-
tween newly graduated physicians and
hospitals benefit from the discovery
that "Opportunities for strategic ma-
nipulations, are surprisingly small" i.e.
they show how approximate strategy-
proofness makes their matching algo-
rithm more robust in a large market.
Clearly, the implications of the prop-
erty could in many cases be important
when deciding how to implement large
matching markets. For example, by
giving market designers the opportu-
nity to trust the agents to report truth-
fully even though the DA mechanism is
not technically strategy-proof.
In this paper we are running mul-
tiple simulations to analyze how and
when one can approximate strategy-
proofness. Furthermore, we are aim-
ing at simulating different amounts of
correlation between the agents' pref-
erences in the market. In researching
these properties we hope to get new in-
sights into when approximate strategy-
proofness can reliably mitigate gam-
ing of the system, and for what mar-
ket size, length of preference lists, and
preference correlation it appears effec-
tive to a lesser extent.
1
Our hypothesis is that we will wit-
ness a near strategy-proofness for suf-
ficiently large markets and that the
length of the reported preference or-
dering and the amount of correlation
will have an effect on the approxima-
tion to strategy-proofness. All source
code can be found on the GitHub-
repo here.
2 Theory
Our study is based on several pieces
of existing research. The following is a
list of theorems with a description of
how they apply to our paper:
• Simplicity,
Dynamic
Stabil-
ity, and Robustness are some
of many wanted properties of
strategy-proofness
(Lubin and
Parkes, 2012[2]) Thus, we want
to have a strategy-proof
mechanism design
• Theorem 12.5 Truthful report-
ing is a dominant strategy for
students in the student-proposing
DA mechanism.
(Parkes and
Seuken, 2019[3])
• Theorem 12.7 No mechanism for
two-sided matching is both stable
and strategy-proof
(Parkes and
Seuken, 2019[3])
Knowing that deferred accep-
tance returns a stable matching,
and that the proposing side has
truthful reporting as its dom-
inant strategy, we can deduce
that in general there exists use-
ful deviations, and that these will
only exits for the agents that are
being proposed to.
2
Thus, we only have to look
for deviations among agents
being proposed to.
• In simple markets, [...] all suc-
cessful manipulations can [also]
be accomplished by truncations
(Roth and Peranson, 1999[1])
in
This result is highly useful
the implementation of our sim-
ulation.
Instead of looking for
k! possible permutations of pref-
erences, for preference orders of
length k, it is sufficient to trun-
cate the preference order once
at every position for each par-
ticipant on the receiving side.
This makes an intractable prob-
lem tractable. Thus, we only
need to check for truncation
deviations.
the
approximated
mentioned
The
theorems
desirable.
strategy-proofness
search
Furthermore,
in
strategy-
for
proofness, we can conclude that
searching for deviations only in
the form of truncation on the re-
ceiving side of the market can be
done without loss of generality.
This reduces the computational
complexity immensely.
Figure 1: Plots of the three probability distributions we used for modeling dif-
ferent levels of correlation between agents' preferences. Left is the uncorrelated
distribution, middle is the moderate, and right is the high correlation.
3 Method
In short, our simulation generates
matching markets for a given amount
of preference correlation and length of
preference orders, gradually increasing
the market size, and for each market
size, counts the number of agents who
3.1 Preference Generation
For our results to be valid we need
to realistically model agent preferences
in a way that would resemble the dis-
tribution over preferences in a real
market. An important aspect we need
to capture is that some market partic-
ipants tend to be popular among all
agents on the other side, e.g. some men
are usually attractive in most women's
eyes, which makes them more likely to
appear high on women's hypothetical
preference orderings. We found that a
distribution defined by the inverse ex-
ponential function would give us a fit-
ting representation. We model three
specific market scenarios with differ-
can usefully deviate.
Our algorithm can be divided into
three parts: (i) Preference Generation,
(ii) Deferred Acceptance, (iii) Count-
ing Deviations.
ing level of correlation between agents'
preferences: (1) almost no correlation,
(2) moderate correlation, and (3) high
correlation. In figure 1, one can see a
plot of the three different distributions
used, respectively.
Based on these distributions we
first generated preference orderings for
one side of the market, where they
were given a k length preference list
drawn at random according to the dis-
tributions described above. Since it
e.g. does not makes sense for a hos-
pital to have preferences over students
they have not interviewed, we chose to
base the proposal-receiving side's pref-
3
erences on the preferences of the pro-
posers. This will also make sure we
have as high match ratio as possible.
We allowed the proposal receiving side
to have a variable length of preferences.
E.g. if every male was allowed to have
preference orderings of length 10, and
3.2 Deferred Acceptance
In order to find stable matches,
we implemented the deferred accep-
tance algorithm which is known to pro-
duce stable matches. This algorithm
takes incomplete preference orders and
interprets it as the agent preferring
to be unmatched to being matched
agents not on their list. E.g. suppose
m's preference over f1, f2, f3 given as
there are 5 males who have one specific
female among their preferences, that
female will have a preference ordering
of length 5. With everyone's prefer-
ences in order we are ready to proceed
to the actual matching.
f1 (cid:31)m f3. Then this will be inter-
preted as f1 (cid:31)m f3 (cid:31)m ∅ (cid:31)m f2.
The following is the pseudo code of
our deferred acceptance algorithm. We
had to implement this algorithm from
the bottom up,
instead of using li-
braries, because most existing code did
not handle truncations and incomplete
preference orders.
Algorithm 1 DefferedAcceptance
Initialize p in proposers and r in recipients unmatched
while exists p that is unmatched and has someone yet to propose to do
for p in unmatched proposers do
r = most preferred r which p has not proposed yet
if p is not in r's preference list then
break
else if r is unmatched then
(p, r) become matched
else
if r prefers p to current match then
(p, r) become matched
r's previous match p' becomes unmatched
else
no change
end if
end if
end for
end while
4
3.3 Counting Deviations
This paper is trying to find a con-
nection between D(n), i.e. the number
of agents with useful deviations, and
n =, the number of agents on one side
of the market. Subsection 3.3 describes
the calculation of D(n).
As presented in section 2 Theory,
for the stable matchings produced by
a given mechanism, misreports in the
form of truncation deviation by the re-
ceiving side of the market is the only
possibility for useful deviations. For
simplicity, we call a recipient female
in this subsection (without intention
of being hetero-normative). When our
simulation returns a stable matching
based truthful reports (given a prefer-
ence generation dependent on ρ, k, and
n) we analyze this matching in the fol-
lowing way:
1. While there are females not ana-
4 Results
lyzed, select a female.
2. Given a female, truncate her last
preference order entry.
3. Based on this new preference or-
der, run deferred acceptance and
check if this female benefits.
4. Repeat step 2 and 3 until devi-
ation is found (if so set D(n) =
D(n) + 1) or preference list is of
length one (in this case mark fe-
male as analysed and go to step
1).
We are counting how many fe-
males have at least one useful mis-
report. The D(n) returned will be
logged with the corresponding n =
market size, k = preference length and
ρ = preference correlation. Giving us
our final results.
In this section we briefly present the results from our extensive simulations in
the form of a series of plots. This required 300+ CPU-hours.
Figure 2: Simulations for ρ = 0.05, 1.0, 3.0 (from left to right) each with k =
10, 15, 20, 40.
5
Figure 3: Plot for 156 data points to the left and 2112 data points to the right.
With increasing data density, curves converge towards clearer separation.
Figure 4: Plot for ρ = 1.0 and for multiple values of k. Observe that longer
preference orderings tend to make the opportunities for strategic behavior larger.
6
Figure 5: Plot for k = 20 for a range of different ρ. Observe that correlation
between agents' preferences tend to increase the amount of opportunities for
strategic behavior.
Figure 6: Plot for k = 20, ρ = 1, for markets of size up to n = 1000. We assume
that other values for k and ρ will produce proportional results in the limit.
7
5 Discussion
5.1 The General Findings
The first
correlations,
There are two main types of sim-
ulations performed.
is
for four different preference lengths,
k = 10, 15, 20, 40, and three amounts
of preference
ρ =
0.05, 1.0, 3.0, with a market size n,
running from 5 to 400 (figure 2). This
illustrates the effect of different k
and ρ on the approximate strategy-
proofness. The second type shows
how, as the market grows even larger
(up to n = 1000), D(n)/n converges
to 0 (figure 6). We show this while
keeping k and ρ constant at 20 and
1.0, respectively. Because we lack the
resources to run the same simulation
for the other values of k and ρ, we
chose k = 20 and ρ = 1.0 as repre-
sentative values. We assume that the
other curves would also tend towards
0.
As k and n grows, the amount of
computation required grows relatively
quickly. Our deferred acceptance im-
plementation has a running time of
O(n2). We run the DA for n recipients
with an average preference length k,
leaving our total simulation at an av-
erage running time of O(n3k). An un-
realistic, but definite worst case would
be O(n4), for k = n.
As a general finding, the results
show that the market quickly ap-
proaches strategy-proofness when the
markets grow from a size of n = 10 up
to around n = 100. In fact, in this seg-
ment of n, given a high ρ the percent-
age of agents with useful deviations
goes from around 16% to a mere 0.01%.
As the market grows even larger, the
percentage of agents with useful devia-
tions approaches zero, but the speed of
convergence slows down (see e.g. figure
5).
Our last figure shows how with
an increasing number of simulation
runs, the data plot gradually becomes
smoother with less random spikes in
D(n)/n. Since the generation of the
preference orders are partially based
on randomness, having the data based
on multiple simulations give us model
results gradually approaching an aver-
age value. The random based algo-
rithm design, makes our results more
easily applicable to the real world as
they are based on a huge set of dif-
ferent kind of matching market situ-
ations. The plot gradually converges
towards a monotone decreasing func-
tion. This leads us to conclude that
approximate strategy proofness gener-
ally exists in large markets. We can-
not, however, always rule out the exis-
tence of the odd possibility for devia-
tion for some agent.
5.2 Effect of Varying Preference-Order Length
In figure 2 we see how the three
plots: k = 10 in blue, k = 15 in or-
ange, and k = 20 in green. The im-
portant thing to note is that in general
the plots of D(n)/n for a given n in-
creases with k. Intuitively this makes
sense as a larger k quite simply gives
more preferences to truncate away, i.e.
more possible deviations. For a mar-
ket designer, the general advice would
8
be that a smaller preference reporting
allowed restricts the amount of benefi-
cial misreports available for all agents
in total.
5.3 Effect of Varying the Preference Correlation
We note that an increasing correla-
tion, ρ (ρ=0.05, ρ=1 and ρ=3), leads
to smaller ratio of agents who can use-
fully deviate (D(n)/n). This allows
market designers with a knowledge of
low correlation to expand the market
to secure approximate strategy proof-
ness. As for the reason why stronger
correlation leads to less possibility for
useful deviation, having more agents
interested in the same matches makes
the rejection chain shorter, thus mak-
ing it unlikely to return to the deviator.
6 Applications and Improvements
As discussed, a matching market
designer has to choose between a stable
and a strategy proof matching. Parkes
and Seuken (2019[3]) argue that un-
stable markets have a tendency to un-
ravel (possibly with catastrophic con-
sequences) which makes stability an ar-
guably more important property than
strategy-proofness. However, as shown
in this paper, large market can gain
approximate strategy-proofness while
maintaining the stability of the out-
come.
In DA, the proposal-receiving side
gets their least preferred, achievable
match.
In some cases it might actu-
ally be more important for organiz-
ing party to get the better matching.
An example could be a market try-
ing to match medical students to res-
idencies at hospitals. It is potentially
better if the hospital-optimal matching
was achieved instead of the student-
optimal one. This is because the needs
for students with specific talent and
knowledge at the hospitals might be
more important for the society at large
the individual wants of the students.
The results in this paper are likely
applicable to the real world. We have
seen that in two-sided matching mar-
kets it quickly becomes difficult to find
useful deviations for the receiving side.
This effect is particularly evident when
we have relatively short preference or-
derings and a high correlation between
the preferences of the agents in the
market. Both of these assumptions we
think are reasonable in real-world ap-
plications.
As a familiar instance, Norway has
a coordinated admissions system for al-
most every institution of higher edu-
cation, Samordna opptak. There, stu-
dents are ranked according to their
high school grades, and students are
allowed to create a strict preference or-
der over at most 10 schools. Here we
see that every agent on the propos-
ing side has a preference ordering of
length at most 10, while every recipient
has a preference ordering over all stu-
dents who have them on their prefer-
ence ordering. Furthermore, we know
from open statistics that a select few
schools are extremely popular, while
many schools are not as popular and
struggle to fill up their classes. This
9
effect is probably common, and is well
captured by our medium or high cor-
relation distribution. These observa-
tions, together with the fact that the
market is very large (tens of thousand
of agents), make us believe that this
market would potentially reach strong
approximate strategy-proofness.
As a last point, we want to ar-
gue that even though our simulation is
implemented with an even-sized pool
of proposers and proposed agents, a
higher ρ models a matching market
with a different sized sides. When ρ is
large almost all proposed agents are in-
terested in a select few proposers. This
is close to a real world situation where
there are few proposers compared to
the proposed side. As such, an increas-
ing ρ could be used to model a situa-
tion with an increasingly large amount
of agents on the proposed side of the
market relative to the proposers.
7 Potential Issues and Next Steps
The following are some points of is-
sues and possible future implementa-
tions:
• Our simulation assumes an equal
number of proposer and recipi-
ents, and have not modeled the
case where there are more pro-
posers than recipients.
• If we were to use our code to
model more varied markets, we
would gain a lot from having
a more highly optimized algo-
rithm.
• We have modelled one-to-one
two sided matching market, but
8 Conclusion
there are many other form of
matching that might
require
analysis. They include match-
ing with couples, one-to-many
matchings, and many-to-many
matchings.
Next step would
be to expand our model to in-
clude these different matching
markets.
• This paper has given a empiri-
cal proof on the existence of ap-
proximate strategy proofness in
large two sided matching mar-
kets. The next step would be to
provide a theoretical proof as to
how and why this property arise
in large markets.
We have found approximate strategy-proofness to be prevalent in large mar-
kets. Our models show that the percentage of agents with useful deviations
decreases most rapidly between a market size of 10 and 100. Approximate
strategy-proofness decreases with the length of preference orders and increases
with preference correlation. Lastly, there are a number of real world applica-
tions, like the Norwegian university-application process, where these findings
are of interest. There still exists open issues that we could address in future
research.
10
References
[1] Alvin E. Roth and Elliott Peranson. The Redesign of the Matching Market
for American Physicians: Some Engineering Aspects of Economic Design.
https://web.stanford.edu/~alroth/papers/rothperansonaer.PDF
[2] Benjamin Lubin and David C. Parkes. Approximate Strategyproofness.
https://dash.harvard.edu/bitstream/handle/1/11879945/Parkes_
ApproximateStrategyproofness.pdf
[3] David C. Parkes and Sven Seuken Algorithmic Economics: A Design Ap-
proach. Harvard University/University of Zurich, 2019.
[4] Alvin E. Roth. The economics of matching: Stability and incentives. Math-
ematics of Operations Research, pages 617 -- 628, 1982.
11
|
1805.07830 | 4 | 1805 | 2018-08-31T18:36:15 | Learning to Teach in Cooperative Multiagent Reinforcement Learning | [
"cs.MA",
"cs.AI"
] | Collective human knowledge has clearly benefited from the fact that innovations by individuals are taught to others through communication. Similar to human social groups, agents in distributed learning systems would likely benefit from communication to share knowledge and teach skills. The problem of teaching to improve agent learning has been investigated by prior works, but these approaches make assumptions that prevent application of teaching to general multiagent problems, or require domain expertise for problems they can apply to. This learning to teach problem has inherent complexities related to measuring long-term impacts of teaching that compound the standard multiagent coordination challenges. In contrast to existing works, this paper presents the first general framework and algorithm for intelligent agents to learn to teach in a multiagent environment. Our algorithm, Learning to Coordinate and Teach Reinforcement (LeCTR), addresses peer-to-peer teaching in cooperative multiagent reinforcement learning. Each agent in our approach learns both when and what to advise, then uses the received advice to improve local learning. Importantly, these roles are not fixed; these agents learn to assume the role of student and/or teacher at the appropriate moments, requesting and providing advice in order to improve teamwide performance and learning. Empirical comparisons against state-of-the-art teaching methods show that our teaching agents not only learn significantly faster, but also learn to coordinate in tasks where existing methods fail. | cs.MA | cs |
Learning to Teach in Cooperative Multiagent Reinforcement Learning
Shayegan Omidshafiei1,2
[email protected]
Dong-Ki Kim1,2
[email protected]
Miao Liu2,3
[email protected]
Matthew Riemer2,3
[email protected]
Christopher Amato4
[email protected]
Murray Campbell2,3
[email protected]
Gerald Tesauro2,3
[email protected]
Jonathan P. How1,2
[email protected]
1LIDS, MIT
2MIT-IBM Watson AI Lab
3IBM Research
4CCIS, Northeastern University
Abstract
Collective human knowledge has clearly benefited from the
fact that innovations by individuals are taught to others through
communication. Similar to human social groups, agents in
distributed learning systems would likely benefit from com-
munication to share knowledge and teach skills. The problem
of teaching to improve agent learning has been investigated
by prior works, but these approaches make assumptions that
prevent application of teaching to general multiagent prob-
lems, or require domain expertise for problems they can apply
to. This learning to teach problem has inherent complexities
related to measuring long-term impacts of teaching that com-
pound the standard multiagent coordination challenges. In
contrast to existing works, this paper presents the first gen-
eral framework and algorithm for intelligent agents to learn to
teach in a multiagent environment. Our algorithm, Learning
to Coordinate and Teach Reinforcement (LeCTR), addresses
peer-to-peer teaching in cooperative multiagent reinforcement
learning. Each agent in our approach learns both when and
what to advise, then uses the received advice to improve local
learning. Importantly, these roles are not fixed; these agents
learn to assume the role of student and/or teacher at the ap-
propriate moments, requesting and providing advice in order
to improve teamwide performance and learning. Empirical
comparisons against state-of-the-art teaching methods show
that our teaching agents not only learn significantly faster, but
also learn to coordinate in tasks where existing methods fail.
Introduction
In social settings, innovations by individuals are taught to
others in the population through communication channels
(Rogers 2010), which not only improves final performance,
but also the effectiveness of the entire learning process (i.e.,
rate of learning). There exist analogous settings where learn-
ing agents interact and adapt behaviors while interacting in
a shared environment (e.g., autonomous cars and assistive
robots). While any given agent may not be an expert during
learning, it may have local knowledge that teammates may be
unaware of. Similar to human social groups, these learning
agents would likely benefit from communication to share
knowledge and teach skills, thereby improving the effective-
ness of system-wide learning. It is also desirable for agents in
such systems to learn to teach one another, rather than rely on
Copyright c(cid:13) 2019, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
hand-crafted teaching heuristics created by domain experts.
The benefit of learned peer-to-peer teaching is that it can
accelerate learning even without relying on the existence of
"all-knowing" teachers. Despite these potential advantages, no
algorithms exist for learning to teach in multiagent systems.
This paper targets the learning to teach problem in the
context of cooperative Multiagent Reinforcement Learning
(MARL). Cooperative MARL is a standard framework for
settings where agents learn to coordinate in a shared environ-
ment. Recent works in cooperative MARL have shown final
task performance can be improved by introducing inter-agent
communication mechanisms (Sukhbaatar, Fergus, and oth-
ers 2016; Foerster et al. 2016; Lowe et al. 2017). Agents in
these works, however, merely communicate to coordinate in
the given task, not to improve overall learning by teaching
one another. By contrast, this paper targets a new multiagent
paradigm in which agents learn to teach by communicat-
ing action advice, thereby improving final performance and
accelerating teamwide learning.
The learning to teach in MARL problem has unique inher-
ent complexities that compound the delayed reward, credit
assignment, and partial observability issues found in gen-
eral multiagent problems (Oliehoek and Amato 2016). As
such, there are several key issues that must be addressed.
First, agents must learn when to teach, what to teach, and
how to learn from what is being taught. Second, despite
coordinating in a shared environment, agents may be inde-
pendent/decentralized learners with privacy constraints (e.g.,
robots from distinct corporations that cannot share full poli-
cies), and so must learn how to teach under these constraints.
A third issue is that agents must estimate the impact of each
piece of advice on their teammate's learning progress. De-
lays in the accumulation of knowledge make this credit as-
signment problem difficult, even in supervised/unsupervised
learning (Graves et al. 2017). Nonstationarities due to agent
interactions and the temporally-extended nature of MARL
compound these difficulties in our setting. These issues are
unique to our learning to teach setting and remain largely
unaddressed in the literature, despite being of practical impor-
tance for future decision-making systems. One of the main
reasons for the lack of progress addressing these inherent
challenges is the significant increase in the computational
complexity of this new teaching/learning paradigm compared
to multiagent problems that have previously been considered.
Our paper targets the problem of learning to teach in a
multiagent team, which has not been considered before. Each
agent in our approach learns both when and what to ad-
vise, then uses the received advice to improve local learn-
ing. Importantly, these roles are not fixed (see Fig. 1); these
agents learn to assume the role of student and/or teacher
at appropriate moments, requesting and providing advice
to improve teamwide performance and learning. In contrast
to prior works, our algorithm supports teaching of hetero-
geneous teammates and applies to settings where advice ex-
change incurs a communication cost. Comparisons conducted
against state-of-the-art teaching methods show that our teach-
ing agents not only learn significantly faster, but also learn to
coordinate in tasks where existing methods fail.
Background: Cooperative MARL
t = (oi
Our work targets cooperative MARL, where agents exe-
cute actions that jointly affect the environment, then receive
feedback via local observations and a shared reward. This
setting is formalized as a Decentralized Partially Observ-
able Markov Decision Process (Dec-POMDP), defined as
(cid:104)I,S,A,T ,R, Ω,O, γ(cid:105) (Oliehoek and Amato 2016); I is
the set of n agents, S is the state space, A = ×iAi is the joint
action space, and Ω = ×iΩi is the joint observation space.1
Joint action a = (cid:104)a1, . . . , an(cid:105) causes state s ∈ S to transition
to s(cid:48) ∈ S with probability P (s(cid:48)s, a) = T (s, a, s(cid:48)). At each
timestep t, joint observation o = (cid:104)o1, . . . , on(cid:105) is observed
with probability P (os(cid:48), a) = O(o, s(cid:48), a). Given its obser-
vation history, hi
t), agent i executes actions
1, . . . , oi
dictated by its policy ai = πi(hi
t). The joint policy is denoted
by π = (cid:104)π1, . . . , πn(cid:105) and parameterized by θ. It may some-
times be desirable to use a recurrent policy representation
(e.g., recurrent neural network) to compute an internal state
ht that compresses the observation history, or to explicitly
compute a belief state (probability distribution over states);
with abuse of notation, we use ht to refer to all such variations
of internal states/observation histories. At each timestep, the
team receives reward rt = R(st, at), with the objective be-
t γtrts0 = s], given
discount factor γ ∈ [0, 1). Let action-value Qi(oi, ai; hi) de-
note agent i's expected value for executing action ai given
a new local observation oi and internal state hi, and using
its policy thereafter. We denote by (cid:126)Q(oi; hi) the vector of
action-values (for all actions) given new observation oi.
ing to maximize value, V (s; θ) = E[(cid:80)
Teaching in Cooperative MARL
This work explores multiagent teaching in a setting where
no agent is necessarily an all-knowing expert. This section
provides a high-level overview of the motivating scenario.
Consider the cooperative MARL setting in Fig. 1, where
agents i and j learn a joint task (i.e., a Dec-POMDP). In each
learning iteration, these agents interact with the environment
and collect data used by their respective learning algorithms,
Li and Lj, to update their policy parameters, θi and θj. This
is the standard cooperative MARL problem, which we here-
after refer to as PTask: the task-level learning problem. For
1Superscript i denotes parameters for the i-th agent. Refer to the
supplementary material for a notation list.
Figure 1: Overview of teaching via action advising in MARL.
Each agent learns to execute the task using task-level policy
π, to request advice using learned student policy(cid:101)πS, and to
respond with action advice using learned teacher policy(cid:101)πT .
help, agent j advises action(cid:101)aj, which the student executes
Each agent can assume a student and/or teacher role at any
time. In this example, agent i uses its student policy to request
instead of its originally-intended action ai. By learning to
transform the local knowledge captured in task-level policies
into action advice, the agents can help one another learn.
example, task-level policy πi is the policy agent i learns and
uses to execute actions in the task. Thus, task-level policies
summarize each agent's learned behavioral knowledge.
During task-level learning, it is unlikely for any agent to be
an expert. However, each agent may have unique experiences,
skill sets, or local knowledge of how to learn effectively in
the task. Throughout the learning process, it would likely
be useful for agents to advise one another using this local
knowledge, in order to improve final performance and accel-
erate teamwide learning. Moreover, it would be desirable for
agents to learn when and what to advise, rather than rely on
hand-crafted and domain-specific advising heuristics. Finally,
following advising, agents should ideally have learned effec-
tive task-level policies that no longer rely on teammate advice
at every timestep. We refer to this new problem, which in-
volves agents learning to advise one another to improve joint
task-level learning, as (cid:101)PAdvise: the advising-level problem.
The advising mechanism used in this paper is action ad-
vising, where agents suggest actions to one another. By
learning to appropriately transform local knowledge (i.e.,
task-level policies) into action advice, teachers can affect
students' experiences and their resulting task-level policy
updates. Action advising makes few assumptions, in that
learners need only use task-level algorithms (cid:104)Li, Lj(cid:105) that
support off-policy exploration (enabling execution of action
advice for policy updates), and that they receive advising-
level observations summarizing teammates' learning progress
(enabling learning of when/what to advise). Action advising
has a good empirical track record (Torrey and Taylor 2013;
Taylor et al. 2014; Fachantidis, Taylor, and Vlahavas 2017;
da Silva, Glatt, and Costa 2017). However, existing frame-
works have key limitations: the majority are designed for
single-agent RL and do not consider multiagent learning;
their teachers always advise optimal actions to students, mak-
Task-level policy 𝜋𝑖 Teacher policy 𝜋 𝑇𝑖 Request advice Task-level policy 𝜋𝑗 Student policy 𝜋 𝑆𝑗 Teacher policy 𝜋 𝑇𝑗 Advise action 𝑎 𝑗 Original action 𝑎𝑖 Advised action 𝑎 𝑗 Student policy 𝜋 𝑆𝑖 ing decisions about when (not what) to teach; they also use
heuristics for advising, rather than training teachers by mea-
suring student learning progress. By contrast, agents in our
paper learn to interchangeably assume the role of a student
(advice requester) and/or teacher (advice responder), denoted
S and T , respectively. Each agent learns task-level policy π
used to actually perform the task, student policy(cid:101)πS used to
(cid:101)πT used to advise a teammate during task-level learning.2
request advice during task-level learning, and teacher policy
Before detailing the algorithm, let us first illustrate the
multiagent interactions in this action advising scenario. Con-
sider again Fig. 1, where agents are learning to execute a task
(i.e., solving PTask) while advising one another. While each
agent in our framework can assume a student and/or teacher
role at any time, Fig. 1 visualizes the case where agent i
is the student and agent j the teacher. At a given task-level
learning timestep, agent i's task-level policy πi outputs an
action ('original action ai' in Fig. 1). However, as the agents
are still learning to solve PTask, agent i may prefer to execute
an action that maximizes local learning. Thus, agent i uses
S to decide whether to ask teammate j for
advice. If this advice request is made, teammate j checks its
T and task-level policy πj to decide whether
to respond with action advice. Given a response, agent i then
its student policy(cid:101)πi
teacher policy(cid:101)πj
executes the advised action ((cid:101)aj in Fig. 1) as opposed to its
originally-intended action (ai in Fig. 1). This results in a local
experience that agent i uses to update its task-level policy. A
reciprocal process occurs when the agents' roles are reversed.
The benefit of advising is that agents can learn to use local
knowledge to improve teamwide learning.
Similar to recent works that model the multiagent learning
process (Hadfield-Menell et al. 2016; Foerster et al. 2018),
we focus on the pairwise (two agent) case, targeting the issues
of when/what to advise, then detail extensions to n agents.
Even in the pairwise case, there exist issues unique to our
learning to teach paradigm. First, note that the objectives of
PTask and (cid:101)PAdvise are distinct. Task-level problem, PTask, has
((cid:101)PAdvise), however, is a higher-level problem, where agents
them. However, PTask and (cid:101)PAdvise are also coupled, as advis-
a standard MARL objective of agents learning to coordinate
to maximize final performance in the task. Learning to advise
learn to influence teammates' task-level learning by advising
ing influences the task-level policies learned. Agents in our
problem must learn to advise despite the nonstationarities
due to changing task-level policies, which are also a function
of algorithms (cid:104)Li, Lj(cid:105) and policy parameterizations (cid:104)θi, θj(cid:105).
Learning to teach is also distinct from prior works that
involve agents learning to communicate (Sukhbaatar, Fer-
gus, and others 2016; Foerster et al. 2016; Lowe et al. 2017).
These works focus on agents communicating in order to coor-
dinate in a task. By contrast, our problem focuses on agents
learning how advising affects the underlying task-level learn-
ing process, then using this knowledge to accelerate learning
even when agents are non-experts. Thus, the objectives of
communication-based multiagent papers are disparate from
ours, and the two approaches may even be combined.
2Tilde accents (e.g.,(cid:101)π) denote advising-level properties.
Figure 2: LeCTR consists of two iterated phases: task-level
learning (Phase I), and advising-level learning (Phase II). In
Phase II, advising policies are trained using rewards corre-
lated to task-level learning (see Table 1). Task-level, student,
and teacher policy colors above follows convention of Fig. 1.
LeCTR: Algorithm for Learning to
Coordinate and Teach Reinforcement
This section introduces our learning to teach approach, de-
tails how issues specific to our problem setting are resolved,
and summarizes overall training protocol. Pseudocode is pre-
sented in the supplementary material due to limited space.
lem (cid:101)PAdvise. The objective is to learn advising policies that
Overview Our algorithm, Learning to Coordinate and
Teach Reinforcement (LeCTR), solves advising-level prob-
augment agents' task-level algorithms (cid:104)Li, Lj(cid:105) to accelerate
solving of PTask. Our approach involves 2 phases (see Fig. 2):
• Phase I: agents learn PTask from scratch using blackbox
learning algorithms (cid:104)Li, Lj(cid:105) and latest advising policies.
• Phase II: advising policies are updated using advising-level
rewards correlated to teammates' task-level learning.
No restrictions are placed on agents' task-level algorithms
(i.e., they can be heterogeneous). Iteration of Phases I and II
enables training of increasingly capable advising policies.
S(cid:105) and teacher policies (cid:104)(cid:101)πi
S,(cid:101)πj
dent policies (cid:104)(cid:101)πi
T ,(cid:101)πj
Advising Policy Inputs & Outputs LeCTR learns stu-
T(cid:105) for
agents i and j, constituting a jointly-initiated advising ap-
proach that learns when to request advice and when/what to
advise. It is often infeasible to learn high-level policies that
directly map task-level policy parameters (cid:104)θi, θj(cid:105) (i.e., local
knowledge) to advising decisions: the agents may be indepen-
dent/decentralized learners and the cost of communicating
task-level policy parameters may be high; sharing policy pa-
rameters may be undesirable due to privacy concerns; and
learning advising policies over the task-level policy parame-
ter space may be infeasible (e.g., if the latter policies involve
millions of parameters). Instead, each LeCTR agent learns ad-
vising policies over advising-level observations(cid:101)o. As detailed
below, these observations are selected to provide informa-
tion about agents' task-level state and knowledge in a more
compact manner than full policy parameters (cid:104)θi, θj(cid:105).
Each LeCTR agent can be a student, teacher, or both si-
multaneously (i.e., request advice for its own state, while
advising a teammate in a different state). For clarity, we
Phase II: Train student/teacher policies, using task-level learning rate as reward Phase I: Train task-level policies from scratch, using latest student/teacher policies Value Agent 𝑗 Agent 𝑖 Value Task-level learning iteration Agent 𝑗 Agent 𝑖 Task-level learning iteration detail advising protocols when agents i and j are student
and teacher, respectively (see Fig. 1). LeCTR uses distinct
advising-level observations for student and teacher policies.
S for agent i decides when to request ad-
S = (cid:104)oi, (cid:126)Qi(oi; hi)(cid:105),
where oi and (cid:126)Qi(oi; hi) are the agent's task-level observation
S, agent i
observes a measure of its local task-level observation and
policy state. Thus, agent i's student-perspective action is
Student policy(cid:101)πi
vice using advising-level observation(cid:101)oi
and action-value vectors, respectively. Through(cid:101)oi
S =(cid:101)πi
(cid:101)ai
S((cid:101)oi
Similarly, agent j's teacher policy (cid:101)πj
S) ∈ {request advice, do not request advice}.
level observation(cid:101)oj
when/what to advise. (cid:101)oj
student's context (via (cid:126)Qj(oi; hi)). Using(cid:101)πj
space, Ai, or a special no-advice action(cid:101)a∅. Thus, the teacher-
T ) ∈ Ai ∪ {(cid:101)a∅}.
perspective action for agent j is(cid:101)aj
T =(cid:101)πj
T ((cid:101)oj
tion ai. However, given advice(cid:101)aj
Given no advice, student i executes originally-intended ac-
βi((cid:101)aj
T , student i executes action
T ), where βi(·) is a local behavioral policy not known
by j. The assumption of local behavioral policies increases
the generality of LeCTR, as students may locally transform
advised actions before execution.
T uses advising-
T = (cid:104)oi, (cid:126)Qi(oi; hi), (cid:126)Qj(oi; hi)(cid:105) to decide
T provides teacher agent j with a
measure of student i's task-level state/knowledge (via oi and
(cid:126)Qi(oi; hi)) and of its own task-level knowledge given the
T , teacher j de-
cides what to advise: either an action from student i's action
Following advice execution, agents collect task-level ex-
periences and update their respective task-level policies. A
training (cid:104)(cid:101)πi
T ,(cid:101)πj
key feature is that LeCTR agents learn what to advise by
T(cid:105), rather than always advising actions they
would have taken in students' states. These agents may learn
to advise exploratory actions or even decline to advise if they
estimate that such advice will improve teammate learning.
Rewarding Advising Policies Recall in Phase II of
LeCTR, advising policies are trained to maximize advising-
level rewards that should, ideally, reflect the objective of
accelerating task-level learning. Without loss of generality,
we focus again on the case where agents i and j assume stu-
dent and teacher roles, respectively, to detail these rewards.
T must coordi-
nate to help student i learn, they receive identical advising-
T . The remaining issue is to identify
Since student policy(cid:101)πi
S =(cid:101)rj
level rewards,(cid:101)ri
S and teacher policy(cid:101)πj
advising-level rewards that reflect learning progress.
Remark 1. Earning task-level rewards by executing ad-
vised actions may not imply actual learning. Thus, rewarding
advising-level policies with the task-level reward, r, received
after advice execution can lead to poor advising policies.
We evaluate many choices of advising-level rewards,
which are summarized and described in Table 1. The uni-
fying intuition is that each reward type corresponds to a
different measure of the advised agent's task-level learning,
which occurs after executing an advised action. Readers are
referred to the supplementary material for more details.
Note that at any time, task-level action ai executed by
or by a teammate j via advising. In Phase II, pair (cid:104)(cid:101)πi
S,(cid:101)πj
agent i may either be selected by its local task-level policy,
T(cid:105) is
rewarded only if advising occurs (with zero advising reward
otherwise). Analogous advising-level rewards apply for the
T . During
Phase II, we train all advising-level policies using a joint
reverse student-teacher pairing j-i, where(cid:101)rj
advising-level reward(cid:101)r =(cid:101)ri
T to induce cooperation.
S =(cid:101)ri
T +(cid:101)rj
Advising-level rewards are only used during advising-level
training, and are computed using either information already
available to agents or only require exchange of scalar values
(rather than full policy parameters). It is sometimes desirable
to consider advising under communication constraints, which
can be done by deducting a communication cost c from these
advising-level rewards for each piece of advice exchanged.
Training Protocol Recall LeCTR's two phases are it-
erated to enable training of increasingly capable advising
policies. In Phase I, task-level learning is conducted us-
ing agents' blackbox learning algorithms and latest advis-
ing policies. At the task-level, agents may be independent
learners with distinct algorithms. Advising policies are exe-
cuted in a decentralized fashion, but their training in Phase
II is centralized. Our advising policies are trained using the
multiagent actor-critic approach of Lowe et al. (2017). Let
joint advising-level observations, advising-level actions, and
T(cid:105),(cid:101)a = (cid:104)(cid:101)ai
noted by (cid:101)o = (cid:104)(cid:101)oi
S,(cid:101)ai
S,(cid:101)aj
T ,(cid:101)oj
S,(cid:101)oi
S,(cid:101)oj
T ,(cid:101)aj
advising-level policies (i.e., 'actors') be, respectively, de-
T(cid:105), with(cid:101)θ parameterizing(cid:101)π. To induce
(cid:101)π = (cid:104)(cid:101)πi
T ,(cid:101)πj
S,(cid:101)πi
S,(cid:101)πj
T(cid:105), and
(cid:101)π to learn to teach both agents i and j, we use a centralized
T . Critic (cid:101)Q((cid:101)o,(cid:101)a;(cid:101)θ) is trained by minimizing loss,
T +(cid:101)rj
(cid:101)r =(cid:101)ri
(cid:12)(cid:12)(cid:12)(cid:101)a(cid:48)=(cid:101)π((cid:101)o(cid:48))
L((cid:101)θ)=E(cid:101)o,(cid:101)a,(cid:101)r,(cid:101)o(cid:48)∼(cid:102)M[((cid:101)r+γ(cid:101)Q((cid:101)o
(cid:48);(cid:101)θ)−(cid:101)Q((cid:101)o,(cid:101)a;(cid:101)θ))2]
(cid:48),(cid:101)a
where(cid:101)a(cid:48) = (cid:101)π((cid:101)o(cid:48)) are next advising actions computed us-
ing the advising policies, and (cid:102)M denotes advising-level re-
rem (Sutton et al. 2000) is invoked on objective J((cid:101)θ) =
E[(cid:80)T
t=k(cid:101)γt−k(cid:101)rt] to update advising policies using gradients,
(cid:2)∇(cid:101)θ log(cid:101)π((cid:101)a(cid:101)o)(cid:101)Q((cid:101)o,(cid:101)a;(cid:101)θ)(cid:3)
∇(cid:101)θJ((cid:101)θ) = E(cid:101)o,(cid:101)a∼(cid:102)M
(cid:2)(cid:88)
ρ (cid:101)Q((cid:101)o,(cid:101)a;(cid:101)θ)(cid:3),
ρ ((cid:101)aα
log(cid:101)πα
ρ(cid:101)oα
ρ )∇(cid:101)aα
∇(cid:101)θα
=E(cid:101)o,(cid:101)a∼(cid:102)M
where(cid:101)πα
play buffer (Mnih et al. 2015). The policy gradient theo-
action-value function (i.e., 'critic') with advising-level reward
ρ is agent α's policy in role ρ.
α∈{i,j}
ρ∈{S,T}
During training, the advising feedback nonstationarities
mentioned earlier are handled as follows: in Phase I, task-
level policies are trained online (i.e., no replay memory is
used so impact of advice on task-level policies is immediately
observed by agents); in Phase II, centralized advising-level
learning reduces nonstationarities due to teammate learning,
and reservoir sampling is used to further reduce advising
reward nonstationarities (see supplementary material for de-
tails). Our overall approach stabilizes advising-level learning.
Advising n Agents In the n agent case, students must also
decide how to fuse advice from multiple teachers. This is a
complex problem requiring full investigation in future work;
feasible ideas include using majority voting for advice fusion
(as in da Silva, Glatt, and Costa (2017)), or asking a specific
agent for advice by learning a 'teacher score' modulated
based on teacher knowledge/previous teaching experiences.
ρ
,
(1)
(2)
Table 1: Summary of rewards used to train advising policies. Rewards shown are for the case where agent i is student and agent j
teacher (i.e., flip the indices for the reverse case). Each reward corresponds to a different measure of task-level learning after the
student executes an action advice and uses it to update its task-level policy. Refer to the supplementary material for more details.
Advising Reward Name
JVG: Joint Value Gain
QTR: Q-Teaching Reward
LG: Loss Gain
LGG: Loss Gradient Gain
TDG: TD Gain
VEG: Value Estimation Gain
Description
Task-level value V (s; θ) improvement after learning
Teacher's estimate of best vs. intended student action maxa QT (oi, a;hi) − QT (oi, ai;hi)
Student's task-level loss L(θi) reduction
Student's task-level policy gradient magnitude
Student's temporal difference (TD) error δi reduction
Student's value estimate V (θi) gain above threshold τ
V (s; θt+1) − V (s; θt)
Reward Value(cid:101)rj
t) − L(θi
L(θi
t+1)
∇θiL(θi)2
δi
t − δi
t+1
1( V (θi) > τ )
T =(cid:101)ri
S
2
Agent j
a1 a2
0
1
a2 0.1 0
i a1
t
n
e
g
A
(b) Counterexample showing poor advising
(a) Repeated game
payoffs. Each agent
has 2 actions (a1, a2).
reward choice(cid:101)rT = r. LeCTR Phase I and
using(cid:101)rT = r yields poor advising, as teachers learn to advise
Figure 3: Repeated game. (b) shows a counterexample where
II iterations are shown as background bands.
actions that maximize reward (left half, green), but do not
actually improve student task-level learning (right half, blue).
Evaluation
We conduct empirical evaluations on a sequence of increas-
ingly challenging domains involving two agents. In the
'Repeated' game domain, agents coordinate to maximize
the payoffs in Fig. 3a over 5 timesteps. In 'Hallway' (see
Fig. 4a), agents only observe their own positions and re-
ceive +1 reward if they reach opposite goal states; task-level
actions are 'move left/right', states are agents' joint grid po-
sitions. The higher-dimensional 'Room' game (see Fig. 5a)
has the same state/observation/reward structure, but 4 ac-
tions ('move up/right/down/left'). Recall student-perspective
advising-level actions are to 'ask' or 'not ask' for advice.
Teacher-perspective actions are to advise an action from the
teammate's task-level action space, or to decline advising.
For the Repeated, Hallway, and Room games, respectively,
each iteration of LeCTR Phase I consists of 50, 100, and
150 task-level learning iterations. Our task-level agents are
independent Q-learners with tabular policies for the Repeated
game and tile-coded policies (Sutton and Barto 1998) for the
other games. Advising policies are neural networks with in-
ternal rectified linear unit activations. Refer to the supplemen-
tary material for hyperparameters. The advising-level learn-
ing nature of our problem makes these domains challenging,
despite their visual simplicity; their complexity is compara-
ble to domains tested in recent MARL works that learn over
multiagent learning processes (Foerster et al. 2018), which
also consider two agent repeated/gridworld games.
Counterexample demonstrating Remark 1 Fig. 3b
(a) Hallway domain overview.
(b) Simult. learning & teaching
Figure 4: Hallway game. (a) Agents receive +1 reward by
navigating to opposite states in 17-grid hallway. (b) LeCTR
accelerates learning & teaching compared to no-teaching.
(a) Room domain overview.
(b) Teaching heterogeneous agents.
Figure 5: Room game. (a) Agents receive +1 reward by navi-
gating to opposite goals in 17×5 grid. (b) LeCTR outperforms
prior approaches when agents are heterogeneous.
shows results given a poor choice of advising-level reward,
(cid:101)rT = r, in the Repeated game. The left plot (in green) shows
task-level return received due to both local policy actions
and advised actions, which increases as teachers learn. How-
ever, in the right plot (blue) we evaluate how well task-level
policies perform by themselves, after they have been trained
using the final advising-level policies. The poor performance
of the resulting task-level policies indicates that advising poli-
cies learned to maximize their own rewards(cid:101)rT = r by always
advising optimal actions to students, thereby disregarding
whether task-level policies actually learn. No exploratory ac-
tions are advised, causing poor task-level performance after
advising. This counterexample demonstrates that advising-
level rewards that reflect student learning progress, rather
than task-level reward r, are critical for useful advising.
Comparisons to existing teaching approaches Table 2
shows extensive comparisons of existing heuristics-based
teaching approaches, no teaching (independent Q-learning),
and LeCTR with all advising rewards introduced in Table 1.
We use the VEG advising-level reward in the final version of
050010001500Teacher Training Iteration024ReturnWithteachersin-loopAfterteachingFailure to learn0246810LeCTRTrainingEpoch0100#EpisodestoV∗NoTeachingLeCTR0◦90◦180◦270◦Teammateactionrotation0100AUCLeCTRImport.AdvisingCorrectImportantAdHocVisit Table 2: ¯V and Area under the Curve (AUC) for teaching algorithms. Best results in bold (computed via a t-test with p < 0.05).
Independent Q-learning correspond to the no-teaching case. †Final version LeCTR uses the VEG advising-level reward.
Algorithm
Independent Q-learning (No Teaching)
Ask Important (Amir et al. 2016)
Ask Uncertain (Clouse 1996)
Early Advising (Torrey and Taylor 2013)
Import. Advising (Torrey and Taylor 2013)
Early Correcting (Amir et al. 2016)
Correct Important (Torrey and Taylor 2013)
AdHocVisit (da Silva, Glatt, and Costa 2017)
AdHocTD (da Silva, Glatt, and Costa 2017)
LeCTR (with JVG)
LeCTR (with QTR)
LeCTR (with TDG)
LeCTR (with LG)
LeCTR (with LGG)
LeCTR†
Repeated Game
¯V
45±1
45±1
45±1
45±1
AUC
2.75±2.12 272±210
1.74±1.89 178±181
1.74±1.89 170±184
0.45±0.00
0.45±0.00
0.45±0.00
0.45±0.00
2.49±2.04 244±199
1.88±1.94 184±189
4.16±1.17 405±114
4.52±0.00 443±3
3.36±1.92 340±138
3.88±1.51 375±132
4.41±0.69 430±53
4.52±0.00 443±3
Hallway Game
¯V
Room Game
¯V
0.42±0.33
0.00±0.00
0.00±0.00
0.00±0.00
0.57±0.03
0.00±0.00
0.56±0.00
0.43±0.33
0.39±0.33
0.11±0.27
0.20±0.32
0.26±0.34
0.37±0.34
0.56±0.27
0.68±0.07
AUC
22±25
0±0
0±0
0±0
48±8
0±0
51±7
22±26
26±29
6±21
11±22
25±32
30±32
56±23
79±16
0.56±0.35
0.53±0.36
0.00±0.00
0.00±0.00
0.67±0.07
0.00±0.00
0.67±0.07
0.57±0.34
0.49±0.37
0.25±0.37
0.21±0.35
0.19±0.34
0.13±0.29
0.22±0.35
0.77±0.00
AUC
36±24
39±27
0±0
0±0
39±17
0±0
39±16
38±24
26±24
21±31
12±22
15±25
13±22
27±29
71±3
our LeCTR algorithm, but show all advising reward results
for completeness. We report both final task-level performance
after teaching, ¯V , and also area under the task-level learning
curve (AUC) as a measure of rate of learning; higher values
are better for both. Single-agent approaches requiring an ex-
pert teacher are extended to the MARL setting by using team-
mates' policies (pre-trained to expert level) as each agent's
teacher. In the Repeated game, LeCTR attains best perfor-
mance in terms of final value and rate of learning (AUC).
Existing approaches always advise the teaching agent's opti-
mal action to its teammate, resulting in suboptimal returns.
In the Hallway and Room games, approaches that tend to
over-advise (e.g., Ask Uncertain, Early Advising, and Early
Correcting) perform poorly. AdHocVisit and AdHocTD fare
better, as their probabilistic nature permits agents to take
exploratory actions and sometimes learn optimal policies.
Importance Advising and Correct Important heuristics lead
agents to suboptimal (distant) goals in Hallway and Room,
yet attain positive value due to domain symmetries.
LeCTR outperforms all approaches when using the VEG
advising-level reward (Table 2). While the JVG advising-
level reward seems an intuitive measure of learning progress
due to directly measuring task-level performance, its high
variance in situations where the task-level value is sensitive to
policy initialization sometimes destabilizes training. JVG is
also expensive to compute, requiring game rollouts after each
advice exchange. LG and TDG perform poorly due to the
high variance of task-level losses used to compute them. We
hypothesize that VEG performs best as its thresholded binary
advising-level reward filters the underlying noisy task-level
losses for teachers. A similar result is reported in recent work
on teaching of supervised learners, where threshold-based
advising-level rewards have good empirical performance (Fan
et al. 2018). Fig. 4b shows improvement of LeCTR's advising
policies due to training, measured by the number of task-level
(a) Hallway domain.
(b) Room domain.
Figure 6: LeCTR accelerates multiagent transfer learning.
episodes needed to converge to the max value reached, V ∗.
LeCTR outperforms the rate of learning for the no-teaching
case, stabilizing after roughly 4-6 training epochs.
Teaching for transfer learning Learning to teach can
also be applied to multiagent transfer learning. We first pre-
train task-level policies in the Hallway/Room tasks (denote
these T1), flip agents' initial positions, then train agents to
use teammates' T1 task-level policies to accelerate learning
in flipped task T2. Results for Hallway and Room are shown
in Figs. 6a and 6b, respectively, where advising accelerates
rate of learning using prior task knowledge. Next, we test
transferability of advising policies themselves (i.e., use ad-
vising policies trained for one task to accelerate learning
in a brand new, but related, task). We fix (no longer train)
advising policies from the above transfer learning test. We
then consider 2 variants of Room: one with the domain (in-
cluding initial agent positions) flipped vertically (T3), and
one flipped vertically and horizontally (T4). We evaluate the
fixed advising policies (trained to transfer from T1 → T2)
on transfer from T3 → T4. Learning without advising on T4
yields AUC 24 ± 26, while using the fixed advising policy
for transfer T3 → T4 attains AUC 68 ± 17. Thus, learning
is accelerated even when using pre-trained advising policies.
While transfer learning typically involves more significant
differences in tasks, these preliminary results motivate future
0246810LeCTRTrainingEpoch0100#EpisodestoV∗NoTeachingLeCTR0246810LeCTRTrainingEpoch0100#EpisodestoV∗NoTeachingLeCTR (a) No communication cost, c = 0, agents advise opposite actions.
(b) With c = 0.5, one agent leads & advises actions opposite its own.
Figure 7: Hallway game, impact of communication cost c on advising policy behaviors. First and second rows show probabilities
of action advice, P ((cid:101)ai) and P ((cid:101)aj), for agents i and j, respectively, as their advising policies are trained using LeCTR.
work on applications of advising for MARL transfer learning.
Advising heterogeneous teammates We consider het-
erogeneous variants of the Room game where one agent,
j, uses rotated versions of its teammate i's action space;
e.g., for rotation 90◦, agent i's action indices correspond to
(up/right/down/left), while j's to (left/up/right/down). Com-
parisons of LeCTR and the best-performing existing meth-
ods are shown in Fig. 5b for all rotations. Prior approaches
(Importance Advising and Correct Important) work well for
homogeneous actions (0◦ rotation). However, they attain 0
AUC for heterogeneous cases, as agents always advise action
indices corresponding to their local action spaces, leading
teammates to no-reward regions. AdHocVisit works reason-
ably well for all rotations, by sometimes permitting agents to
explore. LeCTR attains highest AUC for all rotations.
Effect of communication cost on advice exchange We
evaluate impact of communication cost on advising by de-
ducting cost c from advising rewards for each piece of advice
exchanged. Fig. 7 shows a comparison of action advice prob-
abilities for communication costs c = 0 and c = 0.5 in the
Hallway game. With no cost (c = 0 in Fig. 7a), agents learn
to advise each other opposite actions ((cid:101)aleft and(cid:101)aright, respec-
tively) in addition to exploratory actions. As LeCTR's VEG
advising-level rewards are binary (0 or 1), two-way advising
nullifies positive advising-level rewards, penalizing excessive
advising. Thus, when c = 0.5 (Fig. 7b), advising becomes
unidirectional: one agent advises opposite exploratory actions
of its own, while its teammate tends not to advise.
Related Work
Effective diffusion of knowledge has been studied in many
fields, including inverse reinforcement learning (Ng and
Russell 2000), apprenticeship learning (Abbeel and Ng
2004), and learning from demonstration (Argall et al. 2009),
wherein students discern and emulate key demonstrated be-
haviors. Works on curriculum learning (Bengio et al. 2009)
are also related, particularly automated curriculum learn-
ing (Graves et al. 2017). Though Graves et al. focus on
single student supervised/unsupervised learning, they high-
light interesting measures of learning progress also used
here. Several works meta-learn active learning policies for
supervised learning (Bachman, Sordoni, and Trischler 2017;
Fang, Li, and Cohn 2017; Pang, Dong, and Hospedales 2018;
Fan et al. 2018). Our work also uses advising-level meta-
learning, but in the regime of MARL, where agents must
learn to advise teammates without destabilizing coordination.
In action advising, a student executes actions suggested by
a teacher, who is typically an expert always advising the opti-
mal action (Torrey and Taylor 2013). These works typically
use state importance value I(s, a) = maxa Q(s, a)−Q(s, a)
to decide when to advise, estimating the performance differ-
ence between the student's best action versus intended/worst-
case action a. In student-initiated approaches such as Ask Un-
certain (Clouse 1996) and Ask Important (Amir et al. 2016),
the student decides when to request advice using heuristics
based on I(s, a). In teacher-initiated approaches such as Im-
portance Advising (Torrey and Taylor 2013), Early Correct-
ing (Amir et al. 2016), and Correct Important (Torrey and
Taylor 2013), the teacher decides when to advise by com-
paring student policy πS to expert policy πT . Q-Teaching
(Fachantidis, Taylor, and Vlahavas 2017) learns when to ad-
vise by rewarding the teacher I(s, a) when it advises. See the
supplementary material for details of these approaches.
While most works on information transfer target single-
agent settings, several exist for MARL. These include im-
itation learning of expert demonstrations (Le et al. 2017),
cooperative inverse reinforcement learning with a human and
robot (Hadfield-Menell et al. 2016), and transfer to parallel
learners in tasks with similar value functions (Taylor et al.
2013). To our knowledge, AdHocVisit and AdHocTD (da
Silva, Glatt, and Costa 2017) are the only action advising
methods that do not assume expert teachers; teaching agents
always advise the action they would have locally taken in the
student's state, using state visit counts as a heuristic to decide
when to exchange advise. Wang et al. (2018) uses da Silva,
Glatt, and Costa's teaching algorithm with minor changes.
Contribution
This work introduced a new paradigm for learning to teach
in cooperative MARL settings. Our algorithm, LeCTR, uses
agents' task-level learning progress as advising policy feed-
back, training advisors that improve the rate of learning with-
out harming final performance. Unlike prior works (Torrey
and Taylor 2013; Taylor et al. 2014; Zimmer, Viappiani, and
Weng 2014), our approach avoids hand-crafted advising poli-
cies and does not assume expert teachers. Due to the many
complexities involved, we focused on the pairwise problem,
targeting the issues of when and what to teach. A natural
avenue for future work is to investigate the n-agent setting,
extending the ideas presented here where appropriate.
01P(ai)aleftarighta∅(Noadvice)012345LeCTRTrainingEpoch01P(aj)01P(ai)aleftarighta∅(Noadvice)012345LeCTRTrainingEpoch01P(aj) Acknowledgements
Research funded by IBM (as part of the MIT-IBM Watson AI
Lab initiative) and a Kwanjeong Educational Foundation Fel-
lowship. The authors thank Dr. Kasra Khosoussi for fruitful
discussions early in the paper development process.
References
Abbeel, P., and Ng, A. Y. 2004. Apprenticeship learning via
inverse reinforcement learning. In Proceedings of the twenty-
first international conference on Machine learning, 1. ACM.
Amir, O.; Kamar, E.; Kolobov, A.; and Grosz, B. J. 2016. Inter-
active teaching strategies for agent training. International Joint
Conferences on Artificial Intelligence.
Argall, B. D.; Chernova, S.; Veloso, M.; and Browning, B. 2009.
A survey of robot learning from demonstration. Robotics and
autonomous systems 57(5):469 -- 483.
Bachman, P.; Sordoni, A.; and Trischler, A. 2017. Learning
algorithms for active learning. In International Conference on
Machine Learning, 301 -- 310.
Bengio, Y.; Louradour, J.; Collobert, R.; and Weston, J. 2009.
Curriculum learning. In Proceedings of the 26th annual inter-
national conference on machine learning, 41 -- 48. ACM.
Clouse, J. A. 1996. On integrating apprentice learning and
reinforcement learning.
da Silva, F. L.; Glatt, R.; and Costa, A. H. R.
2017. Si-
multaneously learning and advising in multiagent reinforce-
ment learning. In Proceedings of the 16th Conference on Au-
tonomous Agents and MultiAgent Systems, 1100 -- 1108. Interna-
tional Foundation for Autonomous Agents and Multiagent Sys-
tems.
Fachantidis, A.; Taylor, M. E.; and Vlahavas, I. 2017. Learning
to teach reinforcement learning agents. Machine Learning and
Knowledge Extraction 1(1):2.
Fan, Y.; Tian, F.; Qin, T.; Li, X.-Y.; and Liu, T.-Y. 2018. Learn-
ing to teach. In International Conference on Learning Repre-
sentations.
Fang, M.; Li, Y.; and Cohn, T. 2017. Learning how to active
learn: A deep reinforcement learning approach. In Proceedings
of the 2017 Conference on Empirical Methods in Natural Lan-
guage Processing, 595 -- 605.
Foerster, J.; Assael, I. A.; de Freitas, N.; and Whiteson, S. 2016.
Learning to communicate with deep multi-agent reinforcement
In Advances in Neural Information Processing Sys-
learning.
tems, 2137 -- 2145.
Foerster, J. N.; Chen, R. Y.; Al-Shedivat, M.; Whiteson, S.;
Abbeel, P.; and Mordatch, I. 2018. Learning with opponent-
In Proceedings of the 17th Conference
learning awareness.
on Autonomous Agents and MultiAgent Systems. International
Foundation for Autonomous Agents and Multiagent Systems.
Graves, A.; Bellemare, M. G.; Menick, J.; Munos, R.; and
Kavukcuoglu, K. 2017. Automated curriculum learning for neu-
ral networks. In International Conference on Machine Learn-
ing, 1311 -- 1320.
Hadfield-Menell, D.; Russell, S. J.; Abbeel, P.; and Dragan, A.
2016. Cooperative inverse reinforcement learning. In Advances
in neural information processing systems, 3909 -- 3917.
Jang, E.; Gu, S.; and Poole, B. 2016. Categorical reparameteri-
zation with gumbel-softmax. arXiv preprint arXiv:1611.01144.
Kingma, D. P., and Ba, J. 2014. Adam: A method for stochastic
optimization. arXiv preprint arXiv:1412.6980.
Le, H. M.; Yue, Y.; Carr, P.; and Lucey, P. 2017. Coordinated
multi-agent imitation learning. In International Conference on
Machine Learning, 1995 -- 2003.
Lowe, R.; Wu, Y.; Tamar, A.; Harb, J.; Abbeel, O. P.; and Mor-
datch, I. 2017. Multi-agent actor-critic for mixed cooperative-
competitive environments. In Advances in Neural Information
Processing Systems, 6382 -- 6393.
Mnih, V.; Kavukcuoglu, K.; Silver, D.; Rusu, A. A.; Veness, J.;
Bellemare, M. G.; Graves, A.; Riedmiller, M.; Fidjeland, A. K.;
Ostrovski, G.; et al. 2015. Human-level control through deep
reinforcement learning. Nature 518(7540):529.
Ng, A. Y., and Russell, S. J. 2000. Algorithms for inverse re-
inforcement learning. In Proceedings of the Seventeenth Inter-
national Conference on Machine Learning, 663 -- 670. Morgan
Kaufmann Publishers Inc.
Oliehoek, F. A., and Amato, C. 2016. A concise introduction to
decentralized POMDPs, volume 1. Springer.
Pang, K.; Dong, M.; and Hospedales, T. 2018. Meta-learning
transferable active learning policies by deep reinforcement
learning.
Rogers, E. M. 2010. Diffusion of innovations. Simon and
Schuster.
Sukhbaatar, S.; Fergus, R.; et al. 2016. Learning multiagent
communication with backpropagation. In Advances in Neural
Information Processing Systems, 2244 -- 2252.
Sutton, R. S., and Barto, A. G. 1998. Reinforcement learning:
An introduction, volume 1. MIT press Cambridge.
Sutton, R. S.; McAllester, D. A.; Singh, S. P.; and Mansour, Y.
2000. Policy gradient methods for reinforcement learning with
function approximation. In Advances in neural information pro-
cessing systems, 1057 -- 1063.
Taylor, A.; Dusparic, I.; Galván-López, E.; Clarke, S.; and
Transfer learning in multi-agent sys-
Cahill, V.
In Workshop on Theoretically
tems through parallel transfer.
Grounded Transfer Learning at the 30th International Conf. on
Machine Learning (Poster), volume 28, 28. Omnipress.
Taylor, M. E.; Carboni, N.; Fachantidis, A.; Vlahavas, I.; and
Torrey, L. 2014. Reinforcement learning agents providing ad-
vice in complex video games. Connection Science 26(1):45 -- 63.
Torrey, L., and Taylor, M. 2013. Teaching on a budget: Agents
In Proceedings of
advising agents in reinforcement learning.
the 2013 international conference on Autonomous agents and
multi-agent systems, 1053 -- 1060. International Foundation for
Autonomous Agents and Multiagent Systems.
Wang, Y.; Lu, W.; Hao, J.; Wei, J.; and Leung, H.-F. 2018.
Efficient convention emergence through decoupled reinforce-
ment social learning with teacher-student mechanism. In Pro-
ceedings of the 17th International Conference on Autonomous
Agents and MultiAgent Systems, 795 -- 803. International Foun-
dation for Autonomous Agents and Multiagent Systems.
Zimmer, M.; Viappiani, P.; and Weng, P. 2014. Teacher-student
In AAMAS
framework: a reinforcement learning approach.
Workshop Autonomous Robots and Multirobot Systems.
2013.
Supplementary Material
Details of Advising-level Rewards
Recall in Phase II of LeCTR, advising policies are trained to
maximize advising-level rewards that should, ideally, reflect
the objective of accelerating task-level learning. Selection of
an appropriate advising-level reward is, itself, non-obvious.
Due to this, we considered a variety of advising-level rewards,
each corresponding to a different measure of task-level learn-
ing after the student executes an action advice and uses it to
T below
are detailed for the case where agent i is student and agent j
teacher (i.e., flip the indices for the reverse case). Recall that
the shared reward used to jointly train all advising policies is
update its task-level policy. Advising-level rewards(cid:101)rj
(cid:101)r =(cid:101)ri
• Joint Value Gain (JVG): Let θt and θt+1, respectively,
denote agents' joint task-level policy parameters before
and after learning from an experience resulting from action
advise. The JVG advising-level reward measures improve-
ment in task-level value due to advising, such that,
T +(cid:101)rj
T .
(cid:101)rj
T = V (s; θt+1) − V (s; θt).
(3)
This is, perhaps, the most intuitive choice of advising-level
reward, as it directly measures the gain in task-level perfor-
mance due to advising. However, the JVG reward has high
variance in situations where the task-level value is sensitive
to policy initialization, which sometimes destabilizes train-
ing. Moreover, the JVG reward requires a full evaluation
of task-level performance after each advising step, which
can be expensive due to the game rollouts required.
• Q-Teaching Reward (QTR): The QTR advising-level re-
ward extends Q-Teaching (Fachantidis, Taylor, and Vla-
havas 2017) to MARL by using
T = IT (oi, ai; hi) = max
QT (oi, a;hi) − QT (oi, ai;hi),
(4)
each time advising occurs. The motivating intuition for
QTR is that teacher j should have higher probability of
advising when they estimate that the student's intended
action, ai, can be outperformed by a different action (the
arg max action).
• TD Gain (TDG): For temporal difference (TD) learners,
the TDG advising-level reward measures improvement of
student i's task-level TD error due to advising,
(cid:101)rj
a
t − δi
t+1,
(5)
t is i's TD error at timestep t. For example, if
where δi
agents are independent Q-learners at the task-level, then,
(6)
a(cid:48) Q(o(cid:48), a(cid:48); θi, h(cid:48)i) − Q(o, a; θi, hi).
δ = r + max
(cid:101)rj
T = δi
The motivating intuition for the TDG advising-level reward
is that actions that are anticipated to reduce student i's task-
level TD error should be advised by the teacher j.
• Loss Gain (LG): The LG advising-level reward applies to
many loss-based algorithms, measuring improvement of
the task-level loss function used by student learner i,
(cid:101)rj
T = L(θi
t) − L(θi
t+1).
(7)
For example, if agents are independent Q-learners using
parameterized task-level policies at the task-level, then
a(cid:48) Q(o(cid:48), a(cid:48); θi, h(cid:48)i) − Q(o, a; θi, hi)]2.
L(θi) = [r + γ max
(8)
The motivating intuition for the LG reward is similar to the
TDG, in that teachers should advise actions they anticipate
to decrease student's task-level loss function.
• Loss Gradient Gain (LGG): The LGG advising-level re-
ward is an extension of the gradient prediction gain (Graves
et al. 2017), which measures the magnitude of student pa-
rameter updates due to teaching,
(9)
The intuition here is that larger task-level parameter up-
dates may be correlated to learning progress.
• Value Estimation Gain (VEG): VEG rewards teachers
when student's local value function estimates exceed a
threshold τ,
T = 1( V (θi) > τ ),
(10)
using V (θi) = maxai Q(oi, ai; θi, hi) and indicator func-
tion 1(·). The motivation here is that the student's value
function approximation is correlated to its estimated perfor-
mance as a function of its local experiences. A convenient
means of choosing τ is to set it as a fraction of the value
estimated when no teaching occurs.
(cid:101)rj
(cid:101)rj
T = ∇θiL(θi)2
2.
1996), where k is a threshold parameter.
Details of Heuristics-based Advising Approaches
Existing works on action advising typically use the state
importance value Iρ(s, a) = maxa Qρ(s, a) − Qρ(s, a) to
decide when to advise, where ρ = S for student-initiated
advising, ρ = T for teacher-initiated advising, Qρ is the
corresponding action-value function, and a is the student's
intended action if known (or the worst-case action otherwise).
Iρ(s, a) estimates the performance difference of best versus
intended student action in state s. The following is a summary
of prior advising approaches:
• The Ask Important heuristic (Amir et al. 2016) requests
advice whenever IS(s, a) ≥ k, where k is a threshold
parameter.
• Ask Uncertain requests when IS(s, a) < k (Clouse
• Early Advising advises until advice budget depletion.
• Importance Advising advises when IT (s, a) ≥ k (Torrey
• Early Correcting advises when πS(s) (cid:54)= πT (s) (Amir et
al. 2016).
• Correct Important advises when IT (s) ≥ k and
πS(s) (cid:54)= πT (s) (Torrey and Taylor 2013), where k is
a threshold parameter.
• Q-Teaching (Fachantidis, Taylor, and Vlahavas 2017)
learns when to advise by rewarding the teacher IT (s, a)
when advising occurs. Constrained by a finite advice bud-
get, Q-Teaching has advising performance similar to Im-
portance Advising, with the advantage of not requiring a
tuned threshold k.
and Taylor 2013), where k is a threshold parameter.
to their paper for additional details).
Policy training is conducted with the Adam optimization
algorithm (Kingma and Ba 2014), using a learning rate of
1e−3. We use γ = 0.95 at the task-level and(cid:101)γ = 0.99 at the
advising-level level to induce long-horizon teaching policies.
Similar to Graves et al. (2017), we use reservoir sampling to
adaptively rescale advising-level rewards with time-varying
and non-normalized magnitudes (all except VEG) to the in-
terval [−1, 1]. Refer to Graves et al. for details on how this is
conducted.
Experimental Procedures
In Table 2, ¯V is computed by running each algorithm until
convergence of task-level policies π = (cid:104)πi, πj(cid:105), and com-
puting the mean value obtained by the final joint policy π.
The area under the learning curve (AUC) is computed by
intermittently evaluating the resulting task-level policies π
throughout learning; while teacher advice is used during
learning, the AUC is computed by evaluating the resulting
π after advising (i.e., in absence of teacher advice actions,
such that AUC measures actual student learning progress).
All results and uncertainties are reported using at least 10 in-
dependent runs, with most results using over 20 independent
runs. In Table 2, best results in bold are computed using a
Student's t-test with significance level α = 0.05.
Pairwise combinations of student- and teacher-initiated ap-
proaches can be used to constitute a jointly-initiated approach
(Amir et al. 2016), such as ours. As shown in our experiments,
application of single-agent teaching approaches yields poor
performance in MARL games.
Optimal Action Advising Note that in the majority of
prior approaches, the above heuristics are used to decide
when to advise. To address the question of what to advise,
these works typically assume that teachers have expert-level
knowledge and always advise optimal action to students.
Optimal action advising has a strong empirical track record
in single-agent teaching approaches (Torrey and Taylor 2013;
Zimmer, Viappiani, and Weng 2014; Amir et al. 2016). In
such settings, the assumed homogeneity of the teacher and
student's optimal policies indeed leads optimal action advice
to improve student learning (i.e., when the expert teacher's
optimal policy is equivalent to student's optimal policy). In
the context of multiagent learning, however, this advising
strategy has primarily been applied to games where behav-
ioral homogeneity does not substantially degrade team per-
formance (da Silva, Glatt, and Costa 2017). However, there
exist scenarios where multiple agents learn best by exhibiting
behavioral diversity (e.g., by exploring distinct regions of
the state-action space), or where agents have heterogeneous
capabilities/action/observation spaces altogether (e.g., coor-
dination of 2-armed and 3-armed robots, robots with different
sensors, etc.). Use of optimal action advising in cooperative
multiagent tasks can lead to suboptimal joint return, particu-
larly when the optimal policies for agents are heterogeneous.
We show this empirically in several of our experiments.
In contrast to earlier optimal advising approaches, our
LeCTR algorithm applies to the above settings in addition
to the standard homogeneous case; this is due to LeCTR's
ability to learn a policy over not only when to advise, but
also what to advise. As shown in our experiments, while
existing probabilistic advising strategies (e.g., AdHocTD and
AdHocVisit) attain reasonable performance in heterogeneous
action settings, they do so passively by permitting students
to sometimes explore their local action spaces. By contrast,
LeCTR agents attain even better performance by actively
learning what to advise within teammates' action spaces; this
constitutes a unique strength of our approach.
Architecture, Training Details, and
Hyperparameters
At the teaching level, our advising-level critic is parameter-
ized by a 3-layer multilayer perceptron (MLP), consisting
of internal rectified linear unit (ReLU) activations, linear
output, and 32 hidden units per layer. Our advising-level
actors (advice request/response policies) use a similar param-
eterization, with the softmax function applied to outputs for
discrete advising-level action probabilities. Recurrent neural
networks may also be used in settings where use of advising-
level observation histories yields better performance, though
we did not find this necessary in our domains. As in Lowe et
al. (2017), we use the Gumbel-Softmax estimator (Jang, Gu,
and Poole 2016) to compute gradients for the teaching poli-
cies over discrete advising-level actions (readers are referred
Notation
The following summarizes the notation used throughout the
paper. In general: superscripts i denote properties for an agent
i (e.g., ai); bold notation denotes joint properties for the team
(e.g., a = (cid:104)ai, aj(cid:105)); tilde accents denote properties at the
advising-level (e.g.,(cid:101)ai); and bold characters with tilde accent
denote joint advising-level properties (e.g.,(cid:101)a = (cid:104)(cid:101)ai,(cid:101)aj(cid:105)).
Symbol Definition
Task (a Dec-POMDP)
Task-level learning algorithm
Joint task-level policy
Agent i's task-level policy
Agent i's task-level policy parameters
Task-level value
Agent i's action value function
Agent i's action value vector (i.e., vector of
action-values for all actions)
State space
State
State transition function
Joint action space
Joint action
Agent i's action space
Agent i's action
Joint observation space
Joint observation
Agent i's observation space
Agent i's observation
Observation function
Reward function
Task-level reward
Discount factor
T
L
π
πi
θi
V
Qi
(cid:126)Qi
S
s
T
A
a
Ai
ai
Ω
o
Ωi
oi
O
R
r
γ
M Task-level experience replay memory
Task-level temporal difference error
δ
Agent's advising-level role, where ρ = S for
student role, ρ = T for teacher role
Agent i's advice request policy
Agent i's advice response policy
(cid:101)πi
(cid:101)πi
Joint advising policy(cid:101)π = (cid:104)(cid:101)πi
(cid:101)π
S,(cid:101)πi
S,(cid:101)πj
T ,(cid:101)πj
(cid:101)θ
T(cid:105)
(cid:101)ai
(cid:101)a∅
T ∈ Aj ∪ {(cid:101)a∅}
Agent i's advice response(cid:101)ai
(cid:101)ai
Joint advising action(cid:101)a = (cid:104)(cid:101)ai
(cid:101)a
S,(cid:101)aj
T ,(cid:101)aj
S,(cid:101)ai
T(cid:105)
(cid:101)oi
(cid:101)oi
Joint advising obs.(cid:101)o = (cid:104)(cid:101)oi
(cid:101)o
S,(cid:101)oj
S,(cid:101)oi
T ,(cid:101)oj
Advising-level reward(cid:101)r =(cid:101)ri
T +(cid:101)rj
(cid:101)r
T(cid:105)
(cid:101)γ
(cid:102)M Advising-level experience replay memory
Agent i's behavioral policy
Agent i's student-perspective advising obs.
Agent i's teacher-perspective advising obs.
Joint advising-level policy parameters
Agent i's advice request action
Special no-advice action
Advising-level discount factor
βi
ρ
S
S
T
S
T
T
T
LeCTR Algorithm
Algorithm 1 Get advising-level observations
1: function GETADVISEOBS(o, θ)
2:
3:
4:
5:
6:
7:
8: end function
for agents α ∈ {i, j} do
(cid:101)oα
Let −α denote α's teammate.
(cid:101)oα
S =(cid:104)oα, Qα(oα; hα)(cid:105)
T =(cid:104)o−α, Q−α(o−α; h−α), Qα(o−α; h−α)(cid:105)
return(cid:101)o = (cid:104)(cid:101)oi
S,(cid:101)oj
S,(cid:101)oi
T ,(cid:101)oj
T(cid:105)
end for
Algorithm 2 LeCTR Algorithm
1: for Phase II episode(cid:101)e = 1 to (cid:101)E do
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
Initialize task-level policy parameters θ
for Phase I episode e = 1 to E do
end if
if No advising occurred then
for agents α ∈ {i, j} do
Select action aα via local policy πα
o ← initial task-level observation
for task-level timestep t = 1 to tend do
(cid:101)o ← GETADVISEOBS(o, θ)
Exchange advice(cid:101)aα via advising policies
a ← (cid:104)ai, aj(cid:105),(cid:101)a ← (cid:104)(cid:101)ai,(cid:101)aj(cid:105)
(cid:101)o(cid:48) ← GETMETAOBS(o(cid:48), θ(cid:48))
(cid:101)ri
T ,(cid:101)rj
T ,(cid:101)o(cid:48)(cid:105) in buffer (cid:102)M
Store (cid:104)(cid:101)o,(cid:101)a,(cid:101)r =(cid:101)ri
T +(cid:101)rj
T ← Compute advising-level rewards
(cid:48);(cid:101)θ)−(cid:101)Q((cid:101)o,(cid:101)a;(cid:101)θ))2]
L((cid:101)θ) = E(cid:101)o,(cid:101)a,(cid:101)r,(cid:101)o(cid:48)∼(cid:102)M[((cid:101)r+γ(cid:101)Q((cid:101)o
(cid:48),(cid:101)a
end for
r, o(cid:48) ← Execute action a in task
θ(cid:48)i ← Li(θ(cid:48)i,(cid:104)oi, ai, r, o(cid:48)i(cid:105))
θ(cid:48)j ← Lj(θ(cid:48)j,(cid:104)oj, aj, r, o(cid:48)j(cid:105))
end for
Update advising-level critic by minimizing loss,
end for
(cid:12)(cid:12)(cid:12)(cid:101)a(cid:48)=(cid:101)π((cid:101)o(cid:48))
Update advising policy parameters(cid:101)θα
ρ (cid:101)Q((cid:101)o,(cid:101)a;(cid:101)θ)(cid:3),
ρ ((cid:101)aα
log(cid:101)πα
ρ(cid:101)oα
ρ )∇(cid:101)aα
(cid:2)(cid:88)
∇(cid:101)θα
ρ via,
ρ
for agents α ∈ {i, j} do
for roles ρ ∈ {S, T} do
∇(cid:101)θJ((cid:101)θ) ==E(cid:101)o,(cid:101)a∼(cid:102)M
α∈{i,j}
ρ∈{S,T}
end for
26:
27:
28: end for
end for
|
1905.10490 | 1 | 1905 | 2019-05-25T00:26:39 | Finding new routes for integrating Multi-Agent Systems using Apache Camel | [
"cs.MA",
"cs.AI"
] | In Multi-Agent Systems (MAS) there are two main models of interaction: among agents, and between agents and the environment. Although there are studies considering these models, there is no practical tool to afford the interaction with external entities with both models. This paper presents a proposal for such a tool based on the Apache Camel framework by designing two new components, namely camel-jason and camel-artifact. By means of these components, an external entity is modelled according to its nature, i.e., whether it is autonomous or non-autonomous, interacting with the MAS respectively as an agent or an artifact. It models coherently external entities whereas Camel provides interoperability with several communication protocols. | cs.MA | cs | Finding new routes for integrating
Multi-Agent Systems using Apache Camel∗
Cleber J. Amaral1,2, S´ergio P. Bernardes1, Mateus Conceic¸ ao1
Jomi F. Hubner1, Luis P. A. Lampert1, Ot´avio A. Matoso1, Maicon R. Zatelli1
9
1
0
2
y
a
M
5
2
]
A
M
.
s
c
[
1
v
0
9
4
0
1
.
5
0
9
1
:
v
i
X
r
a
1Universidade Federal de Santa Catarina (UFSC)
Florian´opolis -- SC -- Brazil
2Instituto Federal de Santa Catarina (IFSC)
Sao Jos´e -- SC -- Brazil
[email protected], {sergiopb1998,lp.lampert}@gmail.com
[email protected], {jomi.hubner,maicon.zatelli}@ufsc.br
[email protected]
Abstract. In Multi-Agent Systems (MAS) there are two main models of interac-
tion: among agents, and between agents and the environment. Although there
are studies considering these models, there is no practical tool to afford the
interaction with external entities with both models. This paper presents a pro-
posal for such a tool based on the Apache Camel framework by designing two
new components, namely camel-jason and camel-artifact. By means of these
components, an external entity is modelled according to its nature, i.e., whether
it is autonomous or non-autonomous, interacting with the MAS respectively as
an agent or an artifact. It models coherently external entities whereas Camel
provides interoperability with several communication protocols.
1. Introduction
MAS literature has plenty of research about agents' interactions, i.e., agents sending and
receiving messages to and from other agents (A-A). Many approaches model almost any
entity as an agent and thus the interaction remains something among agents. However
there are new approaches that questioned the agentification method proposing an MAS
where non-autonomous entities are conceived as artifacts in the environment. In these
approaches, the development of an MAS considers the design of both agents and arti-
facts. The environment is not simply what is outside the system (the exogenous envi-
ronment), but it is designed accordingly to the system purpose (the endogenous environ-
ment) [Ricci et al. 2006, Omicini et al. 2008]. In this sense, we have two models of in-
teractions: agent-to-agent (A-A) and agent-to-environment (A-E). In the former, an agent
communicates with another agent using an Agent Communication Language (ACL) and
in the latter, an agent perceives and acts upon artifacts in the environment.
When we consider the integration with other applications, those two models are
adopted by the current development platforms. On the one hand, we have approaches
that use ACL for that purpose and other applications are seen by the agents as other
agents (having a mental state, implied by the ACL semantics). On the other hand, we
∗Supported by Petrobras project AG-BR, IFSC and UFSC.
have approaches where other applications are seen as part of the environment and agents
perceive and act on them. Some platforms provide an A-A approach while others an A-E
approach, but, as far as we know, no platform provides both. The designer is forced to
conceive some other application either as an agent or as an artifact, despite the application
properties.
In this paper, we propose to apply the same argument as [Omicini et al. 2008]
for the integration of MAS and external applications: some external applications are au-
tonomous and should be modelled as agents while others are non-autonomous and should
be modelled as artifacts.
For instance, in the Industry 4.0 context, it is expected the interaction of many
entities such as an autonomous planner sending commands to a non-autonomous machine,
which signalises what was done. Later, the planner must choose a supplier after an auction
to hire a freight to take the product to the destiny, which is usually an human. This
short example gives an idea of how comprehensive and challenging the integration can
be. We can notice that both models of integration are required: the autonomous planner
and the human are better modelled as agents, performing A-A interactions, and the non-
autonomous machine should be integrated as an artifact which when communicating with
and agent performs and A-E interaction. Following this concept, we have developed two
components for JaCaMo platform, for integration among agents, and between agents and
the environment. The referred components are used to set communication routes for the
MAS and external entities, using the framework Apache Camel [Ibsen and Anstey 2010],
a mediation tool to provide interoperability with many technologies.
2. MAS integration approaches
MAS are being applied as a core technology for distributed systems that needs cooperation
and negotiation [Roloff et al. 2016]. The integration of MAS and external entities, i.e.
any entity which was not defined its totality within the MAS itself, regards concerns
such as compatibility with standards, interoperability and portability. We have found
two main forms of integration: (i) among agents (A-A); and, (ii) between agents and the
environment (A-E).
2.1. Integration among agents (A-A)
The communication among agents is usually done by speech-acts which considers utter-
ances as actions, usually intending to change the mental state of recipients. The utterance
can inform beliefs, desires and intentions of rational agents that attempt to influence other
agents. The Knowledge Query and Manipulation Language (KQML) is the first speech-
act based language providing high level communication in the distributed artificial intelli-
gence applications [Vieira et al. 2007]. In fact, once speech acts became widely accepted
in MAS community, the integration among different agent's platforms was facilitated.
Currently FIPA-ACL, which is very similar to KQML, is the main standard for
agents communication. FIPA-ACL uses performatives to make explicit an agent's in-
tention for each sent message, for instance, inform is used to influence the recipient
to believe in something, and request to influence the recipient to add something as a
goal [Vieira et al. 2007].
Another communication aspect is related to the expected sequence of messages.
Conversations among agents usually follow some patterns which are often referred to
as interaction protocols. Typical patterns such as negotiation, auction, and task delega-
tion are defined using FIPA standards [Bellifemine et al. 2005]. In addiction, there are
communication infrastructures that allow agents to be distributed over a network. The
challenge in A-A is the integration between an agent, for instance, using FIPA-ACL, and
another agent using another language, for instance, an human. This situation leads to the
necessity of some tool to make both end-points compatible.
2.2. Integration between agents and the environment (A-E)
There are systems or parts of a system which are better seen as resources or tools that can
be used by agents to achieve their goals. These entities, called artifacts, have no internal
goals, they are not autonomous and neither proactive, but they supply useful functionali-
ties for agents. An analogy for agents and artifacts is the interaction between humans, as
autonomous entities, and tools they exploit in their activities [Ricci et al. 2006]. For in-
stance, a blackboard shared by agents would be modelled as an artifact, being predictable
and deterministic, if not, it would perform undesirable autonomous behaviour.
Artifacts are placed in workspaces which represent areas of the MAS environ-
ment. Agents can perceive changes and act within the workspaces they are occupying,
and on artifacts they are watching, i.e., being aware of events and signals due to interac-
tions with non-autonomous entities held inside of virtual boundaries. The environment
can reflect the effects of agents' actions and other phenomena. It is being treated as a
first-class programming abstraction with similar importance of agents programming ab-
straction [Ricci et al. 2006].
Besides modelling non-autonomous entities, the artifacts can also be used for
other purposes: (i) for agents coordination using shared artifacts such as organisational
boards or coordination marks; (ii) for indirect communication among agents, for in-
stance, by blackboard artifacts; (iii) for implementing the user interface of a system;
(iv) for controlling transactions over environment elements through distribution and
synchronisation facilities; and (v) for integration between the MAS and external enti-
ties [Boissier et al. 2019].
Regarding the use of artifacts to integrate external entities, the integration is done
usually through specific Application Programming Interfaces (APIs). The main concern
with this approach regards to the high programming effort when there are different proto-
cols in scenarios of heterogeneous devices.
3. Integrating A-A and A-E using Camel
Back to our example in the Industry 4.0 context, Figure 1 shows a process that begins with
a packed product, in a production line, up to its delivery to the customer. On the first step,
there is an industrial device, a non-autonomous entity, that communicates using an indus-
trial protocol. Once the device signalises the end of the production, the order is checked
out on the Enterprise Resource Planning (ERP) software, another non-autonomous entity.
The supplier should choose the best offer for a freight, which may be done by access-
ing suppliers' systems to then interact with the winner, an autonomous entity. Later, the
delivery should be tracked by a monitoring system, which is non-autonomous. Finally,
when the product is near the destination, a message must keep the client, an autonomous
entity, informed.
Figure 1. Finishing production and delivering the product in Industry 4.0 context
This scenario illustrates the requirements for the integration: heterogeneous end-
points with different kinds of interactions due to the autonomous and non-autonomous
nature of entities. We propose the use of the framework Apache Camel for both integra-
tion models, A-A and A-E. We have thus two components: camel-jason for integrating
MAS's internal agents with external entities modeled as agents, and camel-artifact for
integrating the former agents with external entities modeled as environmental artifacts.
3.1. Apache Camel
Apache Camel is a lightweight Java-based framework message routing and mediation
engine [Ibsen and Anstey 2010]. Camel achieves high-performance processes handling
multiple messages concurrently, and provides functions such as routing, exception han-
dling, and testing. It uses structured messages and queues based on Enterprise Integration
Patterns (EIP) [Hohpe and Woolf 2003], preserving loose coupling among the resources.
Camel works as a middleware that can be incorporated into an application through the use
of components. Communication among Camel components is defined in so-called routes,
which set and manage how messages will be exchanged, possibly following sets of rules
and using data manipulation.
Figure 2. MAS architecture using camel-jason and camel-artifact components.
Routes define a single endpoint for each entity with an unique address. The de-
fined endpoint may receive data, through a producer and send data through a consumer.
Consumers are entities that admit data in specific formats and encapsulates it in a camel
exchange object, an item that any other producer understands and is able to decode. Pro-
ducers are entities that receive encapsulated data from the consumer decoding it in its
entity's message structure.
In our implemented components, Camel is being embedded in two slightly differ-
ent manners regarding the models of integration, A-A and A-E, as shown in Figure 2. In
the case of A-A, it works as a communication infrastructure that is used when the recipient
is not found locally. In the case of A-E, the external device is usually modelled reflecting
real operations and signals that it generates, typically having their individuals routes. In
both cases, the components are able to define tuned integration, covering a range of end-
points features. Notice that, the complexity of each supported protocol is processed in a
Camel component, which works as a bridge to Camel routes. There are more than two
hundred components available on Camel's website1 and many others on the community's
repositories.
3.2. camel-jason component
The camel-jason component enables agents to communicate with external entities through
ACL, whilst fulfilling the need of understanding those entities as agents when modeling
the MAS. In our proposal, the external entity has a kind of virtual counterpart inside the
MAS, a dummy agent. This counterpart is seen by the agents as an ordinary agent of the
system. Doing so, agents can directly communicate with external entities assuming that
they are other agents (as receivers and senders of ACL messages).
Camel-jason component provides a communication flow that is illustrated in Fig-
ure 3 where an agent interact with a service A (an external entity). Since the agent sees
the service as another agent, it uses ACL for the communication. When the service wants
to contact the agent, the camel-jason component translates the message into ACL and the
agent receives it as if it comes from an agent.
Figure 3. Communication flow using camel-jason component.
1Supported Camel components are listed in http://camel.apache.org/components.html
ACL Ag Service A Camel Router ConsumerService A camel-jason Producercamel-ServiceA Camel Router Consumer camel-ServiceB ACL ProducerService B camel-jason Ag Service B On the top communication flow showed in Figure 3, we have an agent sending a
message to an external autonomous entity: (i) the agent send an ACL message addressed
to a dummy agent, which is created by camel-jason component, referring to Service A;
(ii) the message is consumed by camel-jason component consumer; (iii) the message is
exchanged to the other side of the route, possibly being transformed; and (iv) the message
is processed by Service A component producer which prepares a service A compliance
message, which will be sent to some network address to be effectively consumed by the
Service A.
In the other way around, on the bottom of Figure 3, we have: (i) Service B sends
some data through the network reaching Service B component consumer by its network
address; (ii) the message is exchanged through Camel route, possibly being transformed;
(iii) the message is processed by camel-jason component producer which generates an
ACL message; and (iv) the receiver agent effectively consumes the ACL message.
The component uses a simplistic method to define the communication routes, in
which for many cases no actual programming is required, only XML definitions. The user
should know how to fill camel endpoint parameters according to the compatible endpoint
of the application. In cases data transformation is required, camel brings some tools for
simple transformation as well as complex ones, using embedded programming codes if
needed.
3.3. camel-artifact component
In order to sustain the A-E model, the CArtAgO infrastructure is used, and the camel-
artifact component was developed. This component allows agents to perceive and act
upon artifacts that represent external entities inside the MAS.
Notwithstanding, camel-artifact also allows the definition of communication
routes between CArtAgO artifacts and external entities. Routes for the camel-artifact
component are implemented using the Java language. The user should be aware of regu-
lar camel routes and how to define endpoints and their respective parameters.
Figure 4. Communication flow using camel-artifact component.
On the top communication flow showed in Figure 4, we have an artifact integrated
with some service A, an external non-autonomous entity. The interaction between them
is as follows: (i) the artifact sends a message to to Service A, through specific methods
provided by camel-artifact; (ii) the message, in form of an operation request, is consumed
by camel-artifact component consumer which generates a camel standardised message to
Camel Router ConsumerService A camel-artifact Producercamel-ServiceA Camel Router Consumer camel-ServiceB Service B artifactprop_obs/nprop_obs/noperation/noperation/n Producercamel-artifact Action Ag Ag Perceptionartifactprop_obs/nprop_obs/noperation/noperation/n be sent to the external entity; (iii) the message is exchanged to the other side of the route,
possibly being transformed; and (iv) the message is processed by Service A compatible
component producer which prepares a compliance final message, with the proper format
and structure, which will be sent to some network address to be effectively consumed by
the Service A.
In the other way around, on the bottom of Figure 4, we have: (i) Service B sends
some data through the network reaching Service B component consumer by its network
address; (ii) the message is exchanged through Camel route, possibly being transformed;
(iii) the message is processed by camel-artifact producer which generates an artifact op-
eration request; and (iv) the recipient artifact effectively consumes the operation request
executing the referred method.
4. Illustrative application
For a better understanding of how the camel-jason and camel-artifact components can be
used, we will resume the example of Industry 4.0, presented in Figure 1, and will build an
implementation of this system.
The Figure 5 shows the MAS fully designed, with agents, external entities and the
camel components used to implement the integration. These components are represented
in the middle layer as artifacts and dummy agents. This hypothetical scenario implements
an MAS to integrate the production and distribution stages of a product. The whole course
can be divided in five stages: (i) a Programmable Logic Controller (PLC) finishes the
product manufacturing, (ii) the information about the product is uploaded to an Enterprise
Resource Planning (ERP) software, (iii) a research starts in order to contract the best
freight company, (iv) the hired company starts transporting the product, providing its
tracking information, and (v) warns the client via chat when it is near the final destination.
The MAS is designed to unify those stages and to be responsible for managing each
process. Moreover, Camel components are used as middleware between the MAS and
external entities to integrate them.
One common question when designing the MAS is how many agents should be
used. This is not mandatory, but a natural thought is to divide the process into sections and
designate a single agent to be responsible for each part. In this case, we will consider that
PLC and ERP stages represent the production part of the process, so one agent, named
production agent, will be responsible for managing these processes. Next, there is the
hiring stage, that comprehends searching and hiring the best delivery company, for which
the distribution agent will be designated. The final stage could be thought as the delivery
process, where the last agent, named delivery agent, will be responsible for consulting
the tracking information and sending the message to the costumer.
Another part of the designing process is the identification of which model, agent
or artifact, is more suitable to represent each external entity. A common way to decide is
observing its nature, i.e. autonomous or non-autonomous. Following this idea, it could
be decided that the PLC and ERP software would be modeled as artifacts, since they are
non-autonomous entities; and the customer as an agent, an autonomous entity. Another
possibility to decide is looking at which type of communication the agents will perform
with each external entity, i.e. via message exchanging or perception-action. For example,
the action of hiring the delivery company, informing the company about some new con-
Figure 5. An industrial process illustrating the integration of an MAS with a man-
ufacturing device, updating an enterprise management software, choosing
a supplier for delivering a tracked product to the customer.
tract via email, seems natural to be modeled as an A-A interaction. For this situation, the
best suitable performative for the message is tell, since the agent is telling the supplier
about a new hired delivery. On the other hand, when the distribution agent is searching
for the best delivery company, consulting their prices and conditions, it seems more suit-
able for this information to be perceived, like people do in a websearch. The same idea
can be used when the delivery agent tracks the position of the product, the information
again is perceived by the agent, like looking at a screen. If message exchange was used
in this case the agent would be flooded with unnecessary messages.
In fact, both interpretations, i.e., perception-action vs message exchanging and
autonomous vs non-autonomous, could point to the same conclusion. This statement can
be tested in the integration between the delivery agent and the customer. The chat is
done by message exchanging and it is performed by autonomous entities. Therefore, both
interpretation reinforces the idea that the costumer should be modeled as an agent.
With the external entities modeled as MAS elements we could use camel-artifact
and camel-jason to integrate the internal agents with the artifacts and agents, respectively.
It is worth commenting that the external entities could be easily exchanged, for instance
making the auction via email instead of using a web service. The interaction via email
suggests modelling the participants as agents and use the camel-jason component for the
integration.
Now the camel routes can be developed, depending on which type of technology
the external entities use. As explained before, Camel have more than two hundred end-
points available. In this example, OPC-DA, Rest, email, MQTT and Telegram end-points
are being used to create the routes.
The code, in XML, for the route from the delivery agent to Telegram can be seen
Device using an IndustrialProtocolPLCArtifactLegendERPSoftwareInternalAgentsAgentCamelArtifactTelegram AppIntegrationwith CamelComponentscamel-opcERPArtifactcamel-restSupplier 1 . . . NArtifactcamel-restPrice: $ XLat. & Long.TrackedArtifactcamel-mqttDummyCustomerAgentcamel-telegramProduct is coming!E-mailMQTTMessageOutsideWorldCamelRoutesDummyAgentPrice: $ Ychoose freight supplier. . .. . .. . .camel-restYou are hired!DummyWinnerAgent camel-mailwebserviceproduction_agentdistribution_agentdelivery_agent in Listing 1. The from tag signalises the consumer part of the route, and to signalises
the producer. In this case, the consumer address is the name of the dummy agent to which
the message had been sent (in this example, customer), and the producer address is the
authorisation token followed by the chatId option.
1 <route>
2
3
4 </route>
<from uri="jason:DummyCustomerAgent"/>
<to uri="telegram:bots/sometoken?chatId=-364531"/>
Listing 1. Example of camel-jason route definition
In this example, the delivery agent also uses the camel-artifact in order to ob-
tain the delivery's position from a MQTT server. The route, in Java, is shown in List-
ing 2. When the position is published on the topic of interest (latLong) the route
redirects it to the artifact by specifying its name (TrackedArtifact) and the oper-
ation (giveDistance) as headers. Here, we are assuming that the calculations will be
done by the artifact, but they could be done in the route through a transformation, before
sending to the artifact.
1 from( "mqtt : foo? host=tcp://broker & subscribeTopicName=latLong" )
2
.setHeader( "ArtifactName" , constant ( "TrackedArtifact" ) )
.setHeader( "OperationName" , constant ( "giveDistance" ) )
3
4 .to( "artifact : cartago" );
Listing 2. Example of camel-artifact route definition
5. Related research
In this section, we went over works that have addressed agent technology in an integrat-
ing context. Maturana and Norrie [Maturana and Norrie 1996] have proposed a mediation
and coordination tool for MAS. They have used mediator agents as manufacturing coor-
dinators. Following similar idea, Olaru et al. [Olaru et al. 2013] have developed an agent-
based middleware, which creates a sub-layer of application layer that allows agents to
mediate context-aware exchange of information among entities. We think an autonomous
entity as middleware may increase complexity and compromise performance. Instead of
creating some kind of hierarchy, our approach gives connectivity power to MAS entities.
Leading industrial suppliers are also providing solutions using agents such
as the Agent Development Environment (ADE), designed by Rockwell Automa-
tion [Tich´y et al. 2012]. It provides connectivity with common shop floor devices and
supports the development of agents. The limitation we have seen regards especially con-
nectivity with all sorts of entities (e.g. IoT sensors and mobile devices, ERP and other
software etc), which in our case is provided by Camel.
Other research address the combination of MAS and Service-Oriented Ar-
chitecture (SOA). One way to achieve this merge is based on the creation
of a proxy function to provide interoperability between MAS and SOA, as
found in [Nguyen and Kowalczyk 2005, Shafiq et al. 2005, Greenwood and Calisti 2004,
Fayc¸al et al. 2010]. Another way is by implementing services as agents as we found in
[Mendes et al. 2009, Tapia et al. 2009, Carrascosa et al. 2009, Argente et al. 2011]. The
approaches using SOA are more mature to be applied in practice. The ones that agen-
tified the services have also the advantage to use MAS background, i.e., using ACL
messages they are able to use interaction protocols. In these studies integration is usu-
ally done through specific APIs and they lack differentiation over autonomous and non-
autonoumous entities, and interoperability with heterogeneous entities, both aspects in-
crease development complexity.
Vrba et al. [Vrba et al. 2014] propose a gateway for wrapping an MAS as a ser-
vice to be used as a loosely coupled software component into the Enterprise Service Bus
(ESB). This gateway transforms agent messages to ESB messages and vice versa, en-
abling communication between agents and ESB services. This solution is closely to ours,
only lacking support to the A-E approach, where interaction is based on agents perceiv-
ing and acting upon artifacts in the environment. The indiscriminately agentification may
increase complexity and affect performance.
Cranefield and Ranathunga [Cranefield and Ranathunga 2013] developed a
camel-agent component for Jason agents. It is very similar to our developed camel-jason
component. Essentially, the difference is that we are embedding Apache Camel since
our component works as an infrastructure, being transparent to the agents. In their work,
Jason was actually embedded in an Apache Camel project where agents were smoothly
placed in containers.
6. Conclusion
In this paper, we introduced two Camel components aiming the integration of MAS with
external entities: camel-jason and camel-artifact. The former integrates agents with exter-
nal entities modelled as agents. The latter integrates agents and external entities modelled
as artifacts. The decision of which component to adopt for each entity depends on the
characteristics of the external entity and the MAS developer can choose the most suitable
component. For instance, he/she is not obliged to "agentify" every external entity, even
those that do not have agent properties.
The two components introduced in this paper, along with the communication in-
frastructure provided by Camel and its existing components, makes the integration be-
tween MAS and different entities simpler. Issues related to interoperability, routing, and
data transformations are partially solved in the camel routes. Another advantage of using
such components is that the agent program does not need to deal with integration issues.
Agents continue to interact only with another agents and artifacts.
Finally, this is an ongoing work. In a future step we intend to compare our ap-
proach with related works and to evaluate other aspects of using the developed Camel
components to integrate MAS and external entities, such as the impact on the perfor-
mance, security, openness, scalability, among others.
References
Argente, E., Botti, V., Carrascosa, C., Giret, A., Julian, V., and Rebollo, M. (2011). An
abstract architecture for virtual organizations: The thomas approach. Knowledge and
Information Systems, 29(2):379 -- 403.
Bellifemine, F., Bergenti, F., Caire, G., and Poggi, A. (2005). Jade -- A Java Agent
Development Framework, pages 125 -- 147. Springer US, Boston, MA.
Boissier, O., Bordini, R. H., Hubner, J. F., and Ricci, A. (2019). Dimensions in program-
ming multi-agent systems. The Knowledge Engineering Review, 34:e2.
Carrascosa, C., Giret, A., Julian, V., Rebollo, M., Argente, E., and Botti, V. (2009). Ser-
vice oriented mas: an open architecture. In Proceedings of The 8th International Con-
ference on Autonomous Agents and Multiagent Systems-Volume 2, pages 1291 -- 1292.
International Foundation for Autonomous Agents and Multiagent Systems.
Cranefield, S. and Ranathunga, S. (2013). Embedding Agents in Business Processes
Using Enterprise Integration Patterns. pages 97 -- 116.
Fayc¸al, H., Habiba, D., and Hakima, M. (2010). Integrating legacy systems in a SOA
using an agent based approach for information system agility. 2010 International
Conference on Machine and Web Intelligence, ICMWI 2010 - Proceedings, pages 338 --
343.
Greenwood, D. and Calisti, M. (2004). Engineering web service-agent integration. In
2004 IEEE International Conference on Systems, Man and Cybernetics (IEEE Cat.
No. 04CH37583), volume 2, pages 1918 -- 1925. IEEE.
Hohpe, G. and Woolf, B. (2003). Enterprise Integration Patterns: Designing, Building,
and Deploying Messaging Solutions. Addison-Wesley Longman, USA.
Ibsen, C. and Anstey, J. (2010). Camel in Action. Manning Publications, USA, 1st edition.
Maturana, F. P. and Norrie, D. H. (1996). Multi-agent mediator architecture for distributed
manufacturing. Journal of Intelligent Manufacturing.
Mendes, M., Electric, S., Restivo, F., Colombo, A. W., and Electric, S. (2009). Service-
oriented Agents for Collaborative Industrial Automation and Production Systems.
2744(August).
Nguyen, X. T. and Kowalczyk, R. (2005). Enabling agent-based management of web
services with WS2JADE. Proceedings - International Conference on Quality Software,
2005:407 -- 412.
Olaru, A., Florea, A. M., and El Fallah Seghrouchni, A. (2013). A context-aware multi-
agent system as a middleware for ambient intelligence. Mobile Networks and Applica-
tions, 18(3):429 -- 443.
Omicini, A., Ricci, A., and Viroli, M. (2008). Artifacts in the A&A meta-model for multi-
agent systems. Journal of Autonomous Agents and Multi-Agent Systems, 17(3):432 --
456.
Ricci, A., Viroli, M., and Omicini, A. (2006). Programming MAS with artifacts. Lecture
Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence
and Lecture Notes in Bioinformatics), 3862 LNAI:206 -- 221.
Roloff, M., Amaral, C., Stivanello, M., and Stemmer, M. (2016). Mas4ssp: A multi-agent
reference architecture for the configuration and monitoring of small series production
lines. In INDUSCON.
Shafiq, M. O., Ali, A., Ahmad, H. F., and Suguri, H. (2005). Agentweb gateway-a mid-
dleware for dynamic integration of multi agent system and web services framework.
In 14th IEEE International Workshops on Enabling Technologies: Infrastructure for
Collaborative Enterprise (WETICE'05), pages 267 -- 268. IEEE.
Tapia, D. I., Rodr´ıguez, S., Bajo, J., and Corchado, J. M. (2009). Fusion@, a soa-based
multi-agent architecture. In International Symposium on Distributed Computing and
Artificial Intelligence 2008 (DCAI 2008), pages 99 -- 107. Springer.
Tich´y, P., Kadera, P., Staron, R. J., Vrba, P., and Mar´ık, V. (2012). Multi-agent system
design and integration via agent development environment. Engineering Applications
of Artificial Intelligence.
Vieira, R., Wooldridge, M., and Bordini, R. H. (2007). On the Formal Semantics of
Speech-Act Based Communication in an Agent-Oriented Programming Language.
29:221 -- 267.
Vrba, P., Fuksa, M., and Klima, M. (2014). JADE-JBossESB Gateway: Integration of
Multi-Agent System with Enterprise Service Bus. 2014 Ieee International Conference
on Systems, Man and Cybernetics (Smc), pages 3663 -- 3668.
|
1605.02652 | 1 | 1605 | 2016-05-09T16:50:25 | Global network cooperation catalysed by a small prosocial migrant clique | [
"cs.MA",
"cs.SI",
"physics.soc-ph"
] | Much research has been carried out to understand the emergence of cooperation in simulated social networks of competing individuals. Such research typically implements a population as a single connected network. Here we adopt a more realistic premise; namely that populations consist of multiple networks, whose members migrate from one to another. Specifically, we isolate the key elements of the scenario where a minority of members from a cooperative network migrate to a network populated by defectors. Using the public goods game to model group-wise cooperation, we find that under certain circumstances, the concerted actions of a trivial number of such migrants will catalyse widespread behavioural change throughout an entire population. Such results support a wider argument: that the general presence of some form of disruption contributes to the emergence of cooperation in social networks, and consequently that simpler models may encode a determinism that precludes the emergence of cooperation. | cs.MA | cs |
Global Network Cooperation Catalysed by
a Small Prosocial Migrant Clique
Steve Miller1 and Joshua Knowles2
1 School of Computer Science, University of Manchester, Manchester, UK
2 School of Computer Science, University of Birmingham, Birmingham, UK
[email protected]
Abstract. Much research has been carried out to understand the emer-
gence of cooperation in simulated social networks of competing individ-
uals. Such research typically implements a population as a single con-
nected network. Here we adopt a more realistic premise; namely that pop-
ulations consist of multiple networks, whose members migrate from one
to another. Specifically, we isolate the key elements of the scenario where
a minority of members from a cooperative network migrate to a network
populated by defectors. Using the public goods game to model group-
wise cooperation, we find that under certain circumstances, the concerted
actions of a trivial number of such migrants will catalyse widespread be-
havioural change throughout an entire population. Such results support
a wider argument: that the general presence of some form of disruption
contributes to the emergence of cooperation in social networks, and con-
sequently that simpler models may encode a determinism that precludes
the emergence of cooperation.
Keywords: Evolution of cooperation · Evolutionary game theory · Pub-
lic goods game · Complex networks
1
Introduction
A considerable amount of scientific work has been undertaken to explain the
apparently paradoxical existence of cooperative behaviour in a world defined by
the competitive basis of natural selection [1]. The question of how cooperation
may emerge within a competitive environment is, by definition, predicated on
cooperation being originally absent from the population. On such a basis, the
original appearance of cooperation occurs as a random event, more specifically,
a mutant behaviour in (rare) individual(s). We then consider whether such a
mutation will be extinguished, or will achieve fixation throughout a population.
Within investigations of network-reciprocated cooperation [2,3], models which
abstract social networks to test mechanisms for the emergence of cooperation
broadly follow approaches (implicitly) of this nature (see [4] for a review of such
investigations).
The overwhelming majority of research studies in this field have considered
a population to be one single connected network. However in the real world,
multiple (relatively) discrete dynamic networks exist within populations, and
at times, members of one social network may migrate to another. This is an
aspect of cooperation in real-world scenarios which requires understanding, yet
has thus far received little attention. In the work that follows, we isolate the
key elements of such a scenario: namely, we have a primary network of interest,
predicated on defector behaviour, and we consider the arrival of a very small
group of connected individuals that have emigrated from a cooperative network.
Our investigations here also derive from a second motivating principle. In
earlier work [5], we have described how population size fluctuation has a positive
impact, in promoting the emergence of cooperation in networks. Commenting on
this (ibid.), we suggested the possibility that the observed effect may be viewed
as a generalised response to perturbation of networks, and that population size
fluctuation may be only one way, amongst several, of perturbing a network to
thus yield similar results. This notion hints at a potential issue: that models
of cooperation which are overly deterministic, or lacking in noise, may preclude
the cooperative phenomena we seek to investigate. In the work that follows we
consider whether our findings add further support to this thinking.
2 Background
Here we highlight a few key elements of game theory relevant to this work. We
then briefly consider existing research forming the basis for our investigations.
Within the context of evolutionary game theory, a variety of games are used
to model social behaviours. A model of particular interest for investigating co-
operation is the public goods game (PGG), otherwise referred to as the tragedy
of the commons [6] or the n-person prisoner's dilemma. This game, being based
on group-wise rather than pair-wise behaviour, is arguably more analogous to
the complexity of real-world social interactions, than the standard prisoner's
dilemma (PD), which only models interactions between paired individuals [7].
In the PGG, each participant can choose to contribute, or not, a fixed amount
to a central 'pot'. This pot is then increased by a multiplier and redistributed
amongst all participants, regardless of whether they contributed. The rational
analysis of this game demonstrates that the selfish choice (defection), is the op-
tion which maximises an individual's payoff, however if all individuals exercise
the same rationality, none will contribute and the public good will be minimised,
hence we have a 'social dilemma'. Whilst the rational analysis predicts tragedy,
real-world examples of cooperation (contributing to the public good) are abun-
dant. It is this discrepancy between game theoretic predictions and empirical
findings which research attempts to redress.
The PGG can be implemented within evolving social networks [8], using an
approach where each member of the network in turn, initiates a PGG within a
group which consists of the individuals it is directly connected to -- its 'neigh-
bourhood'. Any given individual in the network will be a neighbour of several
other nodes, hence in addition to the PGG that a particular node initiates itself,
it will also be a participant in PGGs initiated by others. It is this participa-
tion of an individual in multiple games with multiple opponents, i.e. group-wise
interaction, which differentiates the PGG from its cousin in game theory -- the
prisoner's dilemma (PD). In the PD, an individual is able to retaliate or recipro-
cate in response to their partner's behaviour. In the PGG however, participants
are not able to effectively target retaliation directly against defectors, since such
retaliation (i.e. not contributing to the public good) harms cooperator and de-
fector neighbours equally. The classical result for the PGG is that cooperation
becomes less likely as neighbourhood size increases. This result can be appreci-
ated intuitively, by considering that the more the neighbourhood size increases,
(i.e. the closer it gets to having all members of the network participating), the
more the game approximates the mean field scenario, where defection is the Nash
equilibrium [9].
The above approach has been extended to demonstrate the emergence of co-
operation, amongst evolving populations of individuals playing PGG, in dynamic
randomly growing networks [10]. This development differs from earlier work in
its use of two evolutionary elements, rather than one. The two elements are:
1. Strategy updating: This is the primary evolutionary mechanism, present in [8]
and common to the majority of evolutionary game theoretic models used to
investigate cooperation in networks. It represents intrinsic effects within the
population, specifically, direct competition between two competing neigh-
bours. This mechanism's effect is directly responsible for the spread of those
strategies which confer greater fitness upon individuals. It does not however,
in any way, affect the network topology.
2. Population size fluctuation: This secondary evolutionary mechanism [11] rep-
resents widespread 'environmental' effects that are explicitly extrinsic to the
population. In the real world, examples might be disease, predation, food
shortages, drought, many of which may be seasonal. Here a proportion of
the less fit members of a society are periodically 'killed off'. Specifically,
in the case of our implementation, individuals are removed from the pop-
ulation, along with the positions they occupied within the network due to
their connections. This (fitness-based) process causes changes in the net-
work topology, but it does not implement the spread of behaviours from one
individual to another.
In the following, we investigate how a variety of network simulations, all pred-
icated on originally non-cooperative behaviour, are affected by the arrival of a
very small (n ≤ 3) group of cooperative migrants. We initially describe, in de-
tail, the implementation of our models. We then provide 'behaviour profiles' for a
range of network scenarios and growth mechanisms, followed by deeper scrutiny
of phenomena within the actual simulations that are of particular interest.
3 Methods
Our work is based on methodology presented in [8,10,12]. We here give a full
description of our approach for completeness.
Our model describes agents located at the nodes of networks. Each node
in the network has a 'neighbourhood', defined by the nodes its edges connect
to. A PGG occurs for each neighbourhood and hence a network of N nodes
will result in N PGGs. Each agent in the network has a behaviour encoded
by a 'strategy' variable: 'cooperate' or 'defect', which determines whether it
contributes to PGGs, or not, respectively.
The general outline of the evolutionary process, for one generation, is as
follows:
1. Play public goods games: In a round-robin fashion, each agent initiates a
PGG involving its neighbours. An agent's fitness score is the sum of payoffs
from all the individual PGGs that it participates in.
2. Update strategies: Selection occurs. Agents with low scores will have their
strategies replaced, on a probabilistic basis, by comparison with the fitness
scores of randomly selected neighbours.
3. Remove nodes: If the network has reached the nominal maximum size, it is
pruned by a tournament selection process that removes less fit agents.
4. Grow network : A specified number of new nodes are added to the network,
each connecting to m randomly selected distinct existing nodes via m edges.
In the following, we provide more detail on each of the four steps:
Play public goods games. Each node of the network, in turn, initiates a PGG.
Within a single PGG, all cooperator members of a neighbourhood contribute a
cost c to 'the pot'. The resulting collective investment I is multiplied by r, and
rI is then divided equally amongst all members of the neighbourhood, regardless
of strategy.
Since an agent contributes a cost c to each game they participate in, their
overall contribution, in one generation, is therefore c(k+1) where k is the number
of neighbours (degree). The single game individual payoffs of an agent x are
given by the following equations, for scenarios where x is a defector (PD) and a
cooperator (PC) respectively:
PD = crnc/(kx + 1) ,
PC = PD − c ,
(1)
(2)
where c is the cost contributed by each cooperator, r is the reward multiplier,
nc is the number of cooperators in the neighbourhood based around x, and kx
is the degree of x.
Update strategies. Each node i selects a neighbour j at random. If the fitness
of node i, fi is greater or equal to the neighbour's fitness fj, then i's strategy is
unchanged. If the fitness of node i, fi is less than the neighbour's fitness, fj, then
i's strategy is replaced by a copy of the neighbour j's strategy, according to a
probability proportional to the difference between their fitness values. Thus poor
scoring nodes have strategies displaced by those of more successful neighbours.
Hence, at generation t, if fi(t) < fj(t) then i's strategy is replaced with that
of the neighbour j with the following probability:
fj(t) − fi(t)
ΠUi(t) =
f d max(ki(t), kj(t))
,
(3)
where ki and kj are degrees of node i and its neighbour j respectively. The pur-
pose of the denominator is to normalise the difference between the two nodes,
with f d max(ki(t), kj(t)) representing the largest achievable fitness difference
between the two nodes given their respective degrees. In the absence of a mathe-
matical approach to calculate this, we run simulations for all 4 combinations (of
the 2 strategy types at the 2 nodes), to establish maximum possible difference.
Grow network. We add 10 new nodes (7 on the first generation), with randomly
allocated strategies, per generation. Each new node uses m edges to connect to
existing nodes. Duplicate edges and self-edges are not allowed. The probabil-
ity Π(t) that an existing node i receives one of the m new edges is given by
the following equations, for random attachment (RA), degree-based preferen-
tial attachment (PA), and fitness-based evolutionary preferential attachment
(EPA) [12], respectively:
ΠRAi(t) =
1
N (t)
,
(4)
where N (t) is the number of nodes available to connect to at time t in the existing
network. (Given that in our model each new node extends m = 2 new edges, and
multiple edges are not allowed, N is therefore sampled without replacement.)
ΠP Ai(t) =
,
(5)
(cid:80)N (t)
ki(t)
j=1 (kj(t))
where ki(t) is the degree of an existing node i and N (t) is the number of nodes
available to connect to at time t in the existing network.
ΠEP Ai(t) =
,
(6)
(cid:80)N (t)
1 − + fi(t)
j=1 (1 − + fj(t))
where fi(t) is the fitness of an existing node i and N (t) is the number of nodes
available to connect to at time t in the existing network. The parameter ∈ [0, 1)
is used to adjust selection pressure. (We used = 0.99 for 'strong' EPA.)
Growth only occurs at times when the network is below a nominal maximum
size (we used Nmax = 1000 nodes). For all added nodes, other than migrants,
we set m = 2.
Remove nodes (for fluctuation simulations). Whenever the network achieves
or exceeds the nominal maximum size, it is pruned by a percentage X. This is
achieved by tournament selection using a tournament size equivalent to 1% of
the network. Tournament members are selected randomly from the network. The
tournament member having the least fitness is the 'winner' and is added to a
short list of nodes to be deleted. Tournament selection continues until the short
list of X% nodes for deletion is fully populated.
The nodes on the short list (and all of their edges) are removed from the
network. Any nodes that become isolated from the network as a result of this
process are also deleted. (Failure to do this would result in small numbers of
single, disconnected, non-playing nodes, having static strategies and zero fitness
values.) When there are multiple nodes of equivalent low fitness value, oldest
nodes are deleted first. Where X = 0, no deletions occur; in this case, on reach-
ing maximum size, the network structure becomes static.
Migrant clique attachment. At generation 300, the migrant group connects
to the existing primary network. Our migrant groups are small complete net-
works i.e. cliques, consisting of between 1 to 3 nodes (specific details in results
section), all having cooperator strategies. Initial connection to the primary net-
work is via only one of the nodes in the clique. This node extends either 1 or
2 edges (specific details in results section) to existing network nodes chosen at
random. Once connected, the migrants are treated as a part of the primary net-
work and are exposed to all elements of the evolutionary process described above.
General simulation conditions. In networks grown from founder members,
initial nodes were populated with defector strategies. In 'pre-existing' networks,
all nodes were populated with defectors. Strategy types of subsequently added
nodes were allocated independently, uniformly at random (cooperators and de-
fectors with equal probability). All networks had an overall average degree of
approximately k = 4, hence an average neighbourhood size of g = 5 (since
neighbourhood includes self). Simulations were run until 20,000 generations. Fi-
nal 'fraction of cooperators' values we use are means, averaged over the last 20
generations of each simulation. Each simulation consisted of 25 replicates. We
used a shrinkage value of X = 2.5% for all fluctuation simulations. Simulation
data is recorded after step 2 (Update strategies).
4 Results and Discussion
We initially present our results using an approach common for investigations in
this field. We aggregate data from multiple differing sets of simulations, plotting
final fraction of cooperators against the variable, η, which is the PGG reward
multiplier normalised with respect to the average neighbourhood size in the
network. In Fig. 1a we present such 'behaviour profiles' for results from the
'simplified scenario' of pre-existing networks. These networks have initially ran-
dom graph topology [13] and are initialised entirely with defectors. Figure 1b,
illustrates the 'more realistic' scenario where we consider networks grown from
their origins, in this case from 3 founder defector members. In both network
scenarios we provide profiles for the three attachment mechanisms of RA, PA
and EPA.
Fig. 1. Behaviour profile plots illustrating the impact of a migrant cooperator clique
on the emergence of cooperation for three different attachment mechanisms: RA, PA
and EPA. a) shows pre-existing networks having initially random graph topology and
initialised entirely with defectors. b) shows networks grown from 3 defector founders.
Final fraction of cooperators present is plotted against η (the PGG reward multiplier r
normalised with respect to average neighbourhood size, g = 5). Migrant cliques consist
of 3 connected cooperators, one of which attaches to the existing network randomly by
2 nodes. Green lines with circle markers are simulations featuring migrants. Black lines
with square markers are controls (no migrants). Solid lines represent simulations that
are fixed in their network topology. (In the case of b, topology becomes fixed upon
population achieving maximum size.) Dashed lines represent fluctuating simulations.
For the simpler scenario of pre-existing networks, initialised with all defectors
(Fig. 1a) and having a fixed network size, we naturally observe zero cooperation
(solid black lines) for all attachment mechanisms. In comparison, the migrant
scenario (solid green line) precipitates cooperation once the temptation of the
reward achieves a particular threshold (η > 0.7). In the case of fluctuating
network size, we see that the migrants promote higher levels of cooperation than
those seen in their absence (compare green dashed with black dashed lines),
except in the case of EPA, where levels of cooperation have already been elevated
by the increased network heterogeneity associated with this mechanism (see [14]
for detailed information on the role of network heterogeneity in cooperation).
When we consider the more complex scenario of networks grown from founders
(Fig. 1b), we see that our earlier findings still hold. Again, above a reward thresh-
old (η > 0.6), the arrival of the migrants promotes widespread cooperation. We
see this effect for networks that become static on reaching specified maximum
size and also in those that fluctuate in size thereafter. We note that in the case
of fluctuating models, we see little difference in final outcomes when comparing
FractionuofuCooperatorsη1.00.80.60.40.200.20.40.60.81.00.20.40.60.81.00.20.40.60.81.0FractionuofuCooperatorsηa)uPre-existingunetworksb)uNetworksugrownufromufounders1.00.80.60.40.200.20.40.60.81.00.20.40.60.81.00.20.40.60.81.0RAPAEPARAPAEPA pre-existing networks with those grown from founders (compare corresponding
coloured dashed lines in Figs. 1a and b). As described in earlier research [11], the
fluctuation mechanism, by deleting low fitness nodes from within the network,
can overcome the limitations of 'fossilised' (zero-fitness, defector-dominated) re-
gions of the network in a manner that is not achievable by strategy updating
between neighbours. Importantly, cooperation can be supported by a fluctuat-
ing population size, without the requirement for highly heterogeneous network
topology: the fluctuation mechanism drives networks to a topology that has only
moderately heterogeneous connectivity (in the form of a compressed exponential
degree distribution) [10].
Whilst the behaviour profile plots above allow us to neatly characterise and
compare different experimental simulations, they describe derived data which
for the most part is of limited interest, whilst potentially masking more inter-
esting phenomena. More specifically, as the value of the reward variable (η) is
maximised/minimised, the dilemma becomes diminished and the dominant be-
haviour of populations becomes consistent and highly predictable. We suggest
that in presenting abstracted representations of real-world scenarios, such regions
of the behaviour profiles are of limited relevance.
It is the mid-range values of the reward variable that represent the social
dilemma in its strongest form. We suggest that these regions are of particular
importance in investigating the emergence of cooperation, since they represent
the much more realistic challenge faced in nature by individuals attempting
to balance cost versus reward, and in addition, where noise may likely be a
confounding or contributory factor. Where we see transitions in population be-
haviour, where a mixture of competing behaviours exists, where the choice of
cooperate or defect is not clear cut, and where noise may be present -- these are
the areas we are interested in.
We now explore the behaviour of our populations, in these regions of interest,
by focusing on the behaviour of replicate simulations as they transition from
defection to cooperation. From Figs. 1a and b, we see the widest variety of
outcomes in the region approximately around where η = 0.8. Figure 2 illustrates
individual time plots of simulations based on this value, for the simplified case
of pre-existing networks initialised with defectors. The plots show simulations
with the effects of fluctuation and immigration enabled, disabled, and acting in
concert. We summarise from inspection of these plots that:
i The fluctuation mechanism on its own enables a majority of replicates to
transition to cooperation. Similar levels of cooperation are achieved by all
of those replicates that transition. Transition times however remain variable
with some replicates failing to transition over the time period studied.
ii The isolated effect of migrant arrival drives higher levels of cooperation
amongst replicates. In the case of this effect though (in contrast to our previ-
ous observation), it is the levels of cooperation achieved which are variable.
iii The combined impact of migrants together with fluctuating population size,
results in all replicates transitioning to cooperation with consistency in both
final levels of cooperation achieved, and also in transition times (all replicates
transition within 200 generations of the arrival of the migrants).
Fig. 2. Simulation time plots (25 replicates) illustrating the effects of migrant clique
arrival and fluctuation, in pre-existing random networks initialised with defectors, with
η = 0.8. Plots show number of cooperators over 20,000 generations. Migrant groups are
complete networks of 3 cooperator nodes, 1 of which connects to 2 randomly selected
existing network nodes. Network growth is by random attachment. All other details
are as described in Methods section. Number of replicates transitioned to cooperation
is shown in circle inset.
In Fig. 3 we illustrate similar time plots, in this case for the more com-
plex scenario featuring networks grown from founder populations of 3 defectors.
We observe that the findings seen earlier, for the simplified case of pre-existing
networks, still hold: fluctuation alone promotes consistent levels of increased
cooperation albeit with variable transition times; migrants alone promote co-
operation albeit to varying levels; the combination of cooperator migrants and
MIGRANTS: OFFMIGRANTS: ONFLUCTUATION: OFFFraction of Cooperators Generation No. (x 103)1.00.80.60.40.20051015201.00.80.60.40.2005101520FLUCTUATION: ON0131425 fluctuation brings consistency to both transition times and levels of cooperation
achieved.
Fig. 3. Simulation time plots (25 replicates) illustrating the effects of migrant clique
arrival and fluctuation, for networks grown from 3 defector founders, with η = 0.8.
Plots show number of cooperators over 20,000 generations. Migrant groups are complete
networks of 3 cooperators nodes, 1 of which connects to 2 randomly selected existing
network nodes. Network growth is by random attachment. All other details are as
described in Methods section. Number of replicates transitioned to cooperation is shown
in circle inset.
These findings are also robust to attachment mechanisms. For both of the
network models illustrated above, in addition to random attachment (as repre-
sented in Figs. 2 and 3), the same observations also held when tested using both
degree-based attachment (PA), and fitness-based attachment (EPA).
The ability of the of the migrant clique to invade defector networks appears
to arise from benefits conferred on the connecting migrant by the 'back-up'
MIGRANTS: OFFMIGRANTS: ONFLUCTUATION: OFFFraction of Cooperators Generation No. (x 103)1.00.80.60.40.20051015201.00.80.60.40.2005101520FLUCTUATION: ON3152325 provided from its fellow migrants. These back-up migrants are initially immune
to both strategy updating and the impact of defectors in reducing their payoff
values (being as they are initially not directly connected to the network). The
back-up migrants can boost the payoff (fitness) of a connecting migrant, so that
during strategy updating, it can thus readily convert the existing network node it
connects to, into a cooperator. Beyond initial possible payoff calculations, which
can be established analytically, it becomes harder to pin down the details of the
further spread of cooperation. However, it is clear from our investigations that
in the case of migrant-triggered cooperation, it is this back-up which is key.
What is particularly interesting here, is just how small the migrant group can
be, whilst still being able to precipitate the emergence of cooperation through
the entire population. The previous simulations were based on migrant groups of
3 connected individuals, one of which extends 2 connections to random existing
members of the network. In additional work, we have reduced the size of the mi-
grant group to 2 individuals, of which one connects only 1 edge to an existing net-
work node. Tested at the same η (= 0.8), on pre-existing defector-populated ini-
tially random networks, and on networks grown from defector founders (growth
by RA in both cases), our previous findings still hold. (Time plots were highly
similar to those shown in Figures 2 and 3, with the only difference that a delay
in transition was observed infrequently, e.g. 1 or 2 replicates out of 25, for those
simulations combining both migration and fluctuation.) On further reduction to
1 node (extending either 1 or 2 edges), our general findings no longer hold. This
outcome is entirely expected, as this situation is now no different to the standard
attachment process by which all new individuals routinely connect -- 1 node, 2
edges, i.e. no back-up.
These findings based on adjustments to the migrant clique highlight a poten-
tial source of concern regarding models of cooperation in networks, namely that
widely differing outcomes may arise from seemingly small differences in simu-
lation parameters: We can reduce our migrant mechanism to a point where it
appears very similar (2 nodes, 1 edge) to the mechanism by which nodes rou-
tinely attach during network growth (1 node, 2 edges). Given such similarity,
and noting that the migrant effect happens only once in a simulation, whilst
new nodes are added repeatedly in the fluctuation model, we might be inclined
to therefore assume that results due to the migrant clique arrival would be triv-
ial relative to those arising from fluctuation. However, we see in our results that
the isolated, seemingly trivial, migrant event clearly brings about an additional
change to populations, which is not achieved in its absence. The small differ-
ence between these two very similar mechanisms results in markedly different
behavioural dynamics. Importantly, despite their apparent similarities, the at-
tachment mechanism used for routine network growth clearly cannot create the
additional opportunities for cooperation that the migrant clique's arrival can
enable.
These results combined with findings of previous research, reinforce our belief
that fluctuations in the network, or migrant cliques, or alternative mechanisms
to perturb the system, bring an added dimension to models of cooperation in
networks that simpler mechanisms fail to provide: It is these noisy perturbations
of the network that disrupt the 'status quo' and catalyse the spread of coopera-
tion throughout the population. If this assumption is correct then there is a risk
that simpler, more deterministic models of cooperation in networks may lack the
disruptive elements that promote cooperation and may thus preclude or impede
its emergence.
5 Conclusion
Using various models of cooperation, based on the public goods game, we have
investigated a scenario where individuals migrate, from a cooperative network,
to join one that does not demonstrate cooperation. Under certain conditions,
notably around the region where the social dilemma is at its strongest, we
find quite striking results: The effect of a few concerted migrants catalyses a
marked behavioural change, precipitating the widespread emergence of coopera-
tion throughout the entire population. Of particular interest is our finding that
the migrant group size can be extremely small and needs only to form one initial
connection in order to initiate a marked response. The actions of a seemingly
trivial group of concerted cooperators initiate changes throughout a population
that is orders of magnitude larger than the migrant group.
We have hypothesised that perturbation, in the form of population size fluc-
tuation, and also in the form of invading migrants, can promote cooperation.
We have demonstrated this to be the case for both of these effects in isolation,
and to a greater extent, in concert. Clearly other methods, or combinations of
methods, for perturbing or disrupting networks exist that may yield similarly
interesting results.
Our results reinforce previous work proposing that perturbations of networks,
or possible alternative forms of disruption, are an important contributory feature
in the emergence of cooperation. Taken generally, such observations suggest the
potential for oversimplified or strictly deterministic models of cooperation in
social networks, to limit or exclude the phenomena they seek to investigate. We
highlight, in particular, that from a combination of two mechanisms studied here,
there emerged a consistency in outcome that is unlikely to have been anticipated
from studying simpler models of each mechanism in isolation.
Acknowledgements This work has been funded by the Engineering and Phys-
ical Sciences Research Council (Grant reference number EP/I028099/1).
References
1. Axelrod, R., Hamilton, W.D.: The evolution of cooperation. Science 211 (1981)
1390 -- 1396
2. Nowak, M.A., May, R.M.: Evolutionary games and spatial chaos. Nature 359
(1992) 826 -- 829
3. Nowak, M.A.: Five rules for the evolution of cooperation. Science 314 (2006)
1560 -- 1563
4. Perc, M., Szolnoki, A.: Coevolutionary games: a mini review. BioSystems 99
(2010) 109 -- 125
5. Miller, S., Knowles, J.: Population fluctuation promotes cooperation in networks.
Sci. Rep. 5 (2015)
6. Hardin, G.: The tragedy of the commons. Science 162 (1968) 1243 -- 1248
7. Perc, M., G´omez-Gardenes, J., Szolnoki, A., Flor´ıa, L.M., Moreno, Y.: Evolution-
ary dynamics of group interactions on structured populations: a review. Journal
of The Royal Society Interface 10 (2013) 20120997
8. Santos, F.C., Santos, M.D., Pacheco, J.M.: Social diversity promotes the emergence
of cooperation in public goods games. Nature 454 (2008) 213 -- 216
9. Nash, J.: Non-cooperative games. Annals of mathematics (1951) 286 -- 295
10. Miller, S., Knowles, J.: The emergence of cooperation in public goods games on
randomly growing dynamic networks. In Squillero, G., Burelli, P., eds.: Applica-
tions of Evolutionary Computation: 19th European Conference, EvoApplications
2016, Porto, Portugal, March 30 - April 1, 2016, Proceedings, Part I. Springer
International Publishing, Cham (2016) 363 -- 378
11. Miller, S., Knowles, J.: A minimal model for the emergence of cooperation in ran-
domly growing networks. In: Proceedings of the European Conference on Artificial
Life 2015 (ECAL 2015). Volume 13. (2015) 114 -- 121
12. Poncela, J., G´omez-Gardenes, J., Flor´ıa, L.M., S´anchez, A., Moreno, Y.: Com-
plex cooperative networks from evolutionary preferential attachment. PLoS one 3
(2008) e2449
13. Erdos, P., R´enyi, A.: On random graphs. Publicationes Mathematicae Debrecen
6 (1959) 290 -- 297
14. Santos, F.C., Pacheco, J.M.: A new route to the evolution of cooperation. Journal
of Evolutionary Biology 19 (2006) 726 -- 733
|
1109.5712 | 1 | 1109 | 2011-09-26T20:21:57 | Cooperative Information Sharing to Improve Distributed Learning in Multi-Agent Systems | [
"cs.MA"
] | Effective coordination of agents actions in partially-observable domains is a major challenge of multi-agent systems research. To address this, many researchers have developed techniques that allow the agents to make decisions based on estimates of the states and actions of other agents that are typically learnt using some form of machine learning algorithm. Nevertheless, many of these approaches fail to provide an actual means by which the necessary information is made available so that the estimates can be learnt. To this end, we argue that cooperative communication of state information between agents is one such mechanism. However, in a dynamically changing environment, the accuracy and timeliness of this communicated information determine the fidelity of the learned estimates and the usefulness of the actions taken based on these. Given this, we propose a novel information-sharing protocol, post-task-completion sharing, for the distribution of state information. We then show, through a formal analysis, the improvement in the quality of estimates produced using our strategy over the widely used protocol of sharing information between nearest neighbours. Moreover, communication heuristics designed around our information-sharing principle are subjected to empirical evaluation along with other benchmark strategies (including Littmans Q-routing and Stones TPOT-RL) in a simulated call-routing application. These studies, conducted across a range of environmental settings, show that, compared to the different benchmarks used, our strategy generates an improvement of up to 60% in the call connection rate; of more than 1000% in the ability to connect long-distance calls; and incurs as low as 0.25 of the message overhead. | cs.MA | cs | a f Ai(cid:12)
ia e ige
e Reea
h 24 2005 407{463
S bied 02/05; b ihed 10/05
Ceaive fai Shaig ve Diib ed
eaig i i Age Sye
aha S. D a
a01e
..a
. k
i
h a R. eig
je
..a
. k
ea
.ea e
..a
. k
S
h f E e
i
ad C e S
ie
e
Uiveiy f S ha
igh(cid:12)e d S ha S 17 1B U
Aba
E(cid:11)e
ive
diai f age a
i i aia y bevab e dai i a a j
ha ege
f i age ye eea
h. T adde hi ay eea
he have deve ed e
hi e
ha a w he age ake de
ii baed eiae f he ae ad a
i f he
age ha ae yi
a y ea ig e f f a
hie eaig a gih. evehe e
ay f hee aa
he fai vide a a
a ea by whi
h he e
eay ifai
i ade avai ab e ha he eiae
a be ea. T hi ed we ag e ha
ea
ive
i
ai f ae ifai bewee age i e
h e
hai. weve
i a dyai
a y
hagig evie he a
a
y ad ie ie f hi
i
aed
ifai deeie he (cid:12)de iy f he eaed eiae ad he ef e f he a
i
ake baed hee. Give hi we e a ve ifai haig
ak
ei haig f he diib i f ae ifai. We he hw h gh
a fa aa yi he ivee i he a iy f eiae d
ed ig aegy
ve he wide y ed
f haig ifai bewee eae eighb . eve
i
ai he ii
deiged a d ifai haig i
i e ae b je
ed
eii
a eva ai a g wih he be
hak aegie i
dig ia ig
ad Se TT R i a i aed
a ig a i
ai. Thee die
d
ed
a
a age f eviea eig hw ha
aed he di(cid:11)ee be
hak
ed aegy geeae a ivee f 60 i he
a
e
i ae; f
e ha 1000 i he abi iy
e
g dia
e
a ; ad i
a w a 0.25 f
he eage vehead.
1. d
i
A
ea
ha ege i i age ye AS eea
h i deig e
hai f
diaig age ha have aia ib y a y i
ie ad ia
ae view f
he ye ha hey
a geeae
ie i
ex diib ed b e.
h eig he b e vig e f e age
a i(cid:13) e
e he f ahe whee
hey a
a
vea b e e a e f haab e e
e. Th
diae
ef y he age eed
eae by aiig ea
h he ake bee
hi
e ab
he a
i hey ake.
Thi
eai i ade e diAE
be
a e he idivid a age
1
a y have ei
ed
aabi iy i efig exeive
aia a
iviie d e
1. A di(cid:11)ee hi hy died by AS eea
he i ha f ig
eiive age (cid:12)d i
diib ed b e Takahahi Taaka 2003; Wa h We a 2003. ee hweve we e
i(cid:12)
a y
f
eaive age ad h we ae (cid:12) y i he ea f
eaive diib ed b e vig
ye A Bea ee Badhaw D fee i b
k S i Tae Th 2002; ee
Ea 1988; ai e ee ig 2003.
(cid:13)2005 A A
e F dai. A igh eeved.
D a eig ea
iied ey CU
y
e
i
ai badwidh ad/
i
ai ae
y
ad be
a e he age evie i a y
haa
eied by
i ad edi
ab e
hage whi
h i e
eiae
i adaai f he b e vig
e by
he age.
T ve
e hee b e ad
diae e(cid:11)e
ive y he age eed a e
hai
a
adaive y
h ha a
ie vea i i geeaed. hi
ex i
age
diai i yi
a y baed e
hi e f de ig he ae f he age
ad eab ig a age ake a
i baed hee de D a ea eig
2003; D a Se 2003 ee e
i 2 f e deai . weve give ha he age
a y die
y beve a iied be f he ye hey eed be vided wih
e ifai ab he bevab e ae geeae
h de . T a
hieve hi
we be ieve he age h d hae e f hei kw edge ab hei w
a y beved
ae a a iab e eve f aba
i. Thi kw edge
a he be ed by he e
eivig
age ake e ifed a
i f bee
diai. e ha i hi aa
h
he age
eae by v ai y diib ig ifai i he ye fa
i iae he
b e vig
e. evehe e a
i
a e
e b d
h a iied badwidh
ad ae
y hibi he e f exha ive
i
ai ha a age
d be ade
awae f he a f a he age a a ie. Th he
i
ai be e e
ive
ad ai
i
ae he iia a f ifai ha i e
eay f e(cid:11)e
ive
diai.
Give
h
ai i i a
i
a ib e i
ae geeae de
f a age i he ye ha ae a
ae ad dae. evehe e he age
a
a y geeae eiae f he beved ae ake
diai de
ii. F he
i dyai
a y
hagig ye he age h d have a way f daig hee eiae
ada hei b e vig de
ii wih he
hagig evie f geeaig a iy
i.
w i ay a i
ai eif
ee eaig S Ba 1998 R ha bee
ef y ed E G avi
Weheke 2004; ahadeva a
ha e
k Da Gavi
1997; Tg 2002 geeae
h adaive eiae i dyai
evie ee e
i 2
f a di
i f he a eaive. R e i exeie
e f efig ak deve
a de f he evie. Se
i(cid:12)
a y a eif
ee eaig age e
eive a
eai
\ewad" f akig a a
i i a give ae ha a
a feedba
k idi
ae he a iy f
efa
e agai he
ex de(cid:12)ed by he ae a
i ai. Uig
h ewad R i
aab e f i
eea y geeaig b eiae f he
e f di(cid:11)ee a
i i
di(cid:11)ee ae. S
h eiae vide he age wih a geei
ideede f he b e
aheai
a eaig e
hai ake adaive de
ii i dyai
evie.
ai
a he eaig Waki Daya 1992 vaia f R i wide y ed be
a e i
a w eiae be ea wih havig i kw edge ab he ye dyai
.
Uig eaig wih iab e aiig f he age whee i i a wed eeaed y ake
di(cid:11)ee a
i i di(cid:11)ee ae he
e
evie de
a be eaed f e
wad. weve he a i de yig hi e ae ha he age i ab e beve
a eviea ae ad e
eive a ewad a
ae y f ay a
i ake a ay ae.
B i a
i
a AS hi a i i iib e ea ie be
a e a age
a y
beve a a f he
ee ye. Thi i ie ha i
a y e
eive he e f i
a
i wihi i
a evie ad ha i ay be ab e beve iediae y ad
e
y he a
i hei e(cid:11)e
ake by he age.
408
Shaig fai f Diib ed eaig i AS
T ve
e he abve eied iiai we have deve ed a e(cid:11)e
ive ad eAE
ie
i
ai
ha a e f
eaive age
a e ea he eiae f
beved ae i a dyai
evie. ai
a we ee a ve i
i e f
ak
ei TC ifai haig. hi age ake a
i f vig a
give ak ig hei
e eiae f he ye ae ad he diib e hei
a
ae ifai e ahe y afe he ak i
eed.
i
ai i
a ed d ig he eid f ak
eig. The ifai i haed bewee he
age wh
eae
ee he ak. U e
eivig hi ifai he age
be e y dae hei evi eiae f he ae f he he age.
Thi
i
ee y geei
i
e i i deve ed baed ay dai
b e e
i(cid:12)
a i ee aa yi f e
i 5. weve e
i(cid:12)
ia
e f TC
a be i eeed f a give b e dai. F exa e i e
i 4 we de
ibe
h a ia
e f TC i eeed i a
a ig b e whee he age ae
eiae avai ab e badwidh de.
The fa
ha TC i baed ay dai
2
e
i(cid:12)
a i i ie ha i
a be ed i
eaive i age b e he
ha
a ig ad veify hi we ae
e y dyig i a i
abi iy i a diib ed
fa dee
i a i
ai wih iig iiia e D a eig ea 2005.
The TC
i dii
f he e aive y adad aa
h f daig ei
ae ig he ifai f y eae eighb heeafe efeed a
whi e
eig a ak whi
h f he bai f a fai y f ig
edi
k 1988;
Taeba 2003.
T ehaie hi fa
we
he
ae he a iy f eiae
3
eaed ig he TC i
i e agai ha f . Se
i(cid:12)
a y he
a w
ifai be haed bewee die
eighb y wheea
a w i
fai be haed bewee a
eaive g f age. F hee
e e e ie y
i
ai whi
h heefe ead e dae eiae ha
e
i 5 eab ihe hee fa y.
Thee ha a bee he eea
h She ee Cave 2003; Xiag 1996; X a
ee Zi beei 2001 ha ha died hw haig ifai bewee age
a aid
eaive b e vig. Tyi
a y hee aa
he ea
i
ai a a dii
a f a age vea de
ii akig b e ad hw hw i i
ai aid i
vig he ae e
i 2 ha e deai . B hee ha bee i e i he way f a y
eai
dy f deve ig a
i
ai
ha i bh a
i
a y a i
ab e i
e f i beig baed ea ii
a i ad e(cid:11)e
ive i e f ivig ef
a
e i ea ife AS. wk he he had iveigae a e
i(cid:12)
i
ai
ha ha bh he abve deiab e
haa
eii
.
Agai hi ba
kg d i hi ae we eva ae TC ig w ai aa
he.
Fi by a aheai
a aa yi we deae ha he eiae geeaed by a
a
h ae ideed e dae ha . Se
d he e(cid:11)e
ivee f
i
ea ed ig a i aed diib ed e
e a
ai b e.
ai
a we e
4
2. hi a i
ai he \ae" f a age i he badwidh avai abi iy f i de. e deai f w i
e
i 4.
3. exa e a i
ai dai a de i \eae" eighb ae he de ha ae wihi i
aii age he wih wh i
a die
y
i
ae.
4. We
he e
e a
ai be
a e i i a geei
ak dai wide y ed i a
i
a AS Chaib
daa 1995; C
kb eig 1996; eig a Faai 1998. Theefe we be ieve i i
a eaab e
hi
e e ifai haig aegy. A e iiay eii
a iveigai f
ifai haig e
hai i hi dai
a be f d i he wk f D a Daahaaa G
eig ad ea 2004.
409
D a eig ea
a i aed wie e e ehe ewk whee he age have a
ae badwidh
e
a . C i
ai he ii
baed he TC i
i e ae devied be ed
by he age i hi dai ee e
i 6 f a di
i. The efa
e f hee ae
aed agai w we kw a gih ed f ewk ig: Bya ad ia
ig Bya ia 1993 ad Se ad Ve Tea aiied a e Ta
ii Reif
ee eaig TT R Se Ve 1999. The fe i
he
be
a e i i e f he wide y ed be
hak i eaig baed ewk ig a
i
ai Ca Dig 1998a; ehki Savva 2002; Se 2000. The ae ha
aa
ed aei e e
e y by beig hw be ef i a vaiey f dai Se
2000. e
e we be ieve hee a gih ae eaab e be
hak f eii
a vei(cid:12)
a
i f TC. Theefe ig bh he fa aa yi ad eii
a
ai i AE
ie
be ab e (cid:12) y eab ih he ei f TC.
5
Eii
a die have bee
d
ed ve a wide age f eviea eig by
e e
ig di(cid:11)ee ewk gie ewk ad ad dyai
a y
hagig ad ae.
Thee die idi
ae vey iig e f TC. ai
a we beve baia
ivee i he ae f
ef y
e
ed
a 60 ad he abi iy e
a dia deiai wih high ewk ad e ha 1000; bh a
hieved by
i
ig a
h we
i
ai
i e f ifai eage ae a w a
0.25 by TC
aed he be
hak aegie i he exeiea eig
ed; a e beig aii
a y igi(cid:12)
a a he 95
(cid:12)de
e eve .
The f wig aie
ib i wad adva
ig he ae f he a:
(cid:15) We ag e f he e f ifai haig baed ea ii
a i ive
eaive diib ed b e vig.
(cid:15) We e a
i
ai
ideede f b e e
i(cid:12)
fea e
ak
ei haig f geeaig gd eiae ha eaig age
a e
f bee
eai.
(cid:15) We eab ih ig fa aa yi he advaage f he TC
i geeaig
e a
ae eiae by e ig a e ie y diib i f ifai ha he
ifai haig
.
(cid:15) We deae he e(cid:11)e
ivee f he TC
ig eii
a die i a ee
eaive i age e
e a
ai b e de a wide vaiey f eviea
eig ad agai a age f he aegie.
The eaide f hi ae i gaied a f w. Se
i 2 di
e he geea i
i
e f i age
diai ad f
e he e
i(cid:12)
aa
h f ig a
hie eaig
e
hi e i hi
ex. A he e f
i
ai i hee i aa yed. e
i 3
he
haa
eii
f he
eaive i age ye whi
h we exe ify eea
h
5. We a aeed
ae a gih wih a g ba bad
a e
hai i whi
h a age i e
a bad
a f hei
a ae ifai wheeve hei ae
hage. Thi e
hai wa deiged
veify whehe ye efa
e ive by aiig a ae
hage ifai a age.
evehe e i i iab e be ed i a a
i
a a i
ai d e i exbia eage vehead.
Thi wa vei(cid:12)ed whe we de yed i a a ad a e a i
ai a ig e a
hie wih a d a 2.2 G z
AD e
e ad 2GB ey i whi
h i a f ey he a e gy ad
wih he ighe ad ed i exeie. weve a diib ed i eeai f he bad
a
a gih
d be a ae f f e dy.
410
Shaig fai f Diib ed eaig i AS
ae ied. a
ai a bief de
ii f he ewk a i
ai ha we i ae
eii
a y eva ae
i
ai i
i e. Se
i 4 ee a a iaive ag
e f he ia
e f ig
eaive
i
ai ive eaig. A hw
he TC i
i e i i eeed i he
ex f exa e a i
ai i high ighed.
eve bief de
ii f he i eeai f he be
hak a gih ig
ad TT R i hi a i
ai ae eeed. Se
i 5 ee a fa aa yi f he
advaage f aegy ve he eae eighb
i geeaig e a
ae
eiae. S be e y a deai ed de
ii f he i ai i vided i e
i 6. Se
i 7 de
ibe he efa
e ea e agai whi
h he vai aegie ae
aed.
a aa ye he e f eii
a die. Fia y e
i 8 ee
dig
eak ad idei(cid:12)e ave e f f e wk.
2. Re aed Wk
hi e
i we (cid:12) eview he a j heei
a ad eii
a wk
eaive
AS ha ae deve ed a d he hee f geeaig e iab e eiae f beved
ae f iied iea
i ad adaig de
ii i ee dyai
evie
e
i 2.1. The e
i 2.2 di
e
eaive AS a i
ai baed R. A
idei(cid:12)ed i e
i 1 R ake adaive de
ii akig ib e wih ex i
i dai
kw edge e de(cid:12)ed e f
diai. S i i ed a he bai
de
ii akig
faewk f age. Thi eview e
i(cid:12)
a y f
e he e f
i
ai i hee
a i
ai ad aa ye he a
i
a feaibi iy f he ehd ed. The h
ig
f hee aa
he ae idei(cid:12)ed ad he
ib i f eea
h wad a eviaig
he i high ighed. Fia y e
i 2.3 di
e e eva iea e i he aea f ewk
badwidh eiai whi
h i ii a a i
ai dai ad idei(cid:12)e hw
eaig baed aa
h di(cid:11)e f hee.
2.1 Ceaive i Age Sye f Re
e A
ai
he f wig we di
he a j eea
h
ib i i he aea f
eaive AS
deiged f e
e a
ai b e.
2.1.1 F
ia y A
ae Ceai
The f
ia y a
ae
eaive FA/C aa
h adv
ae he ex
hage f aia
eaive i f
a b e ag age geeae
ie aia i
a diib ed ee
h e
gie a i
ai baed hi
e i deve ed by ee
ad Ea 1988. hi he geeae edi
ive ifai ab f e aia
i ha f he he b i d f a
ie g ba i. weve he
bai
ee ad dae ifai b i d a
ee y
ie i
a be
hibiive y age be
a e f
i
ai de ay. e
e i
h i ai i i e
a
i
a
e(cid:11)e
ive a
hieve a g ba i ha ay have a eab e degee f i
ie
y via he ie y ex
hage f aia eaive i. Th FA/C i baed he
e f
i
ai geeae
ie eiae f he g ba b e vig
eai
a idei(cid:12)ed a a key e iee i e
i 1. evehe e FA/C y di
e \wha" i
be de viz. age h d
eae wih aia i ea
h a a
eab e i
a iy. de vide a e
ie f \hw" i
d be a
hieved. wk he he
had adv
ae a e
i(cid:12)
i
ai aegy f he age ive he eaig f
411
D a eig ea
ae eiae whi
h i w d ive he
eaive b e vig. eve we
vide bh a fa aa yi ad eii
a e j ify he bee(cid:12) f aegy
ehig whi
h he FA/C aa
h de d.
2.1.2 gaiaia S
ig
Se eea
he have i
aed gaiaia
e ae f ifai ad
e aihi ha exi bewee he age ad he diib i f b e vig
abi iie ag he i he age de Ca ey Gae 1999. The idea i ha
h
e give a age a bad high eve kw edge ab hw he ye ve
b e he e ha age ay i w ii i he ewk ad hw hey ae
e
ed.
ig hee
e heefe eeia y e ve he e iee ee
e
i 1 f aiaiig high (cid:12)de iy eiae f he i f a ye ha he age
a die
y i. Thi i aa g he way h a gaiai ae fed
ve
ex ak ha ae beyd he
aabi iy f \aia y b ded" a
h Si
1958 idivid a . wk he age d f w e de(cid:12)ed
e f e ad
e aihi. ead hey ea via
eaive
i
ai he eiae f he age
ae ha he deede
ie
a be ifeed ad he ia a
i
a be ake. Th
eai de i a i
ab e a
dai wih e iig ex i
i gaiai
e be e
i(cid:12)ed.
2.1.3 Shii
aed
a C
The hii
aed
a
ehd gy f exa e he aia g ba aig G
aa
h D fee ee 1991 adv
ae ha he
eaig age h d ea ab
hw ex
hage ifai e ve i
ie
ie wh iea
wih ive
eai wha ifai ex
hage
a a
hieve ha b je
ive ad he ike.
w i G he age f
a
a hei a
i ad iea
i egiae ve
a e gaiaia ifai g ide hei aig ad b e vig de
ii
eae i
ie view ad
vege a
eab e ewk efa
e i dyai
eviea
dii. hi de ea
h age aiai i w e f G a e
f
a a ha eee he age view f he g ba b e vig i ai. They
ae daed by he ex
hage f
a aia a ag age ad e(cid:13)e
he e
e
ewk
eai i e f a
hievig he g ba i. e
e hi ehd gy addee
he e iee ee e
i 1 f ae eiai ig
eaive
i
ai.
E abaig he G aa
h he TAES Tak Aa yi Evie de ig
ad Si ai faewk De
ke 1995a 1995b wa deve ed de he ia
ha
he
haa
eii
f a ak evie
a have
diai. Uig TAES
di
ai i a
hieved by hee bad aea f age behavi : hw ad whe
i
ae
ad
a view f he
e b e vig i ai; hw ad whe
ex
hage he aia e f b e vig; hw ad whe ake/beak
ie
ade he age ab wha e wi be avai ab e ad whe. The geea ied a
ia g ba aig GG
i f a g f
diai e
hai baed he
abve bad behavi a ye De
ke ee 1995. Deedig he
haa
eii
f
he ak evie age e e
he aiae
diai e
hai. U ike G
hweve GG diig ihe
a
hed ig f a age f i
diai a
iviie:
he
diai e
hai vide a age
a view f he b e ad he
a
hed e
eae a i
dig bh
a a
i ad
a e(cid:11)e
via
h a
i
412
Shaig fai f Diib ed eaig i AS
ive g ba ye wide i iy by ig ifai f he
diai e
hai.
Th GG i deve ed a d he i
i e f e e
ig a
i baed eiae f he
a ae a idei(cid:12)ed i e
i 1.
weve bh G ad GG e y a bei (cid:13)exib y a e f edeeied
diai
e
hai. S
h e aed
diai
a ve be iade ae agai a f
ige
ie ha
a
i dai whee age aiai i
ee i
e
view
f he w d ae whi
h
hage deeiii
a y ad ay eve fai wih i
idi
ai.
a aa
h f eaig a he age view f he w d ae
he e e
i f a
i whi
h w d g aaee he ivee f he g ba ye
efa
e e ie
h e e
i(cid:12)ed
diai e.
2.1.4 Teawk Baed i ei
bab y he
eheive
eaive AS faewk exiig i
e iea e
i STEA Tabe 1997 ee a eig 1995; Ri
h Side 1997. i deve ed
a d he i
i e f he ji iei hey Che eve e 1991 ad ji
ie eig 1993. T
diae he age aiai a \ji eie ga "
G ha he ea i ji y
ied f dig e ea a
iviy whi e a y
be ievig ha hey ae dig i. Age i STEA aive a a G by ex
hagig ee
h a
:
\e e" ha hey e a
e hei idivid a aia
ie ab aaiig he
g ba ga ad \
(cid:12)" whi
h eab ihe ha a age ha he ae aia
ie
he e wh ade he \e e". F he STEA bw i
i e f he \haed
a" de Gz a 1996 e e ea
hee
e ha a ea ebe
f w a
i ah.
A h gh STEA vide a i
i ed faewk f eaig ab
diai i
eawk a
hievig a ji be ief i age ye f wide y diib ed age i i
ae ike y be a efa
e b ee
k ahe ha a advaage be
a e f he ex
eive
i
ai e ied a
hieve i. Thi i ee
ia y e i evie whee he age
have e a he ak exe
i
e i ee eviea
hage. A h gh
STEA ea hi i e by ig a e aig
i e ie he e eab ihe
f ji
ie ag a he ea ebe. A igi(cid:12)
a a f
i
ai
vehead eage (cid:13)w ad de ay igh be i
ed befe hi i a
hieved whi
h
a degade
he a iy f evi
e igi(cid:12)
a y i e a i
ai. Th i i ia have a
eai de ha a w he age
i e vig he vea ak wih e iig
he eab ih a ye wide
ie wheeve e aig
. eaig baed
eai de ha hi advaage.
2.2 eaig Baed Ceaive i Age Sye
w we eview e key a i
ai ha e a
hie eaig e
hi e f vig i
age
eai b e.
2.2.1 R ig i Dyai
ewk
R i a ied he b e f
eaive diib ed b e vig i a eia ie
e
f wk by Bya ad ia Bya ia 1993 whee hey ve a ewk a
ke
ig b e. hei ae hey de ea
h
i
ai de a ieg a 6 6
gid a a eif
ee eae wh aiai eiae f he de ay i ig a
ke
413
D a eig ea
di(cid:11)ee deiai. T e a a
ke a give deiai a age e e ea
h f
i eighb ig age f hei de ay eiae f ha deiai de. U e
eivig
he de ay eiae f i eighb he e eig age
he he eighb wih he
ii de ay eiae fwad he a
ke. he dae ig adad eaig
i i de ay eiae f ha deiai wih he eiae ha i e
eived f hi
eighb .
The a h f hi ae deae ig eii
a die ha hei aa
h e
ab e he age ea bee i
ie i e f
hig a eighb ig age f ig
a give deiai ha a had
ded he ah a gih. The di(cid:11)ee
e ae e
ed whe he ewk ad i
eae idi
aig ha he eaig a gih i ab e
ada ig de
ii he ah a g whi
h a
ke ae ed de dyai
e
wk
dii. addii he a h e hei a gih wih
hage i he ewk
gy by a a y beakig he ik bewee
eai de ad i he ae f
a
aAE
hagig di(cid:11)ee egi i he ewk whee
a
a igiae ad eiae.
They deae ha hei ig a gih
ef y ada hee
hage ad
ef bee ha he deeiii
he ah a gih.
hei wk heefe Bya ad ia have ed a i e
i
ai
a w he age
eaive y hae hei w kw edge ab he a
ke ig
de ay. evehe e he
i
ai
hey have ed y a w a age
if i iediae eighb ab i w kw edge. Thi ehd w d i
g
ae
y f ifai ea
h age f he away. A a e
ideig ae
hage
i y he ifai
a be
e daed by he ie a age e
eive i. S
h
daed ifai w d he be f i e e geeae e iab e eiae f he
a
ae. hi
ex heefe i i eviaged ha by a wig ae
hage ifai
be haed bewee he g f
eaig age y afe ak
ei he age
a aiai e a
ae eiae f hei
a ae. Thi i
a ive
he vea efa
e f he
eaive AS a a
hieved by Bya ad ia 1993.
eve i hei wk he age dae hei eiae wih he eiae e
eived
f die
eighb . e ha i dig e eae deed he eiae eaed
by ahe. Th hi aa
h eeia y a TD0 ye eaig S 1988 ha he
eia ifa ha \bad" eiae ae agaed d e he eaig f e age.
TC he he had adv
ae aiig he a
a ae f idivid a age he
he i a g . Th TC i eviaged have he h
ig f he aa
h f
Bya ad ia 1993. T veify hi we
he he ig a gih a e f
be
hak f eii
a y eva aig he efa
e advaage f TC.
2.2.2 Tea aiied a e Taii Reif
ee eaig
The Tea aiied a e Taii Reif
ee eaig a gih TT R Se
2000 ha he b je
ive ake he eaig ak eaie i a AS by ed
ig he ae a
e
dieia iy. de hi by aig he ae a iied be f a
i deede
fea e. Aa g ehd f ae aggegai have bee ed i he eif
ee ea
ig a gih e.g.
Ca 1996; Sigh aakk a da 1995 ed
e he ize f
he eaig ak. weve TT R di(cid:11)e f hee aa
he be
a e i ehaie
deivig a e f a ye ifaive fea e f e(cid:11)e
ive eaig. e e
i(cid:12)
a y hee
fea e ae ed eee he h e e(cid:11)e
f a
i ha a age ay ake. Th
he age ea he i iy f e e
ig a
i wih ee
hei w fea e a
e. Thi
414
Shaig fai f Diib ed eaig i AS
i ee
ia y ef whe he age
a beve iediae y i(cid:13) e
e he a
i ake
by he age
h a i ay a
i
a i age eig i ai
a ewk
ig. Tha TT R i a e(cid:11)e
ive a gih i deaed by i
ef a i
ai
a
i e dai Se Ve 1999.
The a h have eva aed he TT R a gih i a i aed ewk ig
evie. The a
i deede fea e i hi
ae ae he ad eve f a de ad
ja
e ik. The age ai hei de ay eiae a g wih a a
ke whi e ig
he ae. F hee hee eiae
e
ed a he
edig deiai de
ae diib ed he de wh ai
iaed i he ig afe (cid:12)xed ie ieva . Th
TT R i fa
e
i
ai diib e ifai ag he age. Thei
eii
a die deae ha TT R ef efa
e ea e i aveage
a
ke de ivey ie he he ah ad ig
whe eaig i de de
wi
hig aAE
dii he a gih i aied de
dii whee he e e
i f
a
ke
e ad deiai ae
haged f w di(cid:11)ee aAE
ae. evehe
e he f wig iiai ae eviaged i hi wk. Fi ideifyig a
i deede
fea e f
a bevai y
a ead f ifai. Thi i be
a e
a
a ae
hage ay be e(cid:13)e
ed i a age iediae ae a
e b
h
ifai ay be e ied by a age e e
a
i. Th i
h
i
a
e ex
i
i kw edge f he
a ae i e
eay. Se
d a a
e e
e f he abve
eied b e he (cid:12)de iy f he deived eiae w d deeiae. Thi i
w d de
eae he vea efa
e f he ye. Thid i he wk f Se 2000
i
e he age dae hei eiae baed he eiae ad ig he a
a
ae a ii a h
ig a idei(cid:12)ed i e
i 2.2.1 f eaig bad eiae
a
.
a wk ae a eviae hee iiai by diib ig he a
a de
ae kee he age ifed f he
a ae. Fia y i TT R ifa
i i diib ed a eg a ieva Se 2000 hweve a fa way f e
ifyig hi
ieva i e
ibed. Thi i a abiay
hee whi
h
a e i age ae
ie i i
fai ea
hig age de. e
e eiae geeaed baed
h ifai ay
be dae. TC he he had diib e ifai iediae y afe ak
ei h aeig iiie he ae
y. Agai be
a e f i
aied e(cid:11)e
ivee ad bad a i
abi iy we
he TT R a he e
d be
hak f eii
a
eva ai agai TC.
2.2.3 i
y Gadie Sea
h
Ahe aa
h f ig R f
eaive diib ed b e vig i ha f i
y
gadie ea
h S
A ee Sigh a 2000. A i
y i a R
ex i
a aig f a ae a a
i. The i
y i h a f
i f a e f aaee
whi
h ae vaiab e de(cid:12)ig he
a ae ad he
e i(cid:13) e
ig he a
i e e
i. A
i
y gadie ea
h i a e
hai ha ie iie he aaee va e
h ha
he aveage g e ewad f he eae i axiied. F exa e i a ewk ig
b e hee aaee
a be he deiai f he a
ke ad he gig ik a e
age e e
f ha deiai e hee aaee ae
a y bevab e a age
whi e he ewad i a ea e f i iy eie he egaive
ha a a
i
a
hieve give he aaee va e. he ewk ig dai he ewad
a be he
egaive i ie f a a
ke ea
h i deiai de. he i
y gadie aa
h
i i a ed ha ea
h age he idivid a eae e
eive he ewad f a a
i ake
415
D a eig ea
by a age a evey ie e.
i y hi ewad ifai ha i g ba y kw
6
by he age. Th he i
y gadie a gih i de fee ideede f dai
de ad kw edge ab he ae ad a
i. divid a age adj hei i
y
aaee i he die
i f he gadie f he aveage ewad ha hey
e ig
he g ba ewad ifai he
e he e i
y gadie. Theefe
i
ai
f ewad va e i key a w he eae iie he aaee va e. weve
he deede
e he g ba ewad ifai dae he i
y aaee
a be a
b ee
k i ye whee he
i
ai badwidh i iied ad hee i a (cid:12)ie
ae
y i eage agae a i a
i
a ye. Thee
ai
a
ead vey w eivee evie
hage i age ig he i
y gadie
aa
h. eve bad
aig ewad by a age i high y dyai
evie
h a ewk exeie
ig heavy ad
a
a e he ewk
ee y a ae
by he ewad eage a beved i i eeai f he g ba bad
a aegy
aed i e
i 1 e ig i a vey ieAE
ie ye.
Thi ehd i ed b i d
eaive AS by Wi ia 1992 Baxe ad Ba e 1999
ad ehki ad Savva 2002 ag he. A f hee wk deae ha he i
y
gadie ea
h a
hieve eaab e efa
e i e f aveage ig de ay
aed
he be
hak a gih
h a he he ah a gih. weve i he wk f
bh Wi ia 1992 ad Baxe ad Ba e 1999 a ex
eedig y age a f ie
i e ied f he eae
vege. Thi i be
a e he eae eed he g ba ewad
ifai dae hei i
y aaee whi
h e i a w iiai f he
aaee. Thi a ei
i he a i
abi iy f hi aa
h b i d a
i
a
ye. A ii a iiai i ike y i he wk f ehki ad Savva 2002 a h gh
he a h d vide hee e . addii wih he high
i
ai vehead
i
ed hi aa
h i ike y be iab e f i eeig a
i
a AS.
a he ihibied
i
ai e ied i he i
y gadie ea
h a
a
h i wk
i
ai i ed a a
ed aegy if he age
ab he i f he g ba ae ha ae e eva hei a
i e e
i. Th we
be ieve wk (cid:11)e a ei a
i
a i he i
y gadie aa
h.
2.2.4 C i
ai De
ii i i Age Cdiai
X a ee ad Zi beei 2001 adv
ae ha
i
ai de
ii ae iega
a age de
ii
diae i a
eaive diib ed AS. The a h
ide
ha ea
h age ve a
a akv De
ii
e D Feibeg S
hwaz 2001
ha geeae bh a
i
ai a
i ad a ae
hagig a
i a evey de
ii
b age. The age ae y give
a bevabi iy i.e. hey
a beve he ae
f he age. weve hey
a beve he
i
ai a
i f he age. The
ia ea f id
ig a
i
ai de
ii i a age
a D a ag ed
i hi ae i ha
i
ai i
. e
e a age h d e y eaig
de
ide whe
i
ai i e ied
h ha he vea i iy eaed f he age
de
ii i axiied. hi
ex hi wk exed he heei
a aa yi f i
age D f B i ie 1999 whee he age ae a ed have g ba ae ifai.
Se
i(cid:12)
a y he a h e w i e he ii
geeaig
i
ai de
ii
6. e ea ii
a y ea
h age ay bad
a he ewad i e
eive a he age. e
e he age
ay e
eive he ewad iga f he eie ye wih e de ay.
416
Shaig fai f Diib ed eaig i AS
ha ai ed
e he
aia
exiy f vig he f D geeae he
ia g ba i
y.
A we ae ieeed i dyig he ia
f
i
ai he efa
e f a
eaive AS he wk f X a e a . 2001 i e aed eea
h. weve whi e
hey aa ye whehe
i
ai i e
eay a ay age f a age de
ii we
ide
i
ai be ieviab e. We aa ye bh aiaive y ad eii
a y
he ia
f a e
i(cid:12)
ifai haig
TC he efa
e f a AS.
wk di(cid:11)e f ha f X a e a . 2001 i he f wig addiia way. Fi y
i hei wk he age ae a ed ieae h gh a e e
e f
i
ae ad a
age y
h y. We ad a e geei
aa
h f
ee y ay
h be
havi . Se
d y hey a e be iaae
i
ai; h ifai e
by a age i e
eived by ahe iediae y. he
ay we
ide a e ea ii
eai whee hee i a (cid:12)ie de ay a
iaed wih
i
ai. Fia y he
i
ai he ii
ha hey e ae baed ea
h age idivid a y iig hei w
ge wad a
hievig a
y ageed ga . wk we
ide f da
ea y diib ed ak
eig whee a age
a ake a
a a
i ad he a
i
f i e age gehe
ee a ak e hi i e
i 3. e
e i wk
idivid a age
a i he ge f a ak exe
i
e wad
ei;
hey ae y
aab e f akig hei
a a
i ig eiae f he beved ae.
2.2.5 he a
hie eaig A gih
The ivai f TC iiia y ay aea be ii a
veia evied eaig
S Widw (cid:11) 1960 whee he a
a
e f a i age edi
i b e i
fed ba
k he idivid a eae edi
. weve hey di(cid:11)e he f wig
ii
a
i e:
(cid:15) S y he (cid:12)a
e
h a whehe a edi
i wa \
e
" a
a he
e f ifai f he eae dae hei edi
i a gih.
a
TC a w a eae gai kw edge ab hw he ae f he age i he
eaive g
hage a g wih he
e.
(cid:15) S a age yi
a y ea a aig f i w a
i he
e f a
i age edi
i b e. TC hweve a w a age
ide he ia
f he ae f he age he (cid:12)a
e f he ak.
(cid:15) Tyi
a y S i ed f edi
i i aiay evie. he he had we
eva ae he
ee
e f TC i aiaiig high (cid:12)de iy eiae i f daea y
dyai
ad
eai evie.
A TC i dii
f he aa
h f ig e igibi iy a
e i whi
h a eae
i vided wih he kw edge f he eie e e
e f ae aii afe a
ee
aiig eide i.e. afe aig f he a ae ad ea
hig he ga ae i
he
1997. he ae a age ea
hig he ga ae afe exe
ig a eie f a
i
dae i evee de i.e. aig f he ga ae ad vig he aig ae
i eiae f ea
h ae aii.
a a
i
a AS hweve i i ib e
f a ig e age beve a aii
ig d ig a ak
eig eide a
a ed i he aa
h ig e igibi iy a
e. F he i a age ad
ex AS
he
aia ad i
ed by a ig e age aeig ake de
ii baed
417
D a eig ea
i kw edge f a ae aii e e
e f evey ak w d be hibiive be
ea ied i a
i
e. hi i ai heefe eea
h
ib e wad deve ig
a a
i
a ad e(cid:11)e
ive ea f diib ig ae ifai ive eaig.
dig i eve he e iee f a age havig aiai he eie
hai f ae
aii i i ey. Rahe i a w he age a
ie a vea i
e f he
ae
hage i he
eaive g ha ef a ak whi
h i a w he
ake de
ii f e(cid:11)e
ive
diai.
2.3 ewk Badwidh Eiai
ewk badwidh avai abi iy die
y ia
he efa
e f ewked a i
ai
h a web evi
e ee ee ye ad bi e ewk. Theefe e(cid:11)e
ive ei
ai ad edi
i f badwidh avai abi iy have aa
ed
ideab e aei i he
ewk
iy. a i
ai f TC d badwidh eiai ad ig hee
fe bea
e eeb a
e hi ie f eea
h. hi ae we e eaig Waki
Daya 1992 f badwidh eiai be
a e i geeae b ad (cid:13)exib e eiae
f bevai. Theefe geeae he eiae i eed beve he badwidh
avai abi iy ae. ewk badwidh eiai
h kw edge i haeed by ei
he a
ive aive ea ee. The fe i de by ije
ig e bai aAE
i he ewk ad he aeig he ae baed he aii
a
haa
eii
f hi
aAE
ai Dv i 2002. weve i a i
ai
h a iied badwidh ad h
ewk
h e bai aAE
waef y
e va ab e badwidh. A e aed a
a
h aded i he age
iy i ha f ig bi e age \a" hae
aAE
dii a
he ewk ad dae de ig ab e Ca Dig 1998a
1998b. w badwidh ewk he id
i f
h exae age
a i i
a
badwidh age. Beide he e
iy ad iva
y b e a
iaed wih he e
f bi e age have ade hei a i
abi iy i ea ife ye debaab e. A eaive y
aive ea ee ae de by (cid:15)ie aa yi f a
a aAE
a
e ai Bake 1999;
Ribei Cae Riedi Savha edi
k Baai k 2000. weve i
ae we
e ie eiai be ie ha we
a
e wih he dyai
dii d ig he
eaia eid f he ye. Theefe TC geeae badwidh eiae by ig
ifai dieiaed by he
eaive age whi e hey e
a . F he he a
ive ea ee aa
he yi
a y eiae badwidh by a ig a ewk de
ha
a a(cid:11)e
he eiai (cid:12)de iy. F ia
e i he wk f Ribei e a . 2000 a
ewk ah i de ed a a ig e e e whi
h diegad he e(cid:11)e
f vaiab e e eig
de ay a g ha ah he eiae. By
hig eaig we ai deve a ai
i
a de f badwidh avai abi iy by
i iig wih geig
aied
by ay ede(cid:12)ed de .
3. Tak Dai Chaa
eii
AS
ex ak ae efed by g f age. Geea y eakig he
b e vig a
iviie f
h age ay be exe
ed i aa e geeae he (cid:12)a
i. B i ay
ae
ai ve
eig di(cid:11)ee b e aed a f a ak
by di(cid:11)ee age ay e ie he ak exe
i
e be aia y e eia ied wih
aiae
hed ig bewee aa e exe
i a ey 1988; a ak Bake C ak
2001. A a i e exa e we
a
ide a
a fa
y whee he i a fa
ig
418
Shaig fai f Diib ed eaig i AS
he di(cid:11)ee a f a
a
a i aa e aifyig a
aibi iy whi e he
(cid:12)a aeb y
e i ay y afe a he a ae
e
y a fa
ed eig
B a 2003. A eaive y he a e f a ak
a ef
e a i
y e eia
eig. F exa e i a diib ed aai ye de ivey f gd bewee w
i e ie a e f
ag vee i e e
e. hi ae we exe ify he a i
a
i f TC i
i e i a e eia dai. weve he
hi
e f a e eia dai
de idi
ae a iiai f TC; TC i baed ay a i f e eia ak
eig e
i 4 ha e deai . evehe e hi i a eaab e
hi
e i
e ev
ea key AS a i
ai
h a e ewk Viwaaha Vahey 1997 y
hai Dekea Zwi
h We k 2004 e e
i
ai badwidh a
ai ia
ae ae 1999 ag he fea e di(cid:11)ee exe ad a di(cid:11)ee eve f
ga aiy e eia ak
eig. ee i h d be
ai(cid:12)ed ha a h gh idivid a
ak ae
ideed e eia he eie i age ye w d yi
a y be efig
i e
h e eia ak i ae y ad ay
h y. F he he idivid a
age a
iviie i exa e a i
ai ae
ideed e iva e a
aig e
e.
Thi i be
a e i ay a
i
a AS a i
ai
h a he e eied abve he
age eeia y a
ae e
e
h a ewk badwidh
e
y
e e
.
ee ak.
Se
i(cid:12)
a y i hi eea
h we
ide a wie e e ehe ewk T a a ee
eaive a i
ai i whi
h i ee
i
ai he ii
baed TC ad
eva ae hee eii
a y. e hweve ha hi
hi
e de ii he a i
abi iy
f e . Thi i be
a e hi a i
ai ha
haa
eii
ha ae
ay
ea w d age
a e diib ed ye ad he
e i ae a e geea y
a i
ab e. Thee i
de: i e age i aed i di(cid:11)ee i f he ye a
ewk de i de ed a a age age havig i
ee view f he a
iviie ad
ae f he ad age havig
diae hei
a ig de
ii i de
ef y a
hieve a g ba ak ig a
a f
e deiai. eve hee
i a
ea ea e f
e whi
h i ia he badwidh age eAE
ie
y
e iva e y he a be f
ef y ed
a .
The T e
i
i wi
hed
i
ai whee de badwidh ha be a
aed
ed ed f a
a
e deiai eab ih
a . Ea
h ig de i eaed
a a age. A de ha a
eai a f badwidh ha
a be a
aed a
eai
axi be f
a i ae y. Ea
h age
a i he ad he a f
badwidh a
aed
a i de ad
a
i
ae y wih he de wihi
i aii age he e f \eighb " de. Ca f (cid:12)ie d ai
a igiae
f ad be deied ay de. Ca igiae eiae
i y. Theefe
he ad he de
i e vay wih ie. The b je
ive f ay
h age i
a
ae badwidh ad fwad a
a e f i eighb
h ha he babi iy f
he
a geig ed via he ea
geed wih he axi avai ab e badwidh ah
i axiied. R ig a
a a a give ie a g he ea
geed ah a ha ie
e e a eAE
ie e f badwidh he
e i
eae he be f
ef y ed
a
i he ye. The fwadig i baed he age eiae f he
gei eve a
he ewk i.e. i eiae f he beved ae. e
e he ak i
eed by a
e e
e f
h a
ai ad fwadig by i e diib ed age. hi
eai
he age
i a y
e ak i.e. e
a f
e deiai a ew
a
igiae. We efe he
e f ig a
a a a ak
eig eide. Se
i 5
e hi i f eidi
ak ex ai he di(cid:11)ee
e bewee he eiae geeaed
419
D a eig ea
by TC ad he
. A deai ed de
ii f he i ai f vai age
a
iviie d ig
h eide i vided i e
i 6.
The ex e
i exad he TC i
i e ad ie hw ifai haig
baed hi i
i e i deiged i he T dai. F hee he i eeai f
he be
hak aegie ig ad TT R ae a ex aied.
4. Shaig fai ve eaig
T i ee e(cid:11)e
ive eaig i dyai
evie by haig
a kw edge bewee
age he
i
ai aegy h d aify he f wig
ieia:
(cid:15) Tie eAE
ie diib i: Thee i a ae
y a
iaed wih
i
ai.
e
e he e ie y he ifai ha i
i
aed a age he e
ike y he ifai i dae.
(cid:15) A
a
y f ifai:
i y
hagig evie i i iib e
f a age eai y
hied wih ae
hage a a ie. evehe e he
e a
ae he ifai e
eived he bee.
Give hee deideaa hee TC i ed a a e(cid:11)e
ive aegy f diib ig he
a ae ifai f age.
De(cid:12)ii 1 ak
ei ifai haig efe he diib i f
a
ae ifai bewee a g f
eaive age by way f a e
hai ha de
ed he a wed age iea
i y afe he
ei f ak deake by he
age.
The ivai f ig hi
hee i e a age ha ai
iaed i
eig a
abaive ak have a idi
ai f he ae
hage f he he age i he g ha
e ed f
eig ha ak. S
h ifai i he ef f akig e ifed
de
ii whi e
eig ay be e ak. a dyai
ye he w d ae
hage
whi e he age
e a give ak. Theefe by de ayig he aii f ifai
i he ak i
eed hi
e e ha a he age wh ai
iaed i
he ak
ei
e ae ifed ab hee ae
hage ad hw ha a(cid:11)e
he
e f he ak. dig we hyheie ha by ig hi e
hai he age
w d be ab e diib e ifai i a ie eAE
ie ae ad ea eaab y a
ae eiae heeby aifyig he e iee idei(cid:12)ed abve. The aa yi eeed
i e
i 5 eab ihe hyhei by
aig he ie ie ad eiae a iie f
TC agai he f he
.
i ia e ha TC i a geea i
i e ad e
i(cid:12)
ia
e f TC
a be
i eeed i a give b e. i eeai f TC i he T dai di
ed i
e
i 4.2 i j
h a ia
e. F he a h gh we exe ify he a i
abi iy f TC
i a e eia dai he
i deiged a d ay
h a i. e ha i
a i
ai whee ak
eig bewee age
ideede f e ahe ad i a
e y aa e fahi he b je
ive f aiaiig eiae f he age ae be
e
ed da. Rahe i
h ye a
ea ak a
a ha a
ae b ak he
aa e y exe
ig age w d be a e iab e aa
h. weve if he
eig f
a b ak by a age e ie eiae f he he d e e deede
ie bewee
420
Shaig fai f Diib ed eaig i AS
he TC
a be ed diib e ifai bewee hee
e. eve if
he b ak i
h ye i e ie e eia
eig ifai haig
a be ed i ha
ex. Theefe we
a ag e ha
i
ai aegie
baed he TC i
i e
a be iab y deiged be a i
ab e a vai eve f
ga aiy i dai he ha he e
he i hi ae.
he eaide f hi e
i we (cid:12) ie he bai
f eaig. S be e y we
de
ibe hw TC i i eeed i a T ive he diib ed eaig f badwidh
avai abi iy. The ig ad TT R i eeai ae a de
ibed i he ae
igh.
4.1 eaig Bai
eaig i a a gih ea he ia a
i a age
a ake
h ha he di
ed
aive i iy f a e eia de
ii b e i axiied.
geea R
a
hieve hi ak by a ig f kw edge f he ewad geeaed f ay a
i i ay
ae ad he e a ae afe a a
i i ake. A fa de
ii f R e ie
ha he evie be eeeed by a 4 e: hS ; A; T ; i whee S i a (cid:12)ie e f
ae A a (cid:12)ie e f age a
i T : S A ! S a babi iy diib i ve
ae geeaed by a give a
i ake i a give ae ad : S A ! R e a
a a
va e he ewad a a e f a a
i ake i a ae. w he i
y f a R age
(cid:25) : S ! A i a aig f ae a
i. Give a i
y (cid:25) he va e f
i V
(cid:25)
efe he
aive ewad ha he age e
eive by f wig (cid:25) aig a ae
ad
a be e ae
; i (cid:21) 1. Th
i
"
#
1
X
(cid:25)
i
(cid:25)
V
= E
(cid:13)
:
1
i
i=0
whee (cid:13) 0 (cid:20) (cid:13) < 1 i a di
fa
ha deeie he e aive weigh f iediae
ad de ayed ewad ad
efe he ewad e
eived a ea
h ae aii f wig
(cid:25)
i
y (cid:25) . e he exe
ed va e i ed be
a e yi
a y de(cid:12)e a babi iy diib i
ve he
e geeaed by akig a a
i i a give ae. he abve exei
he age \ ife" i a ed be i(cid:12)ie he
e he i ve a i(cid:12)ie e e
e. The
ia i
y (cid:25)
i:
(cid:25)
(cid:25)
= agax
V
; 8:
2
(cid:25)
The a
i geeaed a ay ae by he ia i
y (cid:25)
i:
(cid:25)
0
(cid:25)
0
X
"
#
(cid:25)
= agax
E
; a (cid:13)
T ; a;
V
;
3
a
0
whee V
eee he va e f
i
edig he ia i
y (cid:25)
. Exei
(cid:25)
3 idi
ae ha he ia i
y
a be a
ied by eaig he ia va e f
i
f a ae. evehe e i dig
ee kw edge ab he ad T f
i i
e
eay. weve i a
i
a dai a a
ae kw edge f hee w f
i
i ib e.
Agai hi ba
kg d he ef e f eaig i ea he ia i
y wih
havig ea he ia va e f
i. The key a
hievig hi i he f wig
421
D a eig ea
bi i:
(cid:25)
(cid:25)
0
(cid:25)
0
X
; a (cid:17)
; a (cid:13)
T ; a;
V
:
4
0
The ia i
y
a be
a
aed if a age ea he ; a va e a
(cid:25)
= agax
; a;
5
a
ad he ia va e f
i i
ed a
(cid:25)
V
= ax
; a:
6
a
A e
ive de(cid:12)ii f he va e i h
(cid:25)
(cid:25)
0
0
0
X
; a (cid:17)
; a (cid:13)
T ; a;
ax
; a
:
7
0
0
a
Wih ay kw edge f he ewad ad ae aii f
i he e va e
hw i 7
a be eiaed h gh eeaed aiig.
ai
a he f wig
aiig e d e Waki ad Dya a w he eiaed va e
vege hei
^
e va e Waki Daya 1992:
^
^
^
0
0
; a 1 (cid:11)
; a (cid:11)[; a (cid:13) ax
; a
℄:
8
1
0
a
h
ee ad a ae he ae ad a
i ee
ive y daed f he 1
aiig ieai
(cid:11) 0 (cid:20) (cid:11) (cid:20) 1 geeae a de
ayig weighed aveage f he
e
eiae ad he
^
evied va e ad
i he e a ae afe akig a
i a i ae a
dig he
0
e i
y. By adj ig gad a y ed
ig he va e f (cid:11) he
eiae
a be
^
deaed
vege he a
a va e Waki Daya 1992. hi eea
h
we e age e he aiig e 8 ea he va e.
4.2 Saegie f Shaig fai i he Ceaive Badwidh
A
ai b e
A ig a T ha a e f age A we
ide a abiay be f A = fa
j i =
i
1; :::; g whee i he
adia iy f ha
ee a give ak a a give ie.
dig a age i
ide e f i eighb fwad a
a efe he di
i
i e
i 3. Thi de
ii i baed he age eiae. Se
i(cid:12)
a y we e a ab e
f ea
h age a
whee a ey
; k eee he exe
ed i iy f
hig eighb
i
i
a
whe he
a deiai i a
e he ize f he ab e f ea
h age i j A j j j
k
whee i he e f eighb age. ai
a f a T we
he he va e
eee he exe
ed avai abi iy f fee badwidh BW
hae de a g he vai
ah f a
a
. e i hi eeeai f he f
i i he ga ae he
k
fa
ha he
a ha be ed a
f he age ad he
e ae i i he fa
ha he
a i
e y wih a
. S e(cid:11)e
ive y he age ea fwad a
a a g he
i
ah wih he axi avai ab e BW ea
h he ga ae f he
e ae. Thi
eeeai ha he advaage ha a ieediae ae aii he e e
e f de
ha he
a ha be ed h gh ae
aed i e e(cid:11)e
ive aii f he
422
Shaig fai f Diib ed eaig i AS
e ae he ga ae. The va e heefe igi(cid:12)e he \e(cid:11)e
ive i iy" i e
f he avai ab e badwidh; he highe he va e f whi
h he bee i he i iy i e f
ef y ig a
a f e e
ig a give eighb ea
h he ga ae. Thi va e
i eaed f he ifai diib ed by he ifai haig aegie de
ibed
h y. e a ii a f
i ha bee ed evi y by eea
he dyig adaive
ig ig eaig Bya ia 1993.
he abve
ex a age a
wh wa e a
a a deiai de a
i
ig a e f de ha have he highe exe
ed BW avai abi iy
he he eighb a
k
0
h ha a
= agax
; k whee he axiiai i ve a eighb ig age f a
.
i
i
0
k
a
k
eii
a dy we e B za ex ai Waki 1989 a adad
hee f
babi ii
a y
hig a eighb a ed hi deeiii
aegy. A ia
e
f age a
e e i eighb a
fwad a
a wad he deiai de a
1
2
i hw i Fig e 1a. Beig
eaive a
wi a
e he e e if i ha he avai ab e
2
aa
iy. Afe fwadig he e e a
e a
ae e i f
a
hae badwidh
1
Fig e 1b f he aia y
e
ed
a i i i eihe
ef y
e
ed ded
a de
ibed h y. The fwadig
i e Fig e 1b i he deiai de
a
i ea
hed. A hi i a eage i aied ba
k a g he e h gh whi
h
he
a wa ed if ea
h age ha he
a ha a
a y
e
ed. Ea
h age
he a
ae e i f
a
hae badwidh
ee a
i
i f he
e
deiai befe hi he de had y e a
aed badwidh Fig e 1
. A
ig hi eage age ai hei
a ae va e he age he e.
e
e he age ha
eaed a ak ig he
a hae ag hee ve
hei
a ae ifai afe he ak i
eed afe he e e
e f e e ea
h
he deiai de. hi way heefe he TC i
i e ee de(cid:12)ii 1 ha bee
iaiaed e
i(cid:12)
a y i he T dai. e deai hi f w h y.
weve he fwadig
e if a age i
a
ed ha ha a
aed
a
hae badwidh. The he age ai a eage if he he e
d he aia y
e
ed
a ad dea
ae he e a
aed badwidh Fig e 1d.
addii he w
ae e ie f a
a i b ded by a e ii f he ie ha
he age
a
i e wih he fwadig
e. Afe hi ie he
a i ded if i
ha
e
ed. Thi ie i e iva e he axi de ay a
a e w d exeie
e
bewee dia ig a be ad heaig he ig e.
Fia y if a age whi e fwadig
7
he
a dee
a
y
e i he e ake by he e e
e f e e i geeae a eage
wih a ea y ad ai i he age he
y
e. Thi ea y a exeia y
de
eaig f
i wih he dia
e f a de f he ed i be e y ed by
he age dae hei eiae
h ha he
e
e f f he
y
e i ed
ed.
eve whi e ea iig he age he de a
ae he evi y e a
aed
badwidh i
e ae ed da i f a
a ah f e deai ee e
i 6.2.
he e
edig we eeed a bad de
ii f he e eia badwidh a
ai
ak efed by he ig age. he f wig we (cid:12) ex ai hw he TC i
i e i i eeed i hi eig. The he i eeai f he be
hak i hi
7.
ae f he abve w
dii we d e ay ba
ka
kig ea
h f a eaive ah. Thi
kee he ig
i e ad ake he aa yi f he ye behavi eay. eve hi
i i(cid:12)
ai h d ia
he vea
i f hi eea
h i
e i
i f ba
ka
kig
w d ia
a aegie e a y.
423
D a eig ea
Network
a 1 re q u e sts a 2
a2
a1
Call source
node / agent
a2 forwards request
pre−allocate
a2
a3
an
(a)
Call destination
(b)
an
a1
a1
Final route
a2
a3
an sends connect message
(c)
an
Allocate
drop call
a2
a3
no bandwidth with a3
a1
de−allocate
(d)
Fig e 1: The
a fwadig
e
dai ae eeed. A e deai ed di
i f ye i eeai i vided
i e
i 6.
4.2.1 Tak C ei fai Shaig
Cide a de
ibed abve a
a i ed f a
e a
deiai a g he
1
ah a
:::a
. he iaiai f he TC i
i e age a
a
i
aig i
1
w
a ae i a T a de ae i eeeed by he avai ab e BW i ha de
whe he BW a
ai
e
ee a ie a a
heeby eab ihig a
ee
i
i f
e deiai. Age a
i
ae hi ifai a
. Th
1
a
dae i i eiae f a
avai ab e BW i
; wih he ew ae
1
1
ifai ig he adad dae e e ai 8:
; 1 (cid:11)
;
1
1
(cid:11) ; whee ; eee he
a ae f a
a ie ad i he \ewad" f he
eae dae i i eiae. S be e y a
i
ae a
i w
1
2
a ae a ie
1;
6= be
a e f he ae
y i
i
ai bewee
0
0
0
eighb age ad he ifai i had e
eived f a
. A eaive y i
a e i
w ae ifai ad ha e
eived f a
i
ae a ay ifai ha
a e he vea ae f he ah beig ed e he
a . Se
i 6.3 ie w
he ii
f dig hi.
Age a
ii a y dae i i eiae f he badwidh
2
8
avai abi iy
; 1 ig he ifai e
eived f a
. Thi
ed e f
2
1
diib ig hei w ad he evi y e
eived ae ifai
i e i he
e
age hee a
i ea
hed. Tyi
a y he ifai ab he ae f i e de i
1
ed geeae a ay eiae f he ae f he dwea e a de
ibed i
8. S
h aiai f ae ifai i eaae f he bai
TC i
i e whi
h i y ae ha
ifai i haed bewee a
eaig g afe ak
ei.
424
Shaig fai f Diib ed eaig i AS
e
i 6.3. The dii
i bewee
a ae a age w ae ad
a ae
ahe age ae i i aggegaig ifai f i e de aiai a
ay f
a e. weve hi i a b e i a T i
e he kw edge f ah
badwidh avai abi iy i AE
ie f a age ake e(cid:11)e
ive ig de
ii e e
ig a
be e de fwad a
a e e. he he had aiaiig he ifai
f idivid a de w d e
eiae ea
h age vig a
aia y exeive ea
ah b e befe evey ig de
ii. weve i a di(cid:11)ee dai i i eie y
ib e f a age aiai eaae ae eiae f he age i he
eaive
g whi e ig he ae TC i
i e.
The ea iai f TC i he T dai a w e age a a ie
i
ae
ifai i iediae eighb . Thi i be
a e: i e age i a T he dei
ai de dee
he
ei f a ak whi
h be e y a aiig ae
ifai; ii a age
a
i
ae i iediae eighb y; ad iii ae
ifai aii f w a e e
e f he deiai de wad he
e
d e he evi w ea. weve i a di(cid:11)ee i age dai a ea iai
f TC ay iv ve i e age haig ifai a he ae ie deedig he
iea
i ib e bewee age.
hi ai
a iaiai f TC y he de wh ai
iaed i ig a
a hae ifai. S he ae ifai f he he de i he ye i
diib ed. B he de
ii f whi
h age h d be ifed ab whi
h ye ae
i a eaae b e.
hi ae we adv
ae TC a a e
i(cid:12)
ai f diib ig
ifai ag he age wh
eae a ak afe ak
ei. Theefe
diib ig ifai ag he ig de f w hi e
i(cid:12)
ai. evehe e i
gig eea
h we ae iveigaig he b e f hw we
a deeie whi
h age
wihi a
eaive g h d be i(cid:12)ed f a
eai ie
e f ae ifai afe ak
ei ad he
e e
e f
h e e
ive diib i.
The
i
aed ae ifai a
a he \ewad" f eaig. Theefe he
a
a
y ad he ie ie f hi ifai i
ii
a i deeiig he a iy f he
eiae. Thi i die
y ia
he e(cid:11)e
ivee f a age de
ii e
a .
4.2.2 ig
The ig a gih i baed he
a di
ed i e
i 2.2.1.
ig heeafe efeed a R a age a
afe fwadig a
a eighb
i
a
e
eive he ae
e be eiae f he BW avai abi iy ea
h deiai
i1
a
. Th eighb a
if a
wih
= i
; ax
; i 2 whee he
i1
i
i1
i1
i1
^
a
i2
axiiai i de ve a eighb a
f a
. Si
e a give ah i a T
i2
i1
he de wih he ii badwidh avai abi iy deeie he axi be f
a ha
a be a
ed via ha ah a
deeie he ii f i w badwidh
i1
avai abi iy i \ae" deed by
ad i eiae f he be e ah. Age
i1
a
e
eivig hi eiae dae i i eiae
; i 1 a:
; i 1
i
i
i
1 (cid:11)
; i 1 (cid:11)
. Thi
e f akig eighb ad e
eivig he ae
i
i1
^
eiae
i e i he deiai i ea
hed ee Fig e 1
he fwadig
e
i eiaed ee Fig e 1d. The way he eiae ae daed i R i ii a
he dae e ed i TC a hw i e
i 4.2.1. The di(cid:11)ee
e hweve i i he
425
D a eig ea
ewad: wheea i R he ewad i he eiae f he iediae eighb i TC i i
a ay f he a
a ae ifai f a be e ah age.
4.2.3 TT R
TT R i i eeed i T dai f wig he de
ii give by Se 2000.
The
haa
eii
fea e f TT R ee Se 2000 f e deai ae i eeed
a f w: i a aiiig f
i ad ii a a
i deede fea e f
i whee he
a
iviy widw aaee i
he a 100 ad he age heh d f a ik
e
ig
a eighb i e 5.0 ha f he axi badwidh
aa
iy f a de de(cid:12)ed i
e
i 7.2 bh idei
a he f Se 2000; ad iii he ewad dae ieva i
e 100. Se
i 6.2 ex ai i e deai hw hee aaee ae ed by he age
ea eiae ad e
a . Whi e fwadig a
a a de ig TT R ai
he be e de i
e eiae ab he badwidh avai abi iy ea
h he
give
a deiai. Ea
h age a e
d he a f badwidh age ea
h f
he ik
e
ig i eighb . va e eiae he badwidh avai abi iy
ea
h a give deiai via a give eighb f he give ik age eve ied
ve he a a
iviy widw ie e. A
a fwadig de
ii i ake by ig a
B za ex ai ve he eiae. Rewad diib i
i TT R evey
dae ieva ie e. e e
i(cid:12)
a y f a
a ha ae
ef y
e
ed
he
edig deiai de a
ae he ifai ha wa aied by he
fwadig de a g he
a ah. The afe evey dae ieva ie e hee
deiai de a edig he a
aed ifai ba
k a g he
edig
a ah. A age
aed a g
h a
a e dae i va e afe e
eivig hi
ifai. S a de a
a g he ah a
; ; a
; ; a
ge he eiae a ed
k
1
k
he a
a de ae f i be e de a
; ; a
ig whi
h i dae i
k1
va e. e ha we have ed aggegai f he ifai e
eived f be e
age ii a ha i TC e
i 4.2.1.
5. The Advaage f he TC Shaig i
i e
hi e
i we ee a fa aa yi ex ai he advaage f ifai
haig de i geeaig bee eaig ha he
. Whi e hi eab ihe he
bee(cid:12) f TC a heei
a g d i a vide a ex aai f he efa
e
ivee beved i eii
a die. he f wig we (cid:12) ae a
i ad ai ha wi be ed i ae di
i. The e f hi e
i deve a
fa eeeai f he ie ie f diib ig ifai by bh aegie. Thi
eeeai i he ed
ae he a
a
ie f he
a ae ifai i
ea
h
ae.
5.1 Bai
A i ad ai
e
i 3 i wa aed ha we
ide ak eide ha e ie he e eia ai
ia
i f he aiae age f
ef
ei. hi
ex aa yi f
e
a ai
a e f age = fa
; :::; a
g whee age a
iiiae he ak exe
i
1
1
e by e
eivig a ew ak. A we a e wih f geea iy ha he de
i whi
h he age
e ak i: a
! ! a
. Thi a i i ie ha i hi
1
ak
eig ia
e a
e i kw edge f he ae f he age ad e e
a
1
2
426
Shaig fai f Diib ed eaig i AS
fwad he ak a
ii a y e e
a
ad i a
e e
a
whi
h i he age
2
3
1
a whi
h he ak
eig i
eed hi wa de
ibed i e
i 4.2 i he
ex f
a T. e ha eee e ib e e f age ha
a
ee he ak e iv
a e y i a ewk hee
a be i e e h gh whi
h a
a
a be ed
he deiai. The ae eiae ha he age e e e
he be e age ae
geeaed ig
i
aed ifai f he beved ae via he ai
a
i
ai aegy ed by he age TC . Th f
ig e ai
a e
i hi
ae i i(cid:12)e he aa yi f hw a
i
ai aegy a(cid:11)e
he a
a
y
f a give age kw edge f he age i ha e. Si
e i abiai y e e
ed i i
e iva e e e
ig ay he e f age. Theefe he e f aa yi d
deed whi
h i
he.
F he we
ide ha he age
e ak ha ae geeaed
i y a
e e
i 3. T hi ed e he yb
dee he (cid:12)xed ie afe whi
h
eive
ak ae geeaed. S
h a a i f eidi
iy i he ak evie i ade
i ify aa yi. S be e y i e
i 5.6 we hw ha eve de e geea
eidi
evie he ae geea
i h d.
he e e
e f age ha ji y ai
iae i exe
ig he ak hee i he i
f a \ be e" age ad f ha ae a \e
edig" age f ay age ex
e he
a he (cid:12). We eee he age be e a age a
by he idei(cid:12)e a
ad
i
i1
he e e
edig a
by a
.
i
i 1
The age ae a e e
i 4.2 i eeeed by a ea va ed f
i . F exa e
i a ewk hi
a eee he ad eve e iva e y he fa
i f he a
badwidh ed a de. A a di
ed i e
i 4.2 a age ea hee age
ae ig he
i
aed ifai f he age de
ide whi
h be e
age
he. The a
a ae f age a
a beved by a
ie f a ie i eeeed
i
i
0
by i; . Age a
kw edge f a
ae a ie i
i; j; i 6= j . The kw edge ha
i
j
a
ha f he age i a ie i eeeed by:
i
0
S
= f
i; j; j j = 1; :::; g:
9
i
The
edig e f a
a ae va e f he age i i eeeed by:
S
= fj; j j = 1; :::; g:
10
e ha i dyai
ye hee ae
hage wih ie. F exa e he ad eve f a
i
ai de vaie wih ie. Th wih ie y dae he kw va e
a
be di(cid:11)ee f he a
a ae va e.
A ed ea ie age a
e S
e e
he be e age wh i fwad
i
i
a ak. Thi de
ii i a(cid:11)e
he vea i iy eaed f
eig ak i he
ye. The exa
f
i ed by a age deeie he be e age deed
he ak ad he dai
haa
eii
.
a T f exa e a age
a e e
a
be e age f whi
h i eiae ha he aveage ad a de f ha age
he deiai de i iiied. Th hi de
ii ha he e(cid:11)e
f ig he ea
geed ah evey ie a
a ha be e whi
h i axiie he be f
a ed i he ye. aa yi de deed he exa
f f he de
ii
f
i. Rahe i die he de ay bewee
e
ive eif
ee f ae ifai
by a give ifai haig aegy ha geeae he kw edge
i; j; . The
i; j;
0
0
va e a
a he aaee i a age de
ii f
i. e
e i
a be
ded ha
427
D a eig ea
he
e hee va e ae he e ae whi
h a aed ea ie
hage wih ie he
highe i he a
a
y f he age de
ii. iive y he he he de ay i haig
ifai he e dae i he ifai aiaied. Th he e e(cid:11)e
ive
a age de
ii. he f wig be
i we aa ye he ie ie f ifai
haig by he di(cid:11)ee
i
ai aegie.
5.2 Tie ie f fai Diib i i
hi e
i we
e he de ay i
ed by a age ge he ae ifai f
he he i ig he aegy. F hi we f
a
kw edge ab he ae
1
f fa
; :::; a
g a ie . A hi ie a
ha he kw edge f i a
a ae 1; . weve
1
1
i kw edge f he he age
1; j; j = 2; :::; ae di(cid:11)ee f he
edig
0
e ae by a a e a j j;
1; j; j. e ha he ifai ha a
0
age aiai a a give ie i he e f he evi
i
ai ha
ed
bewee he age efe e
i 4.2.2 f a di
i hw a ai
a i eeai
R f he
wk.
Si
e ak igiae evey
ie e a age a
j = 1; :::; 1 e e i b
j
e e age a
f he ae kw edge evey
ie e. F wig a e e a
j1
ay ie a
e
eive he ee f a
afe a de ay f 2 a ig he e e
j
j1
aive a a
afe a de ay f ad he ee f a
e ba
k a
afe a f
j1
j1
j
he de ay f a 2. e hee efe he
i
ai de ay f a eage
bewee die
y
i
aig age. Refeig he di
i i e
i 4.2.2 we e
ha a
vide he ifai ha i ha f he e f age fa
; :::; a
g.
weve
j1
j1
9
a
kw edge ae baed he ifai i e
eived f a
i evi e e
j1
j2
a
. The evi e e f a
a
wa d ig he
eig f he evi ak
j2
j1
j2
a
f whi
h i had e
eived a ee a
2 i.e. afe a de ay
f
2. a ii a way ha ee f a
a
aied ifai
j2
j1
ha a
e
eived f i evi e e a
. Tha e e f a
a
wa a
j2
j3
j2
j3
2 2
f whi
h i had e
eived a ee a 2 2
2 i.e. afe a de ay
f 2
2. Exedig hi
ed e a be e age heefe a ie
2 he ifai ha a
ha f ay he be e age a
i he ae f a
de ayed
j
k
k
by a a d 1
2 whee d = k j . e hee a exa i added
he de ay be
a e a h gh he ee wa e
eived a d 1 d 1
2
b hi
aied ifai ab a
a ie d 1 d 1
2 .
k
The abve de
ii i aied i Tab e 1. hi ab e he w eee age
wih he age be i
eaig f b ad he
eee ie
wih ie fahe i he a a we ve f ef igh. e e
i(cid:12)
a y i Tab e 1
ea
h e ee eee a ie whe a age eeeed by he w be e
eived he
ifai f i be e age. ai
a i f
e age a
ad a e ha
1
i ha e eed a
f he ae kw edge a ie . Theefe a
e
eive ifai
2
1
f a
a 2 w 1
2 i Tab e 1. weve he kw edge ab he be e
2
age ha a
vide a
ae baed he e e ha hee age ade a ie f he
2
1
9. e ha i e
i 4.2.2 we di
ed R whee a ay f he ae f he be e age i
i
aed. hi fa iai we
ide a age aiai eaae e
d f he ae f he
age. S
h a
ideai he ex ai he ia
f a give
i
ai aegy he a
a
y
f a age kw edge.
428
Shaig fai f Diib ed eaig i AS
Tab e 1: Tie diaga f eae eighb haig
Age
Tie
a
2 2
2 2
1
a
2
2 2
2
.
.
.
.
.
.
a
2 2 2
1
de ayed i he a. Thee ae he (cid:12) e ee f ea
h w. e
e a ie 2 he e
f ae ifai f he age fa
; :::; a
g ha a
ha i he f wig:
1
1
2
S
= f1; 2g [ fi; i 2 2 i 2
j i = 2; :::; g:
11
1
e ha a addiia i ba
ed i he va e f a be e age i 11.
Thi i be
a e whi e Tab e 1 hw he a ie a age e
eived ifai f i
be e age hi ifai i i fa
de ayed by a a f ; he
e he i
ba
ed. F
aiy 2 i 11 i e a
ed by
. Th
0
0
0
0
S
= f1;
g [ fi;
i 2 i 2
j i = 2; :::; g:
12
1
The e ae f hee age a ie
ae
0
0
0
S
= fi;
j i = 1; :::; g:
13
Th 12 hw ha he kw edge ha age a
ha f ay he age i i de ayed
1
by a a ha deed he dia
e be f h bewee he. e e
i(cid:12)
a y
he kw edge ha a age ay a
ha f ahe age ay a
ha i k h away i de ayed
i
j
by a a :
= k 1
:
14
de ay
Thi ea e f de ay i
ed i wi be
aed he ae i TC.
5.3 Tie ie f fai Diib i i TC
hi e
i we
e he de ay i
ed by a age ge he ae ifai f
he he i ig he TC
. Sii a he aa yi i e
i 5.2 we a e
ha a ak igiae evey
ie e whe a
iiiae he
eig f he ak ad
1
fwad he e e he eaiig age fa
; :::; a
g. he TC haig
he
2
ae ifai f age ae
i
aed y afe a ak i
eed. Si
e ew ak
ae
eed evey
ie e i
a be ifeed ha he diib i f ae ifai
by he age
evey
ie e i.e. afe evey ak
ei hae ad a ig
ha he
i
ai de ay bewee ay w die
y
i
aig de eai he
ae.
429
D a eig ea
−2 tc
− tc
t + (n−2)
t + (n−2)
t + (n−2)
t
t
t
t
−2 tc
e
m
t− tc T
t
i
1
2
3
n
Agent
2 communicating to 1
n communicating to n−1
Information transmission
Fig e 2: Tie diaga f TC haig
Theefe a
ai i ae ifai a
a ie i.e. whe a ak
ee
1
a ie f ay va e f ad he a
2
ad . Give hi a
ai
1
i w ae ad he ifai e
eived f a
a
a ad he a
2
2
ad
ideig he de ay f f he ifai ea
h a
f a
.
1
Exedig hi
e i
a be ifeed ha a
ai i w ae ad i kw edge f
2
he e f age fa
; :::; a
g a
a 2 ad he a 2
22
3
1
ad . Fig e 2 hw hi
e. hi (cid:12)g e ea
h age i abe ed wih he ie
a whi
h i ai i ae ifai i evi age. Th he ae ifai
ha age a
e
eive f i be e age i hi
ae a
ai he ifai f
1
2
he e f de ayed by i e f . Th a
ha he f wig ifai ab he
1
be e age ae a ig i e
eived ifai f a
a ie
:
2
0
0
0
S
= fi;
i 1 j i = 1; :::; g:
15
1
The e ae f hee age a hi ie
hw i 13 hweve ae di(cid:11)ee f hee
0
va e.
Th 15 hw ha he ifai ha age a
ha f ay he age i i
1
de ayed by a a ha deed he dia
e be f h bewee he. e
e
i(cid:12)
a y he ifai ha a age ay a
ha f ahe age ay a
ha i k h
i
j
away i de ayed by a a :
TC
= k:
16
de ay
he f wig e
i we e he de ay ea e
ed i 16 ad 14 eab ih
he advaage f TC
aed .
5.4 Caig Tie ie f fai Diib i i TC ad
The aa yi f e
i 5.3 hw ha ig he TC aegy a age ay i afe he
ei f a ak eide e
eive he
a ae ifai f ahe age j afe a
de ay f k ee f a 16 whee k i he h
bewee i ad j . he aegy
he he had age i e
eive j ae ifai afe a de ay f k 1
ee f a 14. Caig he de ay he f wig
a be
ded.
ii 1 The de ay f
a ifai ea
h a age i e ig TC ha
ig if he ak evie eidi
iy i geae ha he d i
i
ai
de ay f a eage bewee die
y
i
aig age.
430
Shaig fai f Diib ed eaig i AS
Thi i be
a e f k > 1
> 2 k < k 1
:
17
a yi
a AS he ieva bewee
eive eide f ak exe
i
i
h
10
ge ha he
i
ai ae
y bewee w die
y
i
aig age .
e
e he de ay d e TC i f a a
i
a e
h e ha ha f .
avig eab ihed ha TC diib e ifai i a e ie eAE
ie ae
ha we w f
aa yig hw hi
haa
eii
f TC
eae e dae
ifai.
5.5 ved Eiai A
a
y ig TC
hi e
i he ived ie eAE
ie ifai diib i f TC i aed he
ived a iy f ifai eaed by TC ve . The key idea i ha he he
he de ay bewee
eive ifai eage he e a
ae i he kw edge f he
a
a ae.
A aed befe de ae vay dyai
a y ve ie. weve a a give ie he
ae f a de
a have a
eai va e f a
eai e f va e ay V . A a de
eai i ae
whee
2 V ad 2 f1; ; g whee =j V j f a
eai
i
11
egh f ie ay
. We
ide ha a
eai de
i dyai
a y eiaig he ae
i
f ahe de
ha i a a dia
e f k h f
. Give he abve ifai we
j
i
wa
e he exe
ed va e f a give be f ae
hage ha
a
i
j
i a give ie d ai ay
. We hyheie ha if
i
eae de he exe
ed
D
D
va e f ay be f ae
hage wihi
. Theefe if
e
eive ifai f
D
i
wih highe de ay he i e e ae
hage ifai f he ae. Si
e ha
j
a geae de ay ha TC e
i 5.4 i i
a highe f ae
hage ifai
ha TC. hi
ex we de(cid:12)e he f wig:
De(cid:12)ii 2 Give a e f ae V = f
j = 1; :::; g =j V j whee ea
h ae
va e
a f a ie egh f
ad a ie d ai
whee
<
a
veage
D
D
f ize h f V
eeeed by
h; V ;
i a e f h di(cid:11)ee ae f V h (cid:20)
D
D
h ha
(cid:20)
.
i
D
h
Uig he abve he exe
ed va e f h di(cid:11)ee ae
hage f
wihi a ie ieva
j
i give by:
D
be f ib e
h 1; V ;
D
;
18
h1
whee he ea
a ib e
veage f ize h 1 he
e havig h di(cid:11)ee
ae
hage f V
. The deia e eae a ib e way f
hig h 1
D
di(cid:11)ee ae f V .
10. F exa e i he ye f
i
ai ewk we ae dyig i hi eea
h he yi
a de ay
bewee
eive
a i f he de f i e wheea he
i
ai ae
y bewee adja
e
de i f he de f i ie
d.
11. The eeeai i; ed ea ie i hi e
i ideify he ae va e f age a
a ie i
i
e a
ed wih
. he
e di
i i
e we ae
ideig e age ad he di(cid:11)ee ae
va e ha i
a ake he idei(cid:12)e i ad ie ae ded f a eaie ai. evehe e
8i; 8; 9; i; =
.
431
D a eig ea
w if we
ide a d ai
>
<
he i i ivia ideify
0
0
12
D
D
D
0
ha evey
h; V ;
wi a be a
h; V ;
f a h. Thi i be
a e a
biai
D
D
f
ha \(cid:12)" wihi
w d e
eai y (cid:12) wihi
. Theefe i
a be aid ha
i
D
D
0
b
h; V ;
=
h; V ;
. e
e he ea f 18 wih
e a
ed by
w d a
D
D
D
D
0
0
ea be e a ha f
. The abve eaig big he
i ha he
D
exe
ed va e f bevig di(cid:11)ee ae
hage f ay i
eae wih i
eaig
de ay bewee
eive bevai. Th e ae
hage ifai i a he
de ay bewee bevai i
eae. Si
e a
dig e
i 5.4 TC a
hieve a we
de ay ha bewee
eive bevai he f wig
a be
ded.
ii 2 TC i
a we f ae
hage ifai ha .
The aa yi eeed fa a e a eidi
evie whee he ak eide
eea afe ieva f
a egh. he f wig we ee a ii a aa yi wih
he eidi
a i eved ad deae ha he ae
i h d.
5.6 eidi
Tak Evie
The aa ye eeed i e
i 5.2 ad 5.3 ae baed he a i ha ak
e
i eide eea evey
ie e wih a
a
. Theefe he f a 14 ad
16 wee deived ig y e f hee eide. a e geea eig hweve he ak
eig eide w d be eidi
wih he ie bewee
eive ak
ei
eide vayig. ha
ae hee f a have be
ed
ideig he
eive
eide a ed y e. hi
ex e ha he ifai dieiai
de ay f a e f a 14 deed he va e f
. Theefe he a i f
eidi
eide ia
he de ay e f y . The f wig di
i idi
ae
hw a
f he eidi
iy.
Cideig he
ae de
ibed befe i e
i 5.2 whee age a
aiai he ae f
i
a
whi
h i k h away. a eidi
i ai f a 14
hage
j
k 1
X
j
=
k 2;
19
de ay
=1
whee
f ay j eee he e
e eide
he e
d e
e eide
j 1
j 2
ad .
Uig a ehd ii a ha di
ed i e
i 5.4
give by 16
a be
de ay
TC
aed
give by 19. Theefe we
a
de
<
if:
de ay
de ay
de ay
TC
k 2
X
j
2k 2 <
:
20
=1
The f wig aie hi bevai.
ii 3 a eidi
ak evie he ifai dieiai de ay i
e f TC ha f if he ie bewee ay w
eive ak igiai i geae
ha he d i
i
ai de ay f a eage bewee die
y
i
aig age.
12. f
=
he he exe
ed va e f bevig h ae
hage i e a 1 f a h. f
>
D
D
00
00
he we
a a y he ae eaig a abve f he di(cid:12)ed d ai
whee
(cid:17)
d
D
D
D
ea
h he ae
i.
432
Shaig fai f Diib ed eaig i AS
Thi i be
a e f
dii 20 f (cid:21) 1
j
TC
> 2
<
:
21
de ay
de ay
ii 3 i ii a ii 1. i e f a a
i
a e be
a e he
ie ieva bewee
eive ak
eig eide i yi
a y
h geae ha he
d i
i
ai de ay f a eage bewee die
y
i
aig de.
Si
e he de ay bewee
eive ifai e
eived i a e i TC ha i
i
a be hw ii a e
i 5.5 ha de he eidi
ak a i he
kw edge ab he
a ae geeaed by TC
a e he
hage i hee ae
bee ha ha by .
ii 4 a eidi
ak evie TC i
a we f ae
hage
ifai ha .
The e
edig aa yi deae ha de a a
i
a e he ie ie f
ifai diib i ad he a iy f
a ae ifai eaed by TC
ifai haig aegy ae bee ha he eae eighb aegy de geea
eidi
evie. Wih hee heei
a e i i eaab e ife ha TC
a w he age ake bee ifed de
ii ha whi
h i geeae bee
ye efa
e. T deae hi f he he a
i
a advaage f TC i eva aed
ig eii
a aa yi i a i aed wie e e ehe ewk. The f wig e
i
de
ibe he i ai evie.
6. eeig fai Shaig Saegie i a Si aed Wie e
Te ehe ewk
hi e
i we (cid:12) e eae a be f ia hyi
a eie ad f
ia
haa
eii
f hi a i
ai. Thee eie ae i aed i ye ai
ae y
a e hei e(cid:11)e
i efa
e. S be e y we de
ibe i eeai
f a
eaive e
e a
ai ye f ig
a i a
i
i wi
hed ewk.
ai
a i e abae he i eeai f he TC R ad TT R aegie
de
ibed i e
i 4.2 i he i ai. Fia y we ee w he ii
baed he
TC i
i e f aggegaig ae ifai de
ibed i e
i 4.2.1.
6.1 Dai eie
We a e he f wig
haa
eii
eie f i ai f a T. Thee eie
ae yi
a f he bad
a f wie e ehed ewk ag B ei
h 2004 whee a
e f wie e de wih iied
i
ai badwidh adi
i
ae wih he
wihi hei aii age. e ha hee eie
ed he e geea
de
ii f he T dai eeed i e
i 3.
(cid:15) The
i
ai de have iied badwidh. Theefe he be f
a ha
a be had ed by a de i iied.
(cid:15) A de
a y
i
ae wih he de ha ae wihi i aii age
i iediae eighb .
(cid:15) Ca
a igiae/eiae a ay de. Thee
a igiae h gh he i a
i ad a f a (cid:12)ie d ai h idi
aig a
i age f he ewk.
433
D a eig ea
(cid:15) A de a avai ab e badwidh i divided i w ege: he
a
hae ad
he
hae . The fe i ed e
a ad he ae
i
ae
ifai ad
eage.
(cid:15) The e
e avai ab e a a de i i
a
hae badwidh. e i f hi i
a
aed f ea
h
a ed via he de. The badwidh i f he de a
a
ah ae
ied h gh he d ai f a
a eab ih a
i
i. Th we
ide a
i
i wi
hed ewk whee badwidh i a
aed ed ed eab ih
a .
(cid:15) Ea
h de i de ed a a age. Evey age ha he ai f fwadig a
a he
eighb i be ieve i he (cid:12) de a ah he deiai wih he axi
a
hae badwidh avai abi iy.
(cid:15) We de(cid:12)e he ae f a de a a give ie a he ai f he
a
hae badwidh
i i ha a
aed he axi be f i ha i
a had e. Ea
h
age ha efe
kw edge f he ae f he de i eee ad eiae f he
ae f he age.
6.2 Ceaive Badwidh A
ai f Ca R ig
e
i 4.2 we de
ibed bad y hw he ig age a
ae badwidh i e e
e
e
a
a bewee he
e ad deiai de. ee a e deai ed de
ii f
i eeai f hi ye i eeed. T hi ed we ideify f he di
i
f e
i 4.2 ha he di(cid:11)ee a
i f age ae i ee f ye f ifai
eage: i e e fwad a
a
ii e e
e
a
a
iii e e
d a
a
ad iv e e ea ie age
. he f wig hee
d
a
iviie ae e abaed. T fa
i iae hi di
i we (cid:12) de(cid:12)e he f wig e f
vaiab e ed i de
ii: e f a age i he ewk fa
; :::; a
g whee
1
= j j;
e f
eighb f age a
= j
j; a f
i
f ea
h age
i
i
i
i
i
i
a
whee
:
! [0; 1℄. The f
i
; d i a
eiae f he badwidh
i
i
i
i
i
avai abi iy i he de a ah a
2 via i eighb a
2
; i: a
a
d
i
de ae f a
; B
: a e f de ae fj ; k; :::; z g; E
d: feedba
k ifai
i
j;k ;:::;z
i
vided by a
ab i eiae f
a
hae badwidh i avai ab e ve a e
i
a
f a
; E
d a e f E
d feedba
k eiae f a a
i
; R
a ewad
ed
d
i
j
j
i
i
i
by a
f de ae va e ha i e
eive via
i
ai f he de. TT
i
R hweve hee i a di(cid:11)ee
e i he eeeai f he eiae f ha de
ibed
abve. Thi i be
a e i TT R a age e a a
i deede fea e f
i
whi
h aie he
a e(cid:11)e
f i a
i eiae badwidh avai abi iy. Thi
f
i i baed he badwidh age eve he ik
e
ig a
i eighb .
i
Th if he vaiab e
dee he i f a
a
a
hae badwidh beig ed
i;
i
f
a ha ae ed via i eighb a
he a
eiae f badwidh avai abi iy
i
ea
h a deiai a
via a
i:
e
; ; d. hi e
i \high" if he va e f
d
i
i;
i;
i;
ea ed ve a
eai a
iviy widw ie ieva i he a i e ha a
eai
heh d \ w" hewie e
i 4.2.3 e
i(cid:12)e he va e ed f hee aaee i
exeie.
A eage f ye
ai he f wig ifai: id f he
a
e
d
a
ad deiai a
he e f id f he de h gh whi
h i ha bee ed ah
d
434
Shaig fai f Diib ed eaig i AS
he ie f igiai
he g ba
k ad e
a ab e ieva f he
ive
a ad deedig he ifai haig aegy ed a e f de ae va e
B
. The
e g ba ie i eeeed by . A
ye eage
ai iead f
k ;:::;z
he
a
e id he id f he de whee he a. A a give e a
i1
T
T
a be
d e he age id a a ii e h
e he
a deiai
a
ha he
e age a
whi e a
e he id e h
e he
a
d
T
i
i 1
T
e a
. weve a
dee he
e f he ad he
a
e de.
T
Wih hi i a
e Fig e 3 hw he age a
iviie. U e
eivig a
ie 1 a
age ay a
he
k whehe i ha a
aed badwidh he
a fwadig
e
i
ha aed beyd he axi e ie ii ie 2.
eihe
ae he fwadig
e i ed ad a
ai a
ie 5 geeaed f
ie 4 efe he
i
d
aiae
a eeeed by
he evi age. U e
eivig
ie 33 a
d
age fee he e a
aed
a
hae badwidh ie 34 ad ed he ae
d
he evi age ie 36 i a
i ea
hed.
d
f eihe f he
dii i ie 3 ae ai(cid:12)ed a
(cid:12)
he
k if a ha
ed by
i
he
kig if ah i
de i w id. f i ha ie 7 a
geeae a
ye eage
i
ie 8 ad
e a ea y = 10:9
ie 10 whee x i he h
f a
i
x1
he ed f he i.e. whee he wa (cid:12) dee
ed. Thi ea y a i added
ie 11. A he
e de id i
i e be he id f he de whee he
wa dee
ed ie 12. The hi
eage i aied he evi age
he ie 13. U e
eivig a
ye eage ie 50 a
e he ea y i
i
dae i i eiae f he deiai de; deedig whehe he aegy
ed i TT R e f he dae ie 52 54 ee
ive y ge exe
ed.
S be e y if a
i he age whee he wa dee
ed ie 55 i
e a ew
i
ea y whi
h i a exeia y de
eaig f
i f he dia
e f ha de f he
ed add hi
ad de a
ae he e a
aed badwidh f hi
a ie 57.
Fia y i ai he
he evi age ie 58. The i ii hee i ha he
f he a de i f he ed he e i i eib e f
a ig he ig
.
e
e he we he ea y i ge.
13
Afe
he
kig f
y
e a
he
he
k if i i he deiai f he
e
a ie
i
14. f i i i wi a
ae e badwidh i ie 16 ad ed a
geeaed f
f ea
ied befe he evi age he e f hi
a ie 21
i e
hi
e f a
ai ie 42 ad edig he eage ie 49 i a
i ea
hed;
a whi
h i a
ee
i
i i eab ihed. weve if a age e he TT R
aegy befe edig he
ie 21 i e he eiae baied a g wih he
ie 20; hw
h eiae ae agaed i TT R i de
ibed h y.
he w ifai haig he ii
ha we have deiged baed TC de
ibed
i e
i 6.3.1 ad 6.3.2 a age a
aa
he i w de ae i he eage
14
i
ie 18 ad 47.
Ea
h ah age a
ig TC e
eivig a
ha
i
13. exeie we have beved ha hi he ii
baia y ed
e he be ad ize f .
e
e i e(cid:11)e
ive y ed
e waef e f e
e i
e eee ed da i f a
a
ah.
14. The TC i
i e adv
ae ifai diib i y afe ak
ei. weve a ak exe
i
e
a fai e.g.
a ig fai ig i exa e a i
ai. e ha i hi i ai TC
a e he ak fai e a he eve igge ifai diib i. We have ed hi
e ad
id
ed ifai diib i afe
a fai e. S fa we have beved ha he BW a
ai
a iy i bee ha diib ig ifai y afe
a
ee.
435
D a eig ea
//f a eage i
hae C
f age a
i
i
1. if
// CA FRWARD ESSAGE
2.
if i == 0jj >
// avai ab e badwidh
ive
3.
// ie ex
eeded e ie
4.
deiveF
;//geeae d
a eage
d
5.
if a
;
; // if d
a evi age
i 1
d
6.
e e
7.
if
// dee
ed i
e
a e
8.
deiveF
; //geeae ea y eage
9.
x
: Ed C a
;//dia
e f a
f he ed
i
i
10.
ea y = 10:9
;//ea y a x = 0 f he (cid:12) age
x1
11.
:add ea yea y;
12.
:eS
e dea
; //e he
e de id he de whee he wa dee
ed
i
13.
if a
;
;//ea ie evi age
i 1
14.
ifa
== a
//hi de i he deiai
d
i
15.
deiveF
;//geeae
e
a eage
16.
a
aeU iBadwidh
;
17.
if TC //ee e
i 6.3.1 ad 6.3.2
18.
aed i
;
19.
if TT R//ee e
i 4.2.3
20.
e eage
;// a
ae eiae
21.
if a
;
;// if
e
evi age
i 1
22.
e e
23.
//e e
a eighb baed eiae f he give deiai
24.
a
e e
eighb
; a
;
j
i
d
25.
eA
aeCa
; //e a
ae badwidh
26.
if TT R
27.
aed
e
; a
; a
;// ed w eiae
i
i;i1
d
i1
28.
if a
;
;// fwad
a e e
ed age
j
29.
if R //ee e
i 4.2.2
30.
//e e
ed eighb a
e i
a eiae f badwidh avai abi iy
j
31.
E
a
ij ; ax
a
; a
j
d
j
d
k
a
2
k
j
32.
a
; a
1 (cid:11)
a
; a
(cid:11)E
a
;
i
d
j
i
d
j
j
d
33. e e if
// DR CA ESSAGE
d
34.
deA
aeCa
;//de a
ae badwidh
d
35.
ifa
6= a
//if
e
i
d
36.
if a
;
;
i 1
d
d
37. e e if
// CECT ESSAGE
38.
if TT R ad
//ewad eage i TT R
39.
R
eRewadB
;
i
a
;:::;a
i1
d
40.
e
; a
; a
1 (cid:11)
e
; a
; a
(cid:11)R
;
i
i;i1
d
i1
i
i;i1
d
i1
i
41.
e e
42.
a
aeU iBadwidh
;
43.
if TC //ee e
i 6.3.1 ad 6.3.2
44.
//
e ewad ig w eiae ad he
45.
R
eRewadB
;
i
a
;:::;a
i1
d
46.
a
; a
1 (cid:11)
a
; a
(cid:11)R
;
i
d
i1
i
d
i1
i
47.
aed i
;
48.
ifa
6= a
//if
e
i
49.
if a
;
;
i 1
50. e e if
// EATY ESSAGE
51.
if TT R
52.
e
; a
; a
1 (cid:11)
e
; a
; a
(cid:11)
:ea y ;
i
i;i1
d
i1
i
i;i1
d
i1
53.
e e
54.
a
; a
1 (cid:11)
a
; a
(cid:11)
:ea y ;
i
d
i1
i
d
i1
55.
ifa
6= a
i
56.
d
: Ed C a
;//dia
e f a
f he ed
i
i
57.
ea y = 10:9
;
:add ea yea y; deA
ae
;//de a
ae badwidh
d1
58.
if a
;
;
i 1
59. if TT R //ak e
i(cid:12)
TT R
60.
iikU age
; //ea e age f a
a
2
i
i;j
j
i
61.
if dae ieva == 0
62.
f a a
aed
63.
deiveF
; //
eae ewad eage
64.
if a
;
; //ai ea a g hi
e
i 1
65. 1;
Fig e 3: Age a
i i ee vai eage ye
436
Shaig fai f Diib ed eaig i AS
ai he de ae ifai aied f he age
e a ewad R
a
i
a f
i f he e f
i
aed ae va e B
aied i
ie
a
;:::;a
i1
d
45. Se
i 6.3.1 ad 6.3.2 de(cid:12)e w he ii
f
a
aig hi ewad va e. eihe
ae hweve he ewad i ed dae he i eiae ie 46. The dae e
f w adad eaig whee (cid:11) i he eaig ae ee e
i 4.1. Th he age
eaive y v ee hei
a ifai e ahe ive hei eiae f
he beved de ae.
he he had if a
i he deiai f hi
a i e e
e f i eighb
i
ex
dig he e f whi
h i e
eived he
fwad he
a e e ie 24. Thi
i de by de(cid:12)ig a babi iy diib i ve a
e f eiae f i eighb .
i
ai
a he babi iy f e e
ig a eighb a
i give by:
j
a
=
:
22
j
a
;a
i
j
d
ex
(cid:28)
a
2
;a
6=a
k
k
i
i 1
a
;a
i
d
k
ex
(cid:28)
e ha e ai 22 efe he e e
i e
hai whe TC R i ed.
f
TT R i ed he
e
; a
; a
e a
e he va e i e ai 22. ee (cid:28)
i
i;j
d
j
i he \eea e" aaee age:
[01℄ ad
hw
h he e aive di(cid:11)e
e
e bewee vai eiae w d a(cid:11)e
he e aive babi iie f e e
i he
a e he (cid:28) he age he kewe. Thi i a adad he ii
a ed B za ex
ai Waki 1989 babi ii
a y
he bewee a eaive i. S be e y
a
e a
ae a badwidh i f i
a
hae ie 25 ad fwad
a
ie
i
j
28.
he R aegy he e e
ed eighb a
ed a
wih i w eiae
j
i
E
a
f badwidh avai abi iy e he deiai a
ie 31 whi
h i
j
d
d
e a ij; ax
a
; a
efe e
i 4.2.2. The e eig age a
e
j
d
k
i
a
2
k
j
hi eiae E
a
dae i i eiae
a
; a
ie 32. TT R a
j
d
i
d
j
i
aed i eiae he
ie 27 befe fwadig he ae a
. Thi i hw he
j
e e
e f eiae ge agaed a g wih he
a fwadig e e i TT R.
w we de
ibe a be f a
iviie ha a age ef y if TT R i ed.
Fi a
i he ik age eve f a i eighb ie 60. Thi ifai i ed
i
e he va e f e
j 2
a de
ibed befe. Se
d evey dae ieva
i;j
i
ie e ie 61 a
a edig ewad eage a g he ah f he
a ha
i
eiaed a a
d ig ha eid f he
ha i had ed d ig ha eid.
i
Se
i(cid:12)
a y hee ewad eage ae aa g he
ye eage ex
e ha
badwidh i a
aed whe a age e
eive e a ed badwidh beig a
aed
whe a
i e
eived ie 38. T diig ih f
we dee hee eage a
f de
ii i Fig e 3. Evey dae ieva f ea
h
ed ve he a
ieva a
eae a
ie 63 ad ed hi he eighb ie 64 whi
h i he
i
iediae ea de a g ha
a ah. U e
eivig a
a age
e a
ewad ig a aggegai f he ifai f he be e ah de ie 39 ii a
TC ad dae i va e wih hi ewad ie 40 befe edig he
ea.
h d be ed hee he ifai ed by he age
e he ewad i TT
R ae he eiae ha he age had aeded whi e fwadig he
. Theefe
he e B
i ie 57 eee eiae ad a
a de ae. We have
a
;:::;a
i1
d
ed he ae aggegai ehd f TT R a de
ibed i e
i 6.3.2.
437
D a eig ea
The abve di
i
ed a e deai ed de
ii f he geea dai
de
ii f e
i 3. A idei(cid:12)ed i ha de
ii ad beved i he abve ye de
ii he de
ii e e
a e
i(cid:12)
eighb fwad a
a i
ii
a i deeiig
hw e(cid:11)e
ive he ye i i
ef y ig
a . Si
e hi de
ii i ake baed
he eiae he e a
ae y hey e(cid:13)e
he e de badwidh avai abi iie he
bee ifed ae he de
ii ake by a age. i ehaied ha he ifai
haig aegy ay a key e i deeiig he eiai a
a
y. he f wig we
f ae w i e he ii
f TC ha ae ed de(cid:12)e he
eRewad f
i
i i ai.
6.3 TC Shaig e ii
The di
i age iea
i i e
i 6.2 ex ai hw he age ig TC de ay
aiig he ifai i a
a i
e
ed. hi
ae he age a g he
a
ah
a aggegae he ifai e
eived f he \dwea" ad a ha if
ai he evi ah age. We have f aed w i e he ii
f ifai
aggegai baed TC viz. aveage he ae eiae eed TC A ad ake
he ii ae eiae eed TC . Thee ae de
ibed i he f wig.
15
6.3.1 TC f Aveage Caa
iy TC A
Fig e 3 e
eivig he
a age ig he TC A he ii
e a ewad
va e R
by aveagig he ae f a de he e f a
a
a:
i
i
d
k
k2fa
;:::;a
g
i1
d
R
=
;
23
i
a
; a
i
d
whee a
; a
i he h
hi e f a
a
. Thi de
ibe hw he f
i
i
d
i
d
eRewad f Fig e 3 i i eeed. S be e y hi ewad i ed dae i
i eiae. Th he eiae ae daed wih he ifai ab he e
e
age he \dwea" de he ah f hi
a .
6.3.2 TC f ii Caa
iy TC
Thi i ii a TC A: b iead f aveage avai ab e
aa
iy he ii avai ab e
aa
iy i ed a ewad. F exa e age a
ig he TC he ii
e he
i
ewad a:
R
= ii 1; :::; d:
24
i
Thi i a e
evaive eiae f badwidh avai abi iy ha he aveage
aa
iy
de . Th i ha he advaage ha he babi iy f a ded
a d e age
veeiaig he badwidh avai abi iy i ed
ed. Thi he ii
i a ed aggegae
eiae i TT R i exeie.
15. ee i h d be ed ha he fa aa yi i e
i 5 i baed he age havig eaae
eiae f idivid a age. aiaiig a aggegae eiae a e f age hweve ed
e he
aia
exiy a de
ii ie i i ai. A eaive y f exa e wih eiae
f idivid a de he ie
exiy f deeiig he ea
ah bewee ay w de i
adai
i he be f de i he gah he w
ae ie f Dijka ig e
e he
ah a gih Ce eie Rive Sei 2001.
438
Shaig fai f Diib ed eaig i AS
7. Exeiea Eva ai
Baed he
eai di
ed i e
i 6 we have
d
ed a eie f exeie
eii
a y eva ae he e(cid:11)e
ivee f TC
aed he be
hak R ad TT
R. hi e
i we a by e eaig he ea e
he a idi
a f ye
efa
e. S be e y he e baied f he exeie hee ea e ae
aa yed.
7.1 efa
e ea e
The f wig ea e ae
he eva ae he efa
e f a give
i
ai
aegy i ivig eaig i a T.
7.1.1 be f S
ef Ca
The vea b je
ive f he
eaive age ewk i axiie he be f
e
f y ed
a give he e f e
e
a
hae badwidh avai ab e he ae a
whi
h ew
a igiae ad he d ai
a eai
e
ed h d e
e. Th
he aveage be f
ef
a deeie hw
ef a eaig a gih ha bee
give he abve aaee va e. ye we e
d he a be f
a ha
have igiaed ad he a be f he
a ha have bee
ef y ed
C . Th he va e x =
deeie he fa
i f
ef
a ed ve he
C
ie eid i whi
h he ea ee ae ake.
addii ea ig he be f
a
ef y ed we kee a
k f he
be f
a ha
d be
e
ed if he age had g ba kw edge ab he ewk
badwidh avai abi iy ad if
a ig
d be de iaae y. Thi i
ed
by g ba y ea
hig f he avai abi iy f a ah f ea
h
a . Thi ea
h i de ia
ae y a he begiig f ea
h i ai ie e. We e hi eeia y idea ii
ed e he \aae Ze De ay Sea
h" ZDS. Si
e i a
i
e i ake a (cid:12)ie
be f ie e
e
a
a he ZDS i eeaed a evey ie e i eihe
i (cid:12)d a ah he
a i ded/
e
.
ae f he fe
e he C
izd
16
i i
eeed by 1. The va e x
=
give a had e ii he
a
C
izd
izd
17
e ae de he give
dii.
The eaig aegie ae
aed bh agai
he ab e
e ae x va e ad he e
eage deviai f
e ae f ZDS
x
x
izd
x
izd
. i ifeaib e f ay a
i
a ye aai he ZDS
e ae be
a e i
a
i
e hee i a (cid:12)ie a f de ay
e
a
a a ed he iaae
e
i i ZDS. A he age i a a
i
a ye ae
e
a
a by f
wadig i e h a a ie baed hei idivid a eiae f he w d ae. ZDS
he he had ake a g ba ad a
ae view f he eie ewk (cid:12)d he ea
ah.
16. Th ZDS i g aaeed (cid:12)d a ah if e i f d by he a
a ig a gih b e
eai y
vi
e vea.
17. e ZDS i a ia ea e. Si
e i ie
e
a
a a a e igiae i i
eeia y a geedy aegy. Se he
hed ig aegy ay ha e e f f kahead befe
aeig a
e a
a ay ef ZDS.
439
D a eig ea
7.1.2 S
ef R e f Diffee egh
The
a
e ae ei
de
ibed i e
i 7.1.1 ea e he vea ae f
ef
a i he ye. Thi i a ia efa
e ea e be
a e i idi
ae hw he
ye ef a a bad
a e. S i i a idi
a f hw
ef a
i
ai
aegy i i ivig he ye efa
e. weve i de idi
ae hw e(cid:11)e
ive
he aegy i a
e
ig
a a a give dia
e i
e he
e ae ei
a
a i he ye.
ae
a
a be e ied be ed deiai de ha ae a vai
dia
e f he de f igi. Ca deied f de ha ae a h dia
e ae
e aive y eaie e ha he f he away. Thi i be
a e e a
ae eiae f
he ad a de ha ae eae
a be aiaied. A hw i he aa yi f e
i 5
he fahe a age he ge i he de ay f he ifai aive ad h he e
dae ae he eiae. Theefe i i e bab e f a
a be ed
ef y
h a dia de. A
i
ai aegy ha a w bee
e ae a ge
dia
e
a heefe be
ideed e
ee ha ahe ha a
hieve a e
e ae a g dia
e a he
dii beig e a .
Th we ea e he be f
ef
a
e
i f vai dia
e bewee
he
a
e ad deiai de. e e
i(cid:12)
a y he ii h
ay d
bewee he
e ad deiai f a
a i
ed g ba kw edge f he ewk
gy i ed ea e hi dia
e ad he
e
e C
f
a a dia
e
d
d i i
eeed if
h a
a i
ef y
e
ed. hi ae a
e ae f
a a a dia
e d
a be
ed a x
=
whee
ad f he be f
a
d
d
d
C
d
igiaed wih a
e deiai dia
e f d. F di(cid:11)ee va e f d heefe he
di(cid:11)ee x
va e
d be ed
ae he abi iy f di(cid:11)ee
i
ai aegie
d
a
e
a a di(cid:11)ee egh.
7.1.3 Rewad fai eage
The age i ye e di(cid:11)ee eage ye f
i
ai e
i 6.2 e e
ae hee. weve he ia ag hee ae he e ha
ay he ewad
ifai by whi
h he
a kw edge f e age i aied ahe. Thi i
be
a e ig hi ifai he age dae hei i eiae f he ewk ad
eve . Theefe hee eage
ib e die
y he a iy f eaig ad he ve
a efa
e f he ye i a
aig e
e a
e
a . R f exa e hi
i he eage ha a age e
eive f i eighb afe hadig ve he ae a
a
fwadig e e ee e
i 4.2.2 f deai . Theefe a ea
h
a fwadig e a
ew eage i geeaed by he
a
ed age ad aied he
a
ig age.
he TC baed de he he had hee ae he
ye eage aied afe
a
a
e
ha
ai he ay ewad ifai e
i 6.3.1 ad 6.3.2 have
he deai . e hi i y e eage geeaed by he deiai de afe evey
ef ig ad aied ea a g he
a ah.
A h gh evea he
18
ye f eage e.g.
ad
ae ex
haged bewee he age i i he e
d
age ha
ai he ewad ha a(cid:11)e
he eaig a iy he . he he had
18. We have e
e y
d
ed a dy whee TC hae ifai a afe a
a fai
e
. Thi
di(cid:12)
ai bvi y i
eae he eage ae a h gh i keeig i we ha ha f R ad
TT R. weve i a w bee badwidh a
ai ha he TC eed i hi ai
e ad a
dea wih fai e dee
i ehig he
e TC i
aab e f D a e a . 2005.
440
Shaig fai f Diib ed eaig i AS
Fig e 4: 36 de ieg a gid
i i eeai f TT R a age whi e fwadig a
a ai a g wih
he
a i w eiae f he badwidh avai abi iy a g ah he
a deiai.
S be e y hi ifai i ed dae he va e f he age ee Se
i 7.2
f deai . Theefe hee eage a(cid:11)e
he eaig f age i he way ha
de
i TC. e
e he eage ae f TT R i ea ed by
ig hee eage.
The be f
h eage
a heefe be ed a a ea e f he eAE
ie
y f
a give
i
ai aegy he we he be f eage he highe i he
eAE
ie
y a ig a give va e f e he efa
e ea e
h a
a
e
ae. e e
i(cid:12)
a y he a be f ifai ex
hage eeeed by ay
f aiig he age
a kw edge e ahe i
ed a evey T ie e
d ig a i ai. The va e =
heefe give he ae f eage aied i he
T
eie ye d ig he ieva T . The a i ai i divided i evea ieva ad
he va e f ve ea
h
h ieva geeae a vea ie vaiai f he eage ae.
The di(cid:11)ee
i
ai aegie ae he
aed agai vai eage ae.
a 50 de ad gah
b 100 de ad gah
Fig e 5: Rad ewk gie
7.2 Re ad Aa yi
hi e
i he efa
e f R TC A TC ad TT R ae
aed agai
he ea e de
ibed i e
i 7.1.
441
D a eig ea
Exeie ae
d
ed a be f di(cid:11)ee ewk gie. he f w
ig we e e ad bevai baed e f he. Fig e 4 ad 5 hw
ee
ive y a 36 de ieg a gid ad w ad y geeaed gie a 50 de
ad gah Fig e 5a ad a 100 de ad gah Fig e 5b. The 36 de i
eg a gid gy ha bee ed i evi ae he a i
ai f R i ewk
ig ee e
i 2.2 f a di
i hee ae ad h we
he i ake va id
ai. T veify
i a
a wide age f gie we eed he ae
agai ad gah (cid:12)g e 5a ad 5b hw w
h exa e f di(cid:11)ee ize.
19
a (cid:12)g e he de ae beed f eae f efee
e. The edge bewee de idi
ae
ha he de ae wihi ea
h he adi age. e ha he ad gah ae de
iged
h ha ay de i iked y he wihi a
eai axi adia dia
e
whi
h i ae he aii age f wie e de.
The f wig aaee va e wee ed f a exeie eed he
efh e
hewie aed: eaig ae (cid:11) = 0.03 B za ex ai eea e (cid:28) = 0.1
a
e ie f 36 50 ad 100 ie e f he gie i (cid:12)g e 4 5a ad 5b ee
ive y. We a wed f a age
a e ie i he bigge gie give a wa
e f
he age ize f he ewk. A aveage
a d ai f 20 ie he e ie wa ed.
The \ ad" i he ewk i e by aigig a babi iy wih whi
h
a igiae a evey
ie e i he ewk. Thi
a igiai babi iy wa vaied dy he e(cid:11)e
f
di(cid:11)ee ewk ad he efa
e. Ca wee a wed igiae ad eiae
ay ad y e e
ed de. We have eed he aegie ig bh i a
a
ad i ai whee he
a igiai babi iy i aiaied he ae h gh a
i ai ad ii a dyai
a y
hagig ad i ai whee he
a igiai
babi iy i
haged d ig he
e f a i ai . We have ed di(cid:11)ee be
ad va e f he
a igiai babi iy
hage i a ig e i ai e he
e(cid:11)e
f vai degee f ad (cid:13)
ai he efa
e f TC R ad TT R.
hee dyai
eig he
a igiai babi iy i
haged a e a ieva i
a i ai . A ig e i ai aed f 500000 ie e f he gy i
Fig e 4 f 1000000 ie e f he gy i Fig e 5a ad f 2000000 ie e
f he gy i Fig e 5b. Re ae aveaged ve 10 i ai hee (cid:12)g e
ae aii
a y igi(cid:12)
a a he 95
(cid:12)de
e eve . A evey de had a axi
a
hae
aa
iy f 10 i.
The f wig e
i di
he e he vea
e f he vai
i
ai he ii
i
e
ig
a i e
i 7.2.1 he e(cid:11)e
ivee f hee he ii
i
ef y
e
ig g dia
e
a i e
i 7.2.2 ad he vehead d e
i
aig eage i e
i 7.2.3.
7.2.1 efa
e Ca S
e Rae
We ai
iae ha he highe he a
a
y f eaed eiae f beved ae he
e
aab e he age wi be f ig
a he deiai via he aiae
ah. e
e i he highe wi be he
a
e ae ad he we he deviai
f he ZDS
e ae. he f wig he e f
a ad ae eeed
(cid:12) f wed by he baied f dyai
a y vayig ad.
19. Vai he gie wee ed wih vayig be f de ad
e
iviy ae ad he ae
geea ed i he e wee beved. e
e hee we e hee a e gie.
442
Shaig fai f Diib ed eaig i AS
Ca ad. We exeieed wih a aegie
a
ae: i he
a
e ae
ad ii he e
eage deviai f he ea ed
e ae f he ZDS de eady
ae
dii i.e whe he
a h gh i he ye ea
hed a eady va e.
The
20
aveage
e ae ad he aveage ZDS
e ae bh
ed d ig he eady
ae hae f he i ai ae f he aveaged ve 10 i ai f a give va e
f he
a igiai babi iy. F he
a
e ae i ea ed agai vai
a igiai babi iie e he ia
f ewk ad he
e ae. Thee
ea ee ae eeaed f ea
h f he hee gie.
Tab e 2: Ca
e ae f a aegie gy f Fig e 4
ad
R
TC A
TC
TT R
Avg Sdev ZDS Avg Sdev ZDS Avg
Sdev ZDS Avg Sdev ZDS
0.1
50.94 0.0048 72.66 50.98 0.0076 72.11 51.85 0.0052 70.25 25.74 0.0022 99.57
0.2
32.41 0.0034 60.69 32.64 0.0037 60.56 33.02 0.0039 58.86 19.94 0.0016 97.65
0.4
19.99 0.0018 55.37 20.27 0.0021 55.23 20.38 0.002 52.79 14.82 0.0009 89.44
0.6
14.87 0.0012 54.87 15.07 0.0011 53.44 15.08 0.0014 50.6 11.52 0.0012 85.2
Tab e 3: Ca
e ae f a aegie gy f Fig e 5a
ad
R
TC A
TC
TT R
Avg Sdev ZDS Avg Sdev ZDS Avg
Sdev ZDS Avg Sdev ZDS
0.1
57.31 0.0042 81.75 58.15 0.0052 81.67 58.49 0.0052 80.9 12.18 0.0027 99.99
0.2
37.11 0.002 69.75 37.32 0.0028 68.73 37.65 0.0026 68.33 10.21 0.0011 99.8
0.4
22.98 0.0012 61.44 23.35 0.0012 62.6 23.51 0.0013 60.43 7.46 0.0007 99.8
0.6
17.17 0.001
58.8 17.47 0.0009 61.29 17.49 0.0008 57.53 6.11 0.0004 99.8
Tab e 4: Ca
e ae f a aegie gy f Fig e 5b
ad
R
TC A
TC
TT R
Avg Sdev ZDS Avg Sdev ZDS Avg
Sdev ZDS Avg Sdev ZDS
0.1
35.79 0.0032 66.52 35.86 0.0034 65.58 37.34 0.003 62.65 8.53 0.0031 99.88
0.2
24.17 0.0017 57.41 24.56 0.002 59.18 24.8 0.0014 54.27 6.12 0.0013 99.83
0.4
15.88 0.0011 54.01 16.16 0.0007 55.53 16.18 0.0008 50.72 4.79 0.0006 99.9
e deai Tab e 2 hw he aveage eady ae
e ae a
hieved by he
di(cid:11)ee aegie ad by ZDS ed a gide hee aegie f di(cid:11)ee ewk ad
20. e T a i
ai eee a dyai
ye whee de badwidh avai abi iy
hage wih
ie ew
a ae a
ed ad exiig
a eiae. The va e eiae he badwidh avai abi iy.
S a badwidh avai abi iy
hage d he va e. weve he
a
e ae i a vea
ye ea e whi
h ea
he a eady ae whe a
a
a igiai babi iy i ed.
443
D a eig ea
a igiai babi iie i he gy f Fig e 4. Tab e 3 ad 4 hw he ae
ea ee f he gie i Fig e 5a ad 5b ee
ive y. The e idi
ae
ha he aveage eady ae
a
e ae hw de
\Avg" i hee ab e
a
hieved wih TC diae ha f he he aegie de di(cid:11)ee ewk ad.
F exa e i Tab e 2 whe he ad i 0.1 TC a
hieve a eady ae aveage
a
e ae f 51.85 TC A a
hieve 50.98 R 50.94 ad TT R 25.74. A
idi
aed by he aa yi f e
i 5 TC aiai e dae ifai f he
ewk ae. Th ig TC he age ae
aab e f akig bee ifed de
ii
f fwadig a
a whi
h i e e a highe ike ihd f
ef
e
i. Thi
i e(cid:13)e
ed i he aii
a y igi(cid:12)
a highe
a
e ae a
hieved by TC ve a
he aegie. ai
a i
e he ii
aa
iy he ii
ei
veeiai
f he de badwidh avai abi iy i geeae a igh y bee
e ae ha he aveage
aa
iy he ii
.
0
-2
-4
-6
-8
-10
-12
)
R
Q
o
t
e
v
i
t
a
l
e
r
(
S
D
Z
I
m
o
r
f
n
o
i
t
a
i
v
e
d
f
o
t
n
e
m
e
v
o
r
p
m
i
%
PTC-A / QR
PTC-M / QR
PTC-A / TPOT-RL
PTC-M / TPOT-RL
-15
-20
-25
-30
-35
-40
-45
-50
-55
-60
)
L
R
O
P
T
o
t
e
v
i
t
a
l
e
r
(
S
D
Z
I
m
o
r
f
n
o
i
t
a
i
v
e
d
f
o
t
n
e
m
e
v
o
r
p
m
i
%
-14
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
-65
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Call origination probability
Call origination probability
a TC agai R
b TC agai TT R
Fig e 6: vee f
a
e ae deviai f ZDS e aive R ad TT R
gy f Fig e 4
We have a ea ed he ivee i he
e ae va e a
hieved f he
TC baed ifai aegie e aive he f R ad TT R. T d (cid:12) he
e ae deviai f ZDS i
ed a =
ee e
i 7.1.1. S be e y
x
x
izd
x
izd
R
T C
T T R
T C
ad
ae
aed
ad
a:
ad
T C
T C A
R
T T R
R
T T R
ee
ive y. The gah i Fig e 6 hw he e aive ivee f he aveage
e
ae deviai f ZDS a
hieved by ig TC ve R Fig e 6a ad ve TT R
Fig e 6b whe he gy i Fig e 4 i ed. e ha i hi (cid:12)g e a geae egaive
va e idi
ae e f a deviai f he ZDS e aive R TT R ad he
e
a bee efa
e f TC. F hee gah a igi(cid:12)
a ivee i beved i
TC e aive bh R ad TT R de di(cid:11)ee va e f ewk ad aii
a
igi(cid:12)
a
e i eed a he 95
(cid:12)de
e eve . F exa e ee he w
edig
\ad" = 0.1 i Tab e 2 wih a
a igiai babi iy f 0.1 he deviai f he
a
e ae f he ZDS i 26.2 =
f TC ad 29.9 =
70:25
72:66
70:25 51:85
72:66 50:94
444
Shaig fai f Diib ed eaig i AS
1
0
-1
-2
-3
-4
-5
-6
-7
)
R
Q
o
t
e
v
i
t
a
l
e
r
(
S
D
Z
I
m
o
r
f
n
o
i
t
a
i
v
e
d
f
o
t
n
e
m
e
v
o
r
p
m
i
%
PTC-A / QR
PTC-M / QR
PTC-A / TPOT-RL
PTC-M / TPOT-RL
-20
-25
-30
-35
-40
-45
-50
-55
-60
-65
)
L
R
O
P
T
o
t
e
v
i
t
a
l
e
r
(
S
D
Z
I
m
o
r
f
n
o
i
t
a
i
v
e
d
f
o
t
n
e
m
e
v
o
r
p
m
i
%
-8
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
-70
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Call origination probability
Call origination probability
a TC agai R
b TC agai TT R
Fig e 7: vee f
a
e ae deviai f ZDS e aive R ad TT R
gy f Fig e 5a
)
R
Q
o
t
e
v
i
t
a
l
e
r
(
S
D
Z
I
m
o
r
f
n
o
i
t
a
i
v
e
d
f
o
t
n
e
m
e
v
o
r
p
m
i
%
2
0
-2
-4
-6
-8
-10
-12
-14
0.1
PTC-A / QR
PTC-M / QR
PTC-A / TPOT-RL
PTC-M / TPOT-RL
-25
-30
-35
-40
-45
-50
-55
)
L
R
O
P
T
o
t
e
v
i
t
a
l
e
r
(
S
D
Z
I
m
o
r
f
n
o
i
t
a
i
v
e
d
f
o
t
n
e
m
e
v
o
r
p
m
i
%
0.15
0.2
0.25
0.3
0.35
0.4
Call origination probability
-60
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Call origination probability
a TC agai R
b TC agai TT R
Fig e 8: vee f
a
e ae deviai f ZDS e aive R ad TT R
gy f Fig e 5b
445
D a eig ea
f R. Theefe he
e ae d e TC i 12.37 =j
j
e ZDS
29:9
26:2 29:9
e aive ha f R. The
e ae deviai f he ZDS f TC A a eai we
e aive R. Sii a y
aig TC wih TT R a he ae ad eve TC
a
hieve a 26.2 deviai f ZDS whi e he
edig (cid:12)g e f TT R i 74.15
99:57 25:74
26:2 74:15
=
. Theefe he
e ae d e TC i 64.66 =j
j
e
99:57
74:15
ZDS e aive ha f TT R. The efa
e f TC diae ha f TC A
a h gh bh ef bee ha R ad TT R be
a e f he
evaive a e f
he ii
aa
iy he ii
aed he aveage
aa
iy a aed abve. Thee
bevai f he eghe aa yi i e
i 5 ha by vidig bee a iy
eiae TC ef
e he ZDS ha R. A addiia bevai i ha
wih i
eaig ad he e aive ivee f TC ve R ad TT R ed
e.
Th i Fig e 6a he ivee f TC ve R i 12.37 wih ad f 0.1 whi e
i i y 3.7 wih ad f 0.6. Sii a y i Fig e 6b TC i 64.66 bee ha
TT R a ad 0.1 b 18.82 bee a ad 0.6. e ha a i
eae i ad i ie
ed
i i he ie bewee
eive ak
eig eide i
e
a igiae e
fe e y wih i
eaed ad. A ex aied i e
i 5 he a e he va e f he ieva
bewee
eive ak
eig eide he e dae ae he eiae geeaed.
Theefe wih a i
eae i he ad he efa
e di(cid:11)ee
e bewee TC ad he
he aegie de
eae.
We beve idei
a ed i he efa
e f TC e aive he be
hak
wih he he gie. Fig e 7 hw he e f he gy i Fig e 5a ad
Fig e 8 f he gy i Fig e 5b.
Dyai
a y Chagig ad.
a he eady ae
a
e ae ea ed
wih
a ad hee we ee he ie vaiai f he
a
e ae a he ewk
ad (cid:13)
ae. Thi
a e he eivee f he ye
a
e ae dyai
a y
hagig ad eve give a ai
a ifai haig aegy. T hi ed Fig e 9a
hw he ie vaiai f he
a
e ae f TC R ad TT R a he
a
igiai babi iy i i
eaed f 0.1 0.6 i a i ai ig he gy f
Fig e 4.
deae ha he TC i hi
ae we hw TC ; TC A i ex
ded
21
f he e i
e i ef igh y we ha TC ad bee ha R ad TT
R
a
e ae eai he highe bh whe he ad eve i 0.1 befe ie = 50
ad whe i i
eae 0.6 afe ie = 50. The
a
e ae f a aegie (cid:11)e a
d wih he i
eae i ad eve i
e he de have y iied badwidh a
ae
a a e ha we have beved evi y wih
a ad ee Tab e 2. We f he
exeieed wih i e ad (cid:13)
ai ve a i ai . Fig e 9b hw he
e whe he ad eve
hage bewee (cid:12)ve di(cid:11)ee va e: 0.1 0.2 0.4 0.6 ad 0.8
i ha de.
a
ae TC i beved have he highe
a
e ae f
TC R ad TT R. e ha i hee e we have ed w di(cid:11)ee ye
f ad (cid:13)
ai: e age i
eae 0.1 0.6 ad i
a y i
eaig. Thee
a f
e be evea he ae f ad vaiai
h a ad (cid:13)
ai
f wig e
i(cid:12)
babi iy diib i e.g. i b hee ae
ideed i hi
wk. hi
ex we ideify ha he ia
f a ad
hage i ha he age have
e ea he ew eviea
dii ha
a
a be a
ed e(cid:11)e
ive y. The
21. We have exeieed wih a he he gie de
dii f dyai
a y
hagig ad. The
bad ae beved i he e ae idei
a a
a f he. e
e we
he e a e
gy e e i hi ae.
446
Shaig fai f Diib ed eaig i AS
evee
ae i hi i ai i ha f i
eae i ad be
a e a highe ad dead e
eAE
ie a
ai f iied e
e. Th havig hw ha TC ef he
be
hak de hi
dii we eviage ha ii a bad ed w d be beved
f he ae.
e
t
a
r
s
s
e
c
c
u
s
l
l
a
C
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
PTC-M
QR
TPOTRL
0
10 20 30 40 50 60 70 80 90 100
e
t
a
r
s
s
e
c
c
u
s
l
l
a
C
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
PTC-M
QR
TPOTRL
0
10 20 30 40 50 60 70 80 90 100
Time
Time
a Wih 2 degee f dyai
b Wih 5 degee f dyai
Fig e 9: Tie vaiai f
a
e ae f R TC ad TT R wih ewk
ad (cid:13)
ai gy f Fig e 4
-
M
C
T
P
d
n
a
R
Q
f
o
s
e
t
a
r
s
s
e
c
c
u
s
l
l
a
c
n
i
e
c
n
e
r
e
f
f
i
D
%
-2
-4
-6
-8
-10
-12
-14
-16
-18
-20
1
2
3
4
Degree of dynamism
min
avg
max
5
6
Fig e 10: S ay aii
f
a
e di(cid:11)ee
e bewee R ad TC wih e
wk ad (cid:13)
ai gy f Fig e 4
w we ai aie he e(cid:11)e
f dyai
a y
hagig ad he ewk
a
e ae give a ifai haig aegy. dig we (cid:12) deigae he be
f ad eve i a i ai a he \degee f dyai". F exa e Fig e 9a ha
a degee f dyai f 2 ad Fig e 9b ha 5 degee f dyai. The f a give
degee f dyai we
e he e
eage di(cid:11)ee
e f he
a
e ae f R
447
D a eig ea
TT R ig TC a he bae ie a:
ad
whee
i
T C
T C
R
T C
T T R
T C
he ie vayig
a
e ae. e ha hi di(cid:11)ee
e ea e i a ie deede.
T aie he ivee f
a
e ae ig TC we (cid:12)d he ii he
ea ad he axi f hi di(cid:11)ee
e ve evey ie ieva d ig whi
h he ad eve
eai
a. Thee aii
ee he
a
e ae ivee age wihi he
ie ieva whe he ad eai a a
eai eve . S be e y he ea ad he
adad deviai f ea
h f he ii he ea ad he axi di(cid:11)ee
e ve a
h ieva ae
ed. Thi e geeae he ay f he di(cid:11)ee ivee
age a
a
h ieva . F exa e i Fig e 9a he ii ea ad
axi e
eage di(cid:11)ee
e ae
ed i he ieva whee ad = 0.1 ad i he
ieva whee ad = 0.6. S be e y he ea ad adad deviai f hee w
e f di(cid:11)ee
e aii
ae
ed. Fig e 10 hw he abve eied ea e f
a degee f dyai ed i exeie whe R i
aed TC . Si
e
TT R i a way beved a
hieve a we
a
e ae ae ha bh TC ad
R we ex
de he ay
ai f TT R wih TC i hi ae. A g
he hiza axi f Fig e 10 i he degee f dyai whi e he ay aii
f
he e
eage
e ae di(cid:11)ee
e ae a g he vei
a axi. hi (cid:12)g e a degee f
dyai 2 idi
ae he ad eve i
haged f 0.1 0.2; 3 idi
ae a
hage f 0.1 0.2
0.4; 4 idi
ae 0.1 0.2 0.4 0.6; ad 5 idi
ae 0.1 0.2 0.4 0.6 0.8. The egaive va e f
he
a
e ae di(cid:11)ee
e idi
ae
i we ha ha f
. The (cid:12)g e hw
R
T C
ha he ga bewee he ii ad he axi di(cid:11)ee
e ed
e wih he be f
degee f dyai. F ia
e hi ga i axiae y 8 wih degee f dyai
2 ad i i ab 4.5 whe he degee f dyai i 5. Thi i be
a e wih a highe
be f ad
hage wihi a give i ai hee i e ie f ay aegy
e ea he
hage i he evie. e
e he
a
e ae d aai hei
eady ae va e a i Tab e 2. evehe e he aveage aii
f he
a
e
ae di(cid:11)ee
e hw ha TC a
hieve a igi(cid:12)
a y highe
a
e ae a
a
degee f dyai. F exa e hi i ab 10.2 wih 2 degee f dyai ad 9.5
wih 5 degee f dyai.
Thee bevai eif
e hyhee ha de he give i ai evie
he ak
ei ifai haig de wih ii
aa
iy ewad a
hieve
he be
a
e ae ag a aegie de di(cid:11)ee ewk ad de
dii f
bh ai
ad dyai
a y (cid:13)
aig ad. F exa e wih
a ad he deviai
f he
e ae f he ZDS i 12.37 we i TC ha R ad 65
we i TC ha TT R. F he wih dyai
a y
hagig ad TC a
hieve
a aveage 10 ivee i
a
e ae ve R wih (cid:12)ve di(cid:11)ee ad eve
hage
i a i ai .
7.2.2 efa
e S
e Rae f Ca f Diffee egh
The
a
e ae va e eed i e
i 7.2.1 idi
ae bee efa
e f TC
ve R ad TT R. addii ea e he e(cid:11)e
ivee f a ifai haig
he ii
i
e
ig a
a a deiai ha i a a give dia
e f he
e
he ea e x
ee e
i 7.1.2 f i de(cid:12)ii i ea ed f i
eaig va e f d.
d
Re f he
a ad exeie ae eed i hi e
i. Wih dyai
a y
hagig ad he ay aii
a de
ibed i e
i 7.2.1 f
a
e ae
idi
aed bee efa
e f TC ha he be
hak f a va e f d. Si
e we have
448
Shaig fai f Diib ed eaig i AS
Tab e 5: Ca
e ae a vai dia
e gy f Fig e 4
ad
Saegy
i h
1
2
3
4
5
6
7
8
9
10
0.1
0.2
0.4
0.6
R
0.87 0.76
0.63
0.51
0.42
0.37
0.355 0.343 0.344
0.34
TC A 0.863 0.748 0.632 0.516 0.426 0.374 0.355 0.345 0.348 0.354
TC 0.857 0.753 0.643 0.534 0.438 0.381 0.362 0.356 0.353 0.365
TT R 0.633 0.532 0.356 0.218 0.164 0.148 0.126 0.094 0.071 0.059
R
0.785 0.592 0.44
0.31
0.222 0.178 0.158 0.147 0.147 0.139
TC A 0.773 0.583
0.44
0.315 0.231 0.181 0.164 0.155 0.151 0.151
TC 0.768 0.583 0.444 0.326 0.237 0.187 0.168 0.155 0.151 0.152
TT R 0.584 0.462 0.266 0.139 0.101 0.098 0.086 0.066 0.056 0.048
R
0.693 0.421 0.266 0.161 0.103 0.074 0.064 0.0563 0.053
0.05
TC A 0.675 0.417 0.272
0.17
0.111 0.081 0.068
0.06
0.055 0.058
TC 0.667 0.416 0.274 0.177 0.113 0.081 0.068 0.061 0.057 0.057
TT R 0.496 0.375 0.198 0.086 0.059 0.063 0.054
0.04
0.037 0.029
R
0.632 0.326 0.185 0.103 0.062 0.045 0.037
0.03
0.029 0.027
TC A
0.61
0.323
0.19
0.111 0.068 0.047
0.04
0.035 0.032
0.03
TC 0.596 0.322 0.194 0.113 0.069 0.049 0.039 0.035 0.032 0.03
TT R 0.418 0.301 0.148 0.065 0.043 0.046 0.037 0.023 0.015 0.013
Tab e 6: Ca
e ae a vai dia
e gy f Fig e 5a
ad
Saegy
i h
1
2
3
4
5
6
7
0.1
0.2
0.4
0.6
R
0.886 0.761
0.6
0.504 0.428 0.379
0.353
TC A 0.876 0.746 0.586 0.495 0.426 0.381
0.365
TC 0.873 0.753 0.601 0.511 0.443 0.398 0.369
TT R 0.187 0.129
0.11
0.102 0.104 0.139
0.203
R
0.774 0.57 0.371 0.268 0.197 0.161
0.141
TC A
0.77
0.561 0.369 0.265
0.2
0.163
0.145
TC 0.766 0.564 0.375 0.272 0.207 0.173 0.159
TT R 0.214 0.127 0.085 0.065 0.055
0.06
0.061
R
0.653 0.383 0.203
0.12
0.782
0.58
0.492
TC A 0.648 0.385 0.208 0.126 0.828 0.631
0.53
TC 0.64
0.384 0.212 0.13 0.87 0.662 0.545
TT R 0.202 0.109 0.054 0.025 0.021 0.024
0.048
R
0.58 0.286 0.133 0.712 0.428 0.305
0.249
TC A 0.566 0.289 0.14
0.781 0.481 0.347
0.287
TC 0.557 0.287 0.143 0.802 0.504 0.371 0.305
TT R 0.177 0.095 0.046 0.017 0.011 0.0070 0.0050
a eady hw bee
e ae f TC wih dyai
ad
dii i e
i 7.2.1 we
have ex
ded he
a
e ae a di(cid:11)ee dia
e de he ae
dii i hi
e
i.
449
D a eig ea
Tab e 7: Ca
e ae a vai dia
e gy f Fig e 5b
ad
Saegy
i h
1
2
3
4
5
6
7
8
9
10
11
12
0.1
0.2
0.4
0.6
R
0.894 0.755 0.616 0.452 0.324 0.247 0.208
0.175 0.149
0.12
0.087
0.082
TC A
0.89
0.749 0.589 0.432 0.317 0.247
0.22
0.197 0.169 0.141
0.111
0.1
TC 0.88
0.728 0.598 0.443 0.336 0.267 0.238 0.223 0.191 0.164 0.124 0.122
TT R 0.22
0.172
0.14
0.104 0.079 0.057 0.051
0.045 0.037 0.029
0.028
0.029
R
0.847 0.647 0.47 0.296 0.189 0.123 0.098
0.079 0.061 0.047
0.036
0.029
TC A 0.837 0.628
0.45
0.287 0.178 0.124
0.1
0.087 0.071 0.056
0.046
0.039
TC 0.823 0.622 0.448 0.292 0.192 0.136 0.109 0.096 0.086 0.066 0.053 0.046
TT R 0.202 0.158
0.11
0.072 0.047 0.032 0.029
0.025 0.016 0.012
0.008
0.006
R
0.77 0.502 0.315 0.17
0.094 0.057 0.041
0.031 0.022 0.017
0.011
0.012
TC A 0.766 0.484 0.298 0.161
0.09
0.057 0.042
0.034 0.026 0.021 0.0145 0.015
TC 0.746 0.476 0.303 0.172 0.098 0.061 0.046 0.039 0.031 0.024 0.017 0.018
TT R 0.189 0.141
0.09
0.056 0.034
0.02 0.0167 0.011 0.006 0.003
0.002 1.0E 4
R
0.71
0.4 0.225 0.112
0.06
0.03
0.024
0.018 0.012 0.009
0.006
0.005
TC A 0.706 0.393 0.216 0.109 0.057 0.031 0.025 0.0197 0.014 0.0117 0.008 0.0075
TC 0.688 0.386 0.221 0.113 0.061 0.036 0.028 0.023 0.017 0.013 0.009 0.008
TT R 0.171 0.127 0.079 0.047 0.027 0.015
0.01
0.006 0.003 1.0E 4 1.0E 4 1.0E 4
e deai ab e 5 6 ad 7 hw he va e f x
f di(cid:11)ee va e f d he \i
d
h
"
de di(cid:11)ee ewk ad f he gie f (cid:12)g e 4 5a ad 5b
22
ee
ive y. Thee ab e idi
ae he f wig ed. F vey h dia
e whee he
va e f d i e ha 3 R a
hieve a igh y bee
e ae ha TC . weve
he
e ae a
hieved a ge dia
e wih eihe TC baed aegie fa ef
ha f R. Addiia y TC ef bee ha TT R f a dia
e. T bai a
eae i
e f he e aive advaage f he TC baed aegie ve R ad TT R
i hi
ex we he e aive ivee f x
ha i a
hieved by ig he TC
d
aegie ve R ad TT R
ad
ee
ive y whee
R
T T R
x
d
x
d
T C Y
T C Y
R
T T R
x
x
x
x
d
d
d
d
Y
a be A f he aveage
aa
iy he ii
f he ii
aa
iy he ii
. Fig
e 11 hw he e
eage
hage f x
a
hieved by ig TC ve R Fig e 11a
d
TC A ve R Fig e 11b TC ve TT R Fig e 11
ad TC A ve
TT R Fig e 11d a i
eaig va e f he ii h
bewee
a
e
ad deiai de i he ieg a gid gy f Fig e 4. Ea
h i ea
h f hee
(cid:12)g e i f a di(cid:11)ee va e f he
a igiai babi iy. The gah f Fig e 12
ad 13 hw idei
a ea e ig he gie f Fig e 5a ad 5b ee
ive y.
F
ig (cid:12)g e 11a ad 11b i i beved ha f a give
a igiai b
abi iy whe he
a deiai ae vey
e he
a
e i.e. whe he \ii
h
" axi ha va e f 1 ad 2 R ef igh y bee ha eihe TC
TC A idi
aed by he a egaive deviai. F exa e a ad 0.6 R a
hieve
ab a 5 ivee ve TC i
e
ig
a a de 1 h away ee Fig
e 11a. weve wih i
eaig dia
e bewee he
a
e ad deiai de
22. The axi va e f he ii h
i he ey f he
edig ewk.
450
Shaig fai f Diib ed eaig i AS
i.e. whee he \ii h
" i 3 ad abve he ae f
ef
e
i i
h highe f he TC baed aegie ha R he deviai va e ae iive. F
exa e i Fig e 11a f a ad va e f 0.6 TC a
hieve e ha 15 ive
e ve R i he
e
i ae f
a ha have hei deiai a ea 8 h
away f he
e. A hi e aive ivee i beved geea y i
eae wih
i
eaig dia
e. Th he TC ifai haig aegie ae e e(cid:11)e
ive i
e
ig
a f whi
h he
e ad deiai de ae fahe aa. (cid:12)g e 11
ad 11d TC ad TC A ef bee ha TT R a a dia
e ad de a
ad va e.
20
15
10
5
0
-5
-10
500
400
300
200
100
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
load 0.1
load 0.2
load 0.4
load 0.6
1
2
3
4
5
6
7
8
9
10
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
20
15
10
5
0
-5
load 0.1
load 0.2
load 0.4
load 0.6
1
2
3
4
5
6
7
8
9
10
Minimum hop count
Minimum hop count
a TC agai R
b TC A agai R
load 0.1
load 0.2
load 0.4
load 0.6
load 0.1
load 0.2
load 0.4
load 0.6
500
450
400
350
300
250
200
150
100
50
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
0
1
2
3
4
5
6
7
8
9
10
Minimum hop count
0
1
2
3
4
5
6
7
8
9
10
Minimum hop count
TC agai TT R
d TC A agai TT R
Fig e 11: Ca
e ae a vai dia
e gy f Fig e 4
T ex ai hi advaage f TC e ha a h gh he
a
e
i
e i
exe
ed i e by i e age wh e hei idivid a eiae fwad a
a
he fwadig de
ii f he age wh ae
e he
a igi ae e
ii
a i
deeiig whehe i wi be
ef y ed he deiai. Thi i be
a e if hee
age a fwadig a
a i a die
i whee hee i a high badwidh
a
y he
451
D a eig ea
25
20
15
10
5
0
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
-5
1
1000
800
600
400
200
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
load 0.1
load 0.2
load 0.4
load 0.6
2
3
4
5
6
7
Minimum hop count
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
16
14
12
10
8
6
4
2
0
-2
-4
load 0.1
load 0.2
load 0.4
load 0.6
1
2
3
4
5
6
7
Minimum hop count
a TC agai R
b TC A agai R
load 0.1
load 0.2
load 0.4
load 0.6
load 0.1
load 0.2
load 0.4
load 0.6
1000
800
600
400
200
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
0
1
2
3
4
5
6
7
Minimum hop count
0
1
2
3
4
5
6
7
Minimum hop count
TC agai TT R
d TC A agai TT R
Fig e 12: Ca
e ae a vai dia
e gy f Fig e 5a
452
Shaig fai f Diib ed eaig i AS
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
60
50
40
30
20
10
0
-10
load 0.1
load 0.2
load 0.4
load 0.6
1
2
3
4
5
6
7
8
9 10 11 12
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
50
40
30
20
10
0
-10
load 0.1
load 0.2
load 0.4
load 0.6
1
2
3
4
5
6
7
8
9 10 11 12
Minimum hop count
Minimum hop count
a TC agai R
b TC A agai R
1000
800
600
400
200
0
load 0.1
load 0.2
load 0.4
load 0.6
1
2
3
4
5
6
7
8
9 10 11 12
s
e
t
a
r
s
s
e
c
c
u
s
n
i
t
n
e
m
e
v
o
r
p
m
i
%
1000
800
600
400
200
0
load 0.1
load 0.2
load 0.4
load 0.6
1
2
3
4
5
6
7
8
9 10 11 12
Minimum hop count
Minimum hop count
TC agai TT R
d TC A agai TT R
Fig e 13: Ca
e ae a vai dia
e gy f Fig e 5b
453
D a eig ea
de he ha wi be a b ia de
ii a he begiig f a ak exe
i
e.
hi
ae i i e ike y ha he
a wi be fwaded whee hee i badwidh ef
ad heefe he fwadig
e w d eiae he
a w d be ded. Wih
e dae eiae a
hieved by TC a beved i he aa yi f e
i 5 a
age i
aab e f akig bee ig de
ii i e f fwadig i he aiae
die
i ha ig R. Thi i why we beve highe
a
e ae i he TC baed
aegie ha R whe
a have be ed a ge dia
e. eve wih i
eaig
dia
e
ef ig be
e e diAE
i
e he fahe he deiai he e
dae ae he eiae. Give hi he i
eaig e iive deviai f he
x
va e wih i
eaig dia
e a beved i Fig e 11 idi
ae ha TC i e
d
aab e e a(cid:11)e
ed ha eihe R TT R e
a a g dia
e. TC
i e e(cid:11)e
ive i a
ig g dia
e
a ha TC A a h gh bh ae bee ha R
ad TT R. Thi i evide by he igh y highe iive deviai f TC ve R
Fig e 11a ha ha f TC A ve R Fig e 11b by he highe iive deviai
f TC ve TT R Fig e 11
ha ha f TC A ve TT R Fig e 11d
f he ae
a igiai babi iy.
The abve bevai f Fig e 11 a h d i (cid:12)g e 12 ad 13. Addiia y i
a f hee (cid:12)g e i i beved ha f a give dia
e d he deviai f he x
va e
d
ae geea y highe f a highe
a igiai babi iy. T j ify hi bevai e
ha he
e ae f ay give aegy de
eae wih i
eaig ad ee e
i 7.2.1
f hi e be
a e wih e
a igiaig he be f ded
a i
eae
i
e he de have y a iied a f
a
hae badwidh. evehe e he
i
ai aegy ha geeae bee eiae w d eab e he age
e wih
he i
eaig ad bee by aiaiig a highe
a
e ae. Thi i idi
aed by
he bevai ha i
eaed ad ia
R ad TT R e ha he TC baed
aegie i.e. he de
eae i he
e ae i e i R TT R ha i TC
TC A wih i
eaig ad. e
e he deviai f he x
va e bewee TC ad
d
he be
hak i
eae wih ad. F exa e i Fig e 11a a a h egh f 8 he
deviai f x
i ab 2.5 wih ad 0.1 ad e ha 15 wih ad 0.6.
d
The bevai i hi e
i
a be aied a ak
ei aegie ae
e
aab e f
e
ig
a a ge dia
e ha he be
hak aegie ad exhibi
i
eaed e(cid:11)e
ivee i a
hievig hi agai high ewk ad. A e aive ivee
f e ha 50 i beved i he ae f
ef
a
e
i f TC ve R
a dia
e f 12 h i a 100 de ad gah wih a high ewk ad ee he gah
f a ad f 0.6 i Fig e 13a. F he ae aaee va e TC a
hieve a age
ivee ve TT R f e ha 1000 ee Fig e 13
.
7.2.3 efa
e fai eage Rae
The a be f eage ee e
i 7.1.3 f i de(cid:12)ii aied i he eie
ewk i ea ed evey T ie e f a i ai. The aveage be f
h eage
e i ie i
ed ve he eady ae. Thee ea ee ae eeaed f ea
h
f he hee gie. Re f
a ad exeie ae eed (cid:12) f wed by
he f a dyai
a y
hagig ad.
Ca ad. Tab e 8 hw he eage ae baied i ea
h f he aegie de
eady ae
dii i he gy f Fig e 4. Tab e 9 ad 10 hw he ae f he
gie f Fig e 5a ad Fig e 5b ee
ive y. a f hee e i i beved
454
Shaig fai f Diib ed eaig i AS
ha f ay give
a babi iy he eage ae de
\Avg" i igi(cid:12)
a y
we i he TC aegie ha bh R ad TT R. F exa e i Tab e 8 i i 0.25
f TC 0.26 i TC A 0.39 f R ad 0.523 f TT R wih a ad eve f 0.1.
A addiia bevai i ha he e aive ed
i f he eage ae
aed \ Savig" hw hee va e a
hieved by ig TC be
e e
ed
wih i
eaig ad. Thee va e ae
ed a
R
T T R
T C Y
R
T C Y
T T R
j
j
j
j
whee eee he aveage eady ae eage ae f a give
i
ai aegy
f a give ad ad Y
a be eihe A . F exa e i Tab e 8 a ad 0.1
TC ha ab 35.6 =j
j e ae f eage ha R ad ab 52
0:251 0:39
0:251 0:523
0:39
=j
j e ha TT R. weve hi avig i he eage ae i
eae
0:523
ab 80.3 =j
j e aive R ad ab 72.8 =j
j e aive
1:45
1:05
0:285 1:45
0:285 1:05
TT R whe he
a babi iy i 0.6. Thi i be
a e wih i
eaig ewk ad
he i
eae i he be f eage i bh R ad TT R i
h highe ha he
i
eae i he TC baed aegie e.g. i Tab e 8 he eage ae i
eae f 0.39
1.45 i R a 272 i
eae f 0.523 1.05 i TT R a 101 i
eae ad
f 0.251 0.285 i TC y a 13.5 i
eae a he ad i
eae f 0.1
0.6.
Tab e 8: fai eage ae f a aegie gy f Fig e 4
ad
R
TC A
TC
TT R TC A /
TC /
TC A /
TC /
Avg Sdev Avg
Sdev Avg
Sdev Avg Sdev
R
R
TT R
TT R
Saegie
Savig
0.1
0.39 0.0048
0.26
0.0019 0.251 0.0019 0.523 0.0023
33.33
35.64
50.28
52.00
0.2
0.66 0.0079 0.281
0.002 0.275 0.0022 0.683 0.0029
57.42
58.33
58.85
59.73
0.4
1.1
0.01
0.289 0.0018 0.284 0.0016 0.891 0.0036
73.73
74.18
67.56
68.12
0.6
1.45 0.0125 0.2894 0.0017 0.285 0.0015 1.05 0.0077
80.04
80.34
72.43
72.85
Tab e 9: fai eage ae f a aegie gy f Fig e 5a
ad
R
TC A
TC
TT R TC A /
TC /
TC A /
TC /
Avg Sdev Avg Sdev Avg
Sdev Avg Sdev
R
R
TT R
TT R
Saegie
Savig
0.1
0.33 0.0023 0.23 0.0012 0.23 0.0014 0.759 0.0054
30.30
30.30
69.69
69.69
0.2
0.55 0.0048 0.26 0.0013 0.25 0.0012 1.42 0.0074
52.73
54.55
81.69
82.39
0.4
0.9 0.0049 0.27
0.001 0.26 0.0009 2.163 0.0091
70.0
71.11
87.51
87.97
0.6
1.2 0.0054 0.272 0.0007 0.267 0.0012 2.70 0.0106
77.33
77.75
89.92
90.11
e ha i R a ah age ai a ew ifai eage he fwadig
ea age a ea
h e f he
a fwadig
e. Th wih i
eaed ad hi
aegy i
a age i
eae i he be f eage be
a e hee ae ay e
a
be ed. TT R a de fwad i ae ifai whi e ig a
a . S
a h gh ewad diib i
i TT R evey dae ieva ie e he
ae ifai eage agaed d ig
a ig
ib e wad he eage
455
D a eig ea
Tab e 10: fai eage ae f a aegie gy f Fig e 5b
ad
R
TC A
TC
TT R TC A /
TC /
TC A /
TC /
Saegie
Savig
Avg Sdev Avg Sdev Avg
Sdev Avg Sdev
R
R
TT R
TT R
0.1
0.58 0.0117 0.214 0.0009 0.21 0.0014 1.42 0.018
63.10
63.79
84.92
85.21
0.2
0.83 0.0112 0.235 0.001 0.229 0.001 2.19 0.0091
71.69
72.41
89.27
89.54
0.4
1.26 0.0098 0.249 0.001 0.241 0.0008 2.98 0.0061
80.24
80.87
91.64
91.91
0.6
1.63 0.0095 0.253 0.0007 0.245 0.0008 3.37 0.0079
84.48
84.97
92.49
92.72
vehead i
ed i he ye. S whe age be f
a eed be ed a high
ad va e he be f
h eage i
eae. The TC baed aegie he he
had aai a igi(cid:12)
a avig i he eage ae by de ayig he ifai aii
i a
a
e
ad he aiig a ig e eage f he deiai he
e
de wih y daig he ewad ifai a ea
h e f he eage agai.
Thi eve ay ifai ex
hage f
ig f
a ha fai
e
. Bh
R ad TT R he he had w d i i
he eagig
eve if a
a
(cid:12)a y fai
e
. addii he TC baed aegie ave e eage f beig
ex
haged ha R ad TT R i
he f a
a e. Be
a e he age
ig TC ai eage y afe a
a i
e
ed ad he
a e
ae ded a a
a i fwaded eage ae eveed f beig aied
i.
Dyai
a y Chagig ad. Sii a he
a
e ae ea ee f e
i 7.2.1 he ie vaiai f he ifai eage ae ae e
ded f a aegie
whe he ewk ad (cid:13)
ae dyai
a y i a i ai .
e
t
a
r
e
g
a
s
s
e
M
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
PTC-M
QR
TPOTRL
0
10 20 30 40 50 60 70 80 90 100
e
t
a
r
e
g
a
s
s
e
M
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
PTC-M
QR
TPOTRL
0
10 20 30 40 50 60 70 80 90 100
Time
Time
a Wih 2 degee f dyai
b Wih 5 degee f dyai
Fig e 14: Tie vaiai f eage ae f R TC ad TT R wih ewk ad
(cid:13)
ai gy f Fig e 4
456
Shaig fai f Diib ed eaig i AS
-
M
C
T
P
d
n
a
R
Q
f
o
s
e
t
a
r
e
g
a
s
s
e
m
n
i
e
c
n
e
r
e
f
f
i
D
%
max
avg
min
500
450
400
350
300
250
200
150
100
50
1
2
3
4
Degree of dynamism
5
6
-
M
C
T
P
d
n
a
L
R
T
O
P
T
f
o
s
e
t
a
r
e
g
a
s
s
e
m
n
i
e
c
n
e
r
e
f
f
i
D
%
500
450
400
350
300
250
200
150
100
max
avg
min
1
2
3
4
5
6
Degree of dynamism
a R
aed TC
b TT R
aed TC
Fig e 15: S ay aii
f eage ae di(cid:11)ee
e bewee R TT R ad TC
wih ewk ad (cid:13)
ai gy f Fig e 4
Fig e 14a hw he eage ae vaiai f TC R ad TT R whe
he ad eve
hage f 0.1 0.6 i a i ai ig he gy f Fig e 4.
Fig e 14b hw he ae whe he ad i vaied a 0.1 0.2 0.4 0.6 ad 0.8 afe e a
ieva f ie i he
e f a i ai . Bh f hee (cid:12)g e hw ha TC
ha a igi(cid:12)
a y we eage ae ha bh R ad TT R de (cid:13)
aig
ad
dii. F he a he ewk ad i
eae dyai
a y bh R ad TT R
i
a age i
eae i he eage ae whi e he i
eae i TC i iigi(cid:12)
a. bh
R ad TT R ifai i aied d ig
a e a h gh ewad dae
i TT R afe evey dae ieva ie e. e
e a e
a igiae
wih i
eaig ewk ad he be f
h ifai agai i
eae i hee
aegie. i a beved ha wih i
eaig ad TT R ha a we eage ae
ha R. We ideify ha hi e aive advaage f TT R agai R i d e i
e
a
e ae ha R eed i e
i 7.2.1. Si
e TT R i e eAE
ie
ha R i
e
ig
a hi i ie ha i TT R a age be f
a ae
ae
ef ad i
e bh agae ifai whi e
a fwadig TT R
de i
a
h f a i
eae i he be f eage a i
ed by R wih
i
eaig ad.
a way ii a he
e ae e wih (cid:13)
aig ad he ay f he
di(cid:11)ee
e i eage ae bewee R ad TC ad bewee TT R ad TC
ae eeed i (cid:12)g e 15a ad 15b ee
ive y. bh f hee (cid:12)g e he degee f
dyai i ed a g he hiza axi whi e he e
eage i
eae f eage ae
ig R TT R agai TC i ed a g he vei
a axi. A beved i
he
e ae e f Fig e 10 i bh Fig e 15a ad 15b he ga bewee he
ii ad he axi di(cid:11)ee
e f eage ae ed
e wih i
eaig degee f
feed. F ia
e i Fig e 15a hi ga i ab 140 wih 2 degee f feed
whi e i i 60 wih 5 degee f feed. evehe e he ea di(cid:11)ee
e bewee he
eage ae f TC ad R ha bewee TC ad TT R i
eae wih
457
D a eig ea
i
eaig degee f dyai. F exa e he ea di(cid:11)ee
e i
eae f 151
321 i Fig e 15a whi e i i
eae f 200 284 i Fig e 15b. Thi idi
ae he
advaage f TC i e f aiaiig a iied be f eage vehead
aed
bh R ad TT R.
F hee bevai i
a be
ded ha he ak
ei aegie
y a
hieve a highe
a
e ae e
i 7.2.1 ad highe e(cid:11)e
ivee i
e
ig
a a ge dia
e de high ad e
i 7.2.2 ha he be
hak aegie b
a a
hieve hee a he exee f a igi(cid:12)
a y we ae f eage de bh ai
ad dyai
a y
hagig ad
dii. F exa e TC a
hieve a 80 avig i
eage ae
aed R ad a 72 avig
aed TT R i he gid gy
de high ewk ad. F he i a dyai
ad eig wih (cid:12)ve
hage i he ewk
ad TC ave ab 320 i eage ae ha R ad ab 284 ha TT R.
Th TC i hw be a e eAE
ie aegy wad deve ig a
eaive AS f
he diib ed e
e a
ai b e.
8. C
i ad F e Wk
Thi ae f
e he
ii
a i e f a wig age wih aia bevabi iy ef
fe
ive y
eae
ex ak ha e ie he ai
iai f i e age f
ef
ei. T a
hieve hi ga i i e ied ha he age ake hei a
i
baed eiae f he ae f he age ha a ak i ved i a
ie a
e. evehe e wih beig ab e die
y beve he ye ae i i diAE
f
he age deve
h eiae. A f he diAE
y i a
hievig he b je
ive i ha
ye ae dyai
heefe he ye ae ae b je
ed
i
hage.
e
e i
h
eai he age h d have a way f adaig hei eiae wih
h
hage.
Give hi we hw ha
eaive ifai haig i a a
i
a ad e(cid:11)e
ive e
h
ai a w he age eiae he beved ae. C ed wih hi e
hai
we e eaig d
e b ad (cid:13)exib e eiai f hee ae. ai
a he
e
i(cid:12)
ib i f hi ae ha bee he id
i f a ve ifai diib
i i
i e ak
ei ifai haig ive he eaig f he
ae eiae.
ai he
f eaig f eae eighb ifa
i we have eab ihed ig fa aa yi ha aegy i
aab e f aiaiig
e dae eiae by e ig a e ie y diib i f ifai. F he
we have i eeed a e f
i
ai he ii
baed he TC haig i
i e
be ed i a a
i
a i age e
e a
ai b e
a ig i a e ehe
ewk. We have
d
ed exeive eii
a die by
aig aegy agai
a e f be
hak a gih a
a wide age f eviea eig by e e
ig
di(cid:11)ee ewk gie ewk ad ad dyai
a y
hagig ad ae. Re
f hee die have hw ha
aed he be
hak
a
hieve a
highe
a
e ae 60 ad a ei efa
e f e ha 1000 i he
abi iy
ef y e
a g dia
e. ha bee a deaed ha
aegy aai hee ivee a a igi(cid:12)
a y we eage vehead 300
ha he be
hak.
We ehaie ha he TC haig i a geei
e
hai ive he eaig f
ae eiae by diib ed age. Thi i beved i he aa yi f e
i 5 whee
dai e
i(cid:12)
a i wee ade. weve e
i(cid:12)
i
ai he ii
h
458
Shaig fai f Diib ed eaig i AS
a TC ad TC A i hi ae
a be deiged a d hi e
hai f a give
b e. Th we eviage ha aegy i a i
ab e a wide age f ak dai
a di
ed i e
i 4. T veify hyhei we ae e yig a TC baed he ii
ef diib ed fa dee
i i a ewk ad ae baiig iig iiia e .
A addiia bee(cid:12) f ifai haig aegy i ha i i a
i
a y a i
ab e
i
e i i deve ed a d
h ea ii
a i a
a bevabi iy a
i ei
ed
wihi he
a ae a
e dyai
evie (cid:12)ie ie de ay f a ak be
eed by i e age i e e
e ad ae
y iv ved i he agai f ifai
bewee age. eve
i
ai aegy i a
ib i a
hie eaig
eea
h i geea i
e i i a a
i
a ad e(cid:11)e
ive ea f ivig eaig i ea w d
AS. hi
ex we e ha a key
ha ege i a
hie eaig eea
h i deve
a gih f e eia de
ii akig i
eai dai R e vig 2002. A
a way f vig hi b e i f ae a de
ea ied akv De
ii
e
DEC D Be
ke Zi beei ee G da 2003; Beei Giva ea
Zi beei 2002 vide he e
eay
he age ake a
i agai he
dyai
f he dai. hi
ex we be ieve ha
i
ai aegy
a
ib e igi(cid:12)
a y by geeaig high (cid:12)de iy i a DEC D ha de a
a
i
a dyai
AS. B i dig hi idea we a iveigae way f iegaig
i
ai aegy i
h a de ad dy he ia
he a iy f i
geeaed.
A
kw edge
i
ee hak he eviewe whe ifaive ad iighf
e have e
ab ed ive he ae baia y. We hak X dg Seve G Siada
Daahaaa ad Seve Baihwaie f hei
e ad ggei. Fia y we ae
b iged ESRC f f dig he je
bi e ade Ceaive Age ewk
CA GR/R32697/01 f whi
h hi wk i a a.
Refee
e
A D. . Bea ee . Badhaw . . D fee E. . i . b
k C. A.
S i . Tae A. Th C. W. 2002. Ca ii age exeie: iage
eai i ieaia
a ii. EEE e ige Sye 17 3 26{35.
Baxe . Ba e . . 1999. Die
gadie baed eif
ee eaig: . Ga
die eiai a gih. Te
h. e. Reea
h S
h f fai S
ie
e ad
Egieeig A a ia aia Uiveiy.
Be
ke R. Zi beei S. ee V. G da C. V. 2003. Taii ideede
de
ea ized akv de
ii
ee.
eedig f he Se
d eaia
i Cfee
e A Age ad iage Sye AAAS 2003 .
41{48 e b e. AC e ew Yk.
Beei D. Giva R. ea . Zi beei S. 2002. The
exiy f de
ea ized
f akv de
ii
ee. aheai
f eai Reea
h
27 4 819{840.
459
D a eig ea
B i ie C. 1999. Se eia ia iy ad
diai i iage ye.
eedig f he Sixeeh eaia i Cfee
e Ai(cid:12)
ia e ige
e
CA 99 . 478{485 S
kh .
Bya . A. ia . . 1993. a
ke ig i dyai
a y
hagig ewk: A
eif
ee eaig aa
h. Cwa . D. Tea G. A e
. Ed.
Adva
e i e a fai
eig Sye V . 6 . 671{678.
Ca ey . . Gae . 1999. C aia gaiai hey. Wei G. Ed.
iage Sye: A de Aa
h Diib ed Ai(cid:12)
ia e ige
e . 299{
330. The T e Cabidge A.
Ca G. D. Dig . 1998a. Ae: Diib ed igegei
f
i
ai ewk. a f Ai(cid:12)
ia e ige
e Reea
h 9 317{365.
Ca G. D. Dig . 1998b. Tw a
y a gih f be e(cid:11) ig i
daaga ewk.
eedig f he Teh ASTED eaia Cfee
e
aa e ad Diib ed C ig ad Sye . 541{546. ASTED/ACTA e.
Chaib daa B. 1995.
d ia a i
ai f diib ed A. C i
ai f he
AC 38 11 49{53.
C
kb D. eig . R. 1996. A
h: A diib ed ai(cid:12)
ia ie ige
e ye
f id ia a i
ai. ae G. . . eig . R. Ed. F dai
f Diib ed Ai(cid:12)
ia e ige
e . 319{344. Wi ey.
Che . R. eve e . . 1991. Teawk. 35 4 487{512. Se
ia e
Cgiive S
ie
e ad Ai(cid:12)
ia e ige
e.
Ce T. . eie C. E. Rive R. . Sei C. 2001. d
i a gih
2d edii.
ha. 24: Sig e S
e She ah. T e.
De
ke . 1995a. TAES: A faewk f evie
eeed aa yi ad deig
f
diai e
hai.
ae G. eig . Ed. F dai f
Diib ed Ai(cid:12)
ia e ige
e
ha. 16. Wi ey e S
ie
e.
De
ke . S. 1995b. Evie
eeed aa yi ad deig f
diai e
hai.
h.D. hei Uiveiy f aa
h e Ahe aa
h e.
De
ke . S. ee V. R. 1995. Deigig a fai y f
diai a gih.
eedig f he Fi eaia Cfee
e i age Sye . 73{80
Sa Fai
.
Dekea B. Zwi
h . We k . 2004. i
ad i Age Sye f a fa
ig V . 2744/2004 f e
e e i C e S
ie
e
ha. iage baed
e aig ad
hed ig i
ex f y
hai . 100 { 109. Sige
Ve ag eide beg.
D fee E. . ee V. 1991. aia g ba aig: A
diai faewk f di
ib ed hyhei fai. EEE Taa
i Sye a ad Cybeei
21 5 1167{1183.
D a . S. Daahaaa S. G S. R. eig . R. ea . 2004. Ce
aive ifai haig ive diib ed eaig. AAAS 04 wkh
eaig ad Ev i i Age Baed Sye . 18{23.
460
Shaig fai f Diib ed eaig i AS
D a . S. eig . R. ea . 2005. Shaig ifai f eaig baed
ewk badwidh eiai ad ewk fai e dee
i e.
eedig
f he F h eaia i Cfee
e A Age ad iage
Sye aea.
D a . S. ea . eig . R. 2003. Fidig iea
i ae ig
gii baed de
ii aegie. Wkig e f he CA 2003 wkh
Cgiive de ig f Age ad i Age ea
i . 46{55.
D a . S. Se S. 2003. Fig ab e aehi. Cgiive Sye Reea
h
4 3 211{221.
E D. G avi
. Weheke . 2004. we ye abi iy
: Reif
ee
eaig faewk. EEE Taa
i we Sye 19 1 427{ 435.
Feibeg E. S
hwaz A. 2001. adbk f akv De
ii
ee: de ad
A i
ai. we A
adei
b ihe.
Gz B. a S. 1996. C abaive a f
ex g a
i. Ai(cid:12)
ia
e ige
e 86 2 269{357.
edi
k C. 1988. R ig fai
RFC 1058. ewk Wkig G
ETF. h://www.ief.g/f
/f
1058.x.
ai . Dv i C. 2002. ah ad: A ea ee f ed ed avai ab e
badwidh.
eedig f he aive ad A
ive ea ee A Wkh
. 14{25.
eig . R. 1993. Cie ad
vei: The f dai f
diai i
i age ye. The w edge Egieeig Review 8 3 223{250.
eig . R. 1995. C ig
eaive b e vig i id ia i age
ye ig ji iei. Ai(cid:12)
ia e ige
e 75 2 195{240.
eig . R. B a S. 2003. Age baed
ye. EEE C Sye
agazie 23 3 61{74.
eig . R. a T. . Faai . 1998. Ade: A age baed aa
h
b ie
e aagee. AC SGD 27 4 32{39.
ag T. B ei
h S. 2004. Wie e eh ewkig. Rei y Wie e Dev
ee.
h://www.ei ye.
/ b/a/wie e/2004/01/22/wie eeh.h .
ai . Bake . 1999. ea ig badwidh.
eedig f he EEE FC
1 . 235{245.
ee V. R. Ea . D. 1988. Diib ed ieeai: A de ad exeie.
Bd A. . Gae . Ed. Readig i Diib ed Ai(cid:12)
ia e ige
e .
120{139. ga a fa Sa ae CA.
ahadeva S. a
ha e
k . Da T. Gavi A. 1997. Se f ivig fa
y i
ai ig
i ie aveage ewad eif
ee eaig.
eedig
f he F eeh eaia a
hie eaig Cfee
e . 202{210. ga
a fa.
ai e R. ee V. ig B. 2003. Ceaive egiai f f ea ie
diib ed e
e a
ai.
eedig f he Se
d eaia i C
461
D a eig ea
fee
e A Age ad iAge Sye AAAS 2003 . 576{583
e b e. AC e.
a ey . 1988. aagig he (cid:13)w f ie ige a. Rbi
ad C e egaed
a fa
ig 4 3/4 525{530.
a
h . G. Si . A. 1958. gaizai. h Wi ey S.
Ca R. A. 1996. idde ae ad eif
ee eaig wih ia
e baed
ae idei(cid:12)
ai.
EEE Taa
i Sye a ad Cybeei
a
B Cybeei
26 464 { 473.
ia . ae . . ae . 1999. Ceaig bi e age f dyai
ewk
ig.
eedig f he Sfwae Age f F e C i
ai Sye.
Sige Ve ag. SB 3 540 65578 6.
i
he T. . 1997. a
hie eaig
ha. 13: Reif
ee eaig.
Gaw i .
a ak . V. D. Bake A. D. C ak S. . 2001. The aaia age a
hie
e: F
a fa
ig e iee age baed ye deig. egaed C e Aided
Egieeig 8 1 45{58.
ehki . Savva V. 2002. Reif
ee eaig f adaive ig.
eedig
f he eaia i Cfee
e e a ewk C.
Ribei V. Cae . Riedi R. Savha S. edi
k B. Baai k R. 2000.
ifa
a
aAE
eiai. TC Cfee
e TaAE
de ig ad
aagee.
Ri
h C. Side C. . 1997. CAGE: Whe age
abae wih e e.
eedig f he Fi eaia Cfee
e A Age Age 97
. 284{291.
R e S. . vig . 2002. Ai(cid:12)
ia e ige
e: A de Aa
h 2d edii.
ha. 17: akig C ex De
ii. ei
e a .
She . ee V. Cave . 2003. iiizig
i
ai
i a diib ed
Bayeia ewk ig a de
ea ized D.
eedig f he Se
d eaia
i Cfee
e A Age ad iAge Sye AAAS 2003 .
678{685. AC e.
Sigh S. . aakk a T. da . . 1995. Reif
ee eaig wih f ae
aggegai.
Tea G. T ezky D. ee T. Ed. Adva
e i e a
fai
eig Sye V . 7 . 361{368. The T e.
Se . 2000. TT R a ied ewk ig.
eedig f C 2000 .
935{942.
Se . Ve . 1999. Tea aiied a e aii eif
ee eaig.
eedig f he Thid A a Cfee
e A Age . 206{212.
S R.
A ee D. Sigh S. a Y. 2000. i
y gadie ehd f
eif
ee eaig wih f
i axiai. Adva
e i e a fai
eig Sye 12 1057{1063.
S R. S. 1988. eaig edi
by he ehd f ea di(cid:11)ee
e. a
hie
eaig 3 9{44.
462
Shaig fai f Diib ed eaig i AS
S R. S. Ba A. G. 1998. Reif
ee eaig: A d
i Adaive
C ai ad a
hie eaig. T e.
Takahahi E. Taaka Y. 2003. A
i baed e(cid:11)e
ive badwidh a
ai e
hai.
Te e
i
ai Sye 24 2 323{338.
Tabe . 1997. Twad (cid:13)exib e eawk. a f Ai(cid:12)
ia e ige
e Reea
h 7
83{124.
Taeba A. S. 2003. C e ewk 4
edii.
ha. 5: The ewk aye.
h
ei
e a TR.
Tg . 2002. Reif
ee eaig f
a adii
ad ig de a iy
f evi
e
ai i iedia ewk. a
hie eaig 49 2 111{139.
Viwaaha R. Vahey . . 1997. Diib ed dee
i wih i e e: a
f daea .
eedig f he EEE V . 85 1 . 54{63.
Wa h W. E. We a . . 2003. De
ea ized y
hai fai: A ake
ad
eiive e i ibi aa yi. a f Ai(cid:12)
ia e ige
e Reea
h
19 513{567.
Waki C. . C. . 1989. eaig f de ayed ewad. h.D. hei y
h gy De
ae Uiveiy f Cabidge.
Waki C. . C. . Daya . 1992. Te
hi
a e: eaig. a
hie eaig 8
279{292.
Widw B. (cid:11) . E. 1960. Adaive wi
hig
i
i. WESC Cvei
Re
d a V . 96{104.
Wi ia R. . 1992. Si e aii
a gadie f wig a gih f
e
ii
eif
ee eaig. a
hie eaig 8 3 229{256.
Xiag Y. 1996. A babi ii
faewk f
eaive i age diib ed iee
ai ad iizai f
i
ai. Ai(cid:12)
ia e ige
e 87 1 2 295{342.
X a . ee V. Zi beei S. 2001. C i
ai de
ii i i age
eai: de ad exeie.
eedig f he Fifh eaia Cfee
e
A Age Age 01 . 616{623 ea .
463
|
1412.7223 | 2 | 1412 | 2016-03-21T01:37:57 | Safe Sequential Path Planning of Multi-Vehicle Systems via Double-Obstacle Hamilton-Jacobi-Isaacs Variational Inequality | [
"cs.MA",
"eess.SY"
] | We consider the problem of planning trajectories for a group of $N$ vehicles, each aiming to reach its own target set while avoiding danger zones of other vehicles. The analysis of problems like this is extremely important practically, especially given the growing interest in utilizing unmanned aircraft systems for civil purposes. The direct solution of this problem by solving a single-obstacle Hamilton-Jacobi-Isaacs (HJI) variational inequality (VI) is numerically intractable due to the exponential scaling of computation complexity with problem dimensionality. Furthermore, the single-obstacle HJI VI cannot directly handle situations in which vehicles do not have a common scheduled arrival time. Instead, we perform sequential path planning by considering vehicles in order of priority, modeling higher-priority vehicles as time-varying obstacles for lower-priority vehicles. To do this, we solve a double-obstacle HJI VI which allows us to obtain the reach-avoid set, defined as the set of states from which a vehicle can reach its target while staying within a time-varying state constraint set. From the solution of the double-obstacle HJI VI, we can also extract the latest start time and the optimal control for each vehicle. This is a first application of the double-obstacle HJI VI which can handle systems with time-varying dynamics, target sets, and state constraint sets, and results in computation complexity that scales linearly, as opposed to exponentially, with the number of vehicles in consideration. | cs.MA | cs | Safe Sequential Path Planning of Multi-Vehicle Systems via
Double-Obstacle Hamilton-Jacobi-Isaacs Variational Inequality
Mo Chen, Jaime F. Fisac, Shankar Sastry, Claire J. Tomlin
emergency supplies to disaster-struck areas that are otherwise
difficult to reach [4].
6
1
0
2
r
a
M
1
2
]
A
M
.
s
c
[
2
v
3
2
2
7
.
2
1
4
1
:
v
i
X
r
a
Abstract -- We consider the problem of planning trajectories
for a group of N vehicles, each aiming to reach its own
target set while avoiding danger zones of other vehicles. The
analysis of problems like this is extremely important practically,
especially given the growing interest in utilizing unmanned
aircraft systems for civil purposes. The direct solution of this
problem by solving a single-obstacle Hamilton-Jacobi-Isaacs
(HJI) variational
inequality (VI) is numerically intractable
due to the exponential scaling of computation complexity with
problem dimensionality. Furthermore, the single-obstacle HJI
VI cannot directly handle situations in which vehicles do not
have a common scheduled arrival time. Instead, we perform
sequential path planning by considering vehicles in order
of priority, modeling higher-priority vehicles as time-varying
obstacles for lower-priority vehicles. To do this, we solve a
double-obstacle HJI VI which allows us to obtain the reach-
avoid set, defined as the set of states from which a vehicle
can reach its target while staying within a time-varying state
constraint set. From the solution of the double-obstacle HJI VI,
we can also extract the latest start time and the optimal control
for each vehicle. This is a first application of the double-obstacle
HJI VI which can handle systems with time-varying dynamics,
target sets, and state constraint sets, and results in computation
complexity that scales linearly, as opposed to exponentially, with
the number of vehicles in consideration.
I. INTRODUCTION
Consider a group of autonomous vehicles trying to per-
form a task or reach a goal which may be time-varying in
their joint state space, while avoiding obstacles and other
vehicles. Providing safety and performance guarantees for
such a multi-agent autonomous system (MAAS) is very
relevant practically: Recently,
there has been a growing
interest in using unmanned aerial vehicles (UAVs) for civil
applications, as companies like Amazon and Google are
looking in the near future to send UAVs into the airspace
to deliver packages [1], [2]. Government agencies such as
the Federal Aviation Administration (FAA) and National
Aeronautics and Space Administration (NASA) of the United
States are also expressing growing interest in analyzing these
problems in order to prevent airspace conflicts that could
arise with the introduction of potentially many UAVs in an
urban environment [3]. In addition, UAVs can be used not
only to deliver packages quickly, but in any situation where
fast response is desired. For example, UAVs can provide
This work has been supported in part by NSF under CPS:ActionWebs
(CNS-931843), by ONR under the HUNT (N0014-08-0696) and SMARTS
(N00014-09-1-1051) MURIs and by grant N00014-12-1-0609, by AFOSR
under the CHASE MURI (FA9550-10-1-0567). The research of J.F. Fisac
has received funding from the "la Caixa" Foundation.
All authors are with the Department of Electrical Engineering and
Computer Sciences, University of California, Berkeley. {mochen72, jfisac,
sastry, tomlin}@eecs.berkeley.edu
In general, MAASs are difficult to analyze due to their
inherent high dimensionality. MAASs also often involve
aspects of cooperation and asymmetric goals among the
vehicles or teams of vehicles, making their analysis particu-
larly interesting. MAASs have been explored extensively in
the literature. Some researchers have done work on multi-
vehicle path planning in the presence of other unknown
vehicles or moving entities with assumptions on their specific
control strategies [5]. In a number of formulations for safe
multi-vehicle navigation,
these assumed strategies induce
velocity obstacles that vehicles must avoid to maintain safety
[6], [7]. Researchers have also used potential functions to
perform collision avoidance while maintaining formation
given a predefined trajectory [8], [9]. However, these bodies
of work have not considered trajectory planning and collision
avoidance simultaneously.
One well-known technique for optimal trajectory planning
under disturbances or adversaries is reachability analysis, in
which one computes the reach-avoid set, defined as the set
of states from which the system can reach a target set while
remaining within a state constraint set for all time. For reach-
ability of systems of up to five dimensions, single-obstacle
Hamilton-Jacobi-Isaacs (HJI) variational
inequalities (VI)
[10], [11] have been used in situations where obstacles and
target sets are static. Another HJI VI formulation [12] is able
to handle problems with moving target sets with no obstacles.
A major practical appeal of the above approaches stems
from the availability of modern numerical tools such as [10],
[13], [14], [15], which can efficiently solve HJI equations
when the problem dimension is low. These numerical tools
have been successfully used to solve a variety of differ-
ential games, path planning problems, and optimal control
problems[10], [16], [17]. Despite the power of the previous
HJI formulations, the approaches become numerically in-
tractable very quickly as the number of vehicles in the system
is increased. This is because the numerical computations are
done on a grid in the joint state space of the system, resulting
in an exponential scaling of computation complexity with
respect to the dimensionality of the problem. Furthermore,
state constraint sets, while useful for modeling unsafe vehicle
configurations, are required to be time-invariant
in [10],
[11], [18]. To solve problems involving time-varying state
constraints, [19] proposed to augment the state space with
time; however, this process introduces an extra state space
dimension, resulting in added computation complexity.
Recently, [20] presented a double-obstacle HJI VI which
handles problems in which the dynamics, target sets, and
state constraint sets are all time-varying, and provided a nu-
merical implementation based on well-known schemes. The
formulation does not introduce any additional computation
overhead compared to the above-mentioned techniques, yet
it still maintains the same guarantees on the system's safety
and performance. In this paper, we provide a first application
of the theory presented in [20]. As a point of clarification,
"obstacles" in the context of HJI VIs refer to the effective
constraints in the HJI VI, while obstacles in the state space
represent physical obstacles that vehicles must avoid.
Our contributions are as follows. First, we formulate
a multi-vehicle collision avoidance problem involving N
autonomous vehicles. Each vehicle seeks to get to its own
target sets while avoiding obstacles and collision with all
other vehicles. To reduce the problem complexity to make
the problem tractable, we assign a priority to each vehicle,
and model higher-priority vehicles as time-varying obstacles
that need to be avoided. We then utilize the double-obstacle
HJI VI proposed in [20] to compute reach-avoid sets to plan
trajectories for vehicles in order of priority. This way, we
are able to offer a tractable solution that scales linearly, as
opposed to exponentially, with the number of vehicles. We
demonstrate the scalability of our approach in a four-vehicle
system.
II. PROBLEM FORMULATION
Consider N vehicles Pi, i = 1 . . . , N, each trying to reach
one of N target sets Ti, i = 1 . . . , N, while avoiding obsta-
cles and collision with each other. Each vehicle i has states
xi ∈ Rni and travels on a domain Ω = Ωobs ∪ Ωf ree ∈ Rp,
where Ωobs represents the obstacles that each vehicle must
avoid, and Ωf ree represents all other states in the domain
on which vehicles can move. Each vehicle i = 1, 2, . . . , N
moves with the following dynamics for t ∈ [tEST
, tSTA
]:
i
i
xi = fi(t, xi, ui), xi(tLST
i
) = x0
i
(1)
the joint state space of all vehicles x ∈ Rn where n =(cid:80)
i represents the initial condition of vehicle i, and
where x0
ui(·) represents the control function of vehicle i. In general,
fi(·,·,·) depends on the specific dynamic model of vehicle i,
and need not be of the same form across different vehicles.
Denote pi ∈ Rp the subset of the states that represent the
i ∈ Ωf ree, we define the
position of the vehicle. Given p0
admissible control function set for Pi to be the set of all
control functions such that pi(t) ∈ Ωf ree∀t ≥ tLST
. Denote
i ni,
and their joint control u.
We assume that the control functions ui(·) are drawn from
] → Ui, measurable1} where
the set Ui := {ui : [tEST
Ui ∈ Rnu
is the set of allowed control inputs. Furthermore,
we assume fi(t, xi, ui) is bounded, Lipschitz continuous
in xi for any fixed t, ui, and measurable in t, ui for each
1 A function f : X → Y between two measurable spaces (X, ΣX )
and (Y, ΣY ) is said to be measurable if the preimage of a measurable set
in Y is a measurable set in X, that is: ∀V ∈ ΣY , f−1(V ) ∈ ΣX, with
ΣX , ΣY σ-algebras on X,Y .
, tSTA
i
i
i
i
Fig. 1: An illustration of the problem formulation with three
vehicles. Each vehicle Pi seeks to reach its target set Ti by
time t = tSTA
, while avoiding physical obstacles Ωobs and
the danger zones of other vehicles.
i
i
i
i and any control
xi. Therefore given any initial state x0
function ui(·), there exists a unique, continuous trajectory
xi(·) solving (1) [21].
The goal of each vehicle i is to arrive at Ti ⊂ Rni
at or before some scheduled time of arrival (STA) tSTA
in
minimum time, while avoiding obstacles and danger with all
other vehicles. The target sets Ti can be used to represent
desired kinematic quantities such as position and velocity
and, in the case of non-holonomic systems, quantities such
as heading angle. tEST
can be interpreted as the earliest start
time (EST) of vehicle i, before which the vehicle may not
depart from its initial state. Further, we define tLST
, the latest
(acceptable) start time (LST) for vehicle i. Our problem can
now be thought of as determining the LST tLST
for each
vehicle to get to Ti at or before the STA tSTA
i
, and finding
a control to do this safely. If the LST is before the EST
tLST
i < tEST
, then it is infeasible for vehicle i to arrive at
Ti at or before the STA tSTA
is
feasibility problem that may arise in practice; however, for
i ∀i.
simplicity of presentation, we will assume that tEST
Danger is described by sets Dij(xj) ⊂ Ω. In general, the
definition of Dij depends on the conditions under which ve-
hicles i and j are considered to be in an unsafe configuration,
given the state of vehicle j. Here, we define danger to be
the situation in which the two vehicles come within a certain
radius RC of each other: Dij(xj) = {xi : (cid:107)pi − pj(cid:107)2 ≤
RC}. Such a danger zone is also used by the FAA [22]. An
illustration of the problem setup is shown in Figure 1.
and tEST
i ≤ tLST
. Comparing tLST
i
i
i
i
i
i
In general, the above problem must be analyzed in the joint
state space of all vehicles, making the solution intractable.
In this paper, we will
instead consider the problem of
performing path planning of the vehicles in a sequential
manner. Without loss of generality, we consider the problem
of first fixing i = 1 and determining the optimal control for
vehicle 1, the vehicle with the highest priority. The resulting
optimal control u1 sends vehicle 1 to T1 in minimum time.
Then, we plan the minimum time trajectory for each of
the vehicles 2, . . . , N, in decreasing order of priority, given
ObstacleTargetsVehicleDanger zoneΩ𝜕Ω𝒯1𝒯3𝒯2𝑃1𝑃2𝑃3Ω𝑜𝑏𝑠𝑅𝐶 .
i
i
the previously-determined trajectories for higher-priority ve-
hicles 1, . . . , i − 1. We assume that all vehicles have com-
plete information about the states and trajectories of higher-
priority vehicles, and that all vehicles adhere to their planned
trajectories. Thus, in planning its trajectory, vehicle i treats
higher-priority vehicles as known time-varying obstacles.
and such that xi(τ ) ∈ Ti, τ ≤ tSTA
With the above sequential path planning (SPP) protocol
and assumptions, our problem now reduces to the following
for vehicle i. Given xj(·), j = 1, . . . , i − 1, determine ui(·)
that maximizes tLST
III. SOLUTION VIA DOUBLE-OBSTACLE HJI VI AND SPP
One direct way of solving the problem formulated in Sec-
tion II is by solving a single-obstacle HJI VI [10], [11], [23],
[24]. In this approach, one considers the joint time-invariant
dynamics of the entire system, f (x, u), and defines the static
goal set and the static avoid set in the joint state space of all
vehicles. The goal set encodes the joint states representing all
vehicles being at their target sets, and the avoid set encodes
the joint states representing all unsafe configurations. These
sets are defined as sub-zero level sets of appropriate implicit
surface functions s(x) where x ∈ S ⇔ s(x) ≤ 0. Having
defined the implicit surface functions, the HJI VI (2) is then
solved backwards in time with the implicit surface function
representing the terminal set l(x) as the initial condition and
the implicit surface function representing the avoid set a(x)
as an effective constraint:
max(cid:8)DtV + min [0, H (x, DxV )] ,−a(x) − V (x, t)(cid:9) = 0,
V (x, 0) = l(x)
(2)
with the optimal Hamiltonian H (x, p) = minu∈U p·f (x, u).
The solution V (x, t) is the implicit surface function rep-
resenting the reach-avoid set RA(t), which defines the set
of states from which the system has a control to drive the
state at time t to the goal set L at time 0 while staying out
of the avoid set A at all times. Note that the joint dynamics,
goal set, and avoid set must be time-invariant. Time-varying
dynamics and sets can be treated by augmenting the state
space with time as an auxiliary state [19]; however, this state
augmentation comes at a large computational expense.
The direct solution described above has been successfully
used to solve a number of problems involving up to a pair
of vehicles [10], [16], [17], [25]. However, since numerical
methods for solving a PDE or a VI involve gridding up the
state space, the computation complexity scales exponentially
with the number of dimensions in the joint state. This
makes the single-obstacle HJI VI inapplicable for problems
involving three or more vehicles. Therefore,
instead of
solving a single-obstacle HJI VI in the joint state space in
i ni, we will consider the problem in in Rni and
solve a sequence of double-obstacle HJI VIs introduced in
[20]. By doing so, we take advantage of the fact that time-
varying targets, obstacles, and dynamics can be handled by
the double-obstacle HJI VIs (but not by the single-obstacle
HJI VI without incurring significant computational expense),
making the analysis of the problem tractable. Furthermore,
Rn = R(cid:80)
even if the dimensionality of the problem is sufficiently low
for computing a numerical solution to the single-obstacle HJI
VI, its inability to handle time-varying systems would still
limit us to only consider problems in which the required time
of arrival is common across all vehicles: tSTA
i = tSTA ∀i.
We first describe the framework for computing reach-avoid
sets with arbitrary terrain, domain, moving obstacles, and
moving target sets based on [20]. As with the single-obstacle
HJI VI, sets are defined as sub-zero level sets of implicit
surface functions; however, crucially, these implicit surface
functions can be time-varying in the double-obstacle HJI VI
without increasing computational complexity. Being able to
compute reach-avoid sets with moving obstacles allows us to
overcome the computational intractability described above
by sequentially performing path planning for one vehicle
at a time in order of priority, while treating higher-priority
vehicles as moving obstacles. The target set is defined in the
same way as in the single-obstacle HJI VI; the avoid set is by
convention defined as the complement of the state constraint
set in the double-obstacle HJI VI.
A. Reachability via HJI VI
We first state the result given in [20], and then specialize
the result to the problem formulation given in Section II.
Consider a general nonlinear system describing the state
evolution of two players in a differential game for t ∈ [0, T ].
x(t) = f (t, x, u, d),
x(0) = x
(3)
where x is the joint state, u is the control input for player 1,
and d is the control input for player 2. Their joint dynamics f
is assumed to be bounded, Lipschitz continuous in x for any
fixed u, d and t, and measurable in t, u, d for each x. Given
control functions u(·), d(·), there exists a unique trajectory
x ((τ ), τ ) [21]. Player 1 wishes to minimize, and player 2
φu,d
wishes to maximize the following cost functional:
V(cid:0)t, x, u(·), d(·)(cid:1)
max(cid:8)l(φu,d
= min
τ∈[t,T ]
x(0)(τ ), τ ), max
s∈[t,τ ]
The value of the game is thus given by
x(0)(s), s)(cid:9)
g(φu,d
(4)
V(cid:0)t, x, u(·), δ[u](·)(cid:1)
inf
u(·)
V (x, t) := sup
δ[u](·)
(5)
where player 2 chooses a nonanticipative strategy d(·) =
δ[u](·), under which the control signal d(t) is chosen in
response to player 1's control function up to time t, u(τ ), τ ≤
t [18]. The value of the game characterizes reach-avoid set,
or all the states from which player 1 can reach the target L
encoded by the implicit surface function l(x, t), while staying
within some state constraint set G encoded by the implicit
surface function g(x, t), despite the adversarial actions of
player 2. The value function is the unique viscosity solution
[26] to the following single-obstacle HJI VI [20]:
max
(cid:110)
min(cid:8)DtV + H (x, DxV, t) , l(x, t) − V (x, t)(cid:9)
V (x, T ) = max(cid:8)l(x, T ), g(x, T )(cid:9),
x ∈ Rn
x ∈ Rn
g(x, t) − V (x, t)
t ∈ [0, T ],
(cid:111)
= 0,
(6)
The proof is given in [20] and is based on viscosity
solution theory [27], [28].
Now consider the system with dynamics given by (1).
Given a time-varying target set Ti(t) and obstacle Ai(t) that
vehicle i must avoid, we define implicit surface functions
l(xi, t), g(xi, t) such that xi ∈ Ti(t) ⇔ li(xi, t) ≤ 0, xi /∈
Ai(t) ⇔ gi(x, t) ≤ 0. Now, the problem formulated in
Section II becomes one in which vehicle i chooses a control
function ui(·) to minimize the following cost functional:
(cid:0)t, xi, ui(·)(cid:1)
Vi
= min
τ∈[t,T ]
max(cid:8)li(xi(τ ), τ ), max
gi(xi(s), s)(cid:9) (7)
s∈[t,τ ]
Note here, we have an optimal control problem involving
only one vehicle and no adversary (given gi(xi(s), s)),
unlike in the case of the HJI VI (6). Now, specializing
(6) to our optimal control problem, the value function that
characterizes the reach-avoid set RAi(t) is Vi(xi, t), where
xi ∈ RAi(t) ⇔ Vi(xi, t) ≤ 0. Vi(xi, t) is the viscosity
solution [26] of the HJI VI
max(cid:8) min{DtVi + Hi (xi, DxiVi, t) , li(xi, t) − Vi(xi, t)}
)(cid:9) , xi ∈ Rni
gi(xi, t) − Vi(xi, t)(cid:9) = 0, t ∈ [tEST
) = max(cid:8)li(xi, tSTA
i
), gi(xi, tSTA
], xi ∈ Rni
Vi(xi, tSTA
(8)
where the Hamiltonian Hi(t, xi, p) and optimal control ui
are given by
, tSTA
i
i
i
i
Hi(t, xi, p) = min
ui∈Ui
u∗
i = arg min
ui
p · fi(t, xi, ui)
Hi(t, xi, p)
(9)
B. Sequential Path Planning
In order to use (8) to perform SPP, we first define
the moving obstacles induced by higher-priority vehicles.
Specifically, for vehicle i, we define the moving obstacles
Oi
j(t) induced by vehicles j = 1, . . . , i−1, given their known
trajectories xj(·), to be Oi
Each vehicle i must avoid being in Oi
j(t) for each j =
1, . . . , i − 1 and for all time t, as well as avoid being in
static obstacles Ωobs in the domain. Therefore, for the ith
vehicle, we compute the reach-avoid set with the following
time-varying avoid set Ai(t) and goal set Li(t):
j(t) := {xi : pi ∈ Dij(xj(t)).
Ai(t) := {xi : pi ∈ Ωobs} ∪(cid:16) (cid:91)
(cid:17)
Oi
j(t)
j=1,...,i−1
(10)
Li(t) := Ti, t ≤ tSTA
The goal set is represented by the implicit surface function
li(x, t), where li(xi, t) ≤ 0 ⇔ xi(t) ∈ Li(t). The state
i
constraint set in the HJI VI is defined as the complement
of the avoid set, Ac
i (t), and is represented by the implicit
surface function g(xi, t), where g(xi, t) ≤ 0 ⇔ xi /∈ Ai(t).
For both li(xi, t) and g(xi, t), we use the signed distance
function (in xi) to the sets Li(t) and Ac
Now, we can solve the double-obstacle HJI VI (8). The
solution V (xi, t) represents the reach-avoid set RA(t):
V (xi, t) ≤ 0 ⇔ xi(t) ∈ RA(t). RA(t) is the set of states
at starting time t from which vehicle i can arrive at Ti at or
before time tSTA
i while avoiding obstacles and danger zones
of all higher-priority vehicles j = 1, . . . , i − 1.
i (t), respectively.
i = inf{t : x0
Alternatively, given an initial state x0
i , we can solve (8)
i ∈ RA(t)}. This represents the
to some tLST
latest time that vehicle i must depart from its initial position
in order to reach Ti while avoiding obstacles and all danger
zones of higher-priority vehicles j = 1, . . . , i − 1.
The optimal control is given by
ui(t) = arg min Hi (t, DxiV (xi, t), V (xi, t))
(11)
Observe that since each vehicle i is guaranteed to be safe
with respect to higher priority vehicles j = 1, . . . , i − 1, the
safety of all vehicles, including lower-priority vehicles, can
also be guaranteed.
IV. RESULTS: FOUR VEHICLES WITH CONSTRAINED
TURN RATE
Consider four vehicles with states xi = [xi, yi, θi](cid:62) mod-
eled using a horizontal kinematics model with the following
dynamics for t ∈ [tEST
], i = 1, 2, 3, 4:
, tSTA
i
i
xi = vi cos(θi)
yi = vi sin(θi)
θi = ωi
xi(tEST
i
) = x0
i
ωi ≤ ¯ωi
(12)
where (xi, yi) is the position of vehicle i, θi is the heading
of vehicle i, and vi is the speed of vehicle i. The control
input ui of vehicle i is the turning rate ωi, whose absolute
value is bounded by ¯ωi. For illustration, we chose ¯ωi = 1∀i
and assume vi = 1 is constant; however, our method can
easily handle the case in which ¯ωi differ across vehicles and
vi is a control input. Optimizing the Hamiltonian associated
with vehicle i, Hi(t, DxiVi(xi, t), Vi(xi, t)), we can obtain
the optimal control
ωi(t) = −¯ωi
DθiVi(xi, t)
DθiVi(xi, t)
(13)
The vehicles have initial conditions and STA as follows:
The target sets Ti of the vehicles are all 4 circles of
radius 0.1 in the domain. The centers of the target sets are at
(0.7, 0.2), (−0.7, 0.2), (0.7,−0.7), (−0.7,−0.7) for vehicles
i = 1, 2, 3, 4, respectively. The obstacles are rectangles near
1 = (−0.5, 0, 0),
x0
x0
2 = (0.5, 0, π),
3 = (−0.6, 0.6, 7π/4) ,
x0
x0
4 = (0.6, 0.6, 5π/4) ,
tSTA
1 = 0
tSTA
2 = 0.2
tSTA
3 = 0.4
tSTA
4 = 0.6
(14)
Fig. 2: Initial configuration of the four-vehicle example.
the middle of the domain. The setup for this example is
shown in Figure 2.
i
.
The joint state space of this system is twelve-dimensional,
intractable for analysis using the single-obstacle HJI VI (2).
Therefore, we will repeatedly solve the double-obstacle HJI
VI (8) to compute the reach-avoid sets from targets Ti
for vehicles 1, 2, 3, 4, in that order, with moving obstacles
induced by vehicles j = 1, . . . , i − 1. We will also obtain
tLST
, i = 1, 2, 3, 4, the LSTs for each vehicle in order to
reach Ti by tSTA
i
1 = −1.12, tLST
3 = −1.48, tLST
Figures 3, 4, and 5 show the results. Since the state space
of each vehicle is 3D, the reach-avoid set is also 3D. To
visualize the results, we slice the reach-avoid sets at the
i . Figure 3 shows the 2D reach-avoid
initial heading angles θ0
set slices for each vehicle at its LSTs tLST
2 =
−0.94, tLST
4 = −1.44 determined from our
method. The obstacles in the domain Ωobs and the obstacles
induced by other vehicles inhibit the evolution of the reach-
avoid sets, carving out thin "channels" that separate the
reach-avoid set into different "islands". One can see how
these channels and islands form by examining the time
evolution of the reach-avoid set, shown in Figure 4 for
vehicle 3.
Finally, Figure 5 shows the resulting trajectories of the
four vehicles. The subplot labeled t = −0.55 shows all four
vehicles in close proximity without collision: each vehicle is
outside of the danger zone of all other vehicles. The actual
arrival times of vehicles i = 1, 2, 3, 4 are 0, 0.19, 0.34, 0.31,
respectively. It is interesting to note that for some vehicles,
the actual arrival times are earlier than the STAs tSTA
, i =
1, 2, 3, 4. This is because in order to arrive at the target by
tSTA
, these vehicles must depart early enough to avoid major
i
delays resulting from the induced obstacles of other vehicles;
these delays would have lead to a late arrival if vehicle i
departed after tLST
i
.
i
V. CONCLUSION
We have presented a problem formulation that allows us
to consider the multi-vehicle trajectory planning problem in
a tractable way by planning trajectories for vehicles in order
of priority. In order to do this, we modeled higher-priority
vehicles as time-varying obstacles. We then solved a double-
obstacle HJI VI to obtain the reach-avoid set for each vehicle.
The reach-avoid set characterizes the region from which each
i
Fig. 3: Reach-avoid sets at t = tLST
for vehicles 1, 2, 3, 4,
i . Black arrows indicate direction
sliced at initial headings θ0
the
of obstacle motion. Due to the turn rate constraint,
presence of static obstacles Ωobs and time-varying obstacles
induced by higher-priority vehicles Oi
j(t) carves "channels"
in the reach-avoid set, dividing it up into multiple "islands".
vehicle is guaranteed to arrive at its target within a time
horizon, while avoiding collision with obstacles and higher-
priority vehicles. The solution also gives each vehicle a latest
start time as well as the optimal control which guarantees that
each vehicle safely reaches its target on time.
REFERENCES
[1] Amazon.com, Inc. (2014) Amazon prime air. [Online]. Available:
http://www.amazon.com/b?node=8037720011
[2] J. Stewart. (2014) Google tests drone deliveries in Project Wing trials.
[Online]. Available: http://www.bbc.com/news/technology-28964260
[3] Jointed Planning and Development Office (JPDO), "Unmanned aircraft
systems (UAS) comprehensive plan -- a report on the nation's UAS
path forward," Federal Aviation Administration, Tech. Rep., Sep 2013.
[4] W. M. Debusk, "Unmanned aerial vehicle systems for disaster relief:
Tornado alley," in Infotech@Aerospace Conferences, 2010.
[5] G. C. Chasparis and J. Shamma, "Linear-programming-based multi-
vehicle path planning with adversaries," in Proceedings of American
Control Conference, June 2005.
[6] P. Fiorini and Z. Shillert, "Motion planning in dynamic environments
using velocity obstacles," International Journal of Robotics Research,
vol. 17, pp. 760 -- 772, 1998.
[7] J. van den Berg, M. Lin, and D. Manocha, "Reciprocal velocity obsta-
cles for real-time multi-agent navigation," in Robotics and Automation,
2008. ICRA 2008. IEEE International Conference on, May 2008, pp.
1928 -- 1935.
[8] R. Olfati-Saber and R. M. Murray, "Distributed cooperative control of
multiple vehicle formations using structural potential functions," in in
IFAC World Congress, 2002.
Initial Setup-0.500.5-0.8-0.6-0.4-0.200.20.40.6ObstacleTarget 1Target 2Target 3Target 4Position/Heading 1Danger zone 1Position/Heading 2Danger zone 2Position/Heading 3Danger zone 3Position/Heading 4Danger zone 4Vehicle 1, t=ti1=-1.12-101-1-0.500.51Vehicle 2, t=ti2=-0.94-101-1-0.500.51Vehicle 3, t=ti3=-1.48-101-1-0.500.51Vehicle 4, t=ti4=-1.44-101-1-0.500.51ObstacleTargetsInitial pos. and headingReach-avoid setDanger zones 3 = 7π
its initial heading θ0
Fig. 4: Time evolution of the reach-avoid set for vehicle
3, sliced at
4 . Black arrows
indicate direction of obstacle motion. Initially, the reach-
avoid set grows unobstructed by obstacles, as shown in
the top subplots. Then, in the bottom subplots, the static
obstacles Ωobs and the induced obstacles of vehicles 1 and
2, O3
2, carve out "channels" in the reach-avoid set.
1,O3
[9] Y.-L. Chuang, Y. Huang, M. D'Orsogna, and A. Bertozzi, "Multi-
vehicle flocking: Scalability of cooperative control algorithms using
pairwise potentials," in Robotics and Automation, 2007 IEEE Interna-
tional Conference on, April 2007, pp. 2292 -- 2299.
[10] I. Mitchell, A. Bayen, and C. Tomlin, "A time-dependent Hamilton-
Jacobi formulation of reachable sets for continuous dynamic games,"
IEEE Transactions on Automatic Control, vol. 50, no. 7, July 2005.
[11] O. Bokanowski, N. Forcadel, and H. Zidani, "Reachability and mini-
mal times for state constrained nonlinear problems without any con-
trollability assumption," SIAM Journal on Control and Optimization,
pp. 1 -- 24, 2010.
[12] E. Barron and H. Ishii, "The Bellman equation for minimizing the
maximum cost," Nonlinear Analysis: Theory, Methods & Applications,
1989.
[13] J. A. Sethian, "A fast marching level set method for monotonically
advancing fronts," Proceedings of the National Academy of Sciences,
vol. 93, no. 4, pp. 1591 -- 1595, 1996.
[14] S. Osher and R. Fedkiw, Level Set Methods and Dynamic Implicit
Surfaces. Springer-Verlag, 2002, ISBN: 978-0-387-95482-0.
[15] I. Mitchell, A Toolbox of Level Set Methods, 2009, http://people.cs.
ubc.ca/∼mitchell/ToolboxLS/index.html.
[16] J. Ding, J. Sprinkle, S. S. Sastry, and C. J. Tomlin, "Reachability
calculations for automated aerial refueling," in IEEE Conference on
Decision and Control, Cancun, Mexico, 2008.
[17] H. Huang, J. Ding, W. Zhang, and C. Tomlin, "A differential game
approach to planning in adversarial scenarios: A case study on
capture-the-flag," in Robotics and Automation (ICRA), 2011 IEEE
International Conference on, 2011, pp. 1451 -- 1456.
[18] I. Mitchell, "Application of level set methods to control and reacha-
bility problems in continuous and hybrid systems," Ph.D. dissertation,
Stanford University, 2002.
Fig. 5: The planned trajectories of the four vehicles. In
the left top subplot, only vehicles 3 (green) and 4 (purple)
have started moving, showing tLST
is not common across
the vehicles. Right
top subplot: all vehicles have come
within very close proximity, but none is in the danger zone
another. Left bottom subplot: vehicle 1 (blue) arrives at T1 at
t = 0. Right bottom subplot: all vehicles have reached their
destination, some ahead of the STA tSTA
.
i
i
[19] O. Bokanowski and H. Zidani, "Minimal time problems with moving
targets and obstacles," 18th IFAC World Congress, 2011.
[20] J. F. Fisac, M. Chen, C. J. Tomlin, and S. S. Sastry, "Reach-Avoid
Problems with Time-Varying Dynamics, Targets and Constraints," in
18th International Conference on Hybrid Systems: Computation and
Controls, 2015.
[21] E. A. Coddington and N. Levinson, Theory of ordinary differential
equations. Tata McGraw-Hill Education, 1955.
[22] M. L. C. Mike M. Paglione and H. F. Ryan, "Generic metrics for the
estimation of the prediction accuracy of aircraft to aircraft conflicts by
a strategic conflict probe tool," Air Traffic Control Quarterly, 1999.
[23] K. Margellos and J. Lygeros, "Hamilton-Jacobi Formulation for
Reach-Avoid Differential Games," IEEE Transactions on Automatic
Control, vol. 56, no. 8, Aug 2011.
[24] K. Margellos and J. Lygeros, "Toward 4-D Trajectory Management
in Air Traffic Control: A Study Based on Monte Carlo Simulation
and Reachability Analysis," IEEE Transactions on Control Systems
Technology, vol. 21, no. 5, Sept 2013.
[25] M. Chen, Z. Zhou, and C. Tomlin, "Multiplayer reach-avoid games via
low dimensional solutions and maximum matching," in Proceedings
of the American Control Conference, 2014.
[26] M. G. Crandall, L. C. Evans, and P. L. Lions, "Some properties of
viscosity solutions of hamilton-jacobi equations," Transactions of the
American Mathematical Society, vol. 282, no. 2, p. 487, Apr. 1984.
[27] L. C. Evans and P. E. Souganidis, "Differential games and repre-
sentation formulas for solutions of Hamilton-Jacobi-Isaacs equations,"
Indiana University Mathematics Journal, vol. 33, no. 5, 1984.
[28] E. Barron, "Differential Games with Maximum Cost," Nonlinear
analysis: Theory, methods & applications, pp. 971 -- 989, 1990.
t=0.40-101-1-0.500.51t=-0.27-101-1-0.500.51t=-0.88-101-1-0.500.51t=-1.48-101-1-0.500.51ObstacleTargetsInitial pos. and headingReach-avoid setDanger zonest=-1.15-0.500.5-0.500.5t=-0.55-0.500.5-0.500.5t=0-0.500.5-0.500.5t=0.34-0.500.5-0.500.5ObstacleTargetsPositions, HeadingsTrajectoriesDanger Zones |
1710.04726 | 2 | 1710 | 2017-11-01T18:56:22 | RAWSim-O: A Simulation Framework for Robotic Mobile Fulfillment Systems | [
"cs.MA",
"cs.RO"
] | This paper deals with a new type of warehousing system, Robotic Mobile Fulfillment Systems (RMFS). In such systems, robots are sent to carry storage units, so-called "pods", from the inventory and bring them to human operators working at stations. At the stations, the items are picked according to customers' orders. There exist new decision problems in such systems, for example, the reallocation of pods after their visits at work stations or the selection of pods to fulfill orders. In order to analyze decision strategies for these decision problems and relations between them, we develop a simulation framework called "RAWSim-O" in this paper. Moreover, we show a real-world application of our simulation framework by integrating simple robot prototypes based on vacuum cleaning robots. | cs.MA | cs |
RAWSim-O: A Simulation Framework for
Robotic Mobile Fulfillment Systems
Marius Merschformann∗1, Lin Xie†2, and Hanyi Li3
1Paderborn University, Paderborn, Germany
2Leuphana University of Luneburg, Luneburg, Germany
3Beijing HANNING ZN Tech Co.,Ltd, Beijing, China
November 3, 2017
Abstract
This paper deals with a new type of warehousing system, Robotic
Mobile Fulfillment Systems (RMFS). In such systems, robots are sent to
carry storage units, so-called "pods," from the inventory and bring them
to human operators working at stations. At the stations, the items are
picked according to customers' orders. There exist new decision problems
in such systems, for example, the reallocation of pods after their visits at
work stations or the selection of pods to fulfill orders. In order to analyze
decision strategies for these decision problems and relations between them,
we develop a simulation framework called "RAWSim-O" in this paper.
Moreover, we show a real-world application of our simulation framework
by integrating simple robot prototypes based on vacuum cleaning robots.
1
Introduction
Due to the rise of e-commerce, the traditional manual picker-to-parts warehous-
ing systems no longer work efficiently, and new types of warehousing systems
are required, such as automated parts-to-picker systems. For details about the
classification of different warehousing systems we refer to [4]. This paper stud-
ies one of the parts-to-picker systems, a so-called Robotic Mobile Fulfillment
Systems (RMFS), such as the Kiva System ([5], nowadays Amazon Robotics),
the GreyOrange Butler or the Swisslog CarryPick. The approach of an RMFS
spares some of the time-intense tasks that human workers are confronted with
in traditional warehouses, such as the search and retrieval of order items. [11]
indicates that the Kiva System increases the productivity two to three times,
∗[email protected]
†[email protected]
1
in contrast to a traditional manual picker-to-parts system. Compared to other
kinds of warehousing systems, the biggest advantages of an RMFS are its flexi-
bility as a result of having virtually no fixed installations, the scalability due to
accessing the inventory in a parallel manner, and the reliability due to the use
of only homogeneous components, i.e., redundant components may compensate
for faulty ones (see [6] and [12]).
The first framework, "Alphabet Soup," was published by [6], and is a first
simulation of the RMFS concept. In this work, we extend the work of [6] to
develop our simulation framework, called "RAWSim-O" (Robotic Automatic
Warehouse Simulation (for) Optimization). Similarly to "Alphabet Soup," we
use an agent-based and event-driven simulation focusing on a detailed view of
the system, but extend our simulation framework with the cases of multiple
floors, which are connected by elevators. The reason for this extension is to
avoid the lack of utilization of vertical space in an RMFS compared to other
parts-to-picker systems, such as shuttle-based solutions (see [10]). Moreover,
we integrate a more realistic robot movement emulation by considering the
robot's turning time and adjusted acceleration or deceleration formulas. This
enables a real world application of our simulation framework by using simple
robot prototypes based on vacuum cleaning robots (iRobot Create 2). The
implementation of the framework was done in C# and is compatible with the
Mono project to allow execution on high throughput clusters. All source codes
are available at https://github.com/merschformann/RAWSim-O.
There exist many decision problems in an RMFS, and they influence each
other. With the publication of RAWSim-O we provide a tool for analyzing ef-
fects of decision mechanisms for these problems and synergies of decision strate-
gies. The framework enables a holistic look at RMFS which helps uncovering
side-effects of strategies, e.g. deciding tasks for robot that cause congestion
issues for path planning methods. Hence, RAWSim-O enables a more reliable
assessment of the system's overall efficiency, e.g., in terms of customer order
throughput. Thus, we hope to further support and push the research on RMFS
with our work. We describe the RMFS in more detail and point out the decision
problems in focus in Section 2. Next, our simulation framework is described in
Section 3. After that, we show a demonstration application of our simulation
in Section 4. Finnaly, we conclude our work in Section 5.
2 The Robotic Mobile Fulfillment System
Instead of using a system of shelves and conveyors as in traditional parts-to-
picker warehouses, the central components of an RMFS are:
• movable shelves, called pods, on which inventory is stored
• storage locations denoting the inventory area where the pods are stored
• workstations, where the pick order items are picked from pods (pick sta-
tions) or replenishment orders are stored to pods (replenishment stations)
2
• mobile robots, which can move underneath pods and carry them to work-
stations.
The pods are transported by robots between the inventory area and worksta-
tions. Figure 1 shows the storage and retrieval process: after the arrival of
a replenishment order (consisting of a number of physical units of one stock
keeping unit (SKU)), robots carry selected pods to a replenishment station to
store units in pods. Similarly, after receiving a pick order (including a set of
order lines, each for one SKU, with corresponding units necessary to fulfill the
line), robots carry selected pods to a pick station, where the units for the order
lines are picked. Note that, in order to fulfill pick orders, several pods may be
needed, since orders may have multiple lines. Although pods typically contain
multiple SKUs with many SKUs in stored in the system it is very unlikely that
a pick order can be completed with only one pod. Then, after a pod has been
processed at one or more stations, it is brought back to a storage location in
the inventory area.
Figure 1: The central process of an RMFS (see [7])
2.1 Decision Problems
In an RMFS environment, various optimization and allocation problems have
to be solved in real time. The system aims at keeping human workers at the
stations busy while minimizing the resources (e.g. robots) to fulfill the incoming
pick orders. These problems were first described by [12]:
• For the bots, the planning of their tasks, paths, and motions
• For the stations, the management of the workflow for each station
3
StorageRetrievalStorageReplenishmentOrder pickingRetrieval • For the resources of the system, their planning, provisioning, and allo-
cation to other components of the system, usually called the Resource
Allocation Problem
As [12] note, the Resource Allocation Problem cannot be treated as a global
optimization problem, but rather as a set of subproblems that should be solved
using specialized methods to enable the decisions in an online e-commerce en-
vironment. We reiterate these decision problems briefly in the context of the
simulation framework developed in this work:
• Order Assignment
– Replenishment Order Assignment (ROA): assignment of re-
plenishment orders to replenishment stations
– Pick Order Assignment (POA): assignment of pick orders to pick
stations
• Task Creation
– Pod Selection
∗ Replenishment Pod Selection (RPS): selection of the pod
∗ Pick Pod Selection (PPS): selection of the pods to use for
to store one replenishment order in
picking the pick orders assigned at a pick station
– Pod Storage Assignment (PSA): assignment of an available stor-
age location to a pod that needs to be brought back to the inventory
area
• Task Allocation (TA): assignment of tasks from Task Creation and
additional support tasks like idling to robots
• Path Planning (PP): planning of the paths for the robots to execute
• Station Activation (SA): controlling active times of the stations (e.g.
switching some off based on work shifts or for emulation of "jumper"
pickers that can be assigned to allow for temporarily increased throughput)
• Method Management (MM): exchange of controlling mechanisms in
a running system (e.g. replacing the PSA controller with a different one
during runtime in order to adapt to changing conditions)
The decisions for the aforementioned operational problems influence each other.
Here we sketch two relationships between decision problems that may exploit
synergy effects or sabotage each others success:
• POA and PPS: one objective is to maximize the average number of picks
per handled pod (called pile-on); in other words, the higher the pile-on is,
the fewer pods are needed at pick stations. Therefore, the selection of a
pod for an order in PPS is dependent on the selection of the pick station
4
of other orders. If similar orders are assigned to the same pick station, a
pod can be selected to maximize the number of picks.
• PSA, TA and PP: One objective is to minimize the travel times of robots to
complete all orders, thus, reducing the number of robots needed to achieve
a certain throughput. For this, the selected storage locations for pods
impact the performance of the assignment of tasks and the pathfinding,
because the trip destinations change.
Some of the operational decision problems described above have been ad-
dressed in the context of RMFS in previous publications. First, the sequencing
of pick orders and pick pods for an integration of PPS and POA is studied in
[1]. The authors propose methods for obtaining sequences minimizing the num-
ber of pod visits at one station. Second, the planning of paths in an RMFS is
subject to the studies in [2], [3], and [9]. The state-of-the-art multi-agent path
planning algorithms were implemented in [9] to suit PP in an RMFS. Besides
the operational problems, we would like to note that more literature exists on
more tactical to strategic decision problems, like layout planning, for example,
in [8].
3 RAWSim-O
RAWSim-O is an agent-based discrete-event simulation framework.
It is de-
signed to study the context of an RMFS while considering kinematic behavior
of the mobile robots and evaluating multiple decision problems jointly. Hence,
the main focus of the framework is the assessment of new control strategies for
RMFS and their mutual effects. In the following, we describe the RAWSim-O
simulation framework in more detail. At first, we describe the general structure
of the framework, and then we give detailed information about how the robotic
movement is emulated.
3.1 Simulation Process
Figure 2a shows an overview of our simulation process, which is managed by
the core simulator instance. The tasks of the simulator contain:
• Updating agents, which can resemble either real entities, such as robots
for emulating order
and stations, or virtual entities like managers, e.g.
processes.
• Passing decisions to controllers, which can either decide immediately or
buffer multiple requests and release the decision later.
• Exposing information to a visualizer, which allows optional visual feedback
in 2D or 3D. Figure 2b illustrates a screenshot of our simulation in 3D.
As mentioned before, RMFS usually lack utilization of vertical space when
compared to other parts-to-picker systems, for example shuttle-based solutions
5
(see [10]). Although the flexibility of the system due to the mobile components
is an advantage over the many fixed components of a shuttle-based solution,
in areas where land costs are high the lack of vertical space utilization can
be a deal-breaker for the RMFS concept. To help mitigate this disadvantage
RAWSim-O allows the study of multi-floor systems for applications where the
integration of mezzanine floors is possible (see Fig. 2b for example). The floors
are connected by elevators that can be used by one robot at a time. One eleva-
tor connects at least two waypoints of the underlying waypoint-system, which
is used for guiding the robots. For each pair of connected waypoints a constant
time is specified to capture the travel time of the lift transporting a robot.
Conceptually, the waypoint system itself is a directed graph which robots travel
along while adhering to their kinematic behavior. I.e., turning is only allowed
on-the-spot at waypoints (essentially nodes of the graph) and is executed by
a specified angular speed. Furthermore, straight travel along the edges of the
graph involves acceleration and deceleration, which is discussed in more detail
in Sec. 3.2. To avoid blocked destination waypoints for path planning (many
robots share the same destination when approaching stations or elevators) way-
points may form a queuing zone. After a robot reaches such a zone guided by a
path planning engine a queue manager takes over handling of the robots paths
within the zone. For this, the manager may make use of shortcuts definable
within the queue, but mainly lets the robots move up until they reach the end
waypoint (e.g., a pick station). On leaving the queuing zone again the defined
path planning engine takes over again.
(a) Overview of the simulation process.
(b) Visualization screenshot
Figure 2: RAWSim-O simulation framework
The framework allows easy exchange of controllers by implementing a base
structure for all mentioned decision problems. For this, all controllers are also
agents of the simulation that may expose a next event to jump to, and also get
updated at each event. This enables the mechanisms to react to every change
occurring in the simulated world as well as to steer it. As a shortcut, controllers
may also subscribe to events that fire, if certain situations occur.
Additionally, the controllers are called whenever a new decision needs to be
6
StartEndInitializeSimulatorControllerVisualizerFrameworkDecision imminentDecideRenderOutput resultsFetch informationAgentUpdateOptimize made, e.g. the POA controller is called when a new pick order is submitted
to the system, and the TA controller is called when a robot finished its task.
In Tab. 1 we include a short outline of the calls to the event-driven decision
controllers. The SA and MM controllers are not event-driven, because they
are considered like management mechanisms. However, all controllers may sub-
scribe to simulation-wide events.
Table 1: Overview of the base events causing calls to the controller per decision
problem and the corresponding assignment responsibility
Problem Decision
ROA
repl. order → repl. station
pick order → pick station
POA
RPS
PPS
PSA
TA
PP
repl. order → pod
pod → pick order(s)
pod → storage location
task → robot
robot → robot
Main triggers
New repl. order,
repl. order stored in pod
New pick order,
pick order completed,
repl. order stored in pod
New repl. order,
SKU unit picked
New task is assigned to robot
Pod needs to be brought back to in-
ventory
Robot needs a new task
Robot has a new destination
In addition to acting ad-hoc according to the triggers mentioned in Tab. 1
strategies may plan ahead, e.g., by sequencing tasks for robots instead of greedily
selecting them as soon as a robot is done with its previous one. For this example
it can be done by preparing tasks ahead and assigning them to the chosen robot
as soon as it becomes ready. Furthermore, the framework also allows some
controllers to run optimization algorithms in parallel. This means that the
controllers are allowed to buffer certain decisions (like the assignment of pick
orders), run an optimization procedure in parallel, and submit the decision
later on. In this case the simulation is paced until the optimization algorithm
returns a solution in order to synchronize the wall time of the algorithm with
the simulation time. I.e., the simulation will only continue for the wall time
that already passed for the optimization algorithm converted to simulation time
(while it is still running), but not beyond it.
3.1.1 Core Decision Hierarchy
In the following, we describe the hierarchy of all core decision problems after new
replenishment or pick orders are submitted to the system (see Figure 3). For
this, the SA and MM decision problems are neglected, since they have a more
supportive role and can even be replaced with default mechanisms that keep
the status quo. If a new replenishment order is received, first the controllers of
ROA and RPS are responsible for choosing a replenishment station and a pod.
7
This technically results in an insertion request, i.e. a request for a robot to
bring the selected pod to the given workstation. A number of these requests are
then combined in an insertion task and assigned to a robot by a TA controller.
Similarly, after the POA controller selects a pick order from the backlog and
assigns it to a pick station, an extraction request is generated, i.e. a request
to bring a suitable pod to the chosen station. Up to this point, the physical
units of SKUs for fulfilling the pick order are not yet chosen.
Instead, the
decision is postponed and taken just before PPS combines different requests
into extraction tasks and TA assigns these tasks to robots. This allows the
implemented controllers to exploit more information when choosing a pod for
picking. Hence, in this work we consider PPS as a decision closely interlinked
with TA. Furthermore, the system generates store requests each time a pod is
required to be transported to a storage location, and the PSA controller decides
the storage location for that pod. The idle robots are located at dwelling points,
which are located in the middle of the storage area to avoid blocking prominent
storage locations next to the stations. Another possible type of task is charging,
if the battery of a robot runs low; however, for this work we assume the battery
capacity to be infinite. All of the tasks result in trips, which are planned by a
PP algorithm. The only exception is when a pod can be used for another task
at the same station, thus, not requiring the robot to move.
Figure 3: Order of decisions to be done triggered by receiving a pick or replen-
ishment order
3.1.2
Input Information
The simulation framework conceptually consists of three different inputs. First,
a layout configuration specifies the characteristics and dimensions of the system
layout itself. Second, a scenario configuration describes how orders are gen-
erated and further settings of the system's surroundings. Finally, a controller
configuration is given to specify the decision mechanisms for all previously de-
scribed decision problems. We distinguish them as three different input files to
8
InsertExtractStoreRequests:Trips:RPSROAPOATAPPPSAPPSRepl. orderreceivedPick orderreceived enable easier assessment of control methods for different systems under diverse
scenarios.
Layout Specification The layout specification can be either an explicit file
specifying the exact positions and individual characteristics for all stations, the
waypoint system, and the robots; or a file providing specifications leading to a
default layout based on the concepts of [8]. This default layout can be seen in
Fig. 4, with pick stations as red circles, replenishment stations as yellow circles,
robots in green, and the pod storage locations in the middle as blue squares.
It is based on the idea of circular flows around storage location blocks that
also align with the entrances and exits of the stations and their queuing areas
(dotted lines). Between the queuing area and the storage locations, a hallway
area (dashed lines) allows the robots to cross between the stations. In order to
generate such a layout, the numbers of pick and replenishment stations on the
north, south, east, and west sides need to be set. Furthermore, the numbers of
vertical and horizontal aisles and the block size need to be specified.
Figure 4: The default layout, with two replenishment and two pick stations
Scenario Specification The next input file specifies information about the
scenario to simulate. This includes the duration of the simulation and settings
about the inventory and order emulation. In contrast with the concept of col-
ored word order emulation of "Alphabet Soup" in [6], we use a concept for
SKU and pick order generation based on typical random distributions. For the
generation of SKU popularity information, we implemented constant, uniform,
normal, and gamma distributions to emulate simple scenarios up to more ABC-
like popularity curves. The SKU for each pick order line is then selected using
the chosen distribution, but limiting the choice to only in-stock products. This
is done to avoid stock-out situations, which are not useful when simulating the
control of an RMFS for such situation, i.e. we assume that no unfulfillable pick
9
orders are submitted to the system. The number of order lines and the num-
ber of units per line are also chosen from a distribution previously specified by
the user. For choosing the SKU for each replenishment order, we use the same
popularity distribution connected to information about the order size in which
a certain product is replenished. For this, we also allow setting a parameterized
amount of replenishment orders to be return orders, i.e. single line and single
unit replenishment orders. The space consumption of one unit of a SKU on a
pod is a one-dimensional factor which is also set according to a user-specified
distribution. This is necessary to enable the simulation of both pick and replen-
ishment operations at the same time, while emulating the inventory situation
in our system.
Moreover, we implement three procedures for generating new replenishment
and pick orders. First, a constant order backlog scenario can be selected. This
means that a completed order is immediately replaced by a new one. This is done
to keep the system under constant pressure. To avoid completely overfilling or
draining the inventory over time, we allow the specification of replenishment and
pick order generation pauses according to given inventory level thresholds. E.g.,
if the inventory reaches a 95% filling, replenishment order generation is paused
until it drops below 50% again. Second, we generate orders according to arrivals
of a configurable Poisson process. This Poisson process can be inhomogeneous to
capture order peak situations during a day or match certain patterns observed at
distribution centers. Finally, we allow the input of files to specify which orders
are generated during the simulation horizon. Additionally, we allow a combined
setting of one of the order generation scenarios above with the submission of
order batches at given periodic time points. This is done to allow emulation
of batch operations or hybrid scenarios. With all of these options, we aim to
resemble most artificial and realistic scenarios.
Controller Configuration The last input file specifies the controller to use
for each decision problem as well as the parameters. This enables a flexible
configuration due to the modular controller concept. Controllers integrating
multiple decision problems can be configured using dummy controllers for the
other components. With the help of reflection most parameter structures only
need to be defined once. Without additional implementation effort new param-
eters are serializable and configurable in the graphical user interface.
3.1.3 Output
In order to investigate the system's behavior in more detail, different output
measures are tracked and logged over time. For this, we distinguish three main
measure types. First, a footprint is written for each simulation execution to al-
low for an easy comparison of multiple executions. This footprint contains most
basic performance measures for the simulation run like the order throughput
rate or the distance traveled by the robots. Second, time-based information is
logged to generate plots after execution of the simulation. These are useful for
post-experiment analysis of the simulation processes, especially if no visualiza-
10
(a) Sample progression plot of picked orders (red)
and distance traveled (blue) per hour
(b) Example of a heatmap show-
ing robot movement behavior over
time (red ≡ high, purple ≡ low)
Figure 5: Examples of the system's output
tion was attached. As an example, Figure 5a shows the pick order throughput
and overall distance traveled per hour for a simulation in which the PSA strat-
egy is switched after half of the simulation horizon. For this experiment, a
random PSA strategy was replaced with a strategy selecting the nearest avail-
able storage location for the pod. This is reflected by the decrease in distance to
cover by the robots, which immediately leads to an increased throughput in pick
orders, because more pods are available at the pick stations. Finally, location-
based information is logged to conduct heatmap analyses. This is useful, for
example, to get insights about congestion effects when looking at robot move-
ment behavior over time (see Fig. 5b). In this example, we can see the queuing
happening at the pick stations at the top and the replenishment stations at the
bottom. Furthermore, we can identify highly frequented areas which are prone
to congestion. Note that the logarithm was applied to the heat values in order
to increase the contrast in color in the resulting heatmap.
3.2 Robot Movement Emulation
The calculations of traveling times and distances for straight movement are
based on uniform acceleration and deceleration. For this, the velocity of a robot
has to be considered to determine its arrival time at a destination node. The
symbols used in the following description are defined in Tab. 2. The fundamental
formulas for all remaining definitions are given by the speed (see Eq. 1), the
11
520 540 560 580 600 620 640 660 680 0 5 10 15 20 25 30 35 40 45 24000 25000 26000 27000 28000 29000 30000 31000 32000Count (per hour)Distance (meter)Simulation time (in hours)Distance traveled (y2)Orders completed (y1) Table 2: Symbol definitions
Symbol Description
vt
st
−→a
←−a
The speed at time t
The position at time t
Acceleration in m
s2
Deceleration in m
s2 (negative)
Top-speed in m
s
v
difference in position while accelerating (see Eq. 2), and the difference in position
during top-speed (see Eq. 3).
(1)
(2)
(3)
vt = −→a t + v0
−→a
2
st =
t2 + v0t + s0
st = vt + s0
For the implementation of the simulation framework, it is required to not only
calculate the time for covering a distance when initially standing still, but also
to calculate distances and times based on an initial speed v0 > 0, because
simulation events may occur at any time during robot travel. For this, the time
(tv0→v) and distance(cid:0)stv0→v
This can analogously be defined for full deceleration (see Eq. 6 & 7).
(cid:1) to reach the top-speed are needed (see Eq. 4 & 5).
v = −→a tv0→v + v0 ⇔ tv0→v =
(tv0→v)2 + v0tv0→v
0 = ←−a tv0→0 + v0 ⇔ tv0→0 =
(tv0→0)2 + v0tv0→0
v − v0−→a
0 − v0←−a
−→a
(4)
(6)
(5)
(7)
2
←−a
2
stv0→v =
stv0→0 =
For the calculation of time for traveling distance d when starting at an initial
speed of v0, four cases need to be considered. In the first case, only decelera-
tion is possible. In the second case, cruising at top-speed and deceleration are
possible. In the third case, acceleration up to top-speed, cruising at top-speed
and deceleration are possible. In the fourth case, the distance is so short that
only acceleration and deceleration phases are possible. The function defined in
Alg. 1 calculates the remaining cruise time for all of the cases. Line 7 uses the
time at which acceleration switches to deceleration, which is described in more
detail below.
Let d(cid:48) be the full distance from the start node to the destination node, thus,
starting with zero speed at the beginning of d(cid:48) and stopping with zero speed at
the end of d(cid:48). For this, the time of switching from acceleration to deceleration
is given by Eq. 8, i.e., the time for driving while accelerating. To calculate this
12
Algorithm 1: CruiseT ime(−→a ,←−a , v, v0, d)
1 if d ≤ stv0→0 then
return tv0→0
2
3 if v0 = v then
d−stv0→0
4
return
+ tv0→0
5 if stv0→0 + stv0→0 ≤ d then
v
6
return tv0→v +
(cid:118)(cid:117)(cid:117)(cid:116) d +
−→a
2
−→a
2 +
d−stv0→v−stv0→0
−→a
(cid:1)2
(cid:118)(cid:117)(cid:117)(cid:116) d +
+
v
2
←−a
2 +
(cid:0) v0−→a
−→a
2←−a
(cid:1)2
(cid:0) v0−→a
←−a
2−→a
+ tv0→0
− v0−→a
7 return
time we make use of the fact that the speed at which we switch from acceleration
to deceleration must match (−→a t1 = ←−a t2), hence, we can substitute t2 with
−→a t1←−a .
t2
1 +
t2
1 +
(cid:19)2
2
←−a
←−a
←−a
2
2
t2
2
(cid:18)−→a t1←−a
(−→a t1)
2←−a
2
2
−→a
−→a
−→a
(cid:115)
d(cid:48)
=
⇔ d(cid:48)
=
⇔ d(cid:48)
⇔ t2
=
1 =
⇔ t1 =
t2
1 +
2
d(cid:48)
−→a
2 +
−→a 2
2←−a
d(cid:48)
−→a
2 +
−→a 2
2←−a
2
(8)
For the calculation, we assume that speed is currently zero and the movement
just starts at the start node. If v0 > 0 and d < d(cid:48), then d(cid:48) has to be calculated
by using d (see Eq. 9).
(cid:17)2
(cid:16) v0−→a
−→a
2
d(cid:48)
= d +
(9)
Analogously to Eq. 8, it is also possible to solve for t2, i.e., the time for driving
while decelerating. The sum of t1 and t2 is the complete time for the cruise
from start node to destination node. The time for accelerating to v0 is being
subtracted, such that the remaining time can be expressed as in line 7 of Alg. 1.
4 Demonstrator
We implemented a demonstrator functionality to investigate how our algorithms
developed within RAWSim-O work with real robots. For example with the help
13
of the iRobot Create 2, a mobile robot platform based on the Roomba vac-
uum cleaning robot. There are several reasons we choose the iRobot Create
2: ease of programming, low complexity, similar movement behavior like typ-
ical RMFS robots and low costs. The robots are equipped with ASUS Eee
PCs through serial-to-USB cables for processing capabilities, webcams for line-
following, blink(1) for visual feedback, and RFID tag readers mounted inside the
former vacuum cleaning compartment for waypoint recognition (see Figure 6a).
Although we cannot emulate the transportation of pods with the robots, we are
still able to study the overall movement behavior in a real situation, which is
more prone to errors and noise than a simulation. Technically, the demonstra-
tor robots replace the simulated robots, hence, a hybrid of a real system and a
simulation is built.
(a) Close-up of a demonstrator robot
(b) The view from RAWSim-O
(c) Multiple robots in an emulated RMFS
Figure 6: Demonstrator example of four robots running in an emulated RMFS
We demonstrate our RMFS with a simple running example of four robots
in a grid-world with 0.45m × 0.45m cells (totally 36 cells), see Figure 6b for
the view from RAWSim-O for this example and Figure 6c for the view of the
demonstration. One replenishment station is set on the left bottom side and one
14
pick station is set on the right bottom side. In total, there are six pods (blue
rectangles) available. The maximum velocity limit of each robot is 0.21m/s,
while the time it takes for each robot to do a complete turn is set to 5.5s. And the
maximum acceleration and deceleration of each robot are set to 0.5 and −0.5. To
adhere to the kinematic constraints (such as turning times and acceleration) in
continious time, we use the MAPFWR-solver from [9] to generate time-efficient
collision- and deadlock-free paths. Those paths are converted to a sequence of
go straight, turn left, and turn right commands and sent to the robot via WiFi.
The robot then executes these commands and sends back the RFID waypoint
tag of each intersection it comes across. By doing this, the movement of the
real robot and the expected movement are synchronized. This needs to be done
in order to avoid errors from the noisy real-world setting to add up. Similarly
to the simulated environment, pick and replenishment operations are emulated
by blocking the robot at the station for a fixed time. With the published source
code and similar components the demonstrator can be rebuild.
(a) Filtered image showing the line
(b) Processing result on original image
Figure 7: Line-following example using a blue line
The robot line-following is implemented using a common PID (Proportional
Integration Differential) controller method. The center of the line is extracted
by using the OpenCV library. The line recognition first translates the image to
a binary one by applying a range filter, and then applies the basic morpholog-
ical operation "erosion" for removing noise (see resulting image in Figure 7a).
Afterwards, contour extraction is used to find the line's borders within the gray
image. Using these lines, the bottom center is estimated to allow a stable target
(see green ring in Figure 7b) for the PID controller.
Using the demonstrator, we are able to show the successful application of the
path planning algorithms developed within RAWSim-O in a real-world situa-
tion. Furthermore, it can be used to demonstrate the basic idea of an RMFS at
a small scale.
15
5 Conclusion
In this work, we outlined core real-time decision problems occurring when oper-
ating an RMFS. To investigate different solution approaches for those decision
problems, we introduce the RMFS simulation framework RAWSim-O and de-
scribe its functionality and capabilities. Alongside this publication, we also
publish the source code of the framework and already present controllers and
algorithms at https://github.com/merschformann/RAWSim-O to support fu-
ture research on RMFS. With this work we aim to support future research
on RMFS. For example, a study of how to control systems involving multiple
mezzanine floors can be done using RAWSim-O. For these, an efficient storage
strategy and task allocation method is expected to be crucial in order to mit-
igate the bottle-neck effect introduced by the elevators. Furthermore, we have
shown demonstrator application capabilities of the framework that enable the
integration of simple robots for real world demonstrations.
6 Acknowledgements
We would like to thank Tim Lamballais for providing us with the concepts and
implementation of the default layout generator used in RAWSim-O.
References
[1] Nils Boysen, Dirk Briskorn, and Simon Emde. Parts-to-picker based order
processing in a rack-moving mobile robots environment. European Journal
of Operational Research, 262(2):550–562, 2017.
[2] Liron Cohen, Tansel Uras, and Sven Koenig. Feasibility study: Using high-
ways for bounded-suboptimal multi-agent path finding. In Eighth Annual
Symposium on Combinatorial Search, 2015.
[3] Liron Cohen, Glenn Wagner, T. K. Satish Kumar, Howie Choset, and Sven
Koenig. Rapid randomized restarts for multi-agent path finding solvers,
2017.
[4] Ren´e de Koster, Tho Le-Duc, and Kees Jan Roodbergen. Design and con-
trol of warehouse order picking: A literature review. European Journal of
Operational Research, 182(2):481–501, 2007.
[5] John Enright and Peter R. Wurman. Optimization and coordinated au-
tonomy in mobile fulfillment systems. In Sanem Sariel-Talay, Stephen F.
Smith, and Nilufer Onder, editors, Automated Action Planning for Au-
tonomous Mobile Robots, 2011.
[6] Christopher J. Hazard, Peter R. Wurman, and Raffaello D'Andrea. Al-
phabet soup: A testbed for studying resource allocation in multi-vehicle
16
systems. In Proceedings of AAAI Workshop on Auction Mechanisms for
Robot Coordination, pages 23–30. Citeseer, 2006.
[7] A. E. Hoffman, M. C. Mountz, M. T. Barbehenn, J. R. Allard, M. E. Kim-
mel, F. Santini, M. H. Decker, R. D'Andrea, and P. R. Wurman. System
and method for inventory management using mobile drive units, 2013.
[8] T. Lamballais, D. Roy, and M.B.M. de Koster. Estimating performance
in a robotic mobile fulfillment system. European Journal of Operational
Research, 2016.
[9] M. Merschformann, L. Xie, and D. Erdmann. Path planning for Robotic
Mobile Fulfillment Systems. ArXiv e-prints, June 2017.
[10] Elena Tappia, Debjit Roy, Ren´e de Koster, and Marco Melacini. Modeling,
analysis, and design insights for shuttle-based compact storage systems.
Transportation Science, 51(1):269–295, 2017.
[11] Marc Wulfraat. Is kiva systems a good fit for your distribution center? an
unbiased distribution consultant evaluation., 2012.
[12] Peter R. Wurman, Raffaello D'Andrea, and Mick Mountz. Coordinating
hundreds of cooperative, autonomous vehicles in warehouses. AI Magazine,
29(1):9, 2008.
17
|
1801.04105 | 1 | 1801 | 2018-01-12T09:42:36 | Active repositioning of storage units in Robotic Mobile Fulfillment Systems | [
"cs.MA"
] | In our work we focus on Robotic Mobile Fulfillment Systems in e-commerce distribution centers. These systems were designed to increase pick rates by employing mobile robots bringing movable storage units (so-called pods) to pick and replenishment stations as needed, and back to the storage area afterwards. One advantage of this approach is that repositioning of inventory can be done continuously, even during pick and replenishment operations. This is primarily accomplished by bringing a pod to a storage location different than the one it was fetched from, a process we call passive pod repositioning. Additionally, this can be done by explicitly bringing a pod from one storage location to another, a process we call active pod repositioning. In this work we introduce first mechanisms for the latter technique and conduct a simulation-based experiment to give first insights of their effect. | cs.MA | cs |
Active repositioning of storage units in
Robotic Mobile Fulfillment Systems
Marius Merschformann∗1
1Paderborn University, Paderborn, Germany
January 15, 2018
Abstract
In our work we focus on Robotic Mobile Fulfillment Systems in e-commerce dis-
tribution centers. These systems were designed to increase pick rates by employing
mobile robots bringing movable storage units (so-called pods) to pick and replenish-
ment stations as needed, and back to the storage area afterwards. One advantage of
this approach is that repositioning of inventory can be done continuously, even during
pick and replenishment operations. This is primarily accomplished by bringing a pod
to a storage location different than the one it was fetched from, a process we call passive
pod repositioning. Additionally, this can be done by explicitly bringing a pod from one
storage location to another, a process we call active pod repositioning. In this work
we introduce first mechanisms for the latter technique and conduct a simulation-based
experiment to give first insights of their effect.
1
Introduction
In today's increasingly fast-paced e-commerce an efficient distribution center is one crucial
element of the supply chain. Hence, new automated parts-to-picker systems have been
introduced to increase throughput. One of them is the Robotic Mobile Fulfillment System
(RMFS). In a RMFS mobile robots are used to bring rack-like storage units (so-called pods)
to pick stations as required, thus, eliminating the need for the pickers to walk and search
the inventory. A task which can take up to 70 % of their time in traditional picker-to-
parts systems (see [6]). This concept was first introduced by [7] and an earlier simulation
work by [2]. The first company implementing the concept at large scale was Kiva Systems,
nowadays known as Amazon Robotics.
One of the features of RMFS is the continuous resorting of inventory, i.e. every time a
pod is brought back to the storage area a different storage location may be used. While
this potentially increases flexibility and adaptability, rarely used pods may block prominent
storage locations, unless they are moved explicitly. This raises the question whether active
repositioning of pods, i.e. picking up a pod and moving it to a different storage location,
can be usefully applied to further increase the overall throughput of the system. In order
to address this issue we focus on two approaches for active repositioning. First, we look at
∗[email protected]
1
Figure 1: Active repositioning move (green arrow) vs. passive repositioning (red arrow)
repositioning done in parallel while the system is constantly active and, second, we look at
repositioning during system downtime (e.g. nightly down periods). While the assignment
of storage locations to inventory is a well studied problem in warehousing (see [1]) the
repositioning of inventory is typically not considered for other systems, because it is usually
very expensive.
2 Repositioning in RMFS
In a RMFS passive repositioning of pods is a natural process, if the storage location chosen
for a pod is not fixed. For example, in many situations using the next available storage
location is superior to a fixed strategy, because it decreases the travel time of the robots and
by this enables an earlier availability for their next tasks. However, this strategy might cause
no longer useful pods to be stored at very prominent storage locations, which introduces
the blocking problem discussed earlier. There are two opportunities to resolve this issue:
on the one hand it is possible to already consider characteristics of the pod content while
choosing an appropriate storage location, while on the other hand it is possible to actively
move pods from inappropriate storage locations to better fitting ones. We call the latter
approach active repositioning of pods. Both repositioning approaches are shown in Fig. 1.
Additionally, the figure shows an excerpt of the basic layout used for the experiments, i.e.
the replenishment stations (yellow circles), pick stations (red circles), pods (blue squares),
the storage locations (blocks of 2 by 4) and the directed waypoint graph used for path
planning. This layout is based on the work by [3].
In order to assess the value of storage locations and pods we introduce the following met-
rics. First the prominence F SL of a storage location w ∈ W SL is determined by measuring
the minimum shortest path time to a pick station m ∈ MO (see Eq. 1). The shortest path
t is computed with a modified A∗ algorithm that considers turning times to achieve
time f A∗
more accurate results. The storage location with the lowest F SL (w) is considered the most
prominent one, since it offers the shortest time for bringing the pod to the next pick station.
In order to assess the value of a pod b at time t we introduce the pod-speed (F P S) and
pod-utility (F P U ) measures. The speed of a pod (see Eq. 2) is calculated by summing up
(across all SKUs) the units of an SKU contained (f C) multiplied with the frequency of it
(f F ). This frequency is a relative value reflecting the number of times a SKU is part of a
customer order compared to all other SKUs. By using the minimum of units of an SKU
contained and the demand for it (f D), the utility of a pod (see Eq. 3) sums the number of
potential picks when considering the customer order backlog. Thus, it is a more dynamic
value. Both scores are then combined in the metric F P C (see Eq. 4). For our experiments
2
Algorithm 1: CalculateWellsortednessCombined
1 L ← Sort( W SL, i ⇒ F SL (i)) , r(cid:48) ← 1, f(cid:48) ← min
i∈W SL
2 foreach i ∈ {0, . . . , Size(L)−1} do
if F SL (i) > f(cid:48) then r(cid:48) ← r(cid:48) + 1, f(cid:48) ← F SL (i)
ri ← r(cid:48)
3
4
5 c ← 0, d ← 0
6 foreach i1 ∈ {0, . . . , Size(L)−1} do
F SL (i)
7
8
9
10
foreach i2 ∈ {i1 + 1, . . . , Size(L)−1} do
if IsPodStored(L[i1]) ∧ IsPodStored(L[i2]) ∧ rL[i1] (cid:54)= rL[i2] then
b1 ← GetPod(L[i1]), b2 ← GetPod(L[i2])
if F P C (b1, t) < F P C (b2, t) then c ← c + 1, d ← d +(cid:0)rL[i2] − rL[i1]
11 return a ← d
c
we consider the weights wS = wU = 1 to value both characteristics equally.
f A∗
t (w, m)
(cid:0)f C (b, d, t) · f F (d)(cid:1)
(cid:0)min(cid:0)f C (b, d, t) , f D (d, t)(cid:1)(cid:1)
(cid:88)
(cid:88)
d∈D
d∈D
F SL (w) := min
m∈MO
F P S (b, t) :=
F P U (b, t) :=
F P C (b, t) :=
F P S (b, t)
b(cid:48)∈B F P S (b(cid:48), t)
max
· wS +
F P U (b, t)
b(cid:48)∈B F P U (b(cid:48), t)
max
· wU
(cid:1)
(1)
(2)
(3)
(4)
For evaluation purposes we can use these measures to determine an overall "well-sortedness"
score for the inventory. The procedure for calculating the well-sortedness score is described
in Alg. 1. At first we sort all storage locations by their prominence score in ascending
order (see line 1 f.). Next, ranks ri are assigned to all storage locations i ∈ W SL, i.e.,
the best ones are assigned to the first rank and the rank is increased by one each time the
prominence value increases (see line 2 f.). Then, we assess all storage location two-tuples
and count misplacements, i.e., both storage locations are not of the same rank and the
score of the better placed pod at i1 is lower than the worse placed pod at i2 (see line 5
f.). In addition to the number of misplacements we track the rank offset. From this we
can calculate the average rank offset of all misplacements, i.e., the well-sortedness. Hence, a
lower well-sortedness value means a better sorted inventory according to the given combined
pod-speed and pod-utility measures.
In this work we investigate the following repositioning mechanisms:
Nearest (N) For passive repositioning this mechanism always uses the nearest available
t ). This mechanism does not
storage location in terms of estimated path time (f A∗
allow active repositioning.
Cache (C) This mechanism uses the nearest 25 % of storage locations in terms of estimated
path time (f A∗
t ) as a cache. During passive repositioning pods with combined score
(F P C) above a determined threshold are stored at a cache storage location and others
are stored at one of the remaining storage locations. In its active variant it swaps pods
from and to the cache.
3
Utility (U) This mechanism matches the pods with the ranks of the storage locations (see
Alg. 1) on the basis of their combined score (F P C). A nearby storage location with a
close by rank is selected during passive repositioning. In its active variant pods with
the largest difference between their desired and their actual storage location are moved
to an improved one.
3 Computational results
For capturing and studying the behavior of RMFS we use an event-driven agent-based
simulation that considers acceleration / deceleration and turning times of the robots (see [5]).
Since diverse decision problems need to be considered in an RMFS we focus the scope of
the work by fixing all remaining mandatory ones to simple assignment policies and the FAR
path planning algorithm described in [4]. A more detailed overview of the core decision
problems of our scope are given in [5]. For all experiments we consider a simulation horizon
of one week, do 5 repetitions to reduce noise and new customer and replenishment orders are
generated in a random stream with a Gamma distribution (k = 1, Θ = 2) used for the choice
of SKU per order line from 1000 possible SKUs. Furthermore, we analyze repositioning for
four layouts. The specific characteristics are set as follows:
Small Wide
4/4
8/8
8x10
673
Layout
Stations (pick / replenish)
Aisles (hor. x vert.)
Pods
16x10
1271
Long Large
4/4
8/8
8x22
1407
16x22
2658
For the evaluation of active repositioning effectiveness we consider two scenarios. At first,
we look at a situation where the system faces a nightly down period (22:00 - 6:00) during
which no worker is available for picking or replenishment, but robots can be used for active
repositioning. In order to keep the replenishment processes from obscuring the contribution
of nightly inventory sorting, replenishment orders are submitted to the system at 16:00 in
the afternoon in an amount that is sufficient to bring the storage utilization back to 75 %
fill level. For pick operations we keep a constant backlog of 2000 customer orders to keep
the system under pressure. Additionally, we generate 1500 orders per station at 22:00 in
the evening to increase information for the pod utility metric about the demands for the
following day. Secondly, we look at active repositioning done in parallel in a system that is
continuously in action. For this, we consider three subordinate configurations distributing
the robots per station as following:
R1P3A0: 1
4 replenishment, 3
4 replenishment, 2
5 replenishment, 3
4 picking, no active repos.
4 picking, 1
4 active repos.
5 picking, 1
5 active repos. (+1 robot per station)
R1P2A1: 1
R1P3A1: 1
This scenario is kept under continuous pressure by keeping a backlog of constant size for
both: replenishment (200) and customer orders (2000).
The main performance metric for the evaluation is given by the unit throughput rate score
(UTRS). Since we use a constant time of T P = 10s for picking one unit an upper bound
for the number of units that can possibly be handled by the system during active hours can
T P with the set of all pick stations MO. Using this we can
be calculated by U B :=(cid:12)(cid:12)MO(cid:12)(cid:12) 3600
determine the fractional score by dividing the actual picked units per hour in average by
this upper bound.
4
Table 1: Unit throughput rate score of the different scenarios, layouts and mechanisms
(values in percent (%), read columns as [passive mechanism-active mechanism])
Small
Wide
Layout
Setup/Mech. C-C U-U N-C N-U C-C U-U N-C N-U C-C U-U N-C N-U C-C U-U N-C N-U
Deactivated 52.9 53.6 55.0 55.0 50.6 50.5 52.7 52.7 43.3 46.0 47.8 47.8 42.6 44.8 46.3 46.3
Activated
52.9 53.5 55.5 55.9 50.5 50.4 53.1 53.5 42.8 46.3 49.0 49.7 42.2 45.1 47.6 48.2
52.6 51.7 55.6 55.6 50.3 48.9 53.3 53.3 43.3 43.4 48.2 48.2 42.4 42.4 46.6 46.6
R1P3A0
40.9 40.7 44.6 43.8 39.7 38.1 43.3 42.3 31.7 33.8 39.1 38.1 31.5 32.8 38.8 37.8
R1P2A1
R1P3A1
51.9 51.5 55.3 54.5 50.3 48.5 53.7 53.0 41.6 44.0 49.3 48.1 41.2 42.6 48.9 47.6
Large
Long
Figure 2: Time-wise comparison of layout Long and mechanisms N-U with (colored lines)
and without (gray lines) active repositioning at night
The results of the experiment are summarized in Tab. 1. For the comparison of resorting
the inventory during the nightly down period (line: Activated) vs. no active repositioning at
all (line: Deactivated) we can observe an advantage in throughput. However, for the parallel
active repositioning it is not possible to observe a positive effect. When moving one robot
per pick station from pick operations to active repositioning (lines: R1P3A0 and R1P2A1)
we observe a loss in UTRS, because less robots bring inventory to the pick stations. Even
with an additional robot per pick station (line: R1P3A1) we cannot observe a substantial
positive effect. For most cases, the effect is rather negative as a result of the increased
congestion potential for robots moving within the storage area.
In the following we take a closer look at the nightly down period scenario. If we keep
the system sorted with the passive repositioning mechanism (C-C and U-U), nightly active
repositioning does not have a noticeable positive effect, because the passive repositioning
mechanism already keeps the inventory sorted for the most part. However, the Nearest mech-
anism which has a better overall performance, can benefit from a nightly active repositioning
(N-C and N-U). Especially for the Large and Long layouts we can observe a reasonable boost
in UTRS. The greater merit for layouts with more vertical aisles suggest that shorter trip
times of the robots are the reason. This can also be observed when looking at the detailed
results of a run with and without active repositioning (see Fig. 2). First, more orders can
be completed per hour after the inventory was sorted over night (first graph). This boost is
eliminated as soon as replenishment operations begin. Thus, when and how replenishment is
done is crucial to the benefit of resorting during down times, because the effect may be lost
5
300 350 400 450 500 550 600CountOrders completed per hour (no repos.)Orders completed per hour (with repos.) 0 50 100 150 200Rank offsetAvg. rank diff. (no repos.)Avg. rank diff. (with repos.) 35 40 45 50 55Day 1Day 2Day 3Day 4Day 5Day 6Day 7EndTrip time (s)Trip time (no repos.)Trip time (with repos.) Figure 3: Comparison of inventory situation before and after nightly active repositioning
(Long layout, N-U mechanisms, top: day 4 22:00, bottom: day 5 06:00)
quite quickly. In the third graph the shorter times for completing trips to the pick stations
after sorting the inventory support the assumption that these are the the main reason for
the boost. Lastly, the second graph provides the well-sortedness measure and shows that
sorting the inventory can be done reasonably fast. The situation before and after sorting
is shown in Fig. 3. In this heatmap the combined score (F P C (b, t)) is visualized with one
colored tile of the size of a storage location. Here the most useful and just replenished pods
can be seen positioned next to the replenishment stations (left side) before sorting. After
sorting most useful pods are positioned on the far right side of the horizontal aisles inbound
to a pick station. Thus, these pods offering a high potential hit-rate (number of picks from
a pod) can be fetched most quickly.
4 Conclusion
The results suggest that active repositioning may boost throughput performance of RMFS.
If the system faces regular down periods, costs for repositioning (energy costs, robot wear)
are reasonable and charging times allow it, active repositioning can make a reasonable contri-
bution to a system's overall performance. Since the introduced mechanisms greedily search
for repositioning moves, more moves are conducted than necessary to obtain a desired inven-
tory well-sortedness. For future research we suggest to predetermine moves before starting
repositioning operations, e.g. by using a MIP formulation matching pods with storage lo-
cations and selecting the best moves. The source-code of this publication is available at
https://github.com/merschformann/RAWSim-O.
6
References
[1] J. Gu, M. Goetschalckx, and L. McGinnis. Research on warehouse operation: A com-
prehensive review. Eur J Oper Res, 177(1):1–21, 2007.
[2] C. Hazard, P. Wurman, and R. D'Andrea. Alphabet soup: A testbed for studying
resource allocation in multi-vehicle systems. In AAAI Workshop on Auction Mechanisms
for Robot Coordination, pages 23–30, 2006.
[3] T. Lamballais, D. Roy, and M.B.M. de Koster. Estimating performance in a robotic
mobile fulfillment system. Eur J Oper Res, 256(3):976–990, 2016.
[4] M. Merschformann, L. Xie, and D. Erdmann. Path planning for robotic mobile fulllment
systems. https://arxiv.org/abs/1706.09347, 2017.
[5] M. Merschformann, L. Xie, and H. Li. Rawsim-o: A simulation framework for robotic mo-
bile fulfillment systems: Working paper. https://arxiv.org/abs/1710.04726, 2017.
[6] J. Tompkins. Facilities planning. John Wiley & Sons, 4th ed. edition, 2010.
[7] P. Wurman, R. D'Andrea, and M. Mountz. Coordinating hundreds of cooperative, au-
tonomous vehicles in warehouses. AI Magazine, 29(1):9, 2008.
7
|
1908.03309 | 1 | 1908 | 2019-08-09T04:33:54 | Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model | [
"cs.MA",
"cs.CY",
"cs.LG"
] | While simulations have been utilized in diverse domains, such as urban growth modeling, market dynamics modeling, etc; some of these applications may require validations based upon some real-world observations modeled in the simulation, as well. This validation has been categorized into either qualitative face-validation or quantitative empirical validation, but as the importance and the accumulation of data grows, the importance of the quantitative validation has been highlighted in the recent studies, i.e. digital twin. The key component of quantitative validation is finding a calibrated set of parameters to regenerate the real-world observations with simulation models. While this parameter calibration has been fixed throughout a simulation execution, this paper expands the static parameter calibration in two dimensions: dynamic calibration and heterogeneous calibration. First, dynamic calibration changes the parameter values over the simulation period by reflecting the simulation output trend. Second, heterogeneous calibration changes the parameter values per simulated entity clusters by considering the similarities of entity states. We experimented the suggested calibrations on one hypothetical case and another real-world case. As a hypothetical scenario, we use the Wealth Distribution Model to illustrate how our calibration works. As a real-world scenario, we selected Real Estate Market Model because of three reasons. First, the models have heterogeneous entities as being agent-based models; second, they are economic models with real-world trends over time; and third, they are applicable to the real-world scenarios where we can gather validation data. | cs.MA | cs | Under review as a paper at Journal of Autonomous Agents and Multi-Agent Systems
Automatic Calibration of Dynamic and Heterogeneous Parameters in
Agent-based Model
Dongjun Kim · Tae-Sub Yun · Il-Chul Moon
9
1
0
2
g
u
A
9
]
A
M
.
s
c
[
1
v
9
0
3
3
0
.
8
0
9
1
:
v
i
X
r
a
Abstract While simulations have been utilized in diverse domains, such as urban growth modeling, mar-
ket dynamics modeling, etc; some of these applications may require validations based upon some real-world
observations modeled in the simulation, as well. This validation has been categorized into either qualitative
face-validation or quantitative empirical validation, but as the importance and the accumulation of data grows,
the importance of the quantitative validation has been highlighted in the recent studies, i.e. digital twin. The
key component of quantitative validation is finding a calibrated set of parameters to regenerate the real-world
observations with simulation models. While this parameter calibration has been fixed throughout a simulation
execution, this paper expands the static parameter calibration in two dimensions: dynamic calibration and
heterogeneous calibration. First, dynamic calibration changes the parameter values over the simulation period
by reflecting the simulation output trend. Second, heterogeneous calibration changes the parameter values per
simulated entity clusters by considering the similarities of entity states. We experimented the suggested cali-
brations on one hypothetical case and another real-world case. As a hypothetical scenario, we use the Wealth
Distribution Model to illustrate how our calibration works. As a real-world scenario, we selected Real Estate
Market Model because of three reasons. First, the models have heterogeneous entities as being agent-based
models; second, they are economic models with real-world trends over time; and third, they are applicable to
the real-world scenarios where we can gather validation data.
Keywords Simulation Validation · Parameter Calibration · Machine Learning · Digital Twin · Agent Based
Model
1 Introduction
Simulation has been useful in market modeling [4], traffic management [34], and urban planning [22], and
these simulations are real-world based, validated simulations on well-defined scenarios. Because of the detail
and the fitness to the real-world, the simulation becomes a meaningful tool for design, management, and
analyses of modeled real-world. While modelers accept the fundamental requirement of the validation [2],
many simulation models are validated through parameter calibrations that can be easily invalidated as time
progresses or heterogeneity arises in the simulations. Some modelers manually calibrate the parameters to
improve the fitness, but this manual effort is often limited to the inital state of the simulation.
A simulation world is naturally diverging from the real-world as the simulation progresses, and the parameter
calibration should accommodate this natural divergence over the simulation period. If we utilize handpicked
or heuristically calibrated parameters, we cannot mitigate the natural divergence as often as needed. Hence,
a common practice in the field is calibrating simulation parameters at the up-front or over the predetermined
period [54] , and the calibrated parameters are used by ignoring the simulation divergence from the real-world in
the simulation run-time. Therefore, if we intend to accommodate the parameter calibration as often as needed,
we need to design an automatic calibration method that can resolve the natural divergence of simulation from
the real-world.
This automatic calibration requires two functionalities: when to calibrate and how to calibrate. As the
simulation diverges from the real-world, the automatic calibration needs to determine when is the best time
Dongjun Kim
E-mail: [email protected]
Tae-Sub Yun
E-mail: [email protected]
Il-Chul Moon
E-mail: [email protected]
2
Dongjun Kim et al.
to calibrate by itself, i.e. calibrations for every simulation timestep in extreme cases. Moreover, if we consider
the heterogeneity of simulated entities in the model as the source of divergence, the automatic calibration
needs to determine which groups of entities to be calibrated. After these decisions on when to calibrate, the
automatic calibration requires a component of how to calibrate. For instance, some parameters are temporal
and simulation-widely applicable, and other parameters are static and selected entity-widely applicable.
This paper categorizes the calibrated parameters into two groups: dynamic parameters and heterogeneous
parameters. First, the dynamic parameters are calibrated temporally, and the parameters are shared across
simulated entities at a certain timestep. To remedy the divergence over simulation time, we extract unobserv-
able regimes [16] that explain the temporal dynamics. In dynamic calibration, a hidden structure (regime)
is extracted from the deviation of observations and simulation results by a variant of Hidden Markov Model
(HMM) [11, 3]. Then, the regimes separately calibrate the dynamic parameters by taking the validation level
into account. These dynamic calibrations occur iteratively, so the iterative calibration cycle would result in
improving the fitness to the validation set.
Second, the heterogeneous parameters are calibrated for selected entities with the parameters. To mitigate
the simulation divergence from agent heterogeneity [39], we customize a static parameter by assigning different
values for different agent sub-populations to impose heterogeneity through the parameter. A hidden structure,
or agent sub-population, in this case is extracted from agent-level simulation results. Particularly, a cluster
model groups agents with similar agent-level state variables. To extract agent sub-populations, due to the curse
of dimension, we first reduce the dimension of agent-level state variables by applying a latent representation
learning algorithm, Variational Autoencoder [25]. Afterwards, a probabilistic mixture model [3, 53] is used
to obtain hidden sub-populations of simulation entities. After the clustering, we adopt the surrogate model
calibration methodology, Gaussian Process based Bayesian Optimization [47], applied to reduce the simulation
divergence.
The suggested calibration methodologies, dynamic calibration and heterogeneous calibration, could be con-
sidered in a single framework that the methods utilize hidden structures extracted from simulation state vari-
ables. According to our framework, hidden structures fundamentally determines the target simulation validity.
Hence, for instance, dynamic calibration fits the dynamic trend with capturing dynamically varying hidden
regimes, and heterogeneous calibration fits the agent heterogeneity by uncovering agent sub-populations. Be-
fore we discuss the framework, we limit ourselves to the in-sample validation [60]. In-sample validation focuses
on constructing a highly descriptive model with calibrated parameters, and we expect the in-sample validated
models to repeat the real-world scenario in the simulation.
2 Previous Research
Parameter calibration can be viewed as an inverse problem of estimating optimal simulation parameters that
make the most consistent result with real-world observations:
P = arg min
P
d(Eω∈Ω[S(M(P; ω))],D)1,
(1)
where P is the set of parameters to calibrate and P is the optimal parameters.
In detail, there are three steps in the parameter calibration [18]. The first step is choosing summary statistics
D from observations that are selected as emprirical validation data. Total S number of summary statistics are
considered, with each of summary statistic is the characteristic quantity of the observational distribution.
Simulation summary statistics are generated from the simulation output O with summary statistics function
S, given by S(M(P; ω) = O), by packing up the output distribution into few principal features, where the
agent-based model is M : P × Ω → O, as described in Tab. 1. For the next step, the simulation distance
formula, d, will be defined to measure the closeness between the simulation result S and the summary statistics
D [40, 31]. Simulation stochasticity is represented by ω, which is a sampled path from the sample path space
Ω. As the final step, the modeler applies an iterative stochastic function optimization algorithm to reduce the
simulation average error, d(Eω∈Ω[S(M(P; ω))],D).
Since the optimization lies at the core of the calibration process [54, 5], modelers have categorized opti-
mization algorithms for simulations into the four taxonomies [13] listed below:
1. Category 1 finds an optimal solution of stochastic function (under certain condition).
2. Category 2 finds an optimal solution of deterministic function (under certain condition).
3. Category 3 finds an optimal solution with probability p less than 1.
4. Category 4 finds a practically good solution stably.
These categories are generalized criteria, so each optimization algorithm is assessed by the criteria in the
following survey.
1 d(Eω∈Ω[S(M(P; ω))], D) represents the distance between the validation data D with the expectation over Ω of summarizing
(with a function of S) the output results (O) from executing an agent based model (as a function of M) with a sample random
path ω and a parameter P.
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
3
Table 1 Description of notations are presented in the following table.
Notation Description
M
δ
Agent-based model
Optimization algorithm
d
P
Pdyn
Phet
O
Ω
D
S
Ddyn
Dhet
Sdyn
Shet
Sagent
Slatent
G
Ndyn
Nhet
Ccal
Cdyn
Chet
Sdyn
Shet
Kdyn
Khet
A
T
R
I
U
Att
H
Simulation error distance measure
Simulation input parameters to optimize
Simulation dynamic parametrs to optimize
Simulation heterogeneous parameters to optimize
Simulation output state variables
Sample path space of stochastic simulation
Empirical validation summary statistics
Simulation summary statistics function
Empirical validation summary statistics data used in dynamic calibration
Empirical validation summary statistics data used in heterogeneous calibration
Dynamic calibration summary statistics function
Heterogeneous calibration summary statistics function
Agent-level simulation result
The representational compressed data of Sagent
Aggregated data of simulation error in heterogeneous calibration
The number of dynamic calibration target parameters
The number of heterogeneous calibration target parameters
The number of calibration framework iterations
The number of dynamic calibration iterations
The number of heterogeneous calibration iterations
The dimension of summary statistics in dynamic calibration
The dimension of summary statistics in heterogeneous calibration
The number of hidden regimes in HMM
The number of agent sub-populations
The number of agents
The number of total simulation time
The number of simulation replications for each candidate hypothesis
The number of candidate hypotheses
The number of merged regimes
The number of selected agent attributes for agent clustering
The dimension of hidden representation of agent attributes
2.1 Calibration with Design of Experiment
Besides of the algorithm-based adaptive calibration, the simplest systematic calibration is using the standard
design of experiments, or DOE [28]. However, the classical DOE algorithms have pre-determined input param-
eters, so those do not adaptively generate the parameters. In addition, the DOEs are not specifically designed
to handle the interference between the input parameters, which affects the simulation outputs, collectively.
The full factorial search, in Tab. 2, is powerful if there are few parameters to calibrate [54]. However, the
Latin hypercube, a near-uniformly spce filling search method by sampling from the equally probable axis-
aligned hyperplanes, could be an alternative of full factorial search if there are too many grids to evaluate. The
random search mainly acts as exploring the search space.
2.2 Calibration with Derivative-based Algorithms
The optimization algorithm could be a derivative-based calibration with gradient descents to minimize expected
simulation errors. This derivative-based optimization belongs to the category of Category 4 in the above tax-
onomy. Since the derivative of simulation error is not attainable, the derivative-based calibration requires to
use a derivative approximation techniques such as finite difference or Infinitesimal Perturbation Analysis (IPA)
[51, 20]. While the finite difference is attainable, this method is not appropriate for the calibration task with the
expensive evaluation cost since it requires a number of evaluations, proportional to the optimization dimension,
to get a derivative at a point. IPA approximates the derivative with a single function evaluation, but this single
4
Dongjun Kim et al.
Table 2 Previous researches on simulation static parameter calibration are listed. The suggested calibration methodologies in
this paper are listed in the last two columns.
The name of al-
gorithms
Search Method
factorial
Grid Search
Full
[54]
Global
optimality
Yes
Design of
Experiments
Latin
cube [54]
hyper-
Space filling search
No
Derivative
-based
Algorithms
Heuristic
Algorithms
Random [54]
Random search
Gradient De-
scent
with
finite difference
Gradient De-
scent with IPA
[51]
Gradient Descent
Gradient Descent
Simulation An-
nealing [17]
Local search with cool-
ing mechanism
Particle Swarm
Optimization
[37]
Update particles' posi-
tions and velocities to-
wards current best so-
lution
No
No
No
No
No
Genetic Algo-
rithm [49]
Selection,
and mutation
crossover,
No
Advantage
Disadvantage
Always find the global
optimum within grid
distance
Sampling from equally
probable
axis-aligned
hyperplanes
Computationally
expensive
Computationally
expensive
Exploration
No exploitation
Model-free algorithm
Multiple evaluation re-
quired for an update
Single
quired for an update
evaluation re-
Limited applicability,
high variance
Able to escape from lo-
cal minimum
Need to control
cooling rate
the
applicable
optimiza-
to
dimensional
Model-free
tion
high
parameters
No guarantee on the
convergence
the
global optimum
to
Robust
tion
model-free
in optimiza-
performance,
No guarantee on the
convergence
the
global optimum
to
Rejection Sam-
pling [52]
Rejection Sampling
Yes
Eventually converge to
the target distribution
Sampling
-based
Algorithms
Monte-Carlo
Markov Chain
[30]
Random walk from the
previous evaluated par-
ticles
Yes
Sequential
Monte
[46]
Carlo
Sampling
learned
distribution
from
a
proposal
Yes
Gaussian
Process
Regression-
based Bayesian
Optimization
[55]
XGBoost-
based Nelder
Mead
Op-
timization
[27]
Proposed Dy-
namic Calibra-
tion
Proposed Het-
erogeneous
Calibration
GPR as
surrogate,
Bayesian Optimization
as parameter search
Yes
XGBoost as surrogate,
Nelder Mead as param-
eter search
No
Sampling from a pos-
terior distribution for
each of merged regime
GPR-based Bayesian
Optimization
No
Yes
Model
-based
Algorithms
Proposed
Framework
(Including
interactions
between
dynamic and
heterogeneous
calibrations)
Highly inefficient if the
prior is dissimilar with
the target distribution,
only feasible for
the
calibration task of few
parameters
points
Sampled
are
highly correlated, slow
convergence rate
Choice of the number
of particles and the
forward-backword ker-
nels affects the perfor-
mance
No standard criteria on
the selection of kernel
and acquisition func-
tion
Eventually converge to
the target distribution,
much faster than rejec-
tion sampling
Avoid sampling ineffi-
ciency through learn-
ing of intermediate dis-
tribution, less prone to
get stuck in regions of
low probability
Theoretic analysis, fast
convergence
Applicable to high di-
mension
Parametric
model
surrogate
Able to calibrate the
dynamic parameter
No theoretical conver-
gence
Able to calibrate the
heterogeneous parame-
ter
Injecting heterogeneity
induces an optimiza-
tion with higher di-
mension than original
static calibration with
the same parameters
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
5
evaluation makes IPA too volatile to be used as a calibration method. In addition, IPA is less applicable than
the variants of finite difference methods.
2.3 Calibration with Heuristic Algorithms
Among derivative-free optimization algorithms, Heuristic algorithm is frequently used for the calibration task.
Heuristic algorithms search a next design point PC+1 with a current parameter PC by PC+1 = δ(PC ), where
δ is the optimization algorithm. Most of Heuristic algorithms, due to its memory-less property and simple
selection process, are not proven to find global optimum. However, some Heuristic algorithms with simple idea
work better than sophisticated designed optimization algorithms in the practice [49, 50], which makes the
Heuristic algorithms to be Category 4.
While some research applies either Simulated Aneealing [17, 26] or Particle Swarm Optimization [37, 24];
most of calibration articles, which construct calibration as an optimization problem, use Genetic Algorithm [8,
63, 35, 19, 50, 58] as an optimization solver. There are four main reasons for using Genetic Algorithm [36]. First,
Genetic Algorithm is robust on the task of global optimization for highly nonlinear multi-modal discontinuous
functions. Second, Genetic Algorithm is designed to align well with parallel computations in multi-processor
computers. Parallelizability is especially advantageous in a computationally expensive simulation optimization.
Furthermore, unlike standard DOEs, Genetic Algorithm is suitable in treating functions with input parameters
intricately entangled in its effect on outputs. Lastly, both continuous and discrete variables could be calibrated
using Genetic Algorithm, which makes both simulation parameter P and simulation structure to be optimized
at one hand.
2.4 Calibration with Sampling-Based Algorithms
In statistical inference algorithms, the Approximate Bayesian Computation(ABC) [1, 52, 30, 46] is a class
of sampling-based algorithms used when the likelihood, p(S(M(P; ω) = O) = DP), is intractible [23]. In
ABC, the likelihood is approximately defined as p(S(M(P; ω)) = DP) (cid:39) c · p(d(S(M(P; ω)),D) < τP),
where S is a summary statistics function, c is a normalization constant, and τ is a pre-difined threshold. The
likelihood might be analytically calculated in a simple stochastic simulation, but it is intractible in the case of
agent-based simulation [18, 54], where the simple agent-level behavior merged up to the emergent system-level
behavior. Also, this calculation is infeasible if the agent behaviors are the collection of discrete logics. Some
sampling-based ABC algorithms use only the previous parameters PC to generate the next set of design points
PC+1, by PC+1 = δ(PC ), and this generated parameters may form a stationary distribution which could be
an inferred distribution that satisfies the calibration task. Hence, under the condition of converging to the
stationary distribution, these sampling based ABC is categorized into either Category 1 or Category 2.
In spite of this optimization property, ABC inference algorithms suffers from the slow convergence rate since
the model requires the tens of thousands of point evaluations [18, 54] in the case of large scale agent-based
simulation. The proposed dynamic calibration algorithm in this article borrows an essential idea of Sequantial
Monte Carlo ABC algorithm [46], in that the next set of evaluation points, a.k.a particles, are sampled from
the proposal distribution generated from the current evaluation points with likelihoods.
2.5 Calibration with Model-Based Algorithms
Surrogate model-based optimization algorithms [55, 27] are derivative-free optimization algorithms to calibrate
with two iterative steps: predicting response surface, p(d(S(M(P; ω)),D)P), and setting new evaluation design
parameter, PC+1 = δ(P1, ...,PC ). The variants of regression models, including parametric and nonparametric
models, could be used in the construction of a response surface. In large scale simulation, nonparametric kernel-
based Gaussian process regression, or a.k.a. kriging, [41] would be flexible in estimating the response surface
when only a small number of evaluations are obtained. Gaussian process regression is a Bayesian model that
predicts a response surface, p(d(S(M(P; ω)),D)P), as a normal distribution, N (µ(P), σ2(P)). The predictive
variance, σ(P), measures the uncertainty of the predictive mean µ(P). As a parameter generation algorithm,
Bayesian optimization [44], using an acquisition function, with a Gaussian process regression is one of the most
efficient tools to find the global optimum in terms of function evaluations [29]. The proposed heterogeneous
calibration could be categorized as one of surrogate model-based algorithm.
Theoretical analysis, in particular, on Bayesian optimization with Gaussian pocess regression has been stud-
ied [32, 56, 15, 7, 42], and these analysis provides the convergence of the stochastic function with conditions[7].
Thus, the calibration with Gaussian process regression could be categorized in either Category 1, Category 2,
or Category 3 by the specifics of the conditions. Having said that, the practical conditions would be different
from the theoretical analyses, so we need to study the feasibility with an actual system with an implemented
calibration model.
6
3 Preliminaries
Dongjun Kim et al.
This section introduces machine learning algorithms that we used for our calibration framework to extract
hidden structures. These machine learning algorithms are used as components in our calibration framework.
We present a set of clustering and latent representation models because the calibration needs to capture the
divergence trend between the validation and the simulation results. Afterwards, we introduce a nonparametric
response surface model and its corresponding optimization model because the calibration needs to suggest the
next experimental point.
3.1 Clustering Models for Divergence Trend Detection
This subsection introduces multple clustering algorithms, Hidden Markov Model (HMM) [3], Gaussian Mixture
Model (GMM) [3] and Dirichlet Process Mixture Model (DPMM) [53] in Section 3.1.1 and 3.1.2. Also, we
present a latent structure modeling method, Variational Autoencoder (VAE) [25], in Section 3.1.3.
3.1.1 Temporal Clustering Models (HMM)
HMM [3] is a statistical graphical model that has an input as time-dependent data, and an output as time-
dependent regimes. The graphical architecture of HMM consists of two random variables: 1) the latent variable
zt for the hidden regime, and 2) the observable variable Ot where the graphical relations zt−1 → zt and zt → Ot
are assumed. The below is the generative process of HMM.
z1 ∼ Categorical(πinit)
ztzt−1 ∼ Categorical(πzt−1 )
Otzt ∼ F (θzt )
(2)
πinit is the initial probability of clustering; πzt−1 is the transition probability; and πzt−1=i = p(ztzt−1 =
i). Categorical denotes a categorical distribution; F is the emission probability distribution; and θzt is the
emission parameter, which specify the type and the shape of the distribution on observation variable Ot. If
an observation is assumed to follow the Gaussian distribution, then the emission distribution F becomes the
Gaussian distribution, with the emission parameter θzt of the mean µzt and the standard deviation σzt . The
Baum-Welch algorithm, which is the Expectation-Maximization algorithm on HMM, estimates the parameters
and the hidden regime assignments.
3.1.2 Clustering Mixture Models (GMM, DPMM)
In this subsection, we introduce an entity-wise static hidden structure extraction algorithm required in hetero-
geneous calibration. We define that a group has agents who share similarities in their state variables. Mixture
models follows a probabilistic approach for representing the presence of sub-populations within the overall pop-
ulation by assuming the observation data is distributed multi-modally. The mixture models have a categorical
cluster assignment variable, zi, as a latent variable for inference, which follows the categorical distribution of
πk. Each mixture component has the emission probability parameter, φk, and the emission probability distribu-
tion, F . The parametric model fixes the number of hidden groups, and the nonparametric model automatically
finds the optimal number of sub-populations.
Gaussian Mixture Model (GMM) [3] is a parametric clustering algorithm that the observation distribution
is assumed to follow the mixture of the Gaussian distributions. The emission parameter, φk, is consisted of the
mean µk and the standard deviation σk. We have the following generative process of GMM:
za ∼ Categorical(π)
Oaza ∼ N (µza , σ2
za )
(3)
A measure theoretic view of the parametric bayesian mixture distribution is GNc = (cid:80)Nc
Again, the Expectation-Maximization algorithm is used to maximize the likelihoods. In the Expectation step,
za is assigned by the expectation of the likelihoods. In Maximization step, we maximize the likelihoods by
optimizing parameters, such as πk, µk, and σk in GMM.
k=1 πkδθk , where
k=1 ∼ G0. Dirichlet process [53] is a prior for the nonparametric
{πk}Nc
mixture clustering model, in which the data-adaptive optimal number of clusters are automatically extracted.
As the mixture weight, {πk}∞
k=1, is no more finite dimension if the model is nonparametric, the infinite weight
prior is not a distribution of a vector, Dirichlet distribution, rather it is a distribution of measures, Dirichlet
process. Dirichlet Process [53] DP (γ) is a prior distribution of the infinitized mixture weight {πk}∞
k=1 that
k=1 πkδθk . The infinite weight, π from
k=1γ ∼ Dir(γ/Nc, ..., γ/Nc) and {θk}Nc
makes the mixture distribution into the infinite weighted sum G =(cid:80)∞
(cid:81)k−1
l=1 (1− π(cid:48)
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
7
k ∼ Beta(1, γ).
the Dirichlet process, is constructed by the stick-breaking process, πk = π(cid:48)
In general, base distribution G0 and emission distribution F are conjugate priors to make posterior computation
tractable. The below specifies the generative process of DPMM.
l) with π(cid:48)
k
π ∼ GEM (α)
za ∼ Categorical(π)
φk ∼ G0
Oaza ∼ F (φza )
(4)
The parameters of DPMM are inferenced by maximizing the posterior distribution. A Gibbs sampling algorithm
is used in maximizing a posterior in our implementation. At each iteration, the Gibbs sampling algorithm
updates za with fixed hidden structures of other assignments. After iterations, the Gibbs sampling method is
known to find the equilibrium distribution.
3.1.3 Variational Autoencoder
Agent-level state variables often have a large dimension that causes failure in extracting meaningful clusters
due to the curse of dimensionality. Assuming the data is distributed normally, as the dimension, n, gets larger,
most of the mass lies in the ellipsoidal shell of radius(cid:112)trace(Σ) measured by the Mahalanobis distance, where
single point with distance less than(cid:112)trace(Σ)/2 from the center of the distribution [9]. Therefore, we introduce
Σ is the covariance of the Gaussian distribution. This requires about 2O(n) sampled data points in order to get a
a latent representation learning algorithm to compress the high dimensional data into low dimensional data.
Variational autoencoder [25] is a probabilistic latent variable model that extracts the latent representation
of data. The structure of Variational Autoencoder is consisted of two neural networks: an encoder network
and a decoder network. An encoder network extracts the hidden latent representation from the input of the
original data. The extracted representation propagated to the decoder network to regenerate the original data.
The vanilla version of Variational Autoencoder assumes the encoder distribution to have a standard Gaussian
distribution N (0, I) as a prior.
VAE parametrizes the posterior distribution with the neural network parameters. To make the reconstructed
output similar to the original input, the log-likelihood of the Variational Autoencoder should be maximized.
However, since the log likelihood is intractable to calculate, an alternative loss function, Evidence Lower BOund
(ELBO), is suggested as a supporting lower bound function. In detail, the below formula shows ELBO.
(cid:90)
(cid:90)
z
z
log p(x) = log
p(xz)p(z)dz
= log
q(xz)
(cid:90)
p(xz)p(z)
q(zx)
p(xz)p(z)
dz
q(xz) log
≥
dz
= Eq[log p(xz)] − KL(qp)
q(zx)
z
(5)
The reconstruction term, Eq[log p(xz)], calculates the expectation of the log-likelihood when the posterior
latent representation follows a variational distribution of q. The regularization term, Kullback-Leibler divergence
KL(qp), measures how much information is lost by using the encoder distribution, q, to represent the prior
distribution, p. With a stochastic gradient descent method, the neural network parameters are learned.
3.2 Gaussian Process Regression-based Bayesian Optimization
After a sub-population of agents is obtained, heterogeneous calibration becomes an optimization problem with
a heterogeneity-embedded loss function. This subsection introduces the surrogate based algorithm to optimize
heterogeneous calibration problem, Gaussian process regression-based Bayesian optimization. Gaussian process
regression (GPR) [41] is a Bayesian nonparametric regression algorithm that estimates a response variable, or
simulation error in this paper, as a black-box function of dependant variables, or simulation parameters in our
case. If P is an arbitrary input parameter, GPR estimates the predictive output distribution as a Gaussian
distribution of N (µerr(P), σ2
is given, where d(ER[S(M(Pc(cid:48) ; ωr))],D) is the
average of summarizing (with a function of S) output results (O) from executing an agent based model (as a
function of M) with a sample random path ωr and a parameter Pc(cid:48) with R replications. Then, the predictive
posterior distribution is specified in the below.
Suppose the data G =(cid:8)(cid:0)Pc(cid:48) , d(ER[S(M(Pc(cid:48) ; ωr))],D)(cid:1)(cid:9)c
err(P)).
c(cid:48)=1
p(Eω∈Ω[S(M(P; ω))]G) ∼ N (κT cov
−1
c F, KP,P − κT cov
−1
c κ)
(6)
8
κ = [KP,P1 , ..., KP,Pc ]T ; F =(cid:2)d(ER[S(M(P1; ωr))],D), ..., d(ER[S(M(Pc; ω))],D)(cid:3)T ; and (covc)i,j = KPi,Pj +
β δ{i=j}. Here, KP,P(cid:48) is the kernel distance between P and P(cid:48). The kernel distance is selected by considering
Dongjun Kim et al.
1
the shape of the true function and the meaning of each input variable.
The kernel function, K, plays the key role in GPR. Stationary kernel functions, such as Squared Exponential
kernel and Matern kernel, are well-studied in theory and experiments. A kernel function has hyperparameters
to determine the exact shape of distances where the hyperparameters could be learned by maximizing the
log likelihood from data. Recently, a fully Bayesian treatment of kernel hyperparameters showed its usefulness
[47, 62].
The Bayesian optimization [47, 6, 12, 29, 43, 38] is an adaptive optimization tool that selects a new design
point to optimize the response. It finds the next evaluation point by optimizing an acquisition function, which
is generated using µerr(·) and σerr(·). This acquisition function optimization problem is represented in the
below.
This is called an inner optimization problem, where AF is an acquisition function, and H =(cid:8)P(cid:12)(cid:12)(cid:107)P − Pc(cid:48)(cid:107)2 ≥
r0 for all Pc(cid:48) ∈ G(cid:9) with given radius r0 ≥ 0. H prevents a calibration optimization being stuck in a local
Pc+1 = arg max
P∈H
AF (PG)
(7)
optimum by forcing an exploration. A single calibration iteration contains an optimization of an easy-to-
optimize inner problem.
The exploration-exploitation ratio affects the convergence speed of the calibration problem. The purely ex-
ploitation search algorithm is using the predictive mean acquisition function, µerr(·). Minimizing the predictive
mean will find a next evaluation point with a pure exploitation. For the pure exploration search algorithm,
the predictive variance acquisition function, σerr(·), is used. In this case, maximizing this function will find a
design parameter that has never been visited before.
The Expected Improvement (EI) is a mixture of the exploration and the exploitation. The below formula
is an acquisition function of EI.
EI(PG) =
P (y(x) ≤ ymin − m)dm
(cid:90) ∞
0
=(min
c(cid:48) d(ER[S(M(Pc(cid:48) ; ω))],D) − µerr(P))Φ
(cid:18) minc(cid:48) d(ER[S(M(Pc(cid:48) ; ωr))],D) − µerr(P)
(cid:19)
+σerr(P)φ
σerr(P)
(cid:18) minc(cid:48) d(ER[S(M(Pc(cid:48) ; ωr))],D) − µerr(P)
(cid:19)
(8)
σerr(P)
φ is a probability density function, and Φ is a cumulative density function of the standard Gaussian distribution,
N (0, 1). The integration in Eq. 8 is marginalized out as the addition of two terms, where the first term (exploita-
tion) is the predicted difference between the minimum of simulation errors, minc(cid:48) d(ER[S(M(Pc(cid:48) ; ωr))],D),
and the predicted mean, µerr(P), at the current point, P, penalized by the probability of improvement,
Φ({minc(cid:48) d(ER[S(M(Pc(cid:48) ; ωr))],D) − µerr(P)}/σerr(P)). The second term (exploration) represents the uncer-
tainty of the surrogate model predictions. Therefore, EI is a balanced exploration and exploitation method in
its nature. By maximizing the EI acquisition function, the next design parameter, Pc+1, is generated. EI gained
its popularity with both theoretical analyses [32, 56, 15, 7, 42] and experimental successes [47, 29, 44].
EI is one-step Bayes optimal [61], but EI becomes heuristic when we consider multi-steps. In calibration
tasks, EI often fails to find a global optimum parameter in the following reasons. First, the effect of hyperpa-
rameters is not well-studied. If we estimate the hyperparameters via MLE [47], EI finds the basin of attraction
only for certain circumstances because the hyperparameters are changing at every iterations [7]. Meanwhile,
if we adopt a fully Bayesian treatment of hyperparameter [47, 62], then the theoretic analysis on asymptotic
consistency of EI in continuous parameter space is not equipped [62]. In practice, EI, with either Bayesian treat-
ment or pointwise treatment on hyperparameters, fails to detect the basin of attraction and stops exploring,
which makes no improvement on the response surface after iterations.
More importantly, looking at the shape of the true error function, a well-modeled simulation is likely to
have a sub-optimal plateau near the global minimum. As the simulation model becomes more insensitive on its
input parameters, the sub-optimal plateau area becomes bigger. On the sub-optimal plateau, if we assume the
predicted difference minc(cid:48) d(ER[S(M(Pc(cid:48) ; ωr))],D)−µerr(P) to be constant, denoted by µ0, then the derivative
of EI w.r.t. σerr is specified in the below.
µ0Φ
∂
(cid:20)
= −(cid:16) µ0
(cid:16) µ0
∂σerr
= φ
σerr
σerr
(cid:17)
(cid:16) µ0
(cid:17)2
(cid:16) µ0
(cid:17)
σerr
φ
σerr
(cid:17)
+ σerrφ
(cid:16) µ0
(cid:16) µ0
σerr
(cid:17)(cid:21)
(cid:17)(cid:20)
+ φ
σerr
1 +
(cid:17)2(cid:21)
(cid:16) µ0
σerr
(9)
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
9
The above formula is turned out to be positive in any cases. This suggests that EI is an increasing function
of σerr, so EI will explore on the sub-optimal plateau that it finds Pc+1 as a next design point that makes
σerr(Pc+1) the biggest. In these reasons, some articles investigate the portfolio of the mixture of acquisition
functions [21] to accelerate optimization, which we follow in this calibration task.
Besides of using EI in its fundamental form, the ratio of the exploitation term and the exploration term
could be controlled by adding weight wc as the below.
w-EI(PG)
= (1 − wc) ×(cid:0) min
+ wc × σerr(P)φ
c(cid:48) d(ER[S(M(Pc(cid:48) ; ω))],D) − µerr(P)(cid:1)Φ
(cid:18) minc(cid:48) d(ER[S(M(Pc(cid:48) ; ω))],D) − µerr(P)
(cid:18) minc(cid:48) d(ER[S(M(Pc(cid:48) ; ω))],D) − µerr(P)
(cid:19)
σerr(P)
(cid:19)
(10)
σerr(P)
We call this acquisition function as weighted Expected Improvement (w-EI) [48].
Parallel Bayesian Optimization algorithms [61, 62, 57] could enhance the calibration speed dramatically,
that they could find a set of multiple heterogeneous parameters PC+1,PC+2, ...,PC+L to test. The distributed
parallel simulation then tests L different set of heterogeneous parameters and yield L different simulation errors.
Parallel Bayesian Optimization also supports asynchronized situation where different simulation setup takes
different error evaluation time.
4 Proposed Calibration Methodology
This section describes our calibration framework for agent-baed models with a validation dataset. The descrip-
tion starts from the overall illustration of the calibration framework. Then, we describe two detailed calibration
components, which are dynamic calibration and heterogeneous calibration, in turn.
4.1 Overall Calibration Framework
Our calibration framework intends to optimize the parameters of a generic agent based model in order to
reduce an error function between a validation dataset and a set of stochastic simulation summary statistics.
Some of these components are already defined in the previous literature in the previous section. For instance,
we represent the agent based model as M; parameter set as P; an error function as d; a validation dataset as
D; and a set of stochsatic simulation output summary statistics as S(M(P; ω) = O).
As all of the surveyed calibration methods can be executed in an iterative setting, our model assumes an
iterative optimization of P, which is illustrated in Fig. 1 and Alg. 1. Before we start our iterations, we create a
set of agent clusters because our heterogeneous calibration will be applied to each group of agents, rather than
an individual agent due to the time complexity, see line 3 in Alg. 1. Also, we assume that we obtained a pre-
processed validation dataset D that is a set of temporal sequence values. Afterwards, we start the calibration
iterations from a randomly initialized parameter. Having said that, our calibration would better work if the
initial parameter is close to the optimal parameter P∗, and we provide this trade-off in Sec. 5.1.3.
Our iterations consist of one minor step and two major components with machine learning models.
1. Our iteration starts from a minor step of calculating d(ER[S(M(Pc; ωr))],D) from the most recent simu-
lation outputs M(Pc; ωr) with the most recent parameter PC , see line 15 in Alg. 2 and line 26 in Alg. 3.
The function d is instantiated for each application case, and d can be either simple Euclidean distance, KL
divergence, negative log-likelihoods, etc. We specify d for each of our examples in Section 4.2 and 4.3.
2. After calculating d, we perform either dynamic calibration or heterogeneous calibration, which are the
two major components of the calibration framework. The proposed calibration framework alternates each
calibration with pre-determined iterations, to stabilize the estimation of each parameter: Cdyn iterations
for the dynamic component, and Chet iterations for the heterogeneous component.
(a) Dynamic calibration takes the result of d, d(ER[Sdyn(M(Pc; ωr))],Ddyn), simulation mean, ER[Sdyn(M(Pc; ωr))],
and the current dynamic parameter candidate hypotheses, P i
dyn,c, to perform the regime detection, see
line 17 in Alg. 2, and to generate the next dynamic parameter hypotheses P i
dyn,c+1, see line 18 in Alg.
2.
(b) Heterogeneous calibration accepts the result of d, d(ER[Shet(M(Pc; ωr))],Dhet), and Phet,c to learn the
response surface, see line 28 in Alg. 3, and to generate the heterogeneous parameter Phet,c+1, see line
29 in Alg. 3.
c+1 of ith candidate hypothesis
for the next iteration, our iterative calibration becomes a loop that measures d and optimizes P, alterna-
tively.
dyn,c+1 and Phet,c+1 becomes a new parameter input P i
3. Since the union of P i
The following sections show the details of two major components.
10
Dongjun Kim et al.
Fig. 1 The suggested calibration framework overview is presented. Agent Clustering Component is executed at the beginning
of the calibration, and next the calibration iteration starts. For each iteration, either dynamic calibration or heterogeneous cali-
bration are activated. Dynamic calibration contains two components: Regime Detection Component and Parameter Generation
Component. Heterogeneous calibration in the iteration also contains two components: Response Surface Learning Component
and Parameter Generation Component
Algorithm 1: Calibration Framework Algorithm
input : Input parameter combination P in = P in
dyn ∪ P in
output: Calibrated parameter combination P out = P out
het
dyn ∪ P out
het
1 Function CalibrationFramework(P in
dyn ∪ P in
het):
dyn ∪ P in
het) (see Alg. 3)
dyn
Pdyn,0 = P in
Phet,0=AgentClustering(P in
for c in range(Ccal) do
if 0 ≤ c −(cid:104)
else if Cdyn ≤ c −(cid:104)
c
Cdyn+Chet
(cid:105)
(cid:105)
(Cdyn + Chet) < Cdyn then
Pdyn,c+1=DynamicCalibration(Pdyn,c ∪ Phet,c) (see Alg. 2)
Phet,c+1 = Phet,c
c
Cdyn+Chet
(Cdyn + Chet) < Cdyn + Chet then
Phet,c+1=HeterogeneousCalibration(Pdyn,c ∪ Phet,c) (see Alg. 3)
Pdyn,c+1 = Pdyn,c
2
3
4
5
6
7
8
9
10
11
12
Set P out = P opt
return P out
dyn ∪ P out
het
dyn ∪ P opt
het to have the lowest simulation error
4.2 Dynamic Parameter Calibration
Dynamic parameter calibration [33] is an inverse problem of finding optimal parameters that yields the most
consistent simulation summary statistics, Sdyn, with given validation empirical data, Ddyn. To further illustrate
the calibration algorithm, we state two assumptions. First, we assume that we have a fixed parameter set of
Phet, which is P\Pdyn. This assumption limit this algorithm to finding only Pdyn, which is a subset of P,
without the joint consideration with Phet. Second, we assume that the dynamic parameter is assigned to each
timestep as a switching regime. Formally, if we assume to calibrate Ndyn number of dynamically switching
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
parameter, Pdyn = {P n,t
dynn = 1 . . . , Ndyn, t = 1, . . . , T}. We describe our optimization task as the below.
P∗
dyn = arg min
Pdyn
d(Eω∈Ω[Sdyn(M(Pdyn ∪ Phet; ω))],Ddyn)
11
(11)
We treat the dynamic parameters for each timestep, P n,t
dyn, as the static input determined prior to the
simultion, so the dynamic parameters become the composition of such parameters multiplied by the execution
timesteps, T . This can be viewed as turning the dynamic parameters into the static parameters by increasing
the number of parameters, which could be viewed as an identical problem of Eq. 11. However, this increment of
parameter numbers is significant if the simulation becomes a long execution. Therefore, we limit the increment
by assuming that the dynamic parameters are identical for identified temporal regimes over the simulation
executions. This assumption enables limiting the multiplication of dynamic parameter numbers to the given
regime number, Kdyn.
Because of the above reasoning, our calibration algorithm divides simulation timesteps into a number of
regimes to generate the dynamic parameters for each regime, separately. Also, we separate well-fitted regimes
from poorly-fitted regimes. This separation enables that the poorly-fitted regimes can explore farther parameter
spaces while the well-fitted regimes can exploit the bounded parameter spaces.
4.2.1 Dynamic Calibration Performance
This subsection explains line 12-15 in Alg. 2, which measures the performance of dynamic calibration. Dynamic
calibration uses a posterior distribution inferred from I number of candidate hypotheses, P i
dyn,c, to generate
a new set of parameters, P i
dyn,c+1 for (c + 1)th iteration. To obtain statistically stable simulation summary
statistics, we run simulation R times for each ith candidate hypothesis, see line 13 in Alg. 2. Then, the simulation
r = M(P i
dyn,c ∪Phet; ωr) of ith candidate hypothesis are post-processed to generate
generates state variables Oi
the simulation summary statistics means as follows.
ER[Sdyn(Oi
r)] = {µi
s,ts = 1, . . . ,Sdyn, t = 1, . . . , T, i = 1, . . . , I}.
(12)
simulation variance. The simulation variance is set to be σi
of true observation Ddyn =(cid:0)(ds,t)S
The distance function d in dynamic calibration is the negative likelihood function, which requires the
s,t in this article. Then, the likelihood
being observed from the normal distribution with the output mean
s,t = 0.1 × µi
(cid:1)T
s=1
t=1
s,t and the variance σi
µi
s,t is calculated as below.
The joint likelihood over all the summary statistics is Li
the log-likelihood of Eq. 13.
s,t. The implementation version utilizes
s,t = N (ds,tµi
Li
s,t)
s,t, σi
t = (cid:81)Sdyn
s=1 Li
(13)
(14)
4.2.2 Regime Detection Component
After calculating the likelihoods, we apply the regime detection algorithm of HMM to obtain temporal clustering
results by regarding the deviations of the simulation result, µi
s,t, and the validation data, ds,t, as observed
variables in Eq. 2, introduced in Section 3.1.1, as below.
(cid:16)
Ot =
1,t − d1,t, . . . , µiSdyn,t − dSdyn,t
µi
(cid:17)
The temporal clustering results are Ri
regimes of all the candidate hypotheses, we define merged regimes {MRu}U
timesteps that satisfy the following conditions: if t1, t2 ∈ MRu,
t ∈ {1, . . . , Kdyn} for each of ith candidate hypothesis. After obtaining
u=1 as the partition of simulation
(R1
t1 , ..., RI
t1 ) = (R1
t2 , ..., RI
t2 )
(15)
We utilize MR as the final regime detection result, which answers to the question of when to calibrate. Also,
MR becomes the input to the algorithm of answering how to calibrate.
12
Dongjun Kim et al.
Algorithm 2: Dynamic Calibration
Input : nth Dynamic parameter P n,i
Input : Simulation summary statistics mean µi
Output: Merged regimes M Ru
1 Function RegimeDetection(P n,i
dyn,c, µi
s,t):
s,t
dyn,c in ith hypothesis at cth iteration
2
3
4
5
for i in range(I) do
Apply Hidden Markov Model to obtain the regime detection results for the ith candidates
Merge regime detection results over the candidate parameters to split the simulation time into merged regimes M Ru
return M Ru,
Input : nth dynamic parameter hypotheses P i
Input : Likelihoods Li
Input : Merged regimes M Ru
Output: nth calibrated parameter hypotheses P n,i
s,t
dyn,c+1
dyn,c at cth iteration
6 Function ParameterGeneration(P n,i
dyn,c, Li
s,t, M Ru):
for u in range(U ) do
Generate the next parameter hypotheses P n,i
return P n,i
dyn,c+1
Estimate the beta distribution Beta(αu, βu), using normalized parameters and normalized joint likelihoods
dyn,c+1 from estimated distribution Beta(αu, βu)
7
8
9
10
12
13
14
15
16
17
18
19
Input : Parameter combination, Pdyn,c ∪ Phet,c, of the cth calibration
Output: Calibrated dynamic parameter Pdyn,c+1
11 Function DynamicCalibration(Pdyn,c ∪ Phet,c):
for i in range(I) do
Run simulation R times with the ith candidate hypothesis, P i
c = P i
dyn,c ∪ Phet,c
Obtain simulation mean µi
Obtain the likelihoods Li
for n in range(Ndyn) do
s,t and variance σi
s,t = N (ds,tµi
s,t, σi
s,t
s,t)
M Ru =RegimeDetection(P n,i
P n,i
dyn,c+1 =ParameterGeneration(P n,i
dyn,c, µi
s,t) (see Section 4.2.2)
dyn,c, Li
s,t, M Ru) (see Section 4.2.3)
return Pdyn,c+1
4.2.3 Parameter Generation Component
We generate the next set of paramter hypotheses from each merged regime, MR. In detail, from the merged
regime MRu, we infer the posterior distribution of the parameter space, Beta(αu, βu), to sample the next
parameter candidate hypotheses P i
dyn,c+1. To estimate the uth beta distribution parameters, we use the current
tt ∈ MRu, i = 1, . . . , I}. In particular,
parameter values {P n,t,i
the parameter and the likelihood values are normalized to estimate the beta distribution parameters αu and
βu via maximum likelihood estimation. The parameter is normalized by
dyn,ct ∈ MRu, i = 1, . . . , I} and the likelihoods {Li
P n,t,i
dyn,c =
dyn,c − P n
P n,t,i
dyn − P n
P
n
dyn
dyn
,
(16)
n
dyn and P n
where P
likelihood is normalized by
dyn are the maximum and the minimum of the nth parameter range, respectively. The
(cid:80)I
i=1
(cid:80)
Li
t
t∈MRu
Li
t =
.
Li
t
(17)
Then, the estimated beta distribution, Beta(αu, βu), generates a next set of parameters by either sampling or
Mode Selection, which is explained, later.
If a simulation is far from the validation for all the candidate hypotheses, additional parameter exploration
is executed in order to avoid searching prohibitive region. If the likelihoods are less than given threshold for all
the candidate hypotheses and for all the simulation timestep t ∈ MRu,
t + ratio × maxt Li
(18)
1 + ratio
t = 1/(I × M Ru). This equally distributed likelihoods
then the normalized likelihoods are set to be equal, Li
make the estimated beta distribution equally distributed, and the next parameter will be randomly picked at
the uth merged regime. The ratio is decreased in this article, with ratio = 0.9c.
mint Li
t <
Li
,
t
Given a posterior distribution, Beta(αu, βu), we have three options to generate new parameters:
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
13
1. Sampling by Time: the nth dynamic parameters P n,t,i
and i = 1, . . . , I.
dyn,c+1 are sampled from Beta(αu, βu) for each t ∈ MRu
dyn,c+1 are set to be constant for all t ∈ MRu, where
dyn,c+1 are sampled from the modes of Beta(αu, βu), where
2. sampling by Regime: the nth dynamic parameters P n,t,i
3. Mode Selection: the nth dynamic parameters P n,t,i
the constant is sampled from Beta(αu, βu) for each i = 1, . . . , I.
the modes are meaningful moments of the distribution. In this article, we use(cid:8)µβ
u +(cid:0)i− I+1
2
(cid:1)σβ
u
(cid:12)(cid:12)i = 1, ..., I(cid:9)
u are the mean and the standard deviation of the Beta distribution,
to be the set of modes, where µβ
respectively.
u and σβ
Sampling by Time searches the most rapidly switching parameters among three generation processes. How-
ever, the small parameter gap at the front timesteps affects to the simulation dynamics afterwards, and this
dependency hinders dynamic calibration to estimate accurate time-varying parameters afterwards. Further-
more, Sampling by Time is expected to estimate the overfitted, unrealistic dynamic parameters. On the other
hand, Sampling by Regime searches parameters with an identical value at the identical regimes, which makes
the calibrated parameter to be less fluctuating. Besides of the sampling-based methods, Mode Selection fo-
cuses on exploitation in parameter generations. Exploration is implemented when the likelihoods are below the
threshold, and Mode Selection do not explore the parameter space, additionally. We experiment all parameter
generating rules in Section 5.1.
4.3 Heterogeneous Parameter Calibration
Heterogeneous calibration resolves the simulation divergence arisen from the agent heterogeneity by transform-
ing the simulation result, Shet, to the validation data, Dhet, through calibrating heterogeneous parameters.
Before introducing the algorithm, we present two assumptions. First, we assume that we have a fixed set of
dynamic parameters, Pdyn, which are the complement of heterogeneous parameters from the set of model pa-
rameters, P. Heterogeneous calibration focuses on calibrating the heterogeneous parameters, Phet. Second, the
heterogeneous parameters, Phet, are static parameters differentiated by agent sub-populations. The heteroge-
neous calibration solves the following inverse problem.
P∗
het = arg min
Phet
d(Eω∈Ω[Shet(M(Phet ∪ Phet; ω))],Dhet)
(19)
The heterogeneous parameters are not differentiated by agents, since the above optimization problem becomes
infeasible. This infeasibility comes from the huge dimensionality of the problem, where the optimization dimen-
sion, A × Nhet, is the multiplication of the number of agents, A, by the number of heterogeneous parameters,
Nhet. The second assumption enables the problem to feasible by reducing the dimension, Khet × Nhet, where
Khet is the number of sub-populations.
4.3.1 Heterogeneous Calibration Performance
In heterogeneous calibrations of our experiments, the simulation error function, d, is the Mean Absolute Percent-
age Error (MAPE), but this selection can change per application case. The experimental error, d(ER[Shet(M(P; ωr))],Dhet),
corresponds to the noisy observation of the true error, d(Eω∈Ω[Shet(M(P; ω))],Dhet). We generate the next
design point Phet,c+1 by Bayesian optimization, with a predicted surrogate response surface obtained from the
gathered data G =
(cid:110)(cid:0)Phet,c(cid:48) , d(ER[Shet(M(Pdyn ∪ Phet,c(cid:48) ; ωr))],Dhet)(cid:1)(cid:111)c
.
c(cid:48)=1
4.3.2 Agent Clustering Component
A clustering algorithm divides agents into several sub-populations, using the simulation agent-level results,
Sagent. Since the classical clustering algorithms, such as GMM and DPMM, suffers from the curse of large
dimensionality, we compress the data through a latent representation learning algorithm, VAE (see Section
3.1.3). The vanilla VAE has an encoder-decoder neural network, Enc : RAtt×T → RH and Dec : RH → RAtt×T ,
where the two functions, embedded in a neural network, are learned by maximizing the ELBO, given in Eq.
5. The latent representation of an agent, LRa = {LRa
hh = 1, . . . , H} when a ∈ {1...A}, is the output of the
encoder-part network of the agent-level simulation results of the agent.
agents, {LR1, . . . , LRA}.
The next step executes the mixture models, introduced in Eq. 3 or 4, with the latent representation of
Oa = (LRa
1 , . . . , LRa
H ),
(20)
We test both parametric mixture model, Eq. 3, and nonparametric mixture model, Eq. 4, in our calibration.
2
3
4
5
6
8
11
12
13
14
15
16
17
18
19
20
21
22
14
Dongjun Kim et al.
Algorithm 3: Heterogeneous Calibration
Input : Union of calibrated dynamic parameters and initial heterogeneous parmaeters Pdyn ∪ Phet
Output: Heterogeneity embedded initial heterogeneous parameters P0
1 Function AgentClustering(Pdyn ∪ Phet):
Run simulation R times with input parameters, Pdyn ∪ Phet, to obtain agent-level simulation result Sagent
Apply latent representation extraction algorithm (VAE) on agent-level simulation result to obtain compressed
representations
Apply clustering algorithm (GMM or DPMM) on the latent representations to obtain agent sub-populations
assignments
Generate the heterogeneous parameters, Phet, by assigning different parameter values in different agent
sub-populations
return Phet
Input : parameter-error aggregated data G
7 Function ResponseSurface(G):
Learn Kernel hyperparameters in Gaussian process regression to fit the given data G
Output: Next heterogeneous parameter Phet,C+1
9 Function ParameterGeneration():
10
if C < C0 then
else if C ≥ C0 then
Randomly select Phet,C+1 as the next parameter inputs
Sample ξ from uniform distribution with support [0, 1]
if ξ < ξrand then
else if ξrand ≤ ξ < ξrand + ξP V then
else if ξrand + ξP V ≤ ξ < ξrand + ξP V + ξP M then
else if ξ ≥ ξrand + ξP V + ξP M then
Randomly select Phet,C+1 as the next parameter inputs
Maximize the Predictive Variance to obatin a new parameter set Phet,C+1
Minimize the Predictive Mean to obtain a new parameter set Phet,C+1
Maximize the weighted Expected Improvement with cooling weight ωC = 0.99C−Cdyn /2 to obtain a new
parameter set Phet,C+1
return Phet,C+1
Input : Parameter combination, Pdyn,c ∪ Phet,c, of the cth iteration
Output: Calibrated heterogeneous parameter, Phet,c+1, for the next iteration
24
25
23 Function HeterogeneousCalibration(Pdyn,c ∪ Phet,c):
dyn,c ∪ Phet,c
Select the best candidate of the dynamic parameter, P best
Run simulation R times with Pc = P best
Calculate the error metric d from the heterogeneous summary statistics and validation data
Add simulation result G = G ∪ {Phet,c, d(ER[Shet(M(Pc; ω))], Dhet)}
ResponseSurface(G) (see Section 4.3.3)
Phet,c+1 =ParameterGeneration() (see Section 4.3.4)
return Phet,c+1
26
27
28
29
30
dyn,c
4.3.3 Response Surface Learning Component
Gaussian process regression estimates the predictive posterior distribution of the true error, Eω∈Ω[d(Shet(M(Pdyn∪
Phet; ω)),Dhet)]. GPR, introduced in Section 3.2, predicts the mean and the variance by
µerr(P) = κT cov
err(P) = KP,P − κT cov
σ2
−1
c F
−1
c κ,
where κ = [KP,P1 , ..., KP,Pc ]T , F =(cid:2)ER
(cid:2)d(Shet(M(Pdyn∪Phet,1; ωr)),Dhet)(cid:3), ..., ER
(cid:2)d(Shet(D(P∪Phet,c; ωr)),Dhet)(cid:3)(cid:3)T ,
and (covc)i,j = KPi,Pj + 1
Matern-5/2 kernel as the prior covariance. The hyperparameters of kernel function are point-estimated by
maximizing the likelihoods of collected data being sampled from the predictive distribution.
β δ{i=j}. Here, KP,P(cid:48) is the kernel distance between points P and P(cid:48). We use the
4.3.4 Parameter Generation Component
As a selective search algorithm, Bayesian optimization optimizes the acquisition function to find a next evalu-
ation design point as follows.
Phet,c+1 = arg max
Phet∈H
AF (PhetG)
(21)
(22)
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
Here, AF stands for the acquisition function, and H =(cid:8)Phet
(cid:12)(cid:12)(cid:107)Phet−Phet,c(cid:107)2 ≥ r0 for all Phet,c in the domain of G(cid:9)
15
with a given radius r0 ≥ 0. This adaptive parameter space, H, forces the search algorithm to avoid the r0-
neighborhood of the previous evaluated parameters. We use L-BFGS-B algorithm to optimize the acquisition
function, with r0 = 0.
We apply three types of Acquisition Functions: weighted Expected Improvement, Predictive Mean and Pre-
dictive Variance. We use search strategy as follows:
-- Choose random parameters for the first C0 iterations
-- For the next iterations,
∗ With probability ξrand, choose random parameters
∗ With probability ξP V , choose parameters that maximize Predictive Variance
∗ With probability ξP M , choose parameters that minimize Predictive Mean
∗ With probability ξw−EI , choose parameters that maximize weighted Expected Improvement with cooling
weight wc = 0.99c/2
We use multiple acquisition functions to search parameter space without being trapped by local minima. In
order to escape from local minima, the search strategy contains the exploration algorithms, such as Random
Search and Predictive Variance, where Predictive Variance finds the least searched parameter region. The
role of exploration in heterogeneous calibration is estimating the surrogate function closer to the true error
function, globally. In order to estimate the optimal parameters, the search strategy executes the exploitation by
Predictive Mean and weighted Expected Improvement. The cooling rate, wC , in weighted Expected Improvement
controls the weight between exploration and exploitation in Eq. 10, to find a narrower region in large iterations.
Predictive Mean is a pure exploitation algorithm to find the global minimum of the surrogate function.
5 Experimental Result
Section 5 introduces the experimental results of the suggested calibration framework in two test cases. First,
we test the feasibility of our algorithm on Wealth Distribution Model [59]. We synthesized a validation dataset
with an arbitrailiy chosen parameter set, and this synthesis limits the source of simulation divergence to random
effects and the pre-determined parameters set. Second, we experiment the real-world applicability with a Real
Estate Market Model which uses a real validation dataset with elaborately designed model structures and real
input scenarios. In such a realistic case, the additional sources of divergence exist, including the imperfect
modeling and the highly noisy socio-economics data.
Both cases experiments the calibration framework suggested in Alg. 1. To investigate the effects of each
component in the framework, we also experiment dynamic calibration and heterogeneous calibration, seperately.
We evaluate dynamic calibration without heterogeneous calibration by setting the heterogeneous parameters to
be either known optimal parameters, in test case 1, or human calibrated parameters, in test case 2. Similarily,
we assess heterogeneous calibration without dynamic calibration by setting the dynamic parameters to be either
optimal parameters or human calibrated parameters.
5.1 Test Case 1: Wealth Distribution ABM
5.1.1 Model Description
Wealth Distribution Model [59], adapted from the sugarscape model [10], is an agent-based model to investigate
the macroscopic wealth distribution via microscopic agent behaviors. In the model, a grid provides its wealth to
agents located at the grid. At the end of each simulation timestep, agents consume their wealth to survive and
move toward the wealthest neighboring grid to maximize their wealth income. Also, grids recover their wealth
at the end of each timestep for future provision. The net-wealth of all grids is proportional to the dynamic
parameter, Wealth Income, and agents consume their wealth, proprortional to the heterogeneous parameter,
Wealth Consumption.
5.1.2 Virtual Experimental Design
5.1.2.1 Synthetic Parameter Setting Tab. 3 presents the parameters used in Wealth Distribution ABM. The
dynamic parameter, Wealth Income, represents the seasonal effects, or the up and down of economics business
cycles. The synthetic dynamic parameter is alternating between 1.5 and 0.5 by a period of ten simulation
timesteps. The heterogeneous parameter, Wealth Consumption, represents the agent level heterogeneous char-
acteristics. Agent clusters are divided according to the initial wealth; agents with top 50% in their initial wealth
are clustered, and the other agents are separately clustered. The synthetic heterogeneous parameter is set to
be 0.9 for the first cluster, and 0.1 for the second cluster.
16
Dongjun Kim et al.
Table 3 Calibration parameters in the first test case are listed. The dynamic parameter, Wealth Income, and the heterogeneous
parameter, Wealth Consumption, are the essential variables that control the model dynamics
Parameters
Parameter
Type
Parameter
Range
Synthetic Parameter Setting
Value
Time or Cluster
Wealth Income
Dynamic
0-2
Wealth
sumption
Con-
Heterogeneous
0-1
1.5
0.5
0.9
0.1
1-10,21-30,41-50
11-20,31-40
Top 50% in Initial Wealth
Bottom 50% in Initial Wealth
Table 4 The validation summary statistics in the first test case are listed. The validation data is the average of the 300
simulation replications with synthetic parameters as input parameters
Type of Sum-
mary Statistics
Validation
Summary
Statistics
Name of Summary Statistics
Variable Description
High Class Wealth Average Average wealth of top 1/3 agents
Middle Class Wealth Aver-
age
Low Class Wealth Average Average wealth of bottom 1/3 agents
Gini Index
Average wealth of middle 1/3 agents
The area ratio of the Lorenz curve to measure the
wealth inequality
Table 5 Experimental variables for each experiment case are listed. Dynamic calibration calibrates the dynamic parameter
for 100 iterations, with the given synthetic heterogeneous parameter. Heterogeneous calibration calibrates the heterogeneous
parameter for 100 iterations, with the given synthetic dynamic parameter. The calibration framework calibrates all the dynamic
and the heterogeneous parameters, with two subcases
Variable
Experiments
Dynamic Calibration
Heterogeneous Calibration
Calibration Framework
Ccal
Cdyn
Chet Kdyn Khet
A
T
R
100
100
200
200
1
0
2
20
0
1
3
30
3
3
3
3
2
2
2
2
100
100
100
100
50
50
50
50
10
10
10
10
I
3
1
3
3
5.1.2.2 Summary Statistics The wealth inequality is the main interest in this model, which leads the validation
data consisted of High Class Wealth Average, Middle Class Wealth Average, Low Class Wealth
Average, and Gini Index as in Tab. 4. Here, High Class Wealth Average is the average wealth of the top
33% agents, and Gini Index [14] is a index of measuring the wealth inequality. The synthetic validation data
is generated from the synthetic parameter setting. The agent-level simulation result Sagent is not used, since
the agent cluster is given a-priori to the calibration task in Wealth Distribution ABM to test the performance
of the suggested Bayesian optimization.
5.1.2.3 Experimental Cases There are three experimental cases: dynamic calibration, heterogeneous calibration,
and calibration framework of joining two calibrations into a single framework. Dynamic calibration, proposed in
Section 4.2, is evaluated with suggested parameter updating schemes: Sampling by Time, Sampling by Regime,
Mode Selection, and Random Search. Dynamic calibration utilizes the synthetic heterogeneous parameter, in
the calibration process. The heterogeneous parameter in dynamic calibration is fixed to the synthetic parameter
in Tab. 3. The heterogeneous calibration, proposed in Sectioin 4.3, applies the surrogate-based calibration with
known clustering assignments. In heterogeneous calibration, the dynamic parameter is fixed to be the synthetic
parameter, presented in Tab. 3. The last experiment calibrates both dynamic and heterogeneous parameters to
investigate the effects of the combination of two components. The two subcases are tested in Wealth Distribution
Model. The first subcase calibrates the dynamic and the heterogeneous parameters alternatvely with the pre-
determined iterations for dynamic calibration, Cdyn = 2, and for heterogeneous calibration, Chet = 3. The
second subcase evaluates with Cdyn = 20, and Chet = 30, to investigate the effects of the pre-determined
hyperparameters of the calibration framework. The model parameters are randomly initialized at the initial
iteration. Tab. 5 presents the list of the experimental variables in each experiment.
5.1.2.4 Performance Measure Since we know the synthetic parameters in Wealth Distribution Model, we cal-
culate the Mean Absolute Error (MAE) to measure the distance from the estimated dynamic parameter to the
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
17
Table 6 Performance evaluations in test case 1 are listed, with the following parameter updating rules: RS for Random Search,
ST for Sampling by Time, SR for Sampling by Regime, MS for Mode selection, BO for Bayesian optimization. Numbers are the
mean and the standard deviation of experiments by replicating 30 times. One tailed Welch's t-tests on the suggested parameter
updating rules, including ST, SR, MS and BO, are implemented with the baseline updating rule, RS.
Experiments
Synthetic
Parameter
Experiment
Random
Search
Parameter Up-
dating Rule
Dyn-
amic
Hetero-
geneous
True
True
RS
True
True
RS
RS
RS
ST
True
Dynamic Cali-
bration
SR
True
MS
True
Heterogeneous
Calibration
True
BO
Calibration
Frameworka
Calibration
Frameworkb
MS
BO
MS
BO
Test Case 1: Wealth Distribution ABM
High
Class
Wealth
Aver-
age
MAPE
Middle
Class
Wealth
Aver-
age
MAPE
Low
Class
Wealth
Aver-
age
MAPE
Gini
Index
MAPE
Total
MAPE
Parameter
Mean
Absolute
Error
Parameter
Euclidean
Error
0.012
(±0.009)
0.096
(±0.015)
0.052
(±0.041)
0.140
(±0.031)
0.072
(±0.008)
0.088
(±0.016)
0.080
(±0.011)
0.025
(±0.018)
0.124
(±0.017)
0.113
(±0.011)
0.016
(±0.011)
0.060
(±0.008)
0.031
(±0.022)
0.089
(±0.018)
0.045
(±0.006)
0.055
(±0.013)
0.048
(±0.007)
0.015
(±0.010)
0.085
(±0.020)
0.070
(±0.007)
0.020
(±0.013)
0.046
(±0.008)
0.035
(±0.027)
0.077
(±0.024)
0.034
(±0.004)
0.043
(±0.010)
0.039
(±0.007)
0.018
(±0.012)
0.074
(±0.031)
0.073
(±0.021)
0.015
(±0.009)
0.024
(±0.007)
0.028
(±0.027)
0.041
(±0.015)
0.018
(±0.004)
0.021
(±0.005)
0.021
(±0.005)
0.013
(±0.007)
0.044
(±0.021)
0.048
(±0.021)
0.016
(±0.007)
0.057
(±0.007)
0.036
(±0.018)
0.087
(±0.013)
0.042*
(±0.004)
0.052*
(±0.010)
0.047*
(±0.006)
0.018*
(±0.008)
0.082
(±0.021)
0.076*
(±0.010)
0
(±0)
0.549
(±0.037)
0
(±0)
0.623
(±0.060)
0.601
(±0.076)
0.537
(±0.104)
0.500
(±0.109)
0
(±0)
0.462
(±0.044)
0.501
(±0.089)
0
(±0)
0
(±0)
0.034
(±0.020)
0.339
(±0.238)
0
(±0)
0
(±0)
0
(±0)
0.021
(±0.011)
0.110
(±0.020)
0.092
(±0.015)
* P < 0.05
a Calibration framework with Cdyn = 2, and Chet = 3
b Calibration framework with Cdyn = 20, and Chet = 30
synthetic dynamic parameter. We also calculate the Euclidean error to measure the distance from the estimated
heterogeneous parameter to the synthetic heterogeneous parameter. To compare the simulation output, we use
Mean Absolute Percentage Error (MAPE) to measure the distance from the simulation result to the validation
data in all the experiments.
5.1.3 Dynamic Calibration Results
Fig. 2 presents the overall mechanism of dynamic calibration. Fig. 2 (a) presents a single dimension of summary
statistics, Gini Index, of a parameter candidate hypothesis. (b) presents the result of HMM, Ri
t on Alg. 2-Line
3, that divides the well-fitted regime, in green points, with the poorly-fitted regime, in red and blue points.
The joint likelihoods, Li
t on Alg. 2-Line 15, are presented in (c). The dotted lines are the thresholds in Eq. 18.
Exploration is executed in the most of the simulation timesteps for the first few iterations. Fig. 2 (d) presents
the regime detection results for all three candidate hypotheses, MRu on Alg. 2-Line 17. The estimated beta
distribution of a single merged regime, Beta(αu, βu) on Alg. 2-Line 8, is presented in (e). Fig. 2 (f) presents one
of the next candidate hypotheses, P i
dyn,C+1 on Alg. 1-Line 6, generated from the estimated beta distributions.
The current dynamic parameter candidates, P i
dyn,C , are plotted as the grey lines, and the generated dynamic
parameter evaluated in the next simulation is plotted as the black line, with different colors represent different
merged regimes.
The well-fitted regime in Fig. 2 (b) makes the parameter estimation at the front period success, as the
red colored dots in (f), because the second candidate hypothesis at the well-fitted regime has high likelihoods
that the next parameter will be generated near the second hypothesis. The simulation has failed to recover
the validation statistics afterwards, and this leads the algorithm to force explorations at the latter simulation
timesteps. Fig. 2 (g) is a simulation result, Sdyn on Alg. 2-Line 14, evaluated with the estimated parameter
in Fig. 2 (f). Iterative dynamic updates will change the simulation result to the synthetic generated validation
data, as in Fig. 2 (h).
18
Dongjun Kim et al.
(a)
(b)
(f)
Normalized
likelihoods Li
t
Merged
Regimes M Ru
(c)
(d)
(e)
Generate Next
Parameter P i
dyn,C+1
Run Simulation
M(P i
dyn,C+1 ∪ Phet; ω)
(g)
(h)
After
Iterations
Beta Distribution
for Regime Beta(αu, βu)
Fig. 2 Dynamic parameter calibration mechanism is plotted. (a)-(g) presents a single dynamic calibration iteration, and (h)
presents a final calibrated simulation result. (a) and (c) present calculating the simulation error, in Alg. 2-Line 12-15. (b) and
(d) present the regime detection process, in Alg. 2-Line 17. (e) and (f) present the parameter generation process, in Alg. 2-Line
18
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
19
(a)
(b)
Fig. 3 (a) The estimated optimal dynamic parameter, x marked line, is plotted. Dynamic calibration generates a dynamic
parameter regime-wisely to avoid overfitting in Mode Selection update rule. (b) The sampled optimal dynamic parameter,
o marked line, is plotted. Random Search finds an overfitted dynamic parameter, which only fits the given validation data,
without following the synthetic dynamic parameter trend
(a)
(b)
Fig. 4 (a) The calibrated dynamic parameter mean absolute errors are plotted. Parameter generation methods Sampling by
Regime and Mode Selection find parameters closer to the synthetic parameter than the other methods. (b) Dynamic calibration
simulation error is plotted by iterations. Parameter generation method Sampling by Time performs the best in terms of
simulation MAPE
Fig. 5 Dynamic calibration simulation errors with different initializations are plotted. The + marked line is the experimental
result of the random initialization and the x marked line is that of the selective initialization, with equally distributed initial
static parameters, having values 0.5, 1.0, and 1.5
20
(a)
(d)
Dongjun Kim et al.
(b)
(e)
(c)
(f)
Initial Result
After Iterations
(g)
(h)
Fig. 6 Heterogeneous parameter calibration mechanism is plotted. (a), (b), and (c) present the learning process of the response
surface as iterations increase, in Alg. 3-Line 28. (d), (e), and (f) are the contour plots of the response surface, with evaluation
path, in Alg. 3-Line 29. The search finds the global optimum as we expected. (g) and (h) present a single summary statistics,
Gini index, to compare before and after in calibration
Fig. 3 presents the best estimated parameters with two updating rules: Mode Selection and Random Search.
Mode Selection in (a) estimates the plausible parameter which shares the tendency with the synthetic parameter.
However, Random Search in (b) finds the highly fluctuating parameter, which is unable to interpret. Fig. 4 (a)
plots the parameter Mean Absolute Error (MAE) for the four updating rules. Mode Selection and Sampling
by Regime performs better than other updating rules in parameter error. However, Fig. 4 (b) shows that the
updating rule Sampling by Time performs the best in terms of simulation error.
Tab. 6 shows the experimental statistics of dynamic calibration. Sampling by Time performs the best in
simulation error, but it has the highest parameter error, which implicates that Sampling by Time estimates the
overfitted parameter. Sampling by Regime performance, on the other hand, is not significantly different from
that of Random Search. Lastly, Mode Selection is the best updating rule in reducing the parameter error.
Fig. 5 shows the effect of the parameter initialization. The plus marked line is the calibration result with the
randomly intialized dynamic parameters. Each parameter candidate has random static value at the initial cali-
bration step, see gray lines in Fig. 2 (f). The x marked line is the simulation result of the selective initialization,
where the parameter candidates are equally distributed in the parameter space, having values 0.5, 1.0, and 1.5.
The selective dynamic parameters are believed to be initialized well, since the synthetic dynamic parameter
is alternating between 0.5 and 1.5, see Fig. 3. Fig. 5 shows the selective initialization saturates faster than
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
21
(a)
(b)
Fig. 7 (a) The calibrated heterogeneous parameter Euclidean errors are plotted. (b) Hetergeneous calibration simulation
MAPE is plotted. Suggested calibration converges to the optimal lower bound, which is the simulation error with the synthetic
parameters as input parameters. The optimal lower bound is not 0 since the simulation is stochastic
the random initialization, but at the same time, the random initialization performs as good as the selective
initialization after iterations, due to the exploration in Eq. 18.
5.1.4 Heterogeneous Calibration Result
Fig. 6 (a), (b), and (c) illustrate the surrogate error functions of Gaussian process regression with 10, 50, and
100 evaluation points, respectively. The surrogate function is gradually sophisticated as the iteration increases.
Fig. 6 (d), (e), and (f) present the contour plot to illustrate the evaluation paths of the suggested Bayesian
optimization. The initial simulation result in Fig. 6 (g) is transformed to be close to the validation data as the
best simulation result in Fig. 6 (h).
Fig. 7 (a) presents the parameter Euclidean error of heterogeneous calibration. The parameter Euclidean
error of heterogeneous calibration saturates at 0.021, due to the observation error. In other words, the stochas-
ticity of the simulation makes the surrogate function biased, and this biasness hinders the Bayesian optimization
to estimate the exact synthetic parameter. However, in terms of the simulation error, heterogeneous calibration
converges to the optimal error as in Fig. 7 (b), where the optimal error is the simulation MAPE evaluated with
the synthetic parameter.
5.1.5 Calibration Framework Result
Tab. 6 demonstrates the calibration frameworkb, with Cdyn = 20, Chet = 30, outperforms the calibration
frameworka, in terms of the simulation error. Although the calibration frameworka has the least dynamic
parameter error among all the experiments in test case 1, the simulation error is large, due to the inaccuracy in
the estimation of the heterogeneous parameter. The calibration frameworkb, on the other hand, performs better
than the calibration frameworka in terms of the simulation error, since the calibration frameworkb estimates
closer heterogeneous parameter than the calibration frameworka. Fig. 8 presents the experimental result of the
calibration frameworkb.
5.2 Test Case 2: Real Estate Market ABM
5.2.1 Model Description
Real Estate Market Model is an agent-based model to predict the housing price and the number of house
transactions. The model consists of three types of agents: households, an external supplier and a realtor. Fig.
9 visualizes the interactions between agents. A household buy or sell houses to meet the residential needs. A
household registers the owned houses in List Process and decides to buy a listed house in Buy Processes.
In List Process, a household register its vacant houses to the market through a realtor agent. A house is
empty if the lease contract2 is over, and no other bargain is made. When no household lives in a vacant house,
2 Lease contract includes Jeonse and rental contract, where Jeonse is the lump-sum housing lease that is Korean specific
residential form of living 2 years by paying a large amount of deposits at the initial contract, which is returned at the end of
contract period, instead of paying rental fee monthly.
22
Dongjun Kim et al.
Fig. 8 The calibration frameworkb simulation MAPE is plotted. The calibration frameworkb first updates the dynamic pa-
rameter for 20 iterations, with fixed heterogeneous parameter, and next updates the heterogeneous parameter for 30 iterations,
with previously estimated optimal dynamic parameter
and if a house owner does not move to the house, then a household register its house in a market looking for a
leaseholder or a buyer. The external supplier lists its houses automatically in the housing market, since it acts
as a house supplier in the model.
In Buy Process, a household purchase houses in three steps: participation step, selection step, and contract
step. First, a household make a decision to participate in the market. A household without any house to
remain unconditionally engage in the housing market, looking for a new house to reside. A household with
a house for the residence participates with probability Market Participation Rate, to search for a new house
to invest. Second, a household in the market selects a house out of the registered houses, according to its
preferences. The preference of a household consists of area, contract type, and house type. A household first
decides between capital area and noncapital area for a new house. A household, then, decides the contract type
between lease and purchase, and the household decides the house type, among a house, an apartment and a
condo. Leaseholders, who have remaining contract period, choose always to purchase a house if they make a
deal with other house owners. Others, including 1) the house owners living in their house and 2) leaseholders
with contract termination, purchase or lease a house with probability Purchase Rate or (1 − Purchase Rate).
A realtor receives the participated agents' preferences, and the realtor replies the candidate houses to each
household with matching preferences. Last, in contract step, a household makes a deal on a house within its
budget from a replied house list. If other household contracts the house before, a household chooses other house
to contract. The heterogeneous parameter, Willing to Pay, controls the budget, which is the maximum ratio
out of the total asset an agent would raise, using savings and loan services.
Update Process updates all of required simulation state variables. A household saves the rest of its monthly
salary after excluding the expenses, such as consumption, taxes, rental fee, and loan repayment. Housing price
is decreased to
100(1 − Price Decrease Rate + Inflation Rate)%
(23)
up to ten times, if it is listed in the market, but no agents make a contract for the house at the timestep.
Housing price is increased to
100(1 + Price Increase Rate + Inflation Rate)%
(24)
if other houses with the same conditions are favored at the timestep. After updating state variables, an external
supplier supplies new houses in the market. The newly produced houses have the state variables sampled from
the real dataset.
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
23
Fig. 9 Real Estate Market Model flowchart of a single timestep is plotted. A single simulation timestep consists of the three
processes: List Process, Buy Process, and Update Process. Each process has sequential interactions between agents, as presented
in the flowchart
6LP(cid:3)(QJLQH(cid:20)(cid:29)(cid:3)LQLWLDWH/LVW3URFHVV$JHQW(cid:29)(cid:3)+RXVHKROG(cid:21)(cid:29)(cid:3)(cid:192)QG9DFDQF\+RXVH(cid:22)(cid:29)(cid:3)PRYLQJ'HFLVLRQ(cid:20)(cid:20)(cid:29)(cid:3)UHJLRQ'HFLVLRQ(cid:20)(cid:22)(cid:29)(cid:3)KRXVH7\SH'HFLVLRQ(cid:20)(cid:25)(cid:29)(cid:3)EX\'HFLVLRQ(cid:21)(cid:19)(cid:29)(cid:3)KRXVHKROG$WWULEXWH8SGDWH(cid:21)(cid:20)(cid:29)(cid:3)KRXVH$WWULEXWH8SGDWH$JHQW(cid:29)(cid:3)([WHUQDOVXSSOLHU(cid:21)(cid:24)(cid:29)(cid:3)KRXVH$WWULEXWH8SGDWH$JHQW(cid:29)(cid:3)5HDOWRU(cid:21)(cid:21)(cid:29)(cid:3)LQLWLDWH8SGDWH3URFHVVBIRU5HODWRU(cid:21)(cid:22)(cid:29)(cid:3)KRXVH$WWULEXWH8SGDWH(cid:24)(cid:29)(cid:3)VHQG/LVW5HTXHVWVHT(cid:3)/RRS(cid:3)IRU(cid:3)+RXVHKROG(cid:25)(cid:29)(cid:3)LQLWLDWH/LVW3URFHVVBIRU(6(cid:26)(cid:29)(cid:3)(cid:192)QG9DFDQF\+RXVH(cid:27)(cid:29)(cid:3)KRXVH/LVW'HFLVLRQ(cid:28)(cid:29)(cid:3)VHQG/LVW5HTXHVW(cid:20)(cid:19)(cid:29)(cid:3)LQLWLDWH%X\3URFHVV(cid:20)(cid:23)(cid:29)(cid:3)PVJ(cid:3)UHTXHVW+RXVH7\SH,QIR(cid:20)(cid:24)(cid:29)(cid:3)PVJ(cid:3)KRXVH,QIR(cid:20)(cid:26)(cid:29)(cid:3)FRQWUDFW+RXVHVHT(cid:3)/RRS(cid:3)IRU(cid:3)+RXVHKROG(cid:20)(cid:28)(cid:29)(cid:3)LQLWLDWH8SGDWH3URFHVVVHT(cid:3)/RRS(cid:3)IRU(cid:3)+RXVHKROG(cid:21)(cid:23)(cid:29)(cid:3)LQLWLDWH8SGDWH3URFHVVBIRU(6VHT(cid:3)/RRS(cid:3)IRU(cid:3)7LPHVWHSLQWHUDFWLRQ(cid:3)6LPXODWLRQ(cid:3)SURFHVV(cid:23)(cid:29)(cid:3)KRXVH/LVW'HFLVLRQ(cid:20)(cid:27)(cid:29)(cid:3)PVJ(cid:3)FRQWUDFW,QIR(cid:20)(cid:21)(cid:29)(cid:3)GHDO7\SH'HFLVLRQ 24
Dongjun Kim et al.
Table 7 Calibration parameters in the second test case are listed. The dynamic parameters and the heterogeneous parameters
are the key control parameters for the model dynamics. The dynamic parameters represents the seasonal effect, or business
cycle, and the heterogeneous parameters stand for the agent investment portfolios
Parameter
Parameter Type
Parameter Range
Market Participation Rate
Market Price Increase Rate
Market Price Decrease Rate
Dynamic
Dynamic
Dynamic
Willing to Pay
Purchase Rate
Heterogeneous
Heterogeneous
0-0.05
0-0.1
0-0.1
0.3-0.9
0.3-0.9
5.2.2 Virtual Experimental Design
5.2.2.1 Calibration Parameters Real Estate Market ABM has three dynamic parameters and two heterogeneous
parameters: Market Participation Rate, Market Price Increase Rate and Market Price Decrease Rate for the
dynamic parameters; and Willing to Pay and Purchase Rate for the heterogeneous parameters. The economic
trends directly influence the supply and the demand in the reality, and the dynamic parameters are selected
to reflect this seasonal bull and bear periods. A model uses an external supplier to reflect the supply, using a
real dataset, and the model adopts Market Participation Rate to represent the demand. The seasonal outcome
from the law of supply and demand is modeled by other dynamic parameters, Market Price Increase Rate and
Market Price Decrease Rate. The agent heterogeneity is embedded in the heterogeneous parameters, Willing to
Pay and Purchase Rate, of the investment portfolio. For instance, agents favor in real estate will be represented
to have high Purchase Ratio.
5.2.2.2 Summary Statistics The main interest in the model is predicting the housing price and the number
of housing transactions closer to the validation dataset. Government official data on the price index and the
transaction number is released in Korea Appraisal Board (KAB). We count the Jevons index [45] as the housing
price index, following the KAB index rule. We scale up the simulation transaction numbers by multipling the
ratio of the real population, 2 × 107, over the simulation population, 104, to adjust the simulation results
compatible with the validation transaction data.
KAB releases 24 types summary statistics, with each types are having the following different housing
characteristics: two house regions (Capital/Noncapital), three house types (Detached/Apartment/Multiplex),
two transaction types (Sales/Lease) and two summary statistics types (Index/Transaction number). We select
eight Apartment summary statistics, as listed in Tab. 8, because Apartment forms the 70% of the housing
transactions in Korea.
In agent clustering, total seven simulation agent-level state variables Sagent are considered. Agents living
in the capital area and the noncapital area are assigned separately in Living Region. Savings, Income and
Loan variables are normalized to have values from zero to one. The one-hot encoding is executed for the
variables House Type and Living Type, where the one-hot encoding is transforming the integer i into a
multi-dimensional zero vector but having one at the ith element.
5.2.2.3 Experimental Cases There are three experimental cases: dynamic calibration, heterogeneous calibration,
and the combined calibration framework. Dynamic calibration experiments three subcases: the first subcase
calibrates the first parameter, the next subcase calibrates the first and the second parameter, and the last sub-
case calibrates all the dynamic parameters. Dynamic calibration accepts the manually calibrated heterogeneous
parameter, in the calibration process. Heterogeneous calibration evaluates 15 subcases: first five subcases cali-
brate the first parameter, next five subcases calibrate the second parameter, and the last five subcases calibrate
all heterogeneous parameters. Each of five subcases tests the different number of clusters: the first subcase
calibrates the undivided parameters, the next four subcases calibrate the divided parameters, with first three
subcases for the parametric clustering, and the last subcase for the nonparametric clustering. Heterogeneous
parameter uses the manually human calibrated dynamic parameter, in the calibration process. The calibration
framwork, combining dynamic and heterogeneous calibrations, investigates the effects of the combination of
each calibration. The framework iteratively calibrates both the dynamic and the heterogeneous parameters, as
in Tab. 9.
5.2.3 Dynamic Calibration Result
The simulation price indices are more stable and better aligned with the validation dataset than the alignment
of the simulation transaction numbers. Therefore, most of the error improvement is achieved by aligning the
simulation transaction numbers to the validation, which heavily depends on the first dynamic parameter,
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
25
Table 8 The validation summary statistics and the agent-level summary statistics in the second test case are listed. The macro-
level validation dataset is collected from the Korea Appraisal Board. Apartment typed validation data, apartment price and
transaction numbers, are used, since the apartment forms 70% of the housing transactions in Korea. The agent-level summary
statistics are the average of the simulation replications
Name of Summary Statistics
Variable Description
Variable Value
Types of Sum-
mary Statistics
Validation-
level Summary
Statistics
Apartment Sales Price Index in Capital
Apartment Sales Price Index in Noncapital
Apartment Lease Price Index in Capital
Apartment Lease Price Index in Noncapital
Apartment Sales Transaction Number in Capital
Apartment Sales Transaction Number in Non-
capital
Apartment Lease Transaction Number in Capital
Apartment Lease Transaction Number in Non-
capital
Living Region
Jevons price index of
Apartment sales price
Jevons price index of
Apartment lease price.
Transaction
of Apartment sales.
numbers
Transaction
of Apartment lease.
numbers
living
between
Agent
gion
tal/noncapital area.
re-
capi-
Agent-level
Summary
Statistics
Savings
Income
Loan
House Type
Living Type
Total savings.
Sum of the labor in-
come and transfer in-
come.
Total amount of money
agent have borrowed
from bank.
Type of house where an
agent lives.
Type of living where an
agent lives
Housing Price is con-
verted into a percent-
age, with base value
as 100 at the initial
timestep.
Simulation transaction
number is scaled up
to be compatible with
the validation transac-
tion number.
1: Capital, 0: Noncapi-
tal
1 unit/1000 KRW
1 unit/1000 KRW
1 unit/1000 KRW
1: Detached House, 2:
Apartment,
3: Mul-
tiplex House, 4: No
House
1: Owner, 2: Lease, 3:
No House
Number of Own Houses
Number
agent owns
of
houses
1 unit/1 House
Table 9 Experimental variables of each experiment in the second test case are listed. Dynamic calibration calibrates the
dynamic parameters, with fixed human calibrated heterogeneous parameters. Heterogeneous calibration calibrates the hetero-
geneous parameters, with fixed human calibrated dynamic parameters. The calibration framework calibrates all the dynamic
and the heterogeneous parameters, with two subcases as in the Table
Variable
Experiments
Dynamic Calibration
Heterogeneous Calibration
Calibration Framework
Ccal
Cdyn
Chet Kdyn
Khet
A
T
R
100
100
200
200
1
0
2
0
1
3
20
30
3
3
3
3
1
10000
24
10
1,2,4,8,
nonparametric
10000
24
10
2, nonparametric
2, nonparametric
10000
10000
24
24
10
10
I
3
1
3
3
Market Participation Rate. Comparing with the human calibration, Tab. 10 shows a significant error reduction
in Total MAPE, which is gained from dynamic calibration experiments.
The second experimental subcase calibrates the first and the second dynamic parameters, where the second
dynamic parameter influences to the housing prices. Comparing with the human calibration, Tab. 10 demon-
strates that the analogous performance, with respect to the housing price indices, is achieved by counting the
second parameter into the calibration parameter.
The last experimental subcase, calibrating all three dynamic parameters, provides an insight when parame-
ters with similar effects, the second and the third parameters, are calibrated. The price indices are less correct
than that of the second subcase. Also, total MAPE is slightly worse than the second experiment, which is
mainly due to Apartment Lease Transaction Number in Noncaptial MAPE. The transaction numbers,
relatively irrelavant with the second and the third dynamic parameters, are disturbed by adding the third
26
E
P
A
M
l
a
t
o
T
-
s
n
a
r
T
e
s
a
e
L
-
s
n
a
r
T
s
e
l
a
S
t
n
e
m
t
r
a
p
A
t
n
e
m
t
r
a
p
A
n
i
r
e
b
m
u
N
n
o
i
t
c
a
n
i
r
e
b
m
u
N
n
o
i
t
c
a
l
a
t
i
p
a
c
n
o
N
l
a
t
i
p
a
c
n
o
N
E
P
A
M
E
P
A
M
E
P
A
M
E
P
A
M
E
P
A
M
E
P
A
M
-
s
n
a
r
T
e
s
a
e
L
-
m
u
N
n
o
i
t
c
a
l
a
t
i
p
a
C
n
i
r
e
b
t
n
e
m
t
r
a
p
A
-
s
n
a
r
T
s
e
l
a
S
-
m
u
N
n
o
i
t
c
a
l
a
t
i
p
a
C
n
i
r
e
b
t
n
e
m
t
r
a
p
A
n
i
e
c
i
r
P
e
s
a
e
L
x
e
d
n
I
n
i
e
c
i
r
P
s
e
l
a
S
x
e
d
n
I
l
a
t
i
p
a
c
n
o
N
l
a
t
i
p
a
c
n
o
N
t
n
e
m
t
r
a
p
A
t
n
e
m
t
r
a
p
A
e
c
i
r
P
e
s
a
e
L
-
i
p
a
C
n
i
x
e
d
n
I
t
n
e
m
t
r
a
p
A
E
P
A
M
l
a
t
e
c
i
r
P
s
e
l
a
S
-
i
p
a
C
n
i
x
e
d
n
I
t
n
e
m
t
r
a
p
A
E
P
A
M
l
a
t
M
B
A
t
e
k
r
a
M
e
t
a
t
s
E
l
a
e
R
:
2
e
s
a
C
t
s
e
T
s
r
e
t
s
u
C
l
f
o
r
e
b
m
u
N
-
a
P
n
o
i
t
a
r
b
i
l
a
C
s
r
e
t
e
m
a
r
s
t
n
e
m
i
r
e
p
x
E
h
c
a
e
n
i
r
o
r
r
e
t
s
e
l
l
a
m
s
e
h
t
s
e
t
a
c
i
d
n
i
e
c
a
f
d
l
o
b
e
h
T
.
s
e
m
i
t
0
3
g
n
i
t
a
c
i
l
p
e
r
y
b
s
t
n
e
m
i
r
e
p
x
e
f
o
n
o
i
t
a
i
v
e
d
d
r
a
d
n
a
t
s
e
h
t
d
n
a
n
a
e
m
e
h
t
e
r
a
s
r
e
b
m
u
N
.
d
e
t
s
i
l
e
r
a
2
e
s
a
c
t
s
e
t
n
i
s
l
e
d
o
m
n
o
i
t
a
r
b
i
l
a
c
d
e
s
o
p
o
r
p
e
h
t
f
o
s
n
o
i
t
a
u
l
a
v
e
e
c
n
a
m
r
o
f
r
e
P
0
1
e
l
b
a
T
.
s
t
l
u
s
e
r
n
o
i
t
a
r
b
i
l
a
c
n
a
m
u
h
e
n
i
l
e
s
a
b
e
h
t
h
t
i
w
d
e
t
n
e
m
e
l
p
m
i
e
r
a
s
e
i
g
o
l
o
d
o
h
t
e
m
n
o
i
t
a
r
b
i
l
a
c
d
e
t
s
e
g
g
u
s
e
h
t
n
o
s
t
s
e
t
-
t
s
'
h
c
l
e
W
d
e
l
i
a
t
e
n
O
.
e
s
a
c
l
a
t
n
e
m
i
r
e
p
x
e
Dongjun Kim et al.
)
5
0
0
.
0
±
4
1
2
.
0
(
)
8
0
0
.
0
±
4
0
2
.
0
(
*
7
6
1
.
0
)
3
0
0
.
0
±
(
*
8
6
1
.
0
)
3
0
0
.
0
±
(
*
4
7
1
.
0
)
4
0
0
.
0
±
(
*
0
0
2
.
0
)
4
0
0
.
0
±
(
*
9
0
1
.
0
)
3
0
0
.
0
±
(
*
4
1
1
.
0
)
2
0
0
.
0
±
(
*
7
0
1
.
0
)
3
1
0
.
0
±
(
*
5
2
1
.
0
)
0
2
0
.
0
±
(
*
4
0
2
.
0
)
4
0
0
.
0
±
(
*
8
7
1
.
0
)
9
0
0
.
0
±
(
*
3
8
1
.
0
)
3
2
0
.
0
±
(
*
5
4
1
.
0
)
3
1
0
.
0
±
(
*
5
2
1
.
0
)
1
2
0
.
0
±
(
*
6
9
1
.
0
)
6
0
0
.
0
±
(
*
6
0
1
.
0
)
5
1
0
.
0
±
(
*
5
1
1
.
0
)
7
1
0
.
0
±
(
*
4
1
1
.
0
)
3
1
0
.
0
±
(
*
7
0
1
.
0
)
0
3
0
.
0
±
(
*
3
2
1
.
0
)
6
1
0
.
0
±
(
*
5
3
1
.
0
)
9
1
0
.
0
±
(
*
5
0
1
.
0
)
2
0
0
.
0
±
(
*
0
3
1
.
0
)
1
0
0
.
0
±
(
)
7
1
0
.
0
±
3
6
4
.
0
(
)
7
2
1
.
0
±
7
9
3
.
0
(
)
8
1
0
.
0
±
9
0
4
.
0
(
)
8
1
0
.
0
±
0
1
4
.
0
(
)
9
2
0
.
0
±
2
5
4
.
0
(
)
6
6
0
.
0
±
9
6
4
.
0
(
)
0
2
0
.
0
±
0
5
1
.
0
(
)
5
1
0
.
0
±
1
6
1
.
0
(
)
1
5
0
.
0
±
8
0
2
.
0
(
)
6
4
1
.
0
±
7
7
1
.
0
(
)
7
5
0
.
0
±
0
6
3
.
0
(
)
4
4
0
.
0
±
6
5
2
.
0
(
)
7
6
0
.
0
±
5
5
2
.
0
(
)
0
8
0
.
0
±
5
0
2
.
0
(
)
8
5
0
.
0
±
1
1
2
.
0
(
)
8
9
0
.
0
±
1
6
4
.
0
(
)
4
5
0
.
0
±
9
7
1
.
0
(
)
7
7
0
.
0
±
1
8
1
.
0
(
)
0
8
0
.
0
±
0
3
2
.
0
(
)
6
7
1
.
0
±
6
8
1
.
0
(
)
8
4
0
.
0
±
2
0
2
.
0
(
)
6
1
1
.
0
±
4
6
2
.
0
(
)
0
0
0
.
0
±
6
3
1
.
0
(
)
3
8
0
.
0
±
5
6
2
.
0
(
)
6
1
0
.
0
±
4
8
2
.
0
(
)
5
0
1
.
0
±
8
3
3
.
0
(
)
6
1
0
.
0
±
3
4
3
.
0
(
)
3
2
0
.
0
±
7
3
3
.
0
(
)
7
1
0
.
0
±
5
3
3
.
0
(
)
6
5
0
.
0
±
9
7
2
.
0
(
)
1
1
0
.
0
±
6
6
1
.
0
(
)
0
2
0
.
0
±
2
0
2
.
0
(
)
7
5
0
.
0
±
8
7
1
.
0
(
)
4
6
0
.
0
±
7
5
3
.
0
(
)
1
2
0
.
0
±
9
4
2
.
0
(
)
0
2
0
.
0
±
4
8
1
.
0
(
)
8
3
0
.
0
±
1
0
2
.
0
(
)
1
4
0
.
0
±
5
8
1
.
0
(
)
8
3
0
.
0
±
2
9
1
.
0
(
)
2
7
0
.
0
±
4
2
4
.
0
(
)
8
2
0
.
0
±
0
0
2
.
0
(
)
9
3
0
.
0
±
8
6
1
.
0
(
)
1
4
0
.
0
±
8
1
2
.
0
(
)
5
6
0
.
0
±
8
9
1
.
0
(
)
5
5
0
.
0
±
2
3
2
.
0
(
)
3
6
0
.
0
±
5
1
2
.
0
(
)
9
4
0
.
0
±
2
5
2
.
0
(
)
4
0
0
.
0
±
5
9
1
.
0
(
)
2
1
0
.
0
±
9
5
1
.
0
(
)
5
8
0
.
0
±
1
9
2
.
0
(
)
2
1
0
.
0
±
7
7
2
.
0
(
)
8
1
0
.
0
±
0
8
2
.
0
(
)
8
1
0
.
0
±
2
7
2
.
0
(
)
8
1
0
.
0
±
2
4
1
.
0
(
)
0
1
0
.
0
±
5
5
2
.
0
(
)
5
0
0
.
0
±
4
4
2
.
0
(
)
4
4
0
.
0
±
4
6
1
.
0
(
)
9
2
1
.
0
±
4
6
1
.
0
(
)
3
3
0
.
0
±
8
1
2
.
0
(
)
9
2
0
.
0
±
4
0
1
.
0
(
)
4
6
1
.
0
±
7
3
1
.
0
(
)
3
1
1
.
0
±
4
6
1
.
0
(
)
5
8
0
.
0
±
4
4
1
.
0
(
)
8
7
0
.
0
±
5
5
2
.
0
(
)
8
2
0
.
0
±
7
2
1
.
0
(
)
6
0
1
.
0
±
1
4
2
.
0
(
)
3
1
1
.
0
±
0
8
1
.
0
(
)
1
5
1
.
0
±
9
9
1
.
0
(
)
7
6
0
.
0
±
8
7
1
.
0
(
)
3
3
0
.
0
±
3
8
1
.
0
(
)
8
1
0
.
0
±
5
9
1
.
0
(
)
1
3
0
.
0
±
7
9
1
.
0
(
)
7
2
0
.
0
±
5
6
7
.
0
(
)
1
8
1
.
0
±
4
5
5
.
0
(
)
5
1
0
.
0
±
6
6
2
.
0
(
)
4
3
0
.
0
±
3
7
2
.
0
(
)
7
2
0
.
0
±
1
8
2
.
0
(
)
0
8
0
.
0
±
7
6
6
.
0
(
)
2
1
0
.
0
±
9
5
2
.
0
(
)
0
3
0
.
0
±
6
6
2
.
0
(
)
9
9
0
.
0
±
3
6
2
.
0
(
)
1
8
1
.
0
±
5
6
2
.
0
(
)
9
3
0
.
0
±
4
6
7
.
0
(
)
8
3
0
.
0
±
5
3
8
.
0
(
)
9
9
0
.
0
±
7
2
8
.
0
(
)
7
9
0
.
0
±
7
6
5
.
0
(
)
2
0
1
.
0
±
4
1
4
.
0
(
)
6
2
1
.
0
±
8
9
3
.
0
(
)
2
9
0
.
0
±
3
0
3
.
0
(
)
4
5
0
.
0
±
9
8
2
.
0
(
)
7
9
0
.
0
±
3
4
2
.
0
(
)
8
9
1
.
0
±
2
3
2
.
0
(
)
3
7
0
.
0
±
6
8
2
.
0
(
)
8
0
1
.
0
±
4
2
3
.
0
(
)
7
0
0
.
0
±
9
1
2
.
0
(
)
4
8
0
.
0
±
0
2
3
.
0
(
)
0
0
0
.
0
±
3
0
0
.
0
(
)
4
0
0
.
0
±
8
0
0
.
0
(
)
1
0
0
.
0
±
5
0
0
.
0
(
)
2
0
0
.
0
±
6
0
0
.
0
(
)
4
0
0
.
0
±
2
1
0
.
0
(
)
0
0
0
.
0
±
3
0
0
.
0
(
)
2
0
0
.
0
±
6
0
0
.
0
(
)
1
0
0
.
0
±
5
0
0
.
0
(
)
3
0
0
.
0
±
6
0
0
.
0
(
)
0
0
0
.
0
±
5
0
0
.
0
(
)
0
0
0
.
0
±
3
0
0
.
0
(
)
0
0
0
.
0
±
3
0
0
.
0
(
)
1
0
0
.
0
±
3
0
0
.
0
(
)
8
0
0
.
0
±
3
0
0
.
0
(
)
0
0
0
.
0
±
3
0
0
.
0
(
)
2
0
0
.
0
±
4
0
0
.
0
(
)
0
0
0
.
0
±
3
0
0
.
0
(
)
1
0
0
.
0
±
4
0
0
.
0
(
)
8
0
0
.
0
±
5
0
0
.
0
(
)
0
0
0
.
0
±
3
0
0
.
0
(
)
3
1
0
.
0
±
1
2
0
.
0
(
)
4
1
0
.
0
±
8
1
0
.
0
(
)
1
0
0
.
0
±
4
0
0
.
0
(
)
5
0
0
.
0
±
1
1
0
.
0
(
)
0
0
0
.
0
±
7
0
0
.
0
(
)
5
0
0
.
0
±
9
0
0
.
0
(
)
0
0
0
.
0
±
3
0
0
.
0
(
)
3
0
0
.
0
±
5
0
0
.
0
(
)
3
0
0
.
0
±
8
0
0
.
0
(
)
1
0
0
.
0
±
7
0
0
.
0
(
)
1
0
0
.
0
±
4
0
0
.
0
(
)
1
0
0
.
0
±
4
0
0
.
0
(
)
3
0
0
.
0
±
4
0
0
.
0
(
)
0
0
0
.
0
±
4
0
0
.
0
(
)
0
0
0
.
0
±
7
0
0
.
0
(
)
0
0
0
.
0
±
7
0
0
.
0
(
)
1
0
0
.
0
±
7
0
0
.
0
(
)
8
0
0
.
0
±
6
0
0
.
0
(
)
0
0
0
.
0
±
6
0
0
.
0
(
)
1
0
0
.
0
±
4
0
0
.
0
(
)
0
0
0
.
0
±
6
0
0
.
0
(
)
1
0
0
.
0
±
5
0
0
.
0
(
)
8
0
0
.
0
±
4
0
0
.
0
(
)
0
0
0
.
0
±
7
0
0
.
0
(
)
9
0
0
.
0
±
0
2
0
.
0
(
)
5
1
0
.
0
±
9
1
0
.
0
(
)
3
0
0
.
0
±
5
0
0
.
0
(
)
6
0
0
.
0
±
9
0
0
.
0
(
)
1
0
0
.
0
±
0
2
0
.
0
(
)
6
0
0
.
0
±
1
1
0
.
0
(
)
2
0
0
.
0
±
5
2
0
.
0
(
)
8
0
0
.
0
±
4
2
0
.
0
(
)
0
1
0
.
0
±
0
2
0
.
0
(
)
1
0
0
.
0
±
0
2
0
.
0
(
)
4
0
0
.
0
±
7
2
0
.
0
(
)
2
0
0
.
0
±
6
2
0
.
0
(
)
7
0
0
.
0
±
9
2
0
.
0
(
)
0
0
0
.
0
±
6
2
0
.
0
(
)
0
0
0
.
0
±
0
2
0
.
0
(
)
0
0
0
.
0
±
9
1
0
.
0
(
)
2
0
0
.
0
±
0
2
0
.
0
(
)
5
1
0
.
0
±
1
2
0
.
0
(
)
0
0
0
.
0
±
0
2
0
.
0
(
)
3
0
0
.
0
±
2
1
0
.
0
(
)
1
0
0
.
0
±
1
2
0
.
0
(
)
2
0
0
.
0
±
3
2
0
.
0
(
)
5
1
0
.
0
±
5
2
0
.
0
(
)
1
0
0
.
0
±
1
2
0
.
0
(
)
5
1
0
.
0
±
6
1
0
.
0
(
)
6
1
0
.
0
±
2
2
0
.
0
(
)
1
0
0
.
0
±
6
1
0
.
0
(
)
2
1
0
.
0
±
1
2
0
.
0
(
2
1
0
.
0
)
1
0
0
.
0
±
(
5
2
0
.
0
)
0
1
0
.
0
±
(
)
1
0
0
.
0
±
5
0
0
.
0
(
9
0
0
.
0
)
5
0
0
.
0
±
(
5
1
0
.
0
)
9
0
0
.
0
±
(
--
--
--
--
--
s
r
e
t
e
m
a
r
a
P
l
l
A
-
i
l
a
C
n
a
m
u
H
n
o
i
t
a
r
b
m
o
d
n
a
R
h
c
r
a
e
S
s
r
e
t
e
m
a
r
a
P
l
l
A
-
e
m
a
r
a
P
t
s
r
i
F
d
n
a
t
s
r
i
F
r
e
t
-
m
a
r
a
P
d
n
o
c
e
S
c
i
m
a
n
y
D
s
r
e
t
e
l
l
A
s
r
e
t
e
m
a
r
a
P
-
i
l
a
C
c
i
m
a
n
y
D
n
o
i
t
a
r
b
2
1
0
.
0
)
1
0
0
.
0
±
(
6
0
0
.
0
)
2
0
0
.
0
±
(
6
0
0
.
0
)
2
0
0
.
0
±
(
)
3
0
0
.
0
±
4
0
0
.
0
(
5
0
0
.
0
)
0
0
0
.
0
±
(
2
1
0
.
0
)
1
0
0
.
0
±
(
3
1
0
.
0
)
1
0
0
.
0
±
(
3
1
0
.
0
)
3
0
0
.
0
±
(
)
6
0
0
.
0
±
1
1
0
.
0
(
2
1
0
.
0
)
1
0
0
.
0
±
(
8
0
0
.
0
)
2
0
0
.
0
±
(
1
1
0
.
0
)
1
0
0
.
0
±
(
9
0
0
.
0
)
3
0
0
.
0
±
(
)
6
0
0
.
0
±
7
0
0
.
0
(
1
1
0
.
0
)
2
0
0
.
0
±
(
8
2
0
.
0
)
2
1
0
.
0
±
(
1
3
0
.
0
)
8
1
0
.
0
±
(
)
1
0
0
.
0
±
6
1
0
.
0
(
4
2
0
.
0
)
3
2
0
.
0
±
(
3
.
5
1
)
4
.
1
±
(
c
i
r
t
e
m
a
r
a
p
n
o
N
1
2
4
8
d
e
d
i
v
i
d
n
U
c
i
r
t
e
m
a
r
a
P
-
e
m
a
r
a
P
t
s
r
i
F
r
e
t
1
2
4
8
d
e
d
i
v
i
d
n
U
c
i
r
t
e
m
a
r
a
P
-
m
a
r
a
P
d
n
o
c
e
S
s
u
o
e
n
e
g
o
r
e
t
e
H
r
e
t
e
n
o
i
t
a
r
b
i
l
a
C
9
.
5
1
)
2
.
1
±
(
c
i
r
t
e
m
a
r
a
p
n
o
N
1
2
4
8
d
e
d
i
v
i
d
n
U
c
i
r
t
e
m
a
r
a
P
-
t
e
H
l
l
A
s
u
o
e
n
e
g
o
r
e
s
r
e
t
e
m
a
r
a
P
5
.
5
1
)
2
.
1
±
(
c
i
r
t
e
m
a
r
a
p
n
o
N
6
.
5
1
)
3
.
1
±
(
c
i
r
t
e
m
a
r
a
p
n
o
N
2
c
i
r
t
e
m
a
r
a
P
4
.
5
1
)
1
.
1
±
(
c
i
r
t
e
m
a
r
a
p
n
o
N
2
c
i
r
t
e
m
a
r
a
P
-
a
P
l
e
d
o
M
l
l
A
s
r
e
t
e
m
a
r
a
k
r
o
w
e
m
a
r
F
n
o
i
t
a
r
b
i
l
a
C
-
a
P
l
e
d
o
M
l
l
A
s
r
e
t
e
m
a
r
b
k
r
o
w
e
m
a
r
F
n
o
i
t
a
r
b
i
l
a
C
0
3
=
t
e
h
C
d
n
a
,
0
2
=
n
y
d
C
h
t
i
w
t
n
e
m
i
r
e
p
x
E
b
3
=
t
e
h
C
d
n
a
,
2
=
n
y
d
C
h
t
i
w
t
n
e
m
i
r
e
p
x
E
a
5
0
.
0
<
P
*
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
27
(a)
(b)
Fig. 10 The clustered agent-level simulation results are plotted in the radar charts. The first cluster is the group of leaseholders
with middle-class salary, no debt, and low savings, who live in apartment. The third cluster is the group of apartment-owners
in capital area with high-class salary, high debt, and low savings.
Fig. 11 The calibration frameworkb simulation MAPE is plotted in the x marked line. The calibration frameworkb first
calibrates all three dynamic parameters for 20 iterations, with fixed heterogeneous parameters, and next calibrates all two
heterogeneous parameters for 30 iterations, with perviously calibrated optimal dynamic parameters. The random search reaches
to the human calibration error after 80 iterations, and saturates on the afterward iterations. The calibration framework achieves
the human calibration error in short iterations, and saturates to nearly half of the human calibration error
parameter to calibrate, since the problem complexity is increased, without attaining any supplementary power
to control the model output.
5.2.4 Heterogeneous Calibration Result
Comparing with the human calibration, no significant improvement is attained by calibrating the heterogeneous
parameters without agent clustering. However, when the heterogeneous parameters are differentiated with the
agent sub-populations, errors are considerably improved. This is because one of the error source, agent hetero-
geneity, is mediated through differentiating parameter values by clusters, which generates a micro controller
for the elaboration of the agent behaviors.
28
(a)
Dongjun Kim et al.
(b)
(c)
(d)
Fig. 12 Dynamic calibration (triangle marked), heterogeneous calibration (square marked), the calibration frameworkb (x
marked) experiments are illustrated, with the human calibration result (circle marked) for comparison. (a) and (b) are the
apartment transaction numbers of the capital area for sales and lease, respectively. (c) and (d) are the apartment transaction
numbers of the noncapital area for sales and lease, respectively. All calibration methodologies outperforms the human calibration
result in any of (a)-(d). Similarly, heterogeneous calibration outperforms dynamic calibration in all (a)-(d). The calibration
frameworkb outperforms heterogeneous calibration in (a) and (d)
Fig. 10 illustrates the detailed characteristics of agent clusters with radar charts. For example, the first
cluster is the group of the leaseholder agents with middle-class salary, no debt, and low savings, who live in the
apartments. As another example cluster, the third cluster is consisted of the house owners with high salary,
high debt, and low savings, who live in the apartments located at the capital area.
When we calibrate all the heterogeneous parameters with agent clusters in Fig. 10, the first and the second
calibrated heterogeneous parameters have the same values in the first cluster, 0.9, which indicates the agents
are trying to buy their houses sooner or later. In the third cluster, Willing to Pay and Purchase Rate have
their values 0.3 and 0.9, respectively. Since the third gouped agents have their own house with high debts, they
are unlikely to have high-risk investment propensity. The low Willing to Pay indeed confirms their low-risk
tendency, and high Purchase Rate reflects the fact that they are already living in their own house.
Calibrating the first parameter enhances the simulation performance, since the first parameter, Willing
to Pay, directly controls the overall up and down of the transaction numbers. The error improvement from
the second parameter is not as significant as from the first parameter. The second parameter, Purchase Rate,
is only applicable for the minor agents whose lease contracts are finished. The simulation error gradually
decreases as the number of the agent clusters increases, leading the nonparamtric clustering result performs
the best. Although the two heterogeneous parameters plays different roles, there are considerable overlaps in
their ramification on the simulation result. Thus, the calibration result with all the heterogeneous parameters
reports the negligible error improvement, compare to the first parameter calibration.
5.2.5 Calibration Framework Result
Tab. 10 shows that the calibration framework outperforms both dynamic calibration alone and heterogeneous
calibration alone. Also, the calibration frameworkb with two clusters is remarkably stable in its error improve-
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
29
ment. However, despite of the lowest error, calibrating all the parameters are not as effective as we expected,
since the first heterogeneous parameter calibration with eight clusters reports the analogous error, 0.107, with
the calibration frameworkb error, 0.105. The reasons are two-folds. First, the price indices are relatively insensi-
tive to the change of the parameter values, which leads that the error reduction relies mostly on the improvement
in the transaction numbers. Second, the combination of the controller parameters on the transaction numbers
hinders the response variables to be close to the validation, due to the increased dimensionality.
Fig. 11 presents the simulation errors of the calibration frameoworkb by calibration iterations. First, nearly
80 iterations for the random search are required to reach the human calibration result, and no significant
improvement is made on the afterward iterations. Second, the calibration frameworkb defeats the human
calibration within 20 iterations, in average. After superseding the human calibration, the calibration frameworkb
saturates fast.
Fig. 12 illustrates the calibration results of dynamic calibration, heterogeneous calibration, and the calibra-
tion frameworkb. All the calibration results fit to the validation data better than the human calibration result.
Heterogeneous calibration generates better fitted simulation results than dynamic calibration, in all transaction
numbers. The calibration frameworkb fits better than heterogeneous calibration in (a) and (d).
6 Conclusion
This paper proposes an automatic data-driven calibration framework for the dynamic and the heterogeneous
parameters, by extracting the hidden structures from the simulation outputs. Dynamic calibration controls the
dynamically switching parameters, by extracting the hidden dynamic regimes, and by treating each regimes
separately. Heterogeneous calibration calibrates the agent cluster-wise heterogeneous parameters, by extracting
the hidden agent-subpopulations. The experimental results on dynamic calibration and heterogeneous calibra-
tion demonstrate that the proposed separate calibrations reduce simulation error, with plausible estimated
parameters. The calibration framework, the combination of dynamic calibration and heterogeneous calibration,
exhibits the expected results in both test cases with both Keep-It-Simple-and-Stupid (KISS) model as well as
elaborated and complex model of housing markets.
References
1. Beaumont MA (2010) Approximate bayesian computation in evolution and ecology. Annual review of
ecology, evolution, and systematics 41:379 -- 406
2. Beisbart C, Saam NJ (2018) Computer Simulation Validation: Fundamental Concepts, Methodological
Frameworks, and Philosophical Perspectives. Springer
3. Bishop CM (2006) Pattern recognition and machine learning. springer
4. Bonabeau E (2002) Agent-based modeling: Methods and techniques for simulating human systems. Pro-
ceedings of the national academy of sciences 99(suppl 3):7280 -- 7287
5. Brenner T, Werker C (2006) A practical guide to inference in simulation models. Tech. rep., Papers on
economics and evolution
6. Brochu E, Cora VM, De Freitas N (2010) A tutorial on bayesian optimization of expensive cost func-
tions, with application to active user modeling and hierarchical reinforcement learning. arXiv preprint
arXiv:10122599
7. Bull AD (2011) Convergence rates of efficient global optimization algorithms. Journal of Machine Learning
Research 12(Oct):2879 -- 2904
8. Calvez B, Hutzler G (2005) Automatic tuning of agent-based models using genetic algorithms. In: Inter-
national Workshop on Multi-Agent Systems and Agent-Based Simulation, Springer, pp 41 -- 57
9. Dasgupta S (1999) Learning mixtures of gaussians. In: 40th Annual Symposium on Foundations of Computer
Science (Cat. No. 99CB37039), IEEE, pp 634 -- 644
10. Epstein JM (2006) Generative social science: Studies in agent-based computational modeling. Princeton
University Press
11. Fox EB, Sudderth EB, Jordan MI, Willsky AS (2008) An hdp-hmm for systems with state persistence. In:
Proceedings of the 25th international conference on Machine learning, ACM, pp 312 -- 319
12. Frazier PI (2018) A tutorial on bayesian optimization. arXiv preprint arXiv:180702811
13. Fu MC (2002) Optimization for simulation: Theory vs. practice. INFORMS Journal on Computing
14(3):192 -- 215
14. Gini C (1936) On the measure of concentration with special reference to income and statistics. Colorado
College Publication, General Series 208:73 -- 79
15. Grunewalder S, Audibert JY, Opper M, Shawe-Taylor J (2010) Regret bounds for gaussian process bandit
problems. In: Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statis-
tics, pp 273 -- 280
30
Dongjun Kim et al.
16. Hamilton JD (2016) Regime switching models. The new palgrave dictionary of economics pp 1 -- 7
17. Hara S, Kita H, Ikeda K, Susukita M (2013) Configuring agents attributes with simulated annealing. In:
Agent-Based Approaches in Economic and Social Complex Systems VII, Springer, pp 45 -- 59
18. Hartig F, Calabrese JM, Reineking B, Wiegand T, Huth A (2011) Statistical inference for stochastic
simulation models -- theory and application. Ecology letters 14(8):816 -- 827
19. Heppenstall AJ, Evans AJ, Birkin MH (2007) Genetic algorithm optimisation of an agent-based model for
simulating a retail market. Environment and Planning B: Planning and Design 34(6):1051 -- 1070
20. Ho YC, Hu JQ (1990) An infinitesimal perturbation analysis algorithm for a multiclass g/g/1 queue.
Operations Research Letters 9(1):35 -- 44
21. Hoffman M, Brochu E, de Freitas N (2011) Portfolio allocation for bayesian optimization. In: Proceedings
of the Twenty-Seventh Conference on Uncertainty in Artificial Intelligence, AUAI Press, pp 327 -- 336
22. Hosseinali F, Alesheikh AA, Nourian F (2013) Agent-based modeling of urban land-use development, case
study: Simulating future scenarios of qazvin city. Cities 31:105 -- 113
23. Jabot F, Faure T, Dumoulin N (2013) Easy abc: performing efficient approximate b ayesian computation
sampling schemes using r. Methods in Ecology and Evolution 4(7):684 -- 687
24. Kennedy J (2010) Particle swarm optimization. Encyclopedia of machine learning pp 760 -- 766
25. Kingma DP, Welling M (2013) Auto-encoding variational bayes. arXiv preprint arXiv:13126114
26. Kirkpatrick S, Gelatt CD, Vecchi MP (1983) Optimization by simulated annealing. science 220(4598):671 --
680
27. Lamperti F, Roventini A, Sani A (2018) Agent-based model calibration using machine learning surrogates.
Journal of Economic Dynamics and Control 90:366 -- 389
28. Lee JS, Filatova T, Ligmann-Zielinska A, Hassani-Mahmooei B, Stonedahl F, Lorscheid I, Voinov A, Pol-
hill JG, Sun Z, Parker DC (2015) The complexities of agent-based modeling output analysis. Journal of
Artificial Societies and Social Simulation 18(4):4
29. Lizotte DJ (2008) Practical bayesian optimization. University of Alberta
30. Marjoram P, Molitor J, Plagnol V, Tavar´e S (2003) Markov chain monte carlo without likelihoods. Pro-
ceedings of the National Academy of Sciences 100(26):15324 -- 15328
31. Marks RE (2013) Validation and model selection: Three similarity measures compared. Complexity Eco-
nomics 2(1):41 -- 61
32. Mockus J (2012) Bayesian approach to global optimization: theory and applications, vol 37. Springer Science
& Business Media
33. Moon IC, Kim D, Yun TS, Bae JW, Kang DO, Paik E (2018) Data-driven automatic calibration for
validation of agent-based social simulations. In: 2018 IEEE International Conference on Systems, Man,
and Cybernetics (SMC), IEEE, pp 1605 -- 1610
34. Naiem A, Reda M, El-Beltagy M, El-Khodary I (2010) An agent based approach for modeling traffic flow.
In: 2010 The 7th International Conference on Informatics and Systems (INFOS), IEEE, pp 1 -- 6
35. Nannen V, Eiben AE (2006) A method for parameter calibration and relevance estimation in evolutionary
algorithms. In: Proceedings of the 8th annual conference on Genetic and evolutionary computation, ACM,
pp 183 -- 190
36. Nguyen AT, Reiter S, Rigo P (2014) A review on simulation-based optimization methods applied to building
performance analysis. Applied Energy 113:1043 -- 1058
37. Noel MM, Jannett TC (2004) Simulation of a new hybrid particle swarm optimization algorithm. In:
Thirty-Sixth Southeastern Symposium on System Theory, 2004. Proceedings of the, IEEE, pp 150 -- 153
38. Preuss R, von Toussaint U (2018) Global optimization employing gaussian process-based bayesian surro-
gates. Entropy 20(3):201
39. Rahmandad H, Sterman J (2008) Heterogeneity and network structure in the dynamics of diffusion: Com-
paring agent-based and differential equation models. Management Science 54(5):998 -- 1014
40. Rand W (2012) When does simulated data match real data? comparing model calibration functions using
genetic algorithms. In: 4th World Congress on Social Simulation, 2012
41. Rasmussen C, Williams C (2006) GaussianProcessesforMachineLearning. AdaptiveComputation and Ma-
chine Learning
42. Ryzhov IO (2016) On the convergence rates of expected improvement methods. Operations Research
64(6):1515 -- 1528
43. Sasena MJ, Papalambros P, Goovaerts P (2002) Exploration of metamodeling sampling criteria for con-
strained global optimization. Engineering optimization 34(3):263 -- 278
44. Schonlau M (1997) Computer experiments and global optimization. PhD thesis, University of Waterloo,
Waterloo, Ont., Canada, Canada, aAINQ22234
45. Silver M, Heravi S (2007) Why elementary price index number formulas differ: Evidence on price dispersion.
Journal of Econometrics 140(2):874 -- 883
46. Sisson SA, Fan Y, Tanaka MM (2007) Sequential monte carlo without likelihoods. Proceedings of the
National Academy of Sciences 104(6):1760 -- 1765
Automatic Calibration of Dynamic and Heterogeneous Parameters in Agent-based Model
31
47. Snoek J, Larochelle H, Adams RP (2012) Practical bayesian optimization of machine learning algorithms.
In: Advances in neural information processing systems, pp 2951 -- 2959
48. S´obester A, Leary SJ, Keane AJ (2005) On the design of optimization strategies based on global response
surface approximation models. Journal of Global Optimization 33(1):31 -- 59
49. Stonedahl F, Wilensky U (2010) Evolutionary robustness checking in the artificial anasazi model. In: 2010
AAAI Fall Symposium Series
50. Stonedahl F, Wilensky U (2010) Finding forms of flocking: Evolutionary search in abm parameter-spaces.
In: International Workshop on Multi-Agent Systems and Agent-Based Simulation, Springer, pp 61 -- 75
51. Suri R (1987) Infinitesimal perturbation analysis for general discrete event systems. Journal of the ACM
(JACM) 34(3):686 -- 717
52. Tavar´e S, Balding DJ, Griffiths RC, Donnelly P (1997) Inferring coalescence times from dna sequence data.
Genetics 145(2):505 -- 518
53. Teh YW (2010) Dirichlet process. Encyclopedia of machine learning pp 280 -- 287
54. Thiele JC, Kurth W, Grimm V (2014) Facilitating parameter estimation and sensitivity analysis of agent-
based models: A cookbook using netlogo and r. Journal of Artificial Societies and Social Simulation 17(3):11
55. Tresidder E, Zhang Y, Forrester A (2012) Acceleration of building design optimisation through the use of
kriging surrogate models. Proceedings of building simulation and optimization pp 1 -- 8
56. Vazquez E, Bect J (2010) Convergence properties of the expected improvement algorithm with fixed mean
and covariance functions. Journal of Statistical Planning and inference 140(11):3088 -- 3095
57. Wang J, Clark SC, Liu E, Frazier PI (2016) Parallel bayesian global optimization of expensive functions.
arXiv preprint arXiv:160205149
58. Whitley D (1994) A genetic algorithm tutorial. Statistics and computing 4(2):65 -- 85
59. Wilensky U (1998) Netlogo wealth distribution model. Center for Connected Learning and Computer-
Based Modeling, Northwestern University, Evanston, IL[En l´ınea] Disponible en: http://ccl northwestern
edu/NetLogo/models/WealthDistribution
60. Windrum P, Fagiolo G, Moneta A (2007) Empirical validation of agent-based models: Alternatives and
prospects. Journal of Artificial Societies and Social Simulation 10(2):8
61. Wu J, Frazier P (2016) The parallel knowledge gradient method for batch bayesian optimization. In:
Advances in Neural Information Processing Systems, pp 3126 -- 3134
62. Wu J, Poloczek M, Wilson AG, Frazier P (2017) Bayesian optimization with gradients. In: Advances in
Neural Information Processing Systems, pp 5267 -- 5278
63. Yang C, Kurahashi S, Kurahashi K, Ono I, Terano T (2009) Agent-based simulation on women's role
in a family line on civil service examination in chinese history. Journal of Artificial Societies and Social
Simulation 12(2):5
|
1410.4754 | 2 | 1410 | 2016-01-14T21:55:40 | Parallel and Distributed Methods for Nonconvex Optimization-Part I: Theory | [
"cs.MA",
"math.OC"
] | In this two-part paper, we propose a general algorithmic framework for the minimization of a nonconvex smooth function subject to nonconvex smooth constraints. The algorithm solves a sequence of (separable) strongly convex problems and mantains feasibility at each iteration. Convergence to a stationary solution of the original nonconvex optimization is established. Our framework is very general and flexible; it unifies several existing Successive Convex Approximation (SCA)-based algorithms such as (proximal) gradient or Newton type methods, block coordinate (parallel) descent schemes, difference of convex functions methods, and improves on their convergence properties. More importantly, and differently from current SCA approaches, it naturally leads to distributed and parallelizable implementations for a large class of nonconvex problems.
This Part I is devoted to the description of the framework in its generality. In Part II we customize our general methods to several multi-agent optimization problems, mainly in communications and networking; the result is a new class of (distributed) algorithms that compare favorably to existing ad-hoc (centralized) schemes (when they exist). | cs.MA | cs |
Parallel and Distributed Methods for Nonconvex
Optimization−Part I: Theory
Gesualdo Scutari, Francisco Facchinei, Lorenzo Lampariello, and Peiran Song
1
Abstract -- In this two-part paper, we propose a general algo-
rithmic framework for the minimization of a nonconvex smooth
function subject to nonconvex smooth constraints. The algorithm
solves a sequence of (separable) strongly convex problems and
mantains feasibility at each iteration. Convergence to a stationary
solution of the original nonconvex optimization is established.
Our framework is very general and flexible and unifies several ex-
isting Successive Convex Approximation (SCA)-based algorithms
More importantly, and differently from current SCA approaches,
it naturally leads to distributed and parallelizable implementations
for a large class of nonconvex problems.
This Part I is devoted to the description of the framework in its
generality. In Part II we customize our general methods to several
multi-agent optimization problems, mainly in communications
and networking; the result is a new class of centralized and
distributed algorithms that compare favorably to existing ad-hoc
(centralized) schemes.
I. INTRODUCTION
T HE minimization of a nonconvex (smooth) objective
function U : K → R subject to convex constraints K
and nonconvex ones gj(x) ≤ 0, with gj : K → R smooth,
U (x)
min
x
s.t.
gj(x) ≤ 0, j = 1, . . . , m
x ∈ K
) , X ,
(P)
is an ubiquitous problem that arises in many fields, ranging
from signal processing to communication, networking, ma-
chine learning, etc.
It
is hardly possible here to even summarize the huge
amount of solution methods that have been proposed for prob-
lem (P). Our focus in this paper is on distributed algorithms
converging to stationary solutions of P while preserving the
feasibility of the iterates. While the former feature needs no
further comment, the latter is motivated by several reasons.
First, in many cases the objective function U is not even
defined outside the feasible set; second, in some applications
one may have to interrupt calculations before a solution has
been reached and it is then important that the current iterate is
feasible; and third, in on-line implementations it is mandatory
that some constraints are satisfied by every iterate (e.g., think
"La Sapienza", Rome,
Italy;
Scutari
is with the School of
email: [email protected]. Facchinei
(Discovery Park), Purdue University, West-Lafayette,
Industrial Engineering and the Cy-
IN,
ber Center
and Lampariello
USA;
and Management Engi-
are with the Dept. of Computer, Control,
neering, University of Rome
emails:
<facchinei, lampariello>@diag.uniroma1.it. Song is with
the School of Information and Communication Engineering, Beijing In-
formation Science and Technology University, Beijing, China; email:
[email protected].
The work of Scutari was supported by the USA National Science Foun-
dation under Grants CMS 1218717, CCF 1564044, and CAREER Award
1254739. The work of Facchinei was partially supported by the MIUR project
PLATINO, Grant n. PON01_01007. Part of this work appeared in [1], [2] and
on arxiv [3] on Oct. 2014.
of power budget or interference constraints). As far as we are
aware of, there exists no method for the solution of P in its
full generality that is both feasible and distributed.
Existing efforts pursuing the above design criteria include:
1) Feasible Interior Point (FIP) methods (e.g., [4], [5]), 2)
Feasible Sequential Quadratic Programming (FSQP) methods
(e.g., [6]); 3) Parallel Variable Distribution (PVD) schemes
(e.g., [7] -- [9]); 4) SCA algorithms (in the spirit of [10] --
[15]); and some specialized algorithms with roots in the
structural optimization field (e.g., [16] -- [18]). FIP and FSQP
methods maintain feasibility throughout the iterations but are
centralized and computationally expensive. PVD schemes are
suitable for implementation over parallel architectures but they
require an amount of information exchange/knowledge that
is often not compatible with a distributed architecture (for
example they cannot be applied to the case study discussed
in Part II of the paper [19]). Furthermore, when applied to
problem P, they call for the solution of possibly difficult
nonconvex (smaller) subproblems; and convergence has been
established only for convex [7], [9] or nonconvex but block
separable gjs [8]. Standard SCA methods are centralized [10],
[11], [15], with the exception of [13], [14] and some instances
of [12] that lead instead to distributed schemes. However,
convergence conditions have been established only in the case
of strongly convex U [11] or convex and separable gjs [12] --
[14]. Finally, methods developed in the structural engineering
field, including [16] -- [18], share some similarities with our
approach, but in most cases they lack reliable mathematical
foundations or do not prove convergence to stationary points
of the original problem P. We refer to Sec. III-B for a more
detailed discussion on existing works.
In this paper we propose a new framework for the general
formulation P which, on one hand, maintains feasibility
and, on the other hand, leads, under very mild additional
assumptions, to parallel and distributed solution methods. The
essential, natural idea underlying the proposed approach is to
compute a solution of P by solving a sequence of (simpler)
strongly convex subproblems whereby the nonconvex objec-
tive function and constraints are replaced by suitable convex
approximations; the subproblems can be then solved (under
some mild assumptions) in a distributed fashion using standard
primal/dual decomposition techniques (e.g., [20], [21]). Addi-
tional key features of the proposed method are: i) it includes as
special cases several classical SCA-based algorithms, such as
(proximal) gradient or Newton type methods, block coordinate
(parallel) descent schemes, Difference of Convex (DC) func-
tions approaches, convex-concave approximation methods; ii)
our convergence conditions unify and extend to the general
class P those of current (centralized) SCA methods; iii) it
offers much flexibility in the choice of the convex approxima-
tion functions: for instance, as a major departure from current
2
SCA-based methods [applicable to special cases of P] [10],
[12] and DC programs [15], the proposed approximation of the
objective function U need not be a tight global upper bound of
U , a fact that significantly enlarges the range of applicability
of our framework; and iv) by allowing alternative choices for
the convex approximants, it encompasses a gamut of novel
algorithms, offering great flexibility to control iteration com-
plexity, communication overhead and convergence speed, and
all converging under the same conditions. Quite interestingly,
the proposed scheme leads to new efficient algorithms even
when customized to solve well-researched problems, including
power control problems in cellular systems [22] -- [25], MIMO
relay optimization [26], dynamic spectrum management in
DSL systems [27], [28], sum-rate maximization, proportional-
fairness and max-min optimization of SISO/MISO/MIMO
ad-hoc networks [13], [29] -- [31], robust optimization of CR
networks [32] -- [34], transmit beamforming design for multiple
co-channel multicast groups [35], [36], and cross-layer design
of wireless networks [37], [38]. Part II of the paper [19]
is devoted to the application of the proposed algorithmic
framework to some of the aforementioned problems (and their
generalizations). Numerical results show that our schemes
compare favorably to existing ad-hoc ones (when they exist).
The rest of this two-part paper is organized as follows. Sec.
II introduces the main assumptions underlying the study of the
optimization problem P and provides an informal description
of our new algorithms. Sec. III presents our novel framework
based on SCA, whereas Sec. IV focuses on its distributed
implementation in the primal and dual domain. Finally, Sec.V
draws some conclusions. In Part II of the paper [19] we
apply our algorithmic framework to several resource allocation
problems in wireless networks and provide extensive numer-
ical results showing that the proposed algorithms compare
favorably to state-of-the-art schemes.
II. TECHNICAL PRELIMINARIES AND MAIN IDEA
In this section we introduce the main assumptions under-
lying the study of the optimization problem P along with
some technical results that will be instrumental to describe
our approach. We also provide an informal description of
our new algorithms that sheds light on the core idea of the
proposed decomposition technique. The formal description of
the framework is given in Sec. III.
Consider problem P, whose feasible set is denoted by X .
Assumption 1. We make the blanket assumptions:
A1) K ⊆ Rn is closed and convex (and nonempty);
A2) U and each gj are continuously differentiable on K;
A3) ∇xU is Lipschitz continuous on K with constant L∇U .
A4) U is coercive on K.
The assumptions above are quite standard and are satisfied
by a large class of problems of practical interest. In particular,
A4 guarantees that the problem has a solution, even when the
feasible set X is not bounded. Note that we do not assume
convexity of U and g1, . . . , gm; without loss of generality,
convex constraints, if present, are accommodated in the set K.
Our goal is to efficiently compute locally optimal solutions
of P, possibly in a distributed way, while preserving the fea-
sibility of the iterates. Building on the idea of SCA methods,
our approach consists in solving a sequence of strongly convex
inner approximations of P in the form: given xν ∈ X
) , X (xν ),
gj(x; xν ) ≤ 0, j = 1, . . . , m
x ∈ K
(Pxν )
U (x; xν )
min
x
s.t.
We introduce next a number of assumptions that will be
where U (x; xν ) and gj(x; xν ) represent approximations of
U (x) and gj(x) at the current iterate xν, respectively, and
X (xν ) denotes the feasible set of Pxν .
used throughout the paper.
Assumption 2 (On U). Let U : K × X → R be a function
continuously differentiable with respect to the first argument
and such that:
B1) U (•; y) is uniformly strongly convex on K with constant
c U> 0, i.e. ∀x, z ∈ K, ∀y ∈ X
(x − z)T(cid:16)∇x U (x; y) − ∇x U (z; y)(cid:17) ≥ c Ukx − zk2;
B2) ∇x U (y; y) = ∇xU (y), for all y ∈ X ;
B3) ∇x U (•;•) is continuous on K × X ;
where ∇x U (u; w) denotes the partial gradient of U with
respect to the first argument evaluated at (u; w).
Assumption 3 (On gjs). Let each gj : K × X → R satisfy
the following:
C1) gj(•; y) is convex on K for all y ∈ X ;
C2) gj(y; y) = gj(y), for all y ∈ X ;
C3) gj(x) ≤ gj(x; y) for all x ∈ K and y ∈ X ;
C4) gj(•;•) is continuous on K × X ;
C5) ∇xgj(y) = ∇xgj(y; y), for all y ∈ X ;
C6) ∇xgj(•;•) is continuous on K × X ;
where ∇xgj(y; y) denotes the (partial) gradient of gj with
respect
to the first argument evaluated at y (the second
argument is kept fixed at y).
For some results we need stronger continuity properties of
the (gradient of the) approximation functions.
Assumption 4
B4) ∇x U (x;•) is uniformly Lipschitz continuous on X with
constant L∇,2;
B5) ∇x U (•; y) is uniformly Lipschitz continuous on K with
constant L∇,1;
C7) Each gj(•;•) is Lipschitz continuous on K × X .
The key assumptions are B1, C1, and C3: B1 and C1 make
Pxν strongly convex, whereas C3 guarantees X (xν ) ⊆ X
(iterate feasibility). The others are technical conditions (easy to
be satisfied in practice) ensuring that the approximations have
the same local first order behavior of the original functions.
In the next section we provide some examples of approximate
functions that automatically satisfy Assumptions 2-4. As a
final remark, we point out that Assumptions 1-3 are in many
ways similar but generally weaker than those used in the
literature in order to solve special cases of problem P [10] --
[14]. For instance, [12] -- [14] studied the simpler case of
convex constraints; moreover, [12] requires the convex approx-
imation U (•; xν) to be a global upper bound of the nonconvex
objective function U (•), while we do not. The upper bound
condition C3 is assumed also in [10], [11] but, differently from
those works, we are able to handle also nonconvex objective
functions (rather than only strongly convex ones). Our weaker
conditions on the approximations U and g along with a more
general setting allow us to deal with a much larger class of
problems than [10] -- [14]; see Part II of the paper [19] for
specific examples.
A. Regularity conditions
We conclude this section mentioning certain standard reg-
ularity conditions on the stationary points of constrained
optimization problems. These conditions are needed in the
study of the convergence properties of our method.
µj ≥ 0, ∀j ∈ ¯J
Definition 1 (Regularity): A point ¯x ∈ X is called regular
for P if the Mangasarian-Fromovitz Constraint Qualification
(MFCQ) holds at ¯x, that is (see e.g. [39, Theorem 6.14]) if
the following implication is satisfied:
) ⇒ µj = 0, ∀j ∈ ¯J,
0 ∈Pj∈ ¯J µj∇xgj(¯x) + NK(¯x)
(1)
where NK(¯x) , {d ∈ K : dT (y − ¯x) ≤ 0, ∀y ∈ K} is the
normal cone to K at ¯x, and ¯J , {j ∈ {1, . . . , m} : gj(¯x) =
0} is the index set of those (nonconvex) constraints that are
active at ¯x.
A similar definition holds for problem Pxν : a point ¯x ∈
X (xν ) is called regular for Pxν if
) ⇒ µj = 0, ∀j ∈ ¯J ν,
0 ∈Pj∈ ¯J µj∇xgj(¯x; xν ) + NK(¯x)
where ¯J ν , {j ∈ {1, . . . , m} : gj(¯x; xν ) = 0}.
We point out that the regularity of ¯x is implied by stronger
but easier to be checked CQs, such as the Linear Independence
CQ, see [40, Sec. 3.2] for more details. Note that if the feasible
set is convex, as it is in Pxν , the MFCQ is equivalent to the
Slater's CQ; for a set like X (xν ), Slater's CQ reads
µj ≥ 0, ∀j ∈ ¯J ν
(2)
(cid:3)
ri(K) ∩ X <
g (xν ) 6= ∅,
where X <
g (xν ) , {x ∈ K : gj(x; xν ) < 0, j = 1, . . . , m}
and ri(K) is the relative interior of K (see, e.g., [41, Sec.
1.4]). In particular, this means that for problem Pxν either the
MFCQ holds at all the feasible points or it does not hold at
any point. Furthermore, because of C2 and C5, a point ¯x is
regular for P if and only if ¯x is regular for P¯x (and, therefore,
if any feasible point of P¯x is regular).
We recall that ¯x is a stationary point of problem P, if
0 ∈ ∇xU (¯x) +Pj∈ ¯J µj∇xgj(¯x) + NK(¯x)
µj ≥ 0, ∀j ∈ ¯J
for some suitable Lagrange multipliers µjs. It is well-known
that a regular (local) minimum point of problem P is also
stationary. Finding stationary points is actually the classical
goal of solution algorithms for nonconvex problems.
In order to simplify the presentation, in the rest of this paper
we assume the following regularity condition.
Assumption 5 All feasible points of problem P are regular.
3
One could relax this assumption and require regularity only at
specific points, but at the cost of more convoluted statements;
we leave this task to the reader. We remark, once again, that
Assumption 5 implies that any feasible point of P¯x is regular.
III. ALGORITHMIC FRAMEWORK
We are now ready to formally introduce the proposed
solution method for P. Note first that, because of B1 and
C1, each subproblem Pxν is strongly convex and thus has a
unique solution, which is denoted by x(xν ) (a function of xν):
x(xν ) , argmin
x∈X (xν)
U (x; xν ).
(3)
The proposed convex approximation method consists in
solving iteratively the optimization problems (3), possibly
including a step-size in the iterates; we named it iNner cOnVex
Approximation (NOVA) algorithm. The formal description of
the NOVA algorithm along with its convergence properties are
given in Algorithm 1 and Theorem 2, respectively.
Algorithm 1: NOVA Algorithm for P.
Data: γν ∈ (0, 1], x0 ∈ X ; set ν = 0.
(S.1) If xν is a stationary solution of P: STOP.
(S.2) Compute x(xν ), the solution of Pxν [cf. (3)].
(S.3) Set xν+1 = xν + γν(x(xν ) − xν).
(S.4) ν ← ν + 1 and go to step (S.1).
Theorem 2: Given the nonconvex problem P under As-
sumptions 1-3 and 5, let {xν} be the sequence generated by
Algorithm 1. The following hold.
(a) xν ∈ X (xν ) ⊆ X for all ν ≥ 0 (iterate feasibility);
(b) If the step-size γν and c U are chosen so that
ν
0 < inf
ν
γν ≤ sup
γν ≤ γmax ≤ 1 and 2c U > γmaxL∇,
(4)
then {xν} is bounded and each of its limit points is a stationary
point of problem P.
(c) If the step-size γν is chosen so that
and Xν
γν ∈ (0, 1],
γν → 0,
γν = +∞,
(5)
then {xν} is bounded and at least one of its limit points
is stationary. If, in addition, Assumption 4 holds and X is
compact, every limit point of {xν} is stationary.
Furthermore, if the algorithm does not stop after a finite
number of steps, none of the stationary points above is a local
maximum of U .
Proof: The proof is quite involved and is given in the
appendix; rather classically, its crucial point is showing that
lim(inf)
ν→∞ kx(xν ) − xνk = 0.
A. Discussions on Algorithm 1
Algorithm 1 describes a novel family of inner convex
approximation methods for problem P. Roughly speaking, it
consists in solving the sequence of strongly convex problems
Pxν wherein the original objective function U is replaced
by the strongly convex (simple) approximation U , and the
nonconvex constraints gjs with the convex upper estimates
4
gjs; convex constraints, if any, are kept unaltered. A step-size
in the update of the iterates xν is also used, in the form of
a convex combination via γν ∈ (0, 1] (cf. Step 3). Note that
the iterates {xν} generated by the algorithm are all feasible
for the original problem P. Convergence is guaranteed under
mild assumptions that offer a lot of flexibility in the choice of
the approximation functions and free parameters [cf. Theorem
2(b) and (c)], making the proposed scheme appealing for many
applications. We provide next some examples of candidate
approximants, covering a variety of situations and problems
of practical interest.
1) On the approximations gjs: As already mentioned, while
assumption C3 might look rather elusive, in many practical
cases an upper approximate function for the nonconvex con-
straints gjs is close at hand. Some examples of gj satisfying
Assumption 3 (and in particular C3) are given next; specific
applications where such approximations are used are discussed
in detail in Part II of the paper [19].
Example #1− Nonconvex constraints with Lipschitz gradients.
If the nonconvex function gj does not have a special structure
but Lipschitz continuous gradient on K with constant L∇gj ,
the following convex approximation function is a global upper
bound of gj: for all x ∈ K and y ∈ X ,
L∇gj
2 kx−yk2 ≥ gj(x).
gj(x; y) , gj(y)+∇xgj(y)T (x−y)+
(6)
Example #2− Nonconvex constraints with (uniformly) bounded
Hessian matrix. Suppose that gj is (nonconvex)C2 with second
order bounded derivatives on K. Then, one can find a matrix
xgj(x) + G (cid:23) 0 for all x ∈ K. For
G ≻ 0 such that ∇2
λmin(∇2
xgj(x)
(which is a negative quantity if gj
is nonconvex). Then,
the unstructured nonconvex constraint gj can be equivalently
written as a DC function:
instance, one can set G =(cid:12)(cid:12)minx∈K λmin(∇2
xgj(x)) denoting the minimum eigenvalue of ∇2
xgj(x))(cid:12)(cid:12) · I, with
1
2
{z
1
2
}
{z
}
gj(x) = gj(x) +
xT Gx
−
xT Gx
,
(7)
,g+
j (x)
,g−
j (x)
j and g−
where g+
j are two convex continuously differentiable
functions. An approximant gj of gj satisfying Assumption
3 can then readily be obtained by linearizing g−
j (x); see
Example #3 below for details.
L∇gj or (cid:12)(cid:12)minx∈K λmin(∇2
xgj(x))(cid:12)(cid:12) −are not known exactly
The two examples above cover successfully quite general
unstructured functions gj. However, in some cases, the func-
tion parameters involved in the approximations−the constants
but need to be estimated; if the estimates are not tight, the
resulting gj might be a loose overestimation of gj, which may
negatively affect the practical convergence of Algorithm 1.
Other approximations can be obtained when gj has further
structure to exploit, as discussed in the next examples.
Example #3− Nonconvex constraints with DC structure. Sup-
pose that gj has a DC structure, that is,
j (x) − g−
gj(x) = g+
j (x)
is the difference of two convex and continuously differentiable
j . By linearizing the concave part −g−
functions g+
j
and keeping the convex part g+
j unchanged, we obtain the
j and g−
following convex upper approximation of gj: for all x ∈ K
and y ∈ X ,
gj(x; y) , g+
j (y)T (x−y) ≥ gj(x). (8)
Example #4− Bi-linear constraints. Suppose that gj has a bi-
linear structure, that is,
j (y)−∇xg−
j (x)−g−
gj(x1, x2) = x1 · x2.
(9)
Observe preliminarily that gj(x1, x2) can be rewritten as a DC
function:
gj(x1, x2) =
1
2
(x1 + x2)2 −
1
2
(x2
1 + x2
2).
(10)
1
2
(y2
1 + y2
2)
A valid gj can be then obtained linearizing the concave part
in (10): for any given (y1, y2) ∈ R2,
gj (x1, x2; y1, y2) ,
1
(x1 + x2)2 −
2
−y1 · (x1 − y1) − y2 · (x2 − y2).
In Part II of the paper [19] we show that
the constraint
functions of many resource allocation problems in wireless
systems and networking fit naturally in Examples 1-4 above.
2) On the approximation U : The function U should be
regarded as a (possibly simple) convex approximation that
preserves the first order properties of U . Some instances of
valid U s for a specific U occurring in practical applications
are discussed next.
Example #5− Block-wise convex U (x1, . . . , xn). In many
applications, the vector of variables x is partitioned in blocks
i=1 and the function U is convex in each block xi
x = (xi)I
separately, but not jointly. A natural approximation for such a
U exploring its "partial" convexity is
U (x; y) =
IXi=1
Ui(xi; y),
(11)
with each Ui(xi; y) defined as
τi
2
Ui(xi; y) , U (xi, y−i) +
+
(xi − yi)T Hi(y)(xi − yi), (12)
where y , (yi)I
i=1, y−i , (yj )j6=i, and Hi(y) is any
uniformly positive definite matrix (possibly depending on y).
Note that the quadratic term in (12) can be set to zero if
U (xi, y−i) is strongly convex in xi, uniformly for all feasible
y−i. An alternative choice for Ui(xi; y) is
Ui(xi; y) , ∇xiU (y)T (xi − yi)
τi
1
2 kxi − yik2 ,
2
xiU (y) is the Hessian of U w.r.t. xi evaluated in
where ∇2
y. One can also use any positive definite "approximation" of
xiU (y). Needless to say, if U (x1, . . . , xn) is jointly convex
∇2
in all the variables' blocks, then U (x; y) can be chosen so
that
xiU (y)(xi − yi) +
U (x; y) , U (x) +Xi
2 kxi − yik2 is not needed if U (xi, x−i) is strongly
where τi
convex in xi, uniformly for all feasible x−i.
τi
2 kxi − yik2,
(xi − yi)T∇2
(13)
Example #6−(Proximal) gradient-like approximations. If no
convexity whatsoever is present in U , mimicking proximal-
gradient methods, a valid choice of U is the first order
approximation of U , that is, U (x; y) =PI
Ui(xi; y), with
Ui(xi; y) , ∇xi U (y)T (xi − yi) +
τi
2 kxi − yik2 .
each
i=1
Note that even though classical (proximal) gradient descent
methods (see, e.g., [21]) share the same approximation func-
tion, they are not applicable to problem P, due to the noncon-
vexity of the feasible set.
Example #7− Sum-utility function. In multi-agent scenarios,
the objective function U is generally written as U (x) ,
PI
the utilities
fi(x1, . . . , xI ) of I agents, each controlling the variables xi.
A typical situation is when the fis are convex in some agents'
variables. To capture this property, let us define by
i=1 fi(x1, . . . , xI ),
the sum of
that
is,
Si , {j : fj(•, x−i) is convex in xi, ∀(xi, x−i) ∈ K}
the set of indices of all the functions fj(xi, x−i) that are
convex in xi, for any feasible x−i, and let Ci ⊆ Si be
any subset of Si. Then, the following approximation function
U (x; y) satisfies Assumption 2 while exploiting the partial
convexity of U (if any): U (x; y) = PI
UCi(xi; y), with
fj(xi, y−i) +Xk /∈Ci
UCi(xi; y) , Xj∈Ci
∇xifk(y)T (xi − yi)
each UCi defined as
i=1
+
τi
2
(xi − yi)T Hi(y)(xi − yi),
where Hi(y) is any uniformly positive definite matrix.
Roughly speaking, for each agent i we built an approximation
function such that the convex part of U w.r.t. xi may be
preserved while the nonconvex part is linearized.
Example #8− Product of functions. The function U is often the
product of functions (see Part II [19] for some examples); we
consider here the product of two functions, but the proposed
approach can be readily extended to the case of three or more
functions or to the sum of such product terms. Suppose that
U (x) = f1(x)f2(x), with f1 and f2 convex and positive. In
view of the expression of the gradient of U , ∇xU = f2∇xf1+
f1∇xf2, it seems natural to consider the approximation
U (x; y) = f1(x)f2(y)+f1(y)f2(x)+
(x−y)T H(y)(x−y),
where, as usual, H(y) is a uniformly positive definite matrix;
this term can be omitted if f1 and f2 are bounded away from
zero on the feasible set and f1 + f2 is strongly convex (for
example if one of the two functions is strongly convex). It is
clear that this U satisfies Assumption 2. In case f1 and f2
are still positive but not necessarily convex, we can use the
expression
τi
2
U (x; y) = f1(x; y)f2(y) + f1(y) f2(x; y),
where f1 and f2 are any legitimate approximations for f1
and f2, for example those considered in Examples 5-7 above.
Finally, if f1 and f2 can take non-positive values, we can write
U (x; y) = h1(x, y) + h2(x, y),
5
where h1(x, y) , f1(x; y)f2(y), h2(x, y) , f1(y) f2(x; y),
and h1 and h2 are legitimate approximations for h1 and h2, for
example, again, those considered in Examples 5-7. Note that in
the last cases we no longer need the quadratic term because
it is already included in the approximations f1 and f2, and
h1 and h2, respectively. As a final remark, it is important to
point out that the the U s discussed above belong to a class of
nonconvex functions for which it does not seem possible to
find a global convex upper bound; therefore, all current SCA
techniques (see, e.g., [10], [12], [15]) are not applicable.
We conclude the discussion on the approximation functions
observing that, in Examples 5-7, the proposed U (x; y)s are
all separable in the blocks xi for any given y; in Example
8, instead, the separability is problem dependent and should
be examined on a case-by-case basis. The separability of the
U s paves the way to parallelizable and distributed versions of
Algorithm 1; we discuss this topic more in detail in Sec. IV.
3) On the choice of the step-size rule γν: Theorem 2 states
that Algorithm 1 converges either employing a constant step-
size rule [cf. (4)] or a diminishing step-size rule [cf. (5)].
If a constant step-size is used, one can set in (4) γν =
γ ≤ γmax for every ν, and choose any γmax ∈ (0, 1] and
c U so that 2c U > γmaxL∇U (recall that c U is the constant of
strong convexity of the approximation U and, thus, is a degree
of freedom). This can be done in several ways. For instance,
if the chosen U contains a proximal term with gain τ > 0,
i.e., a term of the type (τ /2)kx − yk2, then the inequality
2c U > γmaxL∇U is readily satisfied setting 2τ /γmax > L∇U
(we used c U ≥ τ). Note that this simple (but conservative)
condition imposes a constraint only on the ratio τ /γmax,
leaving free the choice of one of the two parameters. An
interesting special case worth mentioning is when γmax = 1
and 2τ > L∇U : the choice γν = 1 leads to an instance of
Algorithm 1 with no memory, i.e., xν+1 = x(xν ), for all ν.
When the Lipschitz constant L∇U cannot be estimated, one
can use a diminishing step-size rule, satisfying the standard
conditions (5). A rule that we found to work well in practice
is, see [13]:
ν ≥ 1,
γν = γν−1(1 − εγν−1),
(14)
with γ0 ∈ (0, 1] and ε ∈ (0, 1). Other effective rules can be
found in [13]. Notice that, while this rule may still require
some tuning for optimal behavior, it is quite reliable, since
in general we are not using a (sub)gradient direction, so
that many of the well-known practical drawbacks associated
with a (sub)gradient method with diminishing step-size are
mitigated in our setting. Furthermore, this choice of step-size
does not require any form of centralized coordination and,
thus, provides a favorable feature in distributed environments.
We remark that it is possible to prove the convergence of
Algorithm 1 also using other step-size rules, such as a standard
Armijo-like line-search procedure. We omit the discussion of
line-search based approaches because such options are not in
line with our goal of developing distributed algorithms, see
Sec. IV. In [11] it is shown that, in the specific case of a
strongly convex U and, in our terminology, U = U and K =
Rn, by choosing γν = 1 at every iteration, one can prove
the stationarity of every limit point of the sequence generated
6
by Algorithm 1 (assuming regularity). We can easily derive
this particular result from our general analysis, see Remark
12 in the Appendix. Here we only mention that, attractive
as this result may be, the strong convexity of U is a very
restrictive assumption, and forcing U = U does not permit
the development of distributed versions of Algorithm 1.
Finally, as a technical note, it is interesting to contrast the
different kinds of convergence that one can obtain by choosing
a constant or a diminishing step-size rule. In the former case,
every limit point of the sequence generated by Algorithm
1 is guaranteed to be a stationary solution of the original
nonconvex problem P, whereas, in the latter case, there exists
at least a limit point being stationary, which thus is a weaker
condition. On the other hand, a diminishing step-size has been
observed to be numerically more efficient than a constant one.
In order to obtain, also with a diminishing step-size rule, the
strong convergence behavior that can be guaranteed when a
constant step-size is employed, one needs extra conditions on
the approximation functions U and g (cf. Assumptions 4);
these conditions are quite standard and easy to be satisfied in
most practical applications (as those studied in Part II [19]).
B. Related works
Our approach draws on the SCA paradigm, which has been
widely explored in the literature, see [10] -- [15]. However,
our framework and convergence conditions unify and extend
current SCA methods in several directions, as outlined next.
−On the approximation functions: Conditions on the approx-
imation function U as in Assumption 2 are relatively weak:
this feature allows us to enlarge significantly the class of utility
functions U s that can be successfully handled by Algorithm
1. A key difference with current SCA methods [applicable to
special cases of P] [10], [12] and DC programs [15] is that the
approximation U (x; y) need not be a tight global upper bound
of the objective function U (x) for every x ∈ K and y ∈ X
[cf. Assumption 2]. This fact represents a big step forward
in the literature of SCA methods; Part II of the paper [19]
provides a solid evidence of the wide range of applicability of
the proposed framework.
−Convergence conditions: There are only a few SCA-based
methods in the literature handling nonconvex constraints,
namely [10], [11], [15], and the existing convergence results
are quite weak. In particular, [10, Th. 1] states that if the
whole sequence converges, then the algorithm converges to
a stationary point; however, in general, it is hardly possible
to show that the sequence generated by the algorithms does
converge. In [11], (subsequence) convergence to regular points
is proved, but only for nonconvex problems with strongly
convex objective functions; this fact restricts considerably the
range of applicability of this result (for instance, none of the
problems that we study in Part II [19] have strongly convex
objective functions). Finally, [15] can handle only (possibly
nonsmooth) nonconvex problems whose objective functions
and constraints have a DC form. To the best of our knowledge,
this work is the first attempt towards the generalization of SCA
methods to nonconvex problems having general nonconvex
objective functions and constraints.
−Distributed implementation: A second key and unique fea-
ture of Algorithm 1, missing in current SCA schemes [10],
[11], [15], is that it easily leads to distributed implementations,
as we will discuss in Sec. IV. This feature, along with the
feasibility of the iterates, represents a key difference also with
classical techniques [6] -- [9] that have been proposed in the
literature to deal with nonconvex optimization problems.
IV. DISTRIBUTED IMPLEMENTATION
j(xi; y), for some gi
j : Ki × X → R.
D3) each approximate function gj(x; y) satisfying Assumption
3 is (block) separable in the x-variables, for any given y, that
In many applications, e.g., multi-agent optimization or dis-
tributed networking, it is desirable to keep users coordination
and communication overhead at minimum level. In this section
we discuss distributed versions of Algorithm 1. Of course, we
need to assume that problem P has some suitable structure,
and that consistent choices on U and g are made. Therefore, in
this section we consider the following additional assumptions.
Assumption 6 (Decomposabilty). Given P, suppose that:
D1) the set K has a Cartesian structure, i.e., K = K1 × ··· ×
KI, with each Ki ⊂ Rni, and Pi ni = n; x , (xi)I
i=1 is
partitioned accordingly, with each xi ∈ Ki;
D2) the approximate function U (x; y) satisfying Assumption
2 is chosen so that U (x; y) =Pi
Ui(xi; y);
is, gj(x; y) =Pi gi
Condition D1 is a very natural assumption on problem P and
is usually satisfied when a distributed implementation is called
for. If problem P does not satisfy this assumption, it is not
realistic to expect that efficient distributed solution methods
can be devised; D2 and D3, instead, are assumptions on our
algorithmic choices. In particular, condition D2 permits many
choices for U . For instance, as already discussed at the end of
the subsection "On the approximation U ", essentially all U s
introduced in Examples 5-7 (and possibly some of the U s in
Example 8) satisfy D2. The critical condition in Assumption
6 is D3. Some examples of constraints functions gj for which
one can find a gj(x; y) satisfying D3 are:
−Individual nonconvex constraints: Each gj (still nonconvex)
depends only on one of the block variables x1, . . . , xI , i.e,
gj(x) = gi
−Separable nonconvex constraints: Each gj has the form
gj(x) =Pi gi
−Nonconvex constraints with Lipschitz gradients: Each gj is
not necessarily separable but has Lipschitz gradient on K. In
this case one can choose, e.g., the approximation gj as in (6).
It is important to remark that, even for problems P [or Pxν ]
for which it looks hard to satisfy D3, the introduction of proper
slack variables can help to decouple the constraint functions,
making thus possible to find a gj that satisfies D3; we refer
the reader to Part II of the paper [19] for some non trivial
examples where this technique is applied.
j : Ki → R and i;
j : Ki → R;
j(xi), for some gi
j(xi), with gi
For notational simplicity, let us introduce the vector func-
j=1, for i = 1, . . . , I. Under
tion gi(xi; xν ) , (gi
Assumption 6, each subproblem Pxν becomes
j(xi; xν ))m
min
x PI
s.t.
Ui(xi; xν )
i=1
g(x; xν ) ,Pi gi(xi; xν) ≤ 0,
i = 1, . . . , I.
xi ∈ Ki,
( Pxν )
With a slight abuse of notation, we will still denote the
feasible set of Pxν by X (xν ).
The block separable structure of both the objective function
and the constraints lends itself to a parallel decomposition of
the subproblems Pxν in the primal or dual domain: hence, it
allows the distributed implementation of Step 2 of Algorithm
1. In the next section we briefly show how to customize
standard primal/dual decomposition techniques (see, e.g., [20],
[21]) in order to solve subproblem Pxν . We conclude this
section observing that, if there are only individual constraints
in P, given xν, each Pxν can be split in I independent
subproblems in the variables xi, even if the original nonconvex
U is not separable. To the best of our knowledge, this is the
first attempt to obtain distributed algorithms for a nonconvex
problem in the general form P.
A. Dual decomposition methods
Subproblem Pxν is convex and can be solved in a distributed
way if the constraints g(x; xν ) ≤ 0 are dualized. The dual
problem associated with each Pxν is: given xν ∈ X (xν ),
7
The dual-problem can be solved, e.g., using well-known gradi-
ent algorithms [41]; an instance is given in Algorithm 2, whose
convergence is stated in Theorem 4. The proof of the theorem
follows from Lemma 3 and standard convergence results for
gradient projection algorithms (e.g., see [42, Th. 3.2] and [41,
Prop. 8.2.6] for the theorem statement under assumptions (a)
and (b), respectively). We remark that an assumption made
in the aforementioned references is that subproblem Pxν has a
zero-duality gap and the dual problem (15) has a non-empty
solution set. In our setting, this is guaranteed by Assumption
that ensures that X (xν ) satisfies Slater's CQ (see the
5,
discussion around Assumption 5).
In (19) [•]+ denotes the Euclidean projection onto R+, i.e.,
[x]+ , max(0, x).
Algorithm 2: Dual-based Distributed Implementation of Step
2 of NOVA Algorithm (Algorithm 1).
Data: λ0 ≥ 0, xν, {αn} > 0; set n = 0.
(S.2a) If λn satisfies a suitable termination criterion: STOP.
(S.2b) Solve in parallel (17): for all i = 1, . . . , I, compute
(S.2c) Update λ according to
bxi(λn; xν ).
λn+1 ,"λn + αn
IXi=1
gi (bxi(λn; xν); xν )#+
.
(19)
max
λ≥0
d(λ; xν )
(S.2d) n ← n + 1 and go back to (S.2a).
(15)
with
d(λ; xν) , min
x∈KnPI
bx(λ; xν ) , (bxi(λ; xν ))I
bxi(λ; xν ) , argmin
i=1(cid:16) Ui(xi; xν) + λT gi(xi; xν )(cid:17)o(16)
Note that, for λ ≥ 0, by Assumptions 2 and 3, the minimiza-
tion in (16) has a unique solution, which will be denoted by
i=1, with
xi∈Ki n Ui(xi; xν ) + λT gi(xi; xν)o .
(17)
let us mention the following standard
Before proceeding,
condition.
D4) g(•; xν) is uniformly Lipschitz continuous on K with
constant Lg;
We remind that D4 is implied by condition C7 of Assumption
4; therefore we do not need to assume it if Assumption 4 is
invoked. But in order to connect our results below to classical
ones, it is good to highlight it as a separate assumption.
The next lemma summarizes some desirable properties of
the dual function d(λ; xν ), which are instrumental to prove
the convergence of dual schemes.
Lemma 3: Given Pxν , under Assumptions 1-3, 5, and 6,
+ , with
the following hold.
(a) d(λ; xν ) is differentiable with respect to λ on Rm
gradient
∇λd(λ; xν ) =Xi
(b) If in addition D4 holds, then ∇λd(λ; xν) is Lipschitz
continuous on Rm
+ with constant L∇d , L2
Proof: See Appendix C.
(18)
gi (bxi(λ; xν ); xν ) .
g √m/c U .
Theorem 4: Given P, under Assumptions 1-3, 5, and 6,
suppose that one of the two following conditions is satisfied:
(a) D4 holds true and {αn} is chosen such that 0 < inf n αn ≤
supn αn < 2/L∇d, for all n ≥ 0;
+ , and αn is
(b) ∇λd(•; xν ) is uniformly bounded on Rm
chosen such that αn > 0, αn → 0, Pn αn = ∞, and
Pn(αn)2 < ∞.
Then, the sequence {λn} generated by Algorithm 2 con-
verges to a solution of (15), and the sequence {bx(λn; xν)}
converges to the unique solution of Pxν .
The implementation of Algorithm 1 based on Algorithm
2 leads to a double-loop scheme: given the current value of
the multipliers λn, the subproblems (17) can be solved in
Remark 5 (On the distributed implementation):
parallel across the blocks; once the new valuesbxi(λn; xν) are
available, the multipliers are updated according to (19). Note
that when m = 1 (i.e., there is only one shared constraint), the
update in (19) can be replaced by a bisection search, which
generally converges quite fast. When m > 1, the potential
slow convergence of gradient updates (19) can be alleviated
using accelerated gradient-based (proximal) schemes; see,
e.g., [43], [44].
As far as the communication overhead is concerned, the
required signaling is in the form of message passing and
of course is problem dependent; see Part II of the paper
[19] for specific examples. For instance, in networking ap-
plications where there is a cluster-head, the update of the
multipliers can be performed at the cluster, and, then, it can
be broadcasted to the users. In fully decentralized networks
instead, the update of λ can be done by the users themselves,
8
by running consensus based algorithms to locally estimate
PI
i=1 gi (bxi(λn; xν ); xν ). This, in general, requires a limited
signaling exchange among neighboring nodes only. Note also
that the size of the dual problem (the dimension of λ) is
equal to m (the number of shared constraints), which makes
Algorithm 2 scalable in the number of blocks (users).
B. Primal decomposition methods
Algorithm 2 is based on the relaxation of the shared con-
straints into the Lagrangian, resulting, in general, in a violation
of these constraints during the intermediate iterates. In some
applications this fact may prevent the on-line implementation
of the algorithm. In this section we propose a distributed
scheme that does not suffer from this issue: we cope with the
shared constraints using a primal decomposition technique.
i=1, with each ti ∈
Introducing the slack variables t , (ti)I
Rm, Pxν can be rewritten as
min
(xi,ti)I
s.t.
i=1 PI
PI
Ui(xi; xν ),
i=1
xi ∈ Ki,
gi (xi; xν ) ≤ ti,
∀i = 1, . . . , I,
∀i = 1, . . . , I,
i=1 ti ≤ 0.
i=1 is fixed, (20) can be decoupled across the
When t = (ti)I
users: for each i = 1, . . . , I, solve
Ui(xi; xν ),
min
xi
s.t. xi ∈ Ki,
gi (xi; xν)
(21)
µi(ti;xν )
≤ ti,
where µi(ti; xν ) is the optimal Lagrange multiplier associated
with the inequality constraint gi (xi; xν) ≤ ti. Note that
the existence of µi(ti; xν ) is guaranteed if (21) has zero-
duality gap [45, Prop. 6.5.8] (e.g., when some CQ hold), but
µi(ti; xν ) may not be unique. Let us denote by x⋆
i (ti; xν )
i=1. The optimal
the unique solution of (21), given t =(ti)I
i=1 of the shared constraints can be found
partition t⋆ , (t⋆
solving the so-called master (convex) problem (see, e.g., [20]):
i )I
i=1
Ui(x⋆
i (ti; xν); xν )
(22)
P (t; xν) ,PI
i=1 ti ≤ 0.
min
t
s.t. PI
Due to the non-uniqueness of µi(ti; xν), the objective function
in (22) is nondifferentiable; problem (22) can be solved by
subgradient methods. The partial subgradient of P (t; xν ) with
respect to the first argument evaluated at (t; xν ) is
∂ti P (t; xν ) = −µi(ti; xν),
i = 1, . . . , I.
We refer to [41, Prop. 8.2.6] for standard convergence results
for subgradient projection algorithms.
V. CONCLUSIONS
In this Part I of the two-part paper, we proposed a novel gen-
eral algorithmic framework based on convex approximation
techniques for the solution of nonconvex smooth optimization
problems: we point out that the nonconvexity may occur both
in the objective function and in the constraints. Some key
novel features of our scheme are: i) it maintains feasibility
(20)
APPENDIX
and leads to parallel and distributed solution methods for
a very general class of nonconvex problems; ii) it offers a
lot of flexibility in choosing the approximation functions,
enlarging significantly the class of problems that can be
solved with provable convergence; iii) by choosing different
approximation functions, different (distributed) schemes can
be obtained: they are all convergent, but differ for (practical)
convergence speed, complexity, communication overhead, and
a priori knowledge of the system parameters; iv) it includes
as special cases several classical SCA-based algorithms and
improves on their convergence properties; and v) it provides
new efficient algorithms also for old problems. In Part II
[19] we customize the developed algorithmic framework to a
variety of new (and old) multi-agent optimization problems in
signal processing, communications and networking, providing
a solid evidence of its good performance. Quite interestingly,
even when compared with existing schemes that have been
designed for very specific problems, our algorithms are shown
to outperform them.
We first introduce some intermediate technical results that
are instrumental to prove Theorem 2. The proof of Theorem
2 is given in Appendix B.
A. Intermediate results
Lemma 6: The following properties hold.
We first prove Lemma 6-Lemma 10, providing some key
properties of the sequence {xν} generated by Algorithm 1
and of the best-response function x(•) defined in (3). Finally,
with Theorem 11 we establish some technical conditions under
which a(t least one) regular limit point of the sequence gen-
erated by Algorithm 1 is a stationary solution of the original
nonconvex problem P; the proof of Theorem 2 will rely on
such conditions. We recall that, for the sake of simplicity,
throughout this section we tacitly assume that Assumptions
1-3 and 5 are satisfied.
Lemma 6. The first lemma shows, among other things, that
Algorithm 1 produces a sequence of points that are feasible
for the original problem P.
(i) y ∈ X (y) ⊆ X for all y ∈ X ;
(ii) x(y) ∈ X (y) ⊆ X for all y ∈ X .
such that:
(iii) xν ∈ X ;
(iv) xν+1 ∈ X (xν ) ∩ X (xν+1).
Proof: (i) the first implication y ∈ X (y) follows from
gj(y; y) = gj(y) ≤ 0, for all j = 1, . . . , m [due to C2].
For the inclusion X (y) ⊆ X , it suffices to recall that, by
C3, we have gj(x) ≤ gj(x; y) for all x ∈ K, y ∈ X , and
j = 1, . . . , m, implying that, if x ∈ X (y), then x ∈ X .
(ii) x(y) ∈ X (y) since it is the optimal solution of Py (and
thus also feasible).
(iii) In view of (i) and (ii), it follows by induction and the
fact that xν+1 is a convex combination of xν ∈ X (xν ) and
x(xν ) ∈ X (xν ), which is a convex subset of X .
(iv) By (iii), xν+1 ∈ X (xν ). Furthermore, we have
gj(xν+1; xν+1) = gj(xν+1) ≤ 0, for all j = 1, . . . , m, where
Moreover, the sequence {xν} generated by Algorithm 1 is
the equality follows from C2 and the inequality is due to
xν+1 ∈ X ; thus, xν+1 ∈ X (xν+1).
Lemma 7. With Lemma 7, we establish some key properties
of the best-response function x(•). We will use the following
definitions. Given y, z ∈ X and ρ > 0, let
wρ(x(y), z) , x(y) − ρ∇x U (x(y); z);
(23)
and let PX (y)(u) denote the Euclidean projection of u ∈ Rn
onto the closed convex set X (y):
PX (y)(u) = argmin
(24)
x∈X (y)kx − uk .
Lemma 7: The best-response function X ∋ y 7→ x(y)
satisfies the following:
(i) For every y ∈ X , x(y) − y is a descent direction for U at
y such that
∇U (y)T (x(y) − y) ≤ −c Ukx(y) − yk2,
(25)
where c U > 0 is the constant of uniform strong convexity of
U (cf. B1);
(ii) For every y ∈ X , it holds that
x(y) = PX (y) (wρ (x(y), y)) ,
(26)
for every fixed ρ > 0.
(iii) Suppose that also B5 holds true. Then, x(•) is continuous
at every ¯x ∈ X such that x(¯x) ∈ X (¯x) is regular.
Proof: (i) By Assumption 2, for any given y ∈ X , x(y)
is the solution of the strongly convex optimization problem
Pxν ; therefore,
(z − x(y))T∇x U (x(y); y) ≥ 0,
(27)
By choosing z = y [recall by Lemma 6(i) that y ∈ X (y)],
we get
∀z ∈ X (y).
(y − x(y))T(cid:16)∇x U (x(y); y) −∇x U (y; y)
+∇x U (y; y)(cid:17) ≥ 0,
which, using B1 and B2, leads to
(y − x(y))T∇xU (y) ≥ c Ukx(y) − yk2.
(ii) It follows readily from the fixed-point characterization of
the solution x(y) of the strongly convex subproblem Py: see,
e.g., [46, Prop. 1.5.8].
(iii) We first observe that, under the assumed regularity of all
the points in X (¯x), X (•) is continuous at ¯x [39, Example
5.10]. It follows from [39, Proposition 4.9] (see also [39,
Example 5.57]) that, for every fixed u ∈ Rn,
the map
x 7→ PX (x)(u) is continuous at x = ¯x. This, together with
B1, B3 and B5 is sufficient for x(•) to be continuous at ¯x
[47, Theorem 2.1].
Lemma 8. Under the extra conditions B4-B5, with the fol-
lowing lemma, which is reminiscent of similar results in [47]
and [48], we can establish a suitable sensitivity property of the
best-response function x(•); Lemma 8 will play a key role in
the proof of statement (c) of Theorem 2.
9
Lemma 8: Suppose that B4-B5 hold and there exist ¯ρ > 0
and β > 0 such that
kPX (y)(wρ(x(z), z)) − PX (z)(wρ(x(z), z))k ≤ βky − zk
2 ,
(28)
for all ρ ∈ (0, ¯ρ] and y, z ∈ X . Then there exists ρ ∈ (0, ¯ρ]
such that
1
kx(y) − x(z)k ≤ ηρky − zk + θρky − zk
1
2 ,
(29)
ηρ ,
for all y, z ∈ X and ρ ∈ (0, ρ], with
ρ L∇,2
1 −q1 + ρ2 L2
1 −q1 + ρ2 L2
θρ ,
β
∇,1 − 2ρc U
∇,1 − 2ρc U
(30)
,
where L∇,1 and L∇,2 are the Lipschitz constants of
∇x U (•; y) and ∇x U (x;•), respectively (cf. B4 and B5);
L∇,1 is assumed to be such that L∇,1 ≥ c U without loss
of generality.
Proof: Using (26) we have, for every ρ > 0,
kx(y) − x(z)k
= kPX (y)(wρ(x(y), y)) − PX (z)(wρ(x(z), z))k
≤ kPX (y)(wρ(x(y), y)) − PX (y)(wρ(x(z), y))k
+kPX (y)(wρ(x(z), y)) − PX (z)(wρ(x(z), z)k.
We bound next the two terms on the RHS of (31).
For every ρ > 0, it holds
(31)
kPX (y)(wρ(x(y), y)) − PX (y)(wρ(x(z), y))k2
(a)
≤ kwρ(x(y), y) − wρ(x(z), y)k2
= kx(y) − x(z)k2
+ρ2k∇x U (x(z); y) − ∇x U (x(y); y)k2
−2ρ (x(y) − x(z))T(cid:16)∇x U (x(y); y) − ∇x U (x(z); y)(cid:17)
≤ kx(y) − x(z)k2 + ρ2 L2
−2 ρ c U kx(y) − x(z)k2
∇,1kx(y) − x(z)k2
(b)
= (1 + ρ2 L2
∇,1 − 2 ρc U )kx(y) − x(z)k2,
(32)
where (a) is due to the non-expansive property of the projec-
tion operator PX (y)(•) and (b) follows from B1 and B5. Note
that 1 + ρ2 L2
∇,1 − 2ρc U > 0 since we assumed L∇,1 ≥ c U .
Let us bound now the second term on the RHS of (31). For
every ρ ∈ (0, ¯ρ], we have
kPX (y)(wρ(x(z), y)) − PX (z)(wρ(x(z), z))k
(a)
≤ kPX (y)(wρ(x(z), y)) − PX (y)(wρ(x(z), z))k
+kPX (y)(wρ(x(z), z)) − PX (z)(wρ(x(z), z))k
≤ kwρ(x(z), y) − wρ(x(z), z)k + βky − zk 1
≤ ρ L∇,2 ky − zk + βky − zk 1
2 ,
(b)
2
(33)
10
∇,1, ¯ρ} (so that 0 < 1 + ρ2 L2
Combining (31), (32) and (33) we obtain the desired result
∇,1 −
where (a) is due to the non-expansive property of the projec-
tion PX (y)(•) and (28), and (b) follows from B4.
(29) with ρ = min{2c U / L2
2ρc U < 1 for every ρ ∈ (0, ρ]).
Lemmas 9 and 10. While Assumptions 1-3 and B4-B5 in
Lemma 8 are quite standard, condition (28) is less trivial and
not easy to be checked. The following Lemma 10 provides
some easier to be checked sufficient conditions that imply (28).
To prove Lemma 10 we need first Lemma 9, as stated next.
Lemma 9: Consider ¯x ∈ X . By assuming C7, the following
hold:
(i) If x ∈ X (¯x) is regular, then X (•) enjoys the Aubin
property at (¯x, x);1
(ii) If in addition X is compact, then a neighborhood V¯x of
¯x, a neighborhood Wx of x, and a constant β > 0 exist such
that
(34)
kPX (y)(u) − PX (z)(u)k ≤ βky − zk
1
2
for all y, z ∈ X ∩ V¯x, and u ∈ Wx.
Proof: (i) Under Assumptions 1-3 and C7, the statement
follows readily from [49, Theorem 3.2] in view of the regu-
larity of x.
(ii) Since X (•) has the Aubin property at (¯x, x), there exist
a neighborhood V¯x of ¯x, a neighborhood Wx of x, and a
constant β > 0 such that [48, Lemma 1.1]:
kPX (y)∩Bx(u) − PX (z)∩Bx(u)k ≤ βky − zk 1
2 ,
(35)
for all y, z ∈ X ∩ V¯x, and u ∈ Wx, where Bx denotes a
closed convex neighborhood of x. Since X is compact, one
can always choose Bx such that X (¯x) ⊂ Bx for every ¯x ∈ X
and, thus,
kPX (y)∩Bx(u) − PX (z)∩Bx(u)k = kPX (y)(u) − PX (z)(u)k,
which proves the desired result.
We can now derive sufficient conditions for (28) to hold.
Lemma 10: Suppose that C7 holds true, X is compact and
x(¯x) ∈ X (¯x) is regular for every ¯x ∈ X . Then, property (28)
holds.
Proof: It follows from Lemma 9(ii) that, for every ¯x ∈ X ,
there exist a neighborhood V¯x of ¯x, a neighborhood Wx(¯x) of
x(¯x), and a constant β > 0 such that:
kPX (y)(u) − PX (z)(u)k ≤ βky − zk
1
2
(36)
for every y, z ∈ X ∩ V¯x, u ∈ Wx(¯x).
Suppose now by contradiction that (28) does not hold. Then,
for all ¯ρν > 0 and βν > 0 there exist ρν ∈ (0, ¯ρν], ¯xν, and
¯yν ∈ X such that:
kPX (¯yν)(wρν (x (¯xν ) , ¯xν )) − PX (¯xν)(wρν (x (¯xν) , ¯xν))k
> βνk¯yν − ¯xνk 1
2 .
(37)
1See [39, Def. 9.36] for the definition of the Aubin property. Note also that
we use some results from [48] where a point-to-set map that has the Aubin
property is called pseudo-Lipschitz [48, Def. 1.1].
Furthermore, in view of the compactness of X , denoting by
DX the (finite) diameter of X , the LHS of (37) can be bounded
by
DX ≥ kPX (¯yν )(wρν (x (¯xν ) , ¯xν))−PX (¯xν )(wρν (x (¯xν) , ¯xν ))k.
(38)
Suppose without loss of generality that βν → +∞, ¯ρν ↓ 0,
and ¯xν →N
¯y ∈ X (¯y) ⊆ X ,
possibly on a suitable subsequence N [recall that ¯xν ∈ X (¯xν )
and ¯yν ∈ X (¯yν )]. From (37) and (38), we obtain
¯x ∈ X (¯x) ⊆ X and ¯yν →N
DX ≥ lim sup
ν→+∞
βνk¯yν − ¯xνk
1
2 ,
which, in turn, considering that βν → ∞ and k¯yν− ¯xνk 1
implies
2 ≥ 0,
(39)
ν→+∞ k¯yν − ¯xνk
lim
1
2 = 0.
Then, it must be ¯x = ¯y.
Invoking now the continuity of x(•) at ¯x [cf. Lemma 7(iii)]
and ∇x U (•;•) on K × X [cf. B3], we have
wρν (x (¯xν) , ¯xν ) = x (¯xν) − ρν∇x U (x (¯xν) , ¯xν) →N
x(¯x).
(40)
Therefore, for every β > 0 and neighborhoods V¯x and
Wx(¯z), there exists a sufficiently large ν such that (37) holds
with βν > β (recall that βν → +∞), ¯xν , ¯yν ∈ V¯x ∩ X [due
to (39)], and wρν (x (¯xν ) , ¯xν ) ∈ Wx(¯z) [due to (40)]; this is
in contradiction with (36).
We recall that the assumption on the regularity of x(¯x) ∈
X (¯x) for every ¯x ∈ X , as required in Lemma 10, is implied
by Assumption 5.
Theorem 11. The last theorem of this section provides techni-
cal conditions under which a(t least one) regular limit point of
the sequence generated by Algorithm 1 is a stationary solution
of the original nonconvex problem P.
Theorem 11: Let {xν} be the sequence generated by Al-
gorithm 1 under Assumptions 1-3 and 5. The the following
hold.
(a) Suppose
ν→∞ kx(xν ) − xνk = 0.
liminf
(41)
Then, at least one regular limit point of {xν} is a stationary
solution of P.
(b) Suppose
ν→∞kx(xν ) − xνk = 0.
lim
(42)
Then, every regular limit point of {xν} is a stationary solution
of P.
Proof: We prove only (a); (b) follows applying the result
in (a) to every convergent subsequence of {xν}.
Let ¯x be a regular accumulation point of the subsequence
{xν}N of {xν} satisfying (41); thus, there exists N ′ ⊆ N
such that limN ′ ∋ν→∞ xν = ¯x. We show next that ¯x is a KKT
point of the original problem. Let ¯J and J ν be the following
sets:
¯J , {j ∈ [1, . . . , m] : gj(¯x) = 0},
J ν , {j ∈ [1, . . . , m] : gj(x(xν ); xν ) = 0}
with ν ∈ N ′. Using limN ′ ∋ν→∞ kx(xν )− xνk = 0 [cf. (41)]
along with the continuity of gj, by C2, we have
lim
N ′ ∋ν→∞
gj(x(xν ); xν ) = gj(¯x; ¯x) = gj(¯x), j = 1, . . . , m.
(43)
The limit above implies that there exists a positive integer
ν ∈ N ′ such that
J ν ⊆ ¯J, ∀ν ≥ ν and ν ∈ N ′.
(44)
Since the functions ∇x U and ∇xgj are continuous, we get,
by B2,
lim
N ′ ∋ν→∞∇x U (x(xν ); xν ) = ∇x U (¯x; ¯x) = ∇U (¯x),
(45)
and, for j = 1, . . . , m, by C5,
lim
N ′ ∋ν→∞ ∇xgj(x(xν ); xν ) = ∇xgj(¯x; ¯x) = ∇gj(¯x).
(46)
We claim now that for sufficiently large ν ∈ N ′, the MFCQ
holds at x(xν ) ∈ X (xν ). Assume by contradiction that the
following implication does not hold for infinitely many ν ∈
N ′:
−Pj∈J ν µν
j∇xgj(x(xν ); xν ) ∈ NK(x(xν ))
j ≥ 0, ∀j ∈ J ν,
µν
}
(47)
It follows that a nonempty index set ¯¯J ⊆ ¯J exists such that,
after a suitable renumeration, for every ν ∈ N ′, we must have
µν
j = 0, ∀j ∈ J ν.
{z
⇓
−Pj∈ ¯¯J µν
j ∇xgj(x(xν ); xν ) ∈ NK (x(xν ))
(48)
j ≥ 0, ∀j ∈ ¯¯J
µν
Pj∈ ¯¯J µν
j = 1.
We may assume without loss of generality that, for each
j ∈ ¯¯J, the sequence {µν
j} converges to a limit ¯µj such that
Pj∈ ¯¯J ¯µj = 1. In view of the inclusion ¯¯J ⊆ ¯J, by taking the
limit N
∋ ν → ∞ in (48), and invoking (46) along with the
outer semicontinuity of the mapping NK(•) [39, Prop. 6.6],
we get
′
−Pj∈ ¯¯J ¯µj∇xgj(¯x) ∈ NK (¯x)
¯µj ≥ 0, ∀j ∈ ¯¯J
Pj∈ ¯¯J ¯µj = 1,
(49)
j )m
j=1 such that
in contradiction with the regularity of ¯x [the MFCQ holds at
¯x, see (2)]. Therefore, (47) must hold for sufficiently large
ν ∈ N ′, implying that the KKT system of problem Pxν has a
solution for every sufficiently large ν ∈ N ′: thus, there exist
(µν
j ∇xgj(x(xν ); xν )i∈ NK(x(xν ))
−h∇x U (x(xν ); xν ) +Pm
j ⊥ gj(x(xν ); xν ) ≤ 0, j = 1, . . . , m.
(50)
Note that by (44) and the complementarity slackness in (50),
j = 0 for all j /∈ ¯J and large ν ∈ N ′. Moreover, the
µν
sequence of nonnegative multipliers {µν , (µν
j )j∈ ¯J}ν∈N ′
must be bounded, as shown next. Suppose by contradiction
0 ≤ µν
j=1 µν
11
′
that limN ′∋ν→∞ kµνk = +∞ for some {x(xν )}N ′ (possibly
over a subsequence). Dividing both sides of (50) by kµνk and
taking the limit N
∋ ν → ∞, one would get
−Pj∈ ¯J
¯¯µj∇gj(¯x) ∈ NK(¯x)
0 ≤ ¯¯µj ⊥ gj(¯x) ≤ 0, j ∈ ¯J,
Therefore, {µν , (µν
for some ¯¯µ , (¯¯µj )j∈ ¯J 6= 0, in contradiction with (2).
j )j∈ ¯J}ν∈N ′ must have a limit; let us
denote by (¯µj)j∈ ¯J such a limit (after a suitable renumeration).
Taking the limit N ′ ∋ ν → ∞ in (50), and using (45) and (46)
along with the outer semicontinuity of the mapping NK(•), we
get
(51)
−h∇U (¯x) +Pj∈ ¯J ¯µj∇gj(¯x)i ∈ NK(¯x)
0 ≤ ¯µj ⊥ gj(¯x) ≤ 0,
j ∈ ¯J.
(52)
It follows from (52) that ¯x is a stationary solution of the
original problem P.
B. Proof of Theorem 2
Proof of statement (a). It follows from Lemma 6.
Proof of statement (b). Invoking Theorem 11(b), it is suffi-
cient to show that (42) in Theorem 11 is satisfied.
By the descent lemma [21, Propo. A.24] and Step 3 of
Algorithm 1, we get:
U (xν+1) ≤ U (xν ) + γν∇U (xν )T (x(xν ) − xν )
+ (γ ν )2L∇U
2
kx(xν ) − xνk2.
Invoking (25) in Lemma 7, we obtain
γνL∇U
2 (cid:19)kx(xν ) − xνk2.
U (xν+1) ≤ U (xν ) − γν(cid:18)c U −
(53)
Since 0 < inf ν γν ≤ supν γν ≤ γmax ≤ 1 and 2c U >
γmaxL∇U , we deduce from (53) that either U (xν) → −∞
or {U (xν)} converges to a finite value and
ν→∞kx(xν ) − xνk = 0.
lim
(54)
By assumption A4, {U (xν )} is convergent and the sequence
{xν} ⊆ X [Lemma 6(iii)] is bounded. Therefore, (54) holds
true and {xν} has a limit point in X . By Theorem 11(b) and
(54), statement (b) of the theorem follows readily. Finally, by
(53), U (xν ) is a decreasing sequence: hence, no limit point
of {xν} can be a local maximum of U .
Proof of statement (c). Invoking Theorem 11(a), it is suffi-
cient to show that (41) in Theorem 11 is satisfied. Following
the same steps as in the proof of statement (b), by (53) and
γν → 0, for ν ≥ ¯ν sufficiently large, there exists a positive
constant ζ such that:
U (xν+1) ≤ U (xν ) − γνζkx(xν ) − xνk2,
which, again, by A4, leads to
lim
ν→∞
γtkx(xt) − xtk2 < +∞.
νXt=¯ν
The desired result (41) follows from (56) and P∞
ν=0 γν =
+∞. Similarly to the previous case, by (56), eventually U (xν )
(55)
(56)
12
is a decreasing sequence: thus, no limit point of {xν} can be
a local maximum of U .
Suppose now that Assumption 4 holds. By Theorem 11(b)
it is sufficient to prove that (42) holds true. For notational
simplicity, we set ∆x(xν ) , x(xν ) − xν. We already
proved that lim inf ν k∆x(xν )k = 0; therefore, (42) holds if
lim supν k∆x(xν )k = 0, as stated next.
First of all, note that, by Assumption 4, Lemma 10 and, by
consequence, Lemma 8 hold true; therefore, there exists ρ > 0
such that (cf. Lemma 8)
1
2 ,
kx(xν ) − x(xt)k ≤ ηρkxν − xtk + θρkxν − xtk
(57)
for any ν, t ≥ 1 and ρ ∈ (0, ρ], with ηρ and θρ defined in (30)
(cf. Lemma 8).
Suppose by contradiction that lim supν k∆x(xν )k > 0.
Then, there exists δ > 0 such that k∆x(xν )k > 2δ +pδ/2
for infinitely many ν, and also k∆x(xν )k < δ +pδ/2 for
infinitely many ν. Thus, there exists an infinite subset of
indices N such that, for each ν ∈ N and some iν > ν,
the following hold:
k∆x(xν )k < δ +pδ/2, k∆x(xiν )k > 2δ +pδ/2
δ +pδ/2 ≤ k∆x(xj)k ≤ 2δ +pδ/2, ν < j < iν.
Hence, for all ν ∈ N , we can write
and, in case iν > ν + 1,
(58)
(59)
(60)
δ < k∆x(xiν )k − k∆x(xν )k
≤ kx(xiν ) − x(xν )k + kxiν − xνk
≤ (1 + ηρ)kxiν − xνk + θρkxiν − xνk 1
(a)
2
(b)
≤ (1 + ηρ)(cid:16)2δ +pδ/2(cid:17)Piν −1
2(cid:16)Piν −1
t=ν γt(cid:17) 1
+θρ(cid:16)2δ +pδ/2(cid:17) 1
t=ν γt
2
,
where (a) is due to (57) and (b) comes from the triangle
inequality and the updating rule of the algorithm. It follows
from (58) and (60) that
t=ν γt
lim inf νh(1 +ηρ)(cid:16)2δ +pδ/2(cid:17)Piν −1
2(cid:16)Piν −1
2i > 0.
(61)
We now prove that k∆x(xν )k ≥ δ/2 for sufficiently large
ν ∈ N . Reasoning as in (60), we have
k∆x(xν+1)k − k∆x(xν )k ≤
+ θρ(cid:16)2δ +pδ/2(cid:17) 1
t=ν γt(cid:17) 1
(1 + ηρ)kxν+1 − xνk + θρkxν+1 − xνk 1
≤ (1 + ηρ)γνk∆x(xν )k + θρ(γν)1/2k∆x(xν )k 1
2 ,
(62)
for any given ν. For large ν ∈ N , so that (1 + ηρ)γνδ/2 +
2 < δ/2 +pδ/2, suppose by contradiction that
θρ(γνδ/2)
k∆x(xν )k < δ/2; this would give k∆x(xν+1)k < δ +pδ/2
and condition (59) (or, in case, (58)) would be violated. Then,
it must be
1
2
(64)
t=ν γt,
Using (63), we can show now that (61) is in contradiction
with the convergence of {U (xν )}. By (55), (possibly over a
subsequence) for sufficiently large ν ∈ N , we have
t=ν γtk∆x(xt)k2
U (xiν ) ≤ U (xν ) − ζPiν −1
4 Piν −1
< U (xν ) − ζ δ2
where, in the last inequality, we have used (59) (or, in case,
(58)) and (63). Thus, since U (xν) converges, (64) implies
t=ν γt = 0, in contradiction with (61).
limν∈NPiν −1
Remark 12: As we already mentioned in subsection III-A3,
in [11] it is shown that, in the specific case of a strongly
convex U , U = U , and K = Rn, one can choose γν = 1 at
every iteration and prove the stationarity of every limit point of
the sequence generated by Algorithm 1 (assuming regularity).
For completeness we sketch how this result can be readily
obtained using our framework (and actually slightly improved
on by also considering the case in which K is not necessarily
Rn). The proof is based on Theorem 11(b) and a result in
[11]. By Theorem 11(b), it is enough to show that (42) holds.
But (42) does indeed hold because of the strong convexity
of U , as shown at the beginning of Proposition 3.2 in [11].
Note that the strong convexity of U plays here a fundamental
role and that, once we remove this restrictive assumption,
things get considerably more difficult, as clearly shown by
the complexity of the proof of Theorem 2.
′
′ , λ
≥ 0
mum principle,
′(cid:17) ≥ 0.
Adding the two inequalities above and summing and subtract-
C. Proof of Lemma 3
(a) It is a consequence of Danskin's theorem [21, Prop. A.43].
(b) The statement follows from the uniform Lipschitz con-
+ with constant L∇d, which is
tinuity of bx(•; xν) on Rm
proved next. For notational simplicity, let us write bxλ ,
; xν). Defining L(x, λ) ,
bx(λ; xν ) and bxλ
′ , bx(λ
i=1(cid:16) Ui(xi; xν ) + λT gi(xi; xν)(cid:17), we have, by the mini-
PI
′ −bxλ(cid:1)T
(cid:0)bxλ
∇xL (bxλ, λ)
∇xL(cid:16)bxλ
(cid:0)bxλ −bxλ
′(cid:1)T
′(cid:17), we obtain
ing ∇xL(cid:16)bxλ, λ
c U · kbxλ −bxλ
′(cid:17)i
′ −bxλ(cid:1)Th∇xL (bxλ, λ) − ∇xL(cid:16)bxλ, λ
≤(cid:0)bxλ
′(cid:17) − ∇xL (bxλ, λ)i ,
′(cid:1)Th∇xL(cid:16)bxλ, λ
=(cid:0)bxλ −bxλ
′(cid:17). Hence, we have
convexity of L(cid:16)•, λ
mXj=1(cid:12)(cid:12)(cid:12)λ
j − λj(cid:12)(cid:12)(cid:12)k∇xgj (bxλ; xν)k
c U · kbxλ −bxλ
j − λj(cid:12)(cid:12)(cid:12) Lg =(cid:13)(cid:13)(cid:13)λ
mXj=1(cid:12)(cid:12)(cid:12)λ
− λ(cid:13)(cid:13)(cid:13)1 · Lg
≤ Lg √m(cid:13)(cid:13)(cid:13)λ
− λ(cid:13)(cid:13)(cid:13)2
(66)
where (a) follows from the uniform Lipschitz continuity of
g. The inequality above proves the Lipschitz property of
(65)
where, in the first inequality, we used the uniform strong
′k ≤
′k2
≤
(a)
,
′
′
′
′
k∆x(xν )k ≥ δ/2.
(63)
bx(•; xν ).
REFERENCES
[1] G. Scutari, F. Facchinei, L. Lampariello, and P. Song, "Parallel and
distributed methods for nonconvex optimization," in Proc. of IEEE Int.
Conf. on Acoustics, Speech, and Signal Process. (ICASSP 14), Florence,
Italy, May 4 -- 9 2014.
[2] P. Song, G. Scutari, F. Facchinei, and L. Lampariello, "D3m: Distributed
multi-cell multigroup multicasting," in Proc. of IEEE International
Conference on Acoustics, Speech, and Signal Processing (ICASSP 16),
Shangai, China, March 20-25 2016.
[3] G. Scutari, F. Facchinei, L. Lampariello, and P. Song, "Distributed
methods for nonconvex constrained multi-agent optimization− Part I:
Theory," arXiv:1410.4754, Oct. 2014.
[4] R. Byrd, J. Nocedal, and R. Waltz, "Feasible interior methods using
slacks for nonlinear optimization," Computational Opt. and Appl.,
vol. 26, no. 1, pp. 35 -- 61, 2003.
[5] A. Fiacco and G. M. Cormick, Nonlinear Programming: Sequential
Unconstrained Minimization Techniques.
John Wiley & Sons, 1968.
[6] C. Lawrence and A. Tits, "A computationally efficient feasible sequential
quadratic programming algorithm," SIAM Journal on Optimization,
vol. 11, no. 4, 2001.
[7] M. Ferris and O. Mangasarian, "Parallel variable distribution," SIAM
Journal on Optimization, vol. 4, no. 4, pp. 815 -- 832, 1994.
[8] C. Sagastizábal and M. Solodov, "Parallel variable distribution for con-
strained optimization," Computational Optimization and Applications,
vol. 22, no. 1, pp. 111 -- 131, 2002.
[9] M. Solodov, "On the convergence of constrained parallel variable
distribution algorithms," SIAM Journal on Optimization, vol. 8, no. 1,
pp. 187 -- 196, 1998.
[10] B. Marks and G. Wright, "A general inner approximation algorithm for
nonconvex mathematical programs," Operations Research, vol. 26, no. 4,
pp. 681 -- 683, 1978.
[11] A. Beck, A. Ben-Tal,
and L. Tetruashvili,
"A sequential
to
parametric
nonconvex
of Global
Optimization, vol. 47, no. 1, pp. 29 -- 51, 2010. [Online]. Available:
http://dx.doi.org/10.1007/s10898-009-9456-5
approximation method with applications
topology
convex
truss
design
problems,"
J.
[12] M. Razaviyayn, M. Hong, and Z.-Q. Luo, "A unified convergence
analysis of block successive minimization methods for nonsmooth
optimization," SIAM J. Optim., vol. 23, no. 2, pp. 1126 -- 1153, 2013.
[13] G. Scutari, F. Facchinei, P. Song, D. P. Palomar, and J.-S. Pang, "De-
composition by partial linearization: Parallel optimization of multiuser
systems," IEEE Trans. on Signal Processing, vol. 62, pp. 641 -- 656, Feb.
2014.
[14] A. Alvarado, G. Scutari, and J.-S. Pang, "A new decomposition method
for multiuser dc-programming and its applications," IEEE Trans. on
Signal Processing, vol. 62, pp. 2984 -- 2998, March 2014.
[15] T. Quoc and M. Diehl, "Sequential convex programming methods for
solving nonlinear optimization problems with dc constraints," ArXiv.org
http://adsabs.harvard.edu/abs/2011arXiv1107.5841D, 2011.
[16] C. Fleury, "CONLIN: an efficient dual optimizer based on convex
approximation concepts," Structural Opt., vol. 1, no. 2, pp. 81 -- 89, 1989.
[17] K. Svanberg, "The method of moving asymptotes -- a new method for
structural optimization," Int. Jour. for Numerical Methods in Engineer-
ing, vol. 24, no. 2, pp. 359 -- 373, 1987.
[18] -- -- , "A class of globally convergent optimization methods based on
conservative convex separable approximations," SIAM Jour. on Opt.,
vol. 12, no. 2, pp. 555 -- 573, 2002.
[19] G. Scutari, F. Facchinei, L. Lampariello, and S. Sardellitti, "Parallel and
distributed methods for nonconvex optimization−Part II: Applications,"
IEEE Trans. on Signal Processing, (submitted). [Online]. Available:
TBD
[20] D. P. Palomar and M. Chiang, "Alternative distributed algorithms for
network utility maximization: Framework and applications," IEEE Trans.
on Automatic Control, vol. 52, no. 12, pp. 2254 -- 2269, Dec. 2007.
[21] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computa-
tion: Numerical Methods, 2nd ed. Athena Scientific Press, 1989.
[22] K. Phan, S. Vorobyov, C. Telambura, and T. Le-Ngoc, "Power control
for wireless cellular systems via D.C. programming," in IEEE/SP 14th
Workshop on Statistical Signal Processing, 2007, pp. 507 -- 511.
[23] H. Al-Shatri and T. Weber, "Achieving the maximum sum rate using
D.C. programming in cellular networks," IEEE Trans. Signal Process.,
vol. 60, no. 3, pp. 1331 -- 1341, March 2012.
13
[24] N. Vucic, S. Shi, and M. Schubert, "DC programming approach for
resource allocation in wireless networks," in 8th International Sympo-
sium on Modeling and Optimization in Mobile, Ad Hoc and Wireless
Networks (WiOpt), 2010, pp. 380 -- 386.
[25] M. Chiang, C. Tan, D. Palomar, D. O. Neill, and D. Julian, "Power
control by geometric programming," IEEE Trans. Wireless Commun.,
vol. 6, no. 7, pp. 2640 -- 2651, July 2007.
[26] A. Khabbazibasmenj, F. Roemer, S. Vorobyov, and M. Haardt, "Sum-
rate maximization in two-way AF MIMO relaying: Polynomial time
solutions to a class of DC programming problems," IEEE Trans. Signal
Process., vol. 60, no. 10, pp. 5478 -- 5493, 2012.
[27] Y. Xu, T. Le-Ngoc, and S. Panigrahi, "Global concave minimization for
optimal spectrum balancing in multi-user DSL networks," IEEE Trans.
Signal Process., vol. 56, no. 7, pp. 2875 -- 2885, July 2008.
[28] P. Tsiaflakis, M. Diehl, and M. Moonen, "Distributed spectrum man-
agement algorithms for multiuser dsl networks," IEEE Trans. Signal
Process., vol. 56, no. 10, pp. 4825 -- 4843, Oct. 2008.
[29] D. Schmidt, C. Shi, R. Berry, M. Honig, and W. Utschick, "Distributed
resource allocation schemes: Pricing algorithms for power control and
beamformer design in interference networks," IEEE Signal Process.
Mag., vol. 26, no. 5, pp. 53 -- 63, Sept. 2009.
[30] R. Mochaourab, P. Cao, and E. Jorswieck, "Alternating rate profile
optimization in single stream mimo interference channels," arXiv
preprint, 2013. [Online]. Available: http://arxiv.org/abs/1303.4683
[31] J. Qiu, R. Zhang, Z.-Q. Luo, and S. Cui, "Optimal distributed beam-
forming for miso interference channels," IEEE Trans. Signal Process.,
vol. 59, no. 11, pp. 5638 -- 5643, Nov. 2011.
[32] S.-J. Kim and G. B. Giannakis, "Optimal resource allocation for MIMO
ad hoc cognitive radio networks," IEEE Trans. on Information Theory,
vol. 57, no. 5, pp. 3117 -- 3131, May 2011.
[33] Y. Zhang, E. Dall'Anese, and G. B. Giannakis, "Distributed optimal
beamformers for cognitive radios robust to channel uncertainties," IEEE
Trans. Signal Process., vol. 60, no. 12, pp. 6495 -- 6508, Dec. 2012.
[34] Y. Yang, G. Scutari, and D. Palomar, "Robust mimo cognitive radio un-
der interference temperature constraints," IEEE J. Sel. Areas Commun.,
vol. 31, no. 11, pp. 2465 -- 2482, Nov. 2013.
[35] N. D. Sidiropoulos, T. N. Davidson, and Z.-Q. Luo, "Transmit beam-
forming for physical-layer multicasting," IEEE Trans. on Signal Pro-
cessing, vol. 54, no. 6, pp. 2239 -- 2251, Jun. 2006.
[36] E. Karipidis, N. Sidiropoulos, and Z.-Q. Luo, "Quality of Service and
Max-Min Fair Transmit Beamforming to Multiple Cochannel Multicast
Groups," IEEE Trans. on Signal Processing, vol. 56, no. 3, pp. 1268 --
1279, March 2008.
[37] M. Chiang, S. Low, A. Calderbank, and J. Doyle, "Layering as optimiza-
tion decomposition: A mathematical theory of network architectures,"
Proc. IEEE, vol. 95, no. 1, pp. 255 -- 312, Jan. 2007.
[38] M. Chiang, P. Hande, T. Lan, and C. W. Tan, Power Control in Wireless
Cellular Networks. Now Publishers Inc, 2008.
[39] R. Rockafellar and J. Wets, Variational Analysis. Springer, 1998.
[40] F. Facchinei and J.-S. Pang, "Exact penalty functions for generalized
Nash problems," G. Di Pillo and M. Roma (eds): Large Scale Nonlinear
Optimization, pp. 115 -- 126, 2006.
[41] D. Bertsekas, A. Nedic, and A. Ozdaglar, Convex Analysis and Opti-
mization. Belmont, MA, USA: Athena Scientific, 2003.
[42] M. Su and H.-K. Xu, "Remarks on the gradient-projection algorithm," J.
on Nonlinear Analysis and Optimization, vol. 1, pp. 35 -- 43, July 2010.
[43] Y. Nesterov, Introductory Lectures on Convex Optimization: A Basic
Course (Applied Optimization). Springer, 2004.
[44] -- -- , "Smooth minimization of non-smooth functions," Mathematical
Programming, vol. 103, no. 1, pp. 127 -- 152, 2005. [Online]. Available:
http://dx.doi.org/10.1007/s10107-004-0552-5
[45] D. Bertsekas, A. Nedic, and A. Ozdaglar, Convex Analysis and Opti-
mization. Athena Scientifica, 2003.
[46] F. Facchinei and J.-S. Pang, Finite-Dimensional Variational Inequalities
and Complementarity Problems. Springer, 2003.
[47] S. Dafermos, "Sensitivity analysis in variational inequalities," Mathe-
matics of Operations Research, vol. 13, no. 3, pp. 421 -- 434, 1988.
[48] N. Yen, "Hölder continuity of solutions to a parametric variational
inequality," Applied mathematics & optimization, vol. 31, no. 3, pp.
245 -- 255, 1995.
[49] R. Rockafellar, "Lipschitzian properties of multifunctions," Nonlinear
Analysis: Theory, Methods & Applications, vol. 9, no. 8, pp. 867 -- 885,
1985.
|
1604.04736 | 1 | 1604 | 2016-04-16T12:11:02 | Intra-Team Strategies for Teams Negotiating Against Competitor, Matchers, and Conceders | [
"cs.MA",
"cs.AI",
"cs.GT"
] | Under some circumstances, a group of individuals may need to negotiate together as a negotiation team against another party. Unlike bilateral negotiation between two individuals, this type of negotiations entails to adopt an intra-team strategy for negotiation teams in order to make team decisions and accordingly negotiate with the opponent. It is crucial to be able to negotiate successfully with heterogeneous opponents since opponents' negotiation strategy and behavior may vary in an open environment. While one opponent might collaborate and concede over time, another may not be inclined to concede. This paper analyzes the performance of recently proposed intra-team strategies for negotiation teams against different categories of opponents: competitors, matchers, and conceders. Furthermore, it provides an extension of the negotiation tool Genius for negotiation teams in bilateral settings. Consequently, this work facilitates research in negotiation teams. | cs.MA | cs | Intra-Team Strategies for Teams Negotiating
Against Competitor, Matchers, and Conceders
Victor Sanchez-Anguix1, Reyhan Aydogan2, Vicente Julian1, and Catholijn M.
Jonker2
1 Universitat Polit`ecnica de Val`encia,
Departamento de Sistemas Inform´aticos y Computaci´on
Cami de Vera s/n, 46022, Valencia, Spain
{sanguix,vinglada}@dsic.upv.es
2 Interactive Intelligence Group
Delft University of Technology
Delft, The Netherlands
{R.Aydogan,C.M.Jonker}@tudelft.nl
Abstract. Under some circumstances, a group of individuals may need
to negotiate together as a negotiation team against another party. Unlike
bilateral negotiation between two individuals, this type of negotiations
entails to adopt an intra-team strategy for negotiation teams in order to
make team decisions and accordingly negotiate with the opponent. It is
crucial to be able to negotiate successfully with heterogeneous opponents
since opponents' negotiation strategy and behavior may vary in an open
environment. While one opponent might collaborate and concede over
time, another may not be inclined to concede. This paper analyzes the
performance of recently proposed intra-team strategies for negotiation
teams against different categories of opponents: competitors, matchers,
and conceders. Furthermore, it provides an extension of the negotiation
tool Genius for negotiation teams in bilateral settings. Consequently,
this work facilitates research in negotiation teams.
Keywords: Negotiation Teams, Collective Decision Making, Agreement
Technologies
1
Introduction
A negotiation team is a group of two or more interdependent individuals that
join together as a single negotiation party because they share some common
goals related to the negotiation at hand [30,5]. This kind of party participates
in many real life situations like the negotiation between a married couple and a
house seller, the negotiation between a group of traveling friends and a booking
agency, and the negotiation between two or more organizations. Despite acting
as a single party, most of the time negotiation teams cannot be considered as
a unitary player. As a matter of fact, team members may have different and
conflicting preferences that need to be conciliated when making a team decision
6
1
0
2
r
p
A
6
1
]
A
M
.
s
c
[
1
v
6
3
7
4
0
.
4
0
6
1
:
v
i
X
r
a
2
V. Sanchez-Anguix, R. Aydogan, V. Julian, C. M. Jonker
regarding the negotiation. Agent-based negotiation teams (ABNT) constitutes
a novel topic of research in automated negotiation, where efforts in the last few
years focused mostly on bilateral and multiparty negotiations with unitary play-
ers [8,13,9]. Mechanisms that allow ABNT to take decisions on the negotiation
process, namely intra-team strategies or team dynamics [24,23], are needed in
order to support multi-agent systems for complex applications like group travel
markets, group buying in electronic commerce, and negotiations between agent
organizations (e.g., organizational merging). An intra-team strategy for a spe-
cific negotiation protocol (e.g., alternating bilateral negotiation protocol) defines
what decisions are taken by the negotiation team, and how and when those de-
cisions are taken.
Although there are some studies investigating negotiation among team mem-
bers [29], automated negotiation between a team and an opponent is open to re-
search. Sanchez-Anguix et al. have proposed several intra-team strategies [24,23]
for ABNT following the alternating-offers protocol in a bilateral setting. The
proposed intra-team strategies have been studied under different environmental
conditions to assess the most appropriate intra-team strategy with respect to
the given environmental setting [24]. However, several assumptions regarding
the opponent exist. For instance, it is assumed that the opponent employs a
time-based concession tactics such as Boulware or Conceder[7] in a cooperative
context. Nevertheless, these assumptions might become inconsistent with some
opponents in open and dynamic environment. For instance, an opponent may
adopt a strategy like "take it or leave it" while another opponent may choose
to observe other negotiating agent's behavior and concede accordingly. An im-
mediate question is which intra-team strategies will negotiate well against other
types of opponents different than those using time-based tactics.
Without a doubt, an opponent's negotiation attitude may affect on the per-
formance of intra-team strategies. Opponent's behavior is not limited to classic
time-based concession strategies. Baarslag et al. classify the negotiation strate-
gies according to their negotiation behavior against the opponent into four cate-
gories [3]. These are inverters, conceders, competitors and matchers. Conceders
always concede regardless of the opponent's strategy, while competitors do not
yield independently of the behavior shown by opponents. A matcher mimics its
opponent's behavior while inverter inverts it. When the opponent concedes, the
matcher would concede accordingly while the inverter would not. Based on this
classification, we investigate how intra-team strategies proposed in the literature
perform against opponents belonging to different families of negotiation strate-
gies. To do this, we extend Genius [14] to allow negotiation teams and enable it
to perform bilateral negotiations between a team (a group of agents) and an in-
dividual agent. The contributions of this paper do not solely focus on the study
of intra-team strategies' performance against different types of opponent, but
we also describe how Genius has been modified to support such negotiations.
This extension will allow researchers to (i) design and test domain independent
intra-team strategies, which is desirable given the increasing number of applica-
tion domains for automated negotiation; (ii) engage negotiation teams in open
Intra-Team Strategies Against Competitor, Matchers, and Conceders
3
environments where any kind of opponent behavior is possible; (iii) make use of
a wide repository of negotiation domains, utility functions, and automated nego-
tiators; (iv) focus on the design of intra-team strategies, while leaving simulation
aspects to be governed by Genius.
Our contributions are two-folds. First, we extend Genius to support ABNT;
thus Genius can facilitate research on ABNT. Second, we analyze the per-
formance of different intra-team strategies proposed by Sanchez-Anguix et al.
against different types of heterogeneous opponents. The rest of this paper is
organized as follows. First, we present our general framework. After that, we
briefly introduce the intra-team strategies analyzed in this paper. Then, we de-
scribe how the extension has been included inside the Genius framework. Then,
we describe how the experiments were carried out and present and discuss the
results of the experiments. Finally, we describe our future work and briefly con-
clude this work.
2 General Framework
In our framework, one negotiation team is involved in a negotiation with an
opponent. Independently of whether or not the other party is also a team, both
parties interact with each other by means of the alternating-offers protocol. Team
dynamics or intra-team strategies define what decisions have to be taken by a
negotiation team, how those decisions are taken, and when those decisions are
taken. In a bilateral negotiation between a team and an opponent, the decisions
that must be taken are which offers are sent to the opponent, and whether or not
opponent's offers are accepted. A general view of our framework is represented
in Fig. 1. Dashed lines depict communications inside the team, while others
represent communications with the opponent.
A team A is formed by a team mediator T MA and team members ai. The
team mediator communicates with the other party following the alternating of-
fers protocol and team members communicate with the team mediator. Com-
munications between the team and the opponent are carried out by means of
the team mediator. This mediator sends team decisions to the opponent and
receives, and later broadcasts, decisions from the opponent to team members.
Thus, the fact that the opponent is communicating with a team is not known
by the opponent, which only interacts with the trusted mediator. In this frame-
work, the mechanisms employed by the team to decide on which offers to send
and whether or not accept offers are carried out during the negotiation process
itself. How these decisions are taken depends on the specific intra-team strategy
that is implemented by the team and the team mediator. Each team media-
tor can implement its own intra-team protocol to coordinate team members as
long as team members know how to play such intra-team protocol. It should be
noted that we assume that team membership remains static during the nego-
tiation process. Thus, members do not leave/enter the team as the negotiation
is being carried out. It is acknowledged that team members may leave or join
4
V. Sanchez-Anguix, R. Aydogan, V. Julian, C. M. Jonker
Fig. 1. This picture shows our general negotiation framework.
the group in certain specific situations. However, membership dynamics is not
considered in this article, and it is designated as future work.
3
Intra-team strategies
In this section we briefly describe the intra-team strategies proposed by Sanchez-
Anguix et al. [24,23] that are the focus of our study. These intra-team strate-
gies have been selected according to the minimum level of unanimity that they
are able to guarantee regarding each team decision: no unanimity guaranteed
(representative), majority/plurality (similarity simple voting), semi-unanimity
(similarity Borda voting), and unanimity (full unanimity mediated).
3.1 Representative (RE)
The intra-team protocol employed by the representative intra-team strategy is
the simplest possible strategy for a negotiation team. Basically, one of the team
members is selected as representative and acts on behalf of the team. Interactions
among team members are non-existent, and therefore, every decision is taken by
the representative according to its own criterion. Obviously, the performance
of the team will be determined by the similarity among the team members'
utility functions and the negotiation skills of the representative. It is expected
that if team members' utility functions are very similar and the representative
negotiation strategy is appropriate, the team performance will be reasonably
good.
The mechanism used by the team to select its representative may vary de-
pending on the domain: trust, past experiences, rational voting based on who
Intra-Team Strategies Against Competitor, Matchers, and Conceders
5
Fig. 2. Screenshot showing the menu for configuring a team negotiation session.
is more similar to oneself, etc. Since Genius is a general simulation framework,
a random team member is selected as representative. This random team mem-
ber will receive the messages from the opponent and act accordingly by sending
an offer/counter-offer or accepting the opponent's offer. Generally, any Genius
agent that knows how to play the alternating bilateral game can act as repre-
sentative.
3.2 Similarity Simple Voting (SSV)
The similarity simple voting intra-team strategy relies on voting processes to
decide on which offer is proposed and whether or not the opponent's offer is ac-
cepted. Being based on voting, the strategy requires the action of team members
and team coordination by means of a mediator. The intra-team strategy goes as
follows every round:
-- Accept/Reject opponents' offer: The team mediator receives an offer from
the other party. Then, the mediator broadcasts this offer to the team mem-
bers, indicating that it comes from the opponent party. The team mediator
opens a voting process, where each team member should respond to the me-
diator with an Accept or Reject depending on the acceptability of the offer
6
V. Sanchez-Anguix, R. Aydogan, V. Julian, C. M. Jonker
from the point of view of the team member. The team mediator gathers the
responses from every team member and applies a majority rule. If the num-
ber of Accept actions received from team members is greater than half the
size of the team, the offer is accepted and the corresponding Accept action is
sent to the opponent. Otherwise, the team mediator starts the offer proposal
mechanism.
-- Offer Proposal: Each team member is allowed to propose an offer to be sent.
This offer is communicated solely to the mediator, who will make public
the offers, and start a voting process. In this voting process, each team
member must state whether or not he considers it acceptable for each of
the offers proposed. For instance, if three different offers (x1, x2, x3) have
been proposed, the team members should state about the acceptability of
the three of them: e.g., (yes, no, yes). The mediator applies a plurality rule
to determine the most supported offer, which is the one that is sent to the
opponent.
Standard team members for SSV employ time tactics to decide on the accep-
tance/rejection of the opponent's offer, and the offer proposal. More specifically,
the current aspiration for team members follows the next expression [13]:
(cid:18) t
(cid:19)1/βai
sai (t) = 1 − (1 − RUai)
TA
.
(1)
where sai(t) is the utility demanded by the team member at time t, RUai is
the reservation utility for the agent, TA is the team's deadline, and βai is the
concession speed. When βai < 1 we have a classic Boulware strategy, when
βai = 1 we have a linear concession, and when βai > 1 we have a conceder
strategy. On the one hand a team members considers an opponent's offer as
acceptable if it reports a utility which is greater than or equal to sai(t). On
the other hand, a team member considers an offer proposed by a teammate
acceptable if it reports a utility which is greater than or equal to the utility
of the offer that he proposed to the team in the same round. As for the offer
proposed to the team, team members attempt to select the offer from the iso-
utility curve which is closer to the last opponent's offer and the offer sent by the
team in the last round (similarity heuristic based on Euclidean distance).
3.3 Similarity Borda Voting (SBV)
This intra-team strategy attempts to guarantee a higher level of unanimity by
incorporating voting mechanisms that select broadly accepted candidates like
Borda count [19], and unanimity voting processes. The intra-team strategy pro-
ceeds as SSV with the following differences:
-- Accept/Reject opponents' offer: Instead of using a majority voting to decide
whether or not the opponent's offer is accepted, the team mediator opens
a unanimity voting. Hence, an opponent offer is accepted if and only if the
number of Accept actions is equal to the number of team members. Other-
wise, the team mediator stars the offer proposal mechanism.
Intra-Team Strategies Against Competitor, Matchers, and Conceders
7
-- Offer Proposal: Each team member is allowed to propose an offer to be sent
by the same mechanisms described in SSV. Then, the team mediator makes
the team members privately score each proposal by means of a Borda count.
The team members give a different score to each offer from the set [0,A−1],
where A is the number of proposals received by the team mediator. Once
the scores have been received by the team mediator, it selects the candidate
offer that received the highest sum of scores and it is sent to the opponent.
Standard team members are governed by an individual time-based conces-
sion tactic like the one in Equation 1. Similarly to SSV, an opponent's offer is
acceptable for a team member if the utility that it reports is equal to or greater
than its current utility demanded sai(t). When scoring team offers, each team
member privately ranks candidates in descending order of utility and then it
assigns a score to each offer which is equal to the number of candidates minus
the position of the offer in the ranking.
3.4 Full Unanimity Mediated (FUM)
FUM is capable of reaching unanimous decisions as long as the negotiation do-
main is composed by predictable issues whose type of valuation function is the
same for team members (e.g., either monotonically increasing or decreasing).
The type of unanimity that it is capable of guaranteeing is strict in the sense
that every decision reports a utility which is greater than or equal to the current
aspiration level of each team member [23]. The team mediator governs intra-team
interactions as follows:
-- Accept/Reject opponents' offer: The interaction protocol followed by the
team in this decision is the same as the one presented in SSV. However,
the decision rule applied by the team mediator in this case is unanimity.
Therefore, an opponent offer is only accepted if it is acceptable to each team
member.
-- Offer Proposal: Every team member is involved in the offer proposal, which
consists in an iterated process where the offer is built attribute per attribute.
The mediator starts the iterated building process with an empty partial offer
(no attribute is set). Then, he selects the first attribute to be set following
an agenda. The mediator makes public the current partial offer and the at-
tribute that needs to be set. Each active team member states privately to
the mediator the value that he wants for the requested issue. When all of
the responses have been gathered, the mediator aggregates the values sent
by team members using the max (monotonically increasing valuation func-
tion) or min (monotonically decreasing valuation function) and makes public
among active team members the new partial offer. Since it is assumed that
team members share the same type of valuation function for predictable at-
tributes, increasing the welfare of one of the members results in other team
members increasing their welfare or staying at the same utility. Then, each
active team member must evaluate the partial offer and state if the partial
8
V. Sanchez-Anguix, R. Aydogan, V. Julian, C. M. Jonker
offer is acceptable at the current state (Accept or Reject action). Those team
members that respond with an Accept action are no longer considered active
in the current construction process. The team mediator selects the next at-
tribute in the agenda and follows the same process until all of the attributes
have been set or until there are no more active team members (the rest of
the attributes are maximized to match the opponent's preferences). It should
be said that the agenda of attributes is set by the mediator observing the
concessions from the opponent in the first interactions. Following a rational
criteria, the opponent should have conceded less in the most important at-
tributes in the first negotiation rounds. The amount of concession in each
attribute during the first rounds is summed up and an attribute agenda is
inferred at each round. The first attributes in the agenda are those inferred
as less important for the opponent (more amount of concession), whereas
the last attributes in the agenda are those considered more important for
the opponent. The heuristic behind this agenda and iterated building pro-
cess is attempting to satisfy team members first with those attributes less
important for the opponent.
As for the standard team member behavior, team members have their de-
mands governed by an individual time-based concession tactic like the one in
Equation 1. In the iterated building process, each team member requests the at-
tribute value which, given the current partial offer, makes the partial offer closer
to its current demands sai(t). Additionally, a partial offer is acceptable when,
considering only those attributes that have been set, the partial offer reports a
partial utility which is greater than or equal to the current demands of the team
member. Each team members considers an opponent offer acceptable when it
has a utility which is greater or equal than its current demands sai (t).
4
Implementation in Genius
Genius [14] is a well-known negotiation simulation framework. It supports sim-
ulation of sessions and tournaments based on bilateral negotiations. Users are
able to design their own agents and test them against a wide variety of dif-
ferent agents designed by the community. The framework provides information
critical for analysis (e.g., utility, Pareto optimality, etc.) which is extremely use-
ful for research tasks. Moreover, the use of Genius as a testbed for bilateral
negotiations is testified by its use in the annual automated negotiating agent
competition (ANAC) [2]. The ANAC competition provided Genius with a large
repository of agents. The repository of available agents contains conceder, in-
verter, matcher, and competitor agents. The integration of ABNT in Genius
additionally facilitates the following objectives:
-- The framework includes several negotiation domains and utility functions
for test purposes. Even though most of these domains are thought for bilat-
eral negotiations with unitary players, it is possible to add new negotiation
domains and utility functions in an easy way. In fact, we are in the process of
Intra-Team Strategies Against Competitor, Matchers, and Conceders
9
adding new team negotiation domains (i.e., advanced hotel group booking)
besides the one employed for the experiments of this paper (i.e., hotel group
booking, see Section 5.2).
-- The use of Genius in ANAC has provided with wide variety of conceder,
matcher, inverter and competitor opponents. Previous research in ABNT
had only considered opponents with time-based tactics [24].
-- Current research in ABNT has only considered team members following the
same kind of homogeneous behavior inside the intra-team strategy, which
may not be the case in some open environments. Due to its open nature,
Genius may be able to simulate ABNT whose team members are heteroge-
neous since they have been designed by different scholars.
-- Genius is a consolidated testbed among the agent community. Thus, the
inclusion of ABNT inside Genius can facilitate research on ABNT by other
scholars, and even give room to a future negotiating competition involving
teams.
-- Researchers can either design new team members for teams following the
intra-team protocols included in the framework (i.e., team mediators), or
they can design new intra-team protocols and team members.
In order to implement negotiation teams, two new classes have been intro-
duced in Genius: TeamMediator and TeamMember. These two classes can be
extended by users to include new intra-team strategies and types of team mem-
bers in the system. Next, we depict the main traits of these classes, and how
they can be used to include new features in Genius:
-- TeamMember : Team members extend the Agent class, so they have all of
their methods available. Actions that come from the opponent party are
received by the ReceiveMessage method, whereas actions that come from
the team mediator are received in the
ReceiveTeamMessage method. The method chooseAction is used to decide
the agent's action independently of whether or not the next action involves
communications with the team mediator or the opponent.
-- TeamMediator : The team mediator is the agent that communicates with the
opponent party, and transmits opponent's decisions to the team members.
Thus, it has access to the public interface of all of the team members. De-
pending on the kind of intra-team strategy, the mediator also coordinates
other processes like voting mechanisms, offer proposal mechanisms, and so
forth. As the Agent class, it receives communications from opponent by the
ReceiveMessage method. In the chooseAction method, the mediator can ei-
ther directly send a decision to the opponent, or communicate with team
members to decide the next action to be taken. When interacting with team
members, the mediator uses the ReceiveTeamMessage method in team mem-
bers' API to send messages to team members. Team members can respond to
the mediator with the chooseAction method in their public API. The Team-
Mediator class is completely flexible, as the only mandatory action is receiv-
ing opponent's decisions and sending decisions to the other party. Therefore,
10
V. Sanchez-Anguix, R. Aydogan, V. Julian, C. M. Jonker
any kind of mediated communication protocol can be implemented extending
the TeamMediator class.
Of course, it should be noted that team members and team mediators are
tightly coupled. For a team member to participate in a negotiation team gov-
erned by a specific mediator, the team members should know the intra-team
communication protocol implemented by such mediator.
Genius provides several measures to assess the quality of negotiating agents.
The current version of Genius is capable of running team negotiation sessions
between two parties and provide online information about the minimum utility
of team members, the average utility of team members, the maximum utility of
team members, the joint utility of team members, current round, and current
negotiation time for each offer exchanged between both parties. A screenshot of
the environment being configured for a team negotiation session can be observed
in Fig. 2. In the upper part of the menu, the user can select the intra-team
strategies to be used by each party, whereas the user add and remove team
members for each party in the lower part of the menu.
5 Experiments and Results
As stated in the Section 1, one of the purposes of this paper is assessing the per-
formance of intra-team strategies against negotiation strategies different from
classic time-based tactics. With that purpose, we tested RE, SSV, SBV and
FUM against agents from the ANAC 2010 competition who have been previ-
ously classified into competitors, conceders, and matchers3 [3]. First, we briefly
describe the agents that we selected from the agent competition to represent the
different families of negotiation strategies. Then, we introduce the negotiation
domain used for the experiments. After that, we describe how the experiments
were carried out, and, finally, we show and analyze the results of the experiments.
5.1 ANAC 2010 Agents
In this section we present the different ANAC 2010 agents employed for our
experiments. These agents pertain to three of the four categories presented pre-
viously: matchers, conceders, competitors.
-- IAMHaggler & IAMCrazzyHaggler [32,2]: On the one hand, IAMCrazyHag-
gler is basically a take-or-leave it agent that proposes offers over a high
threshold. The only aspect taken into consideration for accepting an offer
is the utility of such offer, and not time. Due to this behavior, the experi-
ments carried out in [3] classified IAMCrazyHaggler as the most competitive
strategy in the ANAC 2010 competition. On the other hand, IAMHaggler
is a much more complicated agent. It employs Bayesian learning and non-
linear regression to attempt to model the opponent party and updates its
3 Because of the technical inconsistencies, we could not use ANAC's inverters directly
in our settings. Thus, they are not included in this analysis.
Intra-Team Strategies Against Competitor, Matchers, and Conceders
11
acceptance threshold based on information like time, the model of the oppo-
nent, and so forth. It was classified as a competitor agent in the experiments
carried out in [3]
-- Agent Smith [31,2]: This agent is a conceder agent [3] that starts by demand-
ing the highest utility for himself and slowly concedes to attempt to satisfy
the preferences of the opponent by means of a learning heuristic. When the
timeline is approaching two minutes, it proposes the best offer received up
until that moment in an effort to finish the negotiation.
-- Agent K [11,2]: This agent was the winner of the ANAC 2010 competition. It
adjusts its aspirations (i.e., target utility) in the negotiation process consid-
ering an estimation of the maximum utility that will be offered by the other
party. More specifically, the agent gradually reduces its target utility based
on the average utility offered by the opponent and its standard deviation. If
an offer has been proposed by the opponent that satisfies such threshold, it
is sent back since, rationally, it should be also good enough for the opponent.
In [3], Agent K was classified as a competitor agent.
-- Nice Tit-for-Tat [10,3,4]: This strategy is a matcher agent from the 2011
ANAC competition that reciprocates the other party's moves by means of a
Bayesian model of the other party's preferences. According to the Bayesian
model, the Nice Tit-for-Tat agent attempts to calculate the Nash point and
it reciprocates moves by calculating the distance of the last opponent offer to
the aforementioned point. When the negotiation time is reaching its deadline,
the Nice TFT agent will wait for an offer that is not expected to improve in
the remaining time and accept it in order to secure an agreement.
5.2 Test Domain: Hotel Group Booking
A group of friends who have decided to spend their holidays together has to
book accommodation for their stay. Their destination is Rome, and they want
to spend a whole week. The group of agents engages in a negotiation with a
well-known hotel in their city of destination. Both parties have to negotiate the
following issues:
-- Price per person (pp): The price per person to pay.
-- Cancellation fee per person (cf ): The fee that should be paid in case that
the reservation is cancelled.
-- Full payment deadline (pd): It indicates when the group of friends has to
pay the booking.
-- Discount in bar (db): As a token of respect for good clients, the hotel offers
nice discounts at the hotel bar.
In our experimental setup, preference profiles are represented by means of addi-
tive utility functions in the form:
Upi (X) = wpi,1Vpi,1 (x1) + ... + wpi,nVpi,n (xn) .
(2)
where wpi,j is the weight given by agent pi to attribute j, Vpi,j is the valuation
function for attribute j, and xj is the value of attribute j in the offer X. The
12
V. Sanchez-Anguix, R. Aydogan, V. Julian, C. M. Jonker
wpd wcf wpd wdb
a1 0.5 0.1 0.05 0.35
a2 0.25 0.25 0.25 0.25
a3 0.30 0.50 0.05 0.15
op 0.10 0.50 0.25 0.15
Table 1. Preference profiles used in the experiments. ai represents team members and
op represents the opponent.
domain of the attribute values is continuous and scaled to [0,1]. It should be
noted that all of the team members share the same type of monotonic valuation
function for the attributes (monotonically increasing for payment deadline and
discount, and decreasing for price and cancellation fee) so that there is potential
for cooperation among team members. Despite this, team members give different
weights to the negotiation issues. The type of valuation function for the opponent
is the opposite type (increasing for price and cancellation fee, and decreasing for
the payment deadline and the discount) and weights may be different too. The
preference profiles of the agents can be found in Table 1. Even though SSV, SBV,
and RE are able to handle other types of domain where unpredictable attributes
are present, we only use domains with predictable attributes in our analysis
because FUM does not support domains having unpredictable attributes.
5.3 Experimental Setting
In order to evaluate the performance of the intra-team strategies introduced in
Section 3, we set up a negotiation team consisting of three members that nego-
tiates with each ANAC agent presented in Section 5.1. We tested the intra-team
strategies with different parameters' configurations: the FUM strategy where
the concession speed of each team member is drawn from the uniform distribu-
tion βai = U [0.5, 0.99] (FUM Boulware or FUM B) or βai = U [0.01, 0.4] (FUM
Very Boulware or FUM VB), the SSV strategy where βai = U [0.5, 0.99] (SSV
Boulware or SSV B) or βai = U [0.01, 0.4] (SSV Very Boulware or VB), the SBV
strategy where βai = U [0.5, 0.99] (SBV Boulware or SBV B) or βai = U [0.01, 0.4]
(SBV Very Boulware or VB), and the representative approach employing Agent
K as the negotiation strategy (RE K). Both parties have a shared deadline
T = 180 seconds. If the deadline is reached and no final agreement has been
found, both parties get a utility equal to 0.
In the experiments, each intra-team strategy was faced against each ANAC
agent 10 different times to capture stochastic differences in the results. Out of
those 10 repetitions, half of the times the initiating party was the team, and the
other half the initiating party was the ANAC agent. We gathered information on
the average utility of team members in the final agreement, and the joint utility
of both parties (product of utilities of team members and opponent). A one-way
ANOVA (α = 0.05) and a post-hoc analysis with Tukey's test was carried out
to assess the differences in the averages.
Intra-Team Strategies Against Competitor, Matchers, and Conceders
13
Competitive
Crazy Haggler K
FUM B 0.19 0.38
FUM VB 0.16 0.42
0.26
0.00
RE K
0.14
SSV B
0.34
0.36
SSV VB 0.08
0.35
SBV B 0.14
SBV VB 0.13
0.39
Matcher Conceder
TFT
0.29 0.72
0.65 0.72
0.57 0.70
0.36 0.45
0.57 0.44
0.31 0.49
0.59 0.50
Smith
0.68
0.97
0.86
0.57
0.98
0.55
0.98
Table 2. The table shows the average of the average utility for team members in the
final agreement. Crazy: IAMCrazyHaggler, Haggler: IAMHaggler,K: Agent K, TFT:
Nice Tit-for-Tat, Smith: Agent Smith, B: β = U [0.5, 0.99], VB: β = U [0.01, 0.49]
5.4 Results
One of the goals of this paper is identifying which intra-team strategies work
better against different opponents. Therefore, we start by analyzing the results
for the average utility of team members in Table 2. The results in bold fond
indicate which intra-team strategy obtains statistically better results according
to ANOVA (α = 0.05) and post-hoc analysis with Tukey's test. As expected,
all of the intra-team strategies, especially when their concession speed is very
boulware, get higher average utility for team members while negotiating with an
opponent employing a conceder strategy like Agent Smith than while negotiating
with competitive opponents like AgentK, IAMHaggler, and IAMCrazyHaggler.
This result supports the observation of [3] that a successful negotiating agent, if
we only consider a single negotiation with the opponent (short term relationship),
should behave competitively, especially against cooperative strategies.
When the opponent is a conceder (Agent Smith), we observe that, in our
experiments, the best intra-team strategies are those that wait as much as pos-
sible to concede and exploit the opponent. We refer to FUM, SBV, and SSV
strategies employing a very Boulware time tactic (FUM VB, SBV VB, and SSV
VB). FUM VB, SBV VB, and SSV VB statistically get the same average for the
average utility of team members4. This can be explained due to the fact that
the conceder agent has fully conceded before FUM, SBV, and SSV VB have
started to concede. Since a concession from the opponent generally results in all
of the team members increasing their welfare, these intra-team strategies per-
form similarly even though they ensure different levels of unanimity regarding
team decisions. A representative using agent K also performs reasonably well
due to the same reason. However, since only one of the team members takes
decisions, it may not reach an average utility comparable to the ones obtained
by FUM VB, SBV VB, and SSV VB.
When the opponent is a matcher, it is observed that employing FUM strate-
gies (B and VB) and a representative strategy with a competitor representa-
4 One-way ANOVA alpha=0.05 and a post-hoc Tukey test was carried out to support
our claims.
14
V. Sanchez-Anguix, R. Aydogan, V. Julian, C. M. Jonker
tive (RE K) results in higher average utility for the team than employing SSV
and SBV strategies (B and VB). According to the one-way ANOVA test, the
performances of FUM strategies (B and VB) and RE K are statistically and
significantly better than those of SSV and SBV strategies (B and VB). The fact
that SSV and SBV do not guarantee unanimity regarding team decisions has an
important impact on the average utility of team members when faced against
Nice Tit-for-Tat. There are no significant differences between FUM B, FUM
VB, and RE K. Even though using an Agent K representative guarantees less
unanimity regarding team decisions than other intra-team strategies, it is shown
that, against certain types of opponents, a representative with a competitor ne-
gotiation strategy may be enough in practice to achieve results comparable to
results obtained by strategies that guarantee unanimity like FUM.
When the opponent is a competitor, team strategies that employ FUM VB,
SBV VB, and SSV VB perform better than their correspondents, FUM B, SBV
B, and SSV B respectively. That is, if the opponent is competitive, taking a
competitive approach and conceding less results in better average team utility
than taking cooperative approaches. In any case, we can observe that the average
utility obtained by team members in some competitive settings (i.e., against
Haggler and IAMCrazyHaggler) is way lower than the one obtained by the same
intra-team strategies against conceders or matchers. This suggests the necessity
to explore new intra-team strategies that are able to cope with some competitor
agents. If we compare the performances of FUM VB, SBV VB, and SSV VB,
the results show that the team using FUM VB gathers higher utility on average
than the rest of the cases. The fact that FUM approaches are usually the best
options may be explained due to the fact that it ensures that all of the team
members are satisfied with those offers sent to the opponent and offers sent by
the opponent. Note that when the opponent is IAMCrazyHaggler, which is a
take it or leave it strategy, FUM B gets a higher average for the average utility
of team members, but there is no statistically and significant difference with the
second runner, FUM VB. In any case, FUM B is statistically different that the
rest of intra-team strategies. In this case, RE K is not capable of retaining an
average utility for team members comparable to FUM VB, SBV VB, and SSV
VB. In fact, all of the negotiations between RE K and IAMCrazyHaggler failed.
Our second evaluation metric is the joint utility of the final agreement5. The
joint utility of all of the participants is a crucial metric in situations where both
parties not only want to get a deal, but also build a long term relationship
and engage in multiple negotiations in the future. An agent that has been ex-
ploited in the negotiation process may be reluctant to negotiate with the same
team/opponent in the future. Table 2 shows the average for the joint utility
in the final agreement. The best intra-team strategies for each opponent in the
average utility case, are also the best intra-team strategies in the joint utility
case. The only exception to this rule is the conceder case. In that scenario, the
best results are obtained by employing FUM with a Boulware strategy instead
of exploiting the opponent with very Boulware strategies (FUM VB, SSV VB).
5 The product of the utilities of each team member and the opponent
Intra-Team Strategies Against Competitor, Matchers, and Conceders
15
Competitive
Crazy Haggler K
FUM B 0.005 0.04
FUM VB 0.004 0.06
0.00
RE K
0.05
SSV B
0.002 0.03
SSV VB 0.001 0.04
SBV B 0.002 0.03
SBV VB 0.002 0.05
Matcher Conceder
TFT
0.03 0.17
0.15 0.17
0.11 0.15
0.04 0.07
0.10 0.04
0.03 0.06
0.11 0.06
Smith
0.16
0.04
0.09
0.10
0.02
0.09
0.02
Table 3. The table shows the average for the joint utility (product) of team members
and opponent in the final agreement.
Thus, if a long term relationship is to be built with conceder agents, it may be
wise to employ more concessive intra-team strategies. Very Boulware strategies
exploit the opponent, and get very high results for the average utility of team
members, but they do not allow the opponent to get high utilities, which results
in low joint utilities. On average the highest joint utility is gathered when the
team employs a FUM strategy against to a matcher opponent namely, Nice Tit
for Tat. Since a matcher matches its opponent, TFT matches its behavior with
FUM. Even if FUM concedes slowly over time, TFT will also concede, preclud-
ing both parties from being exploited. In competitive settings, FUM needs to
adjust its concession speed (very Boulware for IAMHaggler and Agent K, and
Boulware for IAMCrazyHaggler) to be able to get the most for the joint utility,
and, still, the results are specially low against Hagglers. This again suggests the
necessity to explore new intra-team strategies that are able to cope with some
competitor agents.
6 Related Work
The artificial intelligence community has focused on bilateral or multi-party
negotiations where parties are composed of single individuals. The most relevant
difference in our work is that we consider multi-individual parties. Next, we
analyze and discuss work related in artificial intelligence.
First, we review some relevant work in bilateral negotiations. Faratin et al. [7]
introduced some of the most widely used families of concession tactics in negoti-
ation. The authors proposed concession strategies for negotiation issues that are
a mix of different families of concession tactics. The authors divide these conces-
sion tactics into three different families: (i) time-dependent concession tactics;
(ii) behavior-dependent concession tactics; and (iii) resource-dependent tactics.
Our negotiation framework also considers time as crucial element in negotiation.
Therefore, team members employ time tactics inspired in those introduced by
Faratin et al.. In another work, Lai et al. [13] propose an extension of the clas-
sic alternating bargaining model where agents are allowed to propose up to k
different offers at each negotiation round. Offers are proposed from the current
16
V. Sanchez-Anguix, R. Aydogan, V. Julian, C. M. Jonker
iso-utility curve according to a similarity mechanism that selects the most similar
offer to the last offer received from the opponent. This present work also consid-
ers extending the bilateral alternating protocol by included layers of intra-team
negotiation among team members. This way, team members can decide on the
actions that should be taken during the negotiation. Robu et al.[21,20] introduce
a bilateral negotiation model where agents represent their preferences by means
of utility graphs. Utility graphs are graphical models that represent binary de-
pendencies between issues. The authors propose a negotiation scenario where
the buyer's preferences and the seller's preferences are modeled through utility
graphs. The seller is the agent that carries out a more thorough exploration of
the negotiation space in order to search for agreements where both parties are
satisfied. With this purpose, the seller builds a model of the buyer's preferences
based on historic information of past deals and expert knowledge about the ne-
gotiation domain. Differently to this work, we introduce multi-individual parties
and add layers of intra-team negotiation to make it possible for team members
to decide on which actions to take during the negotiation.
With regards to multi-party negotiations, several works have been proposed
in the literature [6,12,18,1,9,33,15]. For instance, Ehtamo et al. [6] propose a
mediated multi-party negotiation protocol which looks for joint gains in an it-
erated way and a single agreement should be found to satisfy all of the parties.
The algorithm starts from a tentative agreement and moves in a direction ac-
cording to what the agents prefer regarding some offers' comparison. Klein et al.
[12] propose a mediated negotiation model which can be extended to multiple
parties negotiating the same agreement. Similarly, Ito et al. [9] propose different
types of n-ary utility functions and efficient multiparty models for multiple par-
ties negotiating on the same agreement. Marsa-Maestre et al. [16,17] carry out
further research in the area of negotiation models for complex utility functions.
More specifically, they extend the constraint based model proposed by Ito et al.
[9] by proposing different bidding mechanisms for agents. One-to-many negotia-
tions and many-to-many negotiations also represent special cases of multi-party
negotiations. One-to-many negotiations represent settings where one party ne-
gotiates simultaneously with multiple parties. It can be a party negotiating in
parallel negotiation threads for the same good with different opponent parties
[18,1,33,15] or a party that negotiates simultaneously with multiple parties like
in the Contract-Net protocol, and the English and Dutch auction [27,25,26].
Many-to-many negotiations consider the fact that many parties negotiate with
many parties, the double auction being the most representative example [26].
Differently to the aforementioned concepts, negotiation teams are not related
with the cardinality of the parties but the nature of the party itself. When ad-
dressing a negotiation team, we consider a negotiation party that is formed by
more than a multiple individuals whose preferences have to be represented in
the final agreement. This complex negotiation party can participate in bilateral
negotiations, one-to-many negotiations, or many-to-many negotiations. The rea-
son to model this complex negotiation party instead of as multiple individual
parties is the potential for cooperation. Despite having possibly different indi-
Intra-Team Strategies Against Competitor, Matchers, and Conceders
17
vidual preferences, a negotiation team usually exists because there is a shared
common goal among team members which is of particular importance.
As far as we are concerned, only our previous works [24,23,22,28] have con-
sidered negotiation teams in computational models. More specifically, the four
different computational models introduced in this article are analyzed in dif-
ferent negotiation conditions when facing opponents governed by time tactics.
However, the analysis does not include variability with respect to the strategy
carried out by the opponent like the experiments carried out in this present
article.
7 Conclusions and Future Work
This paper presents preliminary results on the performance of existing intra-
team strategies for bilateral negotiations against heterogeneous opponents: com-
petitors, conceders, and matchers. According to our analysis, the intra-team
strategies like Full Unanimity Mediated (FUM), Similarity Borda Voting (SBV),
Simple Similarity Voting (SSV) and Representative (RE) are able to negotiate
with different success against different types of heterogeneous opponents. For
the average utility of team members on the final agreement and the joint utility
of both parties, we found similar results. In the case of conceders, FUM, SBV,
and SSV seems the best options as long as they wait for the opponent to con-
cede and exploit conceders. In the case of matchers, using either FUM or RE
employing Agent K's negotiation strategy seem the best choices. This suggests
that, for certain types of opponents, a representative approach with an appropri-
ate negotiation strategy may be enough in practice. Finally, the results against
competitors show that while strategies like FUM obtain reasonably good re-
sults against some competitors like Agent K, all of them suffer from exploitation
against other competitor agents like IAMHaggler and IAMCrazyHaggler. Since
existing intra-team strategies such as FUM, SBV, and SSV employ time tactics,
they are inclined to concede during the negotiation. This may suggest that new
intra-team strategies are needed to tackle negotiations against a broader set of
competitors.
Additionally, we have extended the well-known negotiation testbed, GENIUS
to support bilateral negotiations where at least one of the parties is a team. The
extension allows developers to design their own intra-team strategies by extend-
ing the type of mediator used by the team and the type of team member. We
expect that by extending GENIUS with negotiation teams, the research in ne-
gotiation teams will further advance and facilitate research.
For future work, we consider designing intra-team negotiation strategies that
analyze the behavior of opponent and act accordingly. If a team understands that
the opponent is cooperative, the team may act cooperatively and find a mutually
acceptable agreement early. Otherwise, if the opponent is a competitor, the team
may decide to take a strong position and not concede during the negotiation.
18
V. Sanchez-Anguix, R. Aydogan, V. Julian, C. M. Jonker
8 Acknowledgments
One part of this research is supported by TIN2011-27652-C03-01 and TIN2012-
36586-C03-01 of the Spanish government. Other part of this research is sup-
ported by the Dutch Technology Foundation STW, applied science division of
NWO and the Technology Program of the Ministry of Economic Affairs; the
Pocket Negotiator project with grant number VICI-project 08075 and the New
Governance Models for Next Generation Infrastructures project with NGI grant
number 04.17. We would also like to thank Tim Baarslag due to his helpful and
valuable comments and feedback about Genius.
References
1. B. An, K. Sim, L. Tang, S. Li, et al. Continuous-time negotiation mechanism
for software agents. Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE
Transactions on, 36(6):1261 -- 1272, 2006.
2. T. Baarslag, K. Hindriks, C. Jonker, S. Kraus, and R. Lin. The first automated
negotiating agents competition (anac 2010). New Trends in Agent-Based Complex
Automated Negotiations, pages 113 -- 135, 2010.
3. T. Baarslag, K. V. Hindriks, and C. M. Jonker. Towards a quantitative concession-
based classification method of negotiation strategies. In Agents in Principle, Agents
in Practice. Lecture Notes of The 14th International Conference on Principles and
Practice of Multi-Agent Systems, 2011.
4. T. Baarslag, K. V. Hindriks, and C. M. Jonker. A Tit for Tat Negotiation Strat-
egy for Real-Time Bilateral Negotiations, volume 435 of Studies in Computational
Intelligence, pages 229 -- 233. Springer Berlin Heidelberg, 2013.
5. S. Brodt and L. Thompson. Negotiating teams: A levels of analysis. Group Dy-
namics, 5(3):208 -- 219, 2001.
6. H. Ehtamo, E. Kettunen, and R. P. Hamalainen. Searching for joint gains in
multi-party negotiations. Eur. J. Oper. Res., 130(1):54 -- 69, 2001.
7. P. Faratin, C. Sierra, and N. R. Jennings. Negotiation decision functions for au-
International Journal of Robotics and Autonomous Systems,
tonomous agents.
24(3-4):159 -- 182, 1998.
8. P. Faratin, C. Sierra, and N. R. Jennings. Using similarity criteria to make issue
trade-offs in automated negotiations. Artificial Intelligence, 142:205 -- 237, 2002.
9. K. Fujita, T. Ito, and M. Klein. Secure and efficient protocols for multiple interde-
pendent issues negotiation. Journal of Intelligent and Fuzzy Systems, 21(3):175 --
185, 2010.
10. K. Hindriks, C. Jonker, and D. Tykhonov. The benefits of opponent models in
negotiation. In Proceedings of the 2009 IEEE/WIC/ACM international joint con-
ference on Web Intelligence and Intelligent Agent Technology, pages 439 -- 444, 2009.
11. S. Kawaguchi, K. Fujita, and T. Ito. Compromising strategy based on estimated
maximum utility for automated negotiation agents competition. In Modern Ap-
proaches in Applied Intelligence, volume 6704, pages 501 -- 510. Springer Berlin /
Heidelberg, 2011.
12. M. Klein, P. Faratin, H. Sayama, and Y. Bar-Yam. Negotiating complex contracts.
Group Decis. Negot., 12(2):111 -- 125, 2003.
Intra-Team Strategies Against Competitor, Matchers, and Conceders
19
13. G. Lai, K. Sycara, and C. Li. A decentralized model for automated multi-attribute
negotiations with incomplete information and general utility functions. Multiagent
and Grid Systems, 4(1):45 -- 65, 2008.
14. R. Lin, S. Kraus, T. Baarslag, D. Tykhonov, K. Hindriks, and C. M. Jonker.
Genius: An integrated environment for supporting the design of generic automated
negotiators. Computational Intelligence, 2012.
15. K. Mansour and R. Kowalczyk. A meta-strategy for coordinating of one-to-many
In Foundations of Intelligent Systems, volume
negotiation over multiple issues.
122, pages 343 -- 353. Springer Berlin / Heidelberg, 2012.
16. I. Marsa-Maestre, M. A. Lopez-Carmona, J. R. Velasco, and E. de la Hoz. Effective
bidding and deal identification for negotiations in highly nonlinear scenarios. In
Proceedings of The 8th International Conference on Autonomous Agents and Mul-
tiagent Systems (AAMAS'09), pages 1057 -- 1064, Richland, SC, 2009. International
Foundation for Autonomous Agents and Multiagent Systems.
17. I. Marsa-Maestre, M. A. L´opez-Carmona, J. R. Velasco, T. Ito, M. Klein, and
K. Fujita. Balancing utility and deal probability for auction-based negotiations
in highly nonlinear utility spaces. In International Joint Conference on Artificial
Intelligence, pages 214 -- 219, 2009.
18. T. Nguyen and N. Jennings. Coordinating multiple concurrent negotiations. In
Proceedings of the Third International Joint Conference on Autonomous Agents
and Multiagent Systems, pages 1064 -- 1071. IEEE Computer Society, 2004.
19. H. Nurmi. Voting systems for social choice. Handbook of Group Decision and
Negotiation, pages 167 -- 182, 2010.
20. V. Robu and J. A. La Poutr´e. Retrieving the structure of utility graphs used in
multi-item negotiation through collaborative filtering of aggregate buyer prefer-
ences. In Rational, Robust and Secure Negotiations, volume 89 of Computational
Intelligence. Springer, 2008.
21. V. Robu, D. J. A. Somefun, and J. A. La Poutr´e. Modeling complex multi-issue
negotiations using utility graphs. In Proceedings of the fourth international joint
conference on Autonomous agents and multiagent systems (AAMAS'05), pages
280 -- 287, New York, NY, USA, 2005. ACM.
22. V. Sanchez-Anguix, T. Dai, Z. Semnani-Azad, K. Sycara, and V. Botti. Modeling
power distance and individualism/collectivism in negotiation team dynamics. In 45
Hawaii International Conference on System Sciences (HICSS-45), pages 628 -- 637,
2012.
23. V. Sanchez-Anguix, V. Julian, V. Botti, and A. Garc´ıa-Fornes. Reaching unani-
mous agreements within agent-based negotiation teams with linear and monotonic
utility functions. IEEE Transactions on Systems, Man, and Cybernetics, Part B:
Cybernetics, 42(3):778 -- 792, 2012.
24. V. Sanchez-Anguix, V. Julian, V. Botti, and A. Garc´ıa-Fornes. Analyzing Intra-
Team Strategies for Agent-Based Negotiation Teams. In 10th International Con-
ference on Autonomous Agents and Multiagent Systems, pages 929 -- 936, 2011.
25. T. Sandholm. An implementation of the contract net protocol based on marginal
cost calculations. In Proceedings of the eleventh national conference on Artificial
intelligence, pages 256 -- 262. AAAI Press, 1993.
26. Y. Shoham and K. Leyton-Brown. Multiagent systems: Algorithmic, game-
theoretic, and logical foundations. Cambridge University Press, 2009.
27. R. G. Smith. The contract net protocol: High-level communication and control in a
distributed problem solver. Computers, IEEE Transactions on, 100(12):1104 -- 1113,
1980.
20
V. Sanchez-Anguix, R. Aydogan, V. Julian, C. M. Jonker
28. V. S´anchez-Anguix, V. Julian, V. Botti, and A. Garc´ıa-Fornes. Studying the Im-
pact of Negotiation Environments on Negotiation Teams' Performance. Informa-
tion Sciences, 219:17 -- 40, 2013.
29. M. Tambe and H. Jung. The benefits of arguing in a team. AI Magazine, 20:85 -- 92,
1999.
30. L. Thompson, E. Peterson, and S. Brodt. Team negotiation: An examination of
integrative and distributive bargaining. Journal of Personality and Social Psychol-
ogy, 70:66 -- 78, 1996.
31. N. van Galen Last. Agent smith: Opponent model estimation in bilateral multi-
issue negotiation. In New Trends in Agent-Based Complex Automated Negotiations,
volume 383, pages 167 -- 174. 2012.
32. C. R. Williams, V. Robu, E. H. Gerding, and N. R. Jennings. Iamhaggler: A ne-
gotiation agent for complex environments. In New Trends in Agent-based Complex
Automated Negotiations, Series of Studies in Computational Intelligence, volume
383, pages 151 -- 158, 2010.
33. C. R. Williams, V. Robu, E. H. Gerding, and N. R. Jennings. Negotiating concur-
rently with unknown opponents in complex, real-time domains. In 20th European
Conference on Artificial Intelligence, volume 242, pages 834 -- 839, 2012.
|
1905.11764 | 1 | 1905 | 2019-05-28T12:19:47 | Justification Based Reasoning in Dynamic Conflict Resolution | [
"cs.MA"
] | We study conflict situations that dynamically arise in traffic scenarios, where different agents try to achieve their set of goals and have to decide on what to do based on their local perception. We distinguish several types of conflicts for this setting. In order to enable modelling of conflict situations and the reasons for conflicts, we present a logical framework that adopts concepts from epistemic and modal logic, justification and temporal logic. Using this framework, we illustrate how conflicts can be identified and how we derive a chain of justifications leading to this conflict. We discuss how conflict resolution can be done when a vehicle has local, incomplete information, vehicle to vehicle communication (V2V) and partially ordered goals. | cs.MA | cs |
JUSTIFICATION BASED REASONING IN DYNAMIC CONFLICT
RESOLUTION∗
A PREPRINT
Werner Damm Martin Franzle Willem Hagemann
Paul Kroger Astrid Rakow
Department of Computing Science
University of Oldenburg, Germany
{werner.damm, martin.fraenzle, willem.hagemann, paul.kroeger, a.rakow}@uol.de
May 29, 2019
ABSTRACT
We study conflict situations that dynamically arise in traffic scenarios, where different agents try
to achieve their set of goals and have to decide on what to do based on their local perception.
We distinguish several types of conflicts for this setting. In order to enable modelling of conflict
situations and the reasons for conflicts, we present a logical framework that adopts concepts from
epistemic and modal logic, justification and temporal logic. Using this framework, we illustrate
how conflicts can be identified and how we derive a chain of justifications leading to this conflict.
We discuss how conflict resolution can be done when a vehicle has local, incomplete information,
vehicle to vehicle communication (V2V) and partially ordered goals.
1 Introduction
As humans are replaced by autonomous systems, such systems must be able to interact with each other and resolve
dynamically arising conflicts. Examples of such conflicts arise when a car wants to enter the highway in dense traffic or
simply when a car wants to drive faster than the preceding. Such "conflicts" are pervasive in road traffic and although
traffic rules define a jurisdictional frame, the decision, e.g., to give way, is not uniquely determined but influenced
by a list of prioritised goals of each system and the personal preferences of its user. If it is impossible to achieve all
goals simultaneously, autonomous driving systems (ADSs) have to decide "who" will "sacrifice" what goal in order
to decide on their manoeuvres. Matters get even more complicated when we take into account that the ADS has only
partial information. It perceives the world via sensors of limited reach and precision. Moreover, measurements can be
contradicting. An ADS might use V2V to retrieve more information about the world, but it inevitably has a confined
insight to other traffic participants and its environment. Nevertheless, for the acceptance of ADSs, it is imperative to
implement conflict resolution mechanisms that take into account the high dimensionality of decision making. These
decisions have to be explained and in case of an incident, the system's decisions have to be accountable.
In this paper we study conflict situations as dynamically occurring in road traffic and develop a formal notion of con-
flict between two agents. We distinguish several types of conflicts and propose a conflict resolution process where the
different kinds of conflicts are resolved in an incremental fashion. This process successively increases the required
cooperation and decreases the privacy of the agents, finally negotiating which goals of the two agents have to be sacri-
ficed. We present a logical framework enabling the analysis of conflicts. This framework borrows from epistemic and
modal logic in order to accommodate the bookkeeping of evidences used during a decision process. The framework
in particular provides a mean to summarise consistent evidences and keep them apart from inconsistent evidences. We
hence can, e.g., fuse compatible perceptions into a belief b about the world and fuse another set of compatible per-
ceptions to a belief b′ and model decisions that take into account that b might contradict b′. Using the framework we
illustrate how conflicts can be explained and algorithmically analysed as required for our conflict resolution process.
∗This work is partly supported by the German Research Council (DFG) as part of the PIRE SD-SSCPS project (Science of
Design of Societal Scale CPS, grant no. DA 206/11-1, FR 2715/4-1) and the Research Training Group SCARE (System Correctness
under Adverse Conditions, grant no. DFG GRK 1765).
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
B
A
Figure 1: Car A wants to circumvent the ob-
stacle (grey box). Car B is approaching from
behind.
Finally we report on a small case study using a prototype implementation (employing the Yices SMT solver [1]) of
the conflict resolution algorithm.
Outline.
In Sect. 2 we introduce the types of conflict on a running example and develop a formal notion of conflict
between two agents. We elaborate on the logical foundations for modelling and analysing conflicts and the logical
framework itself in Sect. 3. We sketch our case study on conflict analysis in Sect. 4 and outline in Sect. 4.2 an algorithm
for analysing conflict situations as requested by our resolution protocol and for deriving explanation of the conflict for
the resolution. Before drawing the conclusions in Sect. 6, we discuss related work in Sect. 5.
2 Conflict
Already in 1969 in the paper "Violence, Peace and Peace Research" [2] J. Galtung presents his theory of the Conflict
Triangle, a framework used in the study of peace and conflict. Following this theory a conflict comprises three
aspects: opposing actions, incompatible goals, inconsistent beliefs (regarding the reasons of the conflict, knowledge
of the conflict parties,. . . ).
We focus on conflicts that arise dynamically between two agents in road traffic. We develop a characterisation of
conflict as a situation where one agent can accomplish its goals with the help of the other, but both agents cannot
accomplish all their goals simultaneously and the agents have to decide what to do based on their local beliefs. In
Sect. 2.1 we formalise our notion of conflict. For two agents with complete information, we may characterise a
conflict as: Agents A and B are in conflict, if 1. A would accomplish its set of goals ΦA, if B will do what A requests,
while 2. B would accomplish its set of goals ΦB, if A will do what B requests, and 3. it is impossible to accomplish
the set of goals ΦA ∪ ΦB. A situation where A and B both compete to consume the same resource is thus an example
of a conflict situation. Since we study conflicts from the view-point of an agent's beliefs, we also consider believed
conflicts, which can be resolved by sharing information regarding the others observations, strategies or goals. To
resolve a conflict we propose a sequence of steps that require an increasing level of cooperation and decreasing level
of privacy -- the steps require to reveal information or to constrain acting options. Our resolution process defines the
following steps:
(C1) Shared situational awareness
(C2) Sharing strategies
(C3) Sharing goals
(C4) Agreeing on which goals to sacrifice and which strategy to follow
Corresponding to (C1) to (C4), we introduce different kinds of conflicts on a running example -- a two lane highway,
where one car, A, is heading towards an obstacle at its lane and at the lane to its left a fast car, B, is approaching
from behind (cf. Fig. 1). An agent has a prioritised list of goals (like 1. "collision-freedom", 2."changing lane" and 3.
"driving fast"). We assume that an agent's goals are achievable.
An agent A has a set of actions actA and exists within a world. At a time the world has a certain state. The world
"evolves" (changes state) as determined by the chosen actions of the agents within the world and events determined
by the environment within the world. The agent perceives the world only via a set of observation predicates, that are
predicates whose valuation is determined by an observation of the agent. Without an observation the agent has no
(direct) evidence for the valuation of the respective observation predicate.
Example 1. Let car A want to change lane. It perceives that it is on a two lane highway, the way ahead is free for
the next 500 m and B is approaching. Let A perceive B's speed via radar. That is A makes the observation car B is
fast justified by the evidence radar. We annotate this briefly as radar:car B is fast. Further let A derive from
lidar data that B is slow -- lidar:car B is slow.
In this situation we say agent A has contradicting evidences. Certain evidences can be combined without contradiction
and others not. We assume that an agent organises its evidences in maximal consistent sets (i.e., justification graphs
of Sect. 3), where each represents a set of possible worlds:
2
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
Example 2. There are possible worlds of A where it is on a two lane highway, the way ahead is free for the next 500 m
and B is slowly approaching. Analogously A considers possible worlds where B is fast. The state of the world outside
of its sensors' reach is unconstrained.
Observing the world (for some time), an agent A assesses what it can do to achieve its goals in all possible worlds.
That is, A tries to find a strategy that guarantees to achieve its goals in all its possible worlds. A strategy determines at
each state the action of the agent -- the agent decides for an action based on its beliefs formed in the past regarding its
possible worlds. If there is one such strategy for A to accomplish its goals ΦA, then A has a (believed) winning strategy
for ΦA. This strategy might not be winning in the "real" world though, e.g., due to misperceptions.
Example 3. Let A want to drive slowly and comfortably. A wants to avoid collisions and it assumes that also B wants
to avoid collisions. Although A has contradicting evidences on the speed of B and hence believes that it is possible
that "B is fast" and also that "B is slow", it can follow the strategy to stay at its lane and wait until B has passed.
This strategy is winning in all of A's possible worlds.
Even when A has no believed winning strategy, it can have a winning strategy for a subset of possible worlds. Ad-
ditional information on the state of world might resolve the conflict by eliminating possible worlds. We call such
conflicts observation-resolvable conflicts.
Example 4. Let A want to change lane to circumvent the obstacle. It is happy to change directly after B but only if B
is fast. If B is slow, it prefers to change before B passed. Further let A have contradicting evidences on the speed of B.
A considers a conflict with B possible in some world and hence has no believed winning strategy. Now it has to resolve
its inconsistent beliefs. Let B tell A, it is fast, and A trust B more than its own sensors, then A might update its beliefs
by dismissing all worlds where B is slow. Then "changing after B passed" becomes a believed winning strategy.
In case of inconsistent evidences, as above, A has to decide how to update its beliefs. The decision how to update its
beliefs will be based on the analysis of justifications (cf. Sect. 3) of (contradicting) evidences. The lidar contradicts
the radar and B reports on its speed. Facing the contradiction of evidences justified by lidar and radar A trusts the
evidence justified by B.
Let the agents already have exchanged observations and A still have no believed winning strategy. A conflict might be
resolved by communicating part of the other agent's (future) strategy:
Example 5. Let A want to change lane. It prefers to change directly after B, if B passes A fast. Otherwise, A wants
to change in front of B. Let B so far away that B might decelerate, in which case it might slow down so heavily that A
would like to change in front of B even if B currently is fast.
Let A believe "B is fast". Now A has no believed winning strategy, as B might decelerate. According to (C2),
information about parts of the agent's strategies are now communicated. A asks B whether it plans to decelerate.
Let B be cooperative and tell A that it will not decelerate. Then A can dismiss all worlds where B slows down and
"changing after B passed" becomes a believed winning strategy for A.
Let the two agents have performed steps (C1) and (C2), i.e., they exchanged missing observations and strategy parts,
and still A has no winning strategy for all possible worlds.
Example 6. Let now, in contrast to Ex. 5, B not tell A whether it will decelerate. Then step (C3) is performed. So A
asks B to respect A's goals. Since A prefers B to be fast and B agrees to adopt A's goal as its own, A can again dismiss
all worlds where B slows down.
Here the conflict is resolved by communicating goals and the agreement to adopt the other's goals. So an agent's
strategy might change in order to support the other agent. We call this kind of conflicts goal-disclosure-resolvable
conflicts.
The above considered conflicts can be resolved by some kind of information exchange between the two agents, so that
the sets of an agent's possible worlds is adapted and in the end all goals ΦA of A and ΦB of B are achievable in all
remaining possible worlds. The price to pay for conflict resolution is that the agents will have to reveal information.
Still there are cases where simply not all goals are (believed to be) achievable. In this case A and B have to negotiate
which goals ΦAB ⊆ ΦA ∪ΦB shall be accomplished. While some goals may be compatible, other goals are conflicting.
We hence consider goal subsets ΦAB of ΦA ∪ ΦB for which a combined winning strategy for A and B exists to achieve
GAB. We assume that there is a weight assignment function w that assigns a value to a given goal combination
2ΦA∪ΦB → N based on which decision for a certain goal combination is taken. This weighting of goals reflects the
relative value of goals for the individual agents. Such a function will have to reflect, e.g., moral, ethics and jurisdiction.
Example 7. Let A's and B's highest priority goal be collision-freedom, reflected in goals ϕA,col and ϕB,col. Further
let A want to go fast ϕA,f ast and change lane immediately ϕA,lc. Let also B want to go fast ϕB,f ast, so that A cannot
3
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
change immediately. Now in step (C4) A and B negotiate what goals shall be accomplished. In our scenario collision-
freedom is valued most, and B's goals get priority over A's, since B is on the fast lane. Hence our resolution is to
agree on a strategy accomplishing {ϕA,col, ϕB,col, ϕB,f ast}, which is the set of goals having the highest value among
all those for which a combined winning strategy exists.
Note that additional agents are captured as part of the environment here. At each step an agent can also decide to
negotiate with some other agent than B in order to resolve its conflict.
2.1 Formal Notions
In the following we introduce basic notions to define a conflict. Conflicts, as introduced above, arise in a wide variety
of system models, but we consider in this paper only a propositional setting.
Let f1 : X → Y1,. . . , fn : X → Yn, and f : X → Y1 × . . . × Yn be functions. We will write f = (f1, . . . , fn) if
and only if f (x) = (f1(x), . . . , fn(x)) for all x ∈ X. Note that for any given f as above the decomposition into its
components fi is uniquely determined by the projections of f onto the corresponding codomain.
Each agent A has a set of actions AA. The sets of actions of two agents are disjoint. To formally define a (possible)
world model of an agent A, let S be a set of states and V be a set of propositional variables. V represents the set of
belief propositions. A state s ∈ S of a (possible) world is labelled with a subset V ⊂ V that is (assumed to be) true.
V \ V is (assumed to be) false.
A (possible) world model M for an agent A is a transition system over S with designated initial state and current
state, all states are labelled with the belief propositions that hold at that state and transitions labelled with actions
hactA, actB, actEnvi with actA ∈ AA an action of agent A, actB ∈ AB an action of agent B and actEnv ∈ AEnv an action
of the environment.
The set of actions of an agent includes send and receive actions via which information can be exchanged, the en-
vironment guarantees to transmit a send message to the respective receiver. Formally a possible world is MA =
(S, T, λ, π, s∗, sc) with
• T ⊆ S × S,
• λ : T → AA × AB × AEnv,
• π : S → 2V,
• s∗ ∈ S, sc ∈ S s.t.
-- ∀s ∈ S: (s, s∗) 6∈ T
-- ∃s1, . . . , sn+1 ∈ S:∀1 ≤ i ≤ n (si, si+1) ∈ T, ∧ s1 = s∗ ∧ sn+1 = sc
(s∗ is the initial state)
∧ ((si, s′) ∈ T ⇒ s′ = si+1)
(the current state sc is reachable from the initial state)
( M is linear between s∗ and sc)
The part of M between s∗ and sc represents the history of the current state. A finite run in M is a sequence of states
r = s1s2 . . . sn+1 with ∀1 ≤ i ≤ n : (si, si+1) ∈ T.
There is one "special" world model that represents the ground truth, i.e., it reflects how the reality evolves. An agent
A considers several worlds possible at a time. This is, at each state s of the real world, A has a set of possible worlds
MA(s). The real world changes states according to the actions of A, B and Env. The set of possible worlds MA(s)
changes to MA(s′) due to the passing of time and due to belief updates triggered by e.g. observations. For the scope
of this paper though, we do not consider the actual passing of time, but study the conflict analysis at a single state of
the real world from the point of view of an agent. Since at each state an agent A may consider several worlds possible,
it may also consider several histories possible. A strategy is hence a function δA : 2(2V)∗
→ AA, that determines an
action for A based on the set of possible histories. H ∈ 2(2V)∗
represents a set of histories, where a history h ∈ H is
given via the sequence of valuations of V along the path from s∗ to sc. The set of possible histories at state s is the
union of histories of possible worlds MA ∈ MA(s), denoted as H(MA(s)).
Let r = s0s1 . . . sn be a run in MA and γ = υ1υ2 . . . υn ∈ (AB × AEnv)n be a sequence of actions of agent B and
Env along r. r follows strategy δ, i.e., r = r(δ, γ), if λ(si−1, si) = (δ(H(MA(si−1)), υi), ∀0 ≤ i ≤ n. We also write
r(δ, MA) to denote the set of runs of MA that follow δ.
We use linear-time temporal logic (LTL) to specify goals (cf. Def. Def. 9). For a run r and a goal (or a conjunction of
goals) ϕ, we write r = ϕ if r satisfies ϕ, i.e., the valuation of propositions along the state sequence satisfies ϕ.2 We say
2We assume that runs are infinite here. In case of finite runs, we make them infinite by repeating the last state infinitely often.
4
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
δ is a (believed) winning strategy for ϕ in MA, if all runs r of MA that follow δ also satisfy ϕ, ∀r ∈ r(MA, δ) : r = ϕ.
We say that δ is a (believed) winning strategy of A for ϕ at the real world state s if δ is a winning strategy for ϕ in all
possible worlds MA ∈ MA(s).
An agent A has a set of goals Φ and a weight assignment function wA : 2Φ → N that assigns values to a given goal
combination. We write r = Φ as shorthand for r = Vϕ∈Φ ϕ. We say Φ′ ⊂ Φ is a believed achievable goal at
real world s if there is a strategy δ, that is winning for the conjunction of all goals ϕ ∈ Φ′ in all possible worlds
MA ∈ MA(s). We say Φ′ ⊂ Φ is a believed maximal goal at real world state s if its is a believed achievable goal and
for all believed achievable goals Φ′′ ⊂ Φ it holds that w(Φ′) ≥ w(Φ′′). The empty subgoal is defined to be true (⊤).
For each world possible MA ∈ MA(s) agent A also has
1. beliefs on the goals of B, ΦB(MA), and
2. beliefs on the importance of subgoals of ΦB(MA) to B, wB(MA), and
3. a set J(MA) of justifications for MA, ΦB(MA) and wB(MA).
So at state s of the real world an agent A has belief B(A, s) = SMA∈MA(s)(MA, ΦB(MA), wB(MA), J(MA)). The
justifications support decision making by keeping track of (source or more generally meta) information. They hence
can influence decisions on how to update an agent's knowledge, how to negotiate and what resolutions are acceptable.
Our notion of conflict captures the following concept: Let Φmax be the set of maximal goals that A beliefs it can
achieve with the help of B. But since B might choose a strategy to accomplish some of its maximal goals, A believes
that it is in a conflict with B, if it cannot find one winning strategy that fits all possible strategy choices of B.
Definition 1 (Believed Possible Conflict). Let Φmax
winning strategy (δA, δ′
B) : (2VA)∗ → AA × AB in MA exists.
A
be the set of maximal subgoals of A at state s for which a believed
Agent A believes at state s it is in a possible conflict with B, if for each of its winning strategies (δA, δ′
AA × AB for a maximal subgoal ΦA ∈ Φmax
,
A
B) : (2VA)∗ →
• there is a strategy (δ′
A, δB) : (2VA)∗ → AA × AB and a possible world M ∈ MA such that (δ′
A, δB) is a winning
strategy in M for ΦB, a believed maximal subgoal of the believed goals of B in M .
• but (δA, δB) is not a winning strategy for ΦA ∪ ΦB in MA.
The above notion of conflict captures that A analyses the situation within its possible worlds MA(s). It assumes that B
will follow some winning strategy to accomplish its own goals, while Env is assumed to behave fully adversarial. A
believes that B beliefs that one of A's possible worlds represents the reality. It is an interesting future extension to also
allow A having more complicated beliefs about the beliefs of B, as already well supported by the logical framework
introduced in Sect. 3. For instance we can capture situations like A considers it possible that there is an obstacle on the
road, while it believes that B believes there is no obstacle. This extension does change the base line of our contribution
but makes the following presentation more complex. So we refrain from considering beliefs about beliefs for the sake
of comprehensibility.
For an example of the conflict notion, consider a situation where A drives on a highway side by side of B and A just
wants to stay collision-free, A does not believe to be in a conflict situation when it believes that B also prioritizes
collision-freedom, since B will not suddenly choose to crash into A which would violate its own goal. But in case B
has no strategy to accomplish collision-freedom (assume a broken car in front of B) within MA, then A assumes that B
behaves arbitrarily (achieving its remaining goal ⊤) and A believes to be in conflict with B.
2.2 Applying the Formal Notion
In this subsection we consider the formal notions introduced in the previous subsection and illustrate them -- focusing
on the examples given at the start of this section.
Propositional Characterisation of the World For the sake of a small example, let us consider the following propo-
sitional characterisation of a world: For each agent X ∈ {A, B} there is a pair of variables (lX , pX ) storing its position
in the road. Further each agent drives a certain speed sX abstracted to three different levels, sX ∈ {0, 1, 2} encoding
slow, medium and fast speed levels. We consider only time bounded properties. The evolution along the observed
time window is captured via copies of (lX , pX, sX ), (lX,t, pX,t, sX,t) where 0 ≤ t ≤ max obs time encodes the ob-
served time points. Each agent X can change lane, encoded by increasing or decreasing lX , and choose between three
different speeds, that is, (i) decelerate inducing a change from fast to medium, or, medium to slow, respectively, or (ii)
accelerates from slow to medium, or, from medium to fast, respectively.
5
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
past
.
<l,l,l>,
<l,l,l>,
<k,k,k>
s12
.
.
...
future
s18
...
Figure 2: The transition labelling λ is sketched within the figure itself.
The state labelling is omitted there. Let us assume that the initial state
π(s∗) is labelled with {lA = 1, lB = 2, lo = 1, pA = 3, pB = 1,
po = 7, sA = medium, sB = medium, so = slow} describing the
situation of Fig. 1. Currently we are at the time t = 4. A, B and the
environment (determining the moves of the obstacle) have done three
moves: (1) all three stayed at their respective lane and kept their speed,
(2) the same but B accelerates and (3) same as (1). The state labelling
reflects the changes induced by the chosen moves. So the propositions
that are true at, e.g., s−2 differ from the one of s∗ only in terms of the
respective positions: {lA = 1, lB = 2, lo = 1, pA = 4, pB = 4, po = 7,
sA = medium, sB = medium, so = slow}.
t=4
<nl,nl,nl>,
<k,k,k>
<nl,nl,nl>,
<k,a,k>
<nl,nl,nl>,
<k,k,k>
s*
s-2
s-1
sc
<nl,nl,nl>,
<a,a,a>
<l,nl,nl>,
<l,nl,nl>,
<k,a,a>
...
s6
<l,l,nl>,
<k,k,a>
...
...
s1
...
t=2
possible world M2 of A
t=2
real world / ground truth
s2
*
<nl,nl,nl>,
<k,a,k>
l d o f A
p o s s i b l e w o r
s2c
t=2
possible world M1 of A
r l d o f A
s sible w o
p o
<nl,nl,nl>,
<a,a,a>
<l,nl,nl>,
<l,nl,nl>,
<k,a,a>
...
<l,l,nl>,
<k,k,a>
...
s2,6
...
s2,1
...
<l,l,l>,
<l,l,l>,
<k,k,k>
s12
.
.
...
s1
*
.
<nl,nl,nl>,
<k,d,k>
s18
...
s1c
<nl,nl,nl>,
<a,a,a>
<l,nl,nl>,
<l,nl,nl>,
<k,a,a>
...
<l,l,nl>,
<k,k,a>
...
s1,6
...
s1,1
...
s1
...
.
<l,l,l>,
<l,l,l>,
<k,k,k>
s1,12
.
.
...
s1,18
...
<nl,nl,nl>,
<k,a,k>
s*
sc
<nl,nl,nl>,
<a,a,a>
<l,nl,nl>,
<l,nl,nl>,
<k,a,a>
...
s6
<l,l,nl>,
<k,k,a>
...
...
.
<l,l,l>,
<l,l,l>,
<k,k,k>
s12
.
.
...
s18
...
Figure 3: The real world to the right is asso-
ciated with beliefs of agent A. A considers at
time t = 2 two worlds as possible, one, M1,
bisimilar to the real world and a second one,
M2, where B accelerates as its first move.
In this setting each state of the real world model is labelled with propositions {lX, pX , sX
A Real World Model
X ∈ {A, B, o}} and there are transitions from a state s to a state s′ labelled λ(s,s')=(lc1, lc2, lc3, sc1, sc2, sc3), where
lci ∈ {lane_change, ¬lane_change} and sci ∈ {a, d, k}. lci encodes whether agent Xi chooses to perform a lane
change and sci encodes how Xi chooses to change its speed, i.e., to accelerate, decelerate or to keep its speed. The
target state is labelled according to effect of the chosen action.
The initial state encodes the start situation (of the tour) and the subgraph from the initial state to the current state
captures the observed past. A world model has a branching structure from the current state towards the future into the
possible different options of lane changing and choices of speed change. Such a world model describes the past, the
current state of the world and possible future evolutions. For each point in time t there is hence such a world model.
See Fig. 2 for a sketch of an example of a world model at a time t = 4.
Possible Worlds Additional to labelling of states and transitions, the real world is also labelled with beliefs of the
agents at that time. The gist is
R the real world model captures the past, presence and the possible futures at a time t.
B At time t an agent within world model M considers a set of worlds possible. This belief is justified by e.g.
evidences from its sensors.
An example is sketched in Fig. 3, where only A's beliefs are sketched. Note that the state labelling, i.e. the set of true
propositions, is not specified in Fig. 3 in order to declutter the figure. Some state labelling is given in Fig. 4. Let for
Fig. 3 be the initial states of the real and the possible worlds be identically labelled, i.e., the agent believes in the "real"
past.
Let us now consider Ex. 1. Agent A has evidence for B being fast and it also has evidence for B being slow. Fig. 4
illustrates that agent A considers the two (sets of) worlds possible that differ in the valuation of the respective state
propositions. Agent A believes that a world is possible where B is fast -- this is justified by its radar data -- , and A
considers a world possible where B is slow -- justified by its lidar.
We assume that an agent considers any world M possible that can be justified by some non-empty set of consistent
evidences. So a possible world M satisfies e.g. a set of constraints that is derived from the agent's observations, i.e.,
the sensory evidences, and it also has to be compatible to the agent's laws/rules about the world, e.g., physical laws.
6
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
t=2
real world / ground truth
mA, mB, so,
pA=3, pB= 1, po=7,
lA=1, lB=2, lo=1
mA, fB, so,
pA=4, pB= 2, po=7,
lA=1, lB=2, lo=1
<nl,nl,nl>,
<k,a,k>
s*
sc
<nl,nl,nl>,
<a,a,a>
<l,nl,nl>,
<l,nl,nl>,
<k,a,a>
...
s6
<l,l,nl>,
<k,k,a>
...
...
.
<l,l,l>,
<l,l,l>,
<k,k,k>
s12
.
.
...
s18
...
s1
...
possible world M2 of A
<nl,nl,nl>,
<k,a,k>
s1*
s1c
<nl,nl,nl>,
<a,a,a>
<l,nl,nl>,
<l,nl,nl>,
<k,a,a>
...
<l,l,nl>,
<k,k,a>
...
s1,6
...
s1,1
...
mA, mB, so,
pA=3, pB= 1, po=7,
lA=1, lB=2, lo=1
mA, fB, so,
pA=4, pB= 2, po=7,
lA=1, lB=2, lo=1
possible world M1 of A
.
<l,l,l>,
<l,l,l>,
<k,k,k>
s1,12
.
.
...
s1,18
...
<nl,nl,nl>,
<k,d,k>
s1*
s1c
<nl,nl,nl>,
<a,a,a>
mA, mB, so,
pA=3, pB= 1, po=7,
lA=1, lB=2, lo=1
mA, sB, so,
pA=4, pB= 2, po=7,
lA=1, lB=2, lo=1
<l,nl,nl>,
<l,nl,nl>,
<k,a,a>
...
<l,l,nl>,
<k,k,a>
...
s1,6
...
.
<l,l,l>,
<l,l,l>,
<k,k,k>
s1,12
.
.
...
s1,18
...
i
j u s t
p o s s i b l e w o r
p o s s i b l e w o r
p o s s i b l e w o r
p o s s i b l e w o r
j u s t
j u s t
j u s t
i
i
i
l d o f A
c a t i o n : { radar}
l d o f A
c a t i o n : { radar}
l d o f A
c a t i o n : { radar}
l d o f A
c a t i o n : { radar}
A
r l d o f
n : { l i d a r }
p o
ju s
t i
s sible w o
catio
s sible w o
catio
s sible w o
catio
s sible w o
catio
p o
t i
t i
t i
ju s
p o
ju s
p o
ju s
A
A
r l d o f
n : { l i d a r }
r l d o f
n : { l i d a r }
A
r l d o f
n : { l i d a r }
s1,1
...
possible world M2 of A
s1*
s1c
<nl,nl,nl>,
<k,a,k>
...
<nl,nl,nl>,
<a,a,a>
<l,nl,nl>,
<l,nl,nl>,
<k,a,a>
...
<l,l,nl>,
<k,k,a>
...
s1,6
...
s1,1
...
mA, mB, so,
pA=3, pB= 1, po=7,
lA=1, lB=2, lo=1
justifica t
i o n :
l d o f A
t=2
p o s s i b l e w o r
}
{ r a d a r ,
l d o f A
}
p o s s i b l e w o r
{ r a d a r ,
i o n :
mA, fB, so,
pA=4, pB= 2, po=7,
lA=1, lB=2, lo=1
possible world M1 of A
justifica t
.
<l,l,l>,
<l,l,l>,
<k,k,k>
s1,12
.
.
...
s1,18
...
<nl,nl,nl>,
<k,d,k>
s1*
s1c
<nl,nl,nl>,
<a,a,a>
mA, mB, so,
pA=3, pB= 1, po=7,
lA=1, lB=2, lo=1
mA, sB, so,
pA=4, pB= 2, po=7,
lA=1, lB=2, lo=1
<l,nl,nl>,
<l,nl,nl>,
<k,a,a>
...
<l,l,nl>,
<k,k,a>
...
s1,6
...
.
<l,l,l>,
<l,l,l>,
<k,k,k>
s1,12
.
.
...
s1,18
...
s1,1
...
A
r l d o f
,
s sible w o
c atio n: { l i d a r
}
p o
j u s t i fi
A
r l d o f
l i d a r ,
s sible w o
a tio n: {
p o
j u s t i fi c
real world / ground truth
mA, mB, so,
pA=3, pB= 1, po=7,
lA=1, lB=2, lo=1
mA, fB, so,
pA=4, pB= 2, po=7,
lA=1, lB=2, lo=1
<nl,nl,nl>,
<k,a,k>
s*
sc
<nl,nl,nl>,
<a,a,a>
<l,nl,nl>,
<l,nl,nl>,
<k,a,a>
...
s6
<l,l,nl>,
<k,k,a>
...
...
.
<l,l,l>,
<l,l,l>,
<k,k,k>
s12
.
.
...
s18
...
s1
...
}
Figure 4: A considers currently two sets of
worlds possible, one set contains all possible
worlds where B is slow in accordance to the
lidar and all worlds in the other set satisfy that
B is fast in accordance to the the radar.
Figure 5: A derived additional constraints
from B goals that constrain its set of possible
worlds.
The evidences provided by radar and lidar in Ex. 1 imply constraints sA = f ast and sB = slow. These constraints
are contradictory and hence there is no possible world that satisfies both constraints. So there cannot be an arrow in
Fig. 4 from the real world to a possible world that is labelled with a justification set containing both justifications, J =
{radar, lidar}. Nevertheless, the radar and lidar evidences justify that agent A believes in alternative worlds (e.g., B is
fast, so it is possible that (a) B was driving at medium speed and accelerated or (b) B was fast and kept its speed.)
Strategy and Possible Worlds Let us formalise Ex. 3. A considers worlds possible where it has the evidence
radar:car B is fast and hence considers worlds possible where B is fast, and also worlds where B is not fast,
due to its evidence lidar:car B is slow. We already sketched the possible worlds of A above. In order to specify
the goals of A and the goals A believes B has, we use the usual LTL operators3
A wants to drive slowly and comfortably, ϕcf = G≤3(sA = medium ∨ sA = slow) and avoid collisions ϕcl =
G≤10(pA 6= pB ∧ pA 6= po). A also assumes that B wants to avoid collisions, ϕB,cl = G≤10(pB 6= pA ∧ pB 6= po).
The weight assignment to subsets of goals for A is wA({ϕcl, ϕcf }) = 2, wA({ϕcl}) = 1 and wA(Φ) = 0 for all all
other subsets Φ. This expresses that collision-freedom is indispensable. Further A believes collision-freedom is also
indispensable for B. Additionally, A derives from ΦB a constraint that expresses that B will not jeopardize collision-
freedom and hence it will not drive irrationally into A. This constraint further restricts the set of worlds that A considers
possible (cf. Fig. 5).
In this situation, A decides on its next move. It is not aware of the state of real world and decides only based on its
current beliefs regarding the possible worlds and associated goals of B and goal weights. A determines that staying on
its current lane and not changing its speed now is a good move since it can stop and wait in any case, i.e., this move is
the prefix of a winning strategy in M1 and all other possible worlds, in which B is slow, and also in M2 and all other
possible worlds that satisfy that B is fast (cf. Fig. 4).
In Ex. 4 A has the goals • avoid collisions ϕcl = G≤10(pA 6= pB ∧ pA 6= po) and • change lane ϕlc = F≤5
Conflicts
change lane and • change lane before B has passed, if B is slow, ϕf lc = ¬(G≤3sB = f ast) ⇒ pA ≥ pB U
change lane ∧ F≤3change lane and • do not change before B has passed, if B is fast, ϕslc = (G≤3sB = f ast)
⇒ G≤3¬change lane. We assume here that A has only short term goals and global goals are determined at a higher
level.4 A also assumes that B wants to avoid collisions. The weight assignment to subsets of goals for A is spec-
3"F" denotes the finally modal operator, "G" denotes globally, "U" denotes until and in addition we use F≤tϕ to express that
within the next t steps ϕ has to be true and likewise G≤tϕ to specify that at all times up to t ϕ has to hold.
4Note that also collision-freedom might be sacrificed in so-called dilemma situations.
7
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
ified as follows 6 = wA({ϕcl, ϕlc, ϕf lc, ϕslc}) > wA({ϕcl, ϕlc, ϕslc}) = wA({ϕcl, ϕlc, ϕf lc}) > wA({ϕcl, ϕlc}) >
wA({ϕcl, ϕf lc}) = wA({ϕcl, ϕslc})5 > wA({ϕcl, }) = 1, wA(Φ) = 0 for all all other subsets.
Obviously A has a winning strategy δA(B) : 2(2V)∗
→ AA × AB, i.e., if it could determine also B's future moves. In
this case it can achieve {ϕcl, ϕlc, ϕf lc, ϕslc}. If A assumes that B follows a strategy achieving B's own goals under
the assumption that A will cooperate (i.e. B can rule out that A changes lane, forcing B to decelerate), then B can e.g.
make up a winning strategy δ(A)B : 2(2V)∗
→ AA × AB, where A stays at lane 1, B at lane 2 and B chooses its speed
arbitrarily without endangering collision-freedom. A does not have a winning strategy for all these strategies of B,
since A cannot follow the same strategy if (i) B is fast in its next three steps and if (ii) B is not fast in at least one of the
next three steps.
If B tells A how fast it will go in its next three steps, the additional information provided by B, makes A dismiss all
possible worlds that do not satisfy the evidence on B's future behaviour. A can determine an appropriate strategy for
all (remaining) possible worlds and the conflict situation is hence resolved.
3 Epistemic Logic, Justifications and Justification Graph
Conflict analysis demands to know who believes to be in conflict with whom and what pieces of information made
him belief that he is in conflict. To this end we introduce the logic of justification graphs that allows to keep track of
external information and extends purely propositional formulae by so called belief atoms (cf. p. 10), which are used
to label the sources of information. In Sect. 2 we already used such formulae, e.g., "radar:car B is fast". Our
logic provides several atomic accessibility relations representing justified beliefs of various sources, as required for
our examples of Sect. 2. It provides justification graphs as a mean to identify belief entities which compose different
justifications to consistent information even when the information base contains contradicting information of different
sources, as required for analysing conflict situations.
First, this section provides a short overview on epistemic modal logics and multi-modal extensions thereof. Such logics
use modal operators to expressing knowledge and belief stemming from different sources. Often we will refer to this
knowledge and belief as information, especially when focusing on the sources or of the information. Thereupon the
basic principles of justifications logics are shortly reviewed. Justification logics are widely seen as interesting variants
to epistemic logics as they allow to trace back intra-logical and external justifications of derived information. In the
following discussion it turns out that tracing back external justifications follows the same principles as the distribution
of information over different sources.
Consequently, the concept of information source and external justification are then unified in our variant of an epis-
temic modal logic. This logic of justification graphs extends the modal logic by a justification graph. The nodes
of a justification graph are called belief entities and represent groups of consistent information. The leaf nodes of a
justification graphs are called belief atoms, which are information source and external justifications at the same time,
as they are the least constituents of external information. We provide a complete axiomatisation with respect to the
semantics of the logic of justification graphs.
3.1 Justification Graphs
Modal Logic and Epistemic Logic. Modal logic extends the classical logic by modal operators expressing necessity
and possibility. The formula 2φ is read as "φ is necessary" and 3φ is read as "φ is possible". The notions of possibility
and necessity are dual to each other, 3φ can be defined as ¬2¬φ. The weakest modal logic K extends propositional
logic by the axiom K2 and the necessitation rule Nec2 as follows
⊢ 2(φ → ψ) → (2φ → 2ψ),
(K2)
from ⊢ φ conclude ⊢ 2φ.
(Nec2)
The axiom K2 ensures that whenever φ → ψ and φ necessarily hold, then also ψ necessarily has to hold. The
necessitation rule Nec2 allows to infer the necessity of φ from any proof of φ and, hence, pushes any derivable logical
truth into the range of the modal operator 2. This principle is also known as logical awareness. Various modalities
like belief or knowledge can be described by adding additional axioms encoding the characteristic properties of the
respective modal operator. The following two axioms are useful to model knowledge and belief:
⊢ 2φ → φ,
(T2)
⊢ 2φ → 3φ.
(D2)
5Note, that ϕf lc does not imply ϕlc.
8
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
The axiom T2 and D2 relate necessity with the factual world. While the truth axiom T2 characterises knowledge
as it postulates that everything which is necessary is also factual, D2 characterises belief as it postulates the weaker
property that everything which is necessary is also possible. Under both axioms ⊢ 2⊥ → ⊥ holds, i.e. a necessary
contradiction yields also a factual contradiction.
Multi-modal logics are easily obtained by adding several modal operators with possibly different properties and can be
used to express the information of more than one agent. E.g., the formula ei:φ expresses that the piece of information
φ belongs to the modality ei. Modal operators can also be used to represent modalities referring to time. E.g., in the
formula Xφ the temporal modality X expresses that φ will hold in the next time step. An important representative of
a temporal extension is linear temporal logic (LTL).
In multi-agent logics the notions of common information and distributed information play an important role. While
common knowledge captures the information which is known to every agent ei, we are mainly interested in information
that is distributed within a group of agents E = {e1, . . . , en}. The distributed information within a group E contains
any piece of information that at least one of the agent e1, . . . , en has. Consequently, we introduce a set-like notion
for groups, where an agent e is identified with the singleton group {e} and the expression {e1, . . . , en}:φ is used to
denote that φ is distributed information within the group E. The distribution of information is axiomatised by
Note that groups may not be empty. The modal logic for distributed information contains for every group E at least
the axiom KE, the necessitation rule NecE, and the axiom DistE,F for any group F with E ⊆ F .
⊢ E:φ → F :φ, where E is a subgroup of F .
(DistE,F )
Justification Logics.
Justification logics [3] are variants of epistemic modal logics where the modal operators of
knowledge and belief are unfolded into justification terms. Hence, justification logics allow a complete realisation
of Plato's characterisation of knowledge as justified true belief. A typical formula of justification logic has the form
s:φ, where s is a justification term built from justification constants, and it is read as "φ is justified by s". The basic
justification logic J0 results from extending propositional logic by the application axiom and the sum axioms
⊢ s:(φ → ψ) → (t:φ → [s · t]:ψ),
(Appl)
⊢ s:φ → [s + t]:φ,
⊢ s:φ → [t + s]:φ,
(Sum)
where s, t, [s · t], [s + t], and [t + s] are justification terms which are assembled from justification constants using the
operators + and · according to the axioms. Justification logics tie the epistemic tradition together with proof theory.
Justification terms are reasonable abstractions for constructions of proofs. If s is a proof of φ → ψ and t is a proof
of φ then the application axiom postulates that there is a common proof, namely s · t, for ψ. Moreover, if we have a
proof s for φ and some proof t then the concatenations of both proofs, s + t and t + s, are still proofs for φ. In our
framework we were not able to derive any meaningful example using the sum axiom of justification logic. Therefore
this axiom is omitted in the following discussion.
Discussion. All instances of classical logical tautologies, like A ∨ ¬A and s:A ∨ ¬s:A, are provable in justification
logics. But in contrast to modal logics, justification logics do not have a necessitation rule. The lack of the necessitation
rule allows justification logics to break the principle of logical awareness, as s:(A ∨ ¬A) is not necessarily provable
for an arbitrary justification term s. Certainly, restricting the principle of logical awareness is attractive to provide a
realistic model of restricted logical resources. Since we are mainly interested in revealing and resolving conflicts, the
principle of logical awareness is indispensable in our approach.
Nevertheless, justification logic can simulate unrestricted logical awareness by adding proper axiom internalisation
rules ⊢ e:φ for all axioms φ and justification constants e. In such systems a weak variant of the necessitation rule of
modal logic holds: for any derivation ⊢ φ there exists a justification term t such that ⊢ t:φ holds. Since φ was derived
using axioms and rules only, also the justification term t is exclusively built from justification constants dedicated to
the involved axioms. Beyond that, t is hardly informative as it does not help to reveal external causes of a conflict.
Hence, we omit the axiom internalisation rule and add the modal axiom Kt and the modal necessitation rule Nect for
any justification term t to obtain a justification logic where each justification term is closed under unrestricted logical
awareness.
An important consequence of the proposed system is that · becomes virtually idempotent and commutative.6 These
insights allows us to argue merely about justification groups instead of justification terms. It turns out that a proper
reformulation of Appl with regard to justification groups is equivalent to DistE,F , finally yielding the same axioma-
tisation for distributed information and compound justifications.
6For any instance ⊢ s:(φ → ψ) → (s:φ → [s · s]:ψ) of Appl there is an instance ⊢ s:(φ → ψ) → (s:φ → s:ψ) of Ks in the
proposed system. Moreover, it is an easy exercise to show that any instance of ⊢ s:(φ → ψ) → (t:φ → [t · s]:ψ) is derivable in
the proposed system.
9
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
Belief Atoms, Belief Groups, and Belief Entities. So far, we argued that assembling distributed information and
compound justifications follow the same principle. In the following we even provide a unified concept for the building
blocks of both notions. A belief atom e is the least constituent of external information in our logic. To each e we assign
the modal operator e:. Hence, for any formula φ also e:φ is a formula saying "e has information φ". Belief atoms play
different roles in our setting. A belief atom may represent a sensor collecting information about the state of the world,
or it may represent certain operational rules as well as a certain goal of the system. The characteristic property of a
belief atom is that the information of a belief atom has to be accepted or rejected as a whole. Due to its external and
indivisible nature, e is the only source of evidence for its information. The only justification for information of e is e
itself. Consequently, e:φ can also be read as "e is the justification for φ". This is what belief atoms and justifications
have in common: either we trust a justification or not.
The information of a system is distributed among its belief atoms. The modal logic for distributed information allows
us to consider the information which is distributed over a belief group. While belief groups can be built arbitrarily from
belief atoms, we also introduce the concept of belief entities. A belief entity is either a belief atom, or a distinguished
group of belief entities. Belief entities are dynamically distinguished by a justification graph. In contrast to belief
groups, belief entities and belief atoms are not allowed to have inconsistent information. Hence a justification graph
allows us to restrict the awareness of extra-logical evidences -- so we can distinctively integrate logical resources that
have to be consistent.
Justification Graphs. Let V be a set of propositional variables and let E be the set of belief entities. The designated
subset EA of E denotes the set of belief atoms.
Definition 2 (Language of Justification Graphs). A formula φ is in the language of justification graphs if and only if
φ is built according to the following BNF, where A ∈ V and ∅ 6= E ⊆ E:
φ ::= ⊥ A (φ → φ) E:(φ) X(φ) P(φ) (φ)U(φ) (φ)S(φ).
Using the descending sequence of operator precedences (:, ¬, ∨, ∧, →, ↔), we can define the well-known logical
connectives ¬, ∨, ∧ and ↔ from → and ⊥. Often, we omit brackets if the formula is still uniquely readable. We define
→ to be right associative. For singleton sets {e} ⊆ E we also write e:φ instead of {e}:φ. The language allows the
usage of temporal operators for next time (X), previous time (P), until (U), and since (S). Operators like always in the
future (G) or always in the past (H) can be defined from the given ones.
Definition 3 (Justification Graph). A justification graph is a directed acyclic graph G whose nodes are belief entities
of E. An edge e 7→G f denotes that the belief entity e has the component f . The set of all direct components of an
entity e is defined as G(e) := {f e 7→G f }.
The leaf nodes of a justification graph are populated by belief atoms, i.e. for any belief entity e it holds e ∈ EA if and
only if G(e) = ∅.
Definition 4 (Axioms of a Justification Graph). Let G be a justification graph. The logic of a justification graph has
the following axioms and rules.
(i) As an extension of propositional logic the rule of modus ponens MP has to hold: from ⊢ φ and ⊢ φ → ψ
conclude ⊢ ψ. Any substitution instance of a propositional tautology φ is an axiom.
(ii) Belief groups are closed under logical consequence and follow the principle of logical awareness. Information is
freely distributed along the subgroup-relation. For any belief group E the axiom KE and the necessitation rule
NecE hold. For groups E and F with E ⊆ F the axiom DistE,F holds.
(iii) Belief entities are not allowed to have inconsistent information. Non-atomic belief entities inherit all information
of their components. For any belief entity e the axiom De holds. If E is a subgroup of the components of e, then
the axiom DistE,e holds.
(iv) In order to express temporal relation the logic for the justification graph includes the axioms of Past-LTL (LTL
with past operator). A comprehensive list of axioms can be found in [4].
(v) Information of a belief entity e ∈ E and time are related. The axiom (PRE) : ⊢ e:Pφ ↔ Pe:φ ensures that
every belief entity e correctly remembers its prior beliefs and establishes a principle which is also known as
perfect recall (e.g., see [5]).
Definition 5 (Proof). Let G be a justification graph. A proof (derivation) of φ in G is a sequence of formulae
φ1, . . . , φn with φn = φ such that each φi is either an axiom of the justification graph or φi is obtained by ap-
plying a rule to previous members φj1 , . . . , φjk with j1, . . . , jk < i. We will write ⊢G φ if and only if such a sequence
exists.
Definition 6 (Proof from a set of formulae). Let G be a justification graph and Σ be a set of formulae. The relation
Σ ⊢G φ holds if and only if ⊢G (σ1 ∧ · · · ∧ σk) → φ for some finite subset {σ1, . . . , σk} ⊆ Σ with k ≥ 0.
10
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
Definition 7 (Consistency with respect to a justification graph). Let G be a justification graph.
(i) A set Σ of formulae is G-inconsistent if and only if Σ ⊢G ⊥. Otherwise, Σ is G-consistent. A formula φ is
G-inconsistent if and only if {φ} is G-inconsistent. Otherwise, φ is G-consistent.
(ii) A set Σ of formulae is maximally G-consistent if and only if Σ is G-consistent and for all φ 6∈ Σ the set Σ ∪ {φ}
is G-inconsistent.
Semantics. Let S be the state space, that is the set of all possible states of the world. An interpretation π over S is a
mapping that maps each state s to a truth assignment over s, i.e. π(s) ⊆ V is the subset of all propositional variables
which are true in the state s. A run over S is a function r from the natural numbers (the time domain) to S. The set of
all runs is denoted by R.
Definition 8. Let G be a justification graph. A Kripke structure M for G is a tuple M = (S, R, π, (7→e)e∈E) where
(i) S is a state space,
(ii) R is the set of all runs over S,
(iii) π is an interpretation over S,
(iv) each 7→e in (7→e)e∈E is an individual accessibility relation 7→e⊆ S × S for a belief entity e in E.
Definition 9 (Model for a Justification Graph). Let M = (S, R, π, (7→e)e∈E) be a Kripke structure for the justification
graph G, where
(i) 7→e is a serial relation for any belief entity e ∈ E,
(ii) 7→E is defined as 7→E = Te∈E 7→e for any belief group E ⊆ E,
(iii) 7→e ⊆ 7→E holds for all non-atomic belief entities e ∈ E \ EA and any subgroup E ⊆ G(e).
We recursively define the model relation (M, r(t)) =G φ as follows:
:⇐⇒ Q ∈ π(r(t)).
(M, r(t)) 6=G ⊥.
(M, r(t)) =G Q
(M, r(t)) =G φ → ψ :⇐⇒ (M, r(t)) =G φ implies (M, r(t)) =G ψ.
(M, r(t)) =G E:φ
(M, r(t)) =G Xφ
(M, r(t)) =G Pφ
(M, r(t)) =G φUψ
:⇐⇒ (M, r′(t)) =G φ for all r′ with r(t′) 7→E r′(t′) for all t′ ≤ t.
:⇐⇒ (M, r(t + 1)) =G φ.
:⇐⇒ (M, r(t′)) =G φ for some t′ with t′ + 1 = t.
:⇐⇒ (M, r(t′)) =G ψ for some t′ ≥ t and
(M, r(t)) =G φSψ
:⇐⇒ (M, r(t′)) =G ψ for some 0 ≤ t′ ≤ t and
(M, r(t′′)) =G φ for all t′′ with t ≤ t′′ < t′.
(M, r(t′′)) =G φ for all t′′ with t′ < t′′ ≤ t.
When (M, r(t)) =G φ holds, we call (M, r(t)) a pointed model of φ for G. If (M, r(0)) is a pointed model of φ for
G, then we write (M, r) =G φ and say that the run r satisfies φ. Finally, we say that φ is satisfiable for G, denoted
by =G φ if and only if there exists a model M and a run r such that (M, r) =G φ holds.
Proposition 1 (Soundness and Completeness). The logic of a justification graph G is a sound and complete axioma-
tisation with respect to the model relation =G. That is, a formula φ is G-consistent if and only if φ is satisfiable for
G.
While the soundness proof is straightforward, a self-contained completeness proof involve lengthy sequences of vari-
ous model constructions and is far beyond the page limit. However, it is well-known, (e.g., [6]), that KD
n , the n-agent
extension of K with distributive information is a sound and complete axiomatisation with respect to the class of Kripke
structures having n arbitrary accessibility relations, where the additional accessibility relations for groups are given as
the intersection of the participating agents, analogously to Def. 9.(ii). Also the additional extension KDD
n with DE
for any belief group E is sound and complete with respect to Kripke structures having serial accessibility relations,
analogously to Def. 9.(i). The axioms of justification graph are between these two systems. Def. 9.(iii) explicitly
allows belief entities to have more information than its components. Various completeness proofs for combining LTL
and epistemic logics are given e.g., in [5].
Extracting Justifications. Let Σ = {σ1, . . . , σn} be a finite set of formulae logically describing the situation which
is object of our investigation. Each formula σi ∈ Σ encodes information of belief atoms (σi ≡ ei:φi with ei ∈ EA),
facts (σi ≡ φi where φi does not contain any epistemic modal operator), or is an arbitrary Boolean combinations
thereof. Further, let G be a justification graph such that Σ is G-consistent and e be a non-atomic belief entity of G.
For any formula φ we may now ask whether φ is part of the information of e. If there is a proof Σ ⊢G e:φ, then φ is
included in e's information. To extract a justification for e:φ we use that Σ ∪ {¬e:φ} is G-inconsistent and accordingly
11
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
unsatisfiable for G. If we succeed in extracting a minimal unsatisfiable core Σ′ ⊆ Σ ∪ {¬e:φ} a minimal inconsistency
proof can be recovered, from which finally the used justifications are extracted.
The following proposition allows to use SAT/SMT-solvers for a restricted setting and has been used in our case study.
Proposition 2 (SAT Reduction). Let Σ = {σ1, . . . , σn} be a set of formulae such that each element σi is of the form
ei:φi with ei ∈ EA and φi does not contain any epistemic modal operators. Further, let e be an arbitrary belief
entity that does not occur in Σ. Then G = {e 7→G eiei occurs in Σ} is a justification graph for Σ if and only if
Φ = {φ1, . . . φn} is satisfiable over the non-epistemic fragment of the logic of justification graphs.
Proof. The satisfiability relation for the non-epistemic fragment is independent of the accessibility relations 7→e,
e ∈ E and, consequently, also independent of G. In particular, Φ is satisfiable if and only if there exists a model
M ′ = (S, R, π) and a run r such that (M ′, r) = Φ.
Let G = {e 7→G eiei occurs in Σ} be a graph.
Let us first assume that G is a justification graph for Σ. Then according to Def. 9 there exists a Kripke structure
M = (S, R, π, (7→e)e∈E) and a run r such that (M, r) =G ei:φi for all ei:φi ∈ Σ. Hence, we have (M, r′) =G φi
for all r′ with r 7→ei r′. Furthermore, from item (i) and (iii) of Def. 9 we observe that 7→e is not empty and
7→e ⊆ 7→e1 ∩ · · · ∩ 7→en . Hence, there exists at least one run r′ that satisfies all formulae in Φ. Since Φ does not
contain epistemic operators, we found a model M ′ = (S, R, π) and a run r′ such that (M ′, r′) = Φ.
For the other direction, let us assume that there exists some model M ′ = (S, R, π) and a run r such that (M ′, r) = Φ.
We extend M ′ to a Kripke structure M = (S, R, π, (7→e′ )e′∈E) by setting r 7→e′ r′ for all e′ ∈ E if and only if r = r′
for all r, r′ ∈ Π. Then (M, r) =G ei:φi for all ei:φi ∈ Σ since (M ′, r′) = φi holds for all r 7→ei r′ by construction
of the accessibility relations. Moreover, since all accessibility relations are equal and reflexive, 7→e is serial.
4 Identifying and Analysing Conflicts
In this section we first present an abstract algorithm for the conflict resolution of Sect. 2 that starts at level (C1) and
proceeds resolution stepwise up to level (C4). We then sketch our small case study where we applied an implementa-
tion of the abstract algorithm.
4.1 Analysing Conflicts
For the analysis of conflicts we employ SMT solvers. Prop. 2 reduces the satisfiability of a justification graph to a
SAT problem. To employ SMT solving for conflict analysis, we encode the (real and possible) worlds of Sect. 2 via
logic formulae as introduced in Sect. 3. Each state si is represented as a conjunction of literals, si ≡ V v ∧ V ¬v′.
Introducing a dedicated propositional variable vt for each v ∈ V and time step t allows us to obtain a formula
describing a finite run on M . A predicate of the form V(s,s′)∈T(st → s′
t+1) encodes the transition relation T. The
effect of performing an action at at state s is captured by a formula of the form at → (st → s′
t+1). Using this we can
encode a strategy δ in a formula ψδ such that its valuations represent runs of M according to δ. All runs according to
δ achieve goals Φ if and only if ψδ ∧ ¬Φ is unsatisfiable. These logical encodings are the main ingredients for using
a SAT solver for our conflict analysis. Since there are only finitely many possible strategies, we examine for each
strategy which goals can be (maximally) achieved in a world M or in a set of worlds M. Likewise we check whether
A has a winning strategy that is compatible with the strategies A believes B might choose.
Since we iterate over all possible worlds for our conflict analysis, we are interested in summarising possible worlds.
We are usually not interested in all vt -- e.g.
the speed of B may at times t be irrelevant. We are hence free to
ignore differences in vt in different possible worlds and are even free to consider all valuations of vt, even if A does
not consider them possible. This insight leads us to a symbolic representation of the possible worlds, collecting the
relevant constraints. Now the justification graph groups the constraints that are relevant, with other words, e:φ and
e′:¬φ will not be components of the same justification graph if the valuation of φ is relevant. In the following we hence
consider the maximal consistent set of possible worlds, meaning encodings of possible worlds that are uncontradictory
wrt. the relevant propositions, which are specified via the justification graph.
4.2 Algorithmic approach
In this section, we sketch an abstract algorithm for the conflict resolution at levels (C1) to (C4) as in Sect. 2. Note
that we do not aim with Alg. 1 for efficiency or optimal solutions but aim to illustrate how satisfiability checks can be
employed to analyse our conflicts.
12
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
A , Φmax
B , AA, AB )
ΣM ← POSSIBLEWORLDS(Σ, AA, AB)
∆A ← STRATA(AA, AB, ΣM, Φmax
A )
C ← ∅
for all (δA, δ′
Algorithm 1 Determining winning strategy based on observations, goals, and possible actions.
1: function FINDSTRATEGY(Σ, Φmax
2:
3:
4:
5:
6:
7:
8:
9:
E ← TESTIFNOTWINNING((δA, δ′
if E 6= ∅ then
C ← C ∪ {E}
∆A = ∆A \ {(δA, δ′
A
B), ΣM, ΦA, Φmax
B) ∈ ∆A with r((δA, δ′
B), ΣM) = ΦA ∈ Φmax
⊲ construct {(δA, δ′
B , AA, AB)
B) r((δA, δ′
⊲ (δA, δ′
do
B)}
B) is not winning for all ΣM ∈ ΣM, i.e. r((δA, δ′
B), ΣM) = ΦA with ΦA ∈ Φmax
⊲ construct set of possible worlds
A }
⊲ set of conflict causes
⊲ cf. Alg. 2
B), ΣM) 6= ΦA
⊲ memorize justifications E
10:
11:
12:
13:
14:
15:
16:
17:
if ∆A = ∅ then
for i ∈ [1, 2, 3, 4] do
′, Φmax
Σ′, Φmax
if (Σ′ 6= Σ) ∨ (Φmax
B
A
′ ← FIXCONFLICT(C, (Ci), Σ, Φmax
B )
A , Φmax
′ 6= Φmax
A ) ∨ (Φmax
A
B
∆A ← FINDSTRATEGY(Σ′, Φmax
′, Φmax
if ∆A 6= ∅ then
B
A
′ 6= Φmax
B ) then
′, AA, AB)
⊲ A is in conflict with B
⊲ traverse resolution levels
⊲ cf. Alg. 3
⊲ new information generated
⊲ new attempt with new information
⊲ new attempt was successful, stop and return
break
return ∆A
⊲ select (δA, δ′
B) ∈ ∆A to reach some goal in Φmax
A
Ci ← Ci+1
initial justified
information base ΣI
∪
Σ
constraints
required information ΣR
ΣMΣMΣMΣM
ΣM
∆ ⊆ ∆A
winning?
∆A
[yes]
∆
[no] justifications
resolution
at level Ci
Figure 6: Abstract resolution process with information base, possible worlds and strategies.
The following algorithms describe how we deal with logic formulae encoding sets of possible worlds, sets of runs on
them, etc. to analyse conflicts (cf. Def. 1, p. 5) via SMT solving. We use ΣM to refer to a formula that encodes a
maximal consistent set of possible worlds (cf. Sect. 4.1), i.e., that corresponds to a justification graph. We use ΣM to
refer to a set of formulas ΣM ∈ ΣM that encode the set of possible worlds M structured into sets of possible worlds
via justification graphs. We use M and ΣM synonymously. Also we often do not distinguish between ΣM and M --
neglecting that ΣM represents a set of worlds that are like M wrt to the relevant constraints.
Fig. 6 provides an overview of the relation between the initial information base ΣI of agent A, its set ΣM of possible
worlds ΣM , winning strategies, resolution, and stepwise update of the information ΣR during our conflict resolution
process. The initial information base defines the set of possible worlds M, which is organised in sets of maximal
consistent worlds ΣM . Based on ΣM, A's set of strategies ∆A is checked whether it comprises a winning strategies
in presence of an agent B that tries to achieve its own goals. If no such winning strategy exists, A believes to be
in conflict with B. At each level (Ci) the resolution procedure tries to determine information ΣR of level (Ci) to
resolve the conflict. If the possible worlds are enriched by this information, the considered conflict vanishes. The new
information is added to the existing information base and the over-all process is re-started again until either winning
strategies are found or ΣR is empty.
How to find a believed winning strategy. Alg. 1 finds a winning strategy of agent A for a goal ΦA in ΣM tolerating
that B follows an arbitrary winning strategy in ΣM ∈ ΣM for its goals, i.e. it finds a strategy that satisfies ΦA in all
possible worlds ΣM where ΦA is maximal for ΣM and in each possible world ΣM ∈ ΣM agent B may also follow a
winning strategy for one of its maximal goals ΦB. If such a strategy cannot be found, A believes to be in conflict with
B (cf. Def. 1).
Input for the algorithm is (i) a set Σ of formulae describing the current belief of A, e.g. its current observations and its
history of beliefs -- we call it the information base in the sequel -- , (ii) a set of goals Φmax
of A that is maximal in M,
of B that is maximal for a ΣM ∈ ΣM, (iv) a set of possible actions AA for A and
(iii) a set of believed goals Φmax
(v) a set of believed possible actions AB for B.
B
A
First (L. 2 of Alg. 1) is to construct sets of maximal consistent sets of possible worlds that together represent M. In L. 3
the set ∆A is determined, which is the set of strategies accomplishing a maximal goal combination for A assuming B
agrees to help, i.e., all winning strategies (δA, δ′
in all possible worlds ΣM ∈ ΣM, where
ΦA is maximal for M.
B) that satisfy ΦA ∈ Φmax
A
13
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
for all ΣM ∈ ΣM do
∆B ← STRATB(AA, AB, ΣM , Φmax
B )
for all (δ′
Algorithm 2 Test if a strategy is winning in all possible worlds.
1: function TESTIFNOTWINNING((δA, δ′
B), ΣM, ΦA, Φmax
2:
3:
4:
5:
6:
7:
8:
9:
if r((δA, δB), ΣM ) 6= ΦA ∪ ΦB then
for all ΦB ∈ Φmax
A, δB) ∈ ∆B do
B with r((δ′
return ∅
else
⊲ (δA, δ′
return GETJUSTIFICATIONS((δA, δB) 6= ΦA ∪ ΦB)
B , AA, AB)
⊲ construct {(δ′
A, δB) r((δ′
A, δB), ΣM ) = ΦB with ΦB ∈ Φmax
B }
A, δB), ΣM ) = ΦB and ΦB is maximal in ΣM do
B) is not winning for all M and all (δ′
A, δB)
Algorithm 3 Try to fix a conflict by resolving contradictions.
1: function FIXCONFLICT(C, (Ci), Σ, Φmax
2:
3:
4:
C ← C \ {E}
Σ, Φmax
B ← RESOLVE(Σ, E, Φmax
A , Φmax
B )
for E ∈ C do
A , Φmax
A , Φmax
B , (Ci))
⊲ try resolution according to level (Ci)
5:
return Σ, Φmax
A , Φmax
B
B
in ΣM .
In lines 5 ff. we examine whether one of A's strategies (where B is willing to help) works even when B follows its
strategy to achieve one of its maximal goals ΦB ∈ Φmax
To this end TESTIFNOTWINNING is called for all of A's winning strategies (δA, δ′
B) ∈ ∆A (L. 6). The function
TESTIFNOTWINNING performs this test iteratively for one maximal consistent set of worlds ΣM (Alg. 2 L. 2). Let
∆B be the set of joint strategies achieving a goal ΦB ∈ Φmax
that is maximal in ΣM . We check the compatibility
of A's strategy (δA, δ′
B) is compatible to all of B's
A, δB) if all joint strategies (δA, δB) achieve the maximal goals for A and B (Alg. 2 L. 6).7 If the joint
strategies (δ′
strategy (δA, δB) is not a winning strategy for the joint goal ΦA ∪ ΦB (Alg. 2 L. 6), the function GETJUSTIFICATIONS
extracts the set of justifications for this conflict situation (Alg. 2 L. 7). The set of justifications is added to the set of
conflict causes C (Alg. 1 L. 8.). Since strategy (δA, δ′
B) is not compatible to all of B's strategies, it is hence not further
considered as a possible conflict-free strategy for A (Alg. 1 L. 9).
A, δB) ∈ ∆B (Alg. 2 L. 3). A strategy of A (δA, δ′
B) to every (δ′
B
A strategy that remains in ∆A at Alg. 1 L. 10 is a winning strategy for one of A's goals in all possible worlds ΣM
regardless of what maximal goals B tries to achieve in ΣM . However, if ∆A is empty at Alg. 1 L. 10 , A is in a
(believed) conflict with B (Def. 1). In this case, conflict resolution is attempted (cf. lines 10 ff. in Alg. 1). Function
FIXCONFLICT from Alg. 3 is called with the set of conflict causes, the current conflict resolution level, and the current
information base and goals. For each conflict cause, an attempt of resolution is made by function RESOLVE. The
conflict is analysed to identify whether adding/updating information of the current resolution level helps to resolve the
conflict. If there are several ways to resolve a conflict, justifications can be used to decide which resolution should be
chosen. Note that conflict resolution hence means updating of the information base Σ or goal sets Φmax
A
and Φmax
B .
Line 13 of Alg. 1 checks if some new information was obtained from the resolution procedure. If not, resolution will
be restarted at the next resolution level. If new information was obtained, FINDSTRATEGY is called with the updated
information. If the result is a non-empty set of strategies, the algorithm terminates by returning them as (believed)
winning strategies for A. However, if the result is the empty set, resolution is restarted at the next resolution level. If
∆A is empty at level (C4), the conflict cannot be resolved and the algorithm terminates.
Termination. Alg. 1 eventually terminates under the following assumptions. The first assumption is that the set of
variables V and hence the information base Σ is finite. In this case, the construction of maximal consistent possible
worlds ΣM terminates since there is a finite number of possible consistent combinations of formulae and the time
horizon for the unrolling of a possible world ΣM is bounded.
Together with finite sets Φmax
B , AA, and AB, the construction of strategies, i.e. functions STRATA and STRATA,
terminates since there are only finite numbers of combinations of input histories and output action and there is only a
finite number of goals to satisfy. All loops in algorithms 1, 2, and 3 hence iterate over finite sets.
A , Φmax
7Note that according to Sect. 2.1, we have ΦB = true if B cannot achieve any goal. This reflects that A cannot make any
assumption about B's behaviour in such a situation.
14
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
The extraction of justifications terminates since runs are finite and consequently the number of actions involved in the
run, too. Furthermore, for each state in a run, there are only a finite number of propositions that apply. Together with
a finite number of propositions representing the goal, GETJUSTIFICATIONS can simply return the (not necessarily
minimal) set of justifications from all these finite many formulae as a naive approach.
B
Alg. 1 terminates if a non-empty set ∆A is derived by testing and/or resolution, or if a fixed point regarding Σ, Φmax
A ,
and Φmax
is reached. Since all other loops and functions terminate, the only open aspect is the fixed point whose
achievement depends on RESOLVE. We assume that Alg. 1 is executed at a fixed time instance s.t. A's perception of
the environment does not change during execution. Thus, Σ contains only a finite number of pieces of information
to share. If we assume that that sharing information leads only to dismissing possible worlds rather then considering
more worlds possible, then this a monotonic process never removing any information.8 Furthermore, we assume that
the partial order of goals leads, if necessary, to a monotonic process of goal negotiation which itself can repeated finite
many times until no further goals can be sacrificed or adopted from B. Thus, if ∆A remains to be the empty set in
line 15, the fixed point will eventually be reached.
Furthermore, we do not consider any kind of race conditions occurring from concurrency, e.g. deadlock situations
where A can't serve B's request because it does not know what its strategy will be since A wait's for B's respond, and
vice versa.
So in summary, the algorithm terminates under certain artificial assumptions but cannot determine a resolution in case
without an outside arbiter. In practice such a conflict resolution process has to be equipped with time bounds and
monitors. We consider these aspects as future work.
4.3 Case study
We implemented the algorithm sketched above in a Java program employing Yices [1] to determine contradictions and
analysed variations of a toy example to evaluate and illustrate our approach.
We modelled a system of two agents on a two lane highway. Each agent is represented by its position and its lane.
Each agent has a set of actions: it can change lane and drive forward with different speeds. We captured this via a
discrete transition relation where agents hop from position to position. The progress of time is encoded via unrolling,
that is we have for each point in time a corresponding copy of a variable to hold the value of the respective attribute at
that time. Accordingly the transition relation then refers to these copies.
Since we analyse believed conflicts of an agent, we consider several worlds. In other words, we consider several
variations of a Yices model. Each variation represents a justification graph summarising the maximal consistent set of
evidences and thereby representing a set of worlds which is justified by this set of evidences.
We modify the Yices file by adding additional constraints according to the algorithm Sect. 4.2. For the steps (C1) to
(C4) we add constraint predicates, e.g., that encode that information about certain observations have been communi-
cated by say B to A, constraints that specify that B tells A it will decelerate at step 4 and constraints that encode goal
combinations.
We employed Yices to determine whether there is conflict. The key observation is: If Yices determines that it holds
that ¬ϕ is satisfiable in our system model, then there is the possibility that the goal is not achieved -- otherwise each
evolution satisfies ϕ and there is a winning strategy for the model.
5 Related work
Studying Traffic Conflicts. According to Tiwari in his 1998 paper [7] studying traffic conflicts in India, one of the
earliest studies concerned with traffic conflicts is the 1963 paper [8] of Perkins and Harris. It aims to predict crashes
in road traffic and to obtain a better insight to causal factors. The term traffic conflict is commonly used according
to [7] as "an observable situation in which two or more road users approach each other in space and time to such an
extent that a collision is imminent if their movements remain unchanged" [9]. In this paper we are interested in a
more general and formal notion of conflict. We are not only interested in collisions-avoidance but more generally in
situations where traffic participants have to cooperate with each other in order to achieve their goals -- which might be
collision-freedom. Moreover, we aim to provide a formal framework that allows to explain real world observations as
provided by, e.g., the studies of [7, 8].
Tiwari also states in [7] that it is necessary to develop a better understanding of conflicts and conjectures that illusion
of control [10] and optimism bias theories like in [11] might explain fatal crashes. In this paper we develop a formal
8Otherwise the set of already examined worlds can be used to define a fixed point.
15
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
framework that allows us to analyse conflicts based on beliefs of the involved agents, -- although supported by our
framework -- we here do not compare the real world evolution with the evolution that an agent considers possible.
Instead we analyse believed conflicts, that are conflicts which an agent expects to occur based on its beliefs. Such
conflicts will have to be identified and analysed by prediction components of the autonomous vehicles architecture,
especially in settings where misperception and, hence, wrong beliefs are possible.
In [12] Sameh et al. present their approach to modelling conflict resolution as done by humans in order to generate
realistic traffic simulations. The trade-off between anticipation and reactivity for conflict resolution is analysed in
[13] in order to determine trajectories for vehicles at an intersection. Both works [12, 13] focus on conflicts leading
to accidents. Regarding the suggested resolution approaches, our resolution process suggests cooperation steps with
increasing level cooperation. This resolution process is tailored for autonomous vehicles that remain autonomous
during the negation process.
Strategies and Games. For strategy synthesis Finkbeiner and Damm [14] determined the right perimeter of a world
model. The approach aims to determine the right level of granularity of a world model allowing to find a remorse-free
dominant strategy. In order to find a winning (or remorse-free dominant) strategy, the information of some aspects
of the world is necessary to make a decision. We accommodated this as an early step in our resolution protocol.
Moreover in contrast to [14], we determine information that agent A then want requests from agent B in order to
resolve a conflict with B -- there may still be no winning (or remorse-free dominant) strategy for all goals of A.
In [15] Finkbeiner et. al.presented an approach to synthesise a cooperative strategy among several processes, where
the lower prioritised process sacrifices its goals when a process of higher priority achieves its goals. In contrast to
[15] we do not enforce a priority of agents but leave it open how a conflict is resolved in case not all their goals are
achievable. Our resolution process aims to identify the different kinds of conflict as introduced in Sect. 2 that arise
when local information and beliefs are taken into account and which not necessarily imply that actually goals have to
be sacrificed.
We characterize our conflict notion in a game theoretic setting by considering the environment of agents A and B as
adversarial and compare two scenarios where (i) the agent B is cooperative (angelic) with the scenario where (ii) B is
not cooperative and also not antagonistic but reasonable in following a strategy to achieve its own goals. As Brenguier
et al. in [16] remark, a fully adversarial environment (including B) is usually a bold abstraction. By assuming in (ii)
that B maximises its own goals -- we assume that B follows a winning strategy for its maximal accomplishable goals.
So we are in a similar mind set than at assume-guarantee [17] and assume-admissible [16] synthesis. Basically we
consider the type of strategy (winning/admissible/dominant) as exchangeable, the key aspect of our definition is that
goals are not achievable but can be achieved with the help of the other.
Logics.
Justification logic was introduced in [3, 18] as an epistemic logic incorporating knowledge and belief modal-
ities into justification terms and extends classical modal logic by Plato's characterisation of knowledge as justified true
belief. However, even this extension might be epistemologically insufficient as Gettier already pointed out in 1963
[19]. In [20] a combination of justification logics and epistemic logic is considered with respect to common knowl-
edge. The knowledge modality Ki of any agent i inherits all information that are justified by some justification term
t, i.e. t:φ → Kiφ. In such a setting any justified information is part of common knowledge. Moreover, justified com-
mon knowledge is obtained by collapsing all justification terms into one modality J and can be regarded as a special
constructive sort of common knowledge. While our approach neglects the notion of common information, we use a
similar inheritance principle where a belief entity inherits information of its components, cf. Def. 4.(iii). A comparison
of the strength of this approach with different notions of common knowledge can be found in [21]. While justification
logic and related approaches [22, 23], aim to restrict the principle of logical awareness and the related notion of logical
omniscience, we argue in Sec. 3 that the principle of logical awareness as provided by modal logic is indispensable
in our approach. A temporal (LTL-based) extension of justification logic has been sketched in [24]. This preliminary
work differs from our approach wrt. the axiom systems used for the temporal logic part and the justification / modal
logic part, cf. the logic of justification graphs axiomatised in Section 3. Our logic and its axiomatisation incorporates
a partial order on the set of beliefs that underlies their prioritization during conflict resolution, which contrasts with
the probabilistic extension of justification logic outlined in [25].
6 Conclusion
Considering local and incomplete information, we presented a new notion of conflict that captures situations where
an agent believes it has to cooperate with another agent. We proposed steps for conflict resolution with increasing
level of cooperation. Key for conflict resolution is the analysis of a conflict, tracing and identifying contradictory
evidences. To this end we presented a formal logical framework unifying justifications with modal logic. Alas, to the
16
Justification Based Reasoning in Dynamic Conflict Resolution
A PREPRINT
authors' best knowledge there are no efficient satisfiability solvers addressing distributed information so far. However,
we exemplified the applicability of our framework in a restricted but non-trivial setting. On the one hand, we plan to
extend this framework by efficient implementations of adapted satisfiability solvers, on the other hand by integrating
richer logics addressing decidable fragments of first order logic, like linear arithmetic, and probabilistic reasoning.
References
[1] B. Dutertre. Yices 2.2. In Computer Aided Verification, pages 737 -- 744. Springer, 2014.
[2] J. Galtung. Violence, Peace, and Peace Research. Journal of Peace Research, 6(3):167 -- 191, September 1969.
[3] S. N. Artemov. Justified common knowledge. Theor. Comput. Sci., 357(1-3):4 -- 22, 2006.
[4] O. Lichtenstein, A. Pnueli, and L. Zuck. The glory of the past.
In Workshop on Logic of Programs, pages
196 -- 218. Springer, 1985.
[5] R. Fagin, J. Y. Halpern, Y. Moses, and Moshe Y. Vardi. Reasoning About Knowledge. MIT Press, 2003.
[6] J. Gerbrandy. Distributed knowledge. In Twendial 1998: Formal Semantics and Pragmatics of Dialogue, vol-
ume 98, pages 111 -- 124, 1998.
[7] G. Tiwari, D. Mohan, and J. Fazio. Conflict analysis for prediction of fatal crash locations in mixed traffic
streams. Accident Analysis & Prevention, 30(2):207 -- 215, 1998.
[8] J. I. Harris and S. R. Perkins. Traffic conflict characteristics: accident potential at intersections. Highway
Research Board, 225:35 -- 43, 1967.
[9] F. H. Amundsen and C. Hyden. Proc. of the first workshop on traffic conflicts, oslo, norway, 1977. 1st Workshop
on Traffic Conflicts, LTH Lund.
[10] J. E. Langer. The illusion of control. Journal of Personality and Social Psychology, 32:311 -- 328, 08 1975.
[11] D. M. DeJoy. The optimism bias and traffic accident risk perception. Accident Analysis & Prevention, 21(4):333
-- 340, 1989.
[12] S. El hadouaj, A. Drogoul, and S. Espi´e. How to combine reactivity and anticipation: The case of conflicts
resolution in a simulated road traffic. In Multi-Agent-Based Simulation, pages 82 -- 96. Springer, 2001.
[13] N. Murgovski, G. R. de Campos, and J. Sjoberg. Convex modeling of conflict resolution at traffic intersections.
In 2015 54th IEEE Conference on Decision and Control (CDC), pages 4708 -- 4713, Dec 2015.
[14] W. Damm and B. Finkbeiner. Does it pay to extend the perimeter of a world model? In FM 2011: Formal
Methods, pages 12 -- 26. Springer, 2011.
[15] W. Damm, B. Finkbeiner, and A. Rakow. What you really need to know about your neighbor. In Proc. Fifth
Workshop on Synthesis, SYNT@CAV 2016, volume 229 of EPTCS, pages 21 -- 34, 2016.
[16] R Brenguier, J.-F. Raskin, and O. Sankur. Assume-admissible synthesis. Acta Informatica, 54(1):41 -- 83, Feb
2017.
[17] K. Chatterjee and T. A. Henzinger. Assume-guarantee synthesis. In Tools and Algorithms for the Construction
and Analysis of Systems, pages 261 -- 275. Springer, 2007.
[18] S. N. Artemov. The logic of justification. Rew. Symb. Logic, 1(4):477 -- 513, 2008.
[19] E. L. Gettier. Is justified true belief knowledge? Analysis, 23(6):121 -- 123, 1963.
[20] S. Artemov and E. Nogina. Introducing justification into epistemic logic. Journal of Logic and Computation,
15(6):1059 -- 1073, 2005.
[21] E. Antonakos. Justified and common knowledge: Limited conservativity. In International Symposium on Logical
Foundations of Computer Science, pages 1 -- 11. Springer, 2007.
[22] R. Fagin and J. Y. Halpern. Belief, awareness, and limited reasoning. Artificial intelligence, 34(1):39 -- 76, 1987.
[23] S. Artemov and R. Kuznets. Logical omniscience as a computational complexity problem. In Proc. of the 12th
Conference on Theoretical Aspects of Rationality and Knowledge, pages 14 -- 23. ACM, 2009.
[24] S. Bucheli, M. Ghari, and T. Studer. Temporal justification logic. In Proc. of the 9th Workshop on Methods for
Modalities, January 2017, volume 243 of EPTCS, pages 59 -- 74. Open Publishing Association, 2017.
[25] I. Kokkinis, Z. Ognjanovic, and T. Studer. Probabilistic justification logic. In Logical Foundations of Computer
Science - International Symposium, LFCS 2016. Proc., volume 9537 of LNCS, pages 174 -- 186. Springer, 2016.
17
|
1905.04205 | 1 | 1905 | 2019-05-10T15:04:48 | On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes | [
"cs.MA",
"cs.AI",
"cs.LG"
] | Self-adaptation has been proposed as a mechanism to counter complexity in control problems of technical systems. A major driver behind self-adaptation is the idea to transfer traditional design-time decisions to runtime and into the responsibility of systems themselves. In order to deal with unforeseen events and conditions, systems need creativity -- typically realized by means of machine learning capabilities. Such learning mechanisms are based on different sources of knowledge. Feedback from the environment used for reinforcement purposes is probably the most prominent one within the self-adapting and self-organizing (SASO) systems community. However, the impact of other (sub-)systems on the success of the individual system's learning performance has mostly been neglected in this context. In this article, we propose a novel methodology to identify effects of actions performed by other systems in a shared environment on the utility achievement of an autonomous system. Consider smart cameras (SC) as illustrating example: For goals such as 3D reconstruction of objects, the most promising configuration of one SC in terms of pan/tilt/zoom parameters depends largely on the configuration of other SCs in the vicinity. Since such mutual influences cannot be pre-defined for dynamic systems, they have to be learned at runtime. Furthermore, they have to be taken into consideration when self-improving the own configuration decisions based on a feedback loop concept, e.g., known from the SASO domain or the Autonomic and Organic Computing initiatives. We define a methodology to detect such influences at runtime, present an approach to consider this information in a reinforcement learning technique, and analyze the behavior in artificial as well as real-world SASO system settings. | cs.MA | cs |
On the Detection of Mutual Influences and Their Consideration in
Reinforcement Learning Processes
STEFAN RUDOLPH, University of Augsburg, Germany
SVEN TOMFORDE, University of Kassel, Germany
JÖRG HÄHNER, University of Augsburg, Germany
Self-adaptation has been proposed as a mechanism to counter complexity in control problems of technical systems. A major driver
behind self-adaptation is the idea to transfer traditional design-time decisions to runtime and into the responsibility of systems
themselves. In order to deal with unforeseen events and conditions, systems need creativity -- typically realized by means of machine
learning capabilities. Such learning mechanisms are based on different sources of knowledge. Feedback from the environment used for
reinforcement purposes is probably the most prominent one within the self-adapting and self-organizing (SASO) systems community.
However, the impact of other (sub-)systems on the success of the individual system's learning performance has mostly been neglected
in this context.
In this article, we propose a novel methodology to identify effects of actions performed by other systems in a shared environment
on the utility achievement of an autonomous system. Consider smart cameras (SC) as illustrating example: For goals such as 3D
reconstruction of objects, the most promising configuration of one SC in terms of pan/tilt/zoom parameters depends largely on the
configuration of other SCs in the vicinity. Since such mutual influences cannot be pre-defined for dynamic systems, they have to be
learned at runtime. Furthermore, they have to be taken into consideration when self-improving the own configuration decisions based
on a feedback loop concept, e.g., known from the SASO domain or the Autonomic and Organic Computing initiatives.
We define a methodology to detect such influences at runtime, present an approach to consider this information in a reinforcement
learning technique, and analyze the behavior in artificial as well as real-world SASO system settings.
CCS Concepts: • Computer systems organization → Self-organizing autonomic computing; • Theory of computation →
Multi-agent learning;
Additional Key Words and Phrases: Mutual Influences, Self-Organization, Self-Integration, Self-Adaption, Machine Learning, Organic
Computing, Interwoven Systems
ACM Reference Format:
Stefan Rudolph, Sven Tomforde, and Jörg Hähner. 2010. On the Detection of Mutual Influences and Their Consideration in Reinforcement
Learning Processes. ACM Trans. Autonom. Adapt. Syst. 9, 4, Article 39 (March 2010), 38 pages. https://doi.org/0000001.0000001
1 INTRODUCTION
The research field of self-adaptive and self-organizing (SASO) systems emerged as response to the dramatic change
Information and Communication Technology (ICT) has undergone within the last two decades: Back then we started
with single isolated and fully comprehensible systems with well defined system boundaries. Afterwards, ICT components
have been continuously coupled with each other and single monolithic solutions have been replaced by collections of
autonomous entities or agents that communicate with each other and cooperatively solve tasks. Driven by initiatives
Authors' addresses: Stefan Rudolph, University of Augsburg, Eichleitnerstr. 30, Augsburg, 86159, Germany, [email protected];
Sven Tomforde, University of Kassel, Kassel, Germany, [email protected]; Jörg Hähner, University of Augsburg, Augsburg, Germany, joerg.
[email protected].
2009. Manuscript submitted to ACM
Manuscript submitted to ACM
1
2
S. Rudolph et al.
such as Autonomic Computing [1], Organic Computing [2], Proactive Computing [3], Multi-Agent Systems [4], or Collective
Adaptive Systems [5] a shift in responsibilities has been established: from design-time to runtime and from engineers to
systems themselves. As a result, various architectures and techniques have been established that allow for runtime
self-adaptation in technical systems [6].
These autonomous systems typically act in shared environments with the goal to achieve and maintain a given
utility function by means of adapting their configuration or parameterization in response to internal and external
changes. This is often combined with learning mechanisms to improve the adaptation behavior over time, which also
allows for handling unexpected or disturbed conditions [7, 8]. However, these existing techniques mostly neglect that
configuration and behavior of other systems in the environment may have strong impact on the degree of goal or utility
achievement. These impacts become even more challenging when considering system-of-system compositions [9] or
Interwoven System constellations [10] where dependencies between systems emerge that affect different abstraction
layers or application domains. An illustrating example for the latter case is the coupling of ICT with the smart grid [11].
Consider a surveillance network consisting of a potentially large number of smart cameras (SCs) [12, 13] as illustrating
example: Each SC acts autonomously and tries to optimize, e.g., the coverage of the observed area, the tracking success
of conspicuous objects, or the quality of 3D reconstruction of objects. In all three goal functions, there is an impact of
the current configuration of other SCs in the vicinity: The field of view should either have no overlap with the own
(coverage), should have overlap at a specific position and time (tracking), or full overlap as result of a certain ideal
configuration (3D reconstruction).
What we can see in this simple example is that the success of the self-adaptation strategy has not just a local origin
that is under control of the system itself -- it is highly influenced by the current configuration of other systems. We call
these effects "mutual influences" to highlight that these influences stem from other systems and not the environment or
possibly other sources. The shift of decisions from design-time to runtime also implies that such mutual influences
cannot be fully foreseen and planned anymore. Consequently, systems need a methodology to detect them in the first
place. This also requires mechanisms to incorporate this information in the self-adaptation strategy and to consider this
in the self-learning behavior.
In order to be applicable to SASO systems in general, such a mutual influence methodology has to fulfill several
requirements that are outlined in the following:
• Runtime capability: Due to the complexity that is present in modern systems, it is a more and more challenging
task to foresee all situations a system will face during its life time. It becomes virtually impossible because of the
appearance and vanishing of other systems that affect the outcome of certain processes. Therefore, an influence
detection algorithm has to be executable during the runtime of a system to allow for an adaption to unanticipated
interaction partner, such as new devices from other stakeholders, manufacturers, or owners.
• Heterogeneity: To reach the goal of self-integration [14], a system faces the challenge to interact with several
other systems that can be very heterogeneous in several aspects, e.g., scale, virtual/physical components, or
connectedness. A special focus of this article is on the heterogeneity induced by the ownership of a system which
limits the access and control of the other systems.
• Compatibility: Today, we face several long-living systems that are in operation for over a decade. Replacing
such systems would be costly and time intensive. Therefore, to integrate with such systems, it is necessary to
keep the influence detection compatible to a wide-range of systems. This especially includes systems that do not
have a cooperation mechanism, are non-adaptive, or use different adaption techniques.
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
3
• Autonomy: Recently, we can observe a dramatic increase in interconnected ICT system surrounding us. These
structures of systems bear new challenges since it is much more difficult to predict their behavior. To avoid time-
and cost-intensive maintenance of these systems, we require an influence detection mechanism that can become
essential part of autonomous systems, i.e., that can be integrated in a system and afterwards decide on its own
which influences have to be addressed.
In this article, we present a novel methodology that fulfills these requirements. We refine our preliminary work on
mutual influence detection [15 -- 17] and present a novel influence-aware learning mechanism. In particular, this article
goes far beyond preliminary work by providing the following contributions:
• A redefined method for the detection of influences among systems including multi component influences that
extends previous work presented in [17] by introducing a general workflow and more complex use cases (see
Section 4)
• A novel extended method of this approach that is runtime capable and consequently distinguishes correlations
in utility and performance with other's configuration from noisy or coincidental effects (see Section 5)
• A novel approach for incorporation of this mutual influence information within a reinforcement learning
algorithm that is applied to the problem of self-adaptation of the system configuration (see Section 6)
2 SYSTEM MODEL
As basis for this article, we describe the system model and the requirements a system has to fulfill to allow for runtime
influence detection. Furthermore, we introduce three example applications for the methodology that serve for illustration
purposes throughout this article.
2.1 Target Systems
We assume an overall system that is a composition of subsystems A1, . . . , An in a virtual or physical environment.
We refer to the term (sub)system using the terminology from the Organic Computing domain [2], and, for a better
readability, omit the sub if it is clear from the context that not the overall system is meant. However, one might prefer to
use the terms entity or agent instead. Each subsystem in the overall system can assume different configurations. Such a
configuration typically consists of different components. We define the entire configuration space of a subsystem Ai as
Cartesian product Ci = ci1 × · · · × cim, where cij are the components of the configuration.
Consider, e.g., a router A1 in a computer network as an illustrating example following the ideas of [18]. It can take
varying configurations into account, such as the processed network protocol or parameter settings. E.g., an interval
c11 = [0, 100] for the timeout parameter in seconds and the set c12 = {1, 2, . . . , 16} for the buffer size in kilobyte. The
entire configuration space of system A1 would then be C1 = [0, 100] × {1, 2, . . . , 16}.
A further assumption is that the particular configurations of individual systems are non-overlapping, meaning each
subsystem has its own set of configurations and cannot control those of other subsystems. This does not mean that
the configuration components have to be completely disjoint in structure and values of the contained variables. For
instance, two subsystems might have the same capabilities, which would lead to the same set of possible configurations
in these attributes but for different subsystems. Such a relation is explicitly allowed within the model.
Besides the configuration space, we need to consider the local reward. In order to apply the proposed method,
each subsystem has to estimate the success of its decisions at runtime -- as a response to actions taken before. This is
realized based on a feedback mechanism, with feedback possibly stemming from the environment of the subsystem (i.e.
Manuscript submitted to ACM
4
S. Rudolph et al.
Agent
State st
Reward rt +1
Action at
Environment
Fig. 1. The basic reinforcement learning model.
direct feedback) or from manual reward assignments (i.e. indirect feedback). This resembles the classic reinforcement
model [19], where the existence of such a reward is one of the basic assumptions. If there is no obvious way to
create such a signal it can be useful to apply more structured approaches from the field of goal-oriented requirements
engineering [20], such as the knowledge acquisition in automated specification (KAOS) goal model [21] that has been
applied to similar tasks [22]. Looking at the router example, a useful local reward can have different forms depending
on the application scenario. For instance, a useful measure of success could be the throughput (see [23]).
Relations to Reinforcement Learning:
We resemble the basic reinforcement learning [19] (RL) model and its extension to multi-agent reinforcement
learning [24] (MARL). The following paragraphs discuss show how this relates to the model used in this article.
Figure 1 illustrates this basic model for RL. It consists of an agent that interacts with its environment. It has the
ability to sense the current state of the environment and manipulate it by applying an action to it. Furthermore, a
reward is provided to the agent which reflects the usefulness of the applied action, which is possibly related to the
current or resulting state and can have stochastic components. The goal of the agent is to maximize the expectation of
this reward over the runtime of the system by choosing the best actions available in each state. To find such a strategy,
several approaches have been presented [25]. The model is usually formalized as a Markov Decision Process (MDP),
which is a tuple (S, A,T , r) defining the following components:
• a set of states S,
• a set of actions A,
• a transition function T : S × A × S → [0, 1],T(s′, a, s) (cid:55)→ p(s′s, a), i.e., a function that gives the probability that
action a in the state s results in the state s′.
• a reward function r : S × A × S → R, which gives a reward for each state transition.
Its basic form assumes discrete time steps t ∈ N. In each time step t, the agent senses the environment's state st ∈ S. It
then selects and applies an action at ∈ A. In the next step, t + 1, the agent receives a reward rt +1 which reflects the
quality of its action in the given state. It then senses the new state st +1 and starts over again. Looking at the transition
function T , we see that the probability of the appearance of a specific state can depend on the previous state and the
selected action. The agent's goal is to find a policy π(st) that maximizes the expected discounted reward
∞
t =0
γ t r(st , at , st +1),
(1)
where 0 < γ < 1 is a discount factor for the future rewards.
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
5
A widely used and intensively studied RL algorithm is Q-learning [26]. Like most RL techniques, Q-learning tries to
solve the general RL problem, i.e., to find an optimal policy for a given problem with respect to the long term reward.
The main idea is to find a Quality-function Q : S × A → R that approximates the reward for each state-action pair and
takes into account the long term reward. To reach this goal, the value for each state-action-pair is initialized according
to some of the various proposed methods, e.g., they are all set to a fixed value or they are set to a random value, and
updated afterwards according to the rule
(cid:16)
Qt(st +1, a) − Qt(st , at)(cid:17)
,
Qt +1(st , at) = Qt(st , at) + α
rt +1 + γ max
a
(2)
where Qt(s, a) denotes the old Q-Value and Qt +1(s, a) the new one, each for a given state-action pair (s, a). Furthermore,
rt +1 denotes the reward received in time step t + 1 and therefore is the immediate reward for the action at taken in
time step t. The discount factor γ ∈ [0, 1) determines the fraction of estimated future rewards that is taken into account
in the present step. The learning rate α ∈ (0, 1] determines how much the current experience, i.e. the current reward, is
taken into account for approximating the Q-value.
An extension to the single-agent RL is MARL [24], where several agents interact with a single environment. Each of
these agents is considered to observe and manipulate only specific parts of the environment, resulting in the agents
mutually influencing their rewards. The formal approach is a generalization of the MDP and is called a stochastic game
(SG). It is defined by a tuple (S, A1, . . . , An,T , r1, . . . , rn) where n is the number of agents. The components are defined
analogously to the single-agent case. Such SGs are classified depending on the reward structure. They are called fully
cooperative if r1 ≡ . . . ≡ rn and fully competitive if r1(s′, a1, . . . , an, s) + . . . + rn(s′, a1, . . . , an, s) = 0, where ai is the
action of the i-th agent. Games between these border cases are called mixed games.
We see several similarities between the initial system model and that from the RL and MARL domains. The initially
discussed configuration is similar to the action set since both are encoding the choice of the agent/subsystem. However,
while actions are often encoded in a relative manner, the configuration space has to be encoded absolute. E.g., a robot
could have an action turn left which would result in a different orientation depending on its previous orientation. For
the configuration, an absolute encoding is necessary, e.g., the orientation as a cardinal point. This can result in an
implicit integration of the state in the configuration, e.g., when the current orientation is part of the state. Furthermore,
the local reward is a concept adopted from the reward in the RL domain, but the difference is the focus on the locality.
E.g., in a case where the local reward of each agent is equal (as in a fully cooperative game), the influence detection can
still be applied but might be of limited use since it is only possible to measure which agent has influence on the overall
result. This result can still be interesting but, in this article, the focus is on studying systems where the notion of a local
reward is matched.
2.2 Example Applications
In the following, we introduce several example applications for illustration purposes. We start with two elementary
use cases that allow for a simple description and introduction in the appearing challenges. Afterwards, smart camera
networks are introduced, which define a real-world application often used to demonstrate SASO concepts.
2.2.1 Collaborative Box Manipulation. The collaborative box manipulation serves as first, rather simple scenario where
mutual influences occur. It includes two robots and a heavy box. The robots can be configured to push or pull the box,
but none of them is able to move the box alone since it is too heavy. Therefore, the robots have to cooperate, i.e., both
push or pull the box. This means that the configuration space is {PU SH , PU LL}. The robots receive a local reward that
Manuscript submitted to ACM
6
S. Rudolph et al.
Fig. 2. An example smart camera as used in our lab.
is 1 if the box moves forward or 0 otherwise. Obviously, the configuration of one robot has an influence on the success
of the other.
2.2.2 Two-Man Saw. The two-man saw scenario is a second example which is quite similar to the previous one. Again,
it includes two robots, but, in comparison to the box example, a two-man saw is operated instead of pushing a box. The
robots can be configured to push or pull the two-man saw (which results in the same configuration space as before), but
they cannot handle the saw alone. Therefore, the robots have to cooperate, i.e., one has to push and the other one has
to pull the saw (and vice versa). The robots receive a local reward that is 1 if the saw moves or 0 otherwise. Again, a
human can easily recognize that the robots influence each other.
Smart Camera Network. Smart cameras (SCs) possess a build-in computation unit that can be utilized for several
2.2.3
tasks such as image processing, object localization, and object tracking. Most SCs have pan, tilt, and zoom capabilities
and the computation unit is used to determine beneficial alignments for the camera. Beyond, SCs are equipped with
wired or wireless communication devices that allow for communication with neighboring cameras, and, therefore, form
a SC network. Such SCs can be utilized to achieve different goals, including the tracking of objects, the identification of
new objects or the 3D reconstruction of objects. In order to apply the methodology of detection of mutual influences to
SCs, a local reward is required. Obviously, this reward function heavily depends on the chosen goal. Regarding the
configuration„ we consider the three adaptable parameters of the camera's alignment (pan, tilt, zoom). In order to cover
different classes of possible mutual influences in SASO systems several goals can be considered:
• the maximization of the observed space,
• the coverage of fast changing environments [12],
• the detection of previously unknown objects [27] and
• the 3D-reconstruction of objects [17].
Concluding the former discussion and mapping this application to the model of the target system, it is a realistic
assumption that the cameras' configuration space is composed of their pan, tilt and zoom configuration. Allowing
a pan between 0 and 360 degree, a tilt between 0 and 90 degree and a zoom between 12 and 18, this would be
[0, 360) × [0, 90] × [12, 18] for each camera. As stated before, the configuration spaces of the cameras are identical
in structure, but each camera can assume an individual configuration. For the local reward we can chose one of the
introduced functions.
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
7
3 RELATED WORK
In the following, the related work in the area of influence formalization and quantification is presented. Further related
work can be found in Section 4.1 where several dependency measures are introduced.
In literature, several approaches can be found that try to formalize the mutual influences. Most of these approaches
focus on the influence through direct or indirect interactions. For instance, a model for interactions is proposed by Keil
et al. [28], but a method to detect the implicit interactions is not provided. Another common approach is to use stit logic
for modeling the interactions in multi-agent systems [29 -- 31]. The focus of these works is on the system specification
and verification and therefore differs fundamentally from the focus of our work where the goal is an adaption at runtime.
Multi-agent reinforcement learning (MARL) is an active field of research [24, 32]. An overview and a useful taxonomy
has been introduced by Busoni et al. [24], for instance. Following their taxonomy, the approach presented here can be
useful in fully competitive and mixed games (each static or dynamic). It can be used in fully cooperative games as well,
especially if the global payoff is a function of the local payoffs of the agents (cf. the smart camera application). As stated
by the authors, the complexity resulting from coordination is a major issue in MARL systems. The influence detection
mechanism presented here can keep this complexity to a minimum since it allows to only coordinate with relevant
partners.
In the following, we outline the related work from the MARL domain. Kok et al. [33] presented work that is based on
so-called coordination graphs and an approach to solve the global coordination problem on a local basis if it is possible
to decompose the global payoff function into a sum of local payoff functions. A restriction of this method is that it
relies on inference rules that are hand-crafted. Furthermore, they focus on discrete state variables which is not the case
in the approach presented in this article. But, since they assume a given graph, the influence detection could be used to
infer such a graph based on the roles of the agents.
An extension to this approach has been presented in [34]. There, similar to the influence-based approach, the
coordination graph is inferred at runtime creating a transition from independent learners to coordinated action selection.
It is based on a t-test between the maximally possible expected reward when the agents act in common and the expected
reward from independent decisions. The approach is for general sum games and is limited to discrete state and action
spaces, which means that it cannot be simply adopted to a continuous case. Furthermore, the work only shows the
applicability for a Q-learning-based approach and there is no trivial way to make it applicable to a wide range of
algorithms. Methodologically, a major difference here is that the method is focused on finding the states in which agents
should collaborate. This requires that all agents are always willing and able to cooperate.
DeHauwere et al. [35] demonstrated a specialized solution for mazes with two robots. They used a generalized
learning automaton that uses the distance to another robot to learn how to avoid a collision by identifying states in
which they have to coordinate. While this approach might be adapted for other tasks it would be necessary to hand-craft
the states of the learning algorithm each time.
Later, the authors presented an article focusing on a method to generalize the learned behavior for a single state
over several states. However, the presented approach needs to identify the states where coordination is necessary [36].
Their method is similar to Kok's. However, they assume that the agents have already learned an optimal policy if acting
alone since this is necessary to find states that need coordination which renders it unusable for learning at runtime.
Furthermore, the work is focused on sparse interaction, i.e., it identifies the states for cooperation and not the systems.
This means the approach is not appropriate to find useful collaborators from a set of agents. Furthermore, the approach
Manuscript submitted to ACM
8
S. Rudolph et al.
is limited to discrete state and action sets. Another difference to the approach presented here is that it relies on the
Kolmogorov -- Smirnov test used as a goodness-of-fit test, i.e., it tests if the distribution of points fits a model. However,
this approach depends on a model, which is not always present, e.g., if not Q-learning is used but a policy-based RL
approach, which is contrary to the requirement of independence from the control mechanism of the system.
The above approach has been extended to solve delayed coordination problems [37]. However, the limitation regarding
the discrete state and action sets persists. Furthermore, the approach is still not independent from the control algorithm.
Lanctot et al. [38] proposed a measure called joint policy correlation. This has been applied prototypically to a two
agent laser tag scenario. It is based on the repetition of the same scenario with several seeds which leads to different
strategies of the agents. Afterwards, a matrix is formed that compares the average rewards of the agents against the
agents from the other repetitions. The values against the initial opponent and the other opponents are then aggregated
to create a measure that allows to see how much an agent has overfitted to its initial opponent. The goal of the work
is very different from the one presented here since it wants to create a measure on how much an agent overfits to
the behavior of other agents. A measure of influence can not be directly derived from this approach. Furthermore, in
this article, we focus on a runtime learning approach which is not possible if the experiment has to be rerun several times.
The approach presented in this article has similarities with feature selection methods. These are categorized as filter,
wrapper or embedded methods. Wrapper methods are not applicable here since they require multiple repetitions of
the learning tasks which is contrary to the goal of learning at runtime. In the field of filter methods, the minimum-
redundancy-maximum-relevance [39] (MRMR) approach is frequently used. However, it relies on the elimination of
redundancies between the features which is not desired here since it reduces the information one gets from the influence
detection. Furthermore, it comes at an extensive computational cost compared to the methods proposed in this article.
Regarding the embedded methods, the most prominent instance is the regularized least-squares policy [40 -- 42]. This
and similar methods are tied to a specific learning algorithm and therefore breach the restriction of an applicability
independent from the control algorithm.
Furthermore, there are attempts to combine self-organizing systems with learning algorithms, e.g., Boes et al. [43]
introduced a MAS for the control of technical systems that has SASO characteristics. The method is called Escher and
aims to deconstruct complex control problems by using an acyclic network of interacting agents that are able to learn
from experiences. The algorithm has been applied to toy problems and an engine calibration. The approach differs
from the one presented here since they created a MAS to control a single technical system while here we focus on
several systems that are autonomous and have to adapt to the behaviour of the other systems. But, we believe that an
application of the influence detection to the individual controller agents of Escher can be helpful if the different inputs
of the controlled system have to be coordinated to reach optimal results.
Concluding the related work, there is no algorithm that allows to identify subsystems with continuous and discrete
states and actions which influence other subsystems to find useful collaborators and is independent from the control
algorithm.
4 MUTUAL INFLUENCES
In this section, we present our basic method for the detection of influences among autonomous SASO systems. This
method is a refined version of preliminary work that can be found in [15 -- 17] and necessary to understand the subsequent
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
9
Observation
Distribution
Estimation
Evaluation
Adaptation
Fig. 3. The general workflow of influence detection.
variants presented in the following sections. Especially, the refinement brings the method in a wider scope by defining
a general workflow and applies it to new scenarios. Besides extending the basic method in the following sections, we
simultaneously incorporate more complex tasks, i.e. the detection at runtime and the possibilities of an intelligent SASO
system to adapt its own behaviour to detected influences.
4.1 Methodology for Detection
Based on the described target systems, we define the methodology for mutual influence measurement. The goal is
to identify those components of the configuration of the other systems (i.e., value range and considered variables)
that have influence on the system itself. After the identification of influencing configuration components, they can be
addressed by a designer, e.g. by considering them in control algorithms, or by a self-adapting system itself, e.g. using a
learning algorithm.
In general, we are interested in the question whether a system as a whole is influencing other systems. However, to
be more precise in the description of the influence, we want to detect those parameters where the optimal configuration
values are somehow influenced by the current settings of the other systems. The basic idea of the following approach is
to make use of stochastic dependency measures that estimate associations and relations between the configuration
components of a system and the reward of a second system. The basic method assumes that the mutual influence
between the systems is instantaneous, meaning that the reward of the system reacts to the configurations of the other
systems in the same time step. However, the approach can be extended in order to detect delayed influences as well by
measuring the dependency between the configuration components and a later reward [44].
In general, dependency measures are designed to find correlations between two random variables. In the following,
we model the configurations of distributed systems as such random variables, which reflects the system model where we
face autonomous systems whose actions can be uncertain due to non-deterministic behavior or incomplete information,
for instance.
Before introducing relevant candidates for the dependency measures, we outline the general workflow (cf. Figure 3)
for an influence detection for each subsystem in the overall system.
Manuscript submitted to ACM
10
S. Rudolph et al.
(1) Observation: Continuously observe the configuration and estimate the goal achievement of a productive system,
e.g. the pan, tilt, and zoom of a camera and the corresponding reward in terms of the given goals. These
observations will typically be done by the system itself but they could as well come from an external entity.
(2) Distribution: Gather configurations from other systems and provide your own one to them.
(3) Estimation: Estimate the dependency value by relating the own reward to the configurations of the other systems.
The basic idea of the algorithm is to make use of stochastic dependency measures that estimate associations
between the reward of a system A and the configuration components of a second system B. These dependency
measures are designed to identify correlations between two random variables X and Y. The reward of A is
identified with a random variable X and the configuration of entity B with a random variable Y. This mapping
implies that if the association between X and Y is high, we also have a high influence of B on A, since it reflects
that the configurations of B matter for the reward of A. Vice versa, if the association is low, we do not see an
influence. There are several dependency measures available which might be suitable depending on the application.
An overview for the most interesting measures is given in the following. Furthermore, it might be necessary
to take other configuration parts into account. These are most likely the configuration parts of the influenced
agent, i.e., the configuration of system A is relevant for the measurement of the correlation between the reward
of system A the configuration of system B. In this case, we condition the calculation with the configuration of B.
This means that we calculate the dependency of the reward and the configuration for each configuration of B
separately in the discrete case. In the continuous case, we split the configuration in two (or possibly more) parts
and calculate the dependency separately. For a more detailed explanation of the estimation please see [15, 17].
(4) Evaluation: This step compares the values calculated for the different (other) systems and their configuration
components. There are several possibilities to do so: The first approach would be to compare the influence values
to a fixed threshold, which is a valid option but can be difficult since an appropriate threshold might not be
available for each application. Second, the influence values can be compared on a relative basis between the
systems. This leads to a ranking of systems according to their influence. A third possibility is to simulate an
independent system with the same configuration space and compare the "artificial" value to the real values of the
other systems. This allows to decide on a basis that only includes one other system. Additionally, it can be useful
to calculate the p-value, i.e. a value that allows to estimate how likely the given outcome under the assumption
of independence is.
(5) Adaptation: This step addresses the influences in the control strategy, which can take various forms: Especially
when applied at design-time, the designer can decide on a case-by-case basis. For a self-adaptive solution, we
propose to use a learning algorithm that includes the configuration of the influencing systems in the situation
description during runtime.
This process can be used either at design-time with a prototype or continuously at runtime. For the design-time
variant, the system runs for a certain time while the configurations and rewards are logged. Afterwards, the calculation
and decision process starts. During runtime, the steps have to be considered as a loop that runs continuously but one
can still decide on how long samples are gathered since a distribution and recalculation may not be justified for each
newly gathered sample.
Discussion of Dependency Measures. As mentioned before, our method relies on the utilization of basic dependency
measures for the influence detection. As in [15], we briefly introduce the most prominent measures and consider their
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
11
advantages and drawbacks in the following list. The contained measures are then analyzed in the following evaluation
to derive statements about their possible applicability for the ranking task in our method.
• Pearson correlation: The probably most prominent instance is the Pearson correlation coefficient -- sometimes just
called "correlation coefficient" [45]. The main advantages are its simple implementation and its fast calculation.
In the context, the major drawback is that only linear correlations can be detected, i.e., it can fail in case of more
complex dependencies. Moreover, it is necessary to calculate the distance between realisations of the random
variable which might make it not well suited for some problems. It assumes values between −1 and 1, where
−1 indicates a perfect negative linear correlation and 1 a perfect positive correlation. 0 means that there is no
linear correlation. When comparing the influence of different systems, it should be considered to use the absolute
values. The Pearson correlation coefficient r is defined as:
n
(cid:113)n
i =1(xi − x)2(cid:113)n
i =1(xi − x)(yi − y)
i =1(yi − y)2
r =
,
(3)
where the xi and yi are the gathered samples, i.e. the configuration components and the reward. n is the number
of samples and x, y denote the mean values of the random variables.
• Kendall rank correlation: Another measure that is based on calculating the ranks of the gathered samples has
been introduced by Kendall [46]. The measure can be computed rather fast. It can detect monotone dependencies,
which is better than just linear dependencies but still can be not sufficient for many applications. It assumes
values between −1 and 1; where −1 indicates a perfect monotone declining relationship and 1 a perfect monotone
increasing relationship. For independent variables, a value around 0 should be considered. The Kendall rank
correlation τ is calculated using:
(# concordantPairs) − (# discordantPairs)
.
τ =
n(n − 1)/2
(4)
A concordant pair are two samples (xi , yi), (xj , yj) where the ranks of the elements agree, i.e. if xi > xj then
yi > yj or if xi < xj then yi < yj. The opposite of a concordant pair is called discordant. Cases with xi = xj or
yi = yj are neither discordant nor concordant and are not handled in the basic variant.
• Spearman rank correlation: The rank correlation after Spearman is similar to Kendall's since it is based on ranks,
too. Here, the ranks of the samples are calculated and instead of a comparison between the samples the Pearsons
correlation coefficient is calculated on the ranks. Therefore, the values range from −1 to 1 with 0 meaning that
no dependency has been detected. The Spearman correlation is calculated using:
n
(cid:113)n
i =1(rд(xi) − rдx)2(cid:113)n
i =1(rд(xi) − rдx)(rд(yi) − rдy)
i =1(rд(yi) − rдy)2
ρ =
,
(5)
where xi and yi denote the samples and n the number of samples. rд(x) is short for the rank of x, i.e. the position
of x if all samples are ordered by their value. rдx is the average rank of the samples xi and rдx for yi, respectively.
• Distance covariance: This measure is an extension to the Pearson correlation that takes the distance between the
samples into account. Since these distances have to be calculated for the entire sample set, it is not suitable for
an online calculation. The advantage of this method is that it is not limited to linear dependencies but can find
all types of dependencies [47]. The distance covariance is calculated as follows: We first derive the Euclidean
distances aj,k = ∥xj − xk ∥ and bj,k = ∥yj − yk ∥. Afterwards, we calculate Aj,k := aj,k − aj . − a.k + a.. and
Manuscript submitted to ACM
12
S. Rudolph et al.
Bj,k := bj,k − bj . − b.k + b.. where aj . is the mean of the j-th row, a.k is the mean of the k-th column and a.. is
the mean of the whole matrix:
n
n
1
n2
Aj,k Bj,k
(6)
• Mutual information: A quite different approach has been taken by Shannon in the context of information theory
[48]. The basic variant can be used for discrete random variables and can be calculated online. Furthermore, all
types of dependencies can be found. A disadvantage is that the maximal possible value depends on the structure
of the random variables, i.e. the values can be normalized between 0 and 1, but the comparability to other
variables with different structures might be limited. The mutual information is defined as:
k =1
j=1
I(X; Y) :=
p(x, y) ld
(cid:18) p(x, y)
(cid:19)
y∈Y
x ∈X
p(x)p(y)
(7)
where p(x, y) is the joint probability of the events x and y and p(x) and p(y) are the marginal probabilities of
x and y. The probabilities can be easily approximated using a frequency counting (i.e., a maximum likelihood
approach based on discrete finite random variables). There exists also a continuous variant of the mutual
information measure. Here, the calculation is more complex since the estimation from samples needs more
advanced techniques. The most common approach is based on a k-nearest neighbor method [49]. Another
approach is the manual binning of the values in order to use the discrete version, which can bear problems
because the results might depend highly on the chosen binning parameters. An automatic binning can be done
with the maximal information coefficient (see below).
• Maximal information coefficient: This measure is an extension to the mutual information for real-valued data.
It automatically calculates the binning that results in the maximal mutual information for a given data set. A
drawback is that there is no method for the online calculation. The basic formula is:
MIC(X , Y) := max
nx ny <B
I(X; Y)
log(min(nx , ny))
(8)
where nx and ny denote the number of bins for X and Y. This means that the number of bins is limited by
a threshold B that is by default determined depending on the sample size. The denominator serves for the
normalization of the value. However, it is computationally expensive to determine the values for all possible
binnings. Therefore, a heuristic is introduced. For details please see [50].
It should be mentioned that the choice of a dependency measure does not have to be exclusive. It can be useful
to calculate several measures in parallel or to perform an iterative process, which initially uses computationally
light-weighted measures and adds more powerful and computationally expensive methods only if necessary.
4.2 Evaluation
Initially, we aim at investigating which measurement from the previous subsection is most suitable for the detection of
mutual influences. Therefore, we consider the three initially introduced application scenarios that are characterised by
increasing complexity.
4.2.1 Collaborative Box Manipulation. In this scenario, two robots have the task of pushing a box. Each of them can
either PU SH or PU LL the box, but the box only moves if both of them PU SH. This leads to a reward of 1 while all
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
13
(a) An overview of the full 100 steps evaluated.
(b) A more detailed view on the first 30 steps.
Fig. 4. The results for the collaborative box manipulation. The graphs show the fraction of 100 runs in which the influence of robot B
is detected higher as the influence of the notional robot.
other combinations lead to a reward of 0. Since some of the dependency measures do not allow for categorical random
variables, such as the here used PU SH and PU LL, the configuration has been mapped to the numbers 0 and 1.
For the evaluation of this application, we conducted 100 experiments and measured the influence using each of the
previously introduced dependency measures. For each run and each measure, two values have been calculated. The
first one is the influence of robot B on robot A. The second one is the influence of a notional robot. This notional robot
has the same capabilities as the real robot B, but his actions do not influence robot A. The value is calculated to assess
how reliable the detection of the influencing robots is. This can be done by comparing in which fraction of the runs
robot B has been found more influential than the notional robot.
Manuscript submitted to ACM
0 20 40 60 80 100 10 20 30 40 50 60 70 80 90 100Fraction of higher real estimateStepsPercentage of Correct RankingPearson CorrelationKendall CorrelationSpearman CorrelationDistance CorrelationMutual Information (discrete)Mutual Information (continuous)Maximal Information Coefficient 0 20 40 60 80 100 0 5 10 15 20 25 30Fraction of higher real estimateStepsPercentage of Correct RankingPearson CorrelationKendall CorrelationSpearman CorrelationDistance CorrelationMutual Information (discrete)Mutual Information (continuous)Maximal Information Coefficient 14
S. Rudolph et al.
The results are depicted in Figure 4. There, we see that most of the measures perform similar in this scenario. Except
for the continuous MI approximated with the Kraskov method 40 steps are sufficient to distinguish the influencing
robot from a non-influencing in each of the 100 runs. The vast majority is already correctly detected after 15 steps. The
continuous MI eventually finds the influence in each run. However, the detection speed is rather slow. This is due to
the fact that the continuous MI has problems with the discrete values that are assumed in the problem. Furthermore,
the continuous MI has not been calculated for less than ten steps since a minimum number of samples is required.
Therefore, the first three data points should be disregarded.
Concluding the results, the influence in this elementary use case can be detected quite easily with the previously
introduced method. Even though the continuous mutual information shows a little slower detection, the selected
measure for the task is not too important since the detection is over all quite fast.
4.2.2 Two-man Saw. As a second example, we evaluate the conditioned measurement in an elementary use case, the
two-man saw. The example is inspired by two robots that operate a saw that only moves if both of them move it in the
right direction, i.e., one PU SHes and the other PU LLs or vice versa depending on the current position of the saw. If the
saw moves each robot gets a reward of 1 otherwise 0. Again, the categorical configurations PU SH and PU LL have been
mapped on numerical values to make all dependency measures applicable.
The results are depicted in Figure 5. As for the collaborative box manipulation, the actual influence has been calculated
using the seven dependency measures and compared with a measurement of a notional robot that assumes uniformly
distributed random configurations but has no influence on the actual outcome of the experiment. The figure shows in
how many of the 100 independent runs the influencing robot has been found to be more influential than the notional
(not-influencing) robot. The first graph shows the result for the method used in the previous evaluations, i.e., without a
consideration of other configurations. The detection is between 30% and 50% which is below the expected value of 50%.
This due to the fact that the run will only be counted as correctly detected if the value is higher but not if both values
are equal which is quite often the case in this scenario. The second graph shows the results with the consideration of
the other configuration, i.e., the calculation of the dependencies is conditioned under the configuration of the first robot.
We see that influences are almost perfectly detected after 30 steps which is close to the result in the collaborative box
manipulation. Also similar is the small weakness of the continuous MI measure which needs about 90 steps to catch up.
Concluding the evaluation of the two-man saw application, we have seen that even in rather simple applications the
problem with other configurations can lead to influences that can not be detected with the basic variant of the influence
detection. However, the improvement that includes the other configurations by conditioning the measurement shows
very good results for such cases as well.
Smart Camera Network. As a third application for the evaluation, we consider an example for the SC network
4.2.3
domain. To recap briefly, SCs are surveillance cameras that are equipped with computational capabilities that can be
used for several tasks including image processing. Furthermore, these cameras are interconnected via a network that
allows them to exchange data and to coordinate. For this article, we stick to so-called PTZ cameras that allow for an
automatic adjustment of the pan angle, tilt angle, and zoom of the camera. As previously described, there are several
reasonable goals for a SC network. For this evaluation, we stick to the goal of a 3D-reconstruction of the observed
targets. In particular, this means to observe the targets from different perspectives and cameras at the same time. To
achieve this, the cameras get a reward of 1 for each object that is observed by at least two cameras in a time step.
A top-down view on scenario SCN 1 is depicted in Figure 6. There, we see three black dots which represent one
camera each. Around the dots, there are colored circles. Each of them represents the area that is potentially observable
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
15
(a) The detection if a single estimator is used and the own configuration is not considered.
(b) The detection if the own configuration is considered, i.e., there is one estimator for each of
the configurations (Push and Pull).
Fig. 5. The results for the two-man saw use case.
Manuscript submitted to ACM
0 20 40 60 80 100 10 20 30 40 50 60 70 80 90 100Fraction of higher real estimateStepsPercentage of Correct RankingPearson CorrelationKendall CorrelationSpearman CorrelationDistance CorrelationMutual Information (discrete)Mutual Information (continuous)Maximal Information Coefficient 0 20 40 60 80 100 10 20 30 40 50 60 70 80 90 100Fraction of higher real estimateStepsPercentage of Correct RankingPearson CorrelationKendall CorrelationSpearman CorrelationDistance CorrelationMutual Information (discrete)Mutual Information (continuous)Maximal Information Coefficient 16
S. Rudolph et al.
Fig. 6. SCN 1. A top-down view on a smart camera network. The black dots depict cameras surrounded by a circle that marks their
potential observable area. The red shapes show the field of view for an exemplary PTZ configuration. The yellow arrow indicates from
where and in which direction the objects of interest move.
Manuscript submitted to ACM
Camera 3Camera 2Camera 1 On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
17
Fig. 7. SCN 2. A top-down view on a smart camera network. This scenario is only slightly changed from SCN 1 by adjusting the
flow of targets represented by the yellow arrows. The black dots depict cameras surrounded by a circle that marks their potential
observable area. The red shapes show the field of view for an exemplary PTZ configuration.
by one of the cameras if it chooses an according PTZ configuration. Furthermore, we see example areas that are currently
observed marked my red lines. Camera 1 and Camera 2 share a common area which can be observed by both at the
same time. In contrast, Camera 3 is isolated and does not share a common area with one of the others. The yellow arrow
marks the entry point and direction for objects that move through the scene in the area between Camera 1 and Camera
2.
Ten independent runs of this scenario have been conducted in a Mason1 simulation, where each of the cameras
assumes a uniformly sampled PTZ configuration in each time step. The pan angels are between 0 and 360 degree,
the tilt angle between 120 and 180 degree, and the zoom between 12 and 18. In each of the runs, we compared the
influence of Camera 2 on Camera 1 and the influence of Camera 3 on Camera 1. For each of the measures, there are
three comparisons: one for the pan, one for the tilt, and one for the zoom of the cameras. It is expected that there is a
clear trend towards a higher influence of Camera 2 over Camera 3 in general and especially for the pan and tilt since
these configuration determine if it is possible to gather a reward for Camera 1 or not.
The results are depicted in Figure 8. There, we see that after a few 100 steps most measures allow a correct detection
in 100% of the cases. The continuous mutual information and maximal information coefficient on the other hand take
a little longer to reach this level of certainty. For the pan, we see that the distance correlation shows the best result
compared to the other measures. However, the other measures work as well with a higher sample size. Even though the
MIC finds an influence quite reliably, the other measures do not show a definite result which can be explained by the
minor influence of this configuration components, i.e., in most cases, the zoom does not determine if there is a positive
reward at all but only the height of this positive reward.
Concluding the results, we have seen that it is possible to detect the influences in simple examples based on real-world
applications, such as SC networks. We will see how to use this in scenarios with more complex requirements in the
remainder of this article.
1Mason is a multi-agent simulation framework for Java [51].
Manuscript submitted to ACM
Camera 3Camera 2Camera 1 18
S. Rudolph et al.
(a) The results for the pan.
(b) The results for the tilt.
Fig. 8. The results for scenario SCN 1 using the general method. Each graph shows the fraction of runs in which the influence of the
configuration component (pan, tilt, or zoom) of Camera 1 on Camera 0 is detected to be higher than the influence of Camera 2.
(c) The results for the zoom.
After the case that has been solved without the consideration of other configuration components, we focus on a
scenario with few changes from scenario SCN 1. A top-down view on scenario SCN 2 is depicted in Figure 7. In the
figure, we see three black dots that mark the position of the cameras. Camera 1 and Camera 2 are quite close and
Manuscript submitted to ACM
0 20 40 60 80 100 0 200 400 600 800 1000 1200 1400 1600 1800 2000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 0 200 400 600 800 1000 1200 1400 1600 1800 2000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 0 200 400 600 800 1000 1200 1400 1600 1800 2000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
19
potentially share a common field of view which is marked by the colors areas around the cameras. Camera 3 on the other
hand is separated and cannot observe the same areas as the other two cameras. The only difference to the previously
evaluated scenario is where the objects of interest appear and move. In the last scenario, they were limited to the area
between Camera 2 and 3 appearing from the south. Here, we see that they appear from the north and the south on
the entire area around Camera 1 and 2, i.e., the two cameras are surrounded by objects. The influences from Camera 2
and Camera 3 on Camera 1 have been measured for the pan, the tilt, and the zoom. We expect that the pan and the tilt
of Camera 2 have the most influence. Since a reward greater than 0 is much more likely in this scenario than in the
previous, we also expect the zoom to play a more important role.
The results are depicted in Figure 9 and 10. As before, 10 independent runs have been conducted and we see the ratio
of runs in which the configuration components of Camera 2 have been identified as more influencing than those of
Camera 3. In Figure 9, we see the results for the previous method. Even though the camera placement is identical to
the previous scenario, we see that the pan is not detected within the first 2000 steps. Furthermore, the tilt cannot be
detected by the Pearson correlation. This is because it is necessary to consider the other configuration components in
this scenario. Therefore, we adopted the conditioned calculation, i.e., we split each of the other configurations in two
parts and sort the points in different buckets depending on which configuration has been assumed. Since we face 5
other configurations, this leads to 25 = 32 buckets. The values calculated for the buckets are then added up to get an
aggregated result. The results achieved by this method are depicted in Figure 10. We see that using this method the
issues experienced in the first experiment do not appear and a flawless detection is ensured.
Concluding the results, we have seen that small changes in the setting can make it necessary to consider the other
configuration components in the influence detection. A fast and reliable detection can be achieved by conditioning of
the dependency measures.
5 INFLUENCE DETECTION AT RUNTIME
Modern ICT systems often cannot be tested entirely at design-time since not all circumstances the system faces can be
foreseen due to a lack of awareness or time constraints. This applies especially to the type of systems that this article
focuses on: self-adaptive and self-organising systems acting in shared environments. Obviously, it is hardly possible
to foresee all situations a system will face if the considered system has to act in parallel to other systems (that may
be under control of another stakeholder) and both should be able to adapt to new goals or environmental changes.
Therefore, this section shifts the focus towards the detection of influences at runtime. A special examination of this is
satisfied because of the problem of static behavior of the systems that introduces disturbances in the measurement by
causing correlations that are not based on causality but coincidence.
5.1 Methodology for Detection
Regarding the detection at runtime, we have to consider that the proposed influence detection algorithm relies on
dependency measures that estimate the correlation between two random variables. If the method is applied at runtime,
this can lead to wrongly detected influences because such correlations can appear without an underlying causality. The
previous experiments demonstrated that it is possible to infer a causality from the correlation between the actions of
two subsystems, since we enforced the configurations to be randomly selected from independent uniform distributions.
However, for the detection at runtime it has to be factored in that we can face autocorrelations and other disturbances
within the configurations of the systems.
Manuscript submitted to ACM
20
S. Rudolph et al.
(a) The results for the pan.
(b) The results for the tilt.
Fig. 9. The results for the scenario SC2 using no conditioning. Each graph shows the fraction of runs in which the influence of the
configuration component (pan, tilt, or zoom) of Camera 1 on Camera 0 is detected to be higher than the influence of Camera 2.
(c) The results for the zoom.
The approach to avoid this is to rely mostly on randomized configuration that appear naturally in applications that
use reinforcement learning during runtime. This is due to the need of exploration in these tasks to find optimal strategies
and to avoid to "get stuck" in a local optimum. This happen often with greedy approaches because the algorithm only
Manuscript submitted to ACM
0 20 40 60 80 100 0 200 400 600 800 1000 1200 1400 1600 1800 2000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 0 200 400 600 800 1000 1200 1400 1600 1800 2000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 0 200 400 600 800 1000 1200 1400 1600 1800 2000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
21
(a) The results for the pan.
(b) The results for the tilt.
Fig. 10. The results for the scenario SC2 using two parts for conditioning. Each graph shows the fraction of runs in which the influence
of the configuration component (pan, tilt, or zoom) of Camera 1 on Camera 0 is detected to be higher than the influence of Camera 2.
(c) The results for the zoom.
tries one behavior and if it is "good" it sticks to it and misses out on other strategies that may lead to better results.
There are different strategies to avoid this behavior. An easy and reliable approach is to use an ϵ-greedy action selection,
i.e., the algorithm will stick to the action that it has evaluated as the best one so far most of the time. However, with a
Manuscript submitted to ACM
0 20 40 60 80 100 0 200 400 600 800 1000 1200 1400 1600 1800Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 0 200 400 600 800 1000 1200 1400 1600 1800Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 0 200 400 600 800 1000 1200 1400 1600 1800Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 22
S. Rudolph et al.
probability of ϵ it will use a random action regardless of the so far evaluated usefulness of it. This means a natural
approach to the issue of correlation without causality is to only use the samples that are formed from such exploration
steps. But this would lead to a significant lower amount of samples. To avoid this we will also examine in detail how
much "randomness" is necessary to allow the measurement to function properly by conducting experiments with
different levels of ϵ. Identifying the lowest level of ϵ will ensure the fastest detection of influences without running the
risk of falsifying the measurement.
One might expect that a very high level is crucial for a successful detection. However, the correlation mainly appears
due to the repetition of specific patterns in both systems in the same frequency. If such patterns are brought out of sync,
a correct detection is possible despite the repetition, i.e., a high autocorrelation does not necessarily mean that the
measurement is flawed. In example, two systems A and B each switch back and forth between their two configurations 1
and 2. The system A gets a reward of 0.5 if it applies 1 and 0.8 if it applies 2. If both systems switch their configurations
in the same time step the measurement will be falsified because system B's configuration will correlate with the reward
of system A. If one of the systems randomly does not switch the configuration the behavior becomes "asynchronous"
and the measurement works well.
5.2 Evaluation
In the following, the evaluation of the runtime detection and the method to adapt to the influences are evaluated. For the
remainder, the evaluation parts of the article focus on the smart camera scenarios only. At this point, the consideration
of the elementary use cases would not add to the discussion of the approach and are therefore omitted.
We start on SCN 2 that has been introduced in Figure 7. Briefly recapped, we have three cameras, where Camera 1
and Camera 2 influence each other due to an overlap in their potential observable space. Camera 3 does not overlap
with one of the other cameras, and, therefore, it does not influence the other two. Previously, we saw how the detection
works when the configuration is chosen randomly from the full configuration space.
In the following, the experiments are based on control mechanisms that use Q-learning, using discretized states and
relative actions. Q-learning has been chosen since it is widely used and well understood. For the concrete implementation,
we limit the alignment of the camera to 12 pan angles, i.e., it is from sp = {0, 30, 60, . . . , 300, 330}, to 3 tilt angles, i.e., it
is from st = {120, 150, 180}, and to 2 zoom levels, i.e., it is from cz = {12, 18}. This results in 12· 3· 2 = 72 states for each
camera. The camera can increase, decrease or leave the pan, tilt and zoom in each time step resulting in 3 · 3 · 3 = 27
actions that can be applied. This means that the configuration space is C = sp × st × sz for each camera.
The first evaluation step is to find out how the limitation of the configuration space and the order of states that are
visited affect the influence detection. The results can be found in Figure 11 and 12. The system has run for 10, 000 steps
by applying random actions and the figure shows the percentage of runs in which Camera 1 detects Camera 2 as more
influencing than Camera 3. The results for the pan and zoom are depicted in Figure 11. There, we see that each measure
can find the influences very reliably within the 10, 000 steps with the distance correlation being much more unreliable
using less samples. In Figure 12, the results for the tilt are depicted. We can see that only few steps are needed to find
the influence, i.e., in only 600 steps each of the measures can detect very reliable the influence.
In the next step, we analyze how the detection rate changes if we switch from a purely random action selection
to ϵ-greedy. The ϵ is varied between 1 and 0.05 according to Table 1. It reflects that ϵ is set to 1 at start and is then
decreased by 10% each 10, 000 steps. The results can be found in Figure 13. There, we see the detection rate if single-
agent Q-learning with an ϵ-greedy action selection is used and the last 10, 000 samples are used for the detection. The
Q-learning algorithm uses a low α = 0.1 to handle fluctuations in the reward signal and a high γ = 0.9 to allow the
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
23
Step
ϵ
Step
ϵ
Step
ϵ
Step
ϵ
Start
1.0
90k
0.43
170k
0.19
250k
0.08
10k
0.9
100k
0.39
180k
0.17
260k
0.07
20k
0.81
110k
0.35
190k
0.15
270k
0.06
30k
0.73
120k
0.31
200k
0.14
280k
0.06
40k
0.66
130k
0.28
210k
0.12
290k
0.05
50k
0.59
140k
0.25
220k
0.11
300k
0.05
70k
0.53
150k
0.23
230k
0.1
310k
0.05
80k
0.48
160k
0.20
240k
0.09
320k
0.05
Table 1. The values for ϵ in the ϵ-greedy action selection. It starts at 1.0 and decreases by 10% every 10k steps until it reaches 5%.
Manuscript submitted to ACM
24
S. Rudolph et al.
ALGORITHM 1: The main loop for each subsystem using the influence detection and adaption at runtime.
initialization;
for each step do
observe own configuration and reward;
distribute observations to other subsystems;
if enough observations then
estimate influences;
evaluate estimated influences;
if influence found then
adapt learning algorithm;
end
end
learning algorithm decides next action;
update learning algorithm with new observation;
end
learning of a sequence of actions. Figure 13a illustrates that for the pan the detection works well for the first 50, 000
steps, i.e., for ϵ > 0.66. Afterwards, the detection works worse for some of the measures with an average of about 80%.
The continuous mutual information sticks out here with a perfect detection even with low values of ϵ. For the tilt, in
Figure 13b, we see a similar effect, but it is way less distinct and occurs only for lower ϵ. Figure 13c shows the graph for
the zoom, which turns out to be way more sensitive regarding ϵ and makes the measurement unreliable.
Concluding the results, we have seen how the application of a control algorithm affects the influence detection
in contrast to a random selection of configurations. Due to the disturbances and autocorrelations, it is possible that
influences are not detected correctly. However, if it is ensured that an adequate amount of "randomness" is used to select
the actions it results in proper results. For the example examined here, we have seen that the usage of 66% randomly
selected actions allows for a precise identification of influencing systems. If the usage of less randomly selected actions
is necessary, an amount of greedy actions should be removed from the set for the calculation of influences until the
required amount of "randomness" can be ensured.
6 SELF-ADAPTING TO INFLUENCES
In the previous sections, we focused on the detection of the influences, which is a crucial step towards the goal of
optimal behavior. In this section, we introduce an approach to self-adaption to these identified influences, i.e., the
exploitation of the influences if they have been discovered. For this adaption, the focus is not on the presentation of a
single algorithm, but on giving a general methodology based on the RL model and realize it with Q-learning.
6.1 Methodology for Self-adaption
In general, there is a variety of options available to address the influences since the detection is designed to work
independently from the control algorithm of the system. Possible reactions to influences range from a hand-crafted
solution during design-time to a system that adapts during runtime in static patterns and finally to a full self-learning
behavior. For the latter, we get an impression on how the subsystem's learning algorithm and influence detection
workflow (cf. Figure 3) interact by studying Listing 1. Potential candidates for such a learning algorithm can be found
in the area of MARL. However, we rely on a basic principle that unifies three properties:
(1) It is possible to start out with single learners and later on learn the cooperation,
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
25
(a) The results for the pan.
(b) The results for the zoom.
Fig. 11. The results for the detection of influences in SCN 2. The influence detection rate for the pan and zoom of Camera 1 versus
the Camera 2 based on 10 independent runs. The measurement is similar to the graphs before but with the discretization of the state
space and a Q-learning that applies the actions randomly. The graphs show the number of runs, in which the influence of Camera 1
is measured higher than the influence of Camera 2, for different numbers of samples.
Manuscript submitted to ACM
0 20 40 60 80 100 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 26
S. Rudolph et al.
(a) The results for 300-9900 samples.
(b) A detailed few for 50-500 samples.
Fig. 12. The results for the detection of influences in SCN 2. The influence detection rate for the tilt of Camera 1 versus the Camera 2
based on 10 independent runs. The measurement is similar to the graphs before but with the discretization of the state space and a
Q-learning that applies the actions randomly. The graphs show the number of runs, in which the influence of Camera 1 is measured
higher than the influence of Camera 2, for different numbers of samples.
Manuscript submitted to ACM
0 20 40 60 80 100 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 50 100 150 200 250 300 350 400 450 500Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
27
(a) The results for the pan.
(b) The results for the tilt.
(c) The results for the zoom.
Fig. 13. The results for the detection of influences in SCN 2. The influence detection rate for the pan, tilt and zoom of Camera 1
versus the Camera 2 based on 10 independent runs. The measurement is similar to the graphs before but here the camera is controlled
by a Q-learning algorithm with ϵ-greedy strategy, where the ϵ is falling from 1 to 0.05 in 10%-declines each 10k steps. The graphs
show the number of runs, in which the influence of Camera 1 is measured higher than the influence of Camera 2, using the last 10k
steps as samples.
Manuscript submitted to ACM
0 20 40 60 80 100 0 50000 100000 150000 200000 250000 300000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 0 50000 100000 150000 200000 250000 300000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 0 50000 100000 150000 200000 250000 300000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 28
S. Rudolph et al.
Conf. ct
State s′
t
System B
System A
Reward rt +1
Action at
Environment
State st
Fig. 14. The proposed method to adapt to other influencing agents in terms of the reinforcement learning model.
(2) it can be applied to many reinforcement learning algorithms,
(3) it can be used with systems that have heterogeneous configuration spaces and control algorithms, e.g., some of
the systems have a static behavior.
The basic adaption principle is depicted in Figure 14 and shows how the adapt learning algorithm from Listing 1 has
been realized. There, we adopt the basic RL model introduced in Section 2.1. We assume that previous measurements
have resulted in a detection of an influence from system A on system B meaning B should react properly regarding the
configuration of A. Therefore, we integrate the value of the configuration ct at time t into the state st resulting in a
state s′
t . For example, if B's states are given through a position represented as (x, y)-coordinate between 0 and 1 each,
its state space is S = [0, 1] × [0, 1]. Assuming the influence detection has found A's configuration component cB, which
represents the speed of B and is between −1 and 1, as influencing, the state space S will be extended to a new state
space S′ = S × [−1, 1] = [0, 1] × [0, 1] × [−1, 1].
There are two points to discuss in this approach. The first one is how the influenced system B will know the current
configuration of A. This is not an issue with delayed influences but only with immediate. It can be resolved by letting
the influenced system wait until it can observe the decision either by a message of the influencing system or through
sensors. If several influences in the system are detected this can lead to chains of systems that wait on each other and it
is necessary to check for each of the integrations in the state space that the graph they form is acyclic.
As mentioned before, this method is independent from the learning algorithm as long as it fits the RL model. But the
extension of the state space can be seen as a special instance of transfer learning in the RL domain [52]. This makes
algorithms that have a strong transfer capability more suitable for this task. In this article, we rely on Q-learning since
it gives an easy and natural way to realize the transfer during runtime that will be explained in the following: Assuming
that cB can take k different values from each state s, there will be k new states s′
i , i ∈ {1, . . . , k}. Before the expansion
of the state space, the system holds a current Q-value for Q(s, aj), where aj ∈ A are the possible actions. To transfer the
already learned knowledge, the new Q-values are set to Q(s′
i , aj) := Q(s, aj), for each i ∈ {1, . . . , k} and aj ∈ A. The
advantage here is that in states that are not affected by the other system the Q-value will remain at the correct value.
In turn, in states where the system is influenced the values will be updated to their true value using the upcoming
experiences.
If the state space is continuous the standard Q-learning can not be used. An alternative would be for example the
extended classifier system (XCS-R) with continuous states [53, 54] that has good online learning capabilities and is
transferable [55].
A further issue is that in the previous experiments about runtime detection, we have seen that a small amount of runs
might detect the independent systems as more influencing than they actual are. This behavior is temporary and can be
corrected by using a bigger sample size. However, during the runtime of the system, it is not possible to determine if
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
29
it is a malicious detection. Therefore, we introduce a factor by that the real measurement has to be higher than the
calculated notional independent counterpart which is used as baseline. Furthermore, the different configuration parts
are ranked and after each influence calculation only the highest ranked configuration is integrated in the state space
of each camera. Even though this mechanism is likely to catch most erroneous integration, they cannot be avoided
entirely. But, if a wrong integration is detected later, it can be corrected easily. To revert the false state space extension
i , aj)
the new created states are merged back together by averaging the Q-values of the states, i.e., Q(s, aj) =k
i =1 Q(s′
for every action aj ∈ A.
6.2 Evaluation
In the next part, we examine how the influence measurement reacts to the adaption that has been introduced previously.
To do so, we integrated the three configuration components of Camera 2 in the state of Camera 1 at the start of the
simulations. In Figure 15, we see similar results as for the independent learning with especially bad results for the
maximal information coefficient and the distance correlation.
As a last comparison on this scenario, we have a look at the rewards received with the different methods: the
independent learners in Figure 16a, the integration of the configuration of Camera 2 in the state of Camera 1 in
Figure 16b, and the dynamic integration of the configurations at runtime in Figure 16c. We can see that the independent
learners are not able to reach an appropriate result since they lack a coordination mechanism. The integration at start
can reach an optimal result of about 25 after 250k steps. The dynamic expand can reach a similar result but learns
slower. This is because the transfer mechanism cannot play out its advantages here since the Q-values differ for all
states in this example.
In Figure 17, we see scenario SCN 3, another scenario from the SC domain. The figure shows a top-down view on a
SC network with six cameras depicted as black dots. Each of the cameras has a corresponding partial circle around
it that visualizes the space potentially observable space of the cameras. The circle is only partial since the vision is
blocked by pillars that are shown as gray squares. There are several streams of targets entering and crossing the scene
from all sites marked by yellow arrows. We see that not each pair of cameras shares a common area both can observe.
Therefore, only the pairs (1, 2), (2, 4), (3, 4), (3, 5), (5, 6) influence each other and have to consider the configuration of
the other camera.
Again, we apply Q-learning to this scenario but with a small adaption: We do not use relative actions, e.g., increase or
decrease, but absolute ones, i.e., each of the defined states can be accessed from every other state leading to 72 actions.
Since this does not require to learn sequences of action to reach one alignment from the previous the discount factor is
set to γ = 0.
The results are shown in Figure 18. The single-agent learning case in Figure 18a is clearly worse than the other
variants only reaching a reward of 2 on average since the cameras do not coordinate well. We see that the algorithm
gets stuck in local optima when ϵ drops low at around 250k steps since single configurations are valued the highest due
to the lack of knowledge of the influencing systems. The other two variants show the dynamic integration in the state
space. In Figure 18b, we see how the system performs when 3, 000 samples are used for the detection; in Figure 18c
10, 000 samples are used. The graphs are nearly identical and both reach an optimal behavior that gives an average
reward of above 3.
Finally, in Figure 19, we see the cooperation network that evolved after 200, 000 steps in one example run. Each
arrow means that the configuration parts of the camera that is pointed at are integrated in the state space of the first
camera. We see that the structure that formed reflects the structure of influences described before.
Manuscript submitted to ACM
30
S. Rudolph et al.
Concluding the results, we have seen how single learners compete with a-priori adapted systems and a stagewise
integration of influencing configurations in the state space of an RL algorithm. We have seen that the single learner can
not compete on the same level as influence respecting variants. The influence detecting approaches have shown similar
results, which makes them useful options for the design of systems.
7 CONCLUSION
Self-adaptive and self-organising systems typically act in a shared environment with other systems of the same kind.
This results in the need to consider the current strategies followed by other systems within the own decision and
learning process. We call such effects between actions and their effect on the achieved utility "mutual influences".
In this article, we presented a novel methodology to identify and consider these mutual influences by extending
and refining prior work and investigating novel solutions to consider them in runtime self-adaptation. The detection
is realized with dependency measures that have different characteristics and are consequently compared regarding
their applicability in various settings. The adaption to the influences is realized using the example of Q-learning. The
techniques are evaluated on two elementary use cases, the collaborative box manipulation and the two-man saw, and a
smart camera application with multiple scenarios.
Finally, in the remaining paragraphs, we discuss the scalability of the approach and give an outlook on future work.
7.1 Considerations about the Scalability of the Approach
While the scalability is not the main focus of this work, we want to give an impression of this by considering and
discussing two potential bottlenecks of the approach and afterwards give practical advice on how to handle it in
situations where they might become an issue.
The first one is the computational overhead: Besides the calculations for the learning algorithm this leads to n(n − 1)k
calculations of the dependency measures where n is the number of subsystems and k is the number of configuration
components on each agent meaning the overhead will grow quadratic with the number of agents. However, this number
will reduce vastly after the first components have been found influencing and are integrated in the learning algorithm
since many possible paths are not considered any longer. For an experiment in the SC3 scenario with n = 6 and
k = 3 this leads to a maximum of 90 calculations for each estimation of the influence. The experiments have been run
sequentially on an Intel [email protected] and the approximate runtime for 150k steps (which is approximately the
time until an optimal solution has been found) is about 1.9h including the time for the simulation itself and the learning
algorithm. Given the remark before, it is clear that the average runtime of the influence estimation is much higher
initially than in the end.
The second potential bottleneck is the distribution of the realizations of the configuration components. Depending on
the system architecture, it is possible that each subsystem has to send its status to every other subsystem. Denoting the
number of agents as n and the number of components as k this leads to (n − 1)k values send by each agent or n(n − 1)k
values send in the whole system in each step. These values can have different size depending on the nature of the
configuration space. Common value types are boolean, integer, and floating point numbers. Using the values of scenario
SC3 with n = 6 and k = 3 this would result in 15 values send in each step by each agent.
While we believe that these requirements regarding the computational power and network capacity can be satisfied
for a variety of systems that can benefit from our approach it is clear that some systems have limitations that might
prevent the application of the influence detection. In this case, it is obviously possible to tweak the communication
complexity by using algorithms that are more efficient given the systems architecture or to rely on less computational
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
31
(a) The results for the pan.
(b) The results for the tilt.
(c) The results for the zoom.
Fig. 15. The results for the detection of influences in SCN 2. The influence detection rate for the pan, tilt and zoom of Camera 1 versus
the Camera 2 based on 10 independent runs. The cameras are controlled with the same algorithms as before but the pan, tilt, and
zoom configuration of Camera 2 are integrated in the state space of Camera 1 at the start of the runs. The graphs show the number
of runs, in which the influence of Camera 1 is measured higher than the influence of Camera 2, using the last 10k steps as samples.
Manuscript submitted to ACM
0 20 40 60 80 100 0 50000 100000 150000 200000 250000 300000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 0 50000 100000 150000 200000 250000 300000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 0 20 40 60 80 100 0 50000 100000 150000 200000 250000 300000Percentage of RunsStepsDistance CorrelationKendall CorrelationMaximal Information CoefficientMutual Information (continuous)Pearson CorrelationSpearman Correlation 32
S. Rudolph et al.
(a) The results with single learners.
(b) The results if the pan, tilt, and zoom configurations of Camera 2 are integrated in the state space of Camera 1 at the start of the
simulation.
(c) The results if the configurations are integrated in the state space of Camera 1 dynamically during runtime.
Fig. 16. The figure shows the rewards received by the system if the Q-learning algorithm is applied to SCN 2 with single learners,
with an integration at the start, and dynamic integration.
Manuscript submitted to ACM
0 2 4 6 8 10 12 14 16 0 50 100 150 200 250 300 350 400 450 500RewardSteps in 1000Average Reward 0 5 10 15 20 25 30 0 50 100 150 200 250 300 350 400 450 500RewardSteps in 1000Average Reward 0 5 10 15 20 25 30 0 50 100 150 200 250 300 350 400 450 500RewardSteps in 1000Average Reward On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
33
Fig. 17. SCN 3. A top-down view on a larger smart camera network with 6 cameras. The position of the cameras are depicted by the
black dots. Each of them has a colored circle around it marking the area that can be potentially observed by it and a red shape that
shows an exemplary observed area. The gray squares show the position of pillars limiting the potentially observable areas. The yellow
arrows mark entry points of observation targets and there movement direction.
intensive dependency measures. Besides such implementation details, we want to discuss two basic ideas that can lead
to a better scalability in general. These are the sequential analysis of systems and the reduction of network load.
Sequential Analysis of Systems. In the conducted experiments, each system measures the influence of each other
system on itself. This can lead to a high system load and possibly to a backlog in the detection of influences, especially
in large-scale systems.
To avoid such situations, it is possible to analyze each other system in a sequential order instead of simultaneously
(possibly with multiple measures), i.e., a system creates an order and starts analyzing the influence from the first system
in the sequence. As soon as it is done analyzing the first system it can switch to the next and so on. This order can be
randomly or based on the physical or virtual distance, for instance. This reduces the computational peak load heavily.
Reduction of Network Load. In the conducted experiments, we assumed that the influence origins from an arbitrary
node in the overall system, i.e., each of the systems distributes its configuration via the communication network to
the other systems. Depending on the connectivity of the systems, this can lead to a high network load affecting the
communication necessary to operate the systems, e.g., the results in a wireless sensor network cannot be forwarded to
the sink.
Manuscript submitted to ACM
Camera 2Camera 1Camera 3Camera 5Camera 4Camera 6 34
S. Rudolph et al.
(a) The cameras are controlled by single agent Q-learner.
(b) The results if 3k samples are used to determine the candidates for an integration.
(c) The results if 10k samples are used to determine the candidates for an integration.
Fig. 18. The figure shows the average reward over 10 runs in the scenario SCN 3. The cameras are each controlled by a Q-learning
algorithm with a ϵ-greedy action selection, where the ϵ decreases. It uses single learners or the integration of configurations at
runtime.
Manuscript submitted to ACM
0 0.5 1 1.5 2 2.5 3 0 50 100 150 200 250 300 350 400 450 500RewardSteps in 1000Average Reward 0 0.5 1 1.5 2 2.5 3 3.5 4 0 50 100 150 200 250 300 350 400 450 500RewardSteps in 1000Average Reward 0 0.5 1 1.5 2 2.5 3 3.5 0 50 100 150 200 250 300 350 400 450 500RewardSteps in 1000Average Reward On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
35
1
PAN,TILT,ZOOM
2
PAN,TILT
3
PAN,TILT,ZOOM
4
PAN
5
PAN,TILT
6
Fig. 19. A snapshot of the structure of the learning network in an exemplary run on SCN 3 after 200000 steps
Similar to the reduction of computational resources, this can be avoided by analyzing the systems sequentially. Another
effective way that can be used for some system types is to send the reward to the other systems instead of the
configuration. This switches the places where the influences are calculated, i.e., instead of system A receiving the
configurations of system B and calculating the influence of system B on itself, it sends the rewards of itself to system B
that in turn calculates the influence. Since the configuration will most often create more traffic then just the reward this
can reduce the computational overhead but this approach is limited to systems which are trusted and are willing to
share their resources.
7.2 Future Work
In the following, two directions for future work are outlined. They aim mainly at further automating the detection
process by an automatic combination of the dependency measures and to extend the applicability to systems that are
not covered by the current system model.
Automated Weighting of Dependency Measures. In this work, we used multiple dependency measures for the detection.
This is due to the different detection rates for the various classes of systems. An interesting approach to extend the
approach shown here is to combine the dependency measures to get a more reliable detection. This can be achieved by
relying on simple techniques such as majority voting, but a useful approach could be to consider ensemble learning
techniques [56]. Especially methods, where the combination of the learners is done by determining the weights for the
individual learners on a case by case basis regarding the structure of the system is promising. The main challenge here
lies in the engineering of a feedback signal or a training set that allows the combinator to find optimal weights for each
situation.
Detecting Influences in Systems without a Local Reward Signal. In this article, the system model relies on a local reward
that is a feedback signal to the subsystems regarding their decisions. If there is no obvious choice for such a reward
signal, it can be useful to have an influence detection mechanism that is able to work without such. Therefore, an
Manuscript submitted to ACM
36
S. Rudolph et al.
interesting approach for future work is to change the system model to a type where only the configuration and a current
state of the subsystems is necessary. The dependency can then be measured between one system's configuration and
another system's situation. However, this introduces the problem that the situation is very likely not a one dimensional
variable but has several dimensions, e.g., the position is composed of an x- and y-coordinate. Therefore, it will be
necessary to use dimension reduction techniques on the situation of a system or to measure the dependency between
two groups of random variables, e.g., with a canonical correlation analysis [57].
Besides these two concrete directions it is promising to apply the influence detection to further real-world systems,
especially with different learning algorithms and influence structures.
ACKNOWLEDGMENTS
We gratefully acknowledge the financial support of the German Research Foundation (DFG) in the context of the project
CYPHOC (HA 5480/3-1,SI 674/9-1).
Furthermore, we would like to thank our project partners Bernhard Sick, Christian Gruhl, Henner Heck (Uni Kassel,
IES group) and Arno Wacker (Universität der Bundeswehr München) for the fruitful discussions in the project.
REFERENCES
[1] J. O. Kephart and D. M. Chess, "The vision of autonomic computing," vol. 36, no. 1, pp. 41 -- 50.
[2] C. Müller-Schloer and S. Tomforde, Organic Computing -- Technical Systems for Survival in the Real World. Springer International Publishing, 2018.
[3] D. Tennenhouse, "Proactive computing," vol. 43, no. 5, pp. 43 -- 50.
[4] M. Wooldridge, An introduction to multiagent systems. John Wiley & Sons.
[5] S. Kernbach, T. Schmickl, and J. Timmis, "Collective adaptive systems: Challenges beyond evolvability," vol. abs/1108.5643.
[6] D. Weyns, B. Schmerl, V. Grassi, S. Malek, R. Mirandola, C. Prehofer, J. Wuttke, J. Andersson, H. Giese, and K. M. Göschka, "On patterns for
decentralized control in self-adaptive systems," in Software Engineering for Self-Adaptive Systems II, pp. 76 -- 107, Springer, 2013.
[7] S. Tomforde and C. Müller-Schloer, "Incremental design of adaptive systems," vol. 6, no. 2, pp. 179 -- 198.
[8] C. Krupitzer, F. M. Roth, S. VanSyckel, G. Schiele, and C. Becker, "A survey on engineering approaches for self-adaptive systems," Pervasive and
Mobile Computing, vol. 17, pp. 184 -- 206, 2015.
[9] M. W. Maier, "Architecting principles for systems-of-systems," Systems Engineering, vol. 1, no. 4, pp. 267 -- 284, 1998.
[10] S. Tomforde, J. Hähner, H. Seebach, W. Reif, B. Sick, A. Wacker, and I. Scholtes, "Engineering and Mastering Interwoven Systems," in ARCS 2014 -
27th International Conference on Architecture of Computing Systems, Workshop Proceedings, February 25-28, 2014, Luebeck, Germany, University of
Luebeck, Institute of Computer Engineering, pp. 1 -- 8.
[11] S. Tomforde, S. Rudolph, K. L. Bellman, and R. P. Würtz, "An organic computing perspective on self-improving system interweaving at runtime," in
2016 IEEE International Conference on Autonomic Computing, ICAC 2016, Wuerzburg, Germany, July 17-22, 2016, pp. 276 -- 284.
[12] S. Rudolph, S. Edenhofer, S. Tomforde, and J. Hähner, "Reinforcement learning for coverage optimization through PTZ camera alignment in highly
dynamic environments," in Proceedings of the International Conference on Distributed Smart Cameras, ICDSC '14, Venezia Mestre, Italy, November 4-7,
2014, pp. 19:1 -- 19:6.
[13] C. Piciarelli, L. Esterle, A. Khan, B. Rinner, and G. L. Foresti, "Dynamic reconfiguration in camera networks: A short survey," IEEE Transactions on
Circuits and Systems for Video Technology, vol. 26, no. 5, pp. 965 -- 977, 2016.
[14] K. Bellman, J. Botev, A. Diaconescu, L. Esterle, C. Gruhl, C. Landauer, P. Lewis, A. Stein, S. Tomforde, and R. Würtz, "Self-improving system
integration -- status and challenges after five years of sissy," 2018.
[15] S. Rudolph, R. Hihn, S. Tomforde, and J. Hähner, "Comparison of dependency measures for the detection of mutual influences in organic computing
systems," in Architecture of Computing Systems - ARCS 2016 - 29th International Conference, Nuremberg, Germany, April 4-7, 2016, Proceedings,
pp. 334 -- 347.
[16] S. Rudolph, S. Tomforde, and J. Hähner, "A mutual influence-based learning algorithm," in Proceedings of the 8th International Conference on Agents
and Artificial Intelligence (ICAART 2016), Volume 1, Rome, Italy, February 24-26, 2016., pp. 181 -- 189.
[17] S. Rudolph, S. Tomforde, B. Sick, and J. Hähner, "A mutual influence detection algorithm for systems with local performance measurement," in 2015
IEEE 9th International Conference on Self-Adaptive and Self-Organizing Systems, Cambridge, MA, USA, September 21-25, 2015, pp. 144 -- 149.
[18] S. Tomforde, B. Hurling, and J. Hähner, "Dynamic control of mobile ad-hoc networks - Network protocol parameter adaptation using Organic
Network Control," in Proc. of the 7th Int. Conf. on Informatics in Control, Automation, and Robotics (ICINCO'10), held in Funchal, Portugal (June 15 - 18,
2010), (Setubal), pp. 28 -- 35, INSTICC, 2010.
[19] R. S. Sutton and A. G. Barto, Introduction to Reinforcement Learning. MIT Press, 1st ed.
Manuscript submitted to ACM
On the Detection of Mutual Influences and Their Consideration in Reinforcement Learning Processes
37
[20] A. van Lamsweerde, "Goal-oriented requirements engineering: a guided tour," in Proceedings Fifth IEEE International Symposium on Requirements
Engineering, pp. 249 -- 262.
[21] A. van Lamsweerde, Requirements Engineering: From System Goals to UML Models to Software Specifications. Wiley Publishing, 1st ed.
[22] E. M. Fredericks, B. DeVries, and B. H. C. Cheng, "Towards run-time adaptation of test cases for self-adaptive systems in the face of uncertainty," in
Proceedings of the 9th International Symposium on Software Engineering for Adaptive and Self-Managing Systems, SEAMS 2014, (New York, NY, USA),
pp. 17 -- 26, ACM.
[23] S. Tomforde and J. Hähner, Biologically Inspired Networking and Sensing: Algorithms and Architectures, ch. Organic Network Control -- Turning
Standard Protocols Into Evolving Systems, pp. 11 -- 35. IGI, 2011.
[24] L. Busoniu, R. Babuska, and B. De Schutter, "A comprehensive survey of multiagent reinforcement learning," vol. 38, no. 2, pp. 156 -- 172.
[25] M. Wiering and M. van Otterlo, Reinforcement Learning: State-of-the-Art. Springer Publishing Company, Incorporated.
[26] C. J. C. H. Watkins and P. Dayan, "Technical note q-learning.," vol. 8, pp. 279 -- 292.
[27] S. Rudolph, S. Tomforde, B. Sick, H. Heck, A. Wacker, and J. Hähner, "An online influence detection algorithm for organic computing systems," in
ARCS 2015 - The 28th International Conference on Architecture of Computing Systems. Proceedings, pp. 1 -- 8.
[28] D. Keil and D. Q. Goldin, "Modeling indirect interaction in open computational systems," in 12th IEEE International Workshops on Enabling Technologies
(WETICE 2003), Infrastructure for Collaborative Enterprises, 9-11 June 2003, Linz, Austria, pp. 371 -- 376.
[29] R. Logie, J. G. Hall, and K. G. Waugh, "Towards mining for influence in a multi agent environment.," in IADIS European Conf. Data Mining (A. Abraham,
ed.), pp. 97 -- 101, IADIS.
[30] R. Logie, J. G. Hall, and K. G. Waugh, "Investigating agent influence and nested other-agent behaviour," International Journal on Advances in Intelligent
Systems Volume 2, Number 4, 2009, 2010.
[31] J. M. Broersen, "CTL.STIT: enhancing ATL to express important multi-agent system verification properties," in 9th International Conference on
Autonomous Agents and Multiagent Systems (AAMAS 2010), Toronto, Canada, May 10-14, 2010, Volume 1-3, pp. 683 -- 690.
[32] P. Stone and M. Veloso, "Multiagent systems: A survey from a machine learning perspective," vol. 8, no. 3, pp. 345 -- 383.
[33] J. R. Kok, M. T. J. Spaan, and N. Vlassis, "Multi-robot decision making using coordination graphs," in Proceedings of the International Conference on
Advanced Robotics (ICAR) (A. T. de Almeida and U. Nunes, eds.), pp. 1124 -- 1129.
[34] J. R. Kok, P. J. 't Hoen, B. Bakker, and N. Vlassis, "Utile coordination: learning interdependencies among cooperative agents," in Proceedings of the
IEEE Symposium on Computational Intelligence and Games (CIG), pp. 29 -- 36.
[35] Y.-M. De Hauwere, P. Vrancx, and A. Nowé, "Learning what to observe in multi-agent systems," in Proceedings of the 20th Belgian-Netherlands
Conference on Artificial Intelligence, pp. 83 -- 90, 2009.
[36] Y.-M. De Hauwere, P. Vrancx, and A. Nowé, "Learning multi-agent state space representations," in Proceedings of the 9th International Conference on
Autonomous Agents and Multiagent Systems: Volume 1, pp. 715 -- 722, International Foundation for Autonomous Agents and Multiagent Systems.
[37] Y.-M. De Hauwere, P. Vrancx, and A. Nowé, "Solving delayed coordination problems in mas," in The 10th International Conference on Autonomous
Agents and Multiagent Systems - Volume 3, AAMAS '11, (Richland, SC), pp. 1115 -- 1116, International Foundation for Autonomous Agents and
Multiagent Systems.
[38] M. Lanctot, V. Zambaldi, A. Gruslys, A. Lazaridou, K. Tuyls, J. Pérolat, D. Silver, and T. Graepel, "A unified game-theoretic approach to multiagent
reinforcement learning," in Advances in Neural Information Processing Systems, pp. 4190 -- 4203, 2017.
[39] H. Peng, F. Long, and C. Ding, "Feature selection based on mutual information: Criteria of max-dependency, max-relevance, and min-redundancy,"
vol. 27, no. 8, pp. 1226 -- 1238.
[40] A. M. Farahmand, M. Ghavamzadeh, C. Szepesvári, and S. Mannor, "Regularized policy iteration," in Advances in Neural Information Processing
Systems 21, Proceedings of the Twenty-Second Annual Conference on Neural Information Processing Systems, Vancouver, British Columbia, Canada,
December 8-11, 2008, pp. 441 -- 448.
[41] J. Z. Kolter and A. Y. Ng, "Regularization and feature selection in least-squares temporal difference learning," in Proceedings of the 26th Annual
International Conference on Machine Learning, ICML '09, (New York, NY, USA), pp. 521 -- 528, ACM, 2009.
[42] D.-R. Liu, H.-L. Li, and D. Wang, "Feature selection and feature learning for high-dimensional batch reinforcement learning: A survey," vol. 12, no. 3,
pp. 229 -- 242.
[43] J. Boes and F. Migeon, "Self-organizing multi-agent systems for the control of complex systems," Journal of Systems and Software, vol. 134, pp. 12 -- 28,
2017.
[44] S. Rudolph, R. Hihn, S. Tomforde, and J. Hähner, "Towards discovering delayed mutual influences in organic computing systems," in ARCS 2017; 30th
GI/ITG International Conference on Architecture of Computing Systems, pp. 39 -- 46.
[45] K. Pearson, "Note on regression and inheritance in the case of two parents," vol. 58, no. 347-352, pp. 240 -- 242.
[46] M. G. Kendall, "A new measure of rank correlation," Biometrika, vol. 30, no. 1/2, pp. 81 -- 93, 1938.
[47] G. J. Székely, M. L. Rizzo, and N. K. Bakirov, "Measuring and testing dependence by correlation of distances," vol. 35, no. 6, pp. 2769 -- 2794.
[48] C. Shannon and W. Weaver, The Mathematical Theory of Communication. University of Illinois Press.
[49] A. Kraskov, H. Stögbauer, and P. Grassberger, "Estimating mutual information," vol. 69, p. 066138.
[50] D. N. Reshef, Y. A. Reshef, H. K. Finucane, S. R. Grossman, G. McVean, P. J. Turnbaugh, E. S. Lander, M. Mitzenmacher, and P. C. Sabeti, "Detecting
novel associations in large data sets," vol. 334, no. 6062, pp. 1518 -- 1524.
Manuscript submitted to ACM
38
S. Rudolph et al.
[51] S. Luke, C. Cioffi-Revilla, L. Panait, K. Sullivan, and G. Balan, "Mason: A multiagent simulation environment," Simulation: Transactions of the society
for Modeling and Simulation International, vol. 82, no. 7, pp. 517 -- 527, 2005.
[52] M. E. Taylor and P. Stone, "Transfer learning for reinforcement learning domains: A survey," Journal of Machine Learning Research, vol. 10, no. Jul,
pp. 1633 -- 1685, 2009.
[53] S. W. Wilson, "Classifier Fitness Based on Accuracy," vol. 3, no. 2, pp. 149 -- 175.
[54] S. W. Wilson, "Get real! xcs with continuous-valued inputs," in Learning Classifier Systems (P. L. Lanzi, W. Stolzmann, and S. W. Wilson, eds.), (Berlin,
Heidelberg), pp. 209 -- 219, Springer Berlin Heidelberg, 2000.
[55] X. Li and G. Yang, "Transferable xcs," in Proceedings of the Genetic and Evolutionary Computation Conference 2016, GECCO '16, (New York, NY, USA),
pp. 453 -- 460, ACM, 2016.
[56] B. Krawczyk, L. L. Minku, J. Gama, J. Stefanowski, and M. Woźniak, "Ensemble learning for data stream analysis: A survey," Information Fusion,
vol. 37, pp. 132 -- 156, 2017.
[57] W. Härdle and L. Simar, Applied multivariate statistical analysis, vol. 22007. Springer, 2007.
Received September 2018
Manuscript submitted to ACM
|
1810.10862 | 4 | 1810 | 2019-04-14T07:39:06 | Multiparty Dynamics and Failure Modes for Machine Learning and Artificial Intelligence | [
"cs.MA",
"cs.AI"
] | An important challenge for safety in machine learning and artificial intelligence systems is a~set of related failures involving specification gaming, reward hacking, fragility to distributional shifts, and Goodhart's or Campbell's law. This paper presents additional failure modes for interactions within multi-agent systems that are closely related. These multi-agent failure modes are more complex, more problematic, and less well understood than the single-agent case, and are also already occurring, largely unnoticed. After motivating the discussion with examples from poker-playing artificial intelligence (AI), the paper explains why these failure modes are in some senses unavoidable. Following this, the paper categorizes failure modes, provides definitions, and cites examples for each of the modes: accidental steering, coordination failures, adversarial misalignment, input spoofing and filtering, and goal co-option or direct hacking. The paper then discusses how extant literature on multi-agent AI fails to address these failure modes, and identifies work which may be useful for the mitigation of these failure modes. | cs.MA | cs |
Multiparty Dynamics and Failure Modes for Machine
Learning and Artificial Intelligence
David Manheim
Abstract: An important challenge for safety in machine learning and artificial intelligence systems
is a set of related failures involving specification gaming, reward hacking, fragility to distributional
shifts, and Goodhart's or Campbell's law. This paper presents additional failure modes for
interactions within multi-agent systems that are closely related. These multi-agent failure modes
are more complex, more problematic, and less well understood than the single-agent case, and are
also already occurring, largely unnoticed. After motivating the discussion with examples from
poker-playing artificial intelligence (AI), the paper explains why these failure modes are in some
senses unavoidable. Following this, the paper categorizes failure modes, provides definitions,
and cites examples for each of the modes: accidental steering, coordination failures, adversarial
misalignment, input spoofing and filtering, and goal co-option or direct hacking. The paper then
discusses how extant literature on multi-agent AI fails to address these failure modes, and identifies
work which may be useful for the mitigation of these failure modes.
Keywords: multi-agent systems; specification gaming; artificial intelligence safety; Goodhart's Law
MSC: 91E45; 91A06
JEL: C79; D74
1. Background, Motivation and Contribution
When complex systems are optimized by a single agent, the representation of the system
and of the goal used for optimization often leads to failures that can be surprising to the agent's
designers. These failure modes go by a variety of names, Amodei and Clark called them faulty
reward functions [1] but similar failures have been referred to as Goodhart's law [2,3], Campbell's
law [4], distributional shift [5], strategic behavior [6], reward hacking [7], Proxyeconomics [8], and
other terms.
Examples of these failures in the single-agent case are shown by Victoria Krakovna's extensive
list of concrete examples of "generating a solution that literally satisfies the stated objective but
fails to solve the problem according to the human designer's intent." [9] Liu et al. suggest that
"a complex activity can often be performed in several different ways," [10] but not all these ways
should be considered valid. To understand why, Krakovna' s list includes examples of "achieving
a goal" by finding and exploiting bugs in a simulation engine to achieve goals [11 -- 13]; by physical
manipulation of objects in unanticipated ways, such as moving a table instead of the item on the
table [14], or flipping instead of lifting a block [15]; and even by exploiting the problem structure or
evaluation, such as returning an empty list as being sorted [16], or deleting the file containing the
target output [16].
1.1. Motivation
This forms only a part of the broader set of concerns in AI safety, [5,17 -- 19], but the failure modes
are the focus of a significant body of work in AI safety discussed later in the paper. However, as the
systems become more capable and more widely used, Danzig and others have noted that this will
"increase rather than reduce collateral risks of loss of control." [20] The speed of such systems is
almost certainly beyond the point of feasible human control, and as they become more complex, the
systems are also individually likely to fail in ways that are harder to understand.
While some progress has been made in the single-agent case, the systems have continued to
become more capable, corporations, governments, and other actors have developed and deployed
1
1.2 Contribution
2 of 14
machine learning systems. These are not only largely autonomous, but also interact with each other.
This allows a new set of failures, and these are not yet a focus of safety-focused research -- but they
are critical.
1.2. Contribution
The analogues of the earlier-mentioned classes of failure for multi-agent systems are more
complex, potentially harder to mitigate, and unfortunately not the subject of a significant focus
among AI safety researchers. In this paper, we introduce a classification of failures that are not yet
well-addressed in the literature involving multiple agents. These failures can occur even when system
designers do not intend to build conflicting AI or ML systems. The current paper contributes to the
literature by outlining how and why these multi-agent failures can occur, and providing an overview
of approaches that could be developed for mitigating them. In doing so, the paper will hopefully
help spur system designers to explicitly consider these failure modes in designing systems, and urge
caution on the part of policymakers.
As a secondary contribution, the link between ongoing work on AI safety and potential work
mitigating these multi-agent failures incidentally answers an objection raised by AI risk skeptics that
AI safety is "not worth current attention" and that the issues are "premature to worry about" [21].
This paper instead shows how failures due to multi-agent dynamics are critical in the present, as ML
and superhuman narrow AI is being widely deployed, even given the (valid) arguments put forward
by Yudkowsky [22] and Bostrom [7] for why a singleton AI is a more important source of existential
risk.
1.3. Extending Single-Agent Optimization Failures
Systems which are optimized using an imperfect system model have several important failure
modes categorized in work by Manheim and Garrabrant [3]. First, imperfect correlates of the goal
will be less correlated in the tails of the distribution, as discussed by Lewis [23]. Heavily optimized
systems will end up in those regions, and even well-designed metrics do not account for every
possible source of variance. Second, there are several context failures [24], where the optimization
is well behaved in the training set ("ancestral environment") but fails as optimization pressure is
applied. For example, it may drift towards an "edge instantiation" where the system may optimize all
the variables that relate to the true goal, but further gain on the metric is found by unexpected means.
Alternatively, the optimizer may properly obey constraints in the initial stage, but find some "nearest
unblocked strategy" [24] allowing it to circumvent designed limits when given more optimization
power. These can all occur in single-agent scenarios.
The types of failure in multi-agent systems presented in this paper can be related to Manheim
and Garrabrant's classification of single-agent metric optimization failures . The four single-agent
overoptimization failure modes outlined there are:
• Tails Fall Apart, or Regressional inaccuracy, where the relationship between the modeled goal and
the true goal is inexact due to noise (for example, measurement error,) so that the bias grows as
the system is optimized.
• Extremal Model Insufficiency, where the approximate model omits factors which dominate the
• Extremal Regime Change, where the model does not include a regime change that occurs under
• Causal Model Failure, where the agent's actions are based on a model which incorrectly represents
causal relationships, and the optimization involves interventions that break the causal structure
the model implicitly relies on.
certain (unobserved) conditions that optimization creates.
system's behavior after optimization.
Despite the completeness of the above categorization, the way in which these failures occur can
differ greatly even when only a single agent is present. In a multi-agent scenario, agents can stumble
into or intentionally exploit model overoptimization failures in even more complex ways. Despite
this complexity, the different multi-agent failure modes can be understood based on understanding
the way in which the implicit or explicit system models used by agents fail.
1.4 Defining Multi-Agent Failures
3 of 14
1.4. Defining Multi-Agent Failures
In this paper, a multi-agent optimization failure is when one (or more) of the agents which can
achieve positive outcomes in some scenarios exhibits behaviors that negatively affect its own outcome
due to the actions of one or more agents other than itself. This occurs either when the objective
function of the agent no longer aligns with the goal, as occurs in the Regressional and both Extremal
cases, or when the learned relationship between action(s), the metric(s), and the goal have changed,
as in the Causal failure case.
This definition does not require the failure to be due to malicious behavior on the part of any
agent, nor does it forbid it. Note also that the definition does not require failure of the system, as in
Behzadan and Munir's categorization of adversarial attacks [25], nor does it make any assumptions
about type of the agents, such as the type of learning or optimization system used. (The multi-agent
cases implicitly preclude agents from being either strongly boxed, as Drexler proposed [26], or
oracular, as discussed by Armstrong [27].)
2. Multi-Agent Failures: Context and Categorization
Several relatively straightforward failure modes involving interactions between an agent and
a regulator were referred to in Manheim and Garrabrant as adversarial Goodhart [3]. These
occur where one AI system opportunistically alters or optimizes the system and uses the expected
optimization of a different victim agent to hijack the overall system. For example, "smart market"
electrical grids use systems that optimize producer actions and prices with a linear optimization
system using known criteria. If power lines or power plants have strategically planned maintenance
schedules, an owner can manipulate the resulting prices to its own advantage, as occurred (legally) in
the case of Enron [28]. This is possible because the manipulator can plan in the presence of a known
optimization regime.
This class of manipulation by an agent frustrating a regulator's goals is an important case,
but more complex dynamics can also exist, and Manheim and Garrabrant noted that there are
"clearly further dynamics worth exploring." [3] This involves not only multiple heterogenous agents,
which Kleinberg and Raghavan suggest an avenue for investigating, but also interaction between
those agents [6]. An example of a well-understood multi-agent system, the game of poker, allows
clarification of why the complexity is far greater in the interaction case.
2.1. Texas Hold'em and the Complexity of Multi-Agent Dynamics
In many-agent systems, simple interactions can become complex adaptive systems due to agent
behavior, as the game of poker shows. Solutions to simplified models of two-player poker predate
game theory as a field [29], and for simplified variants, two-player draw poker has a fairly simple
optimal strategy [30]. These early, manually computed solutions were made possible both by limiting
the complexity of the cards, and more importantly by limiting interaction to a single bet size, with no
raising or interaction between the players. In the more general case of heads-up limit Texas Hold'em,
significantly more work was needed, given the multiplicity of card combinations, the existence of
hidden information, and player interaction, but this multi-stage interactive game is "now essentially
weakly solved" [31]. Still, this game involves only two players. In the no-limit version of the game,
Brown and Sandholm recently unveiled superhuman AI [32], which restricts the game to "Heads'
Up" poker, which involves only two players per game, and still falls far short of a full solution to the
game.
The complex adaptive nature of multi-agent systems means that each agent needs model not
only model the system itself, but also the actions of the other player(s). The multiplicity of potential
outcomes, betting strategies, and different outcomes becomes rapidly infeasible to represent other
than heuristically. In limit Texas Hold'em poker, for example, the number of card combinations is
immense, but the branching possibilities for betting is the more difficult challenge. In a no-betting
game of Hold'em with P players, there are 52!/((52− 2P − 5)! · 2P · 5!) possible situations. This is 2.8·
1012 hands in the two-player case, 3.3 · 1015 in the three-player case, and growing by a similar factor
when expanded to the four-, five-, or six-player case. The probability of winning is the probability
that the five cards on the table plus two unknown other cards from the deck are a better hand than
any that another player holds. In Texas Hold'em, there are four betting stages, one after each stage
of cards is revealed. Billings et al. use a reduced complexity game (limiting betting to three rounds
2.2 Limited Complexity Models versus the Real World
4 of 14
per stage) and find a complexity of O(1018) in the two-hand case [33]. That means the two-player,
three-round game complexity is comparable in size to a no-betting four-player game, with 4.1 · 1018
card combinations possible.
Unlike a no-betting game, however, a player must consider much more than the simple
probability that the hand held is better than those held by other players. That calculation is
unmodified during the additional branching due to player choices. The somewhat more difficult
issue is that the additional branching requires Bayesian updates to estimate the probable distribution
of hand strengths held by other players based on their decisions, which significantly increases the
complexity of solving the game. The most critical challenge, however, is that each player bets based
on the additional information provided by not only the hidden information provided by their cards,
but also based on the betting behavior of other players. Opponent(s) make betting decisions based on
non-public information (in Texas Hold'em, hole cards) and strategy for betting requires a meta-update
taking advantage of the information the other player reveals by betting. The players must also update
based on potential strategic betting by other players, which occurs when a player bets in a way
calculated to deceive. To deal with this, poker players need to model not just the cards, but also
the strategic decisions of other players. This complex model of strategic decisions must be re-run for
all the possible combinations at each decision point to arrive at a conclusion about what other players
are doing. Even after this is complete, an advanced poker player, or an effective AI, must then decide
not just how likely they are to win, but also how to play strategically, optimizing based on how other
players will react to the different choices available.
Behaviors such as bluffing and slow play are based on these dynamics, which become much
more complex as the number of rounds of betting and the number of players increases. For example,
slow play involves underbidding compared to the strength of your hand. This requires that the
players will later be able to raise the stakes, and allows a player to lure others into committing
additional money. The complexity of the required modeling of other agents' decision processes grows
as a function of the number of choices and stages at which each agent makes a decision. This type of
complexity is common in multi-agent systems. In general, however, the problem is much broader in
scope than what can be illustrated by a rigidly structured game such as poker.
2.2. Limited Complexity Models versus the Real World
In machine learning systems, the underlying system is approximated by implicitly or explicitly
learning a multidimensional transformation between inputs and outputs. This transformation
approximates a combination of the relationships between inputs and the underlying system,
and between the system state and the outputs. The complexity of the model learned is limited by
the computational complexity of the underlying structure, and while the number of possible states
for the input is large, it is typically dwarfed by the number of possible states of the system.
The critical feature of machine learning that allows such systems to be successful is that most
relationships can be approximated without inspecting every available state.
(All models simplify
the systems they represent.) The implicit simplification done by machine learning is often quite
impressive, picking up on clues present in the input that humans might not notice, but it comes
at the cost of having difficult to understand and difficult to interpret implicit models of the system.
Any intelligence, whether machine learning-based, human, or AI, requires similar implicit
simplification, since the branching complexity of even a relatively simple game such as Go dwarfs
the number of atoms in the universe. Because even moderately complex systems cannot be fully
represented, as discussed by Soares [34], the types of optimization failures discussed above are
inevitable. The contrapositive to Conant and Ashby's theorem [35] is that if a system is more complex
than the model, any attempt to control the system will be imperfect. Learning, whether human or
machine, builds approximate models based on observations, or input data. This implies that the
behavior of the approximation in regions far from those covered by the training data is more likely
to markedly differ from reality. The more systems change over time, the more difficult prediction
becomes -- and the more optimization is performed on a system, the more it will change. Worsening
this problem, the learning that occurs in ML systems fails to account for the embedded agency issues
discussed by Demski and Garrabrant [36], and interaction between agents with implicit models of
each other and themselves amplifies many of these concerns.
2.3 Failure modes
2.3. Failure modes
5 of 14
Because an essential part of multi-agent dynamic system modeling is opponent modeling, the
opponent models are a central part of any machine learning model. These opponent models may be
implicit in the overall model, or they may be explicitly represented, but they are still models that are
approximate. In many cases, opponent behavior is ignored -- by implicitly simplifying other agent
behavior to noise, or by assuming no adversarial agents exist. Because these models are imperfect,
they will be vulnerable to overoptimization failures discussed above.
The list below is conceptually complete, but limited in at least three ways. First, examples given
in this list primarily discuss failures that occur between two parties, such as a malicious actor and
a victim, or failures induced by multiple individually benign agents. This would exclude strategies
where agents manipulate others indirectly, or those where coordinated interaction between agents
is used to manipulate the system. It is possible that when more agents are involved, more specific
classes of failure will be relevant.
Second, the below list does not include how other factors can compound metric failures.
These are critical, but may involve overoptimization, or multiple-agent interaction, only indirectly.
For example, O'Neil discusses a class of failure involving the interaction between the system, the
inputs, and validation of outputs [37]. These failures occur when a system's metrics are validated in
part based on outputs it contributes towards. For example, a system predicting greater crime rates in
areas with high minority concentrations leads to more police presence, which in turn leads to a higher
rate of crime found. This higher rate of crime in those areas is used to train the model, which leads it to
reinforce the earlier unjustified assumption. Such cases are both likely to occur, and especially hard
to recognize, when the interaction between multiple systems is complex, and it is unclear whether
the system's effects are due in part to its own actions (This class of failure seems particularly likely in
systems that are trained via "self-play," where failures in the model of the system get reinforced by
incorrect feedback on the basis of the models, which is also a case of model insufficiency failure.).
Third and finally,
the failure modes exclude cases that do not directly involve metric
overoptimizations, such as systems learning unacceptable behavior implicitly due to training data
that contains unanticipated biases, or failing to attempt to optimize for social preferences such as
fairness. These are again important, but they are more basic failures of system design.
With those caveats, we propose the following classes of multi-agent overoptimization failures.
For each, a general definition is provided, followed by one or more toy models that demonstrate
the failure mode. Each agent attempts to achieve their goal by optimizing for the metric, but the
optimization is performed by different agents without any explicit coordination or a priori knowledge
about the other agents. The specifics of the strategies that can be constructed and the structure of the
system can be arbitrarily complex, but as explored below, the ways in which these models fail can
still be understood generally.
These models are deliberately simplified, but where possible, real-world examples of the failures
exhibited in the model are suggested. These examples come from both human systems where parallel
dynamics exist, and examples of the failures in extent systems with automated agents. In the toy
models, Mi and Gi stands for the metric and goal, respectively, for agent i. The metric is an imperfect
proxy for the goal, and will typically be defined in relation to a goal. (The goal itself is often left
unspecified, since the model applies to arbitrary systems and agent goals.) In some cases, the failure is
non-adversarial, but where relevant, there is a victim agent V and an opponent agent O that attempts
to exploit it. Please note that the failures can be shown with examples formulated with game-theoretic
notation, but doing so requires more complex specifications of the system and interactions than is
possible using the below characterization of the agent goals and the systems.
Failure Mode 1. Accidental Steering is when multiple agents alter the systems in ways not
anticipated by at least one agent, creating one of the above-mentioned single-party overoptimization
failures.
Remark 1. This failure mode manifests similarly to the single-agent case and differs only in that agents do not
anticipate the actions of other agents. When agents have closely related goals, even if those goals are aligned,
it can exacerbate the types of failures that occur in single-agent cases.
Because the failing agent alone does not (or cannot) trigger the failure, this differs from the single-agent
case. The distributional shift can occur due to a combination of actors' otherwise potentially positive influences
2.3 Failure modes
6 of 14
by either putting the system in an extremal state where the previously learned relationship decays, or triggering
a regime change where previously beneficial actions are harmful.
Model. 1.1 -- Group Overoptimization. A set of agents each have goals which affect the system in
related ways, and metric-goal relationship changes in the extremal region where x>a. As noted above,
Mi and Gi stands for the metric and goal, respectively, for agent i. This extremal region is one where
single-agent failure modes will occur for some or all agents. Each agent i can influence the metric
by an amount αi, where ∑ αi > a, but ∀αi < a. In the extremal subspace where Mi > a, the metric
reverses direction, making further optimization of the metric harm the agent's goal.
(cid:40)
Mi =
Gi,
where Mi <= a
Mi(a) − Gi, where Mi > a
(1)
Remark 2. In the presence of multiple agents without coordination, manipulation of factors not already being
manipulated by other agents is likely to be easier and more rewarding, potentially leading to inadvertent steering
due to model inadequacy, as discussed in Manheim and Garrabrant's categorization of single-agent cases [3].
As shown there overoptimization can lead to perverse outcomes, and the failing agent(s) can hurt both their
own goals, and in similar ways, can lead to negative impacts on the goals of other agents.
Model. 1.2 -- Catastrophic Threshold Failure.
Mi = xi
Gi =
(cid:40)
a + (∑∀i xi) where ∑∀i xi <= T
a − (∑∀i xi) where ∑∀i xi > T
(2)
Each agent manipulates their own variable, unaware of the overall impact. Even though the agents
are collaborating, because they cannot see other agents' variables, there is no obvious way to limit
the combined impact on the system to stay below the catastrophic threshold T. Because each agent is
exploring a different variable, they each are potentially optimizing different parts of the system.
Remark 3. This type of catastrophic threshold is commonly discussed in relations to complex adaptive systems,
but can occur even in systems where the catastrophic threshold is simple. The case discussed by Michael
Eisen involves pricing on Amazon was due to a pair of deterministic linear pricing-setting bots interacting
to set the price of an otherwise unremarkable biology book at tens of millions of dollars, showing that runaway
dynamics are possible even in the simplest cases [38]. This phenomenon is also expected whenever exceeding
some constraint breaks the system, and such constraints are often not identified until a failure occurs.
Example 1. This type of coordination failure can occur in situations such as overfishing across multiple
regions, where each group catches local fish, which they can see, but at a given threshold across regions the
fish population collapses, and recovery is very slow. (In this case, the groups typically are selfish rather than
collaborating, making the dynamics even more extreme.)
Example 2. Smaldino and McElreath [39] shows this failure mode specifically occurring with statistical
methodology in academia, where academics find novel ways to degrade statistical rigor. The more general
"Mutable Practices" model presented by Braganza [8], based on part on Smaldino and McElreath, has each
agent attempting to both outperform the other agents on a metric as well as fulfill a shared societal goal, allows
agents to evolve and find new strategies that combine to subvert a societal goal.
Failure Mode 2. Coordination Failure occurs when multiple agents clash despite having potentially
compatible goals.
Remark 4. Coordination is an inherently difficult task, and can in general be considered impossible [40].
In practice, coordination is especially difficult when the goals of other agents are incompletely known or not
fully understood. Coordination failures such as Yudkowsky's Inadequate equilibria are stable, and coordination
to escape from such an equilibrium can be problematic even when agents share goals [41].
Model. 2.1 -- Unintended Resource Contention. A fixed resource R is split between uses Rn by
different agents. Each agent has limited funds fi, and Ri is allocated to agent i for exploitation
2.3 Failure modes
7 of 14
in proportion to their bid for the resources cRi. The agents choose amounts to spend on acquiring
resources, and then choose amounts sni to exploit each resource, resulting in utility U(sn, Rn). The
agent goals are based on the overall exploitation of the resources by all agents.
cRi
Gi = ∑
∀i
Ri =
∑∀i cRi
Ui,n(sni, Rn)
(3)
In this case, we see that conflicting instrumental goals that neither side anticipates will cause wasted
funds due to contention. The more funds spent on resource capture, which is zero-sum, the less
remaining for exploitation, which can be positive-sum. Above nominal spending on resources to
capture them from aligned competitor-agents will reduce funds available for exploitation of those
resources, even though less resource contention would benefit all agents.
Remark 5. Preferences and gains from different uses can be homogeneous, so that all agents have no
marginal gain from affecting the allocation, funds will be wasted on resource contention. More generally,
heterogeneous preferences can lead to contention to control the allocation, with sub-optimal individual
outcomes, and heterogeneous abilities can lead to less-capable agents harming their goals by capturing then
ineffectively exploiting resources.
Example 3. Different forms of scientific research benefit different goals differently. Even if spending in
every area benefits everyone, a fixed pool of resources implies that with different preferences, contention
between projects with different positive impacts will occur. To the extent that effort must be directed
towards grant-seeking instead of scientific work, the resources available for the projects themselves are reduced,
sometimes enough to cause a net loss.
Remark 6. Coordination limiting overuse of public goods is a major area of research in economics. Ostrom
explains how such coordination is only possible when conflicts are anticipated or noticed and where a reliable
mechanism can be devised [42].
Model. 2.2 -- Unnecessary Resource Contention. As above, but each agent has an identical reward
function of fi,n. Even though all goals are shared, a lack of coordination in the above case leads
to overspending, as shown in simple systems and for specified algebraic objective functions in the
context of welfare economics. This literature shows many methods for how gains are possible, and in
the simplest examples this occurs when agents coordinate to minimize overall spending on resource
acquisition.
Remark 7. Coordination mechanisms themselves can be exploited by agents. The field of algorithmic game
theory has several results for why this is only sometimes possible, and how building mechanisms to avoid such
exploitation is possible [43].
Failure Mode 3. Adversarial optimization can occur when a victim agent has an incomplete model
of how an opponent can influence the system. The opponent's model of the victim allows it
to intentionally select for cases where the victim's model performs poorly and/or promotes the
opponent's goal [3].
Model. 3.1 -- Adversarial Goal Poisoning.
GV = x
GO = −x
MV = X : X ∼ normal(x, σ2(y))
MO = (X, y)
(4)
In this case, the Opponent O can see the metric for the victim, and can select for cases where y is large
and X is small, so that V chooses maximal values of X, to the marginal benefit of O.
Example 4. A victim's model can be learned by "Stealing" models using techniques such as those explored
by Tramèr et al. [44]. In such a case, the information gained can be used for model evasion and other attacks
mentioned there.
2.3 Failure modes
8 of 14
Example 5. Chess and other game engines may adaptively learn and choose openings or strategies for which
the victim is weakest.
Example 6. Sophisticated financial actors can make trades to dupe victims into buying or selling an asset
("Momentum Ignition") in order to exploit the resulting price changes [45], leading to a failure of the exploited
agent due to an actual change in the system which it misinterprets.
Remark 8. The probability of exploitable reward functions increases with the complexity of the system the
agents manipulate [5], and the simplicity of the agent and their reward function. The potential for exploitation
by other agents seems to follow the same pattern, where simple agents will be manipulated by agents with more
accurate opponent models.
Model. 3.2 -- Adversarial Optimization Theft. An attacker can discover exploitable quirks in the
goal function to make the victim agent optimize for a new goal, as in Manheim and Garrabrant's
Campbell's law example, slightly adapted here [3].
MV = GV + X
MO = GO · X
(5)
O selects MO after seeing V's choice of metric. In this case, we can assume the opponent chooses a
metric to maximize based on the system and the victim's goal, which is known to the attacker. The
opponent can choose their MO so that the victim's later selection then induces a relationship between
X and the opponent goal, especially at the extremes. Here, the opponent selects such that even weak
selection on MO hijacks the victim's selection on MV to achieve their goal, because states where MV is
high have changed. In the example given, if X ∼ normal(µ, σ2), the correlation between GO and MO
is zero over the full set of states, but becomes positive on the subspace selected by the victim. (Please
note that the opponent choice of metric is not itself a useful proxy for their goal absent the victim's
actions -- it is a purely parasitic choice.)
Failure Mode 4. Input spoofing and filtering -- Filtered evidence can be provided, or false evidence
can be manufactured and put into the training data stream of a victim agent.
Model. 4.1 -- Input Spoofing.
the present
world-state, and builds a model to choose actions which return rewards f (xt). The opponent can
generate events xi to poison the victim's learned model.
Victim agent receives public data D(xit) about
Remark 9. See the classes of data poisoning attacks explored by Wang and Chaudhuri [46] against online
learning, and of Chen et al [47]. for creating backdoors in deep-learning verification systems.
Example 7. Financial market participants can (illegally) spoof by posting orders that will quickly be canceled
in a "momentum ignition" strategy to lure others into buying or selling, as has been alleged to be occurring
in high-frequency-trading [45]. This differs from the earlier example in that the transactions are not bona-fide
transactions which fool other agents, but are actually false evidence.
Example 8. Rating systems can be attacked by inputting false reviews into a system, or by discouraging
reviews by those likely to be the least or most satisfied reviewers.
Model. 4.2 -- Active Input Spoofing. As in (4.1), where the victim agent employs active learning. In
this case, the opponent can potentially fool the system into collecting data that seems very useful to
the victim from crafted poisoned sources.
Example 9. Honeypots can be placed, or Sybil attacks mounted by opponents to fool victims into learning from
examples that systematically differ from the true distribution.
Example 10. Comments by users "Max" and "Vincent DeBacco" on Eisen's blog post about Amazon pricing
suggested that it is very possible to abuse badly built linear pricing models on Amazon to receive discounts, if
the algorithms choose prices based on other quoted prices [38].
9 of 14
Model. 4.3 -- Input Filtering. As in (4.1), but instead of generating false evidence, true evidence is
hidden to systematically alter the distribution of events seen.
Example 11. Financial actors can filter the evidence available to other agents by performing transactions they
do not want seen as private transactions or dark pool transactions.
Remark 10. There are classes of system where it is impossible to generate arbitrary false data points, but
selective filtering can have similar effects.
Failure Mode 5. Goal co-option is when an opponent controls the system the Victim runs on, or
relies on, and can therefore make changes to affect the victim's actions.
Remark 11. Whenever the computer systems running AI and ML systems are themselves insecure, it presents
a very tempting weak point that potentially requires much less effort than earlier methods of fooling the system.
Model. 5.1 -- External Reward Function Modification. Opponent O directly modifies Victim V's
reward function to achieve a different objective than the one originally specified.
Remark 12. Slight changes in a reward function may have non-obvious impacts until after the system
is deployed.
Model. 5.2 -- Output Interception. Opponent O intercepts and modifies Victim V's output.
Model. 5.3 -- Data or Label Interception. Opponent O modifies externally stored scoring rules
(labels) or data inputs provided to Victim V's output.
Example 12. Xiao, Xiao, and Eckert explore a "label flipping" attack against support vector machines [48]
where modifying a limited number of labels used in the training set can cause performance to deteriorate
severely.
Remark 13. As noted above, there are cases where generating false data may be impossible or easily detected.
Modifying the inputs during training may create less obvious traces of an attack has occurred. Where this is
impossible, access can also allow pure observation which, while not itself an attack, can allow an opponent to
engage in various other exploits discussed earlier.
To conclude the list of failure modes, it is useful to note a few areas where the failures are
induced or amplified. This is when agents explicitly incentivize certain behaviors on the part of
other agents, perhaps by providing payments. These public interactions and incentive payments are
not fundamentally different from other failure modes, but can create or magnify any of the other
modes. This is discussed in literature on the evolution of collusion, such as Dixon's treatment [49].
Contra Dixon, however, the failure modes discussed here can prevent the collusion from being
beneficial. A second, related case is when creating incentives where an agent fails to anticipate
either the ways in which the other agents can achieve the incentivized target, or the systemic changes
that are induced. These so-called "Cobra effects" [3] can lead to both the simpler failures of the
single-agent cases explored in Manheim and Garrabrant, and lead to the failures above. Lastly, as
noted by Sandberg [50], agents with different "speeds" (and, equivalently, processing power per unit
time,) can exacerbate victimization, since older and slower systems are susceptible, and susceptibility
to attacks only grows as new methods of exploitation are found.
3. Discussion
Multi-agent systems can naturally give rise to cooperation instead of competition, as discussed
in Leibo et al.'s 2017 paper [51]. The conditions under which there is exploitation rather than
cooperation, however, are less well understood. A more recent paper by Leibo proposes that the
competition dynamic can be used to encourage more complex models. This discusses coordination
failures, but the discussion of dynamics leading to the failures does not engage with the literature
on safety or goal-alignment [52]. Leibo's work, however, differs from most earlier work where
multi-agent systems are trained together with a single goal, perforce leading to cooperative behavior,
10 of 14
as in Lowe et al.'s heavily cited work,
pre-programming explicit models of other agent behaviors [53].
in which "competitive" dynamics are dealt with by
The failure modes outlined (accidental steering, coordination failures, adversarial misalignment,
input spoofing or filtering, and goal co-option or direct hacking) are all due to models that do
not fully account for other agent behavior. Because all models must simplify the systems they
represent, the prerequisites for these failures are necessarily present in complex-enough systems
where multiple non-coordinated agents interact. The problems of embedded agents discussed by
Demski and Garrabrant [36] make it particularly clear that current approaches are fundamentally
unable to fully represent these factors. For this and other reasons, mitigating the failures modes
discussed here are not yet central to the work of building better ML or narrow AI systems. At the same
time, some competitive domains such as finance are already experiencing some of these exploitative
failures [45], and bots engaging in social network manipulation, or various forms of more direct
interstate competition are likely engaging in similar strategies.
The failures seen so far are minimally disruptive. At the same time, many of the outlined failures
are more problematic for agents with a higher degree of sophistication, so they should be expected not
to lead to catastrophic failures given the types of fairly rudimentary agents currently being deployed.
For this reason, specification gaming currently appears to be a mitigable problem, or as Stuart Russell
claimed, be thought of as "errors in specifying the objective, period" [54]. This might be taken to
imply that these failures are avoidable, but the current trajectory of these systems means that the
problems will inevitably worsen as they become more complex and more such systems are deployed,
and the approaches used are fundamentally incapable of overcoming the obstacles discussed.
Potential Avenues for Mitigation
Mitigations for these failures exist, but as long as the fundamental problems discussed by
Demski and Garrabrant [36] are unaddressed, the dynamics driving these classes of failure seem
unavoidable. Furthermore, such failures are likely to be surprising. They will emerge as multiple
machine learning agents are deployed, and more sophisticated models will be more likely to trigger
them. However, as argued above, these failures are fundamental to interaction between complex
agents. This means that while it is unclear how quickly such failures will emerge, or if they will
be quickly recognized, it is unquestionable that they will continue to occur. System designers and
policymakers should expect that these problems will become intractable if deferred, and are therefore
particularly critical to address now. It is be expected that any solution involves a combination of
approaches [17], though the brief overview of safety approaches below shows that not all general
approaches to AI safety are helpful for multi-agent failures.
First, there are approaches that limit optimization. This can be done via satisficing, using
approaches such as Taylor's Quantilizers, which pick actions at random from the top quantile of
evaluated choices [55]. Satisficing approaches can help in prevent exploiting other agents, or in
preventing accidental overoptimization, but are not effective as a defense against exploitative agents
or systemic failures due to agent interaction. Another approach limiting optimization is explicit
safety guarantees. In extrema, this looks like an AI-Box, preventing any interaction of the AI with
the wider world and hence preventing agent interaction completely. This is effective if such boxes
are not escaped, but it is unclear if this is possible [27]. Less extreme versions of safety guarantees
are sometimes possible, especially in domains where a formal model of safe behavior is possible,
and the system is sufficiently well understood. For example, Shalev-Shwartz et al. have such a model
for self-driving cars, heavily relying on the fact that the physics involved with keeping cars from
hitting one another, or other objects, is in effect perfectly understood [56]. Expanding this to less well
understood domains seems possible, but is problematic for reasons discussed elsewhere [57].
Without limiting optimization explicitly, some approaches attempt to better define the goals,
and thereby reduce the extent of unanticipated behaviors.
These approaches involve some
version of direct optimization safety. One promising direction for limiting the extent to which
goal-directed optimization can be misdirected is to try to recognize actions rather than goals [58].
Human-in-the-loop oversight is another direction for minimizing surprise and ensuring alignment,
though this is already infeasible in many systems [20]. Neither approach is likely to be more effective
than humans themselves are at preventing such exploitation. The primary forward-looking approach
for safety is some version of ensuring that the goal is aligned, which is the bulk of what Yampolskiy
and Fox refer to as AI safety engineering [59].
11 of 14
In multi-agent contexts there is still a concern that because human values are complex, [18]
exploitation is an intrinsically unavoidable pitfall
Paul Christiano's
"Distillation and Amplification" approach involves safe amplification using coordinated multi-agent
systems [60]. This itself involves addressing some of the challenges with multi-agent approaches,
and work on safe amplification using coordinated multi-agent systems in that context has begun [61].
In that work, the coordinating agents are predictive instead of agentic, so the failure modes are more
restricted. The methods suggested can also be extended to agentic systems, where they may prove
more worrisome, and solving the challenges potentially involves mitigating several failure modes
outlined here.
in multi-agent systems.
Between optimization-limiting approaches and AI safety engineering, it is possible that many of
the multi-agent failures discussed in the paper can be mitigated, though not eliminated. In addition,
there will always be pressure to prioritize performance as opposed to safety, and safe systems are
unlikely to perform as quickly as unsafe ones [20]. Even if the tradeoff resolves in favor of slower,
safer systems, such systems can only be created if these approaches are further explored and the
many challenges involved are solved before widespread deployment of unsafe ML and AI. Once the
systems are deployed, it seems infeasible that safer approaches could stop failures due to exploiting
and exploitable systems, short of recalling them. This is not a concern for the far-off future where
misaligned superintelligent AI poses an existential risk.
It is instead a present problem, and it is
growing more serious along with the growth of research that does not address it.
4. Conclusions: Model Failures and Policy Failures
Work addressing the failure modes outlined in the paper is potentially very valuable, in part
because these failure modes are mitigable or avoidable if anticipated. AI and ML system designers
and users should expect that many currently successful but naive agents will be exploited in the
future. Because of this, the failure modes are likely to become more difficult to address if deferred,
and are therefore particularly critical to understand and address them preemptively. This may take
the form of systemic changes such as redesigned financial market structures, or may involve ensuring
that agents have built-in failsafes, or that they fail gracefully when exploited.
At present, it seems unlikely that large enough and detected failures will be sufficient to slow the
deployment of these systems. It is possible that governmental actors, policymakers, and commercial
entities will recognize the tremendous complexities of multiparty coordination among autonomous
agents and address these failure modes, or slow deployment and work towards addressing these
problems even before they become catastrophic. Alternatively, it is possible these challenges will
become apparent via limited catastrophes that are so blatant that AI safety will be prioritized. This
depends on how critical the failures are, how clearly they can be diagnosed, and whether the public
demands they be addressed.
Even if AI amplification remains wholly infeasible, humanity is already deploying autonomous
systems with little regards to safety. The depth of complexity is significant but limited in current
systems, and the strategic interactions of autonomous systems are therefore even more limited.
However,
just as AI for poker eventually became capable enough to understand multi-player
interaction and engage in strategic play, AI in other systems should expect to be confronted with
these challenges. We do not know when the card sharks will show up, or the extent to which they will
make the games they play unsafe for others, but we should admit now that we are as-yet unprepared
for them.
Funding: This research was funded in large part by a grant from the Berkeley Existential Risk Initiative.
Acknowledgments: I would like to thank a subset of the anonymous reviewers in both the first and second
submission for very helpful comments, and thank Roman Yampolskiy for encouraging me to write and revise
the paper, despite setbacks.
Conflicts of Interest: The author declares no conflict of interest. The funders had no role in the writing of the
manuscript, nor in the decision to publish the results.
Bibliography
1.
Clark, J.; Amodei, D. Faulty Reward Functions in the Wild. 2016. Available online: https://openai.com/
blog/faulty-reward-functions/ (accessed on 12 March 2019).
2.
Goodhart, C.A.E. Problems of Monetary Management: The UK Experience; Papers in Monetary Economics;
Reserve Bank of Australia: Sydney, Australia, 1975.
12 of 14
3. Manheim, D.; Garrabrant, S. Categorizing Variants of Goodhart's Law. arXiv 2018, arXiv:1803.04585.
4.
Eval. Program Plan. 1979, 2, 67 -- 90.
Campbell, D.T. Assessing the impact of planned social change.
[CrossRef]
Amodei, D.; Olah, C.; Steinhardt, J.; Christiano, P.; Schulman, J.; Mané, D. Concrete problems in AI safety.
arXiv 2016, arXiv:1606.06565.
Kleinberg, J.; Raghavan, M. How Do Classifiers Induce Agents To Invest Effort Strategically? arXiv 2018,
arXiv:1807.05307.
Bostrom, N. Superintelligence; Oxford University Press: Oxford, UK, 2017.
Braganza, O. Proxyeconomics, An agent based model of Campbell's law in competitive societal systems.
arXiv 2018, arXiv:1803.00345.
Krakovna, V. Specification Gaming Examples in AI. 2018. Available online: https://vkrakovna.wordpress.
com/2018/04/02/specification-gaming-examples-in-ai/ (accessed on 12 March 2019).
5.
6.
7.
8.
9.
10. Liu, L.; Cheng, L.; Liu, Y.; Jia, Y.; Rosenblum, D.S. Recognizing Complex Activities by a Probabilistic
In Proceedings of the National Conference on Artificial Intelligence (AAAI),
Interval-based Model.
Phoenix, AZ, USA, 12 -- 17 February 2016.
11. Cheney, N.; MacCurdy, R.; Clune, J.; Lipson, H. Unshackling evolution: Evolving soft robots with multiple
12.
materials and a powerful generative encoding. ACM SIGEVOlution 2014, 7, 11 -- 23. [CrossRef]
Figueras, J. Genetic Algorithm Physics Exploiting. 2015. Available online: https://youtu.be/ppf3VqpsryU
(accessed on 12 Mach 2019).
13. Lehman, J.; Clune, J.; Misevic, D.; Adami, C.; Beaulieu, J.; Bentley, P.J.; Bernard, S.; Belson, G.; Bryson, D.M.;
Cheney, N. The surprising creativity of digital evolution: A collection of anecdotes from the evolutionary
computation and artificial life research communities. arXiv 2018, arXiv:1803.03453.
14. Chopra, J. GitHub issue for OpenAI gym environment FetchPush-v0. 2018. Available online: https://
github.com/openai/gym/issues/920 (accessed on 12 Mach 2019).
15. Popov, I.; Heess, N.; Lillicrap, T.; Hafner, R.; Barth-Maron, G.; Vecerik, M.; Lampe, T.; Tassa, Y.; Erez, T.;
arXiv 2017,
Riedmiller, M. Data-efficient deep reinforcement learning for dexterous manipulation.
arXiv:1704.03073.
17.
16. Weimer, W. Advances in Automated Program Repair and a Call to Arms.
In Proceedings of the 5th
International Symposium on Search Based Software Engineering -- Volume 8084; Springer: Berlin/Heidelberg,
Germany, 2013.
Sandberg, A. Friendly Superintelligence. Presentation at Extro 5 Conference. Available online: http:
//www.nada.kth.se/~asa/Extro5/Friendly%20Superintelligence.htm, 2001. (accessed on 12 March 2019).
18. Yudkowsky, E. Complex value systems in friendly AI. In Proceedings of the International Conference on
Artificial General Intelligence, Mountain View, CA, USA, 3 -- 6 August 2011; Springer: New York, NY, USA,
2011; pp. 388 -- 393.
19. Worley, G.G., III. Robustness to fundamental uncertainty in AGI alignment. arXiv 2018, arXiv:1807.09836.
20. Danzig, R. Technology Roulette: Managing Loss of Control as Many Militaries Pursue Technological Superiority;
Technical Report; Center for a New American Security: Washington, DC, USA, 2018.
21. Baum, S. Superintelligence skepticism as a political tool. Information 2018, 9, 209. [CrossRef]
22. Yudkowsky, E. Intelligence explosion microeconomics. Mach. Intell. Res. 2013, 23, 2015.
23. Lewis, G.T. Why the Tails Come Apart Apart. Lesswrong, 2014. Available online: http://lesswrong.com/
lw/km6/whythetailscomeapart/ (accessed on 12 March 2019).
24. Yudkowsky, E. The AI Alignment Problem: Why It's Hard, and Where to Start; Stanford University: Stanford,
CA, USA, 2016.
25. Behzadan, V.; Munir, A. Models and Framework for Adversarial Attacks on Complex Adaptive Systems.
arXiv 2017, arXiv:1709.04137.
26. Drexler, K.E. Engines of Creation; Anchor: New York, NY, USA, 1986.
27. Armstrong, S.; Sandberg, A.; Bostrom, N. Thinking inside the box: Controlling and using an oracle AI.
Minds Mach. 2012, 22, 299 -- 324. [CrossRef]
28. Mulligan, T.S. How Enron Manipulated State's Power Market. Los Angeles Times, 9 May 2002. Available
online: http://articles.latimes.com/2002/may/09/business/fi-scheme9 (accessed on 9 March 2019).
13 of 14
29. Borel, E.; Ville, J. Applications de la théorie des Probabilités aux jeux de Hasard; Gauthier-Villars: Paris,
France, 1938.
30. Kuhn, H.W. A simplified two-person poker. Contrib. Theory Games 1950, 1, 97 -- 103.
31. Bowling, M.; Burch, N.; Johanson, M.; Tammelin, O. Heads-up limit hold'em poker is solved. Science 2015,
347, 145 -- 149. [CrossRef] [PubMed]
32. Brown, N.; Sandholm, T. Superhuman AI for heads-up no-limit poker: Libratus beats top professionals.
Science 2018, 359, 418 -- 424. [CrossRef] [PubMed]
33. Billings, D.; Burch, N.; Davidson, A.; Holte, R.; Schaeffer, J.; Schauenberg, T.; Szafron, D. Approximating
34.
game-theoretic optimal strategies for full-scale poker. IJCAI 2003, 3, 661.
Soares, N. Formalizing Two Problems of Realistic World-Models. Technical Report. Available online:
https://intelligence.org/files/RealisticWorldModels.pdf (accessed on 9 March 2019).
35. Conant, R.C.; Ross Ashby, W. Every good regulator of a system must be a model of that system.
Int. J.
Syst. Sci. 1970, 1, 89 -- 97. [CrossRef]
36. Demski, A.; Garrabrant, S. Embedded Agency. arXiv 2019, arXiv:1902.09469.
37. O'Neil, C. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy; Broadway
Books: New York City, NY, USA, 2016.
38. Eisen, M. Amazon's $23,698,655.93 Book about Flies. 2011. Available online: http://www.michaeleisen.
39.
org/blog/?p=358 (accessed on 9 March 2019).
Smaldino, P.E.; McElreath, R. The natural selection of bad science. Open Sci. 2016, 3, 160384. [CrossRef]
[PubMed]
40. Gibbard, A. Manipulation of Voting Schemes: A General Result. Econometrica 1973, 41, 587 -- 601. [CrossRef]
41. Yudkowsky, E. Inadequate Equilibria: Where and How Civilizations Get Stuck; Machine Intelligence Research
Institute: Berkeley, CA, USA, 2017.
42. Ostrom, E. Governing the Commons: The Evolution of Institutions for Collective Action; Cambridge University
Press: Cambridge, UK, 1990.
43. Nisan, N.; Roughgarden, T.; Tardos, E.; Vazirani, V.V. Algorithmic Game Theory; Cambridge University Press:
Cambridge, UK, 2007.
44. Tramèr, F.; Zhang, F.; Juels, A.; Reiter, M.K.; Ristenpart, T. Stealing Machine Learning Models via Prediction
In Proceedings of the USENIX Security Symposium, Vancouver, BC, Canada, 16 -- 18 August 2016;
APIs.
pp. 601 -- 618.
Shorter, G.W.; Miller, R.S. High-Frequency Trading: Background, Concerns, and Regulatory Developments;
Congressional Research Service: Washington, DC, USA, 2014; Volume 29.
45.
46. Wang, Y.; Chaudhuri, K. Data Poisoning Attacks against Online Learning. arXiv 2018, arXiv:1808.08994.
47. Chen, X.; Liu, C.; Li, B.; Lu, K.; Song, D. Targeted backdoor attacks on deep learning systems using data
poisoning. arXiv 2017, arXiv:1712.05526.
48. Xiao, H.; Xiao, H.; Eckert, C. Adversarial Label Flips Attack on Support Vector Machines. Front. Artif.
Intell. Appl. 2012, 242, doi:10.3233/978-1-61499-098-7-870. [CrossRef]
49. Dixon, H.D. Keeping up with the Joneses: Competition and the evolution of collusion.
J. Econ. Behav.
50.
Organ. 2000, 43, 223 -- 238. [CrossRef]
Sandberg, A. There is plenty of time at the bottom: The economics, risk and ethics of time compression.
Foresight 2018, 21, 84 -- 99. [CrossRef]
51. Leibo, J.Z.; Zambaldi, V.; Lanctot, M.; Marecki, J.; Graepel, T. Multi-agent reinforcement learning in
sequential social dilemmas. In Proceedings of the 16th Conference on Autonomous Agents and MultiAgent
Systems, São Paulo, Brazil, 8 -- 12 May 2017; pp. 464 -- 473.
52. Leibo, J.Z.; Hughes, E.; Lanctot, M.; Graepel, T. Autocurricula and the Emergence of Innovation from Social
Interaction: A Manifesto for Multi-Agent Intelligence Research. arXiv 2019, arXiv:1903.00742.
53. Lowe, R.; Wu, Y.; Tamar, A.; Harb, J.; Abbeel, O.P.; Mordatch, I. Multi-agent actor-critic for mixed
cooperative-competitive environments. Adv. Neural Inf. Process. Syst. 2017, 6379 -- 6390.
54. Russell, S. Comment to Victoria Krakovna, Specification Gaming Examples in AI. 2018. Available online:
https://perma.cc/3U33-W8HN (accessed on 12 March 2019).
55. Taylor, J. Quantilizers: A safer alternative to maximizers for limited optimization.
In Proceedings
of the Workshops at the Thirtieth AAAI Conference on Artificial Intelligence, Phoenix, AZ, USA, 12 -- 17
February 2016.
14 of 14
56.
Shalev-Shwartz, S.; Shammah, S.; Shashua, A. On a formal model of safe and scalable self-driving cars.
arXiv 2017, arXiv:1708.06374.
57. Manheim, D. Oversight of Unsafe Systems via Dynamic Safety Envelopes. arXiv 2018, arXiv:1811.09246.
58. Liu, Y.; Nie, L.; Liu, L.; Rosenblum, D.S. From action to activity: Sensor-based activity recognition.
Neurocomputing 2016, 181, 108 -- 115. [CrossRef]
59. Yampolskiy, R.; Fox, J. Safety engineering for artificial general intelligence.
Topoi 2013, 32, 217 -- 226.
[CrossRef]
60. Christiano, P.; Shlegeris, B.; Amodei, D. Supervising strong learners by amplifying weak experts. arXiv
2018, arXiv:1810.08575.
Irving, G.; Christiano, P.; Amodei, D. AI safety via debate. arXiv 2018, arXiv:1805.00899.
61.
c(cid:13) 2019 by the author; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).
|
1612.04512 | 1 | 1612 | 2016-12-14T07:15:18 | Agent-based Model for Spot and Balancing Electricity Markets | [
"cs.MA",
"q-fin.GN"
] | We present a simple, yet realistic, agent-based model of an electricity market. The proposed model combines the spot and balancing markets with a resolution of one minute, which enables a more accurate depiction of the physical properties of the power grid. As a test, we compare the results obtained from our simulation to data from Nord Pool. | cs.MA | cs | Agent-based Model for Spot and
Balancing Electricity Markets
Florian Kühnlenz, Pedro H. J. Nardelli
Centre for Wireless Communications (CWC)
University of Oulu, Finland
Contact: [Florian.Kuhnlenz,Pedro.Nardelli]@oulu.fi
6
1
0
2
c
e
D
4
1
]
A
M
.
s
c
[
1
v
2
1
5
4
0
.
2
1
6
1
:
v
i
X
r
a
Abstract-We present a simple, yet realistic, agent-
based model of an electricity market. The proposed
model combines the spot and balancing markets with a
resolution of one minute, which enables a more accurate
depiction of the physical properties of the power grid.
As a test, we compare the results obtained from our
simulation to data from Nord Pool.
I. Introduction
In the electricity grid, it is necessary that supply and
demand are in balance all the time, due to physical
constraints [1]. The main part of power generation still
happens with synchronous generators. They can only
slightly deviate from the nominal
frequency for short
amounts of time without damage. To ensure that supply
and demand match, electricity is traded in several stages
with increasing time resolution. The two most important
trading stages are the daily spot market and the balancing
market [2]. All differences that could not be accounted for
in the spot market need to be corrected at the balancing
market in real time.
Due to the still coarse time resolution of most spot
markets (usually one hour) they cannot accurately predict
how suppliers need to run their power plants to match
demand all the time. Therefore, a simulation with high
time resolution is needed to accurately model the behavior
of balancing markets.
Nevertheless, even though balancing markets have be-
come more and more an issue of interest (e.g. [3]–[5]), there
is almost no tools available to get insights in the interplay
between balancing and spot markets. Yet, as demonstrated
in [6], [7] the design of the spot market has great impact
in the balancing market the physical behavior in the grid.
Most openly available simulation cover power flow and
unit commitment models [8], [9], the general energy system
[10], [11] and often include at least an approximation of a
spot market. Even when balancing markets are discussed
they are often analyzed individually [12], [13].
Furthermore in real markets a lot of parties take part
in the trading process already, with the amount most
This work is partly funded by Finnish Academy (n. 271150) and
CNPq/Brazil (n.490235/2012-3) as part of the joint project SUS-
TAIN, by Strategic Research Council/Aka BC-DC project (n.292854)
and by the European Commission through the P2P-SmarTest project
(n.646469).
likely rising in the future. Due to generation becomes more
decentralized with renewable sources and consumption
and production happening in new structural units like
micro-grids [1].
To better understand the interplay between players,
the proposed model is designed as an agent based model
(ABM) [14], in which multiple agents of each type (e.g.
producers or utilities) may co-exist. The trading of elec-
tricity is an integral part of the whole system and shall be-
come even more important with the ongoing introduction
of the smart grid concept. The market price of electricity
is, however, often assumed to be an externality in models
for smart grid communication or physical grid simulations.
In contrast, the flow of information, in form of real time
prices, is assumed to change the way that electricity is
consumed, e.g. in demand response systems. If the con-
sumption is changed by the price, it will inevitably change
the price in-turn since it reflects the point where consump-
tion and production meet. To better understanding these
interactions, one needs to include these intrinsic market
aspects in simulations instead of assuming the price as
something external. Targeting this issue, We introduce
here a simple simulation model of the electricity market
that takes this internal relation into account.
II. Liberalized electricity market structure
The trading of electricity usually takes place in several
stages (e.g [2]) as to be described next.
A. Long-term
In the first one, long term contracts are made between
two parties often covering the time span of multiple years.
These contracts cover the so-called base-load, the very
stable and predictable part of electricity consumption.
B. Spot Market
In the next stage, often called sport market, electricity is
usually traded in a time span of a day with an hourly time
resolution between multiple parties and a market maker.
Since these trades usually end several hours before the
actual delivery hour, there needs to be at least one more
trading stage to cover any changes that happen between
the end of trading and the actual delivery.
Sport markets are typically operated with a single price
for each hour. Each party submits its bids for each hour
of the day. The bids describe either a bid to buy a certain
amount of power or to produce a certain amount of power
for a certain price. The market maker is the only one who
will see all offers and is therefore able to determine which
price will optimize the social welfare of all parties. This
price then determines which producer must produce what
amount of power during any given hour and consequently
how much can be consumed. It is important to remember
that the consumption is not controllable; rather it is based
on forecasts, as is the production of renewable energy
sources like solar and wind.
Since the demand (and production to a growing extend)
that is traded in the spot market stage is based on
forecasts, the importance of the market stages with higher
time resolution and shorter lead times has become more
important over the years [5].
C. Intraday and balancing market
Usually there are two more stages: one before the deliv-
ery hour, often called intraday market, and one final, so
called balancing market stage. The balancing takes place
during the delivery hour and aims to ensure that supply
and demand are matched as close a possible. The balanc-
ing market can be split in two timescales again, one where
power is delivered typically within 15 minutes and one
where power is delivered instantaneous depending on the
physical conditions of the grid for up to 15 minutes. This
last stage happens distributed among the participating
generators and will be compensated for afterwards.
The balancing and intraday markets cannot only be
used to compensate for updated forecasts, but can also
cover an outage of a power plant. The third job of these
market stages is to fine tune the production curve on a
finer timescale.
It is for example possible that the demand for a given
hour is as predicted, yet large amounts of balancing power
are needed [6]. Since the spot market is based on a
hourly time resolution, the participants can only agree
on the power needed during the full hour in aggregate.
If for example most of the power is needed during the
second half of the hour, there will be down-regulation
needed in the first part of the hour and up-regulation in
the second. If the intraday market provides sufficiently
high time resolution this would mean that power would
be sold during the first part of the hour and the same
amount needed to be bought for the second. This intrahour
balancing is a significant cause for balancing [6], [7].
The working of intraday markets can differ between
different market implementations and is not as equal as the
spot market design. However, all intraday markets close
before the delivery hour in contrast to balancing markets.
In balancing markets the bids are submitted ahead of time,
however are called upon during the delivery hour when
needed. While in the balancing and in the spot market
it will be settled before the delivery hour which bids and
offers are matched. Therefore, the price of balancing is
only known after the delivery hour is over and it is known
which up- or down regulation bids were used. Opposite
to the spot market or intraday market prices which are
known before the power is produced and consumed.
III. Proposed model
The presented model of an electricity market aims to
provide the basic concepts of a working electricity markets
with all major participants in a simple form. This serves
two main purposes: 1. it helps to understand the interac-
tion between all parts of the model and second it provides
the freedom to specialize the model for specific purposes
later. As a side effect, it should also speed up execution
time. As a drawback of this approach, the basic form the
model cannot predict specific outcomes but only general
behavior. If any details in the implementation needed to
be very specific it was modeled as close as possible to
the Nord Pool market and therefore to the underlying
matching algorithm Euphemia [15].
The Euphemia algorithm is in fact used across most
European markets, making the model quite universal. The
model is constructed as an agent based model [14] (ABM)
where all market participants are agents. Agent based
models are used in a variety of contexts [16] and in the
context of power grids [17]–[20]. In the context of economic
research, more conservative approaches are, however, the
mostly employed, which has been recently criticized [21].
A big advantage of agent based models in this context
is that they are not confined to equilibrium states and
more accurately capture the complexity of economic re-
alities [22], [23]. More specifically agents make individual
decisions based on their perceived environment and their
internal state. The state of an agent might change during
the simulation based on certain rules so that the decisions
of an agent influence the environment, which in-turn
influences the state of the agent which again influences the
decisions. This makes it possible to capture basic feedback
loops and dynamic behavior.
In our proposed model, we employ the following types
of agents:
‚ Producers: provide power with for a given price per
MWh. Producers have a maximum capacity of how
much power they can deliver in every time step. Every
producer can bid a certain amount of his power into
the balancing market.
‚ Utilities: a utility forecasts and buys the power for
its assigned users for the next day and distributes
balancing costs among the users.
‚ Users: have a certain power demand during the day
which might change due to prices or other internal or
external factors.
A. Producers
The basic producer's agents are simple. Their price and
capacity is fixed. The provide offers for the market and
keep track of their production schedule according to the
market. They also offer a certain amount for balancing
based on globally set percentage or an internal on. There
are currently also two sub-classes of producers for simulat-
ing wind and solar production. Both make an internal fore-
cast about their production that is offered on the market.
The realized output of both is not exactly according to the
market set schedule but can differ up or down, therefore
requiring balancing. Solar production has its peak always
at the same time, while wind production peaks can occur
randomly during the day. Producers keep track of their
income from spot and balancing markets, and balancing
payments when they are producers of renewables.
B. Utilities
The utility agent forecasts the consumption of its as-
signed users based on their prior usage. Therefore, the
utility keeps track of up to 30 days of aggregated usage
data and calculates a weighted average of the data to
forecast the next day. The forecast is also multiplied with
an error. The error is modeled as a random walk with mean
return. The utility keeps track of revenue from its users
and costs for buying at the spot and balancing markets.
Since the balancing costs cannot be attributed to specific
users they are shared by all users. Therefore, the fix costs
of the utility are calculated as:
prevenue ´ cost ´ balancing_costq{number_of_users
C. Users
The user agent mainly generates a load curve currently
based on the "sin" function. It has a fixed minimum
and maximum value however the phase of the sine curve
changes during the simulation. The maximum of the sine
curve is occurring at around 6 p.m. in the evening. A
comparison of a sine curve with the actual consumption
pattern for a winter day can be seen in Fig. 1. There are
two types of users, optimizing users and normal users. For
normal users, the curve might shift randomly by up to 15
minutes in any direction. The optimizing user can shift his
sine curve freely. This allows him to use the price as an
input to optimize his daily usage.
D. Simulation Periods
Every simulation day occurs in three main stages de-
picted in Figs. 2, 3 and 4. During the spot market period,
all producers submit their offers and utilities their bids
according to their forecasts. In the current implementation
utilities have no price flexibility so their bids always
need to be matched, which in the real market would
be represented by the bid having the maximal allowed
price. During the balancing period, all balancing offers are
collected and then called upon if needed, starting with the
cheapest ones.
Fig. 1. Comparison of Finnish consumption during a winter day
with a sine curve.
Fig. 2.
Spot market period. First time step: Utilities forecast
the demand of their users and submit bids accordingly; producers
forecast production capacity and submit offers accordingly. Second
time step: market matches bids and offers to create schedules and a
public price.
The differences between production and consumption
are then compared for every given cycle, typically 15
minutes. If the difference exceeds a certain amount of
power, an up- or down- regulation offer is called upon.
The power plant is then removed from the list of both
up and down regulation, to minimize the fast changes in
output, which are typically though on the equipment.
The adjustments are valid for the full rest of the hour.
For example, if a power plant must adjust its output down
during the first 15 minutes of an hour, it will remain
at that output till the next full hour starts. In the final
period, the balancing costs for every period are calculated
according to the pricing scheme that is used in Nord
Pool [2], [24]. Additionally, all agents account for their
consumed or produced power.
05101520time (h)0.800.850.900.951.00consumption (%)consumptionsine curveProducersMarketUtilitiesUsersOffersScheduleScheduleForecastDemandForecastCapacityBidsPriceHist.Data Fig. 3. Balancing period. Producers submit balancing offers, which
the market calls upon depending on the mismatch between real usage
and production.
Fig. 5. The price and production capacity of the simulated power
plants.
changes therefore the results are expected to be represen-
tative. In the presented simulations, there was a total of
100.000 users with none of them being optimizing users.
All users where spread among 6 utilities, while production
was provided by 11 producers.
Fig. 5 shows the power plant configuration used in the
simulated system. All bigger power plants have a very
low regulation factor and only provide small amounts of
power for balancing. This captures the situation that most
thermal power plants, which provide base load, are not
very flexible.
In Fig. 6 we see the amount of balancing in relation to
the consumption for one day of the simulation compared
to a chosen day from the Nord Pool data. The days are
specifically picked to be very comparable and highlight the
phenomenon of intrahour regulation. For the Nord Pool
plot, both automatic and market based regulation where
considered, as there is no difference between these in the
simulation.
Overall the simulation provides results close to data
from Nord Pool, considering the simplifications e.g. no in-
traday market and only sinusoidal load curves. This shows
the model does not only provide all basic functionality but
also produces data which is comparable to real markets.
Comparison between Nord Pool and the proposed model
TABLE I
Nord Pool (2015)
Simulation (30 days)
171% / 60%
186% / 64%
avg. price
?
σ2
avg. regulation
max. regulation
intra-hour regul.
balancing price
21.00€
7.92€
1.59%
7.14%
13.09h
22.86€
10.62€
1.10%
5.02%
11.68h
Fig. 4. After market period. All players update their costs and
revenues. The market collects balancing payments and pays for called
offers. Balancing costs of utilities are passed to the user as fixed costs.
IV. Comparison
To verify the results of the simulation, we provide a
comparison with data obtained from the Nord Pool market
[25]. It is important to state that it is not the goal of the
simulation to model a specific market, like Nord Pool in
great detail, but rather to capture the general behavior.
Hence, the simulation was not calibrated with Nord Pool
data but rather setup to simulate a much smaller system
with comparable prices. However, the results show very
comparable outcomes at the balancing market, specially
concerning the intrahour balancing that could only be
captured due to the high time resolution of the simulation.
In Table I the comparison between the data from Nord
Pool and a 30-day run can be seen. Some of the bigger
differences might be due to the much shorter run time of
the simulation of only 30 days instead of a whole year.
However, the simulation does not undergo any seasonal
ProducersMarketUtilitiesUsersOffersReal UsageReal ProductionScheduleCalculateReservesProducersMarketUtilitiesUsersCalculateCost/RevenueCalculateCost/RevenueCalculateCostBalancingCostBalancingCostFixCost020040060080010001200power (MW)020406080100120price () [12] G. Santos, T. Pinto, Z. Vale, and H. Morais, "Balancing market
integration in MASCEM electricity market simulator," 2012
IEEE Power and . . . , pp. 1–8, 2012.
[13] M. Mureddu, G. Caldarelli, A. Chessa, A. Scala, and A. Dami-
ano, "Green Power Grids: How Energy from Renewable Sources
Affects Networks and Markets," PLoS ONE, vol. 10, no. 9, pp.
e0 135 312–15, Sep. 2015.
[14] D. Helbing, S. Balietti, T. A. Kohler, W. Frej, and B. Ra-
malingam, "How to do agent-based simulations in the future:
from modeling social mechanisms to emergent phenome-na and
interactive systems design," 2011.
[15] P. PXs, "EUPHEMIA Public Description. PCR Market Cou-
pling Algorithm," Tech. Rep., Jan. 2016.
[16] C. M. Macal, "Everything you need to know about agent-based
modelling and simulation," vol. 10, no. 2, pp. 144–156, Apr.
2016.
[17] E. A. Kremers, "Modelling and simulation of electrical energy
systems through a complex systems approach using agent-based
models," 2013.
[18] E. Kremers, P. Viejo, and O. Barambones, "A complex systems
modelling approach for decentralised simulation of electrical
microgrids," . . . (ICECCS), 2010.
[19] E. Shafiei, H. Thorkelsson, E. I. Ásgeirsson, B. Davidsdottir,
M. Raberto, and H. Stefansson, "An agent-based modeling
approach to predict the evolution of market share of electric
vehicles: A case study from Iceland," Technological Forecasting
& Social Change, vol. 79, no. 9, pp. 1638–1653, Nov. 2012.
[20] F. Kühnlenz and P. H. J. Nardelli, "Dynamics of Complex Sys-
tems Built as Coupled Physical, Communication and Decision
Layers," PLoS ONE, vol. 11, no. 1, pp. e0 145 135–19, Jan. 2016.
[21] J. D. Farmer and D. Foley, "The economy needs agent-based
modelling," Nature, vol. 460, no. 7256, pp. 685–686, 2009.
[22] K. H. van Dam, I. Nikolic, and Z. Lukszo, Agent-based modelling
of socio-technical systems. Springer Science & Business Media,
2012, vol. 9.
[23] M. Cristelli, A. Tacchella, and L. Pietronero, "The Heteroge-
neous Dynamics of Economic Complexity," PLoS ONE, vol. 10,
no. 2, pp. e0 117 174–15, Feb. 2015.
[Online].
http://www.fingrid.fi/en/customers/
Balance%20services/management/powermarket/Pages/
default.aspx
[Online]. Available: http://www.nordpoolspot.com/download-
center/
Available:
[24]
[25]
Fig. 6. Left: balancing needed for a specific day in the simulation.
Right: comparable day from Nord Pool data.
V. Conclusions and Outlook
The presented model of a basic electricity market includ-
ing spot and balancing markets. This provides the basis for
future research on the interaction between different mar-
kets, players and possible other systems like the physical
grid or communication networks. The high time resolution
of the model opens further possibilities for interaction with
physical phenomena in the grid and the optimization of
market operation times.
For example, the effect on the balancing or the in-
tegration of renewables of shortening the spot market
interval to 30 or 15 minutes can be readily tested. We
derived the model from real markets and close to the
implementation of Nord Pool and could show that the out-
put is consequently comparable. Yet, the model remains
very flexible so it can simulate systems that are spanning
several countries or just a micro-grid.
References
[2] Nord Pool
[1] P. H. Nardelli, N. Rubido, C. Wang, M. S. Baptista,
C. Pomalaza-Raez, P. Cardieri, and M. Latva-aho, "Models for
the modern power grid," The European Physical Journal Special
Topics, vol. 223, no. 12, pp. 2423–2437, 2014.
"The Nordic Electricity Exchange
a Liberalized Electricity
and The Nordic Model
Market,"
Rep.,
Available:
http://nordpoolspot.com/globalassets/download-center/rules-
and-regulations/the-nordic-electricity-exchange-and-the-
nordic-model-for-a-liberalized-electricity-market.pdf
[3] Fingrid, "Electricity market needs fixing – What can we do?"
for
2014.
[Online].
Tech.
Spot,
Tech. Rep., May 2016.
[4] L. Hirth and I. Ziegenhagen, "Balancing Power and Variable
Renewables: A Glimpse at German Data ," pp. 1–23, Dec. 2013.
[5] F. Ocker and K. M. Ehrhart, "The "German paradox" in the
balancing power markets," 2015.
[6] T. Weissbach and E. Welfonder, "High frequency deviations
within the European power system: Origins and proposals for
improvement," Power Systems Conference and . . . , pp. 1–6,
2009.
[7] Z. W. Li and O. Samuelsson, "Frequency deviations and gen-
eration scheduling in the nordic system," PowerTech, pp. 1–6,
2011.
[8] T. Brown, J. Hörsch, and D. Schlachtberger, "PyPSA." [Online].
Available: https://github.com/FRESNA/PyPSA
[9] A. Greenhall, "Minpower: A Power Systems Optimization
Toolkit," pp. 1–6, Feb. 2012.
[10] L. Hirth, "The market value of variable renewables," Energy
[11] D. Atabay,
[Online]. Available: https://ficus.
Economics, vol. 38, pp. 218–236, Jul. 2013.
"PyPSA."
readthedocs.io/en/latest/
05101520hour1.51.00.50.00.51.01.5correction (%)day 2305101520hour10.03.15 |
1805.09081 | 3 | 1805 | 2019-10-21T15:39:21 | Local Tomography of Large Networks under the Low-Observability Regime | [
"cs.MA",
"cs.IT",
"cs.IT"
] | This article studies the problem of reconstructing the topology of a network of interacting agents via observations of the state-evolution of the agents. We focus on the large-scale network setting with the additional constraint of $partial$ observations, where only a small fraction of the agents can be feasibly observed. The goal is to infer the underlying subnetwork of interactions and we refer to this problem as $local$ $tomography$. In order to study the large-scale setting, we adopt a proper stochastic formulation where the unobserved part of the network is modeled as an Erd\"{o}s-R\'enyi random graph, while the observable subnetwork is left arbitrary. The main result of this work is establishing that, under this setting, local tomography is actually possible with high probability, provided that certain conditions on the network model are met (such as stability and symmetry of the network combination matrix). Remarkably, such conclusion is established under the $low$-$observability$ $regime$, where the cardinality of the observable subnetwork is fixed, while the size of the overall network scales to infinity. | cs.MA | cs | Local Tomography of Large Networks
under the Low-Observability Regime
Augusto Santos∗, Vincenzo Matta†, and Ali H. Sayed(cid:63)
1
9
1
0
2
t
c
O
1
2
]
A
M
.
s
c
[
3
v
1
8
0
9
0
.
5
0
8
1
:
v
i
X
r
a
Abstract -- This article studies the problem of reconstructing
the topology of a network of interacting agents via observations of
the state-evolution of the agents. We focus on the large-scale net-
work setting with the additional constraint of partial observations,
where only a small fraction of the agents can be feasibly observed.
The goal is to infer the underlying subnetwork of interactions and
we refer to this problem as local tomography. In order to study
the large-scale setting, we adopt a proper stochastic formulation
where the unobserved part of the network is modeled as an
Erdos-R´enyi random graph, while the observable subnetwork is
left arbitrary. The main result of this work is to establish that,
under this setting, local tomography is actually possible with
high probability, provided that certain conditions on the network
model are met (such as stability and symmetry of the network
combination matrix). Remarkably, such conclusion is established
under the low-observability regime, where the cardinality of the
observable subnetwork is fixed, while the size of the overall
network scales to infinity.
Index Terms -- Topology inference, network tomography, graph
learning, low-observability, local tomography, large-scale networks,
Erdos-R´enyi model, random graphs, diffusion networks.
I. INTRODUCTION
In networked dynamical systems [1] -- [4] the state of the
agents comprising the network evolves over time and is
affected by peer-to-peer interactions. In general, information
about the profile of interactions is unavailable. It is the goal
of network tomography to infer network connectivity from
observing the evolution of the graph nodes. Problems of
this type arise in many domains where knowledge of the
underlying topology linking the agents is critical for better
inference and control mechanisms. For example, in distributed
processing over networks, the underlying topology is critical
for the performance delivered by the distributed strategies such
as consensus [5] -- [16] and diffusion [17] -- [22]; in the context
of epidemics, it is well-known that the network topology may
foster or hinder the outbreak of diseases or opinions [23];
in the context of brain functionality, it is also known that
the connectivity among brain regions impacts the efficiency
Fig. 1. Illustration of the local tomography problem. The goal is to design a
mechanism to recover the underlying subnetwork topology by appropriately
processing the observables, i.e., the state-evolution of the observable nodes.
is important
and robustness of the brain dynamics [24] and can help
explain brain functional disorders [25], [26]; in cyber-security
applications it
to determine and understand
the underlying network structure to devise effective counter-
measures [27]; and tomography is also a relevant problem in
economics [28] (in the context of causal inference) and physics
applications [29]. Depending on the particular application, an
appropriate model for the underlying networked dynamical
system must be selected. In this work, we focus on a linear
stochastic dynamical system that will be detailed in Sec. I-A
-- see also Sec. I-C for an example.
This article focuses on the large-scale network setting,
where one can typically observe and/or process limited por-
tions of the network. More formally, we address a local
tomography problem: a subset of the agents is observed
and their subnetwork of interactions is inferred from these
observations. Figure 1 depicts the local tomography paradigm.
There are three main reasons that cause this observability
limitation in the large-scale network setting:
• Accessibility-limit. Some portions of the network are not
accessible and, hence, unobservable. Moreover, in many large-
scale settings the existence of some sources of interactions
(i.e., unobserved network links) might be unknown.
Manuscript received May 23, 2018; revised September 9, 2019; accepted
September 19, 2019. A. H. Sayed was supported in part by the NSF under
Grant CCF-1524250 and Grant ECCS-1407712.
∗ A. Santos was with the Adaptive Systems Laboratory, EPFL,
[email protected],
Switzerland
(email:
CH-1015
[email protected]).
Lausanne,
(email: [email protected]).
† V. Matta is with DIEM, University of Salerno, 84084 Fisciano, Italy
(cid:63) A. H. Sayed is with the ´Ecole Polytechnique F´ed´erale de Lausanne
(EPFL), CH-1015 Lausanne, Switzerland (email: [email protected]).
Communicated by I. Kontoyiannis, Associate Editor At Large.
Color versions of one or more of the figures in this article are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIT.2019.2945033
ities can be smaller than the scale of the network.
ther constrains the size of the data that can be processed.
• Probing-limit. The acquisition of data and storage capac-
• Processing-limit. The complexity of the data-mining fur-
For instance, one may probe the activity of a subset of
nodes -- as it is unfeasible to track the activity of all the
nodes in a large-scale network -- in order to reconstruct
its underlying profile of interactions. This requires that we
partially observe the system and extract information about its
underlying subnetwork of interactions.
Under the aforementioned local tomography setting, the
ScannerLocal TomographyScannerScannerScanner? problem of inferring the subnetwork topology across the
observed agents becomes exceedingly challenging or even ill-
posed. It is therefore important to devise nontrivial condi-
tions (if any) under which the problem is still well-posed,
i.e., the information about the topology can be effectively
inferred from the observable samples. In this article, we
show that under an appropriate setting, the problem of local
tomography becomes well-posed with high probability in the
thermodynamic limit: when the number of interacting agents
N grows, the (fixed) subnetwork topology associated with the
observed agents can be perfectly recovered. We refer to such
framework as "low-observability" to emphasize that we are
interested in studying the local tomography problem in the
thermodynamic limit of large networks while the observed
part is fixed and finite. Besides ascertaining conditions under
which the problem is well-posed with high probability, we
further derive a procedure on the space of observables to
recover the subnetwork topology. Finally, as an application of
these results, we devise a strategy that shows how to learn the
topology sequentially, by partitioning the observable network
into small patches, and launching successive instances of the
local tomography algorithm on these patches.
A. Preview of the Main Result
Network tomography is associated with retrieving the un-
derlying network structure of a distributed dynamical system
via observation of the output measurements of the constituent
elements. The typical formulation of the network tomography
problem involves two main objects: i) the statistical model that
governs the laws of evolution of the (stochastic) dynamical
system of interest; ii) and a set of observables. In this article,
we consider a stochastic dynamical system described by a
first-order Vector Auto-Regressive (VAR) or diffusion model,
which is a popular theoretical model arising across many
disciplines. Under this model, N entities corresponding to the
network agents interact over time n according to the following
law:
(1)
Here, A is a stable N × N matrix with nonnegative entries,
and
yn = A yn−1 + βxn
yn = [y1(n), y2(n), . . . , yN (n)](cid:62),
xn = [x1(n), x2(n), . . . , xN (n)](cid:62),
(2)
(3)
with the vector yn collecting the state (or output measure-
ments) at time n of the N agents comprising the network;
and xn representing a random input (e.g., a source of noise
time n. The ensemble {xi(n)} are
or streaming data) at
independent and identically distributed (i.i.d.) both spatially
(i.e., w.r.t. to index i) and temporally (i.e., w.r.t. to index
n). Without loss of generality, we assume that the random
variables xi(n) have zero mean and unit variance.
The support-graph of A reflects the underlying connections
among the agents. Indeed, we have from (1) that:
yi(n) =
N(cid:88)(cid:96)=1
ai(cid:96) y(cid:96)(n − 1) + βxi(n),
(4)
2
Fig. 2. Uppermost panel. Illustration of the local tomography problem in
connection with (4). The observable set is S = {i, j, (cid:96), k}. The measurements
delivered by these observable nodes are collected by a tomography center.
Lowermost panel. How the combination weights determine the profile of
interaction (i.e., the topology).
which shows that, in order to update its output at time n,
agent i combines the outputs of other agents from time n− 1.
In particular, agent i scales the output of agent (cid:96) by using
a combination weight ai(cid:96). Note that the output of agent (cid:96) is
employed by agent i only if ai(cid:96) (cid:54)= 0. The uppermost panel
in Figure 2 offers a pictorial view of the local tomography
problem, whereas the lowermost panel illustrates the role of
the combination weights in determining the mutual influences
between nodes.
Model (4) further shows how the observation yi(n) is
affected by the source value xi(n), which is available locally
at agent i at time n.
The problem of support-graph recovery addressed in this
work is generally referred to as network tomography in the
literature because only indirect observations are available. In
our framework, only output measurements from a subset of
the nodes are accessible, and no information is available about
the unobserved nodes including their number or connectivity.
We refer to this paradigm as local tomography. Under this
challenging framework, the goal is identifiability of the topol-
ogy linking the observable agents. That is, we consider the
problem of inferring the topology associated with a subset S of
observable interacting agents, by measuring only the outputs
produced by such agents.
Let us ignore for a while the restriction of partial observ-
ability. It is tempting (and actually not that uncommon in the
literature) to estimate the connections between the network
agents by measuring the correlation between their output
measurements. There is, however, one critical issue related
to the use of the correlation measure for topology inference,
arising from the streaming nature of the data. In general, when
an external observer starts collecting output measurements,
the network would have been in operation since some time
already. Therefore, after a transient phase, over a connected
network abiding by a stable linear stochastic dynamical system
as described in (1), all agent pairs will become correlated. In
order to illustrate this point in greater detail, let us introduce
the correlation matrix at time n, namely, R0(n) (cid:44) E[ynyT
n ].
When A is symmetric (which is the case considered in this
Tomography Center 1LargeScaleConsistentTomographyOverPartiallyObservedDiffusionNetworksVincenzoMattaAbstract -- IndexTerms -- Notation.Weuseboldfaceletterstodenoterandomvariables,andnormalfontlettersfortheirrealizations.Capitallettersrefertomatrices,smallletterstobothvectorsandscalars.Sometimesweviolatethislatterconvention,forinstance,wedenotethetotalnumberofnetworkagentsbyN.ThesymbolsPandEareusedtodenotetheprobabilityandexpectationoperators,respectively.I.THEPROBLEMEquationsforfigureyi(1),...,yi(n)(1)yj(1),...,yj(n)(2)yℓ(1),...,yℓ(n)(3)yk(1),...,yk(n)(4)A.TheAdaptiveDiffusionNetworkAnetworkofNagentscollectsstreamingdatafromtheenvironment.Thedatumcollectedbythei-thagentattimenisdenotedbyxi(n),andtheglobalsequenceofdataisassumedtobeformedbyspatially(i.e.,w.r.t.indexi)andtemporally(i.e.,w.r.t.indexn)indepedentandidenticallydistributedrandomvariables,withvarianceσ2.Withoutlossofgenerality,weassumethatthevariableshavezeromean.Inordertotrackdriftsinthephenomenonofinterestthatthenetworkismonitoring,theagentsimplementanadaptivealgorithm.Indoingso,owingtothedistributednatureoftheproblem,eachindividualagentshouldrelyonlocalcooperationwithitsownneighbors,namely,theagentsshouldimplementanadaptivedistributedalgorithm,whichwearenowgoingtoillustrate.vi(n−1)=N!ℓ=1[A]iℓyℓ(n−1),(5)yi(n)=vi(n−1)+µ[xi(n)−vi(n−1)].(6)yi(n)=(1−µ)N!ℓ=1[A]iℓyi(n−1)+µxi(n).(7)A=(1−µ)A.(8)V.MattaiswithDIEM,UniversityofSalerno,viaGiovanniPaoloII,I-84084,Fisciano(SA),Italy.E-mail:{vmatta}@unisa.it.Considerasymmetricand(scaled)doubly-stochasticmatrixA:aij=aji≥0,N!ℓ=1aiℓ=N!ℓ=1aℓj=1−µ.(9)Throughoutthemanuscript,weassumethatthenetworkisstronglyconnected,whichimpliesthatmatrixAisirreducible,andthat:An(1−µ)nn→∞−→TN.(10)Ni={ℓ∈{1,2,...,N}:aiℓ>0}.(11)yn=Ayn−1+µxn(12)yn=µn!i=1An−ixi.(13)B.TheTomographyProblemAnentityexternaltotheagentsnetwork(hereafternamedTomographyCenter,TC)isinterestedinreconstructingtheconnectivityprofileofthenetwork,namely,isinterestedinascertainingwhichagentisconnectedtowhichotheragent.TheTCisassumedtohaveaccesstoasubsetofthenetworkagents,andisabletocollectthestreamsofoutputsexchangedbytheagentsduringtheircommunicationprotocol.Formally,ifO⊂{1,2,...,N}istheaccessiblesubset,thedataavailableattimenare{yi(1),yi(2),...,yi(n)}i∈O.TheTCdoesnotknowtheoverallnumberofagentsinthenetwork.Accordingly,theonlygoaloftheTCisproducinganestimateoftheadjacencymatrixoftheobservedagents.Withoutlosinggenerality(andsincethechoiceofaparticularnumberingisimmaterialtoourpurposes)weshallassumethatagents1,2,...,Kareobserved,whiletheremainingM=N−Kagentsarenot.Moreover,inthisarticlewefocusontheasymptoticregimeKN→η<1(14)whereηtakesonthemeaningoftheasymptoticallyachievablefractionofobservedagents.Thisproblemischallengingforthefollowingreasons.•Letusignoreforamomentthefactthatthenetworkispartiallyobserved,andassumethattheTCiscollectingthewholesequencesofoutputsofthediffusionoutputforalltimeinstantsandforallagents.Inprinciple,thereexistseveralwellassessedstrategiestomakeinferenceabouttheinfluencethatoneagenthasonanotheragentstartingfromtheirdata.Themostintuitiveisclearlyanestimateofthecorrelationbetweentheoutputsof1LargeScaleConsistentTomographyOverPartiallyObservedDiffusionNetworksVincenzoMattaAbstract -- IndexTerms -- Notation.Weuseboldfaceletterstodenoterandomvariables,andnormalfontlettersfortheirrealizations.Capitallettersrefertomatrices,smallletterstobothvectorsandscalars.Sometimesweviolatethislatterconvention,forinstance,wedenotethetotalnumberofnetworkagentsbyN.ThesymbolsPandEareusedtodenotetheprobabilityandexpectationoperators,respectively.I.THEPROBLEMEquationsforfigureyi(1),...,yi(n)(1)yj(1),...,yj(n)(2)yℓ(1),...,yℓ(n)(3)yk(1),...,yk(n)(4)A.TheAdaptiveDiffusionNetworkAnetworkofNagentscollectsstreamingdatafromtheenvironment.Thedatumcollectedbythei-thagentattimenisdenotedbyxi(n),andtheglobalsequenceofdataisassumedtobeformedbyspatially(i.e.,w.r.t.indexi)andtemporally(i.e.,w.r.t.indexn)indepedentandidenticallydistributedrandomvariables,withvarianceσ2.Withoutlossofgenerality,weassumethatthevariableshavezeromean.Inordertotrackdriftsinthephenomenonofinterestthatthenetworkismonitoring,theagentsimplementanadaptivealgorithm.Indoingso,owingtothedistributednatureoftheproblem,eachindividualagentshouldrelyonlocalcooperationwithitsownneighbors,namely,theagentsshouldimplementanadaptivedistributedalgorithm,whichwearenowgoingtoillustrate.vi(n−1)=N!ℓ=1[A]iℓyℓ(n−1),(5)yi(n)=vi(n−1)+µ[xi(n)−vi(n−1)].(6)yi(n)=(1−µ)N!ℓ=1[A]iℓyi(n−1)+µxi(n).(7)A=(1−µ)A.(8)V.MattaiswithDIEM,UniversityofSalerno,viaGiovanniPaoloII,I-84084,Fisciano(SA),Italy.E-mail:{vmatta}@unisa.it.Considerasymmetricand(scaled)doubly-stochasticmatrixA:aij=aji≥0,N!ℓ=1aiℓ=N!ℓ=1aℓj=1−µ.(9)Throughoutthemanuscript,weassumethatthenetworkisstronglyconnected,whichimpliesthatmatrixAisirreducible,andthat:An(1−µ)nn→∞−→TN.(10)Ni={ℓ∈{1,2,...,N}:aiℓ>0}.(11)yn=Ayn−1+µxn(12)yn=µn!i=1An−ixi.(13)B.TheTomographyProblemAnentityexternaltotheagentsnetwork(hereafternamedTomographyCenter,TC)isinterestedinreconstructingtheconnectivityprofileofthenetwork,namely,isinterestedinascertainingwhichagentisconnectedtowhichotheragent.TheTCisassumedtohaveaccesstoasubsetofthenetworkagents,andisabletocollectthestreamsofoutputsexchangedbytheagentsduringtheircommunicationprotocol.Formally,ifO⊂{1,2,...,N}istheaccessiblesubset,thedataavailableattimenare{yi(1),yi(2),...,yi(n)}i∈O.TheTCdoesnotknowtheoverallnumberofagentsinthenetwork.Accordingly,theonlygoaloftheTCisproducinganestimateoftheadjacencymatrixoftheobservedagents.Withoutlosinggenerality(andsincethechoiceofaparticularnumberingisimmaterialtoourpurposes)weshallassumethatagents1,2,...,Kareobserved,whiletheremainingM=N−Kagentsarenot.Moreover,inthisarticlewefocusontheasymptoticregimeKN→η<1(14)whereηtakesonthemeaningoftheasymptoticallyachievablefractionofobservedagents.Thisproblemischallengingforthefollowingreasons.•Letusignoreforamomentthefactthatthenetworkispartiallyobserved,andassumethattheTCiscollectingthewholesequencesofoutputsofthediffusionoutputforalltimeinstantsandforallagents.Inprinciple,thereexistseveralwellassessedstrategiestomakeinferenceabouttheinfluencethatoneagenthasonanotheragentstartingfromtheirdata.Themostintuitiveisclearlyanestimateofthecorrelationbetweentheoutputsofi ℓj k 1LargeScaleConsistentTomographyOverPartiallyObservedDiffusionNetworksVincenzoMattaAbstract -- IndexTerms -- Notation.Weuseboldfaceletterstodenoterandomvariables,andnormalfontlettersfortheirrealizations.Capitallettersrefertomatrices,smallletterstobothvectorsandscalars.Sometimesweviolatethislatterconvention,forinstance,wedenotethetotalnumberofnetworkagentsbyN.ThesymbolsPandEareusedtodenotetheprobabilityandexpectationoperators,respectively.I.THEPROBLEMEquationsforfigureyi(1),...,yi(n)(1)yj(1),...,yj(n)(2)yℓ(1),...,yℓ(n)(3)yk(1),...,yk(n)(4)A.TheAdaptiveDiffusionNetworkAnetworkofNagentscollectsstreamingdatafromtheenvironment.Thedatumcollectedbythei-thagentattimenisdenotedbyxi(n),andtheglobalsequenceofdataisassumedtobeformedbyspatially(i.e.,w.r.t.indexi)andtemporally(i.e.,w.r.t.indexn)indepedentandidenticallydistributedrandomvariables,withvarianceσ2.Withoutlossofgenerality,weassumethatthevariableshavezeromean.Inordertotrackdriftsinthephenomenonofinterestthatthenetworkismonitoring,theagentsimplementanadaptivealgorithm.Indoingso,owingtothedistributednatureoftheproblem,eachindividualagentshouldrelyonlocalcooperationwithitsownneighbors,namely,theagentsshouldimplementanadaptivedistributedalgorithm,whichwearenowgoingtoillustrate.vi(n−1)=N!ℓ=1[A]iℓyℓ(n−1),(5)yi(n)=vi(n−1)+µ[xi(n)−vi(n−1)].(6)yi(n)=(1−µ)N!ℓ=1[A]iℓyi(n−1)+µxi(n).(7)A=(1−µ)A.(8)V.MattaiswithDIEM,UniversityofSalerno,viaGiovanniPaoloII,I-84084,Fisciano(SA),Italy.E-mail:{vmatta}@unisa.it.Considerasymmetricand(scaled)doubly-stochasticmatrixA:aij=aji≥0,N!ℓ=1aiℓ=N!ℓ=1aℓj=1−µ.(9)Throughoutthemanuscript,weassumethatthenetworkisstronglyconnected,whichimpliesthatmatrixAisirreducible,andthat:An(1−µ)nn→∞−→TN.(10)Ni={ℓ∈{1,2,...,N}:aiℓ>0}.(11)yn=Ayn−1+µxn(12)yn=µn!i=1An−ixi.(13)B.TheTomographyProblemAnentityexternaltotheagentsnetwork(hereafternamedTomographyCenter,TC)isinterestedinreconstructingtheconnectivityprofileofthenetwork,namely,isinterestedinascertainingwhichagentisconnectedtowhichotheragent.TheTCisassumedtohaveaccesstoasubsetofthenetworkagents,andisabletocollectthestreamsofoutputsexchangedbytheagentsduringtheircommunicationprotocol.Formally,ifO⊂{1,2,...,N}istheaccessiblesubset,thedataavailableattimenare{yi(1),yi(2),...,yi(n)}i∈O.TheTCdoesnotknowtheoverallnumberofagentsinthenetwork.Accordingly,theonlygoaloftheTCisproducinganestimateoftheadjacencymatrixoftheobservedagents.Withoutlosinggenerality(andsincethechoiceofaparticularnumberingisimmaterialtoourpurposes)weshallassumethatagents1,2,...,Kareobserved,whiletheremainingM=N−Kagentsarenot.Moreover,inthisarticlewefocusontheasymptoticregimeKN→η<1(14)whereηtakesonthemeaningoftheasymptoticallyachievablefractionofobservedagents.Thisproblemischallengingforthefollowingreasons.•Letusignoreforamomentthefactthatthenetworkispartiallyobserved,andassumethattheTCiscollectingthewholesequencesofoutputsofthediffusionoutputforalltimeinstantsandforallagents.Inprinciple,thereexistseveralwellassessedstrategiestomakeinferenceabouttheinfluencethatoneagenthasonanotheragentstartingfromtheirdata.Themostintuitiveisclearlyanestimateofthecorrelationbetweentheoutputsofobservable subnet output measurements from the observable set 1LargeScaleConsistentTomographyOverPartiallyObservedDiffusionNetworksVincenzoMattaAbstract -- IndexTerms -- Notation.Weuseboldfaceletterstodenoterandomvariables,andnormalfontlettersfortheirrealizations.Capitallettersrefertomatrices,smallletterstobothvectorsandscalars.Sometimesweviolatethislatterconvention,forinstance,wedenotethetotalnumberofnetworkagentsbyN.ThesymbolsPandEareusedtodenotetheprobabilityandexpectationoperators,respectively.I.THEPROBLEMEquationsforfigureyi(1),...,yi(n)(1)yj(1),...,yj(n)(2)yℓ(1),...,yℓ(n)(3)yk(1),...,yk(n)(4)A.TheAdaptiveDiffusionNetworkAnetworkofNagentscollectsstreamingdatafromtheenvironment.Thedatumcollectedbythei-thagentattimenisdenotedbyxi(n),andtheglobalsequenceofdataisassumedtobeformedbyspatially(i.e.,w.r.t.indexi)andtemporally(i.e.,w.r.t.indexn)indepedentandidenticallydistributedrandomvariables,withvarianceσ2.Withoutlossofgenerality,weassumethatthevariableshavezeromean.Inordertotrackdriftsinthephenomenonofinterestthatthenetworkismonitoring,theagentsimplementanadaptivealgorithm.Indoingso,owingtothedistributednatureoftheproblem,eachindividualagentshouldrelyonlocalcooperationwithitsownneighbors,namely,theagentsshouldimplementanadaptivedistributedalgorithm,whichwearenowgoingtoillustrate.vi(n−1)=N!ℓ=1[A]iℓyℓ(n−1),(5)yi(n)=vi(n−1)+µ[xi(n)−vi(n−1)].(6)yi(n)=(1−µ)N!ℓ=1[A]iℓyi(n−1)+µxi(n).(7)A=(1−µ)A.(8)V.MattaiswithDIEM,UniversityofSalerno,viaGiovanniPaoloII,I-84084,Fisciano(SA),Italy.E-mail:{vmatta}@unisa.it.Considerasymmetricand(scaled)doubly-stochasticmatrixA:aij=aji≥0,N!ℓ=1aiℓ=N!ℓ=1aℓj=1−µ.(9)Throughoutthemanuscript,weassumethatthenetworkisstronglyconnected,whichimpliesthatmatrixAisirreducible,andthat:An(1−µ)nn→∞−→TN.(10)Ni={ℓ∈{1,2,...,N}:aiℓ>0}.(11)yn=Ayn−1+µxn(12)yn=µn!i=1An−ixi.(13)B.TheTomographyProblemAnentityexternaltotheagentsnetwork(hereafternamedTomographyCenter,TC)isinterestedinreconstructingtheconnectivityprofileofthenetwork,namely,isinterestedinascertainingwhichagentisconnectedtowhichotheragent.TheTCisassumedtohaveaccesstoasubsetofthenetworkagents,andisabletocollectthestreamsofoutputsexchangedbytheagentsduringtheircommunicationprotocol.Formally,ifO⊂{1,2,...,N}istheaccessiblesubset,thedataavailableattimenare{yi(1),yi(2),...,yi(n)}i∈O.TheTCdoesnotknowtheoverallnumberofagentsinthenetwork.Accordingly,theonlygoaloftheTCisproducinganestimateoftheadjacencymatrixoftheobservedagents.Withoutlosinggenerality(andsincethechoiceofaparticularnumberingisimmaterialtoourpurposes)weshallassumethatagents1,2,...,Kareobserved,whiletheremainingM=N−Kagentsarenot.Moreover,inthisarticlewefocusontheasymptoticregimeKN→η<1(14)whereηtakesonthemeaningoftheasymptoticallyachievablefractionofobservedagents.Thisproblemischallengingforthefollowingreasons.•Letusignoreforamomentthefactthatthenetworkispartiallyobserved,andassumethattheTCiscollectingthewholesequencesofoutputsofthediffusionoutputforalltimeinstantsandforallagents.Inprinciple,thereexistseveralwellassessedstrategiestomakeinferenceabouttheinfluencethatoneagenthasonanotheragentstartingfromtheirdata.Themostintuitiveisclearlyanestimateofthecorrelationbetweentheoutputsof6theindicatorfunction,whichisequaltooneifconditionEistrue,andisequaltozero,otherwise.TheN⇥NidentitymatrixisdenotedbyI.ThecomplementofasetSisdenotedbyS0.GivenanN⇥NmatrixZ,thesubmatrixthatliesintherowsofZindexedbythesetS✓{1,2,...,N}andinthecolumnsindexedbythesetT✓{1,2,...,N},willbedenotedbyZST,oralternativelyby[Z]ST.WhenS=T,thesubmatrixZSTwillbeabbreviatedasZS.Intheindexingofthesubmatrixwewillretaintheindexsetoftheoriginalmatrix.Forexample,ifS={2,3}andT={2,4,5},wehavethatthesubmatrixM=ZSTisa2⇥3matrix,indexedasfollows:M=✓z22z24z25z32z34z35◆=✓m22m24m25m32m34m35◆.(15)Thisnotationiscrucialinourtreatment,sinceitwillallowustoidentifynodeswithoutcumbersomeandredundantdouble-indexnotation.B.GraphnotationG(V)isthesetofundirectedgraphsdefinedonasetofnodes(vertexset)V.WhenNisthenumberofnodes,theshortcutG(N)impliesthatthevertexsetisV={1,2,...,N}.WhendealingwithagraphG2G(N),itsconnectionstructure(i.e.,theedgesofthegraphs)canbedescribedthroughitsN⇥Nadjacencymatrix.The(i,j)-thentryoftheadjacencymatrixofgraphGwillbedenotedbygij,withgij=1ifnodesiandjareconnected,andgij=0,otherwise.Self-loops(gii=1)willbepermitted.GivenG2G(N),andasubsetS✓{1,2,...,N},thesubgraphcorrespondingtoSisdenotedbyGS2G(S).ThesupportgraphofamatrixAisdenotedbyG(A).The(i,j)-thentryofitsadjacencymatrixisI{aij>0},namely,nodesiandjareconnectedonG(A)if,andonly,ifaijisstrictlypositive.Apathfromitojisasequenceofedgeswherethefirstedgeoriginatesfromiandthelastedgeterminatesatj.Theexistenceofapathoflengthrcanbeexpressedas:gin1gn1n2...gnr 1j=1,(16)foracertainsequenceofverticesn1,n2,...,nr 1belongingtoV.Accordingtothisdefinition,apathcanalsopassmultipletimesthroughthesamenode,orcanlingerforoneormorestepsatthesamenodewhenithasaself-loop.Ni(G)isthesetofneighborsofnodei(includingiitself)intheundirectedgraphG.ThedegreeofnodeiisthecardinalityofNi(G).dmax(G)isthemaximumdegreeinG. i,j(G)isthedistancebetweenthenodesiandjonthegraphG,i.e.,thelengthoftheshortestpathlinkingiandj.N(r)i(G)isther-thorderneighborhoodofnodei(includingiitself),givenbyN(r)i(G)={j2V: i,j(G)r}.(17)ArandomgraphGobeystheErdos-R´enyimodelifeachedgeofGisdrawn,independentlyfromtheotheredges,withidenticalprobabilitypN.Equivalentlystated,theadjacencyrandomvariablesgij,fori=1,2,...,Nandi<j,areindependentandidenticallydistributed(i.i.d.)Bernoullivariables.ThenotationG⇠G?(N,pN)signifiesthatthegraphGbelongstotheErdos-R´enyiclasswithconnectionprobabilitythatvanishesasN!1,andthatobeysthefollowingscalinglaw:pN=logN+cNN,(18)wherecN!1asN!1(inanarbitraryway,providedthatpN!0).Itisawell-knownresultthatrandomgraphsbelongingtothefamilyG?(N,pN)areconnectedwithhighprobability[21].Beforeproceeding,andasanoteofclarity,weassumethatallrandomvariablesfinddomaininthecommonprobabilityspace(⌦,F,P),where⌦isthesetofrealizations,Fisthesigma-algebraofmeasurablesetsandPistheprobabilitymeasure.Forinstance,theevent{!2⌦: i,j(G(!))r}2F,(19)representsthesetofrealizations!2⌦yieldingagraphG(!)whosedistancebetweenthe(prefixed)nodesiandjdoesnotexceedr.Torenderamorecompactnotation,wehenceforthomittherealization!inthecharacterizationoftheevents.Forinstance,inthecaseoftheevent(19)werepresentitratheras{ i,j(G)r}2F,(20)wheretherandomquantitiesareemphasizedbytheboldfacednotation -- intheeventinequation(20),theonlyrandomobjectisthegraphG.C.UsefulGraphOperationsInourexposition,wewillbeperformingcertainoperationsovergraphs,aswellasevaluatecertainfunctionssuchascom-paringdistancesbetweennodesoverdistinctgraphs.There-fore,itisusefultointroducethefollowinggraphoperationsforlateruse(whichareillustratedinFigure2):1)Graphembedding.GivenavertexsetV,andasubsetthereof,S⇢V,theembeddingofagraphG(1)2G(S)intothelargergraphG(2)2G(V)willbedenotedby:G=G(1) G(2),G2G(V),(21)andresultsinagraphwiththefollowingproperties:i)theconnectionsbetweennodesinSthatarepresentinG(2)arecancelled;ii)thenodesinthevertexsetSofgraphG(1)aremappedintothecorrespondingnodesofgraphG(2),andsoarethepertinentconnections.WestressthattheconnectionsfromS0toSaredeterminedbythegraphG(2).Wenoticethattheoperationin(21)isnotcommutative(becausethefirstgraphisembeddedintothesecondgraph,andnotviceversa),andthattheoutputgraphGdoesnotdependontheconnectionsexistinginG(2)amongnodesbelongingtothesetS.2)Localdisconnection.GivenagraphG2G(V),thenotation:GU1=U22G(V),(22)describesthegraphthatisfromGafterremovingalltheedgesthatconnectnodesinU1tonodesinU2,namely,alltheconnectionsbetweenU1andU2.aji>0 aii>0 aij>0 i j work), using (1), and neglecting transient terms associated
with the initial state y0, we have that:
R0(n) = β2
n−1(cid:88)i=0
A2i n→∞−→ R0 = β2(I − A2)−1
(5)
where the latter series is guaranteed to converge whenever A
is a stable matrix -- all its eigenvalues lie inside the unit disc.
Assuming that the system is observed at steady-state (i.e., that
the system is in operation since some time), we must focus on
the limiting correlation matrix, R0. However, we immediately
see from (5) that, even if the correlation matrix were perfectly
known, direct retrieval of the support graph of A from R0 is
obfuscated by the fact that the correlation matrix depends on (a
superposition of) powers of A, and not only on A. Moreover,
even with full observation, when inversion of the matrix R0
can be performed, in view of equation (5), one would retrieve
A2, and not A. Then, since the mapping from A2 to (the
support graph of) A is not bijective in general, one would be
faced with the inverse problem of retrieving the support graph
of A from A2 -- such inverse problem can still be explored
by properly reinforcing some sparsity constraints, refer, e.g.,
to [30].
Tomography relies primarily in developing a scheme to
properly process the observables -- e.g., the state evolution
of the interconnected agents -- so as to infer the underlying
network structure. The above naıve scheme, based purely on
the correlation R0, can be improved by introducing the one-lag
correlation matrix, which, in view of (1), takes the form:
n−1] = AR0(n − 1) n→∞−→ R1 = AR0.
R1(n) (cid:44) E[ynyT
(6)
Therefore, we obtain the following relationship:
A = R1R−1
0
(7)
In principle, since there exist many ways to estimate R0 and
R1 consistently as n → ∞, expression (7) reveals one possible
strategy to estimate A (and hence its support-graph) from the
observations.
Topology estimation based on relation (7) is viable when-
ever full-observation of the system is permitted. Under a
partial observability restriction, however, when only a subset
S of the network is accessible, only the covariance submatrices
associated with the observable agents, denoted by [R0]S and
[R1]S, are available. One is certainly free to introduce a
truncated version of (7), say, as:
(8)
(cid:98)AS = [R1]S ([R0]S)−1
It is clear from basic linear algebra that (8) is in general distinct
from the ground truth matrix AS = (cid:2)R1R−1
from the combination matrix corresponding to the subnetwork
connecting the observable agents S whose support must be
inferred.
0 (cid:3)S, namely,
Despite this difference, it has been shown in the recent
work [31], [32] that the support of the observable network
can still be recovered (consistent tomography) through the
truncated estimator in (8), under certain conditions that can be
summarized as follows: i) the overall network graph is drawn
3
from a connected Erdos-R´enyi random graph with vanishing
connection probability; ii) the cardinality of the observed
subnetwork grows linearly with the size of the overall network;
iii) the matrix A is a symmetric combination matrix belonging
to a certain class.
The work [31] leads to several insightful conclusions about
network tomography for Erdos-R´enyi models. In this work,
we pursue the same network tomography problem albeit for
a different more demanding network setting, explained below,
which will require new arguments and lead to new results that
extend the results from [31]. In particular, the proof techniques
used here will rely on graph theoretical techniques and on
special graph constructs to arrive at the important conclusion
that the naıve truncated estimator (8) is still able to deliver
consistent tomography under partial network observations and
more relaxed requirements.
It is worth noting that this work establishes a strong consis-
tency result (Theorem 1) that does not require an independence
condition. The main features of the framework proposed
in the current manuscript in comparison with [31] can be
summarized as follows:
-- Topology of the accessible network portion. We assume
that the subnetwork connecting the observable agents has an
arbitrary topology, which is modeled through a deterministic
graph. This subgraph is the object of inference. The remaining
unobserved part is assumed to be drawn from an Erdos-R´enyi
random graph. The overall network construction is therefore
referred to as a partial Erdos-R´enyi model. It is useful to
interpret and motivate this model in the classical "signal plus
noise" paradigm in the following sense. For what concerns
the object of the inference (i.e., the support graph of the
observable nodes), it is modeled as an arbitrary deterministic
signal. For what concerns the undesired component (i.e., the
unobserved subnet), it is modeled as a noisy component. To
get insightful results, we must choose some model for this
random component. In the absence of any prior information,
it is meaningful to opt for a uniform model, namely, the
Erdos-R´enyi random graph, where the presence/absence of
each edge is determined through a sequence of i.i.d. Bernoulli
experiments. Accordingly, few connections (i.e., high sparsity
in the unobserved portion) take on the meaning of a controlled
noise level. In contrast, in [31] it was assumed that the overall
network (observed portion + unobserved portion) is drawn
from an Erdos-R´enyi random graph. Such a construction
poses limitations on the subgraph that we wish to identify,
which cannot be selected in an arbitrary fashion any longer.
Moreover, the network construction used in [31] assumes a
vanishing fraction of connected nodes within the observable
set; a condition that is removed in the current analysis.
-- Cardinality of the accessible network portion. In [31], it
was assumed that the cardinality S of the observed subset S
scales linearly with N, so that the ratio S/N converges to
some positive fraction ξ ∈ ( 0, 1] as N → ∞. In contrast,
we assume here that the subnetwork of observable nodes S
is fixed. This means that, in our framework, we focus on
retrieving the support of a subnetwork S that is embedded
in a network that becomes infinitely larger as N → ∞, i.e.,
the size of the unobserved component becomes asymptotically
dominant. The resulting regime is accordingly referred to
as a low-observability regime. Such a model is particularly
relevant, for example, in circumstances where we have a large
network, and we are constrained to perform probing actions at
few accessible locations. We remark that the case where the
ratio S/N goes to zero is not addressed (nor can be obtained
from the results) in [31].
-- Consistent tomography. The main result of the present
work (Theorem 1 further ahead) is to establish that consistent
tomography is achievable under the aforementioned setting.
We shall prove that, if the unobserved network is drawn from
an Erdos-R´enyi random graph with connection probability
pN =
log N + cN
N
(9)
where cN is a sequence that diverges as N → ∞, satisfying
the condition:
[log(log N + cN )]2
log N
→ 0
(10)
then the (arbitrary) support graph of AS can be recovered
through the truncated estimator (cid:98)AS, with probability tending
to one as N → ∞. More specifically, in this work we are
able to establish consistency in that each entry of the support
graph is recovered perfectly in the limit as N → ∞. In [31],
where the object of estimation has cardinality growing with
N, consistency is not formulated in terms of an entry-by-
entry recovery. Instead, consistency there is formulated in
terms of two macroscopic indicators, namely,
the fraction
of correctly classified interacting pairs, and the fraction of
correctly classified non-interacting pairs. Both fractions are
proved to converge to one as N grows to infinity.
Another difference with respect
to [31] relates to the
connection probability of the Erdos-R´enyi graph. Having a
sufficiently small pN translates into a sufficient degree of
sparsity. In other words, if we interpret the unobserved network
as a noisy component, the noise in the system cannot exceed
a certain threshold to grant perfect reconstruction. For the
setting considered in [31], a connection probability vanishing
as in (9) was sufficient to achieve consistent tomography,
without additional constraints on cN . On the other hand,
for the results of this work to hold, we need the additional
constraint in (10), which corresponds to invoking slightly more
sparsity.
We remark that the results of this work allow drawing some
useful conclusions also in relation to the setting addressed
in [31], namely, in relation to the case of a full Erdos-R´enyi
construction with S growing linearly with N. We find it
convenient to postpone the comments on this particular issue
to Sec. VII, because some technical details are necessary for
a proper explanation.
In summary, in this work we address a network tomography
problem in the low-observability regime, where the cardinality
of the accessible network portion is fixed and the number of
unobserved nodes scales to infinity. This challenging regime
cannot be dealt with by using the analysis provided in [31].
In comparison with [31], where the accessible network was
4
Erdos-R´enyi with vanishing connection probability, in this
work we consider an arbitrary topology for the accessible
network. Having the flexibility of an arbitrary topology is crit-
ical because, in the low-observability regime, an Erdos-R´enyi
accessible subgraph with vanishing connection probability
would become a trivial subgraph that is totally-disconnected
as N → ∞. Moreover, the result in [31] holds in a weak
sense: it shows that the fraction of correctly identified edges
converges to one in the limit and an independence approxi-
mation is required to control the error rate. Remarkably, in
this work consistency holds in a strong sense, meaning that
the whole subgraph is exactly recovered as N → ∞, and no
independence approximation is required.
B. Related work
The existing approaches to network reconstruction can be
categorized based on two major features:
-- F1: Class of networked dynamical systems governing the
state-evolution of the agents, e.g., the diffusion model in (1),
and related observables, e.g., the process yn in (1).
-- F2: Topology-retrieval methods that should exploit the
relation between the observables and the underlying support-
topology. Such methods are sensitive to the dynamics and the
observables arising from the model in F1.
Regarding F1, most works focus on linear systems. Non-
linear dynamics are often dealt with by linearizing via con-
sidering variational characterizations of the dynamics (under
small-noise regimes) [33] -- [35] or by appropriately increasing
the dimension of the observable space [36], [37]. In the context
of linear (or linearized) systems, particular attention is paid to
autoregressive diffusion models [28], [30], [38] -- [40].
For what concerns F2, the majority of the literature con-
siders methods aimed at identifying commonalities between
correlation constructs and graph topologies. We now make a
brief summary of the available results as regards the existing
topology-retrieval methods that are more closely related to our
setting. To get some ideal benchmark, it is useful to start with
the full observation case, and then focus on the case of interest
of partial observation.
-- Tomography under full observations. In [41], the authors
introduce directed information graphs, which are used to
reveal the dependencies in networks of interacting processes
linked by causal dynamics. Such a setting is enlarged in [42],
where a metric is proposed to learn causal relationships by
means of functional dependencies, over a (possibly nonlinear)
dynamical network. Causal graph processes are exploited
in [40], where an algorithm (with a detailed convergence anal-
ysis) is proposed for topology recovery. Recently, the inverse
problem of recovering the topology via correlation structures
has been addressed through optimization-based methods, by
reinforcing some (application-dependent) structural constraints
such as sparsity, stability, symmetry. For instance, in [30],
[39], since the combination matrix and the correlation matrix
share the same eigenvectors, the set of candidate topologies
is reduced by computing these eigenvectors, and the inverse
problem is then tackled with optimization methods under
sparsity constraints.
An account of topology inference from node measurements
(still under the full observations regime) is offered in [43],
where a general linear model is considered and an approach
based on Wiener filtering is proposed to infer the topology.
the combination weights {aij}N
However, as already noted in [43] a Wiener filtering ap-
proach is redundant, since exact topology recovery can be
obtained (with full observations) through the estimator in (7).
As it is well known, this solution admits the following useful
interpretation:
j=1 obtained
through (7) are the coefficients of the best one-step linear
predictor (a.k.a., in the context of causal analysis, as Granger
estimator), i.e., they yield the minimum expected squared error
j=1 --
in estimating yi(n) from the past samples {yj(n − 1)}N
see, e.g., [44]. We remark that the case where (7) is applied
with correlation matrices estimated empirically from the mea-
surements provides the best one-step linear predictor in a least-
squares sense (i.e., when the expected squared error is replaced
by the empirical squared error evaluated on the measurements
collected over time). However, all the aforementioned results
pertain to the case where node measurements from the whole
network are available. It is instead necessary to consider the
case when only partial observation of the network is permitted.
-- Tomography under partial observations, identifiability. The
case of partial observations is addressed in [45], [46], for cases
when the network graph is a polytree.
The case of more general topologies is instead addressed
in [38], [47], where technical conditions for exact or partial
topology identification are provided. It is useful to contrast
such identifiability conditions with the approach pursued in the
present work. Basically, the identifiability conditions offered
in [38], [47] act at a "microscopic" level, namely, they need a
detailed knowledge of the topology and/or the statistical model
(e.g., type of noise, joint distribution of the observable data).
For these reasons, the approach is not practical for large-scale
network settings (which are the main focus of this work).
In contrast, in this work we pursue a statistical asymptotic
approach that is genuinely tailored to the large-scale setting:
the conditions on the network topology are described at a
macroscopic level through average descriptive indicators, such
as the connection probability between any two nodes. Under
these conditions, we focus on establishing an achievability
result that holds (in a statistical sense) as the size of the
network scales to infinity.
-- Tomography under partial observations, methods. As al-
ready noted, the classic, exact solution to the topology problem
under full observation is provided by (7), and arises from the
solution of a one-step linear prediction problem [38], [43].
Under partial observations, we propose to keep the same
approach, except that the best one-step linear prediction is
enforced on the observable nodes only. As a matter of fact,
the combination weights estimated through (8) provide the
best one-step linear prediction of the observable measurement
yi(n) (for i ∈ S) from the past observable measurements
{yj(n − 1)}j∈S. We remark that this solution, which can still
be interpreted as a Granger estimator, is widely adopted in
causal inference from time series, when one ignores and/or
neglects the existence of latent components. However, there is
in principle no guarantee that such an estimator can provide
5
reliable tomography. Our main goal is to establish that it actu-
ally can, under the demanding setting illustrated in Sec. I-A.
-- Connections with graphical models. In a nutshell, a
graphical model can be described as a collection of random
variables indexed by nodes in a network, whose pairwise
relationships (which determine the topology, i.e., the undi-
rected graph) are encoded in a Markov random field. One of
the fundamental problems in graphical models is to retrieve
the network topology by collecting measurements from the
network nodes. It is useful to comment on some fundamental
differences, as well as useful commonalities, between the
graphical model setting and our problem.
In the standard graphical model formulation (and, hence, in
the vast majority of the available related results) the network
evolution over time (e.g., the dynamical system in (1)) is not
taken into account. Rather, the samples {yn}n∈N are assumed
independent across the index n. This difference has at least
two relevant implications.
The first difference pertains to the type of estimators used
for topology retrieval. For example, in a Gaussian graphi-
cal model, the inverse of the correlation matrix R−1
0 , a.k.a.
concentration matrix, contains full information of the graph
topology: the (i, j)-th entry of the concentration matrix is
nonzero if, and only if, nodes i and j are connected. In
contrast, we see from the Granger estimator in (7) that in
our case an additional operation is needed (namely, multipli-
cation with the one-lag correlation matrix R1) to obtain the
matrix that contains the topology information (in our case,
the combination matrix, A). This difference is an inherent
consequence of the system dynamics described by the first-
order VAR model
the dynamical system
ruling the network evolution usually enforces some degree
of dependence between subsequent measurements. For this
reason, while in our case the observations collected over time
are correlated, in the standard graphical model formulation
the samples upon which the topology inference is based are
usually assumed statistically independent. Keeping in mind
these fundamental distinctions, we now list some recent works
about topology recovery on graphical models.
in (1). Second,
The idea of studying the large-network behavior through
an Erdos-R´enyi model has been applied in [48], where the
emergence of "large" paths over the random graph (a property
that we will use in our treatment) has been exploited for
topology inference. However, reference [48] addresses the
case of full observations. Instead, for the case of partial
observations, in [49] an efficient method is proposed, which
is suited for the case of large-girth graphs, such as, e.g., the
bipartite Ramanujan graphs and the random Cayley graphs.
In [50], still for the case of partial observations, an inference
method is proposed under the assumption that the connection
matrix is sparse, whereas the error matrix associated to the
latent-variables component exhibits a low-rank property.
In summary, contrasted with recent results about topology
recovery on graphical models,
the results obtained in the
present work constitute an advance because: i) we deal with a
dynamical system, see (1); ii) the partial observations setting
considered in the present work relies on assumptions different
from those used in [49], [50]: in our case the unobserved
component is Erdos-R´enyi, but the subnetwork of observable
nodes is deterministic and arbitrary, and the combination ma-
trix obeys transparent conditions borrowed from the adaptive
networks literature. We believe that the possibility of working
with dynamical models, the arbitrariness of the monitored
subnetwork, as well as the direct physical meaning of the
conditions on the combination matrix, provide useful novel
insights on the problem of topology inference under partial
observations.
To sum up, our major contribution lies in establishing
technical guarantees for graph structural consistency of the
Granger estimator applied to the subset of observable nodes.
This is formally stated in the main result of this paper,
Theorem 1.
Short versions of this work were reported in [51], [52].
C. Motivating Example: Adaptive Diffusion Networks
A network of N agents observes a spatially and temporally
i.i.d. sequence of zero-mean and unit-variance streaming data
{xi(n)}, for n = 1, 2, . . ., and i = 1, 2, . . . , N. Here, the
letter n refers to the time index while the letter i refers to
the node index. In order to track drifts in the phenomenon
they are monitoring, the network agents implement an adaptive
diffusion strategy [14], [21], [53], where each individual agent
relies on local cooperation with its neighbors. One useful form
is the combine-then-adapt (CTA) rule, which has been studied
in some detail in these references. It involves two steps: a
combination step followed by an adaptation step.
During the first step, agent i combines the data of its
neighbors through a sequence of convex (i.e., nonnegative
and adding up to one) combination weights wi(cid:96), for (cid:96) =
1, 2, . . . , N. The combination step produces the intermediate
variable:
vi(n − 1) =
wi(cid:96) y(cid:96)(n − 1).
(11)
N(cid:88)(cid:96)=1
Next, during the adaptation step, agent i updates its output
variable by comparing against the incoming streaming data
xi(n), and updating its state by using a small step-size µ ∈
(0, 1):
yi(n) = vi(n − 1) + µ[xi(n) − vi(n − 1)].
Merging (11) and (12) into a single step yields:
(12)
(13)
yi(n) = (1 − µ)
wi(cid:96) y(cid:96)(n − 1) + µ xi(n).
N(cid:88)(cid:96)=1
It is convenient for our purposes to introduce a combination
matrix, which we denote by A, whose entries are obtained by
scaling the weights wij as follows:
aij (cid:44) (1 − µ)wij.
(14)
With this definition, we see immediately that (13) corresponds
to (1) or (4) with β = µ. Note that under this diffusion
framework, the matrix A is naturally nonnegative and if we
assume symmetry, its normalized counterpart A/(1 − µ) is
doubly stochastic.
6
II. NOTATION AND DEFINITIONS
We list our notation and some definitions used in later sections
for ease of reference.
A. Symbols
We represent sets and events by upper-case calligraphic
letters, and the corresponding normal font letter will be used
to denote the set cardinality. For instance, the cardinality of
set S is S. The complement of a set S is denoted by S(cid:48).
Standard canonical sets follow a different convention, for
instance,
the set of natural numbers is denoted by N =
{1, 2, 3, . . .}, and the set of N × N symmetric matrices with
nonnegative entries by SN×N
We use boldface letters to denote random variables, and
normal font letters for their realizations. Capital letters refer to
matrices, small letters to both vectors and scalars. Sometimes
we violate the latter convention, for instance, we denote the
total number of network agents by N.
+
.
Given an N × N matrix Z, the submatrix that lies in the
rows of Z indexed by the set S ⊆ {1, 2, . . . , N} and in the
columns indexed by the set T ⊆ {1, 2, . . . , N}, is denoted by
ZST, or alternatively by [Z]ST. When S = T, the submatrix
ZST will be abbreviated as ZS or [Z]S. In the indexing of the
submatrix we will retain the index set of the original matrix.
For example, if S = {2, 3} and T = {2, 4, 5}, we have that
the submatrix M = ZST is a 2× 3 matrix, indexed as follows:
M =(cid:18) z22
m32 m34 m35 (cid:19) . (15)
z35 (cid:19) =(cid:18) m22 m24 m25
z24
z34
This notation is crucial in our treatment, since it will allow us
to identify nodes without cumbersome and redundant double-
index notation.
z25
z32
Finally, 1 denotes a column vector with all its entries equal
to one; 0N×N denotes an N × N matrix with all its entries
equal to zero; IE denotes the indicator function, which is equal
to one if condition E is true, and is equal to zero, otherwise;
the N × N identity matrix is denoted by I; and log(·) denotes
the natural logarithm.
B. Graph notation
The set of all undirected graphs that can be defined on a set
of nodes (vertex set) V is denoted by G(V). When N is the
number of nodes, the notation G(N ) implies that the vertex
set is V = {1, 2, . . . , N}.
When dealing with a graph G ∈ G(N ), its connection
the edges of the graph) can be described
structure (i.e.,
through its N × N adjacency matrix. The (i, j)-th entry of
the adjacency matrix of the graph G will be denoted by the
lower-case symbol gij, with gij = 1 if the nodes i and j
are connected, and gij = 0 otherwise. Henceforth, we assume
that gii = 1, i.e., all nodes exhibit self-loops. This reflects
the fact that usually each agent uses information from its own
output measurement to update its state.
Given G ∈ G(N ), and a subset S ⊆ {1, 2, . . . , N}, the
subgraph corresponding to S is denoted by GS ∈ G(S). The
support graph of a matrix A is denoted by G(A). The (i, j)-th
entry of its adjacency matrix is I{aij >0}, namely, nodes i and
j are connected on G(A) if, and only, if aij is strictly positive.
A path from i to j is a sequence of edges where the first
edge originates from i and the last edge terminates at j. The
existence of a path of length r can be expressed as:
gin1 gn1n2 . . . gnr−1j = 1,
(16)
for a certain sequence of vertices n1, n2, . . . , nr−1 belonging
to V. According to this definition, a path can also pass multiple
times through the same node, or can linger for one or more
steps at the same node when it has a self-loop.
The set of neighbors of the node i (including i itself) in the
undirected graph G will be denoted by Ni(G). The degree of
the node i is the cardinality of Ni(G), whereas dmax(G) is the
maximum degree in G. Likewise, the r-th order neighborhood
of the node i (including i itself) is denoted by N(r)
(G), and
is formally given by:
i
N(r)
i
(G) = {j ∈ V : δi,j(G) ≤ r},
(17)
where δi,j(G) is the distance between the nodes i and j on
the graph G, i.e., the length of the shortest path linking i and
j.
A random graph G obeys the Erdos-R´enyi model if each
edge of G is drawn, independently from the other edges, with
identical probability pN . Equivalently stated, the adjacency
random variables gij, for i = 1, 2, . . . , N and i < j,
are independent and identically distributed (i.i.d.) Bernoulli
variables. The notation G ∼ G (cid:63)(N, pN ) signifies that the
graph G belongs to the Erdos-R´enyi class with connection
probability that vanishes as N → ∞, and that obeys the
following scaling law:
pN =
log N + cN
N
,
(18)
where cN → ∞ as N → ∞ (in an arbitrary way, provided
that pN → 0). It is a well-known result that random graphs
belonging to the family G (cid:63)(N, pN ) are connected with high
probability [54].
Remark 1. As a note of clarity, in the forthcoming treatment,
we assume that all random variables find domain in a common
probability space (Ω, F, P), where Ω is the set of realizations,
F is the sigma-algebra of measurable sets and P is the
probability measure. For instance, the event
{ω ∈ Ω : δi,j (G(ω)) ≤ r} ∈ F,
(19)
represents the set of realizations ω ∈ Ω yielding a graph G(ω)
whose distance between the (fixed) nodes i and j does not
exceed r. To render a more compact notation, we henceforth
omit the realization ω in the characterization of the events. For
instance, in the case of the event (19) we represent it rather
as
{δi,j (G) ≤ r} ,
(20)
where the random quantities are emphasized by the boldface
letter -- in the event in (20), the only random object is the
graph G as it is the only boldface variable.
7
Fig. 3. Summary of the graph operations defined in Sec. II-C related to
embedding, disconnection, and inheritance.
C. Useful Graph Operations
In our exposition, we will be performing certain operations
over graphs, as well as evaluate certain functions such as com-
paring distances between nodes over distinct graphs. There-
fore, it is useful to introduce the following graph operations
for later use (which are illustrated in Figure 3):
1) Graph embedding. Given a vertex set V, and a subset
thereof, S ⊂ V, the embedding of a graph G(1) ∈ G(S)
into the larger graph G(2) ∈ G(V) will be denoted by:1
(21)
G = G(1) ⊕ G(2),
G ∈ G(V),
and results in a graph with the following properties: i)
the connections between nodes in S that are present in
G(2) are cancelled; ii) the nodes in the vertex set S of
graph G(1) are mapped into the corresponding nodes of
graph G(2), and so are the pertinent connections. We
stress that the connections from S(cid:48) to S are determined
by the graph G(2). We notice that the operation in (21)
is not commutative (because the first graph is embedded
into the second graph, and not vice versa), and that the
output graph G does not depend on the connections
existing in G(2) among nodes belonging to the set S.
2) Local disconnection. Given a graph G ∈ G(V), the
(22)
notation:
GU1(cid:61)U2 ∈ G(V),
1In order to avoid confusion, we remark that the symbol ⊕ is also used,
in the graph literature, to denote a different kind of operation called "ring
sum". However, we prefer to denote our embedding operation by the same
summation symbol to emphasize the "signal+noise" structure that is relevant
in our application.
DisconnectionInheritanceEmbedding describes the graph that is obtained from G by removing
all the edges that connect nodes in U1 to nodes in U2,
namely, all the connections between U1 and U2.
3) Connections inheritance. The notation:
8
Gj←U ∈ G(V ),
(23)
describes the graph that is obtained from G through the
following chain of operations: i) all edges within U are
removed; ii) all edges connecting nodes in U to the rest
of the network are removed; iii) all connections from
U to the rest of the network are inherited by node j.
All the above graph operations preserve self-loops unless
otherwise stated.
III. PROBLEM FORMULATION
Consider a graph G(V) and assume we are able to observe
data from a subset S ⊂ V of the nodes. From these obser-
vations, we would like to devise a procedure that allows us
to discover the connections among the nodes in S, under the
assumption that the structure of the graph in the complement
set, S(cid:48), and as well as the connections between S and S(cid:48), will
be random, following i.i.d. drawing of the pertinent edges. The
desired construction can be formally described as follows.
Let G(obs) ∈ G(S) be a deterministic graph on the observable
set S, with some unknown topology (which is not restricted
in any way), and let G(unobs) ∼ G (cid:63)(N, pN ) be an Erdos-R´enyi
random graph on N nodes. We assume that the overall network
graph, G, is of the form:
G = G(obs) ⊕ G(unobs)
(24)
S
Specifically, the connections within the observable set S are
described through the graph G(obs), while the connections
within S(cid:48), as well as the connections between S(cid:48) and S, are
described through the graph G(unobs). Note that G(unobs)
is
an Erdos-R´enyi random graph on N nodes, but its subgraph
G(unobs)
is replaced by G(obs) in characterizing G, in view
of (24) (refer also to Figure 3). Therefore, the structure of
G(unobs) within the observable subnet becomes immaterial.
Equation (24) highlights the "signal+noise" construction, with
the boldface notation emphasizing the random (i.e., noisy)
component that corresponds to the unobserved network por-
tion, and with the normal font emphasizing the deterministic
component that corresponds to the arbitrary topology of the
observed network portion.
The aforementioned construction will be referred to as
a partial Erdos-R´enyi graph. The class of partial Erdos-
R´enyi random graphs with a deterministic graph component
G(obs) placed on the set S, will be formally represented by
the notation G (cid:63)(N, pN , G(obs)). We shall often refer to the
observable graph over S by the simpler notation GS. As such,
we can also write,
G ∼ G (cid:63)(N, pN , GS)
(25)
to denote partial Erdos-R´enyi random graphs with determin-
istic component GS. We assume that G (and hence GS)
is unknown. In this context, the goal of local tomography
is to estimate GS via observing the state evolution of the
Fig. 4. Observations are collected from the red subgraph GS (comprised
by the red nodes along with the red edges). The black edges,
the
edges connecting green-green and green-red nodes are assumed to be drawn
randomly with probability pN .
i.e.,
observable agents in S. Figure 4 illustrates the partial Erdos-
R´enyi construction just described.
Before formulating the tomography problem, we observe
that, under condition (9), the partial Erdos-R´enyi graph is
asymptotically connected with high probability for any choice
of the subgraph GS, as stated in the following lemma.
Lemma 1 (Connectivity of partial Erdos-R´enyi graphs). Given
any graph GS ∈ G(S), the partial Erdos-R´enyi graph
G ∼ G (cid:63)(N, pN , GS)
is connected with high probability, i.e.,
(26)
lim
N→∞
Proof: See Appendix A.
P [G is connected] = 1.
(27)
A. Combination assignment.
A combination assignment
is a rule that builds the
combination matrix A as a function of the underlying graph
G. We assign (positive) weights to the edges of G and
denote the resulting matrix of weights by A. Some useful and
popular choices are the Laplacian and the Metropolis rules,
which arise naturally in the context of adaptive diffusion
networks [53], and are defined as follows2. Let ρ ∈ (0, 1) and
λ ∈ (0, 1 ]:
Laplacian rule.
2Strictly speaking, in the network literature the Laplacian and Metropolis
rules are defined with weights that add up to one, which would correspond
to (28) and (29) without the multiplying factor ρ. The multiplying factor
ρ, which provides the matrix stability,
is usually left separate and not
absorbed into the combination matrix. For instance, in the case of diffusion
networks (14) we have ρ = 1−µ, where µ is the step-size. In our treatment, it
is more convenient to include this scaling factor into the combination matrix,
as done in (28) and (29).
aij = ρ ×
λgij
dmax
,
1 −
λ
dmax(cid:88)(cid:96)(cid:54)=i
for i (cid:54)= j
gi(cid:96),
for i = j
.
(28)
Metropolis rule.
aij = ρ ×
gij
max(di, dj)
1 −(cid:88)(cid:96)(cid:54)=i
,
gi(cid:96)
max(di, d(cid:96))
for i (cid:54)= j
,
for i = j
,
(29)
where di is the degree of agent i and dmax is the maximum
degree in the network. It is useful to remark that different com-
bination rules may have different impact on the performance
of the topology estimators, as we will see in the examples
presented in Sec. V.
In this paper, we shall focus on the family of nonnegative
symmetric combination policies introduced in [31], and whose
characterizing properties we recall next.
Property 1 (Bounded-norm). The maximum row-sum norm,
A∞
∆= max
i
ai(cid:96),
(30)
N(cid:88)(cid:96)=1
is upper bounded by some ρ < 1.
(cid:3)
For nonnegative symmetric matrices, Property 1 becomes:
A∞
= max
i=1,2,...N
N(cid:88)(cid:96)=1
ai(cid:96) = max
i=1,2,...N
N(cid:88)(cid:96)=1
a(cid:96)i ≤ ρ
(31)
From Property 1 we see that (most of) the combination weights
ai(cid:96) typically vanish as N gets large, since a finite mass of value
at most ρ must be allocated across an ever-increasing number
of neighbors -- on an Erdos-R´enyi graph, the average number
of neighbors scales as N pN , and in the regime considered in
this paper we have N pN → ∞ in view of (18).
The next property identifies a useful class of combination
policies, for which degeneracy to zero of the combination
weights is prevented by proper scaling. As highlighted below,
such property is broad enough to encompass typical combina-
tion rules, such as the Laplacian (28) and the Metropolis (29)
rules.
Property 2 (Non-degeneracy under (N pN )-scaling). Consider
a combination policy applied to a partial Erdos-R´enyi graph
G ∼ G (cid:63)(N, pN , GS). The combination policy belongs to class
Cτ for some τ > 0, if for all i, j = 1, 2, ..., N with i (cid:54)= j:
P[N pN aij > τgij = 1] ≥ 1 − N
(32)
where N goes to zero as N → ∞. In other words, if two nodes
i and j are connected (corresponding to gij = 1), then the
scaled combination coefficient N pN aij lies above a certain
threshold value denoted by τ, with high probability, for large
(cid:3)
N.
It is useful to remark that, since condition (32) is applied
to a partial Erdos-R´enyi construction, the nodes belonging to
the observable set S are connected in a deterministic fashion.
This means that, for i, j ∈ S, the random variable gij = gij is
in fact deterministic. In this case, the condition in (32) should
be rewritten, for any connected pair (i, j) in GS, as:
P[N pN aij > τgij = 1] = P[N pN aij > τ ] ≥ 1 − N , (33)
9
because conditioning on a deterministic event becomes imma-
terial.
We now introduce a sufficient condition under which a
combination rule fulfills Property 2. The relevance of this
condition is that it can be readily verified for certain typical
combination assignments such as Laplacian and Metropolis
rules, and it automatically provides one value of τ to identify
the class Cτ .
Lemma 2 (Useful policies belonging to Cτ ). Any policy
obeying the inequality
aij ≥
γ
dmax(G)
gij
(34)
for all i (cid:54)= j and for some γ > 0, satisfies Property 2 with the
choice τ = γ/e.
Proof: See Appendix A.
We denote by Cρ,τ the class of weight-assignment policies
for which Properties 1 and 2 hold simultaneously.
Using the definition of the Laplacian rule in (28), it is
readily verified that this rule possesses Property 1 and ful-
fills (34) with the choice γ = ρλ. Likewise, using (29), it
is readily verified that the Metropolis rule possesses Property
1 and fulfills (34) with the choice γ = ρ. As a result, both
policies belong to the class Cρ,τ , with the following choices
of τ (the meaning of the subscripts should be obvious):
τL =
ρλ
e
,
τM =
ρ
e
.
(35)
Before proving the main result of this work, it is useful
to illustrate the physical meaning of Property 2 in connection
with the network tomography problem. We introduce the S×S
error matrix that quantifies how much the truncated estimator
in (8) differs from the true sub-matrix AS, namely,
ES (cid:44) (cid:98)AS − AS
(36)
The magnified (i, j)-th entry of the truncated estimator in (8),
not vanishing
N pN eij,
N pN [(cid:98)AS]ij, can be written as:
+N pN eij,
N pN aij
if i and j are connected,
(cid:124) (cid:123)(cid:122) (cid:125)
to (37), if we want the nonzero entries N pN(cid:98)aij to stand out
from the error floor, when i and j are interacting, or to be
bounded above (by τ), when i and j are non-interacting, as
N grows large, we must be able to control the impact of the
error term N pN eij.
where eij is an error quantity, and the qualification of being
"not vanishing" is a consequence of Property 2. According
otherwise,
(37)
We are now ready to summarize the main problem treated
in this article.
Local Tomography. Let A be an N × N matrix obtained
from any combination assignment belonging to the class Cρ,τ ,
over a graph G ∼ G (cid:63)(N, pN , GS) on N nodes with a given
(arbitrary) subgraph GS. Let {[yn]S}n∈N be the state-evolution
associated with the observable subset of agents S and obeying
the stochastic dynamical law
yn = Ayn−1 + βxn.
(38)
Problem: given {[yn]S}n∈N, can we determine GS?
IV. MAIN RESULT
The main result of this work is to establish that the truncated
estimator (cid:98)AS = [R1]S ([R0]S)−1 introduced in (8), contains
enough information to recover the true support graph, GS, of
the combination matrix AS that corresponds to the observable
subset, S. More specifically, we establish that a positive
threshold τ exists, such that the graph obtained by comparing
the entries of (cid:98)AS against this threshold matches, with high
probability, the true support graph GS. This implies that the
topology of the subnetwork GS can be fully recovered, with
high probability, via the output measurements {[yn]S}n∈N
used to construct [R0]S and [R1]S, namely, via the observable
nodes only. Even if broader observation is permitted,
the
result enables the possibility of surmounting the curse of
dimensionality by processing smaller S × S matrices [R1]S
and [R0]S, instead of large-scale N × N matrices R0 and R1
-- wherein one of the operations involves the often expensive
inversion of a large matrix R0 -- and yet attaining exact
recovery with high probability.
Before stating the main theorem, let us introduce a useful
thresholding operator. We consider a matrix M ∈ SS×S
+ ,
whose (i, j)-th entry is mij, with i ∈ S and j ∈ S. The
thresholding operator compares the off-diagonal entries against
some threshold τ > 0, and produces as output a graph,
Γτ (M ) ∈ G(S), whose adjacency matrix has (i, j)-th entry
equal to
(39)
I{mij >τ},
In other words,
the thresholding operator returns a graph
whereby two nodes i and j are connected only if the (i, j)-th
entry of the matrix M exceeds the threshold. We assume all
entries of the main diagonal of G equal to one. We are now
ready to state the main theorem.
Theorem 1 (Exact recovery of GS). Let GS be a given
deterministic graph (with arbitrary topology) on S nodes, and
let G ∼ G (cid:63) (N, pN , GS) be a partial Erdos-R´enyi random
graph where the sequence cN that determines the connection
probability, pN = (1/N )(log N + cN ), obeys the condition:
∀i (cid:54)= j.
[log(log N + cN )]2
log N
→ 0.
(40)
Let also A be a combination matrix with support graph G.
If there exists τ > 0 and 0 < ρ < 1 such that A belongs to
Cρ,τ , then the following results hold.
i) If i, j ∈ S are interacting, then the (i, j)-th magnified
entry of the truncated estimator, N pN [(cid:98)AS]ij, exceeds the
threshold τ with high probability as N → ∞.
ii) If i, j ∈ S are non-interacting, then the (i, j)-th magnified
entry of the error matrix, N pN eij, converges to zero in
probability.
iii) The graph obtained by applying the thresholding operator
in (39) to the magnified truncated estimator, N pN (cid:98)AS,
10
matches the true support graph, GS, with high probability
as N → ∞, namely,
lim
N→∞
P[Γτ (N pN (cid:98)AS) = GS] = 1.
Proof: See Appendix B.
(41)
A. Outline of the main proof
We offer here an outline of the proof of Theorem 1.
The detailed proof is reported in Appendix B and related
appendices C, D, and E.
First, we use the fact (proved in [31]) that the entries of the
error matrix in (36) are nonnegative, implying, for i, j ∈ S:
(42)
N pN [ AS]ij = N pN aij + N pN eij ≥ N pN aij.
In view of Property 2, Eq. (42) implies that, when i and
j are interacting nodes, the quantity N pN [ AS]ij exceeds a
positive threshold τ with high probability and, hence, part
i) of Theorem 1 is proved. If, in addition, we show that
the magnified error N pN eij converges to zero in probability
over the non-interacting pairs, i.e., if we prove part ii), then
N pN(cid:104)(cid:98)AS(cid:105)ij → 0 (and, hence, N pN(cid:104)(cid:98)AS(cid:105)ij
stays below τ
as N goes to infinity) over the non-interacting pairs. Now, in
order to prove part iii), we need to show that it is possible
to classify correctly, as N → ∞, each pair of nodes by
comparing the truncated estimator (cid:98)AS against the threshold τ
(or any smaller value): if N pN [(cid:98)AS]ij > τ, then classify (i, j)
as an interacting pair, otherwise classify it as non-interacting.
Since the cardinality of the observable set is finite, parts i) and
ii) imply part iii) by direct application of the union bound.
Accordingly, it remains to examine part ii).
Using one result in [31], we rewrite the entries of the error
where B (cid:44) A2 and H (cid:44) (IS(cid:48) − BS(cid:48))−1. The error in (43)
is determined by three main factors, namely: i) ai(cid:96), which is
nonzero only if node i (from subset S) and agent (cid:96) (from subset
S(cid:48)) are neighbors; ii) bmj, which is nonzero only if node m
(from subnet S(cid:48)) and agent j (from subset S) are second-order
neighbors (i.e., connected in one or two steps); iii) the term
h(cid:96)m, which is the ((cid:96), m)-th entry of the matrix H. Clearly,
in (43), the relevant entries h(cid:96)m are those that are "activated"
by nonzero values of ai(cid:96) and bmj. The ((cid:96), m) pairs for which
ai(cid:96) and bmj are nonzero will be accordingly referred to as
"active pairs". Refer to Figure 5 for a graphical illustration of
the active pairs.
It is also clear from (43) that, in order to get a small error,
small values of h(cid:96)m (over the active pairs) are desirable. In
Theorem 2 -- see Appendix C -- we are able to show that,
for large N, vanishing values of h(cid:96)m are obtained when the
distance between nodes (cid:96) and m gets large. In particular,
Theorem 2 states that the distance between (cid:96) and m that
plays an important role in the magnitude of H is the one
evaluated on the transformed graph GS(cid:61)S (see Figure 6),
where the observable graph GS is replaced by an empty graph.
matrix in (36) as:
eij = (cid:88)(cid:96),m∈S(cid:48)
ai(cid:96)h(cid:96)mbmj,
i, j ∈ S,
(43)
11
Fig. 6. The matrix H in (65) does not depend on the combination submatrix
AS. If two nodes (cid:96), m ∈ S(cid:48) are distant in the graph GS(cid:61)S displayed on the
right, i.e., if δ(cid:96),m(GS(cid:61)S) (cid:29) 1, then h(cid:96)m is small. This does not imply, in
general, that δ(cid:96),m(G) (cid:29) 1, namely, that the same nodes are distant over the
(original) graph G displayed on the left.
Fig. 5. Graphical illustration of the active pairs. Note that the neighborhood
(G) refer to the partial Erdos-R´enyi
constraints (cid:96) ∈ Ni(G) and m ∈ N
graph G and hence the intermediate node k may belong to S.
(2)
j
network with high-probability. In particular,
detection threshold:
introducing a
η (cid:44) τ /(N pN ),
(44)
As observed in the proof of Theorem 2 the magnitude of h(cid:96)m
is not contingent on the particular topology GS. As a result,
removing the dependency on GS is crucial to get a universal
result, namely, a result that holds for any arbitrary GS.
Since, loosely speaking, Theorem 2 implies that the error
is small if the distance on the transformed graph GS(cid:61)S is
large, the remaining part of the proof consists of showing that
the distance over the active ((cid:96), m) pairs is large with high
probability, namely, that small distances are rare as N goes
to infinity. Now, such a conclusion can be proved for a pure
Erdos-R´enyi graph, G (cid:63)(N, pN ), as shown in Lemma 3 -- see
Appendix E. However, proving the same result for a partial
Erdos-R´enyi graph, G (cid:63)(N, pN , GS), presents a nontrivial dif-
ficulty related to the fact that the partial Erdos-R´enyi graph is
not homogeneous3 because the observable subgraph GS can be
arbitrary. In order to overcome this difficulty, we will carefully
implement a procedure of homogenization and coupling --
see Appendix D. The homogenization procedure amounts to
carefully replacing the partial Erdos-R´enyi random graph G
lowing sense: if small distances are rare over the classic (hence
by an Erdos-R´enyi graph (cid:101)G that is coupled with G in the fol-
homogeneous) Erdos-R´enyi graph (cid:101)G, then small distances
are also rare over the coupled partial (hence inhomogeneous)
Erdos-R´enyi graph G. In summary, the homogenization-and-
coupling is a useful formal tool that is used to reduce the
inhomogeneous case to a (simpler) homogeneous graph.
V. APPLYING THEOREM 1
Theorem 1 asserts the possibility of performing local to-
mography over large-scale diffusion networks as it asserts the
existence of a threshold τ such that the entries of the naıve
estimator (cid:98)AS provide correct reconstruction of the observable
3Actually, we will see in Appendix B that a further source of inhomogeneity
arises, which is related to the terms ai(cid:96) and bmj. The homogenization
procedure that we are going to introduce is able to solve this further
inhomogeneity.
the topology of the observable network can be recovered for
nodes i and j as interacting, otherwise classify them as non-
interacting. The inverse dependence of η on N pN can be
explained as follows. Since we compare the unnormalized
sufficiently large N as follows: If [(cid:98)AS]ij > η, then classify
matrix entries, [(cid:98)AS]ij, against the threshold η, and since in
view of Theorem 1 these entries (over the connected pairs)
scale essentially as 1/(N pN ), a proper classification threshold
should exhibit the same type of scaling.
From a practical perspective, it is necessary to select an
appropriate value for η,
in order to correctly classify the
interacting and non-interacting pairs. In this connection, prior
information on the dynamical system in (1) can be useful
to set the detection threshold. Indeed, we see from (44) that
knowledge is needed about: i) the average degree N pN ; and
ii) the parameter τ that characterizes the class Cρ,τ where the
combination matrix stems from. Let us assume that such a
knowledge is available. Now, using the values of τ reported
in (35), for the Laplacian and Metropolis rules we have,
respectively:
ηL =
ρλ
eN pN
,
ηM =
ρ
eN pN
.
(45)
We observe that Theorem 1 is an asymptotic (in the size N)
result. For a numerical practical application of this result, it is
useful to make three remarks.
First, given a detection threshold η that guarantees exact
asymptotic classification, any threshold smaller than η still
guarantees exact asymptotic classification. In order to explain
why, let us consider two thresholds η1 and η2, with η1 <
η2, and assume that η2 is known to provide exact asymptotic
classification. Then we have that: i) for interacting pairs, if
implying correct classification also with threshold η1; ii) for
non-interacting pairs, Theorem 1 ensures that, asymptotically,
[(cid:98)AS]ij is higher than η2, then it is obviously higher than η1,
[(cid:98)AS]ij will be smaller than any small value , implying correct
classification also with threshold η1. In other words, η1 also
provides exact classification.
Second, we observe that a combination rule can fulfill
Property 2 for different values of τ. For example, assume that
we proved that a combination rule fulfills Property 2 with a
certain value τ1. Then, the same policy fulfills Property 2 with
a higher value, e.g., τ2 > τ1.
Third, consider a pair (i, j) of interacting nodes, and let
us examine (42). According to Property 2, the selection of
τ relates only to the properties of the combination matrix,
namely, to the behavior of N pN aij for interacting nodes. On
the other hand, for finite sizes of the network, the error eij is
small, but not zero. As a result the quantity N pN [(cid:98)AS]ij will
be greater than N pN aij, namely, the entries of the estimated
combination matrix are, on average, shifted upward due to
this additional (and positive) error. As a result, one expects
that, for finite values of N, increasing the values of τ may be
beneficial for classification purposes.
The aforementioned issues show that there is freedom in
selecting the threshold parameter to attain exact
topology
recovery asymptotically (i.e., as N grows to infinity). On
the other hand, we remark that different threshold choices
are expected to behave differently for finite network sizes. In
fact, the following trade-off arises: a higher threshold reduces
the likelihood that a zero entry of the combination matrix
gives rise to a ( false) threshold crossing, while concurrently
increasing the likelihood that a nonzero entry gives rise to a
(correct) threshold crossing.
A. Nonparametric Strategies
From the analysis conducted in the previous section, we
have learned the following facts about
tomography based
on the thresholding operator. First, a good threshold tuning
requires some a-priori knowledge of the model (e.g., of the
average degree, N pN , or of the class of combination matrices
to set the constant τ). Second, even with some good a-priori
knowledge, it is not clear how the threshold should be opti-
mized to maximize the performance, because a trade-off arises
for finite network sizes, whose (nontrivial) solution would
require an even more detailed knowledge of the underlying
model.
For all these reasons, it is useful to verify the possibility of
employing some nonparametric pattern recognition strategies,
which work blindly (without any a-priori knowledge), and
which are capable to automatically adapt the classification
threshold to the empirical data. In particular, in the forth-
coming experiments we will consider a k-means clustering
algorithm (with k = 2) that will be fed with the entries of
the truncated estimator matrix in (8). The clustering algorithm
will attempt to find some separation threshold empirically
on the data, and to split accordingly the matrix entries into
two clusters (connected and non-connected). The cluster with
higher arithmetic mean will be then elected as the cluster of
connected nodes.
B. Unknown Correlation Matrices
Until now, we have implicitly assumed that the correlation
matrices, R0 and R1, are perfectly known, and, hence, that
12
the truncated estimator AS in (8) could be evaluated exactly.
However, in practice such correlation matrices are unknown,
and must be estimated from the data. For this reason, we will
consider numerical simulations where we empirically estimate
the truncated correlation matrices [R0]S and [R1]S from the
observed data through the sample-average estimator (boldface
notation is omitted to emphasize that the observed yn refers
to a particular realization):
(46)
(47)
1
nmax + 1
[yn]S[yn](cid:62)S ,
[yn+1]S[yn](cid:62)S .
(cid:100)[R0]S =
(cid:100)[R1]S =
nmax(cid:88)n=0
nmax−1(cid:88)n=0
1
nmax
We remark that it is possible to optimize such estimates by
exploiting prior information on the structural properties of the
system, and such an optimization could boost the performance
of the algorithm. This estimation-tuning is outside the scope
of this paper, but showing that our strategy works with the
(perhaps) simplest correlation estimators is definitely encour-
aging. In the next section, we will additionally display the
performance of the algorithm under the ideal case of known
correlations, where the exact computation of the truncated
estimator in (8) can be accomplished. This ideal case provides
a superior limit in performance also with respect to optimized
correlation estimators.
VI. NUMERICAL EXPERIMENTS
We are now ready to present the results of the numer-
ical experiments. In Figure 7, we display the (empirically-
estimated) topology-recovery probability, with reference to an
overall network with number of nodes N ranging from 10
to 200, and for the case of a Laplacian combination rule
with ρ = 0.9 and λ = 0.95. The observable network is
made up of S = 10 nodes and the connections within the
latent part (and between the latent and observable nodes) are
drawn according to an Erdos-R´enyi model with connection
probability pN = 2 log(N )/N. We see that the probability
of correct recovery gets close to 1 as N increases for all the
considered scenarios: parametric versus k-means thresholding,
and empirically estimated truncated correlation matrices (as
in (46) and (47)) versus known truncated correlation matrices.
We notice that the recovery probability curve is not monotonic.
Such behavior matches perfectly our theoretical results, as
we now explain. First, when N = S = 10 (first point in
Figure 7), all the network is observed, and in view of (6)
(and the comments that follow this equation) the recovery
probability must be equal to 1. Second, Theorem 1 ensures
that a probability of correct recovery equal to 1 must be also
attained asymptotically (in N). Accordingly, since the error
probability curve starts from the value 1, and converges to
1 as N increases, the non-monotonic behavior exhibited in
Figure 7 makes sense.
Let us now compare the performance of the different
strategies shown in Figure 7. As one expects, the strategies that
know the true correlation matrices outperform the strategies
that do not know them. A separate comment is called for
13
Fig. 7. Empirical recovery probability as a function of the size of the overall
network, for the Laplacian combination rule -- see (28). Thresholding stands
for the tomography strategy where the entries of (cid:98)AS are thresholded with
the threshold ηL determined in (45). Empirical correlation means that the
truncated correlations were estimated as in (46) and (47), whereas the curves
with known correlations are also displayed as a superior limit in performance.
Fig. 8. Empirical recovery probability as a function of the size of the overall
network, for the Metropolis combination rule -- see (29). Thresholding stands
for the tomography strategy where the entries of (cid:98)AS are thresholded with
the threshold ηM determined in (45). Empirical correlation means that the
truncated correlations were estimated as in (46) and (47), whereas the curves
with known correlations are also displayed as a superior limit in performance.
while comparing the thresholding estimator and the clustering
algorithm. Perhaps unexpectedly, the latter strategy outper-
forms the former one. However, this behavior matches well
the theoretical considerations made in the previous section.
Indeed,
in the simulations the threshold employed by the
thresholding estimator is not optimized at all, whereas the
nonparametric clustering algorithm might automatically adapt
the threshold to the empirical evidence arising from the data,
thus delivering a better performance.
The above analysis is repeated for the case of a Metropolis
combination rule with ρ = 0.9, and the result
is shown
in Figure 8. The same general conclusions that we draw
for the Laplacian rule apply. However, we see here that the
performance of the thresholding operator seems slightly worse.
One explanation for this behavior is the following. The choice
of the constant τ in (35) is perhaps over-conservative. Indeed,
such choice follows by estimating the asymptotic scaling law
of the maximal degree (see Lemma 2), whereas the nonzero
entries of the Metropolis matrix in (29) are determined only
by the maximum over pairs of degrees. This means that,
on average, the nonzero entries of the Metropolis matrix are
higher than what is predicted by the chosen τ. It is expected
that for the Metropolis rule, a larger threshold can be used
without affecting the identification of connected pairs, while
reducing the errors corresponding to the disconnected pairs.
A. Beyond Theorem 1
Theorem 1 establishes that, under certain technical condi-
tions, it is possible to retrieve the topology of a subset S, even
when the majority of the network nodes are not observed.
This appears to be a nontrivial result, since an observable
measurement yi(n), i ∈ S, is subject to the influence of
nodes from all across the network. This happens because the
diffusion recursion in (4) links the nodes through the overall
N × N matrix A, which takes into account also the influences
that unobserved nodes have on observed nodes.
The possibility of inferring the topology of a subnet by
taking measurements from this subnet only, and by ignoring
the unobserved components, is of paramount importance, in
view of the accessability, probing and processing limitations
that arise unavoidably in practical applications. In particular,
it
is tempting to think about a sequential reconstruction
strategy, where a larger network is reconstructed through local
tomography experiments over smaller network portions, and
where each local experiment obeys some probing/processing
constraints. We start by illustrating the perhaps simplest se-
quential reconstruction strategy.
Assume that
there is an observable subset of nodes S,
which is embedded in a larger network with many unobserved
components. Due to probing and processing limitations, it is
possible to probe and process at most M nodes per single
experiment. Accordingly, the set S is divided into the "patches"
S1, S2, . . . , SP . For simplicity, we assume that the patches do
not overlap each other and that they cover completely the
set S (i.e., the patches form a partition of S). Each local
tomography experiment will correspond to probing the union
of two patches, Si ∪ Sj. For this reason, each patch has car-
dinality Si ≤ M/2, for i = 1, 2, . . . , P , which allows coping
with the probing-and-processing constraint. The process is
pictorially illustrated in Figure 9. Clearly, if a particular pair of
nodes does not belong to the union of patches under test, we
cannot make inference on that pair. The maximum number of
experiments that allows testing all pair of nodes is P (P −1)/2.
Per each experiment, we apply the local tomography strategy
described in the previous section, namely: i) we compute the
empirical correlation matrices, (cid:100)[R0]Si∪Sj and (cid:100)[R1]Si∪Sj , from
which ii) the truncated estimator (cid:98)ASi∪Sj
is computed; and
iii) we apply the k-means algorithm to obtain an estimated
204060801001201401601802000.50.550.60.650.70.750.80.850.90.951Laplacian rulethresholding (known corr.)k-means clust. (known corr.)thresholding (emp. corr.)k-means clust. (emp. corr.)204060801001201401601802000.50.550.60.650.70.750.80.850.90.951Metropolis rulethresholding (known corr.)k-means clust. (known corr.)thresholding (emp. corr.)k-means clust. (emp. corr.) 14
Algorithm 1 Patch&Catch Sequential Tomography
Input: Ensemble of patches {S1, S2, . . . , SP} and observ-
ables {[yn]Si} over the patches i = 1, 2, . . . , P , and for
time epochs n = 0, 1, . . . , nmax.
Output: (cid:98)GS, estimate of the subnet GS of observable nodes.
1: while i ≤ P do
2:
nmax+1(cid:80)nmax
(cid:100)[R0]Si∪Sj
3:
nmax(cid:80)nmax−1
(cid:100)[R1]Si∪Sj
(cid:98)ASi∪Sj = (cid:100)[R1]Si∪Sj(cid:16)(cid:100)[R0]Si∪Sj(cid:17)−1
(cid:98)GSi∪Sj = k-means(cid:16)(cid:98)ASi∪Sj(cid:17)
while j < i do
=
= 1
n=0 [yn]Si∪Sj [yn](cid:62)Si∪Sj
[yn+1]Si∪Sj [yn](cid:62)Si∪Sj
j = j + 1
n=0
1
4:
5:
6:
7:
8:
9:
10:
11: end while
end while
j = 1
i = i + 1
Fig. 9. Pictorial illustration of the sequential reconstruction through patching.
run the Patch&Catch algorithm in a sub-region S, assuming a
strict probing constraint of M = 10 nodes per experiment.
subgraph GSi∪Sj . As more and more pairs of patches are
tested, the connection profile of the network is reconstructed.
A pseudo-code for the sequential reconstruction algorithm,
nicknamed "Patch&Catch", is shown in Algorithm 1.
Before seeing the Patch&Catch algorithm in operation, it
is important to make a fundamental remark. Proving that the
sequential reconstruction strategy retrieves the topology of GS
consistently (as N → ∞) seems a nontrivial task. In partic-
ular, we now explain why consistency of the Patch&Catch
algorithm does not come as a corollary of Theorem 1.
Assume first that GS is an arbitrary deterministic network.
In order to apply Theorem 1 to each local experiment,
the unobserved component should be an Erdos-R´enyi graph.
However, given a union-of-patches under test, Si ∪ Sj, the
unobserved component is a mix of an Erdos-R´enyi graph and
of a portion of GS (refer to Figure 9). Since the latter portion is
not purely Erdos-R´enyi (because GS is arbitrary), Theorem 1
does not directly apply. On the other hand, if we assume that
the whole graph (and, hence, also GS) is Erdos-R´enyi, then
the network GS would be not fixed as N → ∞. In particular,
since S has finite cardinality, it will become asymptotically
disconnected, with high probability as N → ∞.
In summary, we make no claim that the sequential recon-
struction can grant consistent recovery. Therefore, the numer-
ical results we are going to illustrate in this subsection must
be intended as a preliminary test aimed at checking whether,
in the finite network-size regime, a sequential reconstruction
strategy might be successfully applicable.
We are now ready to see an application of the Patch&Catch
algorithm. The overall network is made up of N = 300 nodes,
and is generated according to an Erdos-R´enyi graph with
probability of connection pN = 5(log N )/N. The combination
matrix A is obtained via the Metropolis rule, and the system
is observed over a time scale of nmax = 105 samples. We
In Figure 10 we consider a subset S of cardinality S = 20,
and we display the evolution of the algorithm for an increasing
number of tested patches. Since M = 10, and choosing equal-
sized patches, we get P = S/(M/2) = 4 patches. For each
experiment, we depict the true graph of connections (blue
edges), as well as the overall graph of connections estimated
up to the current experiment (red edges). The network nodes
that form the patches tested in the single experiment are
highlighted in red. We see from Figure 10 that the network
is faithfully reconstructed, sequentially as the number of
experiments grows, until the complete subnetwork topology
is correctly retrieved after P (P − 1)/2 = 6 experiments.
In Figure 11, the same procedure is applied to a larger subset
with S = 60. For this case, we illustrate the performance deliv-
ered by the Patch&Catch algorithm in a more quantitative way.
More precisely, we display the evolution, as more experiments
are performed, of the normalized distance between the true
(48)
graph GS and the estimated graph (cid:98)GS, namely:
gij −(cid:98)gij,
S(S − 1) (cid:88)i,j∈S,i<j
dist(GS,(cid:98)GS) (cid:44)
2
and we assume that initially the estimated graph has no edges.
the aforementioned distance
We see from Figure 11 that
exhibits a desirable decreasing behavior:
the discrepancies
between the true graph and the estimated graph diminish
progressively as more experiments are conducted.
Before concluding this section, it is useful to comment
on two important aspects. First, the algorithm can easily be
generalized to account for overlapping patches. This would
simply require to set a tie-break rule for managing the case
where a particular pair, say (h, k), is present in two distinct
experiments. Since usually the connection probability is small,
one meaningful rule could be an AND rule, where the pair
(h, k) is labeled as connected only if so they are in both
experiments. The simplest tie-break rule might be retaining
just the first classification of one pair of nodes, ignoring the
observable subnet 6theindicatorfunction,whichisequaltooneifconditionEistrue,andisequaltozero,otherwise.TheN⇥NidentitymatrixisdenotedbyI.ThecomplementofasetSisdenotedbyS0.GivenanN⇥NmatrixZ,thesubmatrixthatliesintherowsofZindexedbythesetS✓{1,2,...,N}andinthecolumnsindexedbythesetT✓{1,2,...,N},willbedenotedbyZST,oralternativelyby[Z]ST.WhenS=T,thesubmatrixZSTwillbeabbreviatedasZS.Intheindexingofthesubmatrixwewillretaintheindexsetoftheoriginalmatrix.Forexample,ifS={2,3}andT={2,4,5},wehavethatthesubmatrixM=ZSTisa2⇥3matrix,indexedasfollows:M=✓z22z24z25z32z34z35◆=✓m22m24m25m32m34m35◆.(15)Thisnotationiscrucialinourtreatment,sinceitwillallowustoidentifynodeswithoutcumbersomeandredundantdouble-indexnotation.B.GraphnotationG(V)isthesetofundirectedgraphsdefinedonasetofnodes(vertexset)V.WhenNisthenumberofnodes,theshortcutG(N)impliesthatthevertexsetisV={1,2,...,N}.WhendealingwithagraphG2G(N),itsconnectionstructure(i.e.,theedgesofthegraphs)canbedescribedthroughitsN⇥Nadjacencymatrix.The(i,j)-thentryoftheadjacencymatrixofgraphGwillbedenotedbygij,withgij=1ifnodesiandjareconnected,andgij=0,otherwise.Self-loops(gii=1)willbepermitted.GivenG2G(N),andasubsetS✓{1,2,...,N},thesubgraphcorrespondingtoSisdenotedbyGS2G(S).ThesupportgraphofamatrixAisdenotedbyG(A).The(i,j)-thentryofitsadjacencymatrixisI{aij>0},namely,nodesiandjareconnectedonG(A)if,andonly,ifaijisstrictlypositive.Apathfromitojisasequenceofedgeswherethefirstedgeoriginatesfromiandthelastedgeterminatesatj.Theexistenceofapathoflengthrcanbeexpressedas:gin1gn1n2...gnr 1j=1,(16)foracertainsequenceofverticesn1,n2,...,nr 1belongingtoV.Accordingtothisdefinition,apathcanalsopassmultipletimesthroughthesamenode,orcanlingerforoneormorestepsatthesamenodewhenithasaself-loop.Ni(G)isthesetofneighborsofnodei(includingiitself)intheundirectedgraphG.ThedegreeofnodeiisthecardinalityofNi(G).dmax(G)isthemaximumdegreeinG. i,j(G)isthedistancebetweenthenodesiandjonthegraphG,i.e.,thelengthoftheshortestpathlinkingiandj.N(r)i(G)isther-thorderneighborhoodofnodei(includingiitself),givenbyN(r)i(G)={j2V: i,j(G)r}.(17)ArandomgraphGobeystheErdos-R´enyimodelifeachedgeofGisdrawn,independentlyfromtheotheredges,withidenticalprobabilitypN.Equivalentlystated,theadjacencyrandomvariablesgij,fori=1,2,...,Nandi<j,areindependentandidenticallydistributed(i.i.d.)Bernoullivariables.ThenotationG⇠G?(N,pN)signifiesthatthegraphGbelongstotheErdos-R´enyiclasswithconnectionprobabilitythatvanishesasN!1,andthatobeysthefollowingscalinglaw:pN=logN+cNN,(18)wherecN!1asN!1(inanarbitraryway,providedthatpN!0).Itisawell-knownresultthatrandomgraphsbelongingtothefamilyG?(N,pN)areconnectedwithhighprobability[21].Beforeproceeding,andasanoteofclarity,weassumethatallrandomvariablesfinddomaininthecommonprobabilityspace(⌦,F,P),where⌦isthesetofrealizations,Fisthesigma-algebraofmeasurablesetsandPistheprobabilitymeasure.Forinstance,theevent{!2⌦: i,j(G(!))r}2F,(19)representsthesetofrealizations!2⌦yieldingagraphG(!)whosedistancebetweenthe(prefixed)nodesiandjdoesnotexceedr.Torenderamorecompactnotation,wehenceforthomittherealization!inthecharacterizationoftheevents.Forinstance,inthecaseoftheevent(19)werepresentitratheras{ i,j(G)r}2F,(20)wheretherandomquantitiesareemphasizedbytheboldfacednotation -- intheeventinequation(20),theonlyrandomobjectisthegraphG.C.UsefulGraphOperationsInourexposition,wewillbeperformingcertainoperationsovergraphs,aswellasevaluatecertainfunctionssuchascom-paringdistancesbetweennodesoverdistinctgraphs.There-fore,itisusefultointroducethefollowinggraphoperationsforlateruse(whichareillustratedinFigure2):1)Graphembedding.GivenavertexsetV,andasubsetthereof,S⇢V,theembeddingofagraphG(1)2G(S)intothelargergraphG(2)2G(V)willbedenotedby:G=G(1) G(2),G2G(V),(21)andresultsinagraphwiththefollowingproperties:i)theconnectionsbetweennodesinSthatarepresentinG(2)arecancelled;ii)thenodesinthevertexsetSofgraphG(1)aremappedintothecorrespondingnodesofgraphG(2),andsoarethepertinentconnections.WestressthattheconnectionsfromS0toSaredeterminedbythegraphG(2).Wenoticethattheoperationin(21)isnotcommutative(becausethefirstgraphisembeddedintothesecondgraph,andnotviceversa),andthattheoutputgraphGdoesnotdependontheconnectionsexistinginG(2)amongnodesbelongingtothesetS.2)Localdisconnection.GivenagraphG2G(V),thenotation:GU1=U22G(V),(22)describesthegraphthatisfromGafterremovingalltheedgesthatconnectnodesinU1tonodesinU2,namely,alltheconnectionsbetweenU1andU2.6theindicatorfunction,whichisequaltooneifconditionEistrue,andisequaltozero,otherwise.TheN⇥NidentitymatrixisdenotedbyI.ThecomplementofasetSisdenotedbyS0.GivenanN⇥NmatrixZ,thesubmatrixthatliesintherowsofZindexedbythesetS✓{1,2,...,N}andinthecolumnsindexedbythesetT✓{1,2,...,N},willbedenotedbyZST,oralternativelyby[Z]ST.WhenS=T,thesubmatrixZSTwillbeabbreviatedasZS.Intheindexingofthesubmatrixwewillretaintheindexsetoftheoriginalmatrix.Forexample,ifS={2,3}andT={2,4,5},wehavethatthesubmatrixM=ZSTisa2⇥3matrix,indexedasfollows:M=✓z22z24z25z32z34z35◆=✓m22m24m25m32m34m35◆.(15)Thisnotationiscrucialinourtreatment,sinceitwillallowustoidentifynodeswithoutcumbersomeandredundantdouble-indexnotation.B.GraphnotationG(V)isthesetofundirectedgraphsdefinedonasetofnodes(vertexset)V.WhenNisthenumberofnodes,theshortcutG(N)impliesthatthevertexsetisV={1,2,...,N}.WhendealingwithagraphG2G(N),itsconnectionstructure(i.e.,theedgesofthegraphs)canbedescribedthroughitsN⇥Nadjacencymatrix.The(i,j)-thentryoftheadjacencymatrixofgraphGwillbedenotedbygij,withgij=1ifnodesiandjareconnected,andgij=0,otherwise.Self-loops(gii=1)willbepermitted.GivenG2G(N),andasubsetS✓{1,2,...,N},thesubgraphcorrespondingtoSisdenotedbyGS2G(S).ThesupportgraphofamatrixAisdenotedbyG(A).The(i,j)-thentryofitsadjacencymatrixisI{aij>0},namely,nodesiandjareconnectedonG(A)if,andonly,ifaijisstrictlypositive.Apathfromitojisasequenceofedgeswherethefirstedgeoriginatesfromiandthelastedgeterminatesatj.Theexistenceofapathoflengthrcanbeexpressedas:gin1gn1n2...gnr 1j=1,(16)foracertainsequenceofverticesn1,n2,...,nr 1belongingtoV.Accordingtothisdefinition,apathcanalsopassmultipletimesthroughthesamenode,orcanlingerforoneormorestepsatthesamenodewhenithasaself-loop.Ni(G)isthesetofneighborsofnodei(includingiitself)intheundirectedgraphG.ThedegreeofnodeiisthecardinalityofNi(G).dmax(G)isthemaximumdegreeinG. i,j(G)isthedistancebetweenthenodesiandjonthegraphG,i.e.,thelengthoftheshortestpathlinkingiandj.N(r)i(G)isther-thorderneighborhoodofnodei(includingiitself),givenbyN(r)i(G)={j2V: i,j(G)r}.(17)ArandomgraphGobeystheErdos-R´enyimodelifeachedgeofGisdrawn,independentlyfromtheotheredges,withidenticalprobabilitypN.Equivalentlystated,theadjacencyrandomvariablesgij,fori=1,2,...,Nandi<j,areindependentandidenticallydistributed(i.i.d.)Bernoullivariables.ThenotationG⇠G?(N,pN)signifiesthatthegraphGbelongstotheErdos-R´enyiclasswithconnectionprobabilitythatvanishesasN!1,andthatobeysthefollowingscalinglaw:pN=logN+cNN,(18)wherecN!1asN!1(inanarbitraryway,providedthatpN!0).Itisawell-knownresultthatrandomgraphsbelongingtothefamilyG?(N,pN)areconnectedwithhighprobability[21].Beforeproceeding,andasanoteofclarity,weassumethatallrandomvariablesfinddomaininthecommonprobabilityspace(⌦,F,P),where⌦isthesetofrealizations,Fisthesigma-algebraofmeasurablesetsandPistheprobabilitymeasure.Forinstance,theevent{!2⌦: i,j(G(!))r}2F,(19)representsthesetofrealizations!2⌦yieldingagraphG(!)whosedistancebetweenthe(prefixed)nodesiandjdoesnotexceedr.Torenderamorecompactnotation,wehenceforthomittherealization!inthecharacterizationoftheevents.Forinstance,inthecaseoftheevent(19)werepresentitratheras{ i,j(G)r}2F,(20)wheretherandomquantitiesareemphasizedbytheboldfacednotation -- intheeventinequation(20),theonlyrandomobjectisthegraphG.C.UsefulGraphOperationsInourexposition,wewillbeperformingcertainoperationsovergraphs,aswellasevaluatecertainfunctionssuchascom-paringdistancesbetweennodesoverdistinctgraphs.There-fore,itisusefultointroducethefollowinggraphoperationsforlateruse(whichareillustratedinFigure2):1)Graphembedding.GivenavertexsetV,andasubsetthereof,S⇢V,theembeddingofagraphG(1)2G(S)intothelargergraphG(2)2G(V)willbedenotedby:G=G(1) G(2),G2G(V),(21)andresultsinagraphwiththefollowingproperties:i)theconnectionsbetweennodesinSthatarepresentinG(2)arecancelled;ii)thenodesinthevertexsetSofgraphG(1)aremappedintothecorrespondingnodesofgraphG(2),andsoarethepertinentconnections.WestressthattheconnectionsfromS0toSaredeterminedbythegraphG(2).Wenoticethattheoperationin(21)isnotcommutative(becausethefirstgraphisembeddedintothesecondgraph,andnotviceversa),andthattheoutputgraphGdoesnotdependontheconnectionsexistinginG(2)amongnodesbelongingtothesetS.2)Localdisconnection.GivenagraphG2G(V),thenotation:GU1=U22G(V),(22)describesthegraphthatisfromGafterremovingalltheedgesthatconnectnodesinU1tonodesinU2,namely,alltheconnectionsbetweenU1andU2.1 6theindicatorfunction,whichisequaltooneifconditionEistrue,andisequaltozero,otherwise.TheN⇥NidentitymatrixisdenotedbyI.ThecomplementofasetSisdenotedbyS0.GivenanN⇥NmatrixZ,thesubmatrixthatliesintherowsofZindexedbythesetS✓{1,2,...,N}andinthecolumnsindexedbythesetT✓{1,2,...,N},willbedenotedbyZST,oralternativelyby[Z]ST.WhenS=T,thesubmatrixZSTwillbeabbreviatedasZS.Intheindexingofthesubmatrixwewillretaintheindexsetoftheoriginalmatrix.Forexample,ifS={2,3}andT={2,4,5},wehavethatthesubmatrixM=ZSTisa2⇥3matrix,indexedasfollows:M=✓z22z24z25z32z34z35◆=✓m22m24m25m32m34m35◆.(15)Thisnotationiscrucialinourtreatment,sinceitwillallowustoidentifynodeswithoutcumbersomeandredundantdouble-indexnotation.B.GraphnotationG(V)isthesetofundirectedgraphsdefinedonasetofnodes(vertexset)V.WhenNisthenumberofnodes,theshortcutG(N)impliesthatthevertexsetisV={1,2,...,N}.WhendealingwithagraphG2G(N),itsconnectionstructure(i.e.,theedgesofthegraphs)canbedescribedthroughitsN⇥Nadjacencymatrix.The(i,j)-thentryoftheadjacencymatrixofgraphGwillbedenotedbygij,withgij=1ifnodesiandjareconnected,andgij=0,otherwise.Self-loops(gii=1)willbepermitted.GivenG2G(N),andasubsetS✓{1,2,...,N},thesubgraphcorrespondingtoSisdenotedbyGS2G(S).ThesupportgraphofamatrixAisdenotedbyG(A).The(i,j)-thentryofitsadjacencymatrixisI{aij>0},namely,nodesiandjareconnectedonG(A)if,andonly,ifaijisstrictlypositive.Apathfromitojisasequenceofedgeswherethefirstedgeoriginatesfromiandthelastedgeterminatesatj.Theexistenceofapathoflengthrcanbeexpressedas:gin1gn1n2...gnr 1j=1,(16)foracertainsequenceofverticesn1,n2,...,nr 1belongingtoV.Accordingtothisdefinition,apathcanalsopassmultipletimesthroughthesamenode,orcanlingerforoneormorestepsatthesamenodewhenithasaself-loop.Ni(G)isthesetofneighborsofnodei(includingiitself)intheundirectedgraphG.ThedegreeofnodeiisthecardinalityofNi(G).dmax(G)isthemaximumdegreeinG. i,j(G)isthedistancebetweenthenodesiandjonthegraphG,i.e.,thelengthoftheshortestpathlinkingiandj.N(r)i(G)isther-thorderneighborhoodofnodei(includingiitself),givenbyN(r)i(G)={j2V: i,j(G)r}.(17)ArandomgraphGobeystheErdos-R´enyimodelifeachedgeofGisdrawn,independentlyfromtheotheredges,withidenticalprobabilitypN.Equivalentlystated,theadjacencyrandomvariablesgij,fori=1,2,...,Nandi<j,areindependentandidenticallydistributed(i.i.d.)Bernoullivariables.ThenotationG⇠G?(N,pN)signifiesthatthegraphGbelongstotheErdos-R´enyiclasswithconnectionprobabilitythatvanishesasN!1,andthatobeysthefollowingscalinglaw:pN=logN+cNN,(18)wherecN!1asN!1(inanarbitraryway,providedthatpN!0).Itisawell-knownresultthatrandomgraphsbelongingtothefamilyG?(N,pN)areconnectedwithhighprobability[21].Beforeproceeding,andasanoteofclarity,weassumethatallrandomvariablesfinddomaininthecommonprobabilityspace(⌦,F,P),where⌦isthesetofrealizations,Fisthesigma-algebraofmeasurablesetsandPistheprobabilitymeasure.Forinstance,theevent{!2⌦: i,j(G(!))r}2F,(19)representsthesetofrealizations!2⌦yieldingagraphG(!)whosedistancebetweenthe(prefixed)nodesiandjdoesnotexceedr.Torenderamorecompactnotation,wehenceforthomittherealization!inthecharacterizationoftheevents.Forinstance,inthecaseoftheevent(19)werepresentitratheras{ i,j(G)r}2F,(20)wheretherandomquantitiesareemphasizedbytheboldfacednotation -- intheeventinequation(20),theonlyrandomobjectisthegraphG.C.UsefulGraphOperationsInourexposition,wewillbeperformingcertainoperationsovergraphs,aswellasevaluatecertainfunctionssuchascom-paringdistancesbetweennodesoverdistinctgraphs.There-fore,itisusefultointroducethefollowinggraphoperationsforlateruse(whichareillustratedinFigure2):1)Graphembedding.GivenavertexsetV,andasubsetthereof,S⇢V,theembeddingofagraphG(1)2G(S)intothelargergraphG(2)2G(V)willbedenotedby:G=G(1) G(2),G2G(V),(21)andresultsinagraphwiththefollowingproperties:i)theconnectionsbetweennodesinSthatarepresentinG(2)arecancelled;ii)thenodesinthevertexsetSofgraphG(1)aremappedintothecorrespondingnodesofgraphG(2),andsoarethepertinentconnections.WestressthattheconnectionsfromS0toSaredeterminedbythegraphG(2).Wenoticethattheoperationin(21)isnotcommutative(becausethefirstgraphisembeddedintothesecondgraph,andnotviceversa),andthattheoutputgraphGdoesnotdependontheconnectionsexistinginG(2)amongnodesbelongingtothesetS.2)Localdisconnection.GivenagraphG2G(V),thenotation:GU1=U22G(V),(22)describesthegraphthatisfromGafterremovingalltheedgesthatconnectnodesinU1tonodesinU2,namely,alltheconnectionsbetweenU1andU2.2 6theindicatorfunction,whichisequaltooneifconditionEistrue,andisequaltozero,otherwise.TheN⇥NidentitymatrixisdenotedbyI.ThecomplementofasetSisdenotedbyS0.GivenanN⇥NmatrixZ,thesubmatrixthatliesintherowsofZindexedbythesetS✓{1,2,...,N}andinthecolumnsindexedbythesetT✓{1,2,...,N},willbedenotedbyZST,oralternativelyby[Z]ST.WhenS=T,thesubmatrixZSTwillbeabbreviatedasZS.Intheindexingofthesubmatrixwewillretaintheindexsetoftheoriginalmatrix.Forexample,ifS={2,3}andT={2,4,5},wehavethatthesubmatrixM=ZSTisa2⇥3matrix,indexedasfollows:M=✓z22z24z25z32z34z35◆=✓m22m24m25m32m34m35◆.(15)Thisnotationiscrucialinourtreatment,sinceitwillallowustoidentifynodeswithoutcumbersomeandredundantdouble-indexnotation.B.GraphnotationG(V)isthesetofundirectedgraphsdefinedonasetofnodes(vertexset)V.WhenNisthenumberofnodes,theshortcutG(N)impliesthatthevertexsetisV={1,2,...,N}.WhendealingwithagraphG2G(N),itsconnectionstructure(i.e.,theedgesofthegraphs)canbedescribedthroughitsN⇥Nadjacencymatrix.The(i,j)-thentryoftheadjacencymatrixofgraphGwillbedenotedbygij,withgij=1ifnodesiandjareconnected,andgij=0,otherwise.Self-loops(gii=1)willbepermitted.GivenG2G(N),andasubsetS✓{1,2,...,N},thesubgraphcorrespondingtoSisdenotedbyGS2G(S).ThesupportgraphofamatrixAisdenotedbyG(A).The(i,j)-thentryofitsadjacencymatrixisI{aij>0},namely,nodesiandjareconnectedonG(A)if,andonly,ifaijisstrictlypositive.Apathfromitojisasequenceofedgeswherethefirstedgeoriginatesfromiandthelastedgeterminatesatj.Theexistenceofapathoflengthrcanbeexpressedas:gin1gn1n2...gnr 1j=1,(16)foracertainsequenceofverticesn1,n2,...,nr 1belongingtoV.Accordingtothisdefinition,apathcanalsopassmultipletimesthroughthesamenode,orcanlingerforoneormorestepsatthesamenodewhenithasaself-loop.Ni(G)isthesetofneighborsofnodei(includingiitself)intheundirectedgraphG.ThedegreeofnodeiisthecardinalityofNi(G).dmax(G)isthemaximumdegreeinG. i,j(G)isthedistancebetweenthenodesiandjonthegraphG,i.e.,thelengthoftheshortestpathlinkingiandj.N(r)i(G)isther-thorderneighborhoodofnodei(includingiitself),givenbyN(r)i(G)={j2V: i,j(G)r}.(17)ArandomgraphGobeystheErdos-R´enyimodelifeachedgeofGisdrawn,independentlyfromtheotheredges,withidenticalprobabilitypN.Equivalentlystated,theadjacencyrandomvariablesgij,fori=1,2,...,Nandi<j,areindependentandidenticallydistributed(i.i.d.)Bernoullivariables.ThenotationG⇠G?(N,pN)signifiesthatthegraphGbelongstotheErdos-R´enyiclasswithconnectionprobabilitythatvanishesasN!1,andthatobeysthefollowingscalinglaw:pN=logN+cNN,(18)wherecN!1asN!1(inanarbitraryway,providedthatpN!0).Itisawell-knownresultthatrandomgraphsbelongingtothefamilyG?(N,pN)areconnectedwithhighprobability[21].Beforeproceeding,andasanoteofclarity,weassumethatallrandomvariablesfinddomaininthecommonprobabilityspace(⌦,F,P),where⌦isthesetofrealizations,Fisthesigma-algebraofmeasurablesetsandPistheprobabilitymeasure.Forinstance,theevent{!2⌦: i,j(G(!))r}2F,(19)representsthesetofrealizations!2⌦yieldingagraphG(!)whosedistancebetweenthe(prefixed)nodesiandjdoesnotexceedr.Torenderamorecompactnotation,wehenceforthomittherealization!inthecharacterizationoftheevents.Forinstance,inthecaseoftheevent(19)werepresentitratheras{ i,j(G)r}2F,(20)wheretherandomquantitiesareemphasizedbytheboldfacednotation -- intheeventinequation(20),theonlyrandomobjectisthegraphG.C.UsefulGraphOperationsInourexposition,wewillbeperformingcertainoperationsovergraphs,aswellasevaluatecertainfunctionssuchascom-paringdistancesbetweennodesoverdistinctgraphs.There-fore,itisusefultointroducethefollowinggraphoperationsforlateruse(whichareillustratedinFigure2):1)Graphembedding.GivenavertexsetV,andasubsetthereof,S⇢V,theembeddingofagraphG(1)2G(S)intothelargergraphG(2)2G(V)willbedenotedby:G=G(1) G(2),G2G(V),(21)andresultsinagraphwiththefollowingproperties:i)theconnectionsbetweennodesinSthatarepresentinG(2)arecancelled;ii)thenodesinthevertexsetSofgraphG(1)aremappedintothecorrespondingnodesofgraphG(2),andsoarethepertinentconnections.WestressthattheconnectionsfromS0toSaredeterminedbythegraphG(2).Wenoticethattheoperationin(21)isnotcommutative(becausethefirstgraphisembeddedintothesecondgraph,andnotviceversa),andthattheoutputgraphGdoesnotdependontheconnectionsexistinginG(2)amongnodesbelongingtothesetS.2)Localdisconnection.GivenagraphG2G(V),thenotation:GU1=U22G(V),(22)describesthegraphthatisfromGafterremovingalltheedgesthatconnectnodesinU1tonodesinU2,namely,alltheconnectionsbetweenU1andU2.4 6theindicatorfunction,whichisequaltooneifconditionEistrue,andisequaltozero,otherwise.TheN⇥NidentitymatrixisdenotedbyI.ThecomplementofasetSisdenotedbyS0.GivenanN⇥NmatrixZ,thesubmatrixthatliesintherowsofZindexedbythesetS✓{1,2,...,N}andinthecolumnsindexedbythesetT✓{1,2,...,N},willbedenotedbyZST,oralternativelyby[Z]ST.WhenS=T,thesubmatrixZSTwillbeabbreviatedasZS.Intheindexingofthesubmatrixwewillretaintheindexsetoftheoriginalmatrix.Forexample,ifS={2,3}andT={2,4,5},wehavethatthesubmatrixM=ZSTisa2⇥3matrix,indexedasfollows:M=✓z22z24z25z32z34z35◆=✓m22m24m25m32m34m35◆.(15)Thisnotationiscrucialinourtreatment,sinceitwillallowustoidentifynodeswithoutcumbersomeandredundantdouble-indexnotation.B.GraphnotationG(V)isthesetofundirectedgraphsdefinedonasetofnodes(vertexset)V.WhenNisthenumberofnodes,theshortcutG(N)impliesthatthevertexsetisV={1,2,...,N}.WhendealingwithagraphG2G(N),itsconnectionstructure(i.e.,theedgesofthegraphs)canbedescribedthroughitsN⇥Nadjacencymatrix.The(i,j)-thentryoftheadjacencymatrixofgraphGwillbedenotedbygij,withgij=1ifnodesiandjareconnected,andgij=0,otherwise.Self-loops(gii=1)willbepermitted.GivenG2G(N),andasubsetS✓{1,2,...,N},thesubgraphcorrespondingtoSisdenotedbyGS2G(S).ThesupportgraphofamatrixAisdenotedbyG(A).The(i,j)-thentryofitsadjacencymatrixisI{aij>0},namely,nodesiandjareconnectedonG(A)if,andonly,ifaijisstrictlypositive.Apathfromitojisasequenceofedgeswherethefirstedgeoriginatesfromiandthelastedgeterminatesatj.Theexistenceofapathoflengthrcanbeexpressedas:gin1gn1n2...gnr 1j=1,(16)foracertainsequenceofverticesn1,n2,...,nr 1belongingtoV.Accordingtothisdefinition,apathcanalsopassmultipletimesthroughthesamenode,orcanlingerforoneormorestepsatthesamenodewhenithasaself-loop.Ni(G)isthesetofneighborsofnodei(includingiitself)intheundirectedgraphG.ThedegreeofnodeiisthecardinalityofNi(G).dmax(G)isthemaximumdegreeinG. i,j(G)isthedistancebetweenthenodesiandjonthegraphG,i.e.,thelengthoftheshortestpathlinkingiandj.N(r)i(G)isther-thorderneighborhoodofnodei(includingiitself),givenbyN(r)i(G)={j2V: i,j(G)r}.(17)ArandomgraphGobeystheErdos-R´enyimodelifeachedgeofGisdrawn,independentlyfromtheotheredges,withidenticalprobabilitypN.Equivalentlystated,theadjacencyrandomvariablesgij,fori=1,2,...,Nandi<j,areindependentandidenticallydistributed(i.i.d.)Bernoullivariables.ThenotationG⇠G?(N,pN)signifiesthatthegraphGbelongstotheErdos-R´enyiclasswithconnectionprobabilitythatvanishesasN!1,andthatobeysthefollowingscalinglaw:pN=logN+cNN,(18)wherecN!1asN!1(inanarbitraryway,providedthatpN!0).Itisawell-knownresultthatrandomgraphsbelongingtothefamilyG?(N,pN)areconnectedwithhighprobability[21].Beforeproceeding,andasanoteofclarity,weassumethatallrandomvariablesfinddomaininthecommonprobabilityspace(⌦,F,P),where⌦isthesetofrealizations,Fisthesigma-algebraofmeasurablesetsandPistheprobabilitymeasure.Forinstance,theevent{!2⌦: i,j(G(!))r}2F,(19)representsthesetofrealizations!2⌦yieldingagraphG(!)whosedistancebetweenthe(prefixed)nodesiandjdoesnotexceedr.Torenderamorecompactnotation,wehenceforthomittherealization!inthecharacterizationoftheevents.Forinstance,inthecaseoftheevent(19)werepresentitratheras{ i,j(G)r}2F,(20)wheretherandomquantitiesareemphasizedbytheboldfacednotation -- intheeventinequation(20),theonlyrandomobjectisthegraphG.C.UsefulGraphOperationsInourexposition,wewillbeperformingcertainoperationsovergraphs,aswellasevaluatecertainfunctionssuchascom-paringdistancesbetweennodesoverdistinctgraphs.There-fore,itisusefultointroducethefollowinggraphoperationsforlateruse(whichareillustratedinFigure2):1)Graphembedding.GivenavertexsetV,andasubsetthereof,S⇢V,theembeddingofagraphG(1)2G(S)intothelargergraphG(2)2G(V)willbedenotedby:G=G(1) G(2),G2G(V),(21)andresultsinagraphwiththefollowingproperties:i)theconnectionsbetweennodesinSthatarepresentinG(2)arecancelled;ii)thenodesinthevertexsetSofgraphG(1)aremappedintothecorrespondingnodesofgraphG(2),andsoarethepertinentconnections.WestressthattheconnectionsfromS0toSaredeterminedbythegraphG(2).Wenoticethattheoperationin(21)isnotcommutative(becausethefirstgraphisembeddedintothesecondgraph,andnotviceversa),andthattheoutputgraphGdoesnotdependontheconnectionsexistinginG(2)amongnodesbelongingtothesetS.2)Localdisconnection.GivenagraphG2G(V),thenotation:GU1=U22G(V),(22)describesthegraphthatisfromGafterremovingalltheedgesthatconnectnodesinU1tonodesinU2,namely,alltheconnectionsbetweenU1andU2.3 Tomography Center measurements from patches under test 6theindicatorfunction,whichisequaltooneifconditionEistrue,andisequaltozero,otherwise.TheN⇥NidentitymatrixisdenotedbyI.ThecomplementofasetSisdenotedbyS0.GivenanN⇥NmatrixZ,thesubmatrixthatliesintherowsofZindexedbythesetS✓{1,2,...,N}andinthecolumnsindexedbythesetT✓{1,2,...,N},willbedenotedbyZST,oralternativelyby[Z]ST.WhenS=T,thesubmatrixZSTwillbeabbreviatedasZS.Intheindexingofthesubmatrixwewillretaintheindexsetoftheoriginalmatrix.Forexample,ifS={2,3}andT={2,4,5},wehavethatthesubmatrixM=ZSTisa2⇥3matrix,indexedasfollows:M=✓z22z24z25z32z34z35◆=✓m22m24m25m32m34m35◆.(15)Thisnotationiscrucialinourtreatment,sinceitwillallowustoidentifynodeswithoutcumbersomeandredundantdouble-indexnotation.B.GraphnotationG(V)isthesetofundirectedgraphsdefinedonasetofnodes(vertexset)V.WhenNisthenumberofnodes,theshortcutG(N)impliesthatthevertexsetisV={1,2,...,N}.WhendealingwithagraphG2G(N),itsconnectionstructure(i.e.,theedgesofthegraphs)canbedescribedthroughitsN⇥Nadjacencymatrix.The(i,j)-thentryoftheadjacencymatrixofgraphGwillbedenotedbygij,withgij=1ifnodesiandjareconnected,andgij=0,otherwise.Self-loops(gii=1)willbepermitted.GivenG2G(N),andasubsetS✓{1,2,...,N},thesubgraphcorrespondingtoSisdenotedbyGS2G(S).ThesupportgraphofamatrixAisdenotedbyG(A).The(i,j)-thentryofitsadjacencymatrixisI{aij>0},namely,nodesiandjareconnectedonG(A)if,andonly,ifaijisstrictlypositive.Apathfromitojisasequenceofedgeswherethefirstedgeoriginatesfromiandthelastedgeterminatesatj.Theexistenceofapathoflengthrcanbeexpressedas:gin1gn1n2...gnr 1j=1,(16)foracertainsequenceofverticesn1,n2,...,nr 1belongingtoV.Accordingtothisdefinition,apathcanalsopassmultipletimesthroughthesamenode,orcanlingerforoneormorestepsatthesamenodewhenithasaself-loop.Ni(G)isthesetofneighborsofnodei(includingiitself)intheundirectedgraphG.ThedegreeofnodeiisthecardinalityofNi(G).dmax(G)isthemaximumdegreeinG. i,j(G)isthedistancebetweenthenodesiandjonthegraphG,i.e.,thelengthoftheshortestpathlinkingiandj.N(r)i(G)isther-thorderneighborhoodofnodei(includingiitself),givenbyN(r)i(G)={j2V: i,j(G)r}.(17)ArandomgraphGobeystheErdos-R´enyimodelifeachedgeofGisdrawn,independentlyfromtheotheredges,withidenticalprobabilitypN.Equivalentlystated,theadjacencyrandomvariablesgij,fori=1,2,...,Nandi<j,areindependentandidenticallydistributed(i.i.d.)Bernoullivariables.ThenotationG⇠G?(N,pN)signifiesthatthegraphGbelongstotheErdos-R´enyiclasswithconnectionprobabilitythatvanishesasN!1,andthatobeysthefollowingscalinglaw:pN=logN+cNN,(18)wherecN!1asN!1(inanarbitraryway,providedthatpN!0).Itisawell-knownresultthatrandomgraphsbelongingtothefamilyG?(N,pN)areconnectedwithhighprobability[21].Beforeproceeding,andasanoteofclarity,weassumethatallrandomvariablesfinddomaininthecommonprobabilityspace(⌦,F,P),where⌦isthesetofrealizations,Fisthesigma-algebraofmeasurablesetsandPistheprobabilitymeasure.Forinstance,theevent{!2⌦: i,j(G(!))r}2F,(19)representsthesetofrealizations!2⌦yieldingagraphG(!)whosedistancebetweenthe(prefixed)nodesiandjdoesnotexceedr.Torenderamorecompactnotation,wehenceforthomittherealization!inthecharacterizationoftheevents.Forinstance,inthecaseoftheevent(19)werepresentitratheras{ i,j(G)r}2F,(20)wheretherandomquantitiesareemphasizedbytheboldfacednotation -- intheeventinequation(20),theonlyrandomobjectisthegraphG.C.UsefulGraphOperationsInourexposition,wewillbeperformingcertainoperationsovergraphs,aswellasevaluatecertainfunctionssuchascom-paringdistancesbetweennodesoverdistinctgraphs.There-fore,itisusefultointroducethefollowinggraphoperationsforlateruse(whichareillustratedinFigure2):1)Graphembedding.GivenavertexsetV,andasubsetthereof,S⇢V,theembeddingofagraphG(1)2G(S)intothelargergraphG(2)2G(V)willbedenotedby:G=G(1) G(2),G2G(V),(21)andresultsinagraphwiththefollowingproperties:i)theconnectionsbetweennodesinSthatarepresentinG(2)arecancelled;ii)thenodesinthevertexsetSofgraphG(1)aremappedintothecorrespondingnodesofgraphG(2),andsoarethepertinentconnections.WestressthattheconnectionsfromS0toSaredeterminedbythegraphG(2).Wenoticethattheoperationin(21)isnotcommutative(becausethefirstgraphisembeddedintothesecondgraph,andnotviceversa),andthattheoutputgraphGdoesnotdependontheconnectionsexistinginG(2)amongnodesbelongingtothesetS.2)Localdisconnection.GivenagraphG2G(V),thenotation:GU1=U22G(V),(22)describesthegraphthatisfromGafterremovingalltheedgesthatconnectnodesinU1tonodesinU2,namely,alltheconnectionsbetweenU1andU2.1 6theindicatorfunction,whichisequaltooneifconditionEistrue,andisequaltozero,otherwise.TheN⇥NidentitymatrixisdenotedbyI.ThecomplementofasetSisdenotedbyS0.GivenanN⇥NmatrixZ,thesubmatrixthatliesintherowsofZindexedbythesetS✓{1,2,...,N}andinthecolumnsindexedbythesetT✓{1,2,...,N},willbedenotedbyZST,oralternativelyby[Z]ST.WhenS=T,thesubmatrixZSTwillbeabbreviatedasZS.Intheindexingofthesubmatrixwewillretaintheindexsetoftheoriginalmatrix.Forexample,ifS={2,3}andT={2,4,5},wehavethatthesubmatrixM=ZSTisa2⇥3matrix,indexedasfollows:M=✓z22z24z25z32z34z35◆=✓m22m24m25m32m34m35◆.(15)Thisnotationiscrucialinourtreatment,sinceitwillallowustoidentifynodeswithoutcumbersomeandredundantdouble-indexnotation.B.GraphnotationG(V)isthesetofundirectedgraphsdefinedonasetofnodes(vertexset)V.WhenNisthenumberofnodes,theshortcutG(N)impliesthatthevertexsetisV={1,2,...,N}.WhendealingwithagraphG2G(N),itsconnectionstructure(i.e.,theedgesofthegraphs)canbedescribedthroughitsN⇥Nadjacencymatrix.The(i,j)-thentryoftheadjacencymatrixofgraphGwillbedenotedbygij,withgij=1ifnodesiandjareconnected,andgij=0,otherwise.Self-loops(gii=1)willbepermitted.GivenG2G(N),andasubsetS✓{1,2,...,N},thesubgraphcorrespondingtoSisdenotedbyGS2G(S).ThesupportgraphofamatrixAisdenotedbyG(A).The(i,j)-thentryofitsadjacencymatrixisI{aij>0},namely,nodesiandjareconnectedonG(A)if,andonly,ifaijisstrictlypositive.Apathfromitojisasequenceofedgeswherethefirstedgeoriginatesfromiandthelastedgeterminatesatj.Theexistenceofapathoflengthrcanbeexpressedas:gin1gn1n2...gnr 1j=1,(16)foracertainsequenceofverticesn1,n2,...,nr 1belongingtoV.Accordingtothisdefinition,apathcanalsopassmultipletimesthroughthesamenode,orcanlingerforoneormorestepsatthesamenodewhenithasaself-loop.Ni(G)isthesetofneighborsofnodei(includingiitself)intheundirectedgraphG.ThedegreeofnodeiisthecardinalityofNi(G).dmax(G)isthemaximumdegreeinG. i,j(G)isthedistancebetweenthenodesiandjonthegraphG,i.e.,thelengthoftheshortestpathlinkingiandj.N(r)i(G)isther-thorderneighborhoodofnodei(includingiitself),givenbyN(r)i(G)={j2V: i,j(G)r}.(17)ArandomgraphGobeystheErdos-R´enyimodelifeachedgeofGisdrawn,independentlyfromtheotheredges,withidenticalprobabilitypN.Equivalentlystated,theadjacencyrandomvariablesgij,fori=1,2,...,Nandi<j,areindependentandidenticallydistributed(i.i.d.)Bernoullivariables.ThenotationG⇠G?(N,pN)signifiesthatthegraphGbelongstotheErdos-R´enyiclasswithconnectionprobabilitythatvanishesasN!1,andthatobeysthefollowingscalinglaw:pN=logN+cNN,(18)wherecN!1asN!1(inanarbitraryway,providedthatpN!0).Itisawell-knownresultthatrandomgraphsbelongingtothefamilyG?(N,pN)areconnectedwithhighprobability[21].Beforeproceeding,andasanoteofclarity,weassumethatallrandomvariablesfinddomaininthecommonprobabilityspace(⌦,F,P),where⌦isthesetofrealizations,Fisthesigma-algebraofmeasurablesetsandPistheprobabilitymeasure.Forinstance,theevent{!2⌦: i,j(G(!))r}2F,(19)representsthesetofrealizations!2⌦yieldingagraphG(!)whosedistancebetweenthe(prefixed)nodesiandjdoesnotexceedr.Torenderamorecompactnotation,wehenceforthomittherealization!inthecharacterizationoftheevents.Forinstance,inthecaseoftheevent(19)werepresentitratheras{ i,j(G)r}2F,(20)wheretherandomquantitiesareemphasizedbytheboldfacednotation -- intheeventinequation(20),theonlyrandomobjectisthegraphG.C.UsefulGraphOperationsInourexposition,wewillbeperformingcertainoperationsovergraphs,aswellasevaluatecertainfunctionssuchascom-paringdistancesbetweennodesoverdistinctgraphs.There-fore,itisusefultointroducethefollowinggraphoperationsforlateruse(whichareillustratedinFigure2):1)Graphembedding.GivenavertexsetV,andasubsetthereof,S⇢V,theembeddingofagraphG(1)2G(S)intothelargergraphG(2)2G(V)willbedenotedby:G=G(1) G(2),G2G(V),(21)andresultsinagraphwiththefollowingproperties:i)theconnectionsbetweennodesinSthatarepresentinG(2)arecancelled;ii)thenodesinthevertexsetSofgraphG(1)aremappedintothecorrespondingnodesofgraphG(2),andsoarethepertinentconnections.WestressthattheconnectionsfromS0toSaredeterminedbythegraphG(2).Wenoticethattheoperationin(21)isnotcommutative(becausethefirstgraphisembeddedintothesecondgraph,andnotviceversa),andthattheoutputgraphGdoesnotdependontheconnectionsexistinginG(2)amongnodesbelongingtothesetS.2)Localdisconnection.GivenagraphG2G(V),thenotation:GU1=U22G(V),(22)describesthegraphthatisfromGafterremovingalltheedgesthatconnectnodesinU1tonodesinU2,namely,alltheconnectionsbetweenU1andU2.2 6theindicatorfunction,whichisequaltooneifconditionEistrue,andisequaltozero,otherwise.TheN⇥NidentitymatrixisdenotedbyI.ThecomplementofasetSisdenotedbyS0.GivenanN⇥NmatrixZ,thesubmatrixthatliesintherowsofZindexedbythesetS✓{1,2,...,N}andinthecolumnsindexedbythesetT✓{1,2,...,N},willbedenotedbyZST,oralternativelyby[Z]ST.WhenS=T,thesubmatrixZSTwillbeabbreviatedasZS.Intheindexingofthesubmatrixwewillretaintheindexsetoftheoriginalmatrix.Forexample,ifS={2,3}andT={2,4,5},wehavethatthesubmatrixM=ZSTisa2⇥3matrix,indexedasfollows:M=✓z22z24z25z32z34z35◆=✓m22m24m25m32m34m35◆.(15)Thisnotationiscrucialinourtreatment,sinceitwillallowustoidentifynodeswithoutcumbersomeandredundantdouble-indexnotation.B.GraphnotationG(V)isthesetofundirectedgraphsdefinedonasetofnodes(vertexset)V.WhenNisthenumberofnodes,theshortcutG(N)impliesthatthevertexsetisV={1,2,...,N}.WhendealingwithagraphG2G(N),itsconnectionstructure(i.e.,theedgesofthegraphs)canbedescribedthroughitsN⇥Nadjacencymatrix.The(i,j)-thentryoftheadjacencymatrixofgraphGwillbedenotedbygij,withgij=1ifnodesiandjareconnected,andgij=0,otherwise.Self-loops(gii=1)willbepermitted.GivenG2G(N),andasubsetS✓{1,2,...,N},thesubgraphcorrespondingtoSisdenotedbyGS2G(S).ThesupportgraphofamatrixAisdenotedbyG(A).The(i,j)-thentryofitsadjacencymatrixisI{aij>0},namely,nodesiandjareconnectedonG(A)if,andonly,ifaijisstrictlypositive.Apathfromitojisasequenceofedgeswherethefirstedgeoriginatesfromiandthelastedgeterminatesatj.Theexistenceofapathoflengthrcanbeexpressedas:gin1gn1n2...gnr 1j=1,(16)foracertainsequenceofverticesn1,n2,...,nr 1belongingtoV.Accordingtothisdefinition,apathcanalsopassmultipletimesthroughthesamenode,orcanlingerforoneormorestepsatthesamenodewhenithasaself-loop.Ni(G)isthesetofneighborsofnodei(includingiitself)intheundirectedgraphG.ThedegreeofnodeiisthecardinalityofNi(G).dmax(G)isthemaximumdegreeinG. i,j(G)isthedistancebetweenthenodesiandjonthegraphG,i.e.,thelengthoftheshortestpathlinkingiandj.N(r)i(G)isther-thorderneighborhoodofnodei(includingiitself),givenbyN(r)i(G)={j2V: i,j(G)r}.(17)ArandomgraphGobeystheErdos-R´enyimodelifeachedgeofGisdrawn,independentlyfromtheotheredges,withidenticalprobabilitypN.Equivalentlystated,theadjacencyrandomvariablesgij,fori=1,2,...,Nandi<j,areindependentandidenticallydistributed(i.i.d.)Bernoullivariables.ThenotationG⇠G?(N,pN)signifiesthatthegraphGbelongstotheErdos-R´enyiclasswithconnectionprobabilitythatvanishesasN!1,andthatobeysthefollowingscalinglaw:pN=logN+cNN,(18)wherecN!1asN!1(inanarbitraryway,providedthatpN!0).Itisawell-knownresultthatrandomgraphsbelongingtothefamilyG?(N,pN)areconnectedwithhighprobability[21].Beforeproceeding,andasanoteofclarity,weassumethatallrandomvariablesfinddomaininthecommonprobabilityspace(⌦,F,P),where⌦isthesetofrealizations,Fisthesigma-algebraofmeasurablesetsandPistheprobabilitymeasure.Forinstance,theevent{!2⌦: i,j(G(!))r}2F,(19)representsthesetofrealizations!2⌦yieldingagraphG(!)whosedistancebetweenthe(prefixed)nodesiandjdoesnotexceedr.Torenderamorecompactnotation,wehenceforthomittherealization!inthecharacterizationoftheevents.Forinstance,inthecaseoftheevent(19)werepresentitratheras{ i,j(G)r}2F,(20)wheretherandomquantitiesareemphasizedbytheboldfacednotation -- intheeventinequation(20),theonlyrandomobjectisthegraphG.C.UsefulGraphOperationsInourexposition,wewillbeperformingcertainoperationsovergraphs,aswellasevaluatecertainfunctionssuchascom-paringdistancesbetweennodesoverdistinctgraphs.There-fore,itisusefultointroducethefollowinggraphoperationsforlateruse(whichareillustratedinFigure2):1)Graphembedding.GivenavertexsetV,andasubsetthereof,S⇢V,theembeddingofagraphG(1)2G(S)intothelargergraphG(2)2G(V)willbedenotedby:G=G(1) G(2),G2G(V),(21)andresultsinagraphwiththefollowingproperties:i)theconnectionsbetweennodesinSthatarepresentinG(2)arecancelled;ii)thenodesinthevertexsetSofgraphG(1)aremappedintothecorrespondingnodesofgraphG(2),andsoarethepertinentconnections.WestressthattheconnectionsfromS0toSaredeterminedbythegraphG(2).Wenoticethattheoperationin(21)isnotcommutative(becausethefirstgraphisembeddedintothesecondgraph,andnotviceversa),andthattheoutputgraphGdoesnotdependontheconnectionsexistinginG(2)amongnodesbelongingtothesetS.2)Localdisconnection.GivenagraphG2G(V),thenotation:GU1=U22G(V),(22)describesthegraphthatisfromGafterremovingalltheedgesthatconnectnodesinU1tonodesinU2,namely,alltheconnectionsbetweenU1andU2.4 6theindicatorfunction,whichisequaltooneifconditionEistrue,andisequaltozero,otherwise.TheN⇥NidentitymatrixisdenotedbyI.ThecomplementofasetSisdenotedbyS0.GivenanN⇥NmatrixZ,thesubmatrixthatliesintherowsofZindexedbythesetS✓{1,2,...,N}andinthecolumnsindexedbythesetT✓{1,2,...,N},willbedenotedbyZST,oralternativelyby[Z]ST.WhenS=T,thesubmatrixZSTwillbeabbreviatedasZS.Intheindexingofthesubmatrixwewillretaintheindexsetoftheoriginalmatrix.Forexample,ifS={2,3}andT={2,4,5},wehavethatthesubmatrixM=ZSTisa2⇥3matrix,indexedasfollows:M=✓z22z24z25z32z34z35◆=✓m22m24m25m32m34m35◆.(15)Thisnotationiscrucialinourtreatment,sinceitwillallowustoidentifynodeswithoutcumbersomeandredundantdouble-indexnotation.B.GraphnotationG(V)isthesetofundirectedgraphsdefinedonasetofnodes(vertexset)V.WhenNisthenumberofnodes,theshortcutG(N)impliesthatthevertexsetisV={1,2,...,N}.WhendealingwithagraphG2G(N),itsconnectionstructure(i.e.,theedgesofthegraphs)canbedescribedthroughitsN⇥Nadjacencymatrix.The(i,j)-thentryoftheadjacencymatrixofgraphGwillbedenotedbygij,withgij=1ifnodesiandjareconnected,andgij=0,otherwise.Self-loops(gii=1)willbepermitted.GivenG2G(N),andasubsetS✓{1,2,...,N},thesubgraphcorrespondingtoSisdenotedbyGS2G(S).ThesupportgraphofamatrixAisdenotedbyG(A).The(i,j)-thentryofitsadjacencymatrixisI{aij>0},namely,nodesiandjareconnectedonG(A)if,andonly,ifaijisstrictlypositive.Apathfromitojisasequenceofedgeswherethefirstedgeoriginatesfromiandthelastedgeterminatesatj.Theexistenceofapathoflengthrcanbeexpressedas:gin1gn1n2...gnr 1j=1,(16)foracertainsequenceofverticesn1,n2,...,nr 1belongingtoV.Accordingtothisdefinition,apathcanalsopassmultipletimesthroughthesamenode,orcanlingerforoneormorestepsatthesamenodewhenithasaself-loop.Ni(G)isthesetofneighborsofnodei(includingiitself)intheundirectedgraphG.ThedegreeofnodeiisthecardinalityofNi(G).dmax(G)isthemaximumdegreeinG. i,j(G)isthedistancebetweenthenodesiandjonthegraphG,i.e.,thelengthoftheshortestpathlinkingiandj.N(r)i(G)isther-thorderneighborhoodofnodei(includingiitself),givenbyN(r)i(G)={j2V: i,j(G)r}.(17)ArandomgraphGobeystheErdos-R´enyimodelifeachedgeofGisdrawn,independentlyfromtheotheredges,withidenticalprobabilitypN.Equivalentlystated,theadjacencyrandomvariablesgij,fori=1,2,...,Nandi<j,areindependentandidenticallydistributed(i.i.d.)Bernoullivariables.ThenotationG⇠G?(N,pN)signifiesthatthegraphGbelongstotheErdos-R´enyiclasswithconnectionprobabilitythatvanishesasN!1,andthatobeysthefollowingscalinglaw:pN=logN+cNN,(18)wherecN!1asN!1(inanarbitraryway,providedthatpN!0).Itisawell-knownresultthatrandomgraphsbelongingtothefamilyG?(N,pN)areconnectedwithhighprobability[21].Beforeproceeding,andasanoteofclarity,weassumethatallrandomvariablesfinddomaininthecommonprobabilityspace(⌦,F,P),where⌦isthesetofrealizations,Fisthesigma-algebraofmeasurablesetsandPistheprobabilitymeasure.Forinstance,theevent{!2⌦: i,j(G(!))r}2F,(19)representsthesetofrealizations!2⌦yieldingagraphG(!)whosedistancebetweenthe(prefixed)nodesiandjdoesnotexceedr.Torenderamorecompactnotation,wehenceforthomittherealization!inthecharacterizationoftheevents.Forinstance,inthecaseoftheevent(19)werepresentitratheras{ i,j(G)r}2F,(20)wheretherandomquantitiesareemphasizedbytheboldfacednotation -- intheeventinequation(20),theonlyrandomobjectisthegraphG.C.UsefulGraphOperationsInourexposition,wewillbeperformingcertainoperationsovergraphs,aswellasevaluatecertainfunctionssuchascom-paringdistancesbetweennodesoverdistinctgraphs.There-fore,itisusefultointroducethefollowinggraphoperationsforlateruse(whichareillustratedinFigure2):1)Graphembedding.GivenavertexsetV,andasubsetthereof,S⇢V,theembeddingofagraphG(1)2G(S)intothelargergraphG(2)2G(V)willbedenotedby:G=G(1) G(2),G2G(V),(21)andresultsinagraphwiththefollowingproperties:i)theconnectionsbetweennodesinSthatarepresentinG(2)arecancelled;ii)thenodesinthevertexsetSofgraphG(1)aremappedintothecorrespondingnodesofgraphG(2),andsoarethepertinentconnections.WestressthattheconnectionsfromS0toSaredeterminedbythegraphG(2).Wenoticethattheoperationin(21)isnotcommutative(becausethefirstgraphisembeddedintothesecondgraph,andnotviceversa),andthattheoutputgraphGdoesnotdependontheconnectionsexistinginG(2)amongnodesbelongingtothesetS.2)Localdisconnection.GivenagraphG2G(V),thenotation:GU1=U22G(V),(22)describesthegraphthatisfromGafterremovingalltheedgesthatconnectnodesinU1tonodesinU2,namely,alltheconnectionsbetweenU1andU2.3 patches under test Sequential topology reconstruction
15
Probing experiment no. 1
Probing experiment no. 2
Probing experiment no. 3
Probing experiment no. 4
Fig. 10.
Illustration of the sequential graph reconstruction. We consider S = 20 nodes with probing limit M = 10. Each patch Si has cardinality equal to
Si = 5 nodes. At each experiment two patches are probed. The red nodes represent the nodes being probed at each experiment and the red edges represent
the inferred edges up to the current experiment. All pairs were correctly classified.
Probing experiment no. 5
Probing experiment no. 6
importance on the network. Likewise, some prior knowledge
on a particular network structure (e.g., a tree structure), could
help to optimize the formation of successive patches.
VII. COMPARISON WITH THE RESULTS IN [31]
It is useful to contrast the results of this work with the results
in [31]. As already explained in the introductory sections, the
main differences can be summarized as follows. In [31], the
network is homogeneous, since all the connections (also those
in S) obey a classic Erdos-R´enyi construction with connection
probability pN . Also, the size of the network, S, scales linearly
with N as ξN, meaning that the fraction of observable nodes,
ξ, is constant (and greater than zero). Finally, the consistency
result
the
fraction of correctly identified edges converges to one in the
limit and an independence approximation is used to control
the error rate. In this work, consistency holds in a strong
sense and the independence approximation is not required.
The results obtained here generalize the above framework in
several directions.
in [31] holds in a weak sense:
it shows that
Fig. 11.
of the distance between the true subnetwork GS and the estimator (cid:98)GS. We
Illustration of the monotonic decrease, as more patches are probed,
consider S = 60 nodes in S with probing limit M = 10. Each patch has
Si = 5 nodes, for a total number of P = 2S/M = 12 patches. At each
experiment two patches are probed, yielding a total of P (P − 1)/2 = 66
experiments. The graph displays the distance between GS and the estimated
graph at the experiment (cid:96) = 1, 2, . . . , 66. In the considered experiment, only
two pairs were misclassified.
results possibly arising from subsequent experiments. Second,
in some applications, the reconstruction can be formed sequen-
tially, by exploiting, at each experiment, the information com-
ing from past experiments. For example, having ascertained
the structure of a given subset of nodes might be informative
of important network-level features of some nodes -- e.g.,
high degree nodes -- and hence, informative on their level of
A. The case of fixed S
In this work, we prove that perfect recovery is achievable
even in the extreme case that
the number of observable
nodes is fixed when N diverges, namely, when the observable
network portion is embedded into an infinitely large number
of unobservable nodes. We remark that the case of fixed S
cannot be addressed with the tools used in [31]. Let us now
-1.5-1-0.500.511.5-1.5-1-0.500.511.5-1.5-1-0.500.511.5-1.5-1-0.500.511.5-1.5-1-0.500.511.5-1.5-1-0.500.511.5-1.5-1-0.500.511.5-1.5-1-0.500.511.5-1.5-1-0.500.511.5-1.5-1-0.500.511.5-1.5-1-0.500.511.5-1.5-1-0.500.511.5Distance to true graphPatch index explain why. The result proved in [31] relies essentially on the
following result (Theorem 1 in [31]):
S(cid:88)j=1
eij ≤ ρ,
(49)
irrespectively of the network size. This result
which reveals that the (column-wise) sum of the errors is
limited,
is
obtained by exploiting matrix algebra tools. It is shown in [31]
how (49) leads to the useful conclusion that, on average, the
off-diagonal entries of the error matrix scale as 1/S, which
further implies that4:
P[N pN eij > ] (cid:46) N
S
pN ,
(Ref. [31])
(50)
where the symbol "(cid:46)" here means that the quantity appearing
on the left-hand side is upper bounded by a quantity that
scales, asymptotically with N, as the quantity appearing on
the right-hand side. Equation (50) reveals two useful facts.
First, when S/N stays constant as N grows, and since pN
goes to zero, we see that the magnified error vanishes. This is
one fundamental conclusion ascertained in [31]. At the same
time, Eq. (50) highlights how, for fixed S, we are no longer in
the position of establishing from (50) that the magnified error
converges to zero, because the product N pN diverges with
N. In summary, the matrix-algebra tools taken in [31] are not
powerful enough to address the challenging case when S is
fixed, namely, when the fraction of observable nodes goes to
zero as N grows.
On the other hand, in this work we show how this more
challenging scenario can be addressed, by exploiting matrix-
graph tools, i.e., by evaluating paths and distances over graphs.
One important benefit of the new approach is that the results
now hold for an arbitrary topology of the observable network
portion, while in [31] this latter component was constrained
to be Erdos-R´enyi.
B. The case of S ∼ ξN
We notice that the results of this work can be applied to the
case addressed in [31]. Indeed, when GS is Erdos-R´enyi, we
can repeat the proof of Theorem 1 by essentially skipping the
homogenization-and-coupling step, because the overall graph
is homogeneous ab initio. Then we would get, for any i (cid:54)= j
(also for the connected pairs, in this particular case):
P[N pN eij > ] (cid:46) pN (N pN )rN +2.
(51)
Therefore, both the matrix-algebra approach (used in [31]),
and the matrix-graph approach (used here) lead to the result
that the topology of the observable network portion can be
reconstructed faithfully. However, it must be remarked that the
matrix-graph approach requires some additional conditions on
the connection probability, pN , which translate into a slightly
more restrictive requirement in terms of sparsity.
On the other hand, and interestingly, the matrix-algebra
approach and the matrix-graph approach lead to different
4Actually, the result in [31] is formulated in terms of empirical fraction of
errors. In the case that permutation invariance holds -- see Property 2 in [31]
-- the result is easily rephrased as in (50).
16
estimates on how the error probability in (51) converges to
zero. Indeed, with the approach used in [31], one is able
to see that the rate of decay is at least in the order of pN
(see (50), and observe that N/S ∼ 1/ξ). Moreover, in [31]
it is shown that the decay rate is actually faster than pN .
In contrast, with the approach adopted in the current work
we get the upper bound in (51), which provides the (looser)
asymptotic prediction that the decay rate is slower than pN . In
summary, we conclude that, under the regime S ∼ ξN, and for
a full Erdos-R´enyi construction, the results of [31] are more
powerful in predicting the decay rate of the error probabilities.
It could be interesting at this point to ask whether it is possible
to combine the matrix-algebra approach with the matrix-graph
approach to obtain refined estimates.
APPENDIX A
SOME USEFUL LEMMAS
Proof of Lemma 1:
is internally disconnected,
In order to prove that the partial
Erdos-R´enyi graph G ∼ G (cid:63)(N, pN , GS) is connected,
it
suffices to consider the worst case where the embedded graph,
GS,
i.e., where no edges exist
between nodes in S. We note that the nodes in S, even if
disconnected, can still be connected to nodes belonging to the
unobserved set, S(cid:48). The latter property enables the possibility
that the overall graph, G, is connected, as we are going to
show.
Since we are assuming that GS is internally disconnected,
the overall graph is connected if both GS(cid:48)
is connected,
and any node in S connects to some node in S(cid:48). Refer to
Figure 12 for a graphical illustration. We prove Lemma 1 via
the contrapositive statement: the overall graph is not connected
if either GS(cid:48) is not connected, or if at least one node in S is
not connected to S(cid:48), namely, we have that:
{G not connected}
⊆
(52)
{GS(cid:48) not connected}(cid:83){∃ an isolated node of G in S}.
Therefore, applying the union bound we get:
P[G is not connected]
(53)
P[GS(cid:48) is not connected] + P[∃ an isolated node of G in S]
≤
=
P[GS(cid:48) is not connected] + 1 − (1 − (1 − pN )N−S)S.
Since GS(cid:48) is a classic Erdos-R´enyi G (cid:63) (N − S, pN ), we have
that:
(54)
P[GS(cid:48) is not connected] = 0.
lim
N→∞
Moreover, since S is fixed, we have that:
(1 − pN )N−S = (cid:18)1 −
≤ (cid:18)1 −
log N + cN
N
log N
N (cid:19)N−S
(cid:19)N−S
N→∞−→ 0.
(55)
17
parameter pN defined in equation (18). Actually, for i, j ∈ S
(i.e., when gij = gij is deterministic) a simplified version of
Lemma 1 in [31] does suffice, where the conditioning can be
skipped.
APPENDIX B
PROOF OF THEOREM 1
We first prove part i). It is shown in [31] that the entries
of the error matrix defined in (36) are nonnegative, i.e., ES ≥
0S×S, and, hence, we can write, for i, j ∈ S:
(63)
N pN [(cid:98)AS]ij = N pN aij + N pN eij ≥ N pN aij.
Therefore, from Property 2 we get immediately the claim in
part i). If we further show that the magnified error N pN eij
converges to zero in probability over the non-interacting pairs
(i.e., if we prove part ii) of the present theorem), then we can
attain exact (with high probability) classification via inspection
on the truncated estimator (cid:98)AS: if N pN [(cid:98)AS]ij > τ, then
classify (i, j) as an interacting pair, otherwise classify it as
non-interacting, where τ is the threshold characterizing the
family Cρ,τ of weight assignments from where A is obtained,
in view of Property 2. As a result, and since the cardinality
of the observable set is finite, part iii) would follow if we are
able to prove part ii). The proof of part ii) is demanding and
will be developed through a sequence of five steps.
Step 1: Relating the error to the distance between nodes
belonging to S(cid:48). It is shown in [31] that the error matrix in (36)
can be represented as:5
ES = ASS(cid:48)HBS(cid:48)S
where
B (cid:44) A2,
H (cid:44) (IS(cid:48) − BS(cid:48))−1 .
From (64) we can write, for i, j ∈ S:
eij = (cid:88)(cid:96),m∈S(cid:48)
ai(cid:96)h(cid:96)mbmj
(64)
(65)
(66)
where eij is the error at the pair (i, j). Therefore, in order to
control the size of the error eij, small values of the factors h(cid:96)m,
for (cid:96), m ∈ S(cid:48), would be desirable. In view of the definition
for H in (65), we have that:
,
(67)
=(cid:34) ∞(cid:88)k=0
(BS(cid:48))k(cid:35)(cid:96)m
h(cid:96)m =(cid:104)(IS(cid:48) − BS(cid:48))−1(cid:105)(cid:96)m
as the matrix BS(cid:48) = (cid:2)A2(cid:3)S(cid:48)
is stable, since ρ (BS(cid:48)) <
< 1, from Property 1, where ρ (BS(cid:48)) is the spectral
BS(cid:48)∞
radius of BS(cid:48). It is useful at this point to recall the following
known fact from matrix algebra that relates the entries associ-
ated with the powers of a matrix with the distances between
nodes on its underlying support graph.
Let M ∈ SN×N
be a nonnegative symmetric matrix with
positive diagonal entries, and let G(M ) be its underlying
+
Fig. 12.
If the subnetwork GS(cid:48) is connected and there is no node in S that
is isolated in the network G, then G is connected. Equation (52) conforms to
the corresponding contrapositive statement.
Proof of Lemma 2:
In order to prove the claim of the
lemma, we must show that (34) implies (32) with the choice
τ = γ/e. Let us observe preliminarily that (34) yields the
following implication:
(58)
{dmax(G) < eN pN , gij = 1} ⊆ {N pN aij > τ, gij = 1}.
(56)
Therefore, we can write:
P[N pN aij > τgij = 1] ≥ P[dmax(G) < eN pNgij = 1].
(57)
Now, by trivial upper bounding techniques, we can obtain the
following chain of inequalities:
P [ dmax(G) ≥ eN pN gij = 1]
gnk > eN pN(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤ P(cid:34) max
gij = 1(cid:35)
n /∈S(cid:88)k
gnk > eN pN(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
+P(cid:34) max
gij = 1(cid:35)
n∈S(cid:88)k
gnk > eN pN(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
gij = 1(cid:35)
= P(cid:34) max
n /∈S(cid:88)k
gnk(cid:33) > eN pN(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
n∈S(cid:32)(cid:88)k∈S
+P(cid:34) max
gnk +(cid:88)k /∈S
gnk > eN pN(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
≤ P(cid:34) max
gij = 1(cid:35)
n /∈S(cid:88)k
gnk > eN pN(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
+P(cid:34) S + max
gij = 1(cid:35)
n∈S(cid:88)k /∈S
gnk > eN pN(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
= P(cid:34) max
gij = 1(cid:35)
n /∈S(cid:88)k
gnk > eN pN − S(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
+P(cid:34) max
n∈S(cid:88)k /∈S
N→∞−→ 0,
gij = 1(cid:35)
gij = 1(cid:35)
≤ N
(61)
(60)
(59)
(62)
where the last
inequality follows directly from Lemma 1
in [31], since S is fixed and since the subgraph formed by
the edges gnk with either n /∈ S or k /∈ S is Erdos-R´enyi with
5During this first step the boldface notation will be skipped because we
focus on properties that depend solely on the structure of the matrix, and not
on the statistical model of the underlying graph.
support graph. Consider the powers of the matrix M, namely,
M k, for k = 1, 2, . . .. Then we have that [55]:
δ(cid:96),m(G(M )) = r ⇔ the smallest k with [M k](cid:96)m > 0 is r,
(68)
where δ(cid:96),m(G(M )) represents the distance between the nodes
(cid:96) and m in the graph G(M ) as defined in Sec. II-B. In fact,
note that, if (cid:96) is not connected to m in the support graph
G(M ), then M(cid:96)m = 0. If the smallest path connecting (cid:96) to m
has a length of two hops (in particular, (cid:96) is not connected to
m, hence M(cid:96)m = 0), then there exists k so that M(cid:96)k > 0 and
Reasoning by induction one can establish (68). The following
observation follows: if M is stable, i.e., ρ(M ) < 1, and if the
Mkm > 0. Thus,(cid:2)M 2(cid:3)(cid:96)m =(cid:80)r M(cid:96)rMrm > M(cid:96)kMkm > 0.
distance δ(cid:96)m(G(M )) = r is large, then(cid:2)M k(cid:3)(cid:96)m is small for
all k as for k < r we have (cid:2)M k(cid:3)(cid:96)m = 0 and for k ≥ r the
corresponding power (cid:2)M k(cid:3)(cid:96)m is small since k is large and
Now, examining (67), and using (68) with M = BS(cid:48), one
might be tempted to conclude that a small h(cid:96)m would result if
nodes (cid:96) and m are distant from each other. The reasoning is
correct, but note that, in general, the distance between (cid:96) and
m is dependent on the topology of the network GS, which is
arbitrary. In other words, by relying solely on the elementary
observation in (68), one would not be able to draw useful
conclusions about the magnitude of the entries h(cid:96)m (and,
hence, of the entries in the error matrix ES) in our context
where GS is arbitrary.
M is stable.
As a matter of fact, as stated in Theorem 2 (proved in
Appendix C) the distance affecting h(cid:96)m is the one between
(cid:96) and m on a transformed graph, GS(cid:61)S, which is the graph
obtained from G by removing all the edges connecting nodes
inside the observable subset S, introduced in Sec. II-C -- refer
to Figure 6 for a graphical illustration of the contrast between
G and GS(cid:61)S. Note that the edges possibly connecting nodes
from S to nodes in S(cid:48) are not removed in the graph GS(cid:61)S.
Theorem 2. Given two distinct nodes (cid:96), m ∈ S(cid:48), we have that:
(69)
ρr
1 − ρ2
where h(cid:96)m is the ((cid:96), m)-th entry of the matrix
δ(cid:96),m(GS(cid:61)S) = r ⇒ h(cid:96)m ≤
H = (IS(cid:48) − BS(cid:48))−1 ,
(70)
BS(cid:48) = (cid:2)A2(cid:3)S(cid:48) and 0 < ρ < 1 is an upper-bound for the
maximum row-sum of the matrix A, in view of Property 1,
remarking that A is a combination matrix satisfying Proper-
ties 1 and 2, i.e., obtained from any weight assignment in the
class Cρ,τ , as, e.g., the Metropolis and the Laplacian weight
assignment rules.
In words, Theorem 2 relates the magnitude of the entries
of H with the distance between nodes in a manner that does
not depend on the subnetwork GS. We remark that we do not
assume that the nodes in GS are not connected among each
other. In fact, we impose no restrictions whatsoever on the
topology of GS to prove the main theorem. Reference to the
graph GS(cid:61)S is used only when devising universal bounds on
the terms h(cid:96)m, in view of Theorem 2. In other words, we are
18
able to rule out the role of the subnetwork topology GS in as
much as computing upper bounds for H.
In the next step, we will show in detail how (69) is helpful
to control the size of the error in (66).
Step 2: Large distances vs. small distances. The summation
appearing in (66) can be restricted to nodes that obey the
conditions:
j (G),
m ∈ N(2)
(cid:96) ∈ Ni(G),
(71)
namely, to nodes (cid:96) ∈ S(cid:48) that are neighbors of the node i ∈ S
(so that ai(cid:96) > 0), and to nodes m ∈ S(cid:48) that are second-
order neighbors of the node j ∈ S (so that bmj > 0).
Henceforth, we refer to such pair ((cid:96), m) as an active pair.
Figure 5 depicts the possible configurations of the active
pairs. In words, the summation characterizing the error in
equation (66) runs only over the active pairs. In fact, the error
in (66) can be represented as:
N pN eij = N pN (cid:88)(cid:96),m∈S(cid:48)
= N pN (cid:88)(cid:96),m∈S(cid:48)
ai(cid:96)h(cid:96)mbmj
ai(cid:96)J(cid:96)mh(cid:96)mbmj,
(72)
(73)
where the randomness of the various quantities, arising from
the randomness of the underlying random graph G, has been
now emphasized through the boldface notation, and where we
have introduced the variable:
J(cid:96)m (cid:44) I
{ai(cid:96)>0, bmj >0} = I(cid:110)
(cid:111),
(cid:96)∈Ni(G), m∈N(2)
j
(G)
(74)
as the indicator of an active pair ((cid:96), m) ∈ S(cid:48)×S(cid:48), i.e., J(cid:96)m = 1
if ((cid:96), m) is an active pair and J(cid:96)m = 0 otherwise. Now, in
order to prove part ii) of Theorem 1, we need to prove that,
for two non-interacting nodes i and j, and for any > 0:
P[N pN eij > ] N→∞−→ 0.
(75)
As stated in Theorem 2, in Step 1, the distance between nodes
(cid:96), m ∈ S(cid:48) on the aforementioned reference graph, GS(cid:61)S, plays
a role in the size of h(cid:96)m and hence, in the magnitude of the
error. In addition, we have seen that the relevant nodes are
those obeying (71), i.e., the active pairs. It is therefore useful
to introduce the following events. For (cid:96), m ∈ S(cid:48), with (cid:96) (cid:54)= m,
we define:
D(cid:96),m (cid:44) {δ(cid:96),m(GS(cid:61)S) ≤ rN , (cid:96) ∈ Ni(G), m ∈ N(2)
j (G)},
(76)
where rN is a certain sequence of distances, with rN → ∞
as N → ∞, in a way that will be specified later. The event
in (76) certifies that the distance on the graph GS(cid:61)S between
two distinct nodes, (cid:96), m ∈ S(cid:48), does not exceed a prescribed
value rN , and also certifies the membership of the nodes (cid:96) and
m to the pertinent neighborhoods defined on the graph G, i.e.,
it certifies that ((cid:96), m) is an active pair. We remark that D(cid:96),m
is, formally, a (measurable) set and that the only random
object characterizing D(cid:96),m in equation (76) is the random
graph G -- refer to Remark 1 in Sec. II-B. The observable
subset S, the sequence rN and the indexes i, j, (cid:96), m are fixed
(or deterministic). Accordingly, D(cid:96),m represents the set of
distance) we show that large distances6 imply small errors,
formally:
19
P[N pN eij > , D(cid:48)small] = 0 for sufficiently large N .
(83)
Equations (82) and (83) will imply the desired result in equa-
tion (75) in view of equation (80). Let us start by proving (83).
From the definition in equation (78), we have that
D(cid:48)small = (cid:92)(cid:96),m∈S(cid:48)
D(cid:48)(cid:96),m.
(84)
Using (76) and (77), from (84) we conclude that the comple-
mentary event D(cid:48)small can be compactly expressed through the
indicator variables in (74) as follows:
D(cid:48)small =(cid:110)J(cid:96)mJ (δ)
(cid:96),m = 0 for all (cid:96), m ∈ S(cid:48)(cid:111) ,
J (δ)
(cid:96),m (cid:44) I
J (δ)
(cid:96),(cid:96) = 1,
where we have further introduced the indicator variable:
{δ(cid:96),m(GS(cid:61)S)≤rN} ∀(cid:96) (cid:54)= m.
(86)
That is, the event D(cid:48)small represents the set of all realizations
of the random network G, where each pair ((cid:96), m) ∈ S(cid:48) × S(cid:48)
is either non-active or, if active, then (cid:96) is distant from m,
i.e., δ(cid:96)m(G) > rN . We now show that, in view of (69), the
occurrence of D(cid:48)small implies an upper bound on the entries
h(cid:96)m, namely,
(85)
D(cid:48)small ⊆(cid:26)h(cid:96)mJ(cid:96)m ≤
ρrN +1
1 − ρ2 J(cid:96)m for all (cid:96), m ∈ S(cid:48)(cid:27)
(87)
Indeed, we know from (85) that the occurrence of D(cid:48)small
implies that the product J(cid:96)mJ (δ)
(cid:96),m is equal to zero for all
(cid:96), m ∈ S(cid:48).
Let us consider first the degenerate case (cid:96) = m. Since
J (δ)
(cid:96),(cid:96) = 1, the variable J(cid:96)m must be equal to zero and (87)
holds trivially.
We switch to the case (cid:96) (cid:54)= m. If J(cid:96)m = 0, i.e., ((cid:96), m) is not
an active pair, then (87) holds trivially. If, instead, J(cid:96)m = 1,
then we must have J (δ)
(cid:96),m = 0, i.e.,
δ(cid:96),m(GS(cid:61)S) ≥ rN + 1.
(88)
As a consequence, Eq. (87) holds true in view of (69) (proved
in Theorem 2). Applying now (87) to (73), we conclude that,
when D(cid:48)small occurs, we must have that:
N pN eij ≤ N pN
≤ N pN
ρrN +1
1 − ρ2 (cid:88)(cid:96)∈S(cid:48)
1 − ρ2 ,
ρrN +4
ai(cid:96) (cid:88)m∈S(cid:48)
bmjJ(cid:96)m (89)
where the last inequality holds from the row-sum stability of
A ∈ Cρ,τ (Property 1):
A∞ ≤ ρ,
B∞ ≤ ρ2.
(90)
Accordingly, from (90) we have that:
P[N pN eij > , D(cid:48)small] ≤ P(cid:20)N pN
ρrN +4
1 − ρ2 > (cid:21) ,
(91)
6The terminology "small distances" and "large distances" will be of-
ten coined for simplicity to denote min((cid:96),m)is active δ(cid:96)m (G) ≤ rN and
min((cid:96),m)is active δ(cid:96)m (G) > rN , respectively.
Illustration of random realizations with ω1 ∈ D(cid:96),m and ω2, ω3 ∈
Fig. 13.
D(cid:48)
(cid:96)m. This emphasizes that other than the (edges of the) random graph G,
all other quantities, i.e., rN , i, j, (cid:96) and m are fixed (or deterministic).
realizations of the partial Erdos-R´enyi random graph G where
the constraints of distance and neighborhood among the fixed
nodes i, j, (cid:96), m in equation (76) are met. Refer to Figure 13
for an illustration.
Likewise, for (cid:96) = m the set D(cid:96),m reduces to:
D(cid:96),(cid:96) (cid:44) {(cid:96) ∈ Ni(G) ∩ N(2)
j (G)},
(77)
where we see that the event D(cid:96),(cid:96) simply certifies the member-
ship of the node (cid:96) to the pertinent neighborhoods defined on
the graph G. We finally introduce the union event:
Dsmall (cid:44) (cid:91)(cid:96),m∈S(cid:48)
D(cid:96),m,
(78)
as the event where the distance is small, i.e., δ(cid:96)m (G) ≤ rN ,
for at least one active pair ((cid:96), m) ∈ S(cid:48) × S(cid:48). We can write:
(79)
P[N pN eij > ] = P[N pN eij > , Dsmall]
+P[N pN eij > , D(cid:48)small]
≤ P[Dsmall] + P[N pN eij > , D(cid:48)small],
where D(cid:48)small is the complement of the event (or measurable
set) Dsmall, i.e., D(cid:48)small ∩ Dsmall = ∅, and can be read as the
event (or set of realizations of G) where the distances are
large, i.e., δ(cid:96)m (G) > rN , for all active pairs ((cid:96), m). For the
sake of a compact notation, we write
instead of
P[N pN eij > , D(cid:48)small]
P[{N pN eij > } ∩ D(cid:48)small].
(80)
(81)
Then, the route that we follow to prove the claim in (75) goes
as follows: i) (small-distance) we show that:
P[Dsmall] N→∞−→ 0,
(82)
i.e., the occurrence of a small distance in at least one active
pair ((cid:96), m) is rare, with high probability, and ii) (large-
20
where we remark that the event appearing in the latter prob-
ability is in fact a deterministic event. If we now find a
sequence rN that drives to zero the quantity N pN ρrN +4, then,
for sufficiently large N, the probabilities appearing in (91)
are eventually zero. We will illustrate how to make a proper
selection of rN in the final step (i.e., Step 5) of this proof. It
suffices for now to assume that such a sequence exists, namely,
that:
N pN ρrN +4 N→∞−→ 0
(92)
in view of (91), yields the desired claim in (83).
which,
Therefore, we conclude that, if nodes in S(cid:48) forming active
pairs, i.e., obeying (71), lie sufficiently far apart on the graph
GS(cid:61)S, then the magnified error can be driven to zero as
N → ∞. This observation corroborates the claim that large
distances imply small values of the error. In light of (80),
the claim of Theorem 1 will be proved if we show that the
occurrence of a small distance on at least one active pair ((cid:96), m)
is a rare event, namely, if we prove the claim in (82). We will
address this challenge in the two forthcoming steps.
Step 3: Relating partial Erdos-R´enyi to a standard Erdos-
R´enyi via homogenization. Two sources of asymmetry make
the proof of (82) challenging. First, we see that the events
in (76) refer to different graphs, namely, GS(cid:61)S and G --
the graph G for the neighborhood constraint and GS(cid:61)S for
the distance constraint. Second, both the local disconnection
implied by GS(cid:61)S, and the partial Erdos-Renyi construction
implied by G (recall Figure 6), introduce additional non-
homogeneity across nodes that makes the estimation of the
probability of the event D(cid:96)m in (76), and hence the estimation
of the probability in (82), rather intricate.
In order to overcome this issue, Theorem 3 further ahead
states that, without loss of generality, we can replace the
events D(cid:96)m in (76), by the events
(cid:101)D(cid:96),m (cid:44) {δ(cid:96),m((cid:101)G) ≤ rN , (cid:96) ∈ Ni((cid:101)G), m ∈ N(2)
where (cid:101)G ∼ G (cid:63)(N,(cid:101)pN ) is a standard Erdos-R´enyi graph with
(cid:101)pN = SpN (for sufficiently large N), in that if we prove the
j ((cid:101)G)},
convergence
(93)
(94)
then, the convergence in (82) holds, where we have defined
P[(cid:101)Dsmall] −→ 0,
(cid:101)Dsmall (cid:44) (cid:91)(cid:96),m∈S(cid:48)(cid:101)D(cid:96),m.
D(cid:96),m ⊆ (cid:101)D(cid:96),m
Dsmall ⊆ (cid:101)Dsmall.
P[Dsmall] ≤ P[(cid:101)Dsmall],
Fig. 14. If there is a path P(cid:96)m between (cid:96) and m of rN hops (dashed path),
then there is a path of length rN + 3 hops (red path) connecting i to j:
one can reach (cid:96) in one hop from i (since (cid:96) ∈ Ni((cid:101)G)), follow the path P(cid:96)m
((cid:101)G)).
to reach m and reach j in two more hops from m (since m ∈ N
Therefore, the distance between i and j cannot exceed rN + 3.
(2)
j
probability for the original (non-homogeneous) case. We refer
to this coupling procedure as homogenization as the conditions
We now state Theorem 3 and prove it in Appendix D.
the inhomogeneity that characterizes the events Dsmall is not
characterizing the event (cid:101)Dsmall refer to the same graph (cid:101)G
and, further the graph (cid:101)G is a standard Erdos-R´enyi. That is,
present in the new event (cid:101)Dsmall.
Theorem 3 (Coupling and homogenization). Let G ∼
G (cid:63)(N, pN ; GS) be a partial Erdos-R´enyi random graph, and
let (cid:101)G be a pure Erdos-R´enyi random graph (cid:101)G ∼ G (cid:63)(N,(cid:101)pN )
where, for N sufficiently large:
(99)
If i, j ∈ S are non-interacting (gij = 0), then we have that:
(100)
log N +(cid:101)cN
(cid:101)pN = SpN =
P[Dsmall] ≤ P[(cid:101)Dsmall]
N
.
Step 4: Managing the small-distance pairs. The final step
to prove Theorem 1, in view of inequality (100), consists in
proving that (82) holds true on the homogenized graph (cid:101)G ∼
G (cid:63)(N,(cid:101)pN ), namely, that:
(101)
Using (93) and (95), we observe that:
P[(cid:101)Dsmall] N→∞−→ 0.
(cid:101)Dsmall ⊆(cid:110)δi,j((cid:101)G) ≤ rN + 3(cid:111) ,
i.e., if the distance between (cid:96), m of an active pair ((cid:96), m) is
bounded by rN , then, since (cid:96) and m are neighbor and second
order neighbor of i and j, respectively, the distance between
the nodes i and j cannot exceed rN + 3. Refer to Figure 14
for an illustration. Therefore,
P(cid:104)(cid:101)Dsmall(cid:105) ≤ P(cid:104)δi,j((cid:101)G) ≤ rN + 3(cid:105) ,
and the estimation of P(cid:104)δi,j((cid:101)G) ≤ rN + 3(cid:105) amounts to a
standard analysis of distance scaling on Erdos-R´enyi random
(103)
(102)
(95)
(97)
(98)
This is accomplished, in the proof of Theorem 3, via con-
with G in the sense that
structing a graph (cid:101)G that is Erdos-R´enyi and that is coupled
for all (cid:96), m ∈ S(cid:48) and hence,
(96)
Therefore, the induced coupling yields:
implying that if one is able to prove that the probability for
the homogeneous case vanishes, equation (94), so does the
-hops-hops (included in Appendix E for completeness) asserts that
graphs as the graph (cid:101)G is a pure Erdos-R´enyi. In fact, Lemma 3
P(cid:104)δi,j((cid:101)G) ≤ rN + 3(cid:105) ≤(cid:101)pN (N(cid:101)pN )rN +2(cid:18)
1 − 1/(N(cid:101)pN )(cid:19) .
large, we conclude from (104) that P(cid:104)(cid:101)Dsmall(cid:105) vanishes if we
(104)
Now, since by assumption rN → ∞ and in the Erdos-R´enyi
regime that we are assuming we have N pN → ∞, as N grows
are able to choose a sequence rN yielding:
1
(105)
(cid:101)pN (N(cid:101)pN )rN +2 N→∞−→ 0
Note that the requirement (92) implies that rN cannot diverge
too slow, whereas the requirement in (105) implies that rN
cannot diverge too fast. The next step illustrates how to choose
a sequence rN , with rN → ∞, yielding both (92) and (105).
Remark 2 (More on homogenization). Ultimately, the homog-
enization in Theorem 3 reduces the estimation of P [Dsmall]
to a simple analysis of distance scaling between only one
pair of nodes (i, j) in a pure Erdos-R´enyi random graph, in
view of the subset inequalities (97) and (102). Note that, to
prove the convergence (82), one may be tempted to directly
apply the following inequality (instead of invoking the extra
homogenization, inequality (97), granted by Theorem 3)
21
Fig. 15. Counter-example: the inequality (107) does not hold for Gref :=
GS(cid:61)S.
a certain sequence of distances, rN , exists that fulfills the two
conditions in (92) and (105). We start by examining these
conditions in more detail. Taking the logarithm of the functions
appearing in (92), we get:
log(cid:0)N pN ρrN +4(cid:1)
= log(log N + cN ) + (rN + 4) log(ρ)
N→∞−→ −∞,
(110)
Dsmall ⊆ {δi,j(G) ≤ rN + 3} ,
(106)
where we have used the expression for pN in (9). Observing
that log(ρ) < 0 since ρ < 1, and letting:
for the original heterogeneous event Dsmall and with G in the
RHS instead of the pure Erdos-R´enyi (cid:101)G. But the probabil-
ity P [{δi,j(G) ≤ rN + 3}] does not converge to zero as N
grows large in this case as the distance δi,j(G) depends on
GS which is arbitrary. Therefore, the inequality (106) is not
useful to establish the convergence (82). In this line of thought,
one can attempt to find a reference graph Gref, if any, so that
Dsmall ⊆ {δi,j(Gref) ≤ rN + 3} ,
(107)
and for which P [{δi,j(Gref) ≤ rN + 3}] converges to zero.
Another natural candidate is Gref := GS(cid:61)S, but the referred
inequality does not hold in this case, i.e.,
Dsmall (cid:42) {δi,j(GS(cid:61)S) ≤ rN + 3} .
(108)
Refer to Figure 15 for a graphical counter-example on this.
One can attempt
to simply fill
in the gap of GS(cid:61)S by
considering the reference graph
(109)
Gref := (cid:101)GS ⊕ GS(cid:61)S
where (cid:101)GS ∼ G (cid:63)(S,(cid:101)pN ), but this also does not work as the
inequality (107) is not satisfied (for the same reason as GS(cid:61)S
fails to meet it).
The homogenization in Theorem 3 provides a careful con-
struction of a reference graph that allows us to lip this
difficulty and rigorously reduce the computation of P [Dsmall]
to a simple distance-scaling of a particular pair of nodes in
an Erdos-R´enyi random graph. Such construction may be of
independent interest.
Step 5: Choosing the sequence rN . In the above steps
(specifically, in Step 1 and Step 4), we have maintained that
α (cid:44) log(ρ),
from (110) we can write:
ωN (cid:44) log(log N + cN ),
(111)
ωN − α rN − 4α N→∞−→ −∞
(112)
Let us switch to the analysis of (105). Taking the logarithm
of the functions appearing in (105) we can write:
=
=
N
(cid:19)
log(cid:18) (N(cid:101)pN )rN +3
log(cid:0)(cid:101)pN (N(cid:101)pN )rN +2(cid:1)
(rN + 3) log(N(cid:101)pN ) − log N
which, since (cid:101)pN = SpN , in view of (9) and (111), yields:
(rN + 3)[log S + ωN ] − log N, N→∞−→ −∞
N→∞−→ −∞,
(113)
(114)
We first show why the assumption in (40) is related to (112)
and (114). From (112) and (114) we conclude that, for
sufficiently large N, we must necessarily have:
ωN
α
which in turn implies:
< rN <
log N
ωN
,
(115)
ω2
N
log N
< α ⇔
[log(log N + cN )]2
log N
< log(ρ),
(116)
where in the last step we used the definitions in (111). Now,
if we want to guarantee the verification of (116) irrespectively
of the particular value of 0 < ρ < 1, we need to enforce
condition (40).
andbut, Next we illustrate how to choose a sequence rN that, under
assumption (40), fulfills simultaneously (112) and (114). We
set:
rN =(cid:22) 1
2
log N
ωN (cid:23) .
(117)
Substituting (117) into (112), and observing that (cid:98)x(cid:99) > x− 1,
where (cid:98)x(cid:99) stands for the greatest integer smaller than or equal
to x, we can write:
ωN − α(cid:22) 1
ωN (cid:23) − 4α
+ α − 4α
< ωN −
log N
ωN
log N
α
2
2
= ωN
N→∞−→ −∞,
3α
ωN
−
1 −
α
2
log N
ω2
N
→∞ from (40)
(cid:124) (cid:123)(cid:122) (cid:125)
(118)
which shows that the condition in (112) is met with the choice
in (117).
Likewise, substituting (117) into (114), and observing that
2
2
2
log S
≤
=
(cid:98)x(cid:99) ≤ x, we have:
(cid:18)(cid:22) 1
log N
+ 3 log S +
+ 3(cid:19) [log S + ωN ] − log N
ωN (cid:23) + 3(cid:19) [log S + ωN ] − log N
(cid:18) 1
ωN (cid:20) log S
log N
ωN
log N
ωN
1
log N + 3 ωN − log N
2
ωN
log N −
N→∞−→ −∞,
log N(cid:21)
(119)
with the convergence holding true because ωN → ∞ as N →
∞, while ω2
N / log N and ωN / log N vanish in view of (40).
We have in fact shown that the condition in (114) is met with
the choice in (117).
+ 3 log S
ωN
2
log N
ω2
N
+ 3
=
2
Refer to Figure 16 for a summary of the proof of Theorem 1.
APPENDIX C
PROOF OF THEOREM 2
First, we start by observing that the matrix H in (65) is not
sensitive to the submatrix AS. Note that this is not immediately
clear via simple inspection, since computation of H involves
the matrix B = A2. However, note that
A =(cid:20) AS ASS(cid:48)
AS(cid:48)S AS(cid:48) (cid:21)
(120)
and from the rules for block-matrix multiplication, we can
write:
BS(cid:48) = [A2]S(cid:48) = AS(cid:48)SASS(cid:48) + (AS(cid:48))2
(121)
which highlights that the matrix BS(cid:48) does not depend on the
submatrix AS.
As a corollary to this observation, it follows that the matrix
H in (65) is not a function of the particular submatrix AS,
and, hence, it is insensitive to the particular topology of the
22
subnetwork connecting the observed agents. Since we are to
devise bounds for the terms h(cid:96)m, we can assume without loss
of generality that all the entries of AS are equal to zero7,
namely, that AS = 0S×S. In other words, we can remove
the edges among the observable agents as far as computing
bounds on H goes. This will imply that an appropriate distance
between nodes in S(cid:48) (which will play a role in bounding h(cid:96)m)
is given by δ(cid:96),m(GS(cid:61)S), namely, by the distance between
(cid:96) and m on the graph GS(cid:61)S where the edges among the
observed agents in S have been removed.
Theorem 2, proved next, establishes an exponential bound
on h(cid:96)m, which holds if (cid:96) and m are distant on GS(cid:61)S (not
necessarily on G, and hence not dependent on GS).
Proof of Theorem 2: We remind that:
H = (IS(cid:48) − BS(cid:48))−1 =
∞(cid:88)k=0
(BS(cid:48))k,
(122)
since A is row-sum stable and B = A2 (refer, e.g., to [31],
[53]). Let now (cid:101)A be the matrix obtained from A by replac-
accordingly (cid:101)B = (cid:101)A2. Since, in view of (121), modifying the
ing the submatrix AS with the void matrix, 0S×S, and let
submatrix AS does not alter the submatrix BS(cid:48), we can safely
write:
BS(cid:48) = (cid:101)BS(cid:48)
(123)
Moreover, it is known that, for any two nonnegative matrices
Q and R
[QR]S(cid:48) = QS(cid:48)SRSS(cid:48)
+QS(cid:48)RS(cid:48) ≥ QS(cid:48)RS(cid:48),
(124)
(cid:124)
≥0
(cid:123)(cid:122)
(cid:125)
with entry-wise inequality. Taking Q = R = (cid:101)B, and reasoning
by induction, we have then:
(BS(cid:48))n = ((cid:101)BS(cid:48))n ≤ [(cid:101)Bn]S(cid:48) = [(cid:101)A2n]S(cid:48),
where the first equality follows from (123). Rephrasing (125)
on an entry-wise basis, we get, for all (cid:96), m ∈ S(cid:48):
(125)
[(BS(cid:48))n](cid:96)m ≤ [(cid:101)A2n](cid:96)m.
(126)
by assumption δ(cid:96),m(GS(cid:61)S) = r, then in view of (68), the
We recall that the support graph of (cid:101)A is given by GS(cid:61)S. Since
smallest k yielding [(cid:101)Ak](cid:96)m > 0 is k = r. In view of (126),
this property implies that one can consider only the terms
[(BS(cid:48))n](cid:96)m for 2n ≥ r. Since r is not necessarily an even
number, we could in general consider all the terms for which
n ≥ (cid:100)r/2(cid:101), where (cid:100)x(cid:101) stands for the smallest integer that is
7Assuming that the diagonal entries are equal to zero contradicts our rules
for constructing a combination policy. However, this is immaterial, because
setting the AS block to zero is only a mathematical expedient to compute
suitable bounds, without any physical meaning.
23
This is a relevant assertion as estimating the probabil-
ity P[Dsmall]
is rather intricate. In fact, by examining the
events D(cid:96)m in (131), two sources of asymmetry stick out (as
opposed to the characterization of (cid:101)D(cid:96)m). First, the pertinent
distance, δ(cid:96),m(GS(cid:61)S), is computed with respect to the graph
GS(cid:61)S, while the conditions on the neighborhood member-
ships (cid:96) ∈ Ni(G) and m ∈ N(2)
j (G) characterizing the active
pairs are defined in terms of the original graph, G. Second,
both graphs G and GS(cid:61)S are non-homogeneous. That is, the
connections among nodes in S for the graph G are given by
GS -- whose topology is arbitrary and hence, it has a nature
that differs from the rest of the network G -- while in GS(cid:61)S
the connections among nodes in S are absent.
(132)
Therefore, in order to estimate P[D(cid:96)m], and hence P[Dsmall],
in the proof of Theorem 3, we appropriately modify the
structure of the partial Erdos-R´enyi G in such a way that the
resulting transformed graph (cid:101)G fulfills the following properties:
i) (cid:101)G is a homogeneous (i.e., classic) Erdos-R´enyi graph; ii)
the original event D(cid:96),m on G implies its counterpart (cid:101)D(cid:96),m
defined on the new graph (cid:101)G, i.e.,
D(cid:96),m ⊆ (cid:101)D(cid:96),m
for all (cid:96), m ∈ S(cid:48), and hence,
Dsmall ⊆ (cid:101)Dsmall.
P[Dsmall] ≤ P[(cid:101)Dsmall].
As a result, showing the convergence in (128) for the homoge-
neous system implies the convergence in (129) for the original
heterogeneous partial Erdos-R´enyi. We refer to this coupling
procedure simply as homogenization.
Remark 3. It is tempting at first glance to simply replace the
This further yields the desired inequality
graph GS(cid:61)S in (131) with the standard Erdos-R´enyi (cid:101)G by
reinforcing the coupling (cid:101)GS(cid:48) = GS(cid:48). This, in fact, contracts
(133)
(134)
Fig. 16. Diagram-summary of the proof.
greater than or equal to x. With this choice, the series in (122)
can be truncated as (the term n = 0 is zero because (cid:96) (cid:54)= m):
h(cid:96)m =
(a)
≤
(b)
≤
∞(cid:88)n=1
∞(cid:88)n≥(cid:100)r/2(cid:101)
∞(cid:88)n≥(cid:100)r/2(cid:101)
[(BS(cid:48))n](cid:96)m =
[(BS(cid:48))n](cid:96)m
∞(cid:88)n≥(cid:100)r/2(cid:101)
[A2n](cid:96)m
ρ2n (c)
=
ρ2(cid:100)r/2(cid:101)
1 − ρ2
(d)
≤
ρr
1 − ρ2 ,
(127)
where inequality (a) follows by using (126) with A in place
of (cid:101)A; inequality (b) follows due to (31) as we know that
(cid:80)N
(cid:96)=1[A2n](cid:96)m ≤ ρ2n; equality (c) is the partial sum of the
geometric series; and inequality (d) follows from the known
bound (cid:100)r/2(cid:101) ≥ r/2.
APPENDIX D
HOMOGENIZATION
In this section, we prove that if small distances are rare over
a pure (i.e., standard) Erdos-R´enyi graph (cid:101)G, i.e.,
then small distances are also rare over the partial Erdos-
R´enyi G, i.e.,
P[(cid:101)Dsmall] N→∞−→ 0,
P[Dsmall] N→∞−→ 0,
where we recall the definitions
with
Dsmall (cid:44) (cid:91)(cid:96),m∈S(cid:48)
(cid:101)Dsmall (cid:44) (cid:91)(cid:96),m∈S(cid:48)(cid:101)D(cid:96),m,
(cid:101)D(cid:96),m (cid:44) {δ(cid:96),m((cid:101)G) ≤ rN , (cid:96) ∈ Ni((cid:101)G), m ∈ N(2)
j ((cid:101)G)},
D(cid:96),m (cid:44) {δ(cid:96),m(GS(cid:61)S) ≤ rN , (cid:96) ∈ Ni(G), m ∈ N(2)
D(cid:96),m,
(128)
(129)
(130)
j (G)}.
(131)
Homogenization(Theorem 3)Smalldistances are rare(Lemma 3)Largedistances on imply small error (Theorem 2)Under proper choice of 24
Fig. 17. The original graph G (leftmost panel); the locally-disconnected graph, GS(cid:61)S (middle panel); the transformed graph, G (cid:44) (GS(cid:61)S)j←(cid:101)Sij
from GS(cid:61)S when all the connections from(cid:101)Sij to S(cid:48) are inherited by the node j.
, obtained
(135)
implication
δ(cid:96)m(GS(cid:61)S) ≤ rN ⇒ δ(cid:96)m((cid:101)G) ≤ rN
the distance, i.e., δ(cid:96)m((cid:101)G) ≤ δ(cid:96)m(GS(cid:61)S) which yields the
at the same time as the new graph (cid:101)G is homogeneous, but
the neighborhood constraint m ∈ N(2)
j ((cid:101)G) is jeopardized as
it is not implied by its counterpart on G, since the condition
m ∈ N(2)
j (G) also depends on GS, which is not Erdos-R´enyi,
but arbitrary (hence, the subset inclusion (133) does not follow
from this simple homogenization). The structure modification
on the original graph G is carefully performed in the proof
of Theorem 3 to both grant the contraction of the distances at
the same time as preserving the neighborhood constraints.
Proof of Theorem 3:
In Figure 17, middle panel, we
The basic trick that allows homogenization is defining a new
display the graph GS(cid:61)S. Moreover, we denote by(cid:101)Sij the set
S deprived of the nodes i and j, namely, (cid:101)Sij (cid:44) S \ {i, j}.
graph where all the connections from (cid:101)Sij to S(cid:48) are inherited
by the node j. The transformed graph is denoted by (and is
displayed in the rightmost panel of Figure 17):
G (cid:44) (GS(cid:61)S)j←(cid:101)Sij
(136)
This operation achieves the twofold goal of ensuring that i)
the distance δ(cid:96),m(GS(cid:61)S) between any two nodes (cid:96) and m in
S(cid:48) is reduced, namely,
δ(cid:96),m(G) ≤ δ(cid:96),m(GS(cid:61)S)
(137)
and ii) if node m is second-order neighbor of j on the original
graph G, so is on the transformed graph, namely,
m ∈ N(2)
j (G) ⇒ m ∈ N(2)
j (G)
(138)
Note that equations (137) and (138) induce the desired cou-
pling between the transformed graph, G, and the graphs
GS(cid:61)S, G, in that for all (cid:96), m ∈ S(cid:48):
(cid:110)δ(cid:96),m(GS(cid:61)S) ≤ rN , (cid:96) ∈ Ni(G), m ∈ N(2)
j (G)(cid:111)
j (G)(cid:111) .
(cid:110)δ(cid:96),m(G) ≤ rN , (cid:96) ∈ Ni(G), m ∈ N(2)
⊆
(139)
At this point, we observe that G is still not homogeneous (in
The inequality (137) stems from the following observation.
into a path traversing node j (instead of traversing the cor-
hence, the proof is not finished. Before proceeding on this
point, we first justify equations (137) and (138).
particular, the nodes in (cid:101)Sij on the graph G are isolated) and
The only modification in GS(cid:61)S to get G is related to (cid:101)Sij.
Therefore, if there exists a path from (cid:96) ∈ S(cid:48) to m ∈ S(cid:48) on
GS(cid:61)S, which flows through (cid:101)Sij, such path (or a shortened
version thereof) is also present in G, but now via j. Refer to
Figure 18 for an illustration.
Indeed, each path on GS(cid:61)S hopping across(cid:101)Sij is mapped
responding nodes in (cid:101)Sij), since the node j has inherited all
connections between(cid:101)Sij and S(cid:48).
to j through an intermediate node belonging to (cid:101)Sij, then it
is connected to j in one step on the graph G. One difficulty
might arise if, on graph G, node m is connected to j through
node i, because on G nodes i and j are disconnected. This is
not a problem, however, because to prove our result we need
to examine only the case that i and j are disconnected on the
original graph G (as stated in the theorem). We remark that,
The neighborhood implication (138) results from the follow-
ing observation. If on the graph G, the node m is connected
25
Fig. 19. Homogenizing the graph G. In the graph G, the node j is connected
e.g., the nodes in (cid:101)Sij are not connected among each other. New Bernoulli
to a particular node in S(cid:48) with probability pN = 1− (1− pN )S−1 whereas,
realizations are performed so that the probability of any pair of nodes (u, v) ∈
N = 1− (1− pN )S−1 in G(cid:63). This is formalized
S× S being connected is p(cid:63)
in equations (143) and (144). The final graph G(cid:63) is Erdos-R´enyi.
Moreover, since 1−(1−pN )S−1 ≤ SpN , and in order to obtain
a random graph whose connection probability is explicitly
given by (9), we can further define a graph (cid:101)G with connection
probability given by
(145)
log N + cN
=
(cid:101)pN = SpN = S
N
log N +(cid:101)cN
N
,
uv ≤
via a standard coupling between Bernoulli random variables.
where(cid:101)cN = (S−1) log N +ScN , and with the coupling g(cid:63)
(cid:101)guv (realization-wise) for all u, v -- this can be easily obtained
Therefore G ⊆ (cid:101)G, i.e., G is a subgraph of (cid:101)G, realization-
replaced by (cid:101)G. This implies, in view of (93) and (95), that:
wise. Since fleshing out a graph with new connections can only
decrease distances and favor membership to any neighborhood,
the implications shown in (139) and (140) hold true with G
(146)
which in turn implies the claim of the theorem.
Dsmall ⊆ (cid:101)Dsmall,
APPENDIX E
MANAGING SMALL DISTANCES
Lemma 3. Let G be a pure Erdos-R´enyi random graph
G (N, pN ). We have that
1
Proof:
1 − 1/(N pN )(cid:19) (147)
P[δi,j(G) ≤ r] ≤ pN (N pN )r−1(cid:18)
Since the event {δi,j(G) = r} signifies that the shortest
path connecting i to j has length equal to r, there must exist
a path connecting i to j obeying the following conditions: i)
all intermediate nodes are visited only once through the path
Fig. 18. Illustration of a particular path (blue color) connecting (cid:96) to m in the
graph GS(cid:61)S on the LHS of the figure. The edges linking to nodes in(cid:101)Sij on
nodes in(cid:101)Sij in the graph GS(cid:61)S, only crosses j in the new graph G.
the graph GS(cid:61)S, link only to j on the graph G. A path that crosses M ≥ 1
for the case that (cid:96) = m, condition (137) is redundant and
(cid:110)(cid:96) ∈ Ni(G) ∩ N(2)
j (G)(cid:111) ⊆(cid:110)(cid:96) ∈ Ni(G) ∩ N(2)
(140)
Now, we return to the observation that the transformed
graph, G, is still asymmetrical, because, apart from the fact
that the set S contains disconnected nodes, the probability that
the node j is connected to a node in S(cid:48) is now augmented as j
j (G)(cid:111) .
node i):
1 − (1 − pN )S−1.
partial Erdos-R´enyi construction, these connections follow a
standard Bernoulli law, we conclude that the probability of j
being connected to a particular node in S(cid:48), in the new random
inherited all the connections from (cid:101)Sij to S(cid:48). Since under the
graph G, is simply given by (recall that(cid:101)Sij does not contain
To see why, j is not connected to a particular node in S(cid:48), say
k ∈ S(cid:48), in the graph G, if and only if, j is not connected to
k in G, i.e., gjk = 0, and gαk = 0 for all α ∈(cid:101)Sij. In other
words,
P(cid:2)gjk = 0(cid:3) = P(cid:104)gjk = 0, gαk = 0∀α ∈(cid:101)Sij(cid:105)
(142)
Hence, gjk = 1 with probability pN = 1 − (1 − pN )S−1. But
now homogenizing the transformed graph G is an easy task.
It suffices to augment the connection probabilities of all the
remaining pairs (including those in S), in order to match the
connection probability in (141). More formally, resorting to a
simple coupling between Bernoulli random variables, we can
define a new (random) graph G(cid:63) from G as
= (1 − pN )S−1.
(141)
g(cid:63)
uv = max{guv, quv}
(143)
where {quv}u<v are i.i.d. Bernoulli random variables with
P [quv = 1]=
1 − (1 − pN )S−1, If u ∈(cid:101)Sij
1 − (1 − pN )S−1, If (u, v) = (i, j)
1 − (1 − pN )S−2, If u ∈ {i} ∪ S(cid:48) and v ∈ S(cid:48)
(144)
The resulting graph G(cid:63) is Erdos-R´enyi with p(cid:63)
N = 1 −
(1 − pN )S−1. Figure 19 graphically summarizes the idea.
If u = j and v ∈ S(cid:48).
0,
(otherwise the path itself could be squeezed to one of a shorter
length); ii) along the path, one cannot spend one or more steps
lingering on the same node. Accordingly, we can write:
{δi,j(G) = r} ⊆ E (cid:44)(cid:91)M(cid:8)gin1gn1n2 ··· gnr−1j = 1(cid:9) ,
where the set M is defined as M (cid:44) M1 ∩ M2 with
(148)
M1(cid:44){n := (n1, . . . , nr−1) ∈ Nr−1 : nu (cid:54)= nv ∀u, v},(149)
M2(cid:44){n := (n1, . . . , nr−1) ∈ Nr−1 : nk (cid:54)= i, j ∀k}. (150)
It is useful to remark that the event E in (148) does not
coincide with the event that the shortest path has length equal
to r, because the possibility of having paths longer than r is not
ruled out. The event E in (148) simply underlies the existence
of at least one path of length r with the necessary character-
istics, which explains the one-sided implication in (148), and
yields P[δ(cid:96),m(G) = r] ≤ P[E]. We have
P [E] ≤(cid:88)M
P[gin1gn1n2 ··· gnr−1j = 1] = M pr
N .
where we recall that M stands for the cardinality of the set M
in view of the notation in Sec. II-A, where sets are represented
by calligraphic letters and the corresponding cardinalities are
represented by normal font letters. Observe that
(151)
M = (N − 2)(N − 3) . . . (N − r) ≤ (N − 2)r−1.
(152)
Therefore,
P [δi,j(G) = r] ≤ (N − 2)r−1pr
(153)
and as a result,
P[δi,j(G) ≤ r] =
r(cid:88)α=1
r−1(cid:88)α=0
N ≤ pN (N pN )r−1
r(cid:88)α=1
P[δi,j(G) = α] ≤ pN
(N pN )α−1
= pN
(N pN )α ≤
pN (N pN )r−1
1 − 1/(N pN )
. (154)
ACKNOWLEDGMENT
Short and limited versions of this work appear in the
conference publications [51] and [52].
REFERENCES
[1] A. Barrat, M. Barth´elemy, and A. Vespignani, Dynamical Processes
London, UK: Cambridge University Press,
on Complex Networks.
November 2012.
Berlin Heidelberg, 2005.
[2] T. Liggett, Interacting Particle Systems, 1st edition.
Springer-Verlag
[3] P. Robert, Stochastic Networks and Queues. Springer-Verlag, 2003.
[4] M. Porter and J. Gleeson, Dynamical Systems on Networks: A Tutorial.
Springer International Publishing, 2016.
[5] L. Xiao and S. Boyd, "Fast linear iterations for distributed averaging,"
Systems and Control Letters, vol. 53, no. 1, pp. 65 -- 78, Sept 2004.
[6] J. Tsitsiklis, D. Bertsekas, and M. Athans, "Distributed asynchronous
deterministic and stochastic gradient optimization algorithms," IEEE
Transactions on Automatic Control, vol. 31, no. 9, pp. 803 -- 812, Sep
1986.
[7] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, "Randomized gossip
algorithms," IEEE Transactions on Information Theory, vol. 52, no. 6,
pp. 2508 -- 2530, June 2006.
26
[8] A. G. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and A. Scaglione,
"Gossip algorithms for distributed signal processing," Proceedings of the
IEEE, vol. 98, no. 11, pp. 1847 -- 1864, Nov 2010.
[9] D. Bajovic, D. Jakovetic, J. Xavier, B. Sinopoli, and J. M. F. Moura,
"Distributed detection via Gaussian running consensus: Large deviations
asymptotic analysis," IEEE Transactions on Signal Processing, vol. 59,
no. 9, pp. 4381 -- 4396, Sept 2011.
[10] D. Bajovic, D. Jakovetic, J. M. F. Moura, J. Xavier, and B. Sinopoli,
"Large deviations performance of consensus+innovations distributed
detection with non-Gaussian observations," IEEE Transactions on Signal
Processing, vol. 60, no. 11, pp. 5987 -- 6002, Nov 2012.
[11] S. Kar and J. M. F. Moura, "Convergence rate analysis of distributed
gossip (linear parameter) estimation: Fundamental limits and tradeoffs,"
IEEE Journal of Selected Topics in Signal Processing, vol. 5, no. 4, pp.
674 -- 690, Aug 2011.
[12] P. Braca, S. Marano, and V. Matta, "Enforcing consensus while moni-
toring the environment in wireless sensor networks," IEEE Transactions
on Signal Processing, vol. 56, no. 7, pp. 3375 -- 3380, July 2008.
[13] P. Braca, S. Marano, V. Matta, and P. Willett, "Asymptotic optimality
of running consensus in testing binary hypotheses," IEEE Transactions
on Signal Processing, vol. 58, no. 2, pp. 814 -- 825, Feb 2010.
[14] A. H. Sayed, "Adaptive networks," Proceedings of the IEEE, vol. 102,
no. 4, pp. 460 -- 497, April 2014.
[15] J. Chen and A. H. Sayed, "On the learning behavior of adaptive networks
-- part i: Transient analysis," IEEE Transactions on Information Theory,
vol. 61, no. 6, pp. 3487 -- 3517, June 2015.
[16] -- -- , "On the learning behavior of adaptive networks -- part ii: Per-
formance analysis," IEEE Transactions on Information Theory, vol. 61,
no. 6, pp. 3518 -- 3548, June 2015.
[17] C. G. Lopes and A. H. Sayed, "Diffusion least-mean squares over
adaptive networks: Formulation and performance analysis," IEEE Trans-
actions on Signal Processing, vol. 56, no. 7, pp. 3122 -- 3136, July 2008.
[18] F. S. Cattivelli and A. H. Sayed, "Diffusion lms strategies for distributed
estimation," IEEE Transactions on Signal Processing, vol. 58, no. 3, pp.
1035 -- 1048, March 2010.
[19] -- -- , "Distributed detection over adaptive networks using diffusion
adaptation," IEEE Transactions on Signal Processing, vol. 59, no. 5,
pp. 1917 -- 1932, May 2011.
[20] A. H. Sayed, S. Y. Tu, J. Chen, X. Zhao, and Z. J. Towfic, "Diffusion
strategies for adaptation and learning over networks," IEEE Signal
Processing Magazine, vol. 30, no. 3, pp. 155 -- 171, May 2013.
[21] V. Matta, P. Braca, S. Marano, and A. H. Sayed, "Diffusion-based
adaptive distributed detection: Steady-state performance in the slow
adaptation regime," IEEE Transactions on Information Theory, vol. 62,
no. 8, pp. 4710 -- 4732, Aug 2016.
[22] -- -- , "Distributed detection over adaptive networks: Refined asymp-
totics and the role of connectivity," IEEE Transactions on Signal and
Information Processing over Networks, vol. 2, no. 4, pp. 442 -- 460, Dec
2016.
[23] A. Ganesh, L. Massouli´e, and D. Towsley, "The effect of network
topology on the spread of epidemics," in Proceedings IEEE 24th Annual
Joint Conference of the IEEE Computer and Communications Societies.,
vol. 2, March 2005, pp. 1455 -- 1466 vol. 2.
[24] F. Morone, K. Roth, B. Min, H. E. Stanley, and H. A. Makse, "Model of
brain activation predicts the neural collective influence map of the brain,"
Proceedings of the National Academy of Sciences, vol. 114, no. 15, pp.
3849 -- 3854, 2017.
[25] C. Stam, B. Jones, G. Nolte, M. Breakspear, and P. Scheltens, "Small-
World Networks and Functional Connectivity in Alzheimer's Disease,"
Cerebral Cortex, vol. 17, no. 1, pp. 92 -- 99, 2007.
[26] S. Monajemi, K. Eftaxias, S. Sanei, and S. H. Ong, "An informed
multitask diffusion adaptation approach to study tremor in parkinson's
disease," IEEE Journal of Selected Topics in Signal Processing, vol. 10,
no. 7, pp. 1306 -- 1314, Oct 2016.
[27] C. Rossow, D. Andriesse, T. Werner, B. Stone-Gross, D. Plohmann,
C. J. Dietrich, and H. Bos, "Sok: P2pwned - modeling and evaluating
the resilience of peer-to-peer botnets," in Proc. IEEE Symposium on
Security and Privacy, San Francisco, California, May 2013, pp. 97 -- 111.
[28] A. Moneta, N. Chlass, D. Entner, and P. Hoyer, "Causal search in struc-
tural vector autoregressive models," in Proceedings of the 12th Inter-
national Conference on Neural Information Processing Systems (NIPS)
Mini-Symposium on Causality in Time Series, Vancouver, Canada, 2009,
pp. 95 -- 118.
[29] P.-Y. Lai, "Reconstructing network topology and coupling strengths in
directed networks of discrete-time dynamics," Phys. Rev. E, vol. 95, p.
022311, Feb 2017. [Online]. Available: https://link.aps.org/doi/10.1103/
PhysRevE.95.022311
27
[52] -- -- , "Divide-and-conquer tomography for large-scale networks," in
Proc. IEEE International Data Science Workshop, Lausanne, Switzer-
land, June 2018, pp. 1 -- 5.
[53] A. H. Sayed, "Adaptation, Learning, and Optimization over Networks,"
Found. Trends Mach. Learn., vol. 7, no. 4-5, pp. 311 -- 801, 2014.
[54] P. Erdos and A. R´enyi, "On Random Graphs I," Publicationes Mathe-
maticae (Debrecen), vol. 6, pp. 290 -- 297, 1959.
[55] R. A. Horn and C. R. Johnson, Matrix Analysis, 2nd edition. New
York, NY, USA: Cambridge University Press, 2012.
Augusto Santos received the B.Sc. and M.Sc. from Instituto Supe-
rior T´ecnico (IST), Lisbon, Portugal, in 2007 and 2008, respectively.
He received his Ph.D. from IST and Carnegie Mellon University
(CMU), Pittsburgh, USA, in 2014. All degrees were obtained in
electrical and computer engineering. He held a post-doctoral scholar
position at CMU during 2015-2017 and another at the Adaptive
Systems Laboratory at ´Ecole Polytechnique F´ed´erale de Lausanne
(EPFL), Lausanne, Switzerland, during 2017-2019.
Vincenzo Matta received the Laurea degree (cum laude) in
electronic engineering and the Ph.D. degree in information engi-
neering from the University of Salerno, Fisciano, Italy, in 2001 and
2005, respectively. He is currently an Associate Professor with the
Department of Information and Electrical Engineering and Applied
Mathematics, University of Salerno. His research interests cover the
wide area of statistical signal processing and information theory,
with current emphasis on: Adaptation and learning over networks;
the interplay between inference, communications and security in
distributed systems; multiobject/multisensor tracking and data fusion;
detection of gravitational waves. He has published more than 100
articles on international journals and proceedings of international
conferences. He serves as a Senior Area Editor for the IEEE Signal
Processing Letters, and formerly served as an Associate Editor for
the IEEE Transactions on Aerospace and Electronic Systems, for the
IEEE Signal Processing Letters, and for the IEEE Transactions on
Signal and Information Processing over networks
Ali H. Sayed is Dean of Engineering at EPFL, Switzerland. He
has also served as distinguished professor and former chairman of
electrical engineering at UCLA. He is a member of the US National
Academy of Engineering and recognized as a highly-cited researcher.
An author of over 530 scholarly publications and six books, his
research involves several areas including adaptation and learning,
data and network sciences, and multi-agent systems. Dr. Sayed has
received several awards. He is a Fellow of IEEE, EURASIP, and the
American Association for the Advancement of Science (AAAS). He
is serving as President of the IEEE Signal Processing Society.
[30] B. Pasdeloup, V. Gripon, G. Mercier, D. Pastor, and M. Rabbat, "Charac-
terization and inference of graph diffusion processes from observations
of stationary signals," IEEE Transactions on Signal and Information
Processing over Networks, available as IEEE early access article, 2017.
tomography under partial
observations over adaptive networks," IEEE Transactions on Information
Theory, pp. 1 -- 1, 2018.
[31] V. Matta and A. H. Sayed, "Consistent
[32] -- -- , "Tomography of adaptive multi-agent networks under limited
observation," in Proc. IEEE ICASSP, Calgary, Canada, April 2018, pp.
1 -- 5.
[33] E. S. C. Ching and H. C. Tam, "Reconstructing links in directed
networks from noisy dynamics," Phys. Rev. E, vol. 95, p. 010301, Jan
2017.
[Online]. Available: https://link.aps.org/doi/10.1103/PhysRevE.
95.010301
[34] D. Napoletani and T. D. Sauer, "Reconstructing the topology of sparsely
connected dynamical networks," Physical Review. E, Statistical, Nonlin-
ear, and Soft Matter Physics, vol. 77, p. 026103, 2008.
[35] J. Ren, W.-X. Wang, B. Li, and Y.-C. Lai, "Noise bridges dynamical
correlation and topology in coupled oscillator networks," Phys.
Rev. Lett., vol. 104, p. 058701, Feb 2010.
[Online]. Available:
https://link.aps.org/doi/10.1103/PhysRevLett.104.058701
[36] Y. Yang, T. Luo, Z. Li, X. Zhang, and P. S. Yu, "A robust method for
inferring network structures," in Scientific Reports, vol. 7, no. 5221, July
2017.
[37] A. Mauroy and J. Goncalves, "Linear identification of nonlinear systems:
A lifting technique based on the Koopman operator," in 2016 IEEE 55th
Conference on Decision and Control (CDC), Las Vegas, USA, Dec 2016,
pp. 6500 -- 6505.
[38] P. Geiger, K. Zhang, B. Scholkopf, M. Gong, and D. Janzing, "Causal
inference by identification of vector autoregressive processes with hidden
components," in Proc. International Conference on Machine Learning,
vol. 37, July 2015, pp. 1917 -- 1925.
[39] S. Segarra, M. T. Schaub, and A. Jadbabaie, "Network inference from
consensus dynamics," 2017. [Online]. Available: http://arxiv.org/abs/
1708.05329
[40] J. Mei and J. M. F. Moura, "Signal processing on graphs: causal mod-
eling of unstructured data," IEEE Transactions on Signal Processing,
vol. 65, no. 8, pp. 2077 -- 2092, April 2017.
[41] C. J. Quinn, N. Kiyavash, and T. P. Coleman, "Directed information
graphs," IEEE Transactions on Information Theory, vol. 61, no. 12, pp.
6887 -- 6909, Dec 2015.
[42] J. Etesami and N. Kiyavash, "Measuring causal relationships in dy-
namical systems through recovery of functional dependencies," IEEE
Transactions on Signal and Information Processing over Networks,
vol. 3, no. 4, pp. 650 -- 659, Dec 2017.
[43] D. Materassi and M. V. Salapaka, "On the problem of reconstructing
an unknown topology via locality properties of the Wiener filter," IEEE
Transactions on Automatic Control, vol. 57, no. 7, pp. 1765 -- 1777, July
2012.
[44] A. H. Sayed, Adaptive Filters. Wiley, NJ, 2008.
[45] D. Materassi and M. V. Salapaka, "Network reconstruction of dynam-
ical polytrees with unobserved nodes," in Proc. IEEE Conference on
Decision and Control (CDC), Maui, Hawaii, Dec 2012, pp. 4629 -- 4634.
[46] J. Etesami, N. Kiyavash, and T. Coleman, "Learning minimal latent
directed information polytrees," Neural Computation, vol. 28, no. 9, pp.
1723 -- 1768, August 2016.
[47] D. Materassi and M. V. Salapaka, "Identification of network components
in presence of unobserved nodes," in Proc. IEEE Conference on Decision
and Control (CDC), Osaka, Japan, Dec 2015, pp. 1563 -- 1568.
[48] A. Anandkumar, V. Y. F. Tan, F. Huang, and A. S. Willsky, "High-
dimensional gaussian graphical model selection: Walk summability and
local separation criterion," J. Mach. Learn. Res., vol. 13, no. 1, pp.
2293 -- 2337, Aug. 2012.
[49] A. Anandkumar and R. Valluvan, "Learning loopy graphical models
with latent variables: Efficient methods
and guarantees," Ann.
Statist., vol. 41, no. 2, pp. 401 -- 435, 04 2013. [Online]. Available:
https://doi.org/10.1214/12-AOS1070
[50] V. Chandrasekaran, P. A. Parrilo, and A. S. Willsky, "Latent
variable graphical model selection via convex optimization," Ann.
Statist., vol. 40, no. 4, pp. 1935 -- 1967, 08 2012. [Online]. Available:
https://doi.org/10.1214/11-AOS949
[51] A. Santos, V. Matta, and A. H. Sayed, "Consistent tomography over
diffusion networks under the low-observability regime," in Proc. IEEE
International Symposium on Information Theory, Colorado, USA, June
2018, pp. 1 -- 5.
|
1903.03189 | 1 | 1903 | 2019-03-07T21:32:27 | Incorporating social practices in BDI agent systems | [
"cs.MA",
"cs.CY"
] | When agents interact with humans, either through embodied agents or because they are embedded in a robot, it would be easy if they could use fixed interaction protocols as they do with other agents. However, people do not keep fixed protocols in their day-to-day interactions and the environments are often dynamic, making it impossible to use fixed protocols. Deliberating about interactions from fundamentals is not very scalable either, because in that case all possible reactions of a user have to be considered in the plans. In this paper we argue that social practices can be used as an inspiration for designing flexible and scalable interaction mechanisms that are also robust. However, using social practices requires extending the traditional BDI deliberation cycle to monitor landmark states and perform expected actions by leveraging existing plans. We define and implement this mechanism in Jason using a periodically run meta-deliberation plan, supported by a metainterpreter, and illustrate its use in a realistic scenario. | cs.MA | cs | Incorporating social practices in BDI agent systems
Stephen Cranefield
University of Otago
Dunedin, New Zealand
[email protected]
Frank Dignum
Umeå University
Umeå, Sweden
[email protected]
9
1
0
2
r
a
M
7
]
A
M
.
s
c
[
1
v
9
8
1
3
0
.
3
0
9
1
:
v
i
X
r
a
ABSTRACT
When agents interact with humans, either through embodied agents
or because they are embedded in a robot, it would be easy if they
could use fixed interaction protocols as they do with other agents.
However, people do not keep fixed protocols in their day-to-day
interactions and the environments are often dynamic, making it im-
possible to use fixed protocols. Deliberating about interactions from
fundamentals is not very scalable either, because in that case all pos-
sible reactions of a user have to be considered in the plans. In this
paper we argue that social practices can be used as an inspiration
for designing flexible and scalable interaction mechanisms that are
also robust. However, using social practices requires extending the
traditional BDI deliberation cycle to monitor landmark states and
perform expected actions by leveraging existing plans. We define
and implement this mechanism in Jason using a periodically run
meta-deliberation plan, supported by a metainterpreter, and illus-
trate its use in a realistic scenario.
KEYWORDS
Social practices; BDI agents; Jason
1 INTRODUCTION
Imagine the scenario where a disabled person, living alone, is as-
sisted by a care robot. The robot takes care that the person gets up
every morning and makes sure that he drinks some coffee and takes
his morning pills (if needed). Then they read the newspaper, which
means that the person looks at the pictures in the paper and the ro-
bot reads the articles out loud for the person to hear. (The hearing
of the person is better than his eyesight, so, he cannot read the small
font of the newspaper well, but can hear the robot).
When agents in the role of this type of personal assistant or care
robot have to interact with humans over a longer time period and
in a dynamic environment (that is not controlled by the agent), the
interaction management becomes very difficult. When fixed proto-
cols are used for the interaction they are often not appropriate in
all situations and cause breakdowns and consequent loss of trust
in the system. However, to have real-time deliberation about the
best response during the interaction is not very scalable, because in
real life the contexts are dynamic and complex and thus the agent
would need to take many parameters into consideration at each step.
Thus we need something in between a completely scripted interac-
tion that is too brittle and a completely open interaction that is not
scalable.
As we have done before in the agent community, we take inspira-
tion from human interactions and the way they are managed by indi-
viduals. We classify situations into standard contexts in which a cer-
tain social practice can be applied. Social science has studied this
phenomenon in social practice theory. Social practice theory comes
forth from a variety of different sub-disciplines of social science. It
started from philosophical sociology with proponents like Bourdieu
[3] and Giddens [8]. Later on Reckwitz and Shove [16, 18] have ex-
panded on these ideas, and also Schatzki [17] made some valuable
contributions.
These authors all claim that important features of human life
should be understood in terms of organized constellations of inter-
acting persons. Thus people are not just creating these practices, but
our deliberations are also based on the fact that most of our life is
shaped by social practices. Thus we use social practices to catego-
rize situations and decide upon ways of behaviour based on social
practices. The main intuition behind this is that our life is quite
cyclic, in that many activities come back with a certain regularity.
We have meals every day, go to work on Monday until Friday, go to
the supermarket once a week, etc. These so-called Patterns of Life
([7]) can be exploited to create standard situations and expectations.
It makes sense to categorize recurrent situations as social practices
with a kind of standard behaviour for each of them.
Unfortunately social practice theory has not been widely used in
computer science or in HCI and thus there are no ready-to-use tools
in order to incorporate them in agents. It is clear from the above
description that social practices are more than just a protocol or a
frame to be used by the agent in its deliberation. Therefore, in this
paper we make the following contributions. We propose a mecha-
nism for BDI agents to maintain awareness about active social prac-
tices, and to leverage their existing plans to act in accordance with
these practices. This is presented as a meta-deliberation plan that
can be directly executed by Jason agents, or treated as a specifica-
tion for an optimised implementation in an extended agent platform.
This plan has been deployed in the (simulated) care robot scenario,
to confirm that awareness of and adherence to a social practice en-
ables the robot to have a more successful interaction with the patient
over a longer period of time. As some of the features needed to im-
plement this scenario, and to support our meta-deliberation plan,
are not currently available in Jason, we also present a Jason metain-
terpreter, which provides this extended functionality, but can also
be used independently to support other research on extensions to
practical reasoning in the BDI agent paradigm.
In the next section, we give an introduction to the purpose and
structure of social practices. In Section 3, we elaborate on the care
robot scenario and how we have modelled it in Jason. In Section 4,
we describe the role of social practices in this scenario, and discuss
the requirements this imposes for a BDI agent. In Section 5, we
present our mechanism for extending Jason to leverage social prac-
tices, and the metainterpreter needed to support this. We finish the
paper with some conclusions and suggestions for future work.
2 SOCIAL PRACTICES
Social practices are defined as accepted ways of doing things, con-
textual and materially mediated, that are shared between actors and
routinized over time [16]. They can be seen as patterns which can be
filled in by a multitude of single and often unique actions. Through
(joint) performance, the patterns provided by the practice are filled
out and reproduced.
According to [16, 18] a social practice consists of three parts:
• Material: covers all physical aspects of the performance of
a practice, including the human body and objects that are
available (relates to physical aspects of a context).
• Meaning: refers to the issues which are considered to be rele-
vant with respect to that material, i.e. understandings, beliefs
and emotions (relates to social aspects of a situation).
• Competence: refers to skills and knowledge which are re-
quired to perform the practice (relates to the notion of de-
liberation about a situation).
Whereas the first and third parts intuitively can be made more
precise for an implementation, the second part is rather vague. Let
us consider these three parts of a social practice in the scenario of
the care robot scenario introduced in Section 1. The material refers
to the room where the robot serves morning coffee for the disabled
person. It includes the materials that are needed to make coffee
(such as coffee and a coffee maker) and serve it (such as a cup
and tray). However, it also includes the table and other furniture
in the room, the newspaper (if present), the TV, radio, computer,
tablet, and the robot and person (and possible other people that are
present).
The competence part describes what activities every party can
perform and expectations about what they will actually do. For ex-
ample, the robot is capable of making coffee and serving it. The per-
son can drink his coffee by himself. They both can read the newspa-
per or watch TV. The expectation is that the robot wakes the person
if he is not awake yet, makes the coffee and gives it to the person.
After that they will read the newspaper together to provide mental
stimulation. Note, these are expectations, not a protocol. So, parties
can deviate from it and they can also fill the parts in, in ways they
see fit best. The meaning part has to do with all the social interpre-
tations that come with the social practice, e.g. drinking coffee in the
morning might give the person a sense of well-being that he can use
to face the challenges of the rest of the day. When the coffee is cold
or weak the person might interpret it as disinterest on the part of the
robot in his well-being. The goal of reading the newspaper might
also be not just to get the information from it, but a form of enter-
tainment and feeling related to the robot, because you do something
together.
Given the above descriptions of social practices one can summa-
rize the purpose of them for individuals (informally) as follows:
Perceive(resources) ⇒ Expect (activities) ∧ Expect (competencies)
Done(activities) ⇒ physical (postconditions)∧social(postconditions)
Given a certain situation that is perceived in which the social prac-
tice is activated, the social practice now triggers expectations that
the activities of that social practice will be executed. This implies
that we assume a certain competence of all the people involved in
the social practice (e.g. the robot can make coffee). After the activ-
ities have been executed, we do not just assume the postconditions
of the actions hold, but also assume that certain social effects have
been achieved according to the "meaning" of the practice. Thus
the social practice allows us to make a whole set of assumptions
and have expectations that could otherwise not be readily made or
would take a lot of effort to derive.
From the above description it can already be seen that social
practices are more encompassing than conventions and norms. Con-
ventions focus on the strategic advantage that an individual gets by
conforming to the convention. The reason to follow a convention is
that if all parties involved comply, a kind of optimal coordination is
reached, i.e. if we all drive on the left side of the road, traffic will
be smoother than when everyone chooses the side to drive on freely.
Thus, conventions focus on the actual actions being performed and
how they optimize the coordination. Social practices do not neces-
sarily optimize the coordination. Because they indicate expected ac-
tions and interactions given a social and physical context they will
smoothen the coordination. However, this is not necessarily the op-
timal way the coordination could have been done. For example, if
we go to a presentation, we sit down as soon as we see chairs stand-
ing in rows in the room. However, we could also keep standing (as
is often done outside).
Social practices are also different from norms. Norms are also ap-
plicable in certain situations and for particular people (or roles) and
they also create expectations (namely that the norm is followed).
However, norms usually dictate a very specific behaviour rather
than creating a set of loosely couple expectations as is the case
for social practices. E.g. if the norm states that a car has to stop
for a red light, it gives a very specific directive. If a norm is more
abstract (like "drive carefully") then we need to translate this into
concrete norms for specific situations. This is different from saying
that some parts of a situation are governed by the norm and others
are still free of the normative influence. Basically, when specifying
the norm one indicates exactly when the norm is applicable rather
than general situations for which the norm can be applied in some
part.
One framework that seems very close to social practices is the
notion of scripts. However, social practices are not just mere scripts
in the sense of Minsky [14]. Practices are more flexible than the
classical frames defined by scripts in that they can be extended and
changed by learning, and the "slots" only need to be filled in as
far as they are needed to determine a course of action. Using these
structures changes planning in many common situations to pattern
recognition and filling in parameters. They support, rather than re-
strict, deliberation about behaviour. For example, the social practice
of "going to work" incorporates usual means of transport that can
be used, timing constraints, weather and traffic conditions, etc. So,
normally you take a car to work, but if the weather is exceptionally
bad, the social practice does not force the default action, but rather
gives input for deliberation about a new plan in this situation, such
as taking a bus or train (or even staying home). So, social practices
can be seen as a kind of flexible script. Moreover, scripts do not
incorporate any social meaning for the activities performed in them
as social practices do.
Social practices have been used in applications already in a vari-
ety of ways. In [12, 15] they have been used as part of social simu-
lations. In those applications, social practices are used as a standard
package of actions with a special status. Thus individuals can use
them with a certain probability given the circumstances are right.
However, these applications do not use the internal structure of so-
cial practices for the planning of the individuals. Social practices
have been used for applications in natural language and dialogue
management in [1, 9]. Here, the social practices are used to guide
the planning process, but are geared towards a particular dialogue
rather than as part of a more general interaction. In [13] it is shown
how social practices can be used by a traditional epistemic multi-
agent planner to provide efficient and robust plans in cooperative
settings. However, in this case the planner was not part of a BDI
agent with its own goals and plans, but completely dedicated to
finding a plan for the situation at hand. In [6] a first structure of
social practices was presented that is more amenable for the use by
agents. The paper is, unfortunately, only conceptual and no imple-
mentation was made yet. In this paper we will follow the structure
described in [6], but use only those parts that are relevant for our
scenario. A complete account would be far too large to fit in the
space available here, but we take those parts that seem to be core
to the idea of social practices and show how they work with BDI
agents in the Jason platform.
The complete structure for social practices (based on [6]) is as
follows:
Context
• Roles describe the competencies and expectations about a
certain type of actor. Thus the robot is expected to be able to
make a cup of coffee.
• Actors are all people and autonomous systems involved, that
have capability to reason and (inter)act. This indicates the
agents that are expected to fulfil a part in the practice. In our
scenario, these are the robot and the person.
• Resources are objects that are used by the actions in the prac-
tice, such as cups, coffee, trays, curtains, and chairs. So, they
are assumed to be available both for standard actions and for
the planning within the practice.
• Affordances are the properties of the context that permit so-
cial actions and depend on the match between context con-
ditions and actor characteristics. For example, the bed might
be used as a chair, or a mug as a cup.
• Places indicates where all objects and actors are usually lo-
cated relatively to each other, in space or time: the cups are
in the cupboard in the kitchen, the person is in the chair (or
in bed), etc.
Meaning
• Purpose determines the social interpretation of actions and
of certain physical situations. For example, the purpose of
reading the newspaper is to get information about current
affairs and to entertain the person.
• Promotes indicates the values that are promoted (or demoted,
by promoting the opposite) by the social practice. Giving cof-
fee to the person will promote the value of "caring".
• Counts-as are rules of the type "X counts as Y in C" link-
ing brute facts (X) and institutional facts (Y) in the context
(C). E.g., reading the newspaper with the person counts as
entertaining the person.
Expectations
• Plan patterns describe usual patterns of actions defined by
the landmarks that are expected to occur (states of affairs
around which the inter-agent coordination is structured). For
example, the care robot first checks if the person is awake
then makes sure there is coffee served.
• Norms describe the rules of (expected) behaviour within the
practice. E.g., the robot should ask the person if he wants
coffee, before starting to make it.
• Strategies indicate condition-action pairs that can occur at
any time during the practice. E.g. if the person drops the cof-
fee, the robot will clean it up. If the robot notices the person
is asleep (again) it will try to wake him.
• A Start condition, or trigger, indicates how the social prac-
tice starts, e.g., The practice of having morning coffee starts
at 8 am.
• A Duration, or End condition, indicates how the social prac-
tice ends, e.g., the morning routine takes around 45 minutes
and ends when the newspaper is read and the coffee is fin-
ished.
Activities
• Possible actions describes the expected actions by actors in
the social practice, e.g. making coffee, reading the newspa-
per, and opening curtains.
• Requirements indicate the type of capabilities or competences
that the agent is expected to have in order to perform the
activities within this practice. For example, the robot is ex-
pected to know how to make coffee and read the newspaper.
In [5] there is a first formalization of all these aspects based on dy-
namic logic. Due to space limitations we will not include this whole
formalization here, but just discuss a few points that are important
for the current implementation of social practices in Jason.
The core element of the social practice for an agent is the plan pat-
tern, which gives it handles to plan its behaviour. Plan patterns are
defined as follows:
Definition 2.1. Plan Patterns Language
A Plan Pattern of a social practice is an element of the set PP, which
is the smallest set closed under
γϕ ∈ PP
γ1ϕ1 , γ2ϕ2 ∈ PP ⇒ γ1ϕ1 + γ2ϕ2 ∈ PP
γ1ϕ1 , γ2ϕ2 ∈ PP ⇒ γ1ϕ1&γ2ϕ2 ∈ PP
γ1ϕ1 , γ2ϕ2 ∈ PP ⇒ γ1ϕ1; γ2ϕ2 ∈ PP
Here γϕ stands for any sequence of actions γ that contain actions
contributing towards the achievement of ϕ (starting from a particu-
lar situation). ϕ is the purpose of that part of the practice. There can
be more effects, but they are not all specified. So, in our morning
routine practice the plan pattern can be defined as:
γ1ϕ1; (γ2ϕ2&γ3ϕ3); γ4ϕ4
ϕ1 is the person is awake
ϕ2 is the coffee is served and ϕ3 is the pills are taken
ϕ4 is the person is mentally stimulated.
So, the purpose of the first part of the morning routine is that the per-
son is awake. This might be done by opening the curtains, giving
some loud noise or otherwise. If the purpose is achieved by opening
the curtains, not only is the person awake, but the curtains are also
open. The latter is merely a side effect of achieving the purpose in
this way.
Two more things should be noted about these patterns. One is
that the overall pattern is supposed to achieve the overall purpose of
the social practice. This is a formal constraint, but we only treat this
implicitly. The other is that after a part of the plan pattern is finished,
it automatically triggers the start of the next part of the pattern. In
the full formalism this is assured, but is not explicit from only this
fragment. In the same way, a social practice is started when the start
condition becomes true. It then becomes available for execution and
can be used by any agent present in the situation.
Finally, the formalism of social practices also guarantees that
there is a common belief in the elements of the social practice and
if actions are taken everyone has the at least a common belief about
the effects in as far as they are important for the social practice.
Thus it guarantees a common situation awareness.
3 THE CARE ROBOT SCENARIO
In this section we elaborate on the care robot scenario outlined in
the introduction, and describe how we have modelled and imple-
mented it using Jason.
We assume the high-level operation of the robot is based on a
BDI interpreter, and that it comes equipped with goals and plans
to trigger and enact its care activities (most likely with some cus-
tomisation of key parameters possible). In this section we consider
only a small subset of the robot's duties: to wake the patient at a
certain time in the morning, to provide coffee as required, and to
provide mental stimulation. We do not specify any goals of the ro-
bot outside the practice here, but normally the care robot would also
have its own goals such as powering its battery, (vacuum) cleaning
a room and taking care of the health of the patient. The morning
routine can be seen as a part of the plan to take care of the health
of the patient, while there might be no social practice for vacuum
cleaning and this is completely handled by the standard BDI part of
the care robot.
Social practices provide patterns of coordination for multiple
agents in terms of landmark states rather than explicit sequences
of actions. Therefore they do not make limiting assumptions about
the temporal aspects of actions and their effects leading up to a
landmark. Only the landmarks themselves are explicitly temporally
ordered. To illustrate this we include some temporal complexity in
the scenario by including durative actions (i.e. those that take place
across an interval of time), an action with a delayed effect, and a
joint durative action, which has its desired effect only if two par-
ticipants perform it during overlapping time intervals. Durative and
joint actions are implemented using a Jason metainterpreter1 that is
described in Section 5. To simulate the passing of time, we use a
"ticker" agent with a recursive plan that periodically performs a tick
action to update the time recorded in the environment. We use Ja-
son's synchronous execution mode, so the robot, patient and ticker
agents perform a single reasoning step in every step of the simula-
tion.
Figure 1 shows the robot's initial beliefs, rules and plans. It has
four sets of plans (lines 25 onwards). These have declarative goals
(i.e. their triggering goals express desired states) and use Jason pre-
processing directives to transform them according to a predefined
declarative achievement goal pattern [2].
continueMakingPodCoffee).
1 // Initial beliefs and rules
2
3 durative(makePodCoffee).
4 durative(readNewspaper).
5 joint(readNewspaper).
6 durative_action_continuation_pred
7
8 durative_action_continuation_pred(makePodCoffee,
9
10 continueReadingNewspaper :-
11
12
(readNewspaper, continueReadingNewspaper).
started(readNewspaper, T1) &
not started_durative_action(readNewspaper, patient, _) &
time(T2) &
T2 <= T1 + 20.
started_durative_action(readNewspaper, patient, _) &
not stopped_durative_action(readNewspaper, patient, _).
13
14
15 continueReadingNewspaper :-
16
17
18
19 continueMakingPodCoffee :- state(coffee, not_made).
20
21 wake_up_phrase("Good morning sleepyhead!").
22
23 // Plans
24
25 {begin ebdg(state(patient,awake))}
26 +!state(patient,awake) : wake_up_phrase(P) <- talkToPatient(P).
27 +!state(patient,awake) <- shakePatient.
28 +!state(patient,awake) <-
29
openCurtains;
.wait(state(patient, awake), 30000).
"", readNewspaper[participants([patient,robot])])]).
30
31 {end}
32
33 {begin ebdg(state(patient,mentally_stimulated))}
34 +!state(patient, mentally_stimulated) <-
.wait(state(patient, awake)); .fail.
35
36 +!state(patient, mentally_stimulated) <- play_mozart.
37 +!state(patient, mentally_stimulated) <-
38
!solve([body_term(
!state(coffee, made);
serveCoffee.
39
40 {end}
41
42 +!state(coffee, served) <-
43
44
45
46 {begin ebdg(state(coffee,made))}
47 +!state(coffee, made) :
48
49
50
51 +!state(coffee, made) : resource(instant_coffee) <-
52
53 {end}
54
55 { include("metainterpreter.asl") }
makePodCoffee;
.wait(state(coffee, made), 10000).
makeInstantCoffee.
resource(coffee_pods) & resource(coffee_pod_machine) <-
Figure 1: Plans for the care robot domain
The first set of plans (lines 25 -- 31) are for achieving a state where
the patient is awake, with alternative plans for talking to the patient,
shaking him, and opening the curtains and waiting for the light to
wake him. The exclusive backtracking declarative goal ("edbg")
pattern specifies that additional failure-handling logic should be
added to ensure that all the plans will be tried (once each) until the
goal is achieved, or all plans fail. The action of opening the curtains
has a delayed effect: it will eventually wake the patient2.
The second set of plans (lines 33 -- 40) are used for the goal of
having the patient mentally simulated, and also use the ebdg pattern.
The first plan waits for the patient to be awake, and then fails so
that the other plans will be tried. The other two alternatives involve
playing the music of Mozart to the patient, and initiating the joint
action of reading the newspaper with the patient. As joint actions
are not directly supported by Jason, lines 38 -- 39 call this action via
the solve goal that is handled by our metainterpreter.
1A metainterpreter is a programming language interpreter written in the same, or a
similar, language. It can be used to prototype extensions to the base language.
2Actions are implemented in Jason by defining an execute method in Java class mod-
elling the environment. The delay is currently hard-coded in this class.
These plans are followed by a single plan for serving coffee. This
has the subgoal of having the coffee made, and then the action of
serving the coffee is performed.
The final set of plans are for reaching a state in which the coffee
is made. The options are to use a coffee pod and wait for it to finish
(up to a time limit)3, or to make instant coffee.
The initial segment of the listing contains initial beliefs and rules
related to the processing of durative actions: declarations of which
actions are declarative and/or joint, and predicates and associated
rules defining the circumstances in which the robot will continue
performing the durative actions.
The environment sends a percept to all participants of a joint ac-
tion when any other participant performs the action for the first time
or performs a stop action with the joint action as an argument. The
patient agent has a plan to take his pills once he is awake. He also
has a plan that will respond to the robot beginning the joint newspa-
per reading action by also beginning that action. He will continue
reading the newspaper for 40 time units if he is in a good mood,
but only 20 if it is in a bad mood. Being woken by daylight (after
the curtains are opened) leaves him in a good mood; being shaken
awake leaves him in a bad mood, and talking will not wake him up.
Thus, if the robot begins with goals to have the patient awake and
mentally stimulated, the patient will be left in a bad mood by being
shaken awake and the newspaper reading will be shorter (and less
stimulating) that if he were in a good mood.
4 A CARE ROBOT WITH SOCIAL
PRACTICES
Section 3 introduced the care robot scenario. In this section, we
consider how the robot could be enhanced using social practices.
As noted previously, it is assumed that the robot comes equipped
with appropriate goals and plans, and that it is possible to customise
certain parameters such as the time the user likes to wake up, and
the time and style of coffee that he likes to have. However, customis-
ing each plan in isolation will not easily provide the coordination
between activities and dynamic adaptability to different contexts
that can be provided by social practices. To perform most effec-
tively, the robot should choose, for a given context, the plans for
each goal that will achieve the best outcomes for the patient, and
furthermore, consider constraints on goal orderings that arise from
preferences and habit. For example, if the patient prefers to be wo-
ken at a certain time in a given context (e.g. when his family is due
to visit) and/or in a certain way (e.g. by the curtains being opened),
his mood is likely to be adversely affected if he is woken at a dif-
ferent time, and his engagement with subsequent activities (such as
reading the newspaper together) may be reduced. In this section we
describe how this type of contextual information can be addressed
by the use of a social practice.
In Section 3, we described the various plans and actions available
to the robot. We now assume that the following "morning routine"
social practice has emerged4. We present this as a set of beliefs in
the form used by our social practice reasoning plans that will be
discussed in Section 5. Note that we only illustrate a small subset
3The wait timeouts (in ms) in these plans are for simulation purposes only, and would
be much longer in a real-world application.
4It is beyond the scope of this paper to consider how social practices might be learned
and/or communicated.
of what would be likely to be a real morning routine for a patient
and his/her care robot, but this is sufficient to highlight the nature
of social practices and their relation to BDI agents.
social_practice(morningRoutine,
[state(location, home), resource(coffee_pods),
resource(coffee_pod_machine), resource(pills),
resource(newspaper_subscription),
(time(T) & T < 1200)]).
landmark(morningRoutine, pa,
[], [action(robot, openCurtains)],
state(patient, awake)).
landmark(morningRoutine, pt,
[pa], [action(patient, takePills)],
state(pills, taken)).
landmark(morningRoutine, cs,
[pa], [action(robot, makePodCoffee)],
state(coffee, served)).
landmark(morningRoutine, ms,
[pt,cs], [action([robot,patient], readNewspaper)],
state(patient, mentally_stimulated)).
The first belief above encodes the name of the social practice and
a list of conditions that must all hold for it to become active: there
are constraints on the location, the resources available, and the time
(here, the number 1200 is a proxy for some real-world time that
ends the morning routine period).
The other four beliefs model the landmarks, specifying the social
practice they are part of, an identifier for the landmark, a list of
landmarks that must have been reached previously, a list of actions
and their actors that are associated with the landmark, and finally,
a goal that is the purpose of the landmark. The landmarks are: (1)
to have the patient awake due to the robot opening the curtains, (2)
for the patient to have taken his pills, (3) to have the coffee served,
which should involve the robot making pod coffee, and (4) for the
patient to be mentally stimulated due to the newspaper being read
jointly. These landmarks are partially ordered with 1 before 2 and
3, which both precede 4.
Comparing this social practice to the robot plans shown in Fig-
ure 1, it can be seen that it avoids an ineffective attempt to wake
the patient by talking to him, and prevents him from being left in a
bad mood after being shaken awake. It agrees with the first-ordered
plan for making coffee (by making pod coffee), and avoids an ill-
fated attempt by the robot to provide mental stimulation by playing
Mozart. Furthermore, it specifies an ordering on these activities that
is not intrinsic to the plans themselves. Note also, that the social
practice does not provide complete information on how to reach the
landmark of having coffee served: it indicates that the robot should
make pod coffee, but doesn't specify the action of serving the cof-
fee. While a planning system could deduce the missing action using
a model of actions and their effects [13], a BDI agent does not have
this capability. Instead, a BDI agent using social practices must rea-
son about how its existing plans could be used to satisfy landmarks
given potentially incomplete information about the actions it must
perform.
Furthermore, the robot may already have goals to wake the pa-
tient, provide mental stimulation, etc., and the activation of a social
practice should not create additional instances of those goals. Thus,
the activation of a social practice should override the agent's normal
behaviour (for the relevant goals) during the period of activation.
As social practices are structured in terms of ordered landmarks,
which model expected states to be reached in a pattern of inter-agent
coordination, it is necessary for the agent to actively monitor the
status of landmarks once their prior landmarks have been achieved,
and to actively work towards the fulfilment of the current landmarks
for which it has associated actions. In the next section, we present
a meta-deliberation cycle for Jason agents that addresses this and
the other issues outlined above, and which enables the successful
execution of our care robot enhanced with social practices.
5 IMPLEMENTATION
5.1 Meta-level reasoning about social practices
Maintaining awareness of social practices (SPs), and contribut-
ing to them in an appropriate way, requires agents to detect when
each known social practice becomes active or inactive, to monitor
the state of the landmarks in an active social practice, and to trigger
the appropriate activity if an active SP has an action for the agent as-
sociated with the next landmark. This is a type of meta-level reason-
ing that the agent should perform periodically, and it may override
the performance of any standard BDI processing of goals, which is
not informed by social practices. We note that, on an abstract level,
the same was done in [1] where the plan pattern was translated into
a global pattern in Drools (Java based expert system) and the spe-
cific interactions within each phase were programmed in a chatbot.
The question then arises of how best to implement such a meta-
level reasoner in a BDI architecture. While the best performance
can, no doubt, be achieved by extending a BDI platform using its
underlying implementation language, this approach requires signif-
icant knowledge of the implementation and requires using an imper-
ative coding style that is not best suited to reasoning about goals [10]
and for rapid prototyping and dissemination of new reasoning tech-
niques. Therefore, in this work we define the meta-level reasoner
as a plan for a metadeliberate goal that reasons about social
practices, sleeps and then calls itself recursively. This, and some
other plans it triggers, are shown in Figure 2. The plans make use
of some extensions to Jason, handled by a metainterpreter that is
described in the following subsection5.
First, we give a brief overview of the syntax of the AgentSpeak
syntax, as implemented (and extended) in Jason. Based on logic-
programming languages like Prolog, the basic language constructs
are atoms (beginning with lower case letters), variables (beginning
with upper case letters), and structured terms with a functor and
terms as arguments. Jason also allows terms to have annotations:
lists of terms in square brackets, and these are treated specially
during unification6. Agent programs consist of initial beliefs, rules
(Horn clauses) and plans. Plans are expressed using the syntax @Label
Trigger : ContextCond <- PlanBody, where Label, ContextCond
and PlanBody are optional. Triggers are events such as the creation
of a new goal (+!GoalTerm) or a new belief (+BeliefTerm). Con-
textCond is a logical formula stating when the plan is applicable,
and is evaluated using beliefs and rules. A plan body contains a se-
quence of actions (terms with no prefix), built-in internal actions
(terms whose functor contains a "."), belief additions and deletions
(with prefix '+' and '-', respectively), queries over beliefs and rules
(with prefix '?'), and subgoals (with prefix '!' or '!!' -- the latter cre-
ates a separate intention for the subgoal). Some control structures
such as if-then-else are also supported within a plan body.
The social practice reasoner runs in response to the goal meta-
Deliberate (line 10 in Figure 2). Lines 13 to 31 show the plan
for this goal. The atomic annotation on the plan label ensures that
steps of this plan are not interleaved with steps of other plans. The
plan begins by (re)considering which social practice (if any) should
be active. It uses the rules in lines 3 to 7 to find social practices
that are relevant (i.e. all their requirements hold), and to select one
(currently, the first option is always selected). If none are relevant
(lines 17 -- 20), any existing belief about the currently selected social
practice is retracted. Otherwise (lines 22 -- 28), if the selection has
changed, the belief about the selection is updated. Any monitored
landmarks are then checked to see if their purpose has been fulfilled
(lines 26 -- 28). If so, a belief about their completion is added. The
plan then sleeps for period, before triggering itself to be re-run in
a new intention (lines 30 -- 31). The new intention is needed for the
recursive call as the plan is atomic, and the agent's other plans must
be allowed to run).
A new belief about a selected social practice is handled by the
plan in lines 33 -- 46. This loops through the landmarks to check if
the agent already has intentions to achieve any of their purposes7. If
so, these intentions are suspended, and this is recorded in a belief so
the intentions can be later marked as successful if the landmark is
completed (see line 73). A plan is also temporarily added (lines 40-
42) to ensure that if some other active plan of the agent separately
creates this intention, it will be immediately suspended (the new
plan is placed before any existing plans for that goal). For each
landmark in the social practice that has no prior landmarks, a goal
is created to activate it (lines 44 -- 46).
Landmark activations are handled by the plan in lines 48 -- 69. A
belief recording that the landmark's purpose should be monitored
is added, then the action associated with the landmark is processed
(only a single action is supported currently). If the action is to be
performed by the agent, three options are considered. First (line 54),
a query is made to find a solution for achieving the landmark's pur-
pose that involves performing the specified action. A set of rules
(not shown) handle this query by searching for the action recur-
sively (up to a prespecified depth bound) through the plans that
achieve the purpose, and the subgoals in those plans, and so on.
The plans' context conditions are checked for the top level plans
(those for the landmark's purpose), but the recursive calls do not,
as, in general, it cannot be known how the state of the world will
change as these plans are executed. If such a solution is found, it is
recorded as a goal-plan tree "path" (see Section 5.2) and passed to
a call to our Jason metainterpreter (line 56). If no such solution is
found, and the action is a joint, durative or internal one, the metain-
terpreter is called to handle this (lines 58 -- 65. Otherwise, the action
is performed directly (line 67).
Finally, the plan in lines 71 -- 87 handles completed landmarks --
those for which the purpose has been achieved. Any suspended in-
tentions for the purpose are succeeded, the belief stating that the
landmark should be monitored is retracted, and the temporary plan
5See https://github.com/scranefield/jason-social-practices for source code.
6https://github.com/jason-lang/jason/blob/master/doc/tech/annotations.adoc
7The unifications in lines 35 and 50 instantiate the variable on the left with the value
of the variable on the right, but with any annotations removed.
added in lines 40-42 is removed. The plan then checks for subse-
quent landmarks that should now be activated (if all their prior land-
marks are completed), and finally adds a belief that the social prac-
tice has completed if all its landmarks are completed. Another plan
(not shown) handles social practices that become inactive when
their relevance conditions cease to hold. In this case, any active land-
marks should be abandoned, and original intentions to achieve their
purposes can be resumed.
With these plans and the metadeliberation goal in place, our ro-
bot and patient agent can successfully coordinate their actions across
the landmarks of the social practice, ensuring that the patient re-
mains in a good mood, and engages in the newspaper reading for
longer.
5.2 A Jason metainterpreter
Figure 3 shows our Jason metainterpreter, which extends the
AgentSpeak metainterpreter defined by Winikoff [19], and specialises
it for use with Jason. The metainterpreter is initiated by calling a
solve goal with a list of plan body terms, i.e. terms represent-
ing the various types of goals and actions that can appear in a plan
body. In each plan_body term, the Prefix argument identifies
the type of the goal or action by a string (e.g. '?' for a query to the
belief base, '+' for a belief addition, and '!' for a subgoal). From
line 16 onwards, each solve trigger event has additional argu-
ments that: (a) identify the current intention as a stack of current
subgoal indices within each active plan body, interleaved with the
labels for the plans currently active to solve those subgoals, and
(b) a final Path argument, explained below. The intention identi-
fier is used in lines 64 to 71, which sequentially try the plans for
a goal, asserting beliefs about the plans that have been tried. Lines
70 and 71 leverage Jason's failure-handling mechanisms (posting
achievement goal deletion events upon goal failure) to detect that
an attempt to "solve" a plan failed, and to try the next plan. Finally,
note that there are two work-arounds for current restrictions of Ja-
son. First, as Jason does not provide a way to decompose a plan
body from within plans, line 30 calls a custom internal action we
have implemented in Java. Given a trigger event (e.g. a new goal
event), this action returns a list of relevant plans, encoded as list
of plan terms, each including a list of plan_body terms. Second,
internal (in-built) actions cannot be called dynamically via instanti-
ated higher order variables (as used for other actions: see line 56).
Therefore, lines 57 to 60 enumerate specific internal actions that are
supported (and more can be added).
We made the following extensions to support new capabilities;
(1) Durative actions, as required by our scenario, are supported
(lines 38 -- 55 and 73 -- 86)8. A continuation predicate, and op-
tionally a clean-up goal9, for the action are looked up (lines
40 -- 45), the time the action was started is recorded as a belief
(line 53), and a solve_durative goal is created (line 55)
to trigger the performance of the action. The plan for this
goal (lines 73 -- 84) checks the continuation condition (passed
as variable Query). It is intended that the query is a 0-arity
predicate defined by a rule in the agent's program. If the
8The first context condition on line 39 instantiates variable Act to the action term, with
any Jason annotations removed.
9Cleanup goals can be used to remove any temporary state recorded as beliefs during
a durative action's execution, but are not important for the discussion in this paper.
query succeeds, the action is executed with a "durative" an-
notation (which the environment should check for), and pos-
sibly an annotation listing the action participants if it is a
joint action (see below). The goal is then called recursively.
If the query fails, stop(Act) is executed (again, with the
appropriate annotations). Thus, durative actions are imple-
mented by repeated execution of an action until the corre-
sponding stop action is called.
(2) Joint actions are also supported. These are durative actions
with an annotation listing the intended action participants.
The environment should notify all intended participants (via
a percept) when a durative action is called for the first time or
is stopped, thus enabling the participants to coordinate their
actions. It should also keep a history of the time intervals
over which the participants perform the action, as its out-
come will depend on the existence and length of a period
of overlap.
(3) As explained in Section 5.1, when a landmark in a social
practice includes an action associated with the current agent,
the plan to activate a landmark attempts to find an existing
plan that can achieve the landmark's purpose while also in-
cluding the specified action. This is a recursive search through
plans and their subgoals, and it results in a pre-selected path
through the goal-plan tree [11] corresponding to the search
space for satisfying the landmark's purpose. This path can be
passed to the metainterpreter (line 12), to guide it directly to
the pre-chosen subplans, and eventually the desired action.
This feature is useful for plan pre-selection in other meta-
reasoning contexts as well, e.g. choosing plans based on their
effect on the values of a human user [4].
6 CONCLUSIONS
We have argued that for interactive settings, as sketched in our sce-
nario, the use of social practices is a good compromise between us-
ing a fixed interaction protocol and deliberation and planning from
scratch at each point during the interaction. We proposed a mech-
anism for a BDI agent to maintain awareness about and contribute
towards the completion of social practices, and presented this as a
meta-deliberation plan for Jason agents. To extend Jason with fea-
tures required for this plan and our care robot scenario, we also pre-
sented a Jason metainterpreter. These contributions can serve as a
specification of potential extensions to the BDI reasoning cycle, but
also allow the approach to be directly applied within Jason agents.
Our approach allows BDI agents to use their existing plans to
achieve social practice landmarks that do not detail all actions re-
quired to achieve the landmark. However, there are some subtleties
that remain to be explored. For example, suppose that an agent has
a plan for a goal that is the purpose of a landmark, but that one of
that plan's subgoals is the purpose of a prior landmark. In that case,
the execution of the plan should be adapted to exclude this subgoal.
In future work we intend to investigate more complex cases such
as this. We also intend to develop elaborate scenarios that use all
aspects of a social practice, and compare these with agent imple-
mentations where no social practice is used, both in terms of the
outcomes of the agent and the ease of design of the agents.
social_practice(SP, Requirements) & forall(Requirements).
selected_sp(CurrentSP) & .member(CurrentSP, Options).
1 /* Rules */
2 // Omitted: has_plan_generating_action/4 and for_all/1
3 relevant_sp(SP) :-
4
5 sp_selection(Options, CurrentSP) :-
6
7 sp_selection([SP_], SP).
8
9 /* Initial goal */
10 !metaDeliberate.
11
12 /* Plans */
13 @metaplan[atomic]
14 +!metaDeliberate <-
15
16
17
18
19
20
if (selected_sp(CurrentlySelectedSP)) {
-selected_sp(CurrentlySelectedSP)
if (RelevantSPs == []) {
RelevantSPs);
}
.findall(SP, ( relevant_sp(SP) & not completed_sp(SP) ),
1 /* Rules */
2 context_ok(plan(_,_,ContextCond,_)) :- ContextCond.
3
4 // Rules for filter_list/3 omitted
5
6 /* Plans */
7
8 // Entry points: solve body term list, with or without a
9 // prespecified path through the goal-plan tree
10
11 +!solve(PlanBodyTerms) <- !solve(PlanBodyTerms, [], 1, no_path).
12 +!solve(PlanBodyTerms, Path) <- !solve(PlanBodyTerms, [], 1, Path).
13
14 // solve body term list
15
16 +!solve([], _, _, _).
17 +!solve([body_term(Prefix, Term)BTs], Intn, N, Path) <-
18
19
20
Intn2 = [NIntn];
!solve(Prefix, Term, Intn2, Path);
!solve(BTs, Intn, N+1, Path).
21
22
23
24
25
26
27
28
29
} else {
if ( sp_selection(RelevantSPs, SelectedSP) &
not selected_sp(SelectedSP) ) {
-+selected_sp(SelectedSP)
}
for (monitored(Purpose, SP, ID)) {
if (Purpose) { +completed_landmark(SP, ID, Purpose) }
}
}
.wait(500);
!!metaDeliberate.
30
31
32
33 +selected_sp(SP) <-
34
35
36
37
38
for (landmark(SP, ID, _, _, Purpose)) {
PurposeNoAnnots[dummy] = Purpose[dummy];
if (.intend(PurposeNoAnnots)) {
.suspend(PurposeNoAnnots);
+suspended_intention(SP, ID, PurposeNoAnnots)
39
40
41
42
43
44
45
46
}
.add_plan({@suspend_purpose(SP,ID)
+!PurposeNoAnnots <- .suspend(PurposeNoAnnots)},
landmark(SP,ID), begin)
}
for (landmark(SP, ID, [], Actions, Purpose)) {
!activate_landmark(SP, ID, Actions, Purpose)
}.
47
48 @activate_landmark[atomic]
49 +!activate_landmark(SP, ID, Actions, Purpose) <-
50
51
52
53
54
55
PurposeNoAnnots[dummy] = Purpose[dummy];
+monitored(PurposeNoAnnots, SP, ID)
if (Actions = [action(Actors, Act)] &
if (has_plan_generating_action(
{+!Purpose}, Act, BodyTerms, Path)) {
(Actors = Me (.list(Actors) & .member(Me, Actors)))) {
56
57
58
59
60
61
62
63
64
!!solve([body_term("!",Purpose)], Path)
} else {
if (joint(Act)) {
!!solve([body_term("", Act[participants(Actors)])])
} else {
if (durative(Act)) {
!!solve([body_term("", Act)])
} else {
if (Act =.. [F, _, _] & .substring(".", F)) {
} else { .print("Multiple actions are not yet supported"); }.
!!solve([body_term(".", Act)])
Act
}}}}
} else {
65
66
67
68
69
70
71 @completed_landmark[atomic]
72 +completed_landmark(SP, ID, Purpose) <-
73
74
75
76
77
78
79
80
81
CompletedPrecIDs) &
.succeed_goal(Purpose);
-monitored(Purpose, SP, ID);
.remove_plan(suspend_purpose(SP,ID));
for ( landmark(SP, ID2, PrecedingLMs, Actions, Purpose2) &
not completed_landmark(SP, ID2, _) &
.findall(PrecID, (.member(PrecID, PrecedingLMs) &
.difference(PrecedingLMs, CompletedPrecIDs, []) ) {
completed_landmark(SP, PrecID, _)),
82
83
84
85
86
87
!activate_landmark(SP, ID2, Actions, Purpose2)
}
.findall(ID2, ( landmark(SP, ID2, _, _,_) &
not completed_landmark(SP, ID2, _) ),
PendingLandmarks);
if (PendingLandmarks == []) { +completed_sp(SP) }.
Figure 2: Rules and plans for social practice reasoning
!solve(PBTs, Intn, 1, Path).
21
22 // solve body terms
23
24 +!solve("?", B, _, _) <- ?B.
25 +!solve("+", B, _, _) <- +B.
26 +!solve("-", B, _, _) <- -B.
27 +!solve("!", solve(PBTs), Intn, Path) <-
28
29 +!solve("!", G, Intn, Path) <-
30
31
32
33
34
35
36
37
38 +!solve("", AnnotatedAction, _, _) :
39
40
41
42
43
44
45
46
CUGoal = CleanupGoal;
CUGoal = true;
} else {
} else {
}.
meta.relevant_plan_bodies_as_terms({+!G}, RPlans);
if (.list(Path) & Path = [NPathTail] &
.nth(N, RPlans, plan(Label,_,_,PlanBodyTerms))) {
!solve(PlanBodyTerms, [LabelIntn], 1, PathTail);
?filter_list(RPlans, context_ok, APlans);
!solve_one(APlans, Intn, Path);
Act[dummy] = AnnotatedAction[dummy] & durative(Act) <-
?durative_action_continuation_pred(Act, Query);
if (durative_action_cleanup_goal(Act, CleanupGoal)) {
}
if (joint(Act) & Act[participants(P)] = AnnotatedAction) {
} else {
ParticipantAnnotation = [];
}
if (time(T)) {
ParticipantAnnotation = [participants(P)];
Act[durativeParticipantAnnotation];
-+started(Act, T)[source(meta)];
}
!solve_durative(Query, Act, ParticipantAnnotation, CUGoal).
47
48
49
50
51
52
53
54
55
56 +!solve("", Action, _, _) <- Action.
57 +!solve(".", .fail, _, _) <- .fail.
58 +!solve(".", .print(S), _, _) <- .print(S).
59 +!solve(".", .wait(Cond), _, _) <- .wait(Cond).
60 +!solve(".", .wait(Cond, Timeout), _, _) <- .wait(Cond, Timeout).
61
62 // Solve some plan in a list of plans
63
64 +!solve_one([plan(Label,_,_,PlanBodyTerms)_], Intn, Path)
65
66
67
68
69
70 -!solve_one([_PlanTerms], Intn, Path) <-
71
72
73 +!solve_durative(Query, Act,
74
75
76
77
78
79
80
81
+tried_plan(Label, Intn);
!solve(PlanBodyTerms, [LabelIntn], 1, Path);
-tried_plan(Label, Intn).
stop(Act)[durativeParticipantAnnotation];
if (CleanupGoal \== true) {
Act[durativeParticipantAnnotation];
!solve_durative(Query, Act,
ParticipantAnnotation, CleanupGoal) <-
ParticipantAnnotation, CleanupGoal);
!solve_one(PlanTerms, Intn, Path).
: not tried_plan(Label, Intn) <-
if (Query) {
} else {
}
!CleanupGoal;
82
83
84
85 -!solve_durative(_, Act, _) <-
86
-started(Act, _).
}.
Figure 3: A Jason metainterpreter
REFERENCES
[1] Agnese Augello, Manuel Gentile, and Frank Dignum. 2016. Social Practices
for Social Driven Conversations in Serious Games. In GALA 2015, Alessandro
De Gloria and Remco Veltkamp (Eds.). Springer International Publishing, 100 --
110.
[2] Rafael H. Bordini, Jomi Fred Hübner, and Michael Wooldridge. 2007. Program-
ming multi-agent systems in AgentSpeak using Jason. Wiley.
[3] P. Bourdieu (trans. R. Nice). 1972. Outline of a theory of practice. Cambridge
University Press.
[4] Stephen Cranefield, Michael Winikoff, Virginia Dignum, and Frank Dignum.
2017. No Pizza for You: Value-based Plan Selection in BDI Agents. In Proceed-
ings of the Twenty-Sixth International Joint Conference on Artificial Intelligence.
ijcai.org, 178 -- 184.
[5] F. Dignum. 2018. Interactions as Social Practices: towards a formalization. arXiv
(2018). https://arxiv.org/abs/1809.08751
[6] V. Dignum and F. Dignum. 2015. Contextualized Planning Using Social Prac-
tices. In Coordination, Organizations, Institutions and Norms in Agent Systems
X: COIN 2014, A. Ghose, N. Oren, P. Telang, and Thangarajah J (Eds.). Springer-
Verlag, 36 -- 52.
[7] J.T. Folsom-Kovarik and S. Schatz. 2014. AI challenge problem: Scalable models
for patterns of life. AI Magazine 35, 1 (2014), 10 -- 14.
[8] A. Giddens. 1979. Central problems in social theory: Action, structure and con-
tradiction in social analysis. University of California Press.
[9] Raoul Harel, Zerrin Yumak, and Frank Dignum. 2018. Towards a Generic
Framework for Multi-party Dialogue with Virtual Humans. In Proceedings of the
31st International Conference on Computer Animation and Social Agents (CASA
2018). ACM, New York, NY, USA, 1 -- 6.
[10] Brian Logan. 2018. An agent programming manifesto. International Journal of
Agent-Oriented Software Engineering 6, 2 (2018), 187 -- 210.
[11] Brian Logan, John Thangarajah, and Neil Yorke-Smith. 2017. Progressing Inten-
tion Progression: A Call for a Goal-Plan Tree Contest. In Proceedings of the 16th
Conference on Autonomous Agents and MultiAgent Systems, AAMAS 2017, São
Paulo, Brazil, May 8-12, 2017. ACM, 768 -- 772.
[12] Rijk Mercuur, Frank Dignum, and Yoshihisa Kashima. 2017. Changing Habits
Using Contextualized Decision Making. In Advances in Social Simulation 2015,
Wander Jager, Rineke Verbrugge, Andreas Flache, Gert de Roo, Lex Hoogduin,
and Charlotte Hemelrijk (Eds.). Springer International Publishing, 267 -- 272.
[13] Tim Miller, Virginia Dignum, and Frank Dignum. 2018. Planning for Human-
Agent collaboration using Social Practices. In First international workshop on
socio-cognitive systems at IJCAI 2018.
[14] Marvin Minsky. 1988. A framework for representing knowledge. In Readings in
Cognitive Science, Allan Smith and Edward Collins (Eds.). Morgan Kaufmann,
156 -- 189.
[15] Kavin Narasimhan, Thomas Roberts, Maria Xenitidou, and Nigel Gilbert. 2017.
Using ABM to Clarify and Refine Social Practice Theory. In Advances in Social
Simulation 2015, Wander Jager, Rineke Verbrugge, Andreas Flache, Gert de Roo,
Lex Hoogduin, and Charlotte Hemelrijk (Eds.). Springer International Publishing,
307 -- 319.
[16] A. Reckwitz. 2002. Toward a Theory of Social Practices. European Journal of
Social Theory 5, 2 (2002), 243 -- 263.
[17] Theodore R. Schatzki. 2012. A Primer on Practices. SensePublishers, Rotterdam,
13 -- 26.
[18] E. Shove, M. Pantzar, and M. Watson. 2012. The Dynamics of Social Practice.
Sage.
[19] Michael Winikoff. 2006. An AgentSpeak Meta-interpreter and Its Applications.
In Programming Multi-Agent Systems, Third International Workshop, ProMAS
2005. Lecture Notes in Computer Science, Vol. 3862. Springer, 123 -- 138.
|
1304.5545 | 1 | 1304 | 2013-04-19T21:00:37 | Designing Electronic Markets for Defeasible-based Contractual Agents | [
"cs.MA"
] | The design of punishment policies applied to specific domains linking agents actions to material penalties is an open research issue. The proposed framework applies principles of contract law to set penalties: expectation damages, opportunity cost, reliance damages, and party design remedies. In order to decide which remedy provides maximum welfare within an electronic market, a simulation environment called DEMCA (Designing Electronic Markets for Contractual Agents) was developed. Knowledge representation and the reasoning capabilities of the agents are based on an extended version of temporal defeasible logic. | cs.MA | cs |
Designing Electronic Markets for
Defeasible-based Contractual Agents
Adrian Groza
Technical University of Cluj-Napoca
Department of Computer Science
Baritiu 28, RO-3400 Cluj-Napoca, Romania
[email protected]
Abstract. The design of punishment policies applied to specific domains
linking agents actions to material penalties is an open research issue. The
proposed framework applies principles of contract law to set penalties: ex-
pectation damages, opportunity cost, reliance damages, and party design
remedies. In order to decide which remedy provides maximum welfare
within an electronic market, a simulation environment called DEMCA
(Designing Electronic Markets for Contractual Agents) was developed.
Knowledge representation and the reasoning capabilities of the agents
are based on an extended version of temporal defeasible logic.
Keywords: contractual agents, legal remedies, defeasible logic.
1
INTRODUCTION
According to [?], there are five different philosophies of punishment from which
all punishment policies can be derived: deterrence, retribution, incapacitation,
rehabilitation and restoration. Retribution considers that the remedy should be
as severe as the wrongful act, making this doctrine most suitable for multi-
agent systems. Consequently, we drive our attention toward four legal doctrines
from contract law that aim to equal the victim's harm: expectation damages,
opportunity costs, reliance damages, and party designed remedies. The output
of this research formalizes the above remedies for multi-agent systems within a
framework for computing penalties for B2B disputes.
The remedies imposed by law affect the agents' behaviour [?]: (1) searching
for trading partners; (2) negotiating exchanges; (3) keeping or breaking com-
mitments; (5) taking precaution against breach causing events; (6) acting based
on reliance on promises; (7) acting to mitigate damages caused by broken com-
mitments; (8) settle disputes caused by broken promises. In the supply chain
context a contract breach can propagate over the entire chain. The damages im-
posed by legal institutions can positively influence breach propagation. Usually,
a contract breach appears when some perturbation arises on the market1. The
1 For instance, the market price of a raw material could rise so much that, for the
agent who had planed to achieve it in order to produce an item, is more efficient to
breach the contract with its buyer.
question is which of the above remedies is adequate for an efficient functionality
of the electronic market. Since normative reasoning is defeasible by nature, we
developed a framework which provide defeasible reasoning capabilities for agents
to act in case of breach.
The problem in hand is approached by: (i) formalising contractual clauses
and contract law remedies for multi-agent systems; (ii) providing a system for
designing experiments in order to decide which legal doctrine suits an electronic
market; iii) providing defeasible logic-based mechanisms for enhancing reasoning
capabilities of the agents about penalties. The paper is organized as follows: The
next section introduces contracts within the task dependency network model.
Section 3 formalises four types of remedies for multi-agent systems according
to contract law. In section 4 the functions used by the market for penalties
are implemented. Section 5 describes the implemented system. Section 6 details
related work and section 7 concludes the paper.
2 PROBLEM SPECIFICATION
2.1 Commitment Dependency Network
The task dependency network model [?], used in the analysis of the supply chain,
was adapted [?] as follows: commitment dependency network is a directed, acyclic
graph, (V,E), with vertices V = G∪A, where: G = the set of goods, A = S∪P ∪C
the set of agents, S = the set of suppliers, P = the set of producers, C = the set
of consumers, and a set of edges E (commitments) connecting agents with their
input and output goods. With each agent a ∈ A we associate an input set Ia and
s1
s2
s3
s4
s5
s6
g1
g2
p1
p2
p3
g3
g4
g5
p4
p5
g6
p6
g7
c1
c2
c3
Fig. 1. Task dependency network: goods are indicated by circles, suppliers and
consumers are represented by boxes, while producers by curved boxes.
an output set Oa: Ia = {g ∈ G ≺ g, a ≻∈ E} and Oa = {g ∈ G ≺ a, g ≻∈ E}.
Agent a is a supplier if Ia = ∅, a consumer if Oa = ∅, and a producer in all other
cases. Without any generalization lost, we consider that a consumer c ∈ C needs
a single item (Ic = 1) and every supplier s ∈ S or producer p ∈ P build one
single item (Os = 1 and Op = 1)
An agent must have a contract for all of its input goods in order to produce
its output, named presumable 2 and denoted by p. If we note np = Ip, the agent
has to sign np + 1 contracts in order to be a member in the supply chain. For
each input good gk ∈ Ip the agent p bids its own item valuation vk
p . The auction
for the good gk sets the transaction price at pk. The agent's investments are
p the agent's
investments but without considering the investments made for the current good
k=1 vk
p
and this value without considering the bid for good g as V g
p . For the output good,
the agent p signs a contract at reliance price Rp. We consider that there are no
production costs and when perturbation or unexpected events occur, agents need
protocols for repairing or reforming the supply chain.
Ip = Pnp
g. Similarly, we note all bids values submitted by the agent p as Vp = Pnp
k=1 pk where k are the winning input goods. We note by I g
2.2 Contracts
By extending social commitments [?], we define a set of contractual patterns
used to declaratively specify the contracts signed between parties. The classical
definition of a conditional commitment states that a commitment is a promise
from a debtor x to a creditor y to bring about a particular sentence p under a
condition q. Starting from this definition we formalise a set of contractual clauses
inspired from contract law:
In a Gratuitous Promise the debtor x promises the creditor y to bring
about p until tmaturity without requesting anything: C 0
1 (x, y, 1, p1, tmaturity).
A Unilateral Contract involves an exchange of the offerer's promise p for the
oferee's act q, where the debtor x promises the creditor y to bring about p
until tmaturity if condition q holds at time tissue. For instance, in the contract
C =≺ as, ab, gi, Pc, tissue, tmaturity ≻, as represents the seller agent, ab the buyer
agent, gi the good or the transaction subject, Pc the contract price, tissue is the
time when the offer is accepted and tmaturity is the time when the transaction
occurs. In a Bilateral Contract both sides make promises, the debtor x promises
the creditor y to bring about p if the creditor y promises x to bring about p1,
formalised as C(x, y, C(y, x, 1, p1), p).
During its life cycle, a commitment may be in one of the following states: open
offer, active, released, breached, fulfilled, canceled, or failed, which are also useful
to be considered from a legal perspective. The state transition between open offer
and active contract depends on each type of commitment: the acceptance of a
gratuitous commitment means reliance and acting upon it, the acceptance of a
unilateral contract means the execution of the required task, the acceptance of
a bilateral contract means the creation of the required promise.
2 Note that when someone breaches a contract with a presumable agent it has to pay
more damages.
3 REMEDIES
The remedies described in this section try to equal the victim's harm. In the
first three cases3, the system estimates the harm according to current market
conditions, while in the last case, the agents themselves compute the damages
and generate their own penalties.
Expectation Damages. The courts reward damages that place the victim of
breach in the position he or she would have been in if the other party had
performed the contract [?]. Therefore, in an ideal situation, the expectation dam-
ages does not affect the potential victims whether the contract is performed or
breached. Ideal expectation damages remain constant when the promisee relies
on the performance of the contract more than it is optimal.
Reliance Damages. Reliance increases the loss resulting from the breach of the
contract. Reliance damages put the victim in the same position after the breach
as if he had not signed a contract with the promisor or anyone else [?]. In
an ideal situation, the reliance damages do not affect the potential victims if
the contract is breached or there was no initial contract. No contract provides a
baseline for computing the injury. Using this baseline, the courts reward damages
that place the victims in the position that they would have been, if they had
never contracted with another agent. Reliance damages represent the difference
between profit if there is no contract and the current profit.
Opportunity Cost. Signing a contract often entails the loss of an opportunity to
make an alternative. The lost opportunity provides a baseline for computing the
damage. Using this baseline, the courts reward damages that place victims of
breach in the position that they would have been if they had signed the contract
that would have been the best alternative to the one that was breached [?]. In
the ideal situation, the opportunity cost damages does not affect the potential
victims whether the contract is breached or the best alternative contract is per-
formed4. If breach causes the injured party to purchase a substitute item, the
opportunity cost formula equals the difference between the best alternative con-
tract price available at the time of contracting and the price of the substitute
item obtained after the breach.
Party-Designed Remedies. The contract might stipulate a sum of money that
the breaker will pay to the party without guilt. These "leveled commitment
contracts" [?] allow self-interested agents to face the events that unfolded since
the contract started. A rational person damages others whenever the benefit is
large enough to pay an ideal compensation and have some profit, as required to
3 Expectation damages, reliance damages and opportunity cost are analyzed from an
economical point of view in [?,?].
4 Opportunity cost and expectation damages approach equality as markets approach
perfect competition.
increase efficiency. Game theoretic analysis has shown that leveled committed
contracts increase the Pareto efficiency. One contract may charge a high price and
offer to pay high damages if the seller fails to deliver the goods, while another
contract may charge a low price and offer to pay low damages, the types of
contracts separating the set of buyers and allowing "price discrimination."
4 CASE ANALYSIS
The conclusions from the last sections are: (i) The amount of expectation dam-
ages must place the victim in the same position as if the actual contract had
been performed; (ii) The amount of reliance damages must place the victim in the
same position as if no contract had been signed; (iii) The amount of opportunity-
cost damages must place the victim in the same position as if the best alternative
contract had been performed; (iv) Party designed remedies specify themselves
the amount of damages in case of a breach.
4.1 No substitute
12
p5
s5
11
12
g5
s6
13
15
c3
c3
p5
Cp=12$
$
15
14
13
12
11
s6
s5
−1 1
Items
Fig. 2. Supplier-Consumer contract
Supplier-Consumer
The consumer breaches the contract. In fig. 2a) the suppliers s5 and s6 want to
sell good g5 at price 11 and respectively 13, while the agents p5, and c3 try to
buy it at prices 12 and 15. According to (M+1)st price protocol the transaction
price is Pc = 12$5. The auction clears at every round. In fig. 2b) a single uni-
lateral contract is signed: C 1
g5 =≺ s5, c3, g5, 12, tissue, tmaturity ≻. Consider c3
breaches the contract. In this case, the remedies will be: i) Expectation damages:
if the agent c3 performs, the s5's estimated profit is the difference between the
5 The goods are transacted using the (M+1)st price auction protocol, which has the
property to balance the offer and the demand at each level in the supply chain
(otherwise the supply demand equilibrium cannot be achieved globally). It provides
a uniform price mechanism: all contracts determined by a particular clearing are
signed at the same price.
contract price Pc = 12 and its own valuation6 vg5
a6 = 11 (victim valuation). The
remedies compensate this value: De = Pc − vg
a. ; ii) Opportunity damages: first,
the auctioneer has to compute the opportunity cost Po, which is the transac-
tion cost in case the breacher was absent from the auction. In fig. 2, if agent c3
is not present Po = 11. The s5's bid is one who wins. The contract would be
g5 =≺ s5, p5, g5, 11, tissue, tmaturity ≻ and the agent's profit would be Po − vg
C 1
a.
But, when there is no contract for agent s5, his profit would be null. The op-
portunity damages should reflect this. We define opportunity cost damage Do
which is received by the agent a as: Do = max(Po − vg
a, 0); iii) Reliance dam-
ages: if the victim does not have any input good, the supplier's investments in
performing are null: Dr = 0; iv) Party-designed remedies: the remedies may be
a fraction from the contract price (Dp = α · Pc), a fraction from the expected
profit (Dp = α · De), or constant (Dp = C).
The supplier breaches the contract. Consider s5 breaches contract C 1
g5 : i) Expec-
tation damages: De = vg
a − Pc; ii) Opportunity damages: if the breacher had not
bid and the victim had signed a contract at the opportunity price Po, than it's
profit would have been vg
a − Po. If the victim has no contract when the breacher
is not bidding, it receives no damages. Hence, Do = max(vg
a − Po, 0). In the
depicted case, if the agent s5 had not existed, c3 would have signed a contract
with s6 for an opportunity cost Po = 12. Therefore, Do = 3; and iii) Reliance
damages: because the client does not produce any output goods, it's reliance is
null: Dr = 0.
Supplier-Producer
The supplier breaches the contract. Consider the contract from fig. 3 C =≺
s5, p5, g5, 12, tissue, tmaturity ≻: i) Expectation damages: Observe that the victim
is a presumable agent because it has contracts for all its input goods. Its invest-
ments are Ip = 5 + 9 + 12 = 26 and I g5
p = 9 + 5 = 14. The producer p5 has also
a contract for its output item, so Rp5 = 34. Its profit is Rp5 − Ip = 8. When bad
contracts have been signed this value can be negative, therefore no damages are
imposed. Otherwise, expectation damages equals the difference between its bid
and the contract price:
p, ∃Rp
otherwise
Dr =(V g
p − Po, 0),
max(Rp − I g
vg
p − Po,
0,
p, ∃Rp
otherwise
p + Rp − vgo
p − I g
p ,
V g
p − I gk
p ,
p, ∃Rp, ∃Po
¬p, ∃Rp, ∃Po
¬∃Po
vg
p − Pc,
De =(max(Rp − Ip, 0),
Do =
Recall p means that agent p is presumable, go is the output good of the agent
represents all k contracts signed for input goods, where k < np. In the
p and I gk
p
6 (M+1)st price auction has the following property: the dominant strategy for each
agent is to reveal its real valuation.
5
g3
9
g4
12
g5
34
32
34
g6
35
5
p4
10
p5
15
13
c3
s5
s6
10
12
Fig. 3. Supplier-Producer contract
depicted case p5 is presumable and there is a contract with a buyer. Therefore, it
has to receive, as a victim, the next reliance damages Dr = V g5
p5 =
(10 + 5) − (9 + 5) + 34 − 32 = 3. In case of Opportunity costs damages, one seller
less implies Po ≥ Pc.
p5 +Rp5 −vg6
p5 −I g5
In some cases damages can be higher than the contract value itself (Dr ≥
Pc). According to current practice in law, these damages are the right ones if
the victim gives a previous notification about the risks faced by the potential
breacher. This is a clear situation when information propagation improves the
supply chain performance. In the light of the above facts, their reliance damages
should remain the mentioned ones if the victim has notified its partner, but
should be maxim Pc otherwise. Hence, we define D′
r equals Dr when the breacher
receives a notice, and D′
r = min(Dr, Pc), otherwise.
p4
34
p5
32
5
10
15
34
g6
35
p6
g7
37
c1
c2
Fig. 4. Producer-Consumer contract
Producer-Consumer
The consumer breaches the contract Consider the unilateral contract C =≺
p5, c2, g6, 34, tissue, tmaturity ≻ from fig. 4, where c2 breaches: i) In the case of
Expectation damages, p5 is presumable and De = 34 − (12 + 9 + 5) = 8. Suppose
the agent p5 does not have any contract for one input good. Therefore, it is not
presumable and it will receive De = 34 − 32; ii) Reliance damages: Dr = Vp − Ip;
and iii) Opportunity cost : one buyer less implies Po ≤ Pc:
De =(max(Pc − I g
Pc − vg
p ,
p , 0),
4.2 Substitute
p
otherwise
p , 0),
max(Po − I g
Po − vg
p ,
0,
p, ∃Po
¬p, ∃Po
∃Po)
Do =
The common law requires the promisee to mitigate damages. Specifically, the
promisee must take reasonable actions to reduce losses occurred by the promisor's
breach. The market can force the victim to find substitute items, in this case
the imposed damages reflect only the difference between original contract and
substitute contract. With a substitute contract, the victim signs for the identical
item, with the same deadline or tmaturity, but at a different price. Let Ps be the
value of the substitute contract. For the general case Producer-Producer, when
the buyer breaches the contract, the equations become:
p , 0),
max(Pc − I g
Pc − vg
p ,
max(Pc − Ps),
p, ¬∃Ps
p, ¬∃Ps
∃Ps
Dr =(Vp − Ip,
max(Pc − Ps, 0), ∃Ps
¬∃Ps
De =
max(Po − I g
max(Po − vg
max(Po − Ps, 0), ∃Ps
p, ∃Po, ¬∃Ps
p , 0),
p , 0), ¬p,∃Po, ¬∃Ps
Do =
5 DEMCA: A SIMULATION ENVIRONMENT FOR
CONTRACT BREACH EXPERIMENTS
5.1 System Architecture
Designing Electronic Markets for Contractual Agents (DEMCA) tool7 consists
of the next three components (figure 5):
Agents. For defining a contractual agent in DEMCA one has to instantiate
a generic agent with the following structure (figure 6). The type of the agent
identifies its role in the supply chain: supplier, producer, or consumer. The
agent's strategy is defined by the function program, which points to a set of
7 Available at http://cs-gw.utcluj.ro/∼adrian/odr.html
Fig. 5. System architecture.
(defun generic_agent (name &optional (program (program name)) stored_items
active_contracts input_goods output_goods valuation_in
valuation_out money type))
(MAKE-agent :name name :type type :program program
:legal-actions '(bid breach) :stored_items stored_items
:active_contracts active_contracts :score money
:input_goods input_goods :output_goods output_goods
:valuation_in valuation_in :valuation_out valuation_out))
Fig. 6. Agent's structure in DEMCA.
temporalised defeasible rules coded in the TeDeLo (Temporal Defeasible Logic)
framework. Examples of strategies include: reactive/proactive agents, breaching
often/seldom agents, information propagation/hiding agents. The agent has a list
o percepts consisting in public information provided by the market: the current
open auctions, the existing bids, the agents, the public contracts. Each agent has
attributes keeping stored items, active contracts, or the amount of money avail-
able at the beginning of each experiment round (figure 6). Other information
that can be provided is: the goods which the agent is interested in achieving, the
good that the agent produces, the own valuation of these items. The public or
private character if these information depends on the current simulation.
Environment. The simulation environment introduces auctions for each item
within the market, to which the agents manifest their buying or selling interests.
The market is governed by a discrete time step. At the beginning of each round
the agent perceives the public information, and after a reasoning process, it
conveys one of the following actions: bid, breach, no action. A contract might
be signed in case the agent wins the M+1 price auction. If the agent has all
the necessary preconditions, the environment simulates the contract execution.
Otherwise, if the preconditions are not met at the time of maturity and no breach
action was executed, the environment automatically computes the amount of
remedy according to the legal doctrine in force.
The auctions run in parallel, following the same M+1 price protocol. After
the bids were sent, contracts are generated for the winning agents, the public
information being broadcasted to all the agents within the task dependency net-
work. The environment is responsible for computing the profit for each agent,
profit which includes the possible penalties. During DEMCA experiments five
types of remedies can be used: expectation damages, opportunity costs, reliance
damages, fixed penalty, and percentage from the contract value. The output of
the simulation is analysed by some scripts coded in R language, responsible for
generating graphics presenting the evolution of the agents' score during experi-
ments.
Reasoning capabilities. The framework provides defeasible reasoning capa-
bilities to the DEMCA agents to flexibly represent market strategies under time
constraints and levels of certainty for handling preferences when deciding to
breach or not a contract. By enhancing agents with defeasible reasoning, they
are able to reason with incomplete information, including confidential contrac-
tual clauses. On the other hand, defeasible logic has already been proved to be
suitable for legal reasoning [?].
The expressivity of defeasible logic is given by three types of rules: Strict rules
are rules in the classical sense, that is whenever the premises are indisputable,
then so is the conclusion, while defeasible rules are rules that can be defeated by
contrary evidence. For "sending the goods means the goods were delivered", if we
know that the goods were sent then they reach the destination, unless there is
other, not inferior, rule suggesting the contrary. Defeaters are rules that cannot
be used to draw any conclusions. Their only use is to prevent some conclusions,
as in "if the customer is a regular one and he has a short delay for paying, we
might not ask for penalties". This rule cannot be used to support a "not penalty"
conclusion, but it can prevent the derivation of the penalty conclusion. Defeaters
have a particular role in this context due to their capability to explicitly represent
exception, or contract breaches in our case.
Defeasible derivations have an argumentation like structure: firstly, we choose
a supported rule having the conclusions q we want to prove, secondly we consider
all the possible counterarguments against q, and finally we rebut all the above
counterarguments showing that, either some of their premises do not hold, or
the rule used for its derivation is weaker than the rule supporting the initial
conclusion q.
The TeDeLo (Temporal Defeasible Logic) module implements an extended
version of temporal defeasible logic ETDL, based on LISA (Lisp based Intelligent
Time step 1:
Agent S15, 100, NIL, NIL, (G5) perceives:
((#S(GOOD :NAME G5
:SELL_AGENTS (S15, 100, NIL, NIL, (G5))
:BUY_AGENTS (C4, 100, NIL, (G5), NIL)
:SELL_BIDS NIL
:BUY_BIDS NIL
:CONTRACT NIL)))
and does:
(((BID #S(GOOD :NAME G5
:SELL_AGENTS (S15, 100, NIL, NIL, (G5))
:BUY_AGENTS (C4, 100, NIL, (G5), NIL)
:SELL_BIDS NIL
:BUY_BIDS NIL
:CONTRACT NIL))))
Fig. 7. Trace of the agent's percepts and actions in DEMCA.
Software Agents). Compared to the existing systems, the current implementation
extends the defeasible logic with: (1) temporal defeasible reasoning, (2) certainty
factor instead of a partial order relation, (3) open world assumption applied to
the rules, (4) dynamic rules, (5) dynamic priorities, (6) identifying of hidden
rules, (7) pattern of rules for representing cyclic events or facts, (8) rollback ac-
tivation to better handle exceptions. The formalism was extended with temporal
aspects in order to reason with contract deadlines. Dynamic rules and priorities
allow agents to adjust their market strategies according to the changes in the
environment.
5.2 Running scenarios
Basic Example In order to trace a simple scenario, two agents were added to
the framework, by calling the function add agent :
(add_agent 'reactive_agent 's15 '() '() '()
(add_agent 'reactive_agent 'c4
'() '() '(g5) '()
'(g5) 5 11 100)
15 16 100)
Here, the supplier s15 is interested in selling the item g5 for a minimum
price of 11 and needs 5 units to produce it, while the client c4 is interested in
buying the item g5 at a maximum price of 15, its own valuation being 16. At the
beginning of the experiment both agents have 100 monetary units and no stored
item.
Figure 7 presents part of a DEMCA session. During the first round, the
environment automatically starts an auction for the item g5. The agent s15
perceives this open auction, the list of agents which are interested in selling
the item (only itself in this case), the list of agents interested in buying it (the
consumer c4), the current buying or selling offers, at the output in case the
auction clears. Consider, the agent s15 has a reactive strategy, in which if it has
an item for selling and no active contract for it, the agent bids its own valuation
of the item. This is formalised as a defeasible theory as follows:
(defeasible 'r1 0.8 'neg 'contract 1 1 'poz bid 1 3)
(defeater 'r2 0.9 'poz 'win 1 1 'neg bid 2 3)
Here, the defeasible rule r1 supports with a certainty factor of 0.8 the consequent
bid for all time instances within the interval [1, 3] in case the agent has not signed
a contract for the desired item ('neg contract ) during the first round. In defeasible
reasoning, the conclusion of a defeasible rule can be blocked by contrary evidence,
in case it has less degree of support. In the depicted case, the defeater r2 may
rebuttaly block the derivation of the bid consequent, by supporting the opposite
conclusion 'neg 'bid. In case the agent won the auction at timestep 1 ('win 1 1 ),
the bid action is no longer derived because the defeater r2 provides a stronger
degree of support. In DEMCA, one can dynamically adjust these parameters,
in order to tune the strategy to the market conditions. By using patterns of
rules facility provided by the platform, the agent can specify cyclic activations
of specific events.
Browsing the percepts sensed at the beginning of each round, the agent can
incorporate a part of them into his defeasible theory, using the function (fact
'neg 'contract 1.0 1 1), which stipulates that the agent does not have a contract
at round 1. Because the information is considered sure, the certainty factor of
this fact is set to 1.0. Having this fact asserted, the supplier s15, according to
rule r1, will execute the bid action (figure 7).
Large scale experiments First, the framework can be used as a tool for au-
tomated online dispute resolution (ODR). There are three situations: (i) The
market may have substantial authority, hence one remedy is imposed to all
agents. In this case, the amount of penalties can be automatically computed
with this framework; (ii) Consistent with party autonomy, the agents may set-
tle on different remedies at contracting time. This approach increases flexibility
and efficiency, because the agents are the ones who know what type of remedy
better protects their interests; (iii) All the above remedies influence the amount
of penalties: in this approach the role of the framework is to monitor the market
and collect data such as: the expected profit, the opportunity cost, the amount
of investments made, if there is a substitute at tbreach. All these information is
used as arguments when the dispute is arbitrated [?] in an architecture which
combines rule based reasoning (laws) and case based reasoning (precedent cases).
Second, knowing the bids, the actual contracts, the amount of potential reme-
dies, and the available offers on the market, the framework can identify situations
in which for both agents is more profitable to breach the contract when a for-
tunate or an unfortunate contingency appears. It computes pairs of suggestions,
helping to increase total welfare towards Pareto frontier.
Third, as a simulation tool, the market designer may obtain results regarding
the following questions: what types of remedies assure flexibility in the supply
chain? or how information sharing influences total revenues or can be used to
compute optimum reliance? In the developed prototype we are currently mak-
ing experiments with different types of agents: low-high reliance, breach often-
seldom, sharing information-not sharing, risk seeking-averse (when they are risk
averse, the penalties do not need to be so high to force breachers behave appro-
priately).
6 RELATED WORK
The task dependency network model was proposed [?] as an efficient market
mechanism in achieving supply chain coordination. The authors analyze pro-
tocols for agents to reallocate tasks for which they have no acquired rights.
However, this approach is rather a timeless-riskless economy. On real markets
a firm seldom signs contracts with its buyers and its suppliers simultaneously.
Moreover, the breach of a contract implies no penalties, which is an unrealistic
assumption in real world. In contrast, in the DEMCA framework auctions end
independently, and we introduce penalties in case of contract breaching.
The role of sanctions in multi-agent systems [?] is the enforcement of a so-
cial control mechanism for the satisfaction of commitments. We focus only on
material sanctions and we do not include social sanctions which affect trust,
credibility or reputation. Moreover, we have applied four types of material reme-
dies in a specific domain. In the same spirit of computing penalties according
to the level of harm produced, the amount of remedies may depend on the time
when the contract was breached [?]. Expectation damages, reliance damages,
and opportunity costs have also been studied [?,?,?] and how contracts influence
the supply chain coordination or strategic breach appear in [?,?].
Experiments regarding breach penalties were attacked from a game theory
viewpoint in [?]. The contract specifies decommitment penalties for both parties.
The resulting leveled commitment contracts enable agents to sign agreements
that would be inefficient in case of full commitment contracts. The efficiency
increases according to the fitness of the penalty to market conditions. Our hy-
pothesis is that, even if the amount is not known apriori the penalties introduced
in this paper can be proved efficiently because they perfectly reflect the environ-
ment conditions. Further experiments are required to validate this idea. There
are two issues using levelled commitment contracts in real life scenarios. On the
one hand, a contract has penalties associated to the main contractual clauses.
When a situation which has not been enclosed into the penalty clauses occurs,
the system still has to provide a mechanism of computing remedies. On the other
hand, the agents are legally bind to a penalty if there is a normative theory be-
yond that penalty. Otherwise, the agent can invoke many legal doctrines in order
to avoid a penalty. This is not the case in DEMCA framework. It is based on
such normative theory, and in case of unexpected exceptions one can apply legal
doctrines in order to compute the fair amount of penalty.
As commitments appear to be sometimes too restrictive (direct obligations)
and sometimes too flexible, enriched agents with defeasible reasoning mechanism,
the resulting defeasible commitments are more flexible than usual obligations
but also more constrained than permissions [?]. Dr-contract [?] also provides
defeasible reasoning mechanism for representing business rules. By introducing
temporised position, DEMCA framework can handle more realistic contracts
deadlines. Another difference consists in the contract representation, in our case
the contract is an extension of a social commitment in accordance with practice
in law: gratuitous promise, unilateral contract, bilateral contract. By introducing
defeasible logic when reasoning about commitments, we obtain two main advan-
tages. On the one hand, agents can reason with incomplete information. Also,
this property of nonmonotonic logics permits, in our case, to model confidential
contractual clauses. Expected exceptions can be captured by defining a prefer-
ence structure over the runs within the commitments dependency network [?].
Having the superiority relation from defeasible logic, we can easily define such a
structure in our framework.
7 CONCLUSIONS
This paper introduces the DEMCA framework for designing electronic markets
for contractual agents. While the amount of existing literature on e-commerce
focuses on contract negation or contract formation, we focus on the contract
breach phase during the lifecycle of a contract.
The contribution contains two ideas. On the one hand, we apply the princi-
ples of contract law in the task dependency network model [?]. As a result, we
enrich that model by including different types of penalties when agents breach,
thus bringing the model closer to the real world. On the other hand, the imple-
mented framework can be used for simulations when designed electronic markets,
knowing that the dispute resolution mechanism is a key factor in the success of
e-commerce applications [?]. Such a framework is useful for automated online
dispute resolution. The data obtained can be used as arguments in a mediated
dispute or the remedies can be computed in real time in case the agents agreed
with the market policy.
|
1906.00401 | 2 | 1906 | 2019-06-17T05:06:19 | Multiple Drones driven Hexagonally Partitioned Area Exploration: Simulation and Evaluation | [
"cs.MA",
"cs.RO"
] | In this paper, we simulated a distributed, cooperative path planning technique for multiple drones (~200) to explore an unknown region (~10,000 connected units) in the presence of obstacles. The map of an unknown region is dynamically created based on the information obtained from sensors and other drones. The unknown area is considered a connected region made up of hexagonal unit cells. These cells are grouped to form larger cells called sub-areas. We use long range and short range communication. The short-range communication within drones in smaller proximity helps avoid re-exploration of cells already explored by companion drones located in the same subarea. The long-range communication helps drones identify next subarea to be targeted based on weighted RNN (Reverse nearest neighbor). Simulation results show that weighted RNN in a hexagonal representation makes exploration more efficient, scalable and resilient to communication failures. | cs.MA | cs |
MULTIPLE DRONES DRIVEN HEXAGONALLY PARTITIONED AREA
EXPLORATION: SIMULATION AND EVALUATION
Ayush Datta
Rahul Tallamraju
Kamalakar Karlapalem
Agents and Applied Robotics Group
International Institute of Information Technology, Hyderabad
Gachibowli, Hyderabad, India
{ayush.datta,rahul.t}@research.iiit.ac.in, [email protected]
ABSTRACT
In this paper, we simulated a distributed, cooperative path planning technique for multiple drones (∼200)
to explore an unknown region (∼10,000 connected units) in the presence of obstacles. The map of an
unknown region is dynamically created based on the information obtained from sensors and other drones.
The unknown area is considered a connected region made up of hexagonal unit cells. These cells are grouped
to form larger cells called sub-areas. We use long range and short range communication. The short-range
communication within drones in smaller proximity helps avoid re-exploration of cells already explored by
companion drones located in the same subarea. The long-range communication helps drones identify next
subarea to be targeted based on weighted RNN (Reverse nearest neighbor). Simulation results show that
weighted RNN in a hexagonal representation makes exploration more efficient, scalable and resilient to
communication failures.
Keywords: Multi Robot Coordination, Drones, Cooperation
1
INTRODUCTION
1.1 Motivation
These days, drones have been used increasingly for exploration of an unknown area. There are various
applications of area exploration like search and rescue, map building, intrusion detection, and planetary
exploration. Building co-ordinated multiple basic drones is cheaper than one expensive and complex drone.
The idea behind using co-ordinated drones is that if one drone takes a certain amount of time to explore a
region, then two co-ordinated drones should take half the time taken by a single drone. Similarly, if more
drones are exploring a certain region, they should reduce the exploration time drastically. Assuming that
some identical drones are equipped with sensing, localization, mapping and communication capabilities,
they need to explore the unknown area efficiently and reliably.
The challenge of exploring an unknown area with obstacles is that the drones tend to explore the same
area multiple times, which further tends to increase due to lesser communication between them. Since
SummerSim-SCSC, 2019 July 22-July 24, Berlin, Germany; c⃝2019 Society for Modeling & Simulation International (SCS)
Datta, Tallamraju, and Karlapalem
centralized coordination algorithms usually suffer from a single point of failure problem, it is desirable to
have a distributed co-ordinated system to improve system reliability.
The drones have their own memory in which they keep track of the area explored and the information about
the position of the obstacles. If they do not communicate this information among themselves, then they tend
to explore the area which is already covered by the other drone. This further increases the exploration time.
If they communicate less and exchange less information, then it leads to a more redundant exploration of the
cells. At the same time, we need to reduce the amount of communication between the drones and make the
system more resilient to communication failures. Further, the mechanism of co-ordination should be such
that it is scalable to a large area with multiple drones.
Efficient exploration of the area means that drones do not explore the same area multiple times and they
explore the complete area in a minimum amount of time and with minimum drone traveling distance and
with minimal communication between them. To address these issues, we propose Hexagonally Partitioned
Area Exploration using Reverse Nearest Neighbors. It is a modification of the distributed frontier based
algorithm (Wang, Liang, and Guan 2011) with Particle Swarm Optimization (PSO). In this paper, we show
that our solution outperforms the PSO model based approach.
The area to be explored is tessellated into hexagonal subareas, unlike the traditional quadrangular grid. First,
the drones explore their own subarea using the traditional frontier based algorithm with a limited or short
range communication with drones present in the same subarea. After exploration of the own subarea, the
drones decide which subarea to explore based on weighed reverse nearest neighbor (RNN). This is aided
by the long range communication to figure out which drones are planning to approach which subarea. The
RNN assigns the drones to the subareas in such a way that the communication and area exploration time
reduces drastically.
1.2 Previous Work
The initial work in the field of area exploration started with a single autonomous robot with sensing, local-
ization and mapping. The work on multi-robot exploration was started by Yamauchi (1997) who introduced
distributed frontier algorithm. Since there is no explicit co-ordination among the robots, they tend to move
towards the same frontier cells which introduces inefficiency. Based on similar frontier concept, Simmons
et al. (2000) developed a semi-distributed multi-robot exploration algorithm which requires a central agent
to evaluate the bidding from all the other robots to obtain the most information gain while reducing the
cost, or the traveling distance. Berhault et al. (2003) used Combinatorial Auction for the area exploration.
The disadvantage of the bidding algorithms is that they have a single point of failure and since the com-
munication cost increases drastically with the increase in the number of robots, these algorithms are not
scalable.
There has also been work on heterogeneous robots exploring the area by Singh and Fujimura (1993). How-
ever, they do not focus on efficiency. Market based map exploration was also proposed by Zlot et al. (2002).
Ant-inspired algorithm to divide the area into square cells on which the robots leave trails of their passage
was done by Koenig and Liu (2001). All these strategies require a central shared memory where the robots
can leave their marks in the environment. This introduces a single point of failure which makes the sys-
tem less robust. Fu, Bandyopadhyay, and Ang (2009) made use of local Voronoi decomposition for task
allocation in map exploration.
Kennedy and Eberhart (1995), proposed the PSO model which was inspired by the earlier research by bird
flocks. Al-Obaidy and Ayesh (2008) added the power cost as the indicator in PSO model to the WSN
optimization. Dasgupta, Cheng, and Fan (2009) built a robot team based on Reynolds' flocking model
to improve the efficiency of the multi-robot map exploration. Wang, Liang, and Guan (2011) proposed
Datta, Tallamraju, and Karlapalem
PSO model for frontier based exploration. Dornhege and Kleiner (2013) combined the concept of voids
with frontier based approach to search for entombed victims in confined structures. Yoder and Scherer
(2016) improvises frontier algorithm performance by using only state-changed space in the 3D map in
each iteration. Mahdoui, Frémont, and Natalizio (2018) modifies frontier based approach where instead of
sharing local maps, robots share their local frontier points with a so-called Leader.
In Quadrangular RNN, we show that RNN outperforms the traditional PSO model as it keeps a track of
which drones are approaching which subarea instead of keeping track of drones that are in the path of the
subarea to be explored. As proposed by Sheng et al. (2006), we need to use a nearness measure so that the
agents remain within the communication radius. This approach leads to duplicate exploration of the area. To
overcome this, in Quadrangular RNN with SRC we show that using two kinds of communication improves
the exploration efficiency. Recently, a lot of research has been done on Multi-Robot area exploration and
swarm intelligence, but most of these algorithms use a quadrangular grid to divide the area into subareas.
In Hexagonal RNN with SRC, we show that using a hexagonal grid over a quadrangular grid reduces the
exploration time as there are six degrees of directions to move instead of four. Finally, Weighted HRNN
with SRC also highlights how this system is failure resilient, scalable and avoids the possible deadlocks.
2 PROBLEM FORMULATION
We assume that identical drones are equipped with sensing, localization and communication capabilities.
The task is to explore an unknown area avoiding the obstacles efficiently and reliably, at the same time
making it resilient to communication failure and hardware malfunctioning of the drone. Each drone is
assumed to have its own memory which it updates after every step of exploration. In our testing scenario,
the obstacles are randomly distributed in the bounded exploring area and each drone is assumed to correctly
locate itself on the map. We have divided the complete environment into either square cells or hexagonal
cells. Moreover, each drone is equipped to communicate and share information with companion drones.
In case of the map being divided into a quadrangular grid, we assume that the drones are equipped with eight
sensors that can sense the environment in eight directions. The sensors are equidistant around the drone,
so the drone can detect local environment in eight directions: Front, Right-front, Right, Right-back, Back,
Left-Back, Left, and Left-front. The drones can move only in four directions: Front, Right, Back, Left. In
case of the map being divided into hexagonal cells, we assume that the drones are equipped with six sensors
which can sense the environment. The hexagonal grid allows the drones to move in six directions around it.
Each drone stores the environment in its memory in a cell based map. The cell based map is stored in a [n x
m] matrix for both hexagonal and square grid. Each cell in the map is either unexplored, obstacle, visited
or it is a frontier cell. The frontier cell is the cell which has been sensed by drone, but it has not yet been
visited. The whole environment is divided into subareas either in rectangular regions composed of square
cells in it or in hexagonal subareas composed of hexagonal cells in it.
The drones are initially placed randomly into cells. The goal is to explore all the cells in the area at least
once. The drones can communicate with companion drones and co-ordinate to reduce the time to cover
the complete area. The communication allows the robots to exchange their local maps with the companion
drones.
3 MULTI-ROBOT AREA EXPLORATION ALGORITHM
We present the algorithm in four steps where at each step we modify the algorithm in such a way such that
the efficiency is progressively improved.
Datta, Tallamraju, and Karlapalem
In all the approaches we divide our algorithm into two stages. Initially, the drones are randomly distributed
among the subareas. In the first part, Exploration state the drones explore their own local subareas and
later in the second part, Moving State the drones move towards the unexplored cells in the different subarea
determined by RNN or PSO model. Each drone runs its own algorithm and maintains exploration or moving
state. It is done in such a way that the total time to explore the complete area is minimized.
3.1 Quadrangular Reverse Nearest Neigbour (QRNN)
3.1.1 Exploring local subarea (Exploration state)
In the exploration state, the drones keep moving towards the frontier cells recording the terrain of the nearby
cells until there are no frontier cells left in the same subarea.
Initially, the cell where a drone is located is marked visited and the neighboring cells are marked frontier
cells or obstacle cells in the local memory of each drone. After each time step, the drones move to the nearby
frontier cell marking it covered and marking the new neighbor cells as frontier cells. The drone selects the
frontier cell to move in this order North, South, East and, West. If there are no more neighboring frontier
cells, then the drones move towards the nearest frontier cell. This goes on until there are no more frontier
cells left in the subarea. The complete details are written in Algorithm1.
To find the nearest frontier cell in a grid with obstacles in between, we use A* search to find the nearest
frontier cell.
The function of the A* search is given by:
f (i, j) = g(i, j) + h(i, j)
(1)
Mark the current cell as a visited cell.
Mark the neighbor unexplored cells in the same subarea as frontier cells or obstacles.
if there are frontier cells around then
In equation (1), g(i, j) is the distance along the shortest path that connects the two cells 'i' and 'j'. Here
h(i, j) is the displacement or crow fly distance between the two cells as explained by Nosrati, Karimi, and
Hasanvand (2012). Algorithm 1 explains the Exploration State.
Algorithm 1 Subarea Coverage
1: procedure SUB -- AREA COVERAGE
2:
3:
4:
5:
6:
7:
8:
9:
10:
11: end procedure
else Moving state
end if
go to 2
go to one of them
else if frontier cells in the same subarea present then
go to nearest frontier cell along the shortest path
3.1.2 RNN based co-ordination (Moving State)
In this state, the drones have to move towards an appropriate subarea. The drones should move towards a
subarea ensuring that a minimal number of companion drones are approaching the same subarea.
Datta, Tallamraju, and Karlapalem
Previous papers use PSO based cooperation model to avoid multiple drones heading towards the same cell.
The PSO model uses the velocity which is updated according to previous best positions and the global best
position achieved by its neighbors. The global best position is defined as the cell with a minimum number
of other drones in its direction.
The disadvantage of the PSO model and can be explained using Figure 1.
Figure 1: Quadrangular Grid of size 9X9 with 9 subareas.
In the above quadrangular grid, D1 and D2 represent the drones, S1 and S2 are unexplored subareas and U1,
U2 are the unit cells of subarea S1. Considering a case where only subareas S1 and S2 have unexplored cells
left and rest all area is explored. If we follow the PSO model, then both drones D1 and D2 will move towards
S1, as unexplored cells U1, U2 are more closer to them than the cells of subarea S2 but ideally, since there is
nobody to move to unexplored region S2, the D1 should move towards S2 while the drone D2 should move
towards S1.
To overcome this problem, we use the concept of reverse nearest neighbor. The intuition behind this concept
is to keep a track of the number of drones approaching that subarea. If a greater number of drones are
approaching that subarea then the cost to travel to that subarea should be increased. RNN outperforms the
traditional PSO model as it keeps a track of which drones are approaching which subarea instead of keeping
track of drones that are approaching a particular cell. We call it reverse nearest neighbor because instead of
focusing on the nearest area to a drone we focus on the nearest drone to a subarea.
The cost for the Drone Di at i to travel to a cell Uj at j is calculated as shown in (2):
Cost(Di,Uj) = g(i, j) + h(i, j) + α ∗ S(Uj)
(2)
Here, S(Uj) is the number of drones that were approaching the subarea in which Uj is located. α is the
constant just to scale this factor. Functions 'g' and 'h' are part of A* search as explained earlier.
Using RNN, we can see that initially, both the drones will move towards the nearby unexplored region S1
but after a time step, the cost in RNN for D1 to move towards S1 will increase thereby, pushing D1 towards
S2 and D2 towards S1.
Long range communication facilitates that all the drones exchange information towards which subarea they
tend to approach after each time stamp. Even if the communication fails with some far away drones, then
also the algorithm keeps working based on the previous information. Hence, this algorithm is distributed
and resilient to communication failure.
In Algorithm 2, the drone in Moving State makes a step towards selected subarea. After making a step it
comes back to Exploration State if there are frontier cells around it.
Datta, Tallamraju, and Karlapalem
Algorithm 2 RNN based coordination
1: procedure RNN BASED COORDINATION
2:
3:
Initialise, drones are not approaching any subarea in present iteration.
Communicate(Long Range) with companion drones to find which subarea they were approaching
in previous iteration.
Calculate the cost Cost(Di,Uj) to approach each cell based on number of drones S(Uj) moving
towards subarea of Uj in previous iteration.
Move towards the cell with minimum cost.
Update which subarea drone is approaching.
go to Exploration State (Algorithm 1)
4:
5:
6:
7:
8: end procedure
3.2 QRNN with Short Range Communication (QRNN-SRC)
This approach is similar to the previous approach, showing and improving upon the demerits of QRNN. In
QRNN during the exploration state, it is possible to have multiple drones in the same subarea. This is more
likely if the number of drones exploring the area is high. In QRNN, communication was only taking place
in the movement state. The drones might follow one another to explore their own subarea. As all the drones
follow the order of direction to move that is North, South, East and, West, the drones might end up covering
a lot of cells multiple times.
As shown in Figure 2 if the drones are in the same subarea, then one drone might follow the path of a
companion drone as the other drone is unaware of what the first drone has already explored.
Figure 2: Quadrangular Grid.
Therefore, drones within a subarea need to communicate more frequently. The short range communication
amongst the drones in smaller proximity helps avoid obstacles and re-exploration of cells already explored
by companion drones located in the same subarea. The short range communication can also be multi-hop
communication between the nearby drones leading to cascading information exchange.
The long range communication like in QRNN helps in deciding optimal next sub area to be targeted by
individual drones based on RNN (Reverse nearest neighbor).
Datta, Tallamraju, and Karlapalem
3.3 Hexagonal RNN with SRC (HRNN-SRC)
One of the most common and simplest representations of the area to be explored is the quadrangular grid
map. In a quadrangular grid, the diagonal distance between two squares is bigger than the horizontal or
vertical distance between two squares. Thus, it is harder to manage distances when drones move in diagonal,
vertical and horizontal ways.
To simplify that problem, drones are frequently restricted to only make vertical or horizontal moves when
quadrangular map representation is used. In a hexagonal grid, diagonal and vertical distances between two
hexagons are the same. Thus, there is no need to worry about managing different distances scales or restrict
drones to only make vertical or horizontal moves. Therefore, in this approach, we use the hexagonal cells
and hexagonal subareas to represent the area to be explored.
The hexagonal approach gives drones six directions to move. Later simulation results show that there is an
improvement in the efficiency using the hexagonal grid cells.
3.4 Weighted RNN-SRC (WHRNN-SRC)
In this approach, we improve upon the suggested RNN method in QRNN-SRC.
In the scenario, as shown in Figure 3, if there are drones D1−4 and there are two subareas S1 and S2. More
drones are approaching subarea 1, therefore using Reverse Nearest Neighbor D4 will move towards S2. But,
if S2 has only one cell left to be unexplored and the complete S1 is unexplored then D4 should move towards
subarea 1. Figure 4 explains the deadlock scenario which is explained later.
Figure 3: Drones are approaching subareas.
Figure 4: Deadlock Scenario.
To address the above demerit, we propose a weighted RNN algorithm, wherein each subarea has a weight
proportional to the number of cells presently unexplored in it. Now, the cost for the drone Di to travel to a
cell Uj is calculated as shown in formula (3):
Cost(Di,Uj) = g(i, j) + h(i, j) + α ∗ S(Uj)− β ∗
WS j
subarea.height ∗ subarea.width
(3)
Here, WS j represents the number of cells unexplored in the subarea containing cell Uj in the previous iter-
ation. The above formula penalizes agents more if they move towards a subarea where drones are already
approaching by increasing cost. It also reduces the penalty if they move towards a completely unexplored
subarea. β is the constant which is multiplied by the percentage of unexplored cells in that subarea.
The weighted RNN algorithm improves the efficiency keeping the system resilient to failure of communica-
tion. The efficiency is defined as minimizing number of times the cells are covered.
Datta, Tallamraju, and Karlapalem
In this algorithm, we can see that towards the end of the simulation when only a few unexplored cells are left
then, all the drones move towards few remaining unexplored cells unnecessarily. Hence, towards the end
when the number of unexplored cells is comparable to the number of drones, we apply the bidding algorithm
to further optimize the results.
Since the RNN works at every time step, many times this can result in deadlocks.
In the scenario, as shown in Figure 4, initially the drones D3 and D4 move towards unexplored region U1.
In the next time step, since both of the drones proceed towards U1. Now, using RNN the cost to the U1
increases as two drones are trying to approach it. So, both the drones start moving towards unexplored area
U2. This way both the drones get stuck in the cycle causing a deadlock. To avoid this problem, we have used
a timestamp mechanism, wherein the decision to move is delayed by an iteration. This way the deadlocks
are also handled.
4 EXPERIMENTAL RESULTS
We have tested our proposed approaches and compared them with the traditional PSO model. The experi-
ments were simulated with a large number of grid cells (Max 10,000) and with a maximum of 200 drones.
In the simulation experiment, we compared our approaches with the traditional PSO based model.
The various parameters for the simulation experiment are shown in Table 1:
Table 1: Parameters of the Simulation Experiment.
Parameters Values
Grid-size{height, width}
Size of Sub-area taken
Number of Agents
Number of Obstacles
Radius of short range communication (Euclidean Distance)
40x40, 60x60, 80x80, 100x100
{4x4},{5x5},{8x8},{10x10}
5, 10, 25, 50, 100, 150, 200
20%
5, 10, 15 , 20
Constant 'α' which gives weight to RNN 2-10
The traditional PSO based model, QRNN and QRNN-SRC are simulated on the quadrangular grid as shown
in Figure 5. The HRNN-SRC and WHRNN-SRC are simulated on the hexagonal grid as shown in Figure 6.
Figure 5: Quadrangular grid.
Figure 6: Hexagonal Grid.
Datta, Tallamraju, and Karlapalem
In Figure 5 and Figure 6, the cells outlined in black color represents the obstacles, visited cells are outlined
in red color and the unvisited cells are outlined in green color. The simulation environment shown above
was created in SWIFT using OpenGL for visualization. The number inside the cell represents how many
times that cell has been explored.
Figure 7 compares the performance of PSO, QRNN-SRC, HRNN-SRC and, WHRNN-SRC, with the in-
crease in drones keeping the area to be explored constant. Here the grid size is 1600(40 x 40) cells. It
is clear from the figure that, for Weighted RNN, with increasing drones, the average number of cells re-
explored is much less compared to the traditional PSO model. This proves that the algorithm is efficient and
scalable.
Figure 7: Number of drones vs Average Re-explored Cells.
Figure 8 represents the performance of Weighted RNN vs traditional PSO model, with an increase in grid
size keeping the number of drones as constant. Here the approaches are compared taking 50 drones with an
increase in the grid size. In this case also we can see that the weighted RNN outperforms the PSO based
model. Our solution performs better not only in terms of the average number of times a cell is visited but
also it has much less variance.
Figure 8: Grid Size vs Average Re-Explored Cells.
Datta, Tallamraju, and Karlapalem
The above box plots are generated with extensive grid search on the parameters as shown in Table 1. For
each combination of number of drones and size of grid, the simulation environment is created 20 times
with 20% random obstacles. For each of the simulation environment grid search on parameters as shown in
Table 1 is performed. For each combination of algorithm type, number of drones and size of grid the best
parameters achieved will be different. For instance, with an increase in the number of drones, the value of
'α' should increase to spread the drones farther away.
For all the experiments, parameters which give the best result after applying grid search are used.
4.1 EVALUATION OF FAILURES
In real time, it is very probable that failures might occur in the exploration process which can lead to massive
delays in task completion.
There are two kinds of failures that can occur: 1. Temporary Failures: Drones are affected by temporary
failures due to weather and wind conditions, terrain features, communication latency, etc. The system
can recover from such failures by incorporating a robust control algorithm into the drones 2. Permanent
Failures in Drones: Permanent failures in communication between drones can lead to massive task delays
or permanent sensor failures can lead to partial exploration and mapping or motor and controller failures
can lead to odometry/state estimation errors and accidents causing mid-air drone collisions.
Communication failures are usually simulated using MTBF (Mean Time Between Failures). At the time 't'
the reliability of communication system R(t) for a drone is given by (4):
−
t
MT BF
R(t) = e
(4)
Failure in communication leads drones to re-exploration of the areas already explored by the companion
drones. The plot as shown in Figure 9 is simulated taking MTBF of 3 for all drones.
Figure 9: Number of drones vs Avg Re-explored Cells with MTBF=3.
After a lapse in communication as soon as the communication starts working the drone gets the latest in-
formation regarding the area explored by the companion drones. The results above show that our algorithm
outperforms the traditional PSO even in case of communication failures.
Datta, Tallamraju, and Karlapalem
5
SUMMARY
In this paper, we improved upon the existing PSO algorithm for coordinating a swarm of drones while they
are exploring an unknown environment. The goal of the paper is to reduce the redundancy and achieve
complete map coverage. Our proposed algorithm is scalable and outperforms even in case of failure in
communication among drones. Using two kinds of communication on a hexagonal grid with the weighted
RNN algorithm improved the efficiency drastically. In the future, the same approach can also be applied to
the continuous region rather than the discrete cells taken presently.
REFERENCES
Al-Obaidy, M., and A. Ayesh. 2008. "Optimizing autonomous mobile sensors network using PSO algo-
rithms". In Computer Engineering & Systems, 2008. ICCES 2008. International Conference on, pp.
199 -- 203. IEEE.
Amigoni, F., N. Basilico, and A. Q. Li. 2013. "How much worth is coordination of mobile robots for explo-
ration in search and rescue?". In RoboCup 2012: Robot Soccer World Cup XVI, pp. 106 -- 117. Springer.
Berhault, M., H. Huang, P. Keskinocak, S. Koenig, W. Elmaghraby, P. Griffin, and A. Kleywegt. 2003.
"Robot exploration with combinatorial auctions". In Intelligent Robots and Systems, 2003.(IROS 2003).
Proceedings. 2003 IEEE/RSJ International Conference on, Volume 2, pp. 1957 -- 1962. IEEE.
Burgard, W., M. Moors, D. Fox, R. Simmons, and S. Thrun. 2000. "Collaborative multi-robot exploration".
In Robotics and Automation, 2000. Proceedings. ICRA'00. IEEE International Conference on, Vol-
ume 1, pp. 476 -- 481. IEEE.
Burgard, W., M. Moors, C. Stachniss, and F. E. Schneider. 2005. "Coordinated multi-robot exploration".
IEEE Transactions on robotics vol. 21 (3), pp. 376 -- 386.
Dasgupta, P., K. Cheng, and L. Fan. 2009. "Flocking-based distributed terrain coverage with dynamically-
formed teams of mobile mini-robots". In Swarm Intelligence Symposium, 2009. SIS'09. IEEE, pp. 96 --
103. IEEE.
Dornhege, C., and A. Kleiner. 2013. "A frontier-void-based approach for autonomous exploration in 3d".
Advanced Robotics vol. 27 (6), pp. 459 -- 468.
Fu, J. G. M., T. Bandyopadhyay, and M. H. Ang. 2009. "Local Voronoi decomposition for multi-agent task
allocation". In 2009 IEEE International Conference on Robotics and Automation, pp. 1935 -- 1940. IEEE.
Kennedy, J and Eberhart, R 1995. "Particle swarm optimization, IEEE International of first Conference on
Neural Networks".
Koenig, S., and Y. Liu. 2001. "Terrain coverage with ant robots: a simulation study". In Proceedings of the
fifth international conference on Autonomous agents, pp. 600 -- 607. ACM.
Mahdoui, N., V. Frémont, and E. Natalizio. 2018. "Cooperative Frontier-Based Exploration Strategy for
Multi-Robot System". In 2018 13th Annual Conference on System of Systems Engineering (SoSE), pp.
203 -- 210. IEEE.
Matignon, L., L. Jeanpierre, and A.-I. Mouaddib. 2012. "Distributed value functions for multi-robot explo-
ration". In Robotics and Automation (ICRA), 2012 IEEE International Conference on, pp. 1544 -- 1550.
IEEE.
Middleton, L., and J. Sivaswamy. 2006. Hexagonal image processing: A practical approach. Springer Sci-
ence & Business Media.
Nieto-Granda, C., J. G. Rogers, and H. I. Christensen. 2014. "Coordination strategies for multi-robot explo-
ration and mapping". The International Journal of Robotics Research, pp. 0278364913515309.
Datta, Tallamraju, and Karlapalem
Nosrati, M., R. Karimi, and H. A. Hasanvand. 2012. "Investigation of the*(star) search algorithms: Charac-
teristics, methods and approaches". World Applied Programming vol. 2 (4), pp. 251 -- 256.
Rogers III, J. G., C. Nieto-Granda, and H. I. Christensen. 2013. "Coordination strategies for multi-robot
exploration and mapping". In Experimental Robotics, pp. 231 -- 243. Springer.
Senthilkumar, K., and K. Bharadwaj. 2012. "Multi-robot exploration and terrain coverage in an unknown
environment". Robotics and Autonomous Systems vol. 60 (1), pp. 123 -- 132.
Sheng, W., Q. Yang, J. Tan, and N. Xi. 2006. "Distributed multi-robot coordination in area exploration".
Robotics and Autonomous Systems vol. 54 (12), pp. 945 -- 955.
Simmons, R., D. Apfelbaum, W. Burgard, D. Fox, M. Moors, S. Thrun, and H. Younes. 2000. "Coordination
for multi-robot exploration and mapping". In AAAI/IAAI, pp. 852 -- 858.
Singh, K., and K. Fujimura. 1993. "Map making by cooperating mobile robots". In Robotics and Automa-
tion, 1993. Proceedings., 1993 IEEE International Conference on, pp. 254 -- 259. IEEE.
Wang, Y., A. Liang, and H. Guan. 2011. "Frontier-based multi-robot map exploration using particle swarm
optimization". In Swarm Intelligence (SIS), 2011 IEEE Symposium on, pp. 1 -- 6. IEEE.
Wurm, K. M., C. Stachniss, and W. Burgard. 2008. "Coordinated multi-robot exploration using a segmenta-
tion of the environment". In Intelligent Robots and Systems, 2008. IROS 2008. IEEE/RSJ International
Conference on, pp. 1160 -- 1165. IEEE.
Yamauchi, B. 1997. "A frontier-based approach for autonomous exploration". In Computational Intelligence
in Robotics and Automation, 1997. CIRA'97., Proceedings., 1997 IEEE International Symposium on,
pp. 146 -- 151. IEEE.
Yoder, L., and S. Scherer. 2016. "Autonomous exploration for infrastructure modeling with a micro aerial
vehicle". In Field and service robotics, pp. 427 -- 440. Springer.
Zlot, R., A. Stentz, M. B. Dias, and S. Thayer. 2002. "Multi-robot exploration controlled by a market
economy".
AUTHOR BIOGRAPHIES
AYUSH DATTA is a research student of "Agents and Applied Robotics Group" at "International Institute
of Information Technology, Hyderabad". His interests lie in Robotics, Simulation and Path Planning. His
email address is [email protected].
RAHUL TALLAMRAJU is a Ph.D. student of "Agents and Applied Robotics Group" at "International
Institute of Information Technology, Hyderabad". His interests lie in Multi-Robot Systems, Multi-Agent
Motion Planning. His email address is [email protected].
KAMALAKAR KARLAPALEM is a Full Professor at "International Institute of Information Technology,
Hyderabad". He is the head of Data Sciences and Analytics Centre and Agents and Applied Robotics Group.
He holds a Ph.D. from (Georgia Institute of Technology, USA), His research interests include Database
system, data visualization, data analytics, multi agent systems, workflows and electronic contracts. His
email address is [email protected].
|
cs/0108008 | 1 | 0108 | 2001-08-14T14:51:23 | Using Methods of Declarative Logic Programming for Intelligent Information Agents | [
"cs.MA",
"cs.AI"
] | The search for information on the web is faced with several problems, which arise on the one hand from the vast number of available sources, and on the other hand from their heterogeneity. A promising approach is the use of multi-agent systems of information agents, which cooperatively solve advanced information-retrieval problems. This requires capabilities to address complex tasks, such as search and assessment of sources, query planning, information merging and fusion, dealing with incomplete information, and handling of inconsistency. In this paper, our interest is in the role which some methods from the field of declarative logic programming can play in the realization of reasoning capabilities for information agents. In particular, we are interested in how they can be used and further developed for the specific needs of this application domain. We review some existing systems and current projects, which address information-integration problems. We then focus on declarative knowledge-representation methods, and review and evaluate approaches from logic programming and nonmonotonic reasoning for information agents. We discuss advantages and drawbacks, and point out possible extensions and open issues. | cs.MA | cs | Under consideration for publication in Theory and Practice of Logic Programming
1
Using Methods of Declarative Logic
Programming for Intel ligent Information Agents
THOMAS EITER, MICHAEL FINK,
GIULIANA SABBATINI, HANS TOMPITS
Technische Universitat Wien,
Institut fur Informationssysteme,
Abt. Wissensbasierte Systeme 184/3,
Favoritenstrasse 9-11, A-1040 Vienna, Austria
(e-mail: [eiter,michael,giuliana,tompits]@kr.tuwien.ac.at)
Abstract
At present, the search for specific information on the World Wide Web is faced with
several problems, which arise on the one hand from the vast number of information sources
available, and on the other hand from their intrinsic heterogeneity, since standards are
missing. A promising approach for solving the complex problems emerging in this context
is the use of multi-agent systems of information agents, which cooperatively solve advanced
information-retrieval problems. This requires advanced capabilities to address complex
tasks, such as search and assessment of information sources, query planning, information
merging and fusion, dealing with incomplete information, and handling of inconsistency.
In this paper, our interest lies in the role which some methods from the field of declara-
tive logic programming can play in the realization of reasoning capabilities for information
agents. In particular, we are interested to see in how they can be used, extended, and fur-
ther developed for the specific needs of this application domain. We review some existing
systems and current pro jects, which typically address information-integration problems.
We then focus on declarative knowledge-representation methods, and review and eval-
uate approaches and methods from logic programming and nonmonotonic reasoning for
information agents. We discuss advantages and drawbacks, and point out the possible
extensions and open issues.
Contents
1 Introduction
2 Intelligent Information Agents
3 Problems and Challenges
4 Systems and Frameworks
4.1 Cohen’s Information System for Structured Collections of Text
4.2
Information Manifold
4.3 Carnot
InfoSleuth
4.4
4.5
Infomaster
4.6 COIN
3
4
8
9
10
11
11
12
12
13
2
Eiter et al.
4.7 HERMES
4.8
IMPACT
4.9 SIMS
4.10 TSIMMIS
4.11 Further Pro jects
5 Declarative Methods
5.1 Overview
5.2 Short Review of Logic Programs and Answer Sets
6 Preference Handling
6.1 Prioritized Defaults by Gelfond and Son
6.2 Preferred Answer Sets by Brewka and Eiter
6.3 Prioritized Logic Programs by Inoue and Sakama
6.4 Prioritized Logic Programs by Foo and Zhang
6.5 Compiled Preferences by Delgrande et al.
6.6 Disjunctive Programs with Inheritance by Buccafurri et al.
6.7 Preference Default Theories by Brewka
6.8 Ordered Logic Programs
6.9 Further Approaches
7 Revision and Update
7.1 Revision Programs by Marek and Truszczy´nski
7.2 Update Rules as Logic Programs by Pereira et al.
7.3 Abductive Updates by Inoue and Sakama
7.4 Updates by Means of PLPs by Foo and Zhang
7.5 Dynamic Logic Programming by Alferes et al.
7.6 Updates and Preferences by Alferes and Pereira
7.7
Inheritance Programs and Updates
7.8 Revision of Preference Default Theories by Brewka
7.9 Arbitration
8 Quantitative Information
8.1 Disjunctive Programs with Weak Constraints by Buccafurri et al.
8.2 Weight Constraint Rules by Niemela et al.
8.3 Weighted Logic Programming by Marek and Truszczy´nski
8.4 Probabilistic Programs by Subrahmanian et al.
9 Temporal Reasoning
9.1 Reasoning about Actions
9.2 Temporal Logics for BDI agents
9.3 LUPS, a Language for Specifying Updates
10 Evaluation
10.1 Preference Handling
10.2 Logic Programs with Quantitative Information
10.3 Revision and Update
10.4 Temporal Reasoning
11 Conclusion
References
13
13
14
14
15
17
17
19
21
22
23
23
25
26
26
27
28
28
29
31
31
32
34
35
36
38
39
40
41
42
42
42
43
43
44
45
45
45
46
49
50
54
55
57
Using Methods of Declarative LP for Intel ligent Information Agents
3
1 Introduction
Currently, a user who is looking for some desired information on the World Wide
Web faces several problems, which arise from the vast amount of available infor-
mation sources, and from their heterogeneity, since standards are missing. First
of all, the user has to identify the relevant sources which should be queried, since
bounds on resources, time, and/or cost of services usually do not permit to query
al l sources which are available. The generation of a query plan is often quite ex-
pensive and cannot be optimal when the user has no additional information about
the knowledge contained in each source, but merely a short description of it. Then,
once the user has made up his or her plan, for each source the query must be formu-
lated in an appropriate way, depending on the interfaces available, as well as on the
data organization and presentation. Furthermore, the user must possibly adapt the
query more than once for each source in order to get the desired information, and
must learn several different query approaches. Furthermore, it may happen that
some sources provide as a result only partial or incomplete information to a query.
The user has then to merge all data retrieved, taking into account that the infor-
mation provided by different sources may be inconsistent. In such a situation, the
user needs some notion of reliability for the sources in order to choose the proper
information.
A promising approach for solving the complex problem outlined above is the
use of multi-agent systems for accessing several heterogeneous information sources.
The user is presented a uniform interface for accessing all available services and
information sources, without having to bother with the heterogeneity underneath.
It is the system as a whole which takes care of searching the appropriate sources,
accessing them, and returning to the user the required information, and this to an
extent as complete and consistent as possible.
The realization of such systems requires special functionalities and capabilities,
which emerge from specialized tasks like the ones outlined above. Such capabilities
are provided as services on request by various kinds of information agents, which
form a society of agents for cooperatively solving complex information-retrieval
problems.
In this paper, we discuss some problems arising for the successful realization of
information agents in a multi-agent environment, namely:
• search and assessment of information sources;
• query planning;
• information merging and fusion;
• dealing with incomplete information; and
• handling of inconsistency.
It is not hard to imagine that an advanced approach to any of these problems
must involve, in some form, logical reasoning tasks, based on ad hoc knowledge
about the task in question and on background knowledge of the domain, suitably
represented in a knowledge base.
A number of different models and methods for knowledge representation and rea-
soning have been developed in the past, which may be used for this purpose. In this
4
Eiter et al.
paper, our interest is in the role which some methods from the field of “pure” logic
programming, and in particular answer set semantics and its derivatives, can play
in the realization of reasoning capabilities for information agents. These methods
lack richer algorithmic control structures as those typical of Prolog, and permit
knowledge representation in a purely declarative way, without the need for paying
much attention to syntactical matters such as the way in which rules are listed in
a program or literals occur in the body of a rule. This seems to be well-suited in
the context of rule-based decision making, where semantical aspects for a “ratio-
nal” decision component are the main concern rather than the control structure. In
fact, the rational decision component will be embedded into some control structure,
which calls the component.
The main focus of this paper is to discuss how the methods from pure (or “declar-
ative”) logic programming can be used, extended, and further developed for the
specific needs of information agents. As we shall see, the methods which are avail-
able to date do not meet all the needs which are required, which leaves the way
open for interesting research.
The paper is organized as follows. Section 2 gives a detailed description of the role
of information agents in the field of advanced information access, and discusses the
architecture of a prototypical agent system. In Section 3, we identify a set of rea-
soning capabilities which are mandatory for an advanced information-integration
system, some of which are critical for an “intelligent” information agent. Section 4
contains a brief review of some existing systems and current pro jects which address
typical information integration problems, most of them by means of ad hoc pro-
cedural techniques. Although we are more interested in declarative methods, the
different approaches implemented in these systems provide relevant information
concerning possible solutions and hint to intrinsic difficulties.
We then focus on declarative knowledge representation methods. First of all,
Section 5 discusses some subtasks amenable for information agents, which promise
to have a successful solution based on declarative methods. In Sections 6–9, we
then review some approaches and methodologies from the field of declarative logic
programming and nonmonotonic reasoning, which are likely to be of particular
interest for one or more of the subtasks. In Section 10, we evaluate the presented
methods, taking into account their applicability in the field of information agents.
We discuss advantages and drawbacks, and address the possibility of extensions.
Section 11 concludes the paper, containing some open issues which emerge from
this analysis, and outlines directions for future research.
2 Intelligent Information Agents
Several models of multi-agent systems have been developed during the last years,
together with a number of successful applications, mainly in the fields of hetero-
geneous information and software integration, electronic commerce, and Web in-
terfaces. In this paper, we focus our attention on know ledge-based approaches for
developing reasoning components for intelligent information access (for an extensive
overview of agent programming approaches, cf. (Sadri & Toni, 2000)). In particu-
Using Methods of Declarative LP for Intel ligent Information Agents
5
lar, we are interested in information agents (sometimes also called midd le agents
(Decker et al., 1997)) which are involved in the following tasks:
• finding, selecting, and querying relevant sources;
• managing and possibly processing the retrieved information; and
• updating their own knowledge about other agent services and features.
In this context, different scenarios—which to some extent overlap—have been
suggested and implemented. Following Flores-Mendez (1999), these can be classified
as follows:
Facilitators: Agents which take control over a set of subordinated agents and
coordinate the services they offer and the use of the resources they require.
Brokers: Agents often used for matching between a set of different data sources
and user requests. Brokers receive requests, look for relevant sources matching
these requests, and then perform actions using services from other agents (com-
bining them with their own resources or information).
Mediators: In the mediator approach (Wiederhold, 1992), meta-knowledge about
a number of other agents (sometimes called provider agents ) is available to the
mediator, which exploits this knowledge to create higher-level services (not pro-
vided by the underlying agents) for user applications. These new services result
by the combination and merging of low-level services on the basis of the compre-
hensive meta-information which the mediator has. In a sense, mediators may be
seen as enhanced, high-level brokers.
Yellow Pages: A yellow pages dictionary helps users and other agents in finding
the right providers (other agents or information sources) for the kind of service
they need, possibly performing a match between advertised services, an ontology
of the domain in question, and service requests.
Blackboards: A blackboard serves as a temporal repository for service requests
which remain to be processed. Agents offering some service can access the black-
board and look for (and retrieve from it) service requests which they can satisfy.
Note that in a multi-agent system several of these agents with different features
may be available at the same time, performing different tasks.
We focus our attention on information agents which act as mediators; in par-
ticular, we are interested in the rational capabilities of such agents. For this pur-
pose, we assume that the agent has knowledge about the application domain of
interest, together with some meta-knowledge about the contents and features of
the distributed, heterogeneous information sources the system has access to. This
knowledge is represented in the form of declarative rules, determining the beliefs
and actions of the agent. For example, in an application where bibliographic refer-
ences should be found, the agent must have some knowledge about the structure
of references, the information fields which are necessary to identify a work, and so
on, together with further knowledge about the information sources which it might
consult, such as their reliability and availability, what kind of works they offer, and
others.
For the goal of satisfying a user request, an information agent of the above kind
may have to solve various subgoals:
6
Eiter et al.
• identifying and possibly ranking relevant information sources (e.g., selection
of Web pages);
• retrieving the required information (or ask some appropriate provider agent
for it);
• processing the information returned by the sources by combining, merging,
and integrating them (i.e., if more than one source has been queried);
• optimizing the number of accessed sources, or the total cost and time required
for the search, or the precision and completeness of the results.
For executing these tasks, the agent has to make decisions about what actions
should be performed or how to proceed at certain points. An “intelligent” agent
makes these decisions based on reasoning methods using its knowledge about the
application domain and the particular requests to be answered. Rather than hav-
ing the decision mechanism implicitly hard-coded into the agent’s program, we
shall assume that the agent’s decision-making procedure is strictly modularized
and attached to a thinking component. Such a thinking component can be realized
in many ways, by using one of the numerous approaches that have been devel-
oped in the agent and AI literature. Here, we restrict our attention to the use
of non-Prolog logic programming (mostly based on the answer set semantics) for
this purpose. It is important to point out, however, that we suggest the use of
declarative logic programming for realizing some rational components of the agent,
whereas other programming languages and environments may be more suitable for
implementing other components, like control and communication (infra-)structures
(cf. Section 4.11.2).
As for an agent-based information system, different architectures can be envis-
aged, in which information agents have different capabilities and roles. In this paper,
we assume the following prototypical system architecture, as illustrated in Figure 1:
• We have a multi-agent system providing access to several heterogeneous in-
formation sources. In such a system, the existence of an interface agent is
assumed, which offers an homogeneous user interface, collects user requests
(typically from the Web), and formulates them in a language suitable for sub-
sequent processing. The user may be a person accessing a graphical interface,
or may be some software.
• For each information source a specific provider agent takes care of access to
that source (or a single provider agent manages the accesses to all sources
of a given type; this is not relevant at this point). It receives a request in a
standard communication language, e.g., KQML (Finin et al., 1994)—ensuring
homogeneity and implementation transparency—and accesses the underlying
source as needed for retrieving the desired information. This access can consist
of a query to a relational database in a standard query language, of an ex-
tensive search over some files, of a retrieval from databases containing special
data (e.g., geographical data related to a given address), of a more complex
search over the content of a set of Web pages, and so on. The provider agent
composes then the information thus obtained into an answer in the standard
Using Methods of Declarative LP for Intel ligent Information Agents
7
User 1
User 2
User 3
... ...
U
S
E
R
I
N
T
E
R
F
A
C
E
I
N
T
E
R
F
A
C
E
A
G
E
N
T
Figure 1. System Architecture
Information
Agent
Thinking
Component
Domain
Knowledge
Meta-
Knowledge
about
Sources
Provider
Agent 1
Source 1
Provider
Agent 2
Source 2
... ...
... ...
communication language and passes it to the interface agent, which in turn
informs the user in a suitable way.
Information sources and the corresponding provider agents may of course not be
part of the system itself; the only requirement is the possibility of communicating
with them in a standard language.
In such a context, (sets of ) information agents act between interface and provider
agents, in order to determine which sources are to be accessed and in which order,
based on the user query, and to generate the appropriate answer based on the
answers of each accessed source, maybe resolving conflicts or handling incomplete
information. Meta-knowledge about available sources and the application domain,
as well as advanced reasoning and planning capabilities, are fundamental for this
purpose.
An example of a multi-agent environment into which such a model of an infor-
mation agent fits into is IMPACT, the Interactive Maryland Platform for Agents
Collaborating Together (Subrahmanian et al., 2000), which we briefly discuss in
Section 4.8. Note, however, that for the purposes of this paper, the particular ar-
chitecture of an agent-based information system is not a primary issue. Rather, we
are dealing with certain capabilities that should somehow be realized in the system;
for us, it is less important where exactly this capability resides.
8
Eiter et al.
3 Problems and Challenges
In a multi-agent system for intelligent information access, information agents are
supposed to provide advanced capabilities which help in improving the quality of
query results. The following is a list of such capabilities, whose features partially
overlap:
• Decompose a request in its “atomic” parts on the basis of available knowl-
edge about the information sources, and reformulate it according to the corre-
sponding data structure. E.g., if the user looks for a restaurant and wants to
see where it is located, the query can be decomposed into two parts: (i) search
for a restaurant satisfying certain criteria, and (ii) search for a picture show-
ing a map of the corresponding location. This decomposition task can be of
varying complexity and can generate a set of possible decompositions, based
on the information about the underlying sources and the domain of inter-
est (Levy et al., 1996a; Levy et al., 1996b; Levy et al., 1995; Cannata et al.,
1997; Huhns & Singh, 1992; Bargmeyer et al., 1999; Bayardo et al., 1997;
Bohrer et al., 1999; Arens et al., 1993; Arens et al., 1996; Garcia-Molina
et al., 1995; Garcia-Molina et al., 1997; Quass et al., 1995; Abiteboul et al.,
1997a; Abiteboul et al., 1997b; Adali et al., 1995; Adali & Emery, 1995).
• Integrate the query with additional user information, if available. This
could require asking first a profiling agent for the user profile, habits, rights
and interests. E.g., if a user is looking for Web pages about a given topic, and
the system knows that the user does not speak English, then pages containing
text in his/her native language should be preferred. This is useful not only
for human users but for servicing other software as well. E.g., it is important
to know if the user software poses requirements over the data format to be
retrieved, if it has the possibility of directly communicating with the selected
information source, if it has specific rights, and so on.
• Select the information sources to be queried, using the meta-knowledge
about them and about the application domain to determine which of the
sources contain relevant data to answer the query (Levy et al., 1996a; Levy
et al., 1996b; Levy et al., 1995; Cannata et al., 1997; Huhns & Singh, 1992;
Bargmeyer et al., 1999; Bayardo et al., 1997; Bohrer et al., 1999; Adali et al.,
1995; Adali & Emery, 1995; Arens et al., 1993; Arens et al., 1996; Garcia-
Molina et al., 1995; Garcia-Molina et al., 1997; Quass et al., 1995). If possible,
determine further preferred information sources to be queried.
• Create a query plan. Determine in which order sub-queries are to be per-
formed (Levy et al., 1996a; Levy et al., 1996b; Levy et al., 1995; Arens et al.,
1993; Arens et al., 1996), on the basis of query decomposition, of the data
available in every source, and of preference over sources, in order to optimize
the execution. E.g., it could be required to minimize the costs or the expected
waiting time for the user, to maximize the quality of the expected results, or
to minimize the number of accessed sources and the network load, and so on.
• Execute the plan, asking the corresponding provider agents for the re-
quired services and waiting for their answers, possibly adapting the query
Using Methods of Declarative LP for Intel ligent Information Agents
9
plan dynamically to run-time information and to the answers obtained so
far (Levy et al., 1996a; Levy et al., 1996b; Arens et al., 1993; Arens et al.,
1996; Knoblock, 1995).
• Compose and merge the answers. This task can have varying complexity,
from simply collecting all the answers and passing them to the interface agent
without processing them, to organizing the retrieved data in some form (e.g.,
eliminating multiple instances of the same information, ranking the answers
after their quality or after other criteria), to merging them in a single com-
prehensive answer, and so on (Cannata et al., 1997; Huhns & Singh, 1992;
Bargmeyer et al., 1999; Bayardo et al., 1997; Adali et al., 1995; Adali &
Emery, 1995; Garcia-Molina et al., 1995; Garcia-Molina et al., 1997; Quass
et al., 1995).
• Detect and possibly remove inconsistencies among retrieved data (Can-
nata et al., 1997; Huhns & Singh, 1992; Adali et al., 1995; Adali & Emery,
1995) on the basis of an inconsistency removal strategy and of meta-knowledge
about the sources, or meta-knowledge contained in the answers themselves.
E.g., some source may return data together with a reliability or probability
value for those data.
• Integrate incomplete information by means of internal reasoning capa-
bilities (Garcia-Molina et al., 1995; Garcia-Molina et al., 1997; Quass et al.,
1995). E.g., if in a geographical database there is no information available
about a specific address, but the address is located in a certain district or
near a known place, it could be useful to search for a map of the correspond-
ing district, giving the user at least an approximative information.
• Start a learning procedure in order to improve or update internal models
and reasoning rules (Arens et al., 1993; Arens et al., 1996), perhaps providing
input to the profiling agent as well. E.g., if a certain source is classified as
providing an answer in one second, and it required for each of the last ten
queries more than ten seconds, than the classification or the reliability of that
source should be updated.
4 Systems and Frameworks
Some of the tasks for information agents identified in the previous section have
already been widely addressed and some feasible solutions have been developed.
Following the grouping suggested by Levy and Weld (2000), existing intelligent
systems (in particular Web systems) have their main focus in some of the following
fields:
• User modeling and profiling: addressing the issue of deploying adaptive
user interfaces and recommender systems, integrating user queries on the basis
of the user profile, or directly suggesting the user items of interest. We do not
address this topic explicitly in our review; hints and relevant literature can
be found in (Levy & Weld, 2000).
• Analysis and preprocessing of information sources: building up the
meta-knowledge on which reasoning and decision making is based, on the basis
10
Eiter et al.
of application domain description. As far as information agents are concerned,
we will suppose that such meta-knowledge is available.
• Information integration and information management: covering a
wide area of applications and different problems. We will focus on this global
issue from the point of outlined in Section 2.
In the following, we give an overview of some existing systems and frameworks as
well as of current pro jects, which address different topics in the field of information
integration. We briefly describe the problems they face, their approach and focus
of interest, and their key features. Note that many of these systems use ad hoc
procedural techniques, rather than declarative methods in which we are interested
in. Nonetheless, they are highly relevant, since the kind of problems attacked by
these approaches and the proposed solutions are of interest for the whole field of
heterogeneous information integration. In the following sections, we consider how
declarative formalisms (mostly based on declarative logic programming under the
answer set semantics) can help in improving some rational capabilities of infor-
mation agents. For a comprehensive overview of existing systems in the field of
information integration and of the role of logic programming in some of them,
cf. (Dimopoulos & Kakas, 2000).
The common scenario of such systems is usually as follows. There is some in-
termediary software, which performs the tasks of provider agents as described in
Section 2. The system can access several data sources and has some means of trans-
lating data in the required format for internal processing, as well as of translating
requests in the format required by the information source. Users access the system
and the available information sources through an interface which, like the interface
agent we described in Section 2, hides from the user most of the implementation
aspects and data representation issues. The user can query the system based on
the semantics of the information he or she is looking for in a given application
domain and receives from the system the required information, without having to
care about different data representation schemas, formats, or languages.
4.1 Cohen’s Information System for Structured Col lections of Text
The system of Cohen (1998b) processes movie review data extracted from Web
information sources. Text sources are converted into a highly structured collection
of fragments of text, on which automatic data processing is facilitated. Queries
over this structured collection are then answered using an ad hoc “logic” called
WHIRL (Cohen, 1998a), which combines inference with methods for ranked re-
trieval. The approximation is obtained using similarity metrics for text from infor-
mation retrieval.
The system uses a relational data model in which every element of every tuple is
assumed to consist of free text, with no additional structure. As two descriptions
of the same ob ject will not necessarily be identical, approximation of database
operations with a similarity metric is performed. The output tuples are then sorted
in decreasing order according to the similarity of corresponding fields. The method
Using Methods of Declarative LP for Intel ligent Information Agents
11
is proven to be robust to errors, e.g., in case of some field being incomplete due to
extraction errors.
The main advantage of this approach is that it requires little preprocessing work
compared to similar tools. As well, no key normalization and no complete extraction
of the documents are necessary.
4.2 Information Manifold
The Information Manifold (Levy et al., 1996a; Levy et al., 1996b; Levy et al.,
1995) provides uniform access to autonomous, heterogeneous, structured informa-
tion sources on the Web, which make use of different data models and vocabularies.
The language CARIN, due to Levy and Rousset (1996) and based on descrip-
tion logics, is used to describe the structure and properties of existing information
sources and to add new ones. The query answering algorithm guarantees that only
relevant sources are accessed, while further pruning on the basis of run-time infor-
mation is performed. Queries are expressed by means of classes and roles in CARIN,
ordinary predicates of every arity, and some specific predicates.
The user is presented with a world view consisting of a set of relations with a
class hierarchy, modeling the available information about the domain of interest.
Presented relations are only virtual: the actual data remain stored in the external
sources with their own structure. Only the descriptions are centralized and re-
organized. Queries can be formulated in terms of the world view relations, and the
system uses the available descriptions of the sources to generate a query-answering
plan (cf. (Levy et al., 1996a; Levy et al., 1996b)). The aim of the planning tool is
the minimization of the number of information sources to be accessed.
4.3 Carnot
Carnot (Huhns & Singh, 1992; Cannata et al., 1997) is a system of federated hetero-
geneous databases based on static data integration, and constitutes the foundation
for the InfoSleuth pro ject presented in the next subsection.
Many heterogeneous information models (i.e., schemas of databases, knowledge
bases and process models) are statically integrated by means of an ad hoc method,
achieving integration at the semantic level by using ontologies, which describe the
semantical structure of the application domain, to resolve possible inconsistencies.
An existing global schema is combined with the schemas of the individual sources,
comparing and merging them with the global one but not with each other. Axioms
are then produced stating the equivalence between some components of two theo-
ries, capturing the mapping between the data structure of each individual source
and the global schema. In this way, a global view of all information sources—and
at the same time local views corresponding to each source—are available.
Carnot is reported to have been successfully applied in the fields of scientific deci-
sion support over several large scientific databases, workflow automation and man-
agement from a database perspective for telecommunication applications (Huhns
& Singh, 1994), data validation and cleaning of large databases, and information
12
Eiter et al.
retrieval from structured databases in response to text queries through model inte-
gration of heterogeneous resources.
4.4 InfoSleuth
InfoSleuth (Bargmeyer et al., 1999; Bayardo et al., 1997; Bohrer et al., 1999) is
a Web agent system based on the results of the Carnot pro ject. It uses domain
ontologies which provide a declarative description of the semantic information in-
herent to a specific domain, independently of different system representation. A
semantic matching between requests and sources is then performed, based on the
corresponding domain ontology.
Information sources are thus considered at the level of their relevant semantic
concepts. User requests can be specified independent of structure, location, and
existence of data, and are then matched with the relevant information sources at
processing. This allows the matching procedure also to take into account run-time
information, e.g., the availability of a Web database.
Traditional querying and schema mappings are used together with the KIF
(Know ledge Interchange Format ) knowledge representation language (Genesereth,
1991), which enables programs with different internal knowledge representation
schemas to exchange information in a common language. The internal format is
converted to KIF and then wrapped in KQML (for Know ledge Query and Manip-
ulation Language ) performatives.
Queries are routed by mediators and brokers to data retrieval agents. Each agent
can process the requests it receives by making inferences on local knowledge in order
to answer some part of the query, or by routing the request to other agents, or by
decomposing it into sub-requests. Decisions are based on the domain ontology and
on the meta-data describing agent knowledge and relationships with other agents.
InfoSleuth has been applied in the field of health care over distributed heteroge-
neous databases of hospitals for comparing results of applied treatments, and in the
environmental domain as a component of EDEN (Environmental Data Exchange
Network).
4.5 Infomaster
Infomaster (Genesereth et al., 1997) is a system for information integration which
provides the user with homogeneous access to multiple distributed heterogeneous
information sources. These may be either databases, which are accessed by means
of wrappers, or structured Web sources.
In Infomaster, a facilitator dynamically determines, for each user request, the
most efficient query plan, trying to minimize the number of accessed sources and
coping with their heterogeneity, both in structure and content. An overview of
query planning capabilities in Infomaster can be found in (Duschka & Genesereth,
1997). Infomaster has been used at Stanford University, and is now commercially
available.
Using Methods of Declarative LP for Intel ligent Information Agents
13
4.6 COIN
COIN (COntext INterchange) (Bressan et al., 1997; Bressan & Goh, 1998) is an
information integration framework addressing semantic heterogeneity of data and
possible inconsistencies arising from different semantic interpretations of informa-
tion. It is based on a shared application domain model and on declarative definitions
of how different pieces of information have to be integrated. Utilizing a uniform
internal data model and language, the so-called context mediator rewrites and pro-
vides answers to user queries, which do not need to be expressed in the language
of the respective information source. Query optimization features are also provided
by the system.
4.7 HERMES
The Heterogeneous Reasoning and Mediator System (HERMES) (Adali et al., 1995;
Adali & Emery, 1995; Adali & Subrahmanian, 1996; Subrahmanian, 1994) is based
on a declarative language for the definition of a mediator (Wiederhold, 1992), which
expresses semantic integration of information from diverse data sources and rea-
soning systems. A facility enables the incremental integration of new systems into
an already existing mediator.
The system addresses two main issues:
1. domain integration, i.e. the physical linking of the available data sources and
reasoning systems; and
2. semantic integration, i.e., the coherent extraction and combination of the
provided information.
Methods to detect potential conflicts and how to resolve them, to pool informa-
tion together, and to define new compositional operations over information can be
specified and stored.
External data sources and software packages are accessed by means of the API
they provide, such that HERMES has to manage a translation of user queries into
the required syntax for these APIs, and to convert the delivered answers into a
suitable format for further processing.
In a specific mediator programming environment, a mediator compiler and some
yellow page servers are made available through a user interface.
Applications have been developed for the integration of terrain maps on the basis
of relational databases, as well as for face recognition in law enforcement.
4.8 IMPACT
In the Interactive Maryland Platform for Agents Collaborating Together (IM-
PACT) (Arisha et al., 1999; Subrahmanian et al., 2000; Eiter et al., 1999; Eiter
& Subrahmanian, 1999), several heterogeneous agents can interact with each other
and with some IMPACT servers by means of wrappers, one for each piece of soft-
ware serving as agent. IMPACT servers provide some infrastructural services which
14
Eiter et al.
the agents require in order to cooperate, including yellow pages, a thesaurus, and
registration facilities.
The decision-making component of each agent is declarative and comprises several
items, among them an agent program. This program contains rules which govern
the agent’s behavior. They are based on a logical abstraction of the internal func-
tions and data structures managed by the agent, together with the current state of
knowledge. Agents can also hold beliefs about other agents and make their decisions
depending on these beliefs, as well as on general constraints and action policies. A
precise semantics for agent programs is provided, formally describing the behavior
of agents.
External software can, by means of a logical abstraction, act and interact with
other agents present in the platform: IMPACT offers in this way the possibility of
integrating and reusing legacy software, making its services available in such a way
that implementation aspects are not relevant for the interaction.
The system has been successfully used in the development of a logistic applica-
tion for the U.S. Army. Other applications have been developed for department
stores, for solving the control led flight into terrain problem, and for supply chain
management.
4.9 SIMS
The Single Interface to Multiple Sources effort (SIMS) (Arens et al., 1993; Arens
et al., 1996) exploits the semantic model of a domain to integrate the information
from various sources, making use of techniques for knowledge representation and
modeling, for planning and searching, as well as for learning.
The system can be combined with Oracle databases as well as with LOOM (Mac-
Gregor, 1991) knowledge bases. Specific tools are used for generating database
queries in the appropriate language and for query planning.
The main aim of reasoning tasks in the system is to determine which information
sources are to be accessed, by matching and reformulating the queries, and then
to create and optimize the corresponding query plan. This is achieved through a
common domain model, consisting of ob jects and possible actions, which are then
mapped to the single databases.
The approach has been applied to transportation planning problems for the move-
ment of personnel and material from one location to another (Arens & Knoblock,
1992), to the control of the status of the U.S. Pacific Fleet (Arens, 1990), and to
decision support systems for medical trauma care.
4.10 TSIMMIS
The Stanford-IBM Manager of Multiple Information Sources (TSIMMIS) (Garcia-
Molina et al., 1995; Garcia-Molina et al., 1997; Quass et al., 1995) integrates het-
erogeneous information sources including structured and unstructured data, to be
used in decision support systems.
The architecture of the system includes several components:
Using Methods of Declarative LP for Intel ligent Information Agents
15
1. a classifier/extractor for the extraction of properties from unstructured
ob jects, based on the RUFUS System (IBM Almaden Research Center);
2. a translator/wrapper for the translation of information into a common
ob ject model, also performing data and query conversion;
3. a mediator combining information from the several sources;
4. a browser.
LOREL (Quass et al., 1995; Abiteboul et al., 1997b; Abiteboul et al., 1997a)
is the specially developed query language and is used in TSIMMIS for handling
semistructured data. The choice of developing a new language instead of using an
already existing one was motivated by the necessity of covering some features which
are not addressed by other languages. Queries should be
• answered meaningfully, even when some data is missing, by means of partial
type/ob ject assignments;
• handled uniformly over single- and set-valued attributes, to manage data with
irregular structure;
• manipulated uniformly over data of different types, in order to deal with
databases where different types refer to the same concept;
• implemented flexibly, to deal with the possibility of a query resulting in het-
erogeneous sets; and
• executable even if the ob ject structure is not fully known, by means of wild-
cards over attribute labels.
4.11 Further Projects
Apart from the pro jects described in the previous subsections, a number of other
pro jects are in progress which explore alternative promising techniques in the field
of information integration and management. We give here a short overview of some
of them.
4.11.1 MENTAL
This pro ject at Lisbon University, directed by L.M. Pereira, consists of four main
tasks, which partially overlap:
Agent architecture, coordination and task integration, dealing with:
• integration of abduction and tabling mechanisms;
• expressing updates by means of abduction;
• constraints, strong negation, and abduction.
The reference architecture (Dell’Acqua & Pereira, 1999b) is built upon the XSB-
Prolog system (Ramakrishnan et al., 1995; Sagonas et al., 1994) and combines
KS-Agents (Kowalski & Sadri, 1996), the dynamic logic programming paradigm
for updates (cf. Section 7.5), and tabling and abduction.
A framework for communicating agents has been suggested, in which agents have
the ability of dynamically updating their knowledge base and their goals as a re-
sult of interactions with other agents, based on forward and backward reasoning.
16
Eiter et al.
Knowledge management and reasoning, dealing with negation and paracon-
sistency, updates of logic programs (Alferes et al., 1998; Alferes et al., 2000a;
Alferes et al., 1999a), belief revision, and abduction. A prototype for alarm-
correlation in cellular-phone networks has been implemented (Frohlich et al.,
1999).
Actions, interactions and cooperation. Specialized sub-agents have been sug-
gested taking care of different aspects in multi-agent systems:
• actions and observations (Alferes et al., 1999b);
• coordination of interaction protocols;
• individual and joint planning of actions;
• non-interference behavior;
• resource sharing;
• self and mutual diagnosis of imperfect knowledge.
Learning, addressing the following issues (Lamma et al., 2000):
• reinforcement learning;
• causal deductive back-propagation;
• genetic algorithms applied to logic programming;
• degrees of belief based on the strength of supports of beliefs;
• truth maintenance based learning in logic programming.
4.11.2 Logic Programming for Multi-Agent Systems
The role of logic programming in multi-agent systems is usually confined to the
reasoning and decision components of agents, whereas other functionalities required
in the traditional observe-think-act cycle of the agent are provided by means of
other programming approaches. As suggested by Kowalski and Sadri (1999), it is
nevertheless possible to extend logic programming techniques for providing other
features required in multi-agent systems, with the purpose of unifying the rational
and the reactive components of an agent in a single framework. From this point of
view, the approach extends logic programming to handle multi-theories in a shared
environment, in order to take the knowledge bases of multiple agents into account,
and to model not only queries and updates in form of observations but also message
exchange and actions in the system.
The BDI architecture of Rao and Georgeff (Georgeff & Rao, 1991) can also be
seen as a special case of this framework, in which agent programs handle uniformly
different concepts such as beliefs and goals, queries, obligations and prohibitions,
constraints, actions, and commands.
Interesting applications of this extension of logic programming would be in the
field of distributed information integration, and in the field of network management.
4.11.3 MIA: Mobile Information Agents
The goal of this pro ject is to develop information systems which provide a mobile
user with information extracted from the Web, related to the spatial position of the
Using Methods of Declarative LP for Intel ligent Information Agents
17
user. The location of the user is checked and updated through a GPS receiver, so
that the system is able to answer user queries on a specific topic with information
of local relevance. For instance, a user arrives late in the night in some foreign city
and wants to know which restaurants are still open offering some particular kind
of food.
Information search and extraction from the Web is performed by information
agents ahead of the user requests. When the user (i.e., the client software) asks
for some information, the server side provides it with the knowledge present in the
system, with respect to the user location. The user is directly presented with the
required information, preprocessed and merged.
The agents building up the knowledge of the system are autonomous in detecting
Web pages which are of interest with respect to some specific topic, and in extracting
relevant information, based on its structure and content. The representation and
reasoning components are based on logic programming and automated theorem
proving methods, together with machine learning techniques. For more details about
the pro ject, cf. (Thomas, 1998; Thomas, 1999).
5 Declarative Methods
5.1 Overview
When looking for feasible or improved solutions to the main problems of advanced
information access, which we recalled in Section 3, some central reasoning sub-tasks
can be identified, in the form of representation and reasoning capabilities which
the information agent (or some module) must have. For some of these reasoning
tasks, listed below, methods of declarative logic programming (mainly based on the
answer set semantics) seem to be promising, and more complex control structures,
such as in Prolog, are not necessarily required. An advantage of these formalisms
is their well-defined semantics, which supports a strict formalization of reasoning
procedures and their outcomes.
Priority handling. Priorities may result from information the agent possesses
about reliability or accuracy of knowledge pieces or knowledge sources. Priority
information makes it possible, e.g., to select the appropriate sources, to merge an-
swers from different sources, and to remove possible inconsistencies. Dealing with
priority in the logic programming context has received considerable attention, cf.,
e.g., (Gelfond & Son, 1998; Brewka & Eiter, 1999; Inoue & Sakama, 1996; Foo
& Zhang, 1997a; Delgrande et al., 2000c; Brewka, 2000; Buccafurri et al., 1999a;
Buccafurri et al., 1999b; Alferes & Pereira, 2000). Priority information needs to
be encoded in the knowledge base of the agent, possibly in the ob ject language
itself in which the knowledge is expressed. The preference relation may be static
or dynamic, in the latter case reasoning procedures have to account for possible
changes.
Revision and update. The knowledge base of the agent can be sub ject to change
on the basis of information from the environment (e.g., the application domain
18
Eiter et al.
itself, other agents inside or outside the system, or sensing actions), or from in-
ternal learning procedures. When incorporating this new knowledge into the cur-
rent knowledge base, conflicts with available information may arise, and revision
or update strategies are required to yield a consistent knowledge base (Brewka,
2000; Marek & Truszczy´nski, 1994; Leite & Pereira, 1997; Inoue & Sakama, 1995;
Sakama, 2000; Inoue & Sakama, 1999; Foo & Zhang, 1997b; Foo & Zhang, 1998;
Alferes et al., 1998; Alferes et al., 2000a; Alferes & Pereira, 1997; Alferes &
Pereira, 2000; Alferes et al., 1999a). The epistemic state and the intended belief
set of the agent have thus to be continuously revised, and the revision or up-
date policy governing this process could in turn be explicitly described in the
knowledge base and may be dynamically updated too.
Inconsistency removal. The agent should, in the presence of conflicts in the
knowledge base, be able to detect them and to find an acceptable fal l-back knowl-
edge configuration, in order to ensure that the decision making process is not
stopped or inconsistent as a whole. The fall-back configuration is usually re-
quired to preserve “as much as possible” of the available knowledge. Strategies
for inconsistency removal may be based on explicit or implicit preference infor-
mation, e.g., the application domain may require old information to be preserved
against new one, thus old knowledge is to be preferred to solve inconsistencies.
Decision making with incomplete information. Under the manifestation of
incomplete information, some reasoning and deduction capabilities may provide
candidate hypotheses for the missing information pieces, in order to ensure that
the process of decision making can derive plausible results. We are not addressing
this task in this paper.
Temporal reasoning. The evolution of a dynamic knowledge base could be sub-
ject to external specifications (Alferes et al., 1999a), describing the content of a
sequence of update programs under dynamic conditions, or to internal update
policies, defining the way in which the knowledge base is going to evolve for some
specific situation. Corresponding forms of reasoning about knowledge evolution
have to be provided in order to ensure that the agent’s behavior is appropriate,
e.g., that some undesired status cannot be reached (Emerson, 1996; Clarke et al.,
2000), or that the agent may arrive at some goal state.
Learning. Based on the history of the agent (like a sequence of changes in its
knowledge base, or a sequence of observations), some form of inductive learn-
ing could be implemented (Sakama, 2000; Inoue & Sakama, 1999; Dell’Acqua &
Pereira, 1999b). This topic is also not addressed in this paper.
In the following, we deal with formalisms which offer partial or tentative solutions
for the implementation of such capabilities. We shall point out in which way and
for which purposes the individual approaches seem to be more appropriate (cf. also
Section 10), and we usually provide examples illustrating the different approaches.
Note that besides methods from logic programming, other knowledge represen-
tation and reasoning techniques could be suitable as well (e.g., description logics).
However, to keep this review focused, we do not consider such methods in this re-
view. Furthermore, we do not focus here on applying particular modes of reasoning,
Using Methods of Declarative LP for Intel ligent Information Agents
19
which have been extensively studied in the literature and proven to be useful, such
as:
• abduction and abductive logic programming (Kakas et al., 1992; Kakas et al.,
1998; Sakama, 2000); the relationship between abductive and disjunctive resp.
normal logic programming (Inoue & Sakama, 1994; Inoue & Sakama, 2000a;
Kowalski & Toni, 1995); learning with abduction (Kakas & Riguzzi, 1997);
cf. also references in Section 7.3;
• induction and abduction (Lamma et al., 1999; Flach & Kakas, 2000); cf.
also (Sakama, 2000);
• argumentation (Dung, 1993; Bondarenko et al., 1997; Kowalski & Toni, 1996);
argumentation for reasoning about actions and change (Kakas et al., 1999);
argumentation and priorities (Prakken & Sartor, 1997; Prakken & Sartor,
1999); argumentation and negotiation among agents (Jennings et al., 1997).
A broad overview with further references can be found in (Sadri & Toni, 2000).
5.2 Short Review of Logic Programs and Answer Sets
We recall some notions of logic programming and the answer set semantics, which
shall be useful in the following sections. We will deliberatively not focus on all
technical details, however.
A logic program consists of a (finite) set of rules Head ← Body , where Head is
a literal and Body is a (possibly empty) conjunction of literals. Rules with empty
Body are called facts. Literals are built over a set of atoms using default negation
not (also referred to as negation as failure, or weak negation ), and strong negation
¬ (also called classical negation ).
Example 1
Consider the following rules, representing the knowledge of an agent reasoning about
which of two information sites should be queried:
up (S ) ← not ¬up (S );
r1 :
¬query (S ) ← ¬up (S );
r2 :
query (index ) ← not ¬query (index ), up (index );
r3 :
r4 : query (cite seer ) ← not ¬query (cite seer ), ¬up (index ), up (cite seer );
flag error ← ¬up (index ), ¬up (cite seer ).
r5 :
Informally, rule r1 says that a site is up by default, and r2 expresses that if a site
is down, then it cannot be queried. Rule r3 states that the science citation index
site (index ) is queried by default, providing the site is up, while rule r4 states that,
by default, the cite seer site is queried if index is down but cite seer is up. Rule r5
says that if both index and cite seer are down, then an error is flagged.
If, as in this example, default negation does not occur in the heads of rules, then
P is called an extended logic program (ELP). If, additionally, no strong negation
occurs in P , i.e., the only form of negation is default negation in rule bodies,
then P is called a normal logic program (NLP). The generalization of an NLP by
20
Eiter et al.
allowing default negation in the heads of rules is called generalized logic program
(GLP). Further variants of logic programs exist, allowing rules with empty heads
(constraints ) and/or rules with a disjunction L1 ∨ · · · ∨ Ln of literals in the head.
In the latter case, programs are called disjunctive logic programs (DLPs).
The semantics of a logic program P is defined in terms of its ground instantiation
ground (P ), which contains all ground instances of rules over the Herbrand universe
that is generated by the function and constant symbols occurring in P .
There exist several semantics of ELPs (for an overview, cf. (Dix, 1996)). One
of the most widely used is the answer set semantics (Gelfond & Lifschitz, 1991),
which generalizes the stable model semantics for NLPs (Gelfond & Lifschitz, 1988).
Similar definitions for GLPs and other classes of programs can be found in the
literature (cf., e.g., (Lifschitz & Woo, 1992)).
A (partial ) interpretation is a set I of classical ground literals, i.e., literals without
not , such that no opposite literals A, ¬A belong to I simultaneously. For a classical
literal L, where L = A or L = ¬A for some atom A, we say that L is true in I if
L ∈ I , while not L is true in I if L /∈ I . An interpretation I satisfies a ground rule
r, denoted by I = r, if either the head of r is true in I , or some literal in the body
of r is not true in I .
Given an interpretation I and an ELP P , the Gelfond-Lifschitz reduct, P I , of P
with respect to I is the program obtained from ground (P ) as follows:
1. remove every rule r from ground (P ) which contains in the body a weakly
negated literal that is not true in I ;
2. remove all weakly negated literals from the bodies of the remaining rules.
An interpretation I is an answer set of a program P if it coincides with the
smallest set of classical literals which is closed under the rules of P I .
Example 2
Consider the program P from Example 1. The set
I = {up (index ), up (cite seer ), query (index )}
is an interpretation of P . Moreover, it is an answer set of P : The reduct P I contains
the rules
up (index ) ← ;
up (cite seer ) ← ;
¬query (index ) ← ¬up (index );
¬query (cite seer ) ← ¬up (cite seer );
query (index ) ← up (index );
flag error ← ¬up (index ), ¬up (cite seer ).
Each set of literals which is closed under the rules of P I must contain the lit-
erals up (index ), up (cite seer ), and, by virtue of the penultimate rule of P I , also
query (index ). No further literals need to be added. Thus, I is closed, and therefore
it is an answer set of P .
In general, an ELP may possess none, one, or several answer sets.
Using Methods of Declarative LP for Intel ligent Information Agents
21
Example 3
Suppose we add to the program P of Example 1 the rule
r6 : ¬query (S ) ← not query (S ), ¬rel (S )
and the facts ¬up (index ) ← and ¬rel (cite seer ) ← . Intuitively, the rule r6 states
that an unreliable source is not queried by default. Then, the corresponding ELP
Q has two answer sets:
J1 = {¬up (index ), ¬rel (cite seer ), up (cite seer ),
¬query (index ), query (cite seer )},
J2 = {¬up (index ), ¬rel (cite seer ), up (cite seer ),
¬query (index ), ¬query (cite seer )}.
The set of consequences of a program P is usually defined as the set of literals
which are cautiously derived from it, i.e., which are true in all answer sets of P . Thus,
in Example 1, query (index ) is a consequence of P , while neither query (cite seer )
nor ¬query (cite seer ) is a consequence of P .
If, as in Example 3, multiple answer sets exist, preferences might be used to single
out either a unique “preferred” answer set, or a collection of preferred ones. If, for
instance, rule r4 would have preference over the rule r6 , then J1 would be the single
preferred answer set of Q, and query (cite seer ) would be among the conclusions
of Q. Different ways to represent and handle preferences on logic programs are
reviewed in the next section.
Concerning updates, suppose a new fact ¬up (index ) ← should be incorporated
into P . If we simply add it, then the resulting program P ′ has the answer set
I ′ = {¬up (index ), up (cite seer ), ¬query (index ), query (cite seer )}.
Thus, query (index ) is retracted and query (cite seer ) is added. Simply adding a new
piece of information to a logic program is, in general, not a suitable strategy for
updating it, since conflicts may arise. We discuss some approaches to updates in
Section 7.
6 Preference Handling
An important requirement for an advanced information agent is the capability of
representing and handling priority information. Such information may be either
represented as meta-data associated with the rules of the knowledge base and used
to compute the intended (or preferred) belief set, or associated with single rules or
whole programs when revising or updating the knowledge base itself.
As meta-data associated to the rules of the knowledge base, preference informa-
tion can be useful to facilitate choices in case of multiple possibilities (e.g., if a logic
program, encoding an agent’s decision component, has more than one answer set),
or to determine rules which should be overridden in case of conflicts (e.g., the logic
program as a whole has no answer set, but, if a rule with low priority is retracted,
then the resulting program has an answer set).
22
Eiter et al.
When associated with (sets of ) rules serving as updates of the knowledge base,
preferences clearly express the priority of the new information with respect to the
current knowledge. These preferences can be obtained as meta-knowledge provided
by other source agents (e.g., indicating the precision of the required information)
or can be part of the meta-knowledge of the agent itself (e.g., knowledge about the
credibility of a certain source).
Several declarative methods have been developed in the field of logic programming
to cope with prioritized information, which differ in the following respects.
Meta- vs. ob ject-level encoding. Priorities can be explicitly expressed in the
ob ject language of the knowledge base (Sections 6.1, 6.4, 6.5, 6.7, 6.8), usually
by means of a special symbol in the alphabet and of a naming function for
rules; alternatively, they can be given as meta-knowledge associated to the rules
(Sections 6.2, 6.4, 6.6, and also 8.1) or to the literals (Section 6.3), in form of a
(partial) order among them.
Static vs. dynamic priorities. Priorities can be static or dynamic: static prior-
ities are not explicitly contained in the knowledge base and are not sub ject to
change (Sections 6.2-6.4, 6.6, and also 8.1); on the other hand when priorities are
expressed in the ob ject language (Sections 6.1, 6.4, 6.5, 6.7, 6.8), the preference
relation can itself be dynamically revised or updated together with the remaining
knowledge.
Fixed vs. variable strategies. Strategies for priority handling can be fixed, usu-
ally described through an appropriate ad hoc semantics taking care of respecting
the preference relation (Sections 6.2-6.4, 6.7, and also 8.1), or they can be encoded
in the knowledge base of the agent by means of appropriate program rewriting
(Sections 6.5, 6.6) or by addition of appropriate rules (Section 6.1). In this case,
no special semantics is usually needed, and the handling policy could itself be
sub ject to change, or could easily be modified for other purposes.
6.1 Prioritized Defaults by Gelfond and Son
In (Gelfond & Son, 1998), a method is suggested for reasoning with prioritized
defaults in extended logic programs under the answer set semantics. To this purpose,
a set of axioms, in the form of an ELP defining how prioritized defaults are to
be handled, is introduced. This set of axioms, added to any logic program with
priorities between the defaults, produces the desired results. No new semantics is
needed, as the answer set semantics of the union of the given program and of the
set of axioms works as intended.
A simple example from (Gelfond & Son, 1998) shows that the formalism works
as intended. Consider the following rules:
¬flies (tweety ) ← not flies (tweety ), penguin (tweety );
r1 :
flies (tweety ) ← not ¬flies (tweety ), bird (tweety );
r2 :
r3 : penguin (tweety ) ← not ¬penguin (tweety ), bird (tweety );
together with the fact bird (tweety ) ←, and with the information that r1 is preferred
over r2 , r2 over r3 , and r1 over r3 . Then, the approach computes only one answer
Using Methods of Declarative LP for Intel ligent Information Agents
23
set, containing the literals bird (tweety ), penguin (tweety ), ¬flies (tweety ) , as expected
(cf. (Gelfond & Son, 1998) for the necessary technical definitions).
In this approach, defaults and rules are technically distinguished, as rules are non
defeasible, while defaults express defeasible properties of some elements. Dynamic
priorities are addressed, since defaults and rules about the preference relation are
permitted. Moreover, changes in the properties of the preference relation, in the
definitions of conflicting defaults, or in the way of reasoning require only simple
changes in the formalism. Alternative strategies for priority handling can thus be
implemented by necessary changes in the set of axioms. For more complex examples
implementing these features, cf. (Gelfond & Son, 1998).
6.2 Preferred Answer Sets by Brewka and Eiter
In (Brewka & Eiter, 1999), two general principles for prioritized knowledge repre-
sentation formalisms are formulated and analyzed, and a notion of static preferences
over rules is introduced, satisfying these principles and extending the answer set
semantics for ELPs (an extension to dynamic preferences was given in (Brewka &
Eiter, 2000)).
Based on a given strict partial order over rules, some of the answer sets of the
ELP are selected as “preferred”, therefore increasing the set of conclusions which
can be cautiously derived from it.
Consider again the Tweety example, expressed by the following program P , where
the rules are listed with descending priority (cf. (Brewka & Eiter, 1999)):
P = {penguin (tweety ) ← ;
bird (tweety ) ← ;
¬flies (tweety ) ← not flies (tweety ), penguin (tweety );
flies (tweety ) ← not ¬flies (tweety ), bird (tweety )}.
Then, P has a unique preferred answer set, containing, as intended, the literals
penguin (tweety ), bird (tweety ), ¬flies (tweety ) .
A strong and a weak notion of preferred answer sets is considered, respectively,
where the second one ensures that if a program has an answer set, then it also
has at least a preferred one, which is not guaranteed with the strong notion, i.e., a
program may have answer sets but no preferred one.
The basic idea behind this approach is that the order in which rules are applied
has to be compatible with the explicitly given preference information among rules.
A rule can only be applied in case its body is not defeated by another rule having
higher priority.
6.3 Prioritized Logic Programs by Inoue and Sakama
In (Inoue & Sakama, 1996), the problem of reasoning with explicit representation of
priorities in logic programming is addressed. The proposed approach serves mainly
to reduce nondeterminism in choosing one out of several answer sets of a given pro-
24
Eiter et al.
gram, and can capture several forms of reasoning, e.g., abduction, default reasoning,
and prioritized circumscription.
The semantics of the so-called prioritized logic programs (PLPs) is given by means
of appropriate preferred answer sets. PLPs are constructed as general extended
disjunctive programs (GEDPs), including normal programs and able to express
abductive programs as well, together with a reflexive and transitive priority relation
over literals of the language.
The semantics of a PLP is defined in terms of the answer sets of the corresponding
GEDP, through an appropriate notion of preferred answer sets. Intuitively, from the
notion of priority over the elements of the language, priorities over answer sets are
derived. In case the set of priority relations is empty, the given definitions reduce
to the standard answer set semantics.
In the framework of PLP, the usual default reasoning problem about Tweety can
be represented by the following program P (cf. (Inoue & Sakama, 1996)):
¬flies (tweety ) ← penguin (tweety );
bird (tweety ) ← penguin (tweety );
penguin (tweety ) ← ;
flies (tweety ) ← default (tweety ), bird (tweety );
default (tweety ) ∨ not default (tweety ) ← ,
together with the preference information not default (tweety ) ≺ default (tweety ), in-
dicating that not default (tweety ) is preferred over default (tweety ). The unique pre-
ferred answer set is thus {bird (tweety ), penguin (tweety ), ¬flies (tweety )}, as desired.
Priorities over literals of the language can also be used to express more general
priorities, introducing new elements as heads of suitable rules and priorities over
these new elements; examples can be found in (Inoue & Sakama, 1996).
In (Inoue & Sakama, 1999), abduction is used to derive appropriate priorities
yielding intended conclusions. In particular, an appropriate set of priorities is de-
rived which can explain an observation under skeptical inference, thus addressing
the multiple extension problem of answer set semantics for ELPs. To this purpose,
the notion of preference abduction is introduced, in order to infer a sufficient pri-
ority relation to make the intended conclusion hold. This is achieved in terms of
an integrated framework for the abduction of both literals and priorities, using
prioritized and abductive ELPs.
The method can be described as follows. In a first step, a basic framework is intro-
duced. This step selects, through abduction, from a given set of candidate priorities
those which are sufficient to derive an observation skeptically, if the given observa-
tion is credulously but not skeptically inferred by the initial program. In a second
step, an extended framework is obtained by combining preference abduction with
ordinary abduction, providing the given observation cannot even be credulously
inferred by the initial program. In this case, given a set of abducibles and a set of
priority relations, subsets of abducibles and of priorities are selected such that the
initial program, together with new literals and the given priorities, is sufficient to
explain the observation skeptically.
In case the set of abducibles is empty, the extended framework collapses to the
Using Methods of Declarative LP for Intel ligent Information Agents
25
basic one, and if the set of candidate priorities is empty, the framework collapses to
abductive logic programming. If the set of abducibles is made up of rules instead
of literals, and the set of priorities contains relations over those rules, a naming
technique can be used to yield the original framework. Moreover, the extended
framework can also be used to derive possible new abducibles starting from the
rules which are the source of the nondeterminism. These are in turn converted to
abducible rules and then renamed, in order to arrive at new abducibles. Preferences
over abductive hypotheses can also be used to acquire new priority information.
6.4 Prioritized Logic Programs by Foo and Zhang
In (Foo & Zhang, 1997a), PLPs with static priorities under an extended answer set
semantics are introduced. In this approach, a PLP consists of an ELP, a naming
function for rules, and a strict partial order over names. The idea is to reduce PLPs
to ELPs by progressively deleting rules that, due to the defined priority relation, are
to be ignored. A rule is ignored if a rule of higher priority exists such that keeping
the latter and deleting the former still results in defeating the rule with lower
priority. If this is not the case, even if one rule is more preferred over the other,
the two rules do not affect each other during the evaluation, i.e., the preference
relation between these two rules does not play any role. Therefore, by progressively
checking all the rules with respect to the given order, the rules that are going to
be ignored can be deleted, and for the remaining ones the preference order plays
no role at all. This can be carried out until a fixed point is reached, in the form of
an ELP with no priority relation. The answer sets of this program are the intended
answer sets of the initial PLP.
As a formalization of the usual example about Tweety, consider the following
program P (cf. (Foo & Zhang, 1997a)):
flies (tweety ) ← bird (tweety ), not ¬flies (tweety );
¬flies (tweety ) ← penguin (tweety ), not flies (tweety );
bird (tweety ) ← ;
penguin (tweety ) ← ,
where the second rule has higher priority than the first one. The single answer set
of the program is given by
{¬fly (tweety ), bird (tweety ), penguin (tweety )}.
In (Foo & Zhang, 1997a), PLPs are extended to express dynamic priorities as
well. The semantics of a dynamic PLP is defined in terms of the answer sets of a
corresponding (static) PLP. More specifically, a dynamic PLP is a pair consisting of
a program over a language containing a symbol expressing a partial order between
names of rules, and a naming function for rules. In order to ensure the consistency of
the given partial order, the presence of some special rules in the program (expressing
antisymmetry and transitivity of the preference relation) is assumed. The semantics
is given in terms of a transformation from dynamic PLPs into (static) PLPs, through
a sequence of reductions with respect to the ordering. Starting from a dynamic PLP,
26
Eiter et al.
successive PLPs are built up until a fixed point is reached. Each new obtained
program is a reduct of the previous one with respect to the order relation over
the previous program, and the new order relation contains all pairs such that the
corresponding rule belongs to all answer sets of the previous PLP. The answer sets
of the dynamic program are the answer sets of the result of this transformation.
6.5 Compiled Preferences by Delgrande et al.
In (Delgrande et al., 2000c), a methodology is introduced for computing the an-
swer sets of ELPs with dynamic priorities between rules, explicitly expressed in
the program itself (such a program is called ordered ). On the basis of the encoded
preferences, an ordered program is transformed into a standard ELP whose an-
swer sets correspond to the intended preferred answer sets of the original one. The
transformation is realized by decomposing each rule of the program into a group of
associated rules, in which newly introduced control atoms guarantee that successive
rule applications are compatible with the given preference information.
The approach is sufficiently general to allow the specification of preferences
among preferences, preferences holding in a particular context, and preferences
holding by default. Moreover, the concept of static preferences, where the prefer-
ence information about rules is fixed (e.g., given at the meta level like in most
approaches to preference handling) is easily realizable.
Alternative strategies for preference handling can be achieved as well by chang-
ing the specification of the transformed program. For instance, in (Delgrande et al.,
2000a) it is shown that the method of Brewka and Eiter (1999) can be encoded
within this framework. Furthermore, it is straightforward implementing the method-
ology, since the result of the transformation is a standard ELP and therefore exist-
ing logic programming systems can be used as underlying reasoning engines. The
description of such an implementation is given in (Delgrande et al., 2000b).
6.6 Disjunctive Programs with Inheritance by Buccafurri et al.
In (Buccafurri et al., 1999b; Buccafurri et al., 1999a), a framework for disjunc-
tive logic programs with inheritance (denoted DLP <) on the basis of the answer
set semantics is introduced. Given a hierarchy of ob jects, represented by disjunc-
tive programs, possible conflicts in determining the properties of each ob ject in
the hierarchy are resolved by favoring rules which are more specific according to
the hierarchy. The hierarchy itself is given in terms of a strict partial order over
ob jects. If, for simplicity, we identify each ob ject with an indexed program, a hier-
archical knowledge base consists of a set of programs together with a partial order
over them. The program for a given ob ject in the hierarchy is then given by the
collection of programs at and above the corresponding indexed program. The se-
mantics of inheritance programs is defined as an extension of the standard answer
set semantics.
Using Methods of Declarative LP for Intel ligent Information Agents
27
6.7 Preference Default Theories by Brewka
In (Brewka, 2000), the problem of handling preference information in epistemic
states is addressed. Preference information is considered as explicitly expressed in
the representation language of epistemic states, viewed as nonmonotonic theories.
The desired results are achieved in two steps. The first step consists in an ex-
tension of default systems in order to express preference information in the ob ject
language, together with an appropriate definition of theory extensions. In a second
step, a notion of prioritized inference is introduced, as the least fixed-point of an
appropriate monotone operator, thus identifying epistemic states with preferential
default theories under this ad hoc semantics. Due to a potential self-referentiability
of preference information, not all theories expressed in this formalism possess ex-
tensions (i.e., acceptable sets of beliefs).
The main differences between the proposed formalism and the original default
notions consist in the following features:
• a single set is considered containing all information, in which there is no
absolutely unrevisable information; and
• preferences and other meta-information are expressed in the ob ject language,
by means of a naming mechanism for formulas and a strict total order between
names.
A preferred extension of a preference default theory is given by the deductive clo-
sure of a maximal consistent subset of the theory, which can be generated through
a total preference ordering on the formulas of the theory, compatible with the pref-
erence information explicitly expressed in the formulas belonging to the extension.
A formula is an accepted conclusion of a preference default theory if it belongs
to the least fixed-point of the operator computing the intersection of all preferred
extensions of the theory.
The idea is to iteratively compute, starting with the empty set, the intersection
of the extensions which are compatible with the information obtained so far. Since
the set of formulas in each step may contain new preference information, the num-
ber of compatible extensions may be progressively reduced, and their intersection
may thus grow. This process is continued until a fixed point is reached. If a formula
is an accepted conclusion of the theory, then it is contained in all preferred exten-
sions of the theory. The set of accepted conclusions of a theory is consistent and,
by identifying the theory with an epistemic state, the set of accepted conclusions
defines the corresponding belief set.
For illustration, consider how the usual Tweety example works in this framework
(cf. (Brewka, 2000)). Let T be the following theory (where ⊃ stands for implication):
bird (tweety ) ⊃ flies (tweety )
f1 :
f2 : penguin (tweety ) ⊃ ¬flies (tweety )
bird (tweety ) ∧ penguin (tweety )
f3 :
f4 :
f3 < f1
f2 < f1
f5 :
where f4 intuitively states that f3 is preferred over f1 , and f5 states that f2 is
28
Eiter et al.
preferred over f1 . Then, the only accepted conclusions of T are, as intended, given
by the set {bird (tweety ), penguin (tweety ), ¬flies (tweety )}.
In Section 7.8, we discuss the use of this formalism for representing revision
strategies.
6.8 Ordered Logic Programs
Another approach for resolving conflicts between rules is given by ordered logic
programs (Laenens & Vermeir, 1990; Gabbay et al., 1991; Buccafurri et al., 1996;
Buccafurri et al., 1999c) (cf. also Section 7.9).
Ordered logic programs offer the possibility of specifying preferences between
(sets of ) rules at the ob ject level, by means of a naming function for rules and
of special atoms expressing priority between rule names. These special atoms can
appear everywhere in a rule, thus permit to express dynamic preferences.
Considering again the example about Tweety, it can be formalized by the program
P containing the following rules:
r1 : penguin (tweety ) ←
bird (tweety ) ←
r2 :
r3 : flies (tweety ) ← bird (tweety ), not ¬flies (tweety )
r4 : ¬flies (tweety ) ← penguin (tweety ), not flies (tweety )
r3 ≺ r4 ←
r5 :
where the last rule intuitively states that r4 is preferred over r3 . Then, as intended,
the preferred answer set of P is given by
{penguin (tweety ), bird (tweety )¬flies (tweety ) .
While ordered logic programs are similar to logic programs with inheritance as
defined by Buccafurri et al. (1999a), there is a salient difference: in the former,
contradictions between rules r1 and r2 from different sets of rules R1 and R2 ,
respectively, with no preference among them are removed, while in the latter they
lead to inconsistency. For example, for R1 = {rain ← } and R2 = {¬rain ← }
having lower priority than R0 = ∅, the respective ordered logic program has two
preferred answer sets, {rain } and {¬rain }, while the respective inheritance program
has no answer set.
6.9 Further Approaches
Besides the methods discussed above, a number of other approaches for adding
priorities in extended logic programs are available. We briefly address some of them
in this section.
Under the answer set semantics, ELPs may be viewed as a fragment of Reiter’s
default logic, as there is a one-to-one correspondence between the answer sets of an
ELP P and the consistent extensions of a corresponding default theory (Gelfond
& Lifschitz, 1991). A semantics for prioritized ELPs is thus inherited from any
semantics for prioritized default logic, such as the proposals discussed in (Marek
Using Methods of Declarative LP for Intel ligent Information Agents
29
& Truszczy´nski, 1993; Brewka, 1994; Baader & Hollunder, 1995; Rintanen, 1998).
Applied to ELPs, all these semantics select, by using priority information on rules,
particular answer sets of an ELP from the collection of all answer sets. They fail,
however, to satisfy the principles for this selection process which have been proposed
in (Brewka & Eiter, 1999).
Further semantics for priorities in extended logic programming, which are not
based on answer sets, have been proposed in (Nute, 1994; Analyti & Pramanik,
1995; Brewka, 1996; Prakken & Sartor, 1997). They have quite different founda-
tions such as logical entrenchment and specificity (Nute, 1994), reliability handling
(Analyti & Pramanik, 1995), well-founded semantics (Brewka, 1996), or defeasible
argumentation (Prakken & Sartor, 1997).
In (Pradhan & Minker, 1996), it is shown how priorities can be used to combine
different potentially conflicting Datalog databases. Preferences are used to deter-
mine that information which is given up in the merging process. Three different
semantics of priorities are defined, two of which are equivalent. However, Pradhan
and Minker do not consider negation (neither weak nor strong) at all, and the
approach has thus limited expressive capability.
In (Kowalski & Sadri, 1991), rules with negation in the head are considered as
exceptions to more general rules, and they are given higher priority. Technically,
this is achieved by a redefinition of answer sets in which answer sets as defined in
Section 5.2 are also answer sets according to the definition of (Kowalski & Sadri,
1991). The main achievement is that programs which merely admit an inconsistent
answer set (as possible in the definition of (Gelfond & Lifschitz, 1991)), become
consistent in the new semantics. Thus, the approach in (Kowalski & Sadri, 1991)
has more of a contradiction removal method than of a priority approach.
For further discussions and comparisons concerning different approaches for pref-
erences, we refer the reader to (Brewka & Eiter, 1999; Delgrande & Schaub, 2000;
Inoue & Sakama, 2000b).
7 Revision and Update
An agent’s knowledge base is naturally sub ject to change, which may happen
through different events, including the following:
1. new information is arriving at the agent (this could, e.g, be the result of a
sensing action or a message from some other agent);
2. temporal evolution, e.g., a “tick of the clock” event activating some predefined
evolution (Alferes et al., 1999a); or
3. the tentative merging of the current knowledge base with another one.
Two basic approaches to this change process can be distinguished: revision and
update. The idea of program revision is that of a program forced to include addi-
tional information about a world supposed to be static, by revising the assumptions
it contains. On the other hand, the issue of program updating is to deal with a dy-
namic world (Katsuno & Mendelzon, 1991). A detailed discussion on the different
approaches to changes of the knowledge base can also be found in (Gardenfors &
30
Eiter et al.
Rott, 1995; Nebel, 1998); a clear distinction is not always possible, and “hybrid”
formalisms are possible as well (Eiter et al., 2000a).
Moreover, a plethora of other features distinguish the different methodologies
such that a clear, orthogonal classification is not straightforward. In particular, the
following features can be found:
• Revisions and updates can involve two theories or two sets of rules and facts
(Sections 7.1–7.7, and 7.9), often two logic programs, or the current program
or theory plus a single rule or fact to be inserted in it (Section 7.8). Some
of them eventually consider the case in which a single rule or fact has to be
retracted too (Sections 7.1–7.3).
• Update formalisms usually implicitly assign higher priority to the “new” in-
formation, viewed to represent changes in the external world, which the agent
should be forced to accept (Sections 7.1–7.8). Other formalisms assume no
priority between old and new information; in this case, the term merging is
often used (e.g., Section 7.9).
• As in the case of priority handling, some methods define an ad hoc semantics
(Section 7.8), while others rely on a rewriting of the original programs under
some standard semantics (Sections 7.2, and 7.4–7.7). If the update policy is
explicitly expressed, changes to the processing strategy (e.g., arbitrary pri-
ority assignments to the involved programs instead of fixed ones) could be
easier realized. Sometimes, the update policy may itself be sub ject to change
(Section 7.8).
• Most of the approaches pose no explicit requirements on the consistency of
the old or the new knowledge base (Sections 7.1–7.3, 7.5, and 7.6). As a
consequence of the update process, conflicts may sometime arise, and the for-
malism may implicitly contain a strategy for inconsistency removal leading
to a consistent belief set (Sections 7.3, 7.4, and 7.8), or rely on explicit pri-
ority information (Section 7.6). For other formalisms, addressing consistency
requires changes in the update approach and its semantics (Section 7.5).
• In some approaches, old knowledge, which is overridden or contradicted by the
new one, is physically deleted from the knowledge base (Section 7.1–7.3, 7.9),
while in other approaches, it is only temporarily de-activated (Section 7.8)
and can be activated again by successive updates (Sections 7.5–7.7).
• It is also important to consider whether the formalism imposes some form
of minimality of updates, in order to preserve “as much as possible” of the
current knowledge when new information is incorporated. Some of the meth-
ods include already in their basic definitions some condition for minimality
of change (Sections 7.1–7.4, 7.9), while for other formalisms additional limi-
tations are required or no minimality notion is given at all.
• Some of the formalisms apply only to single updates (Sections 7.3, 7.8, 7.9),
while others can be applied uniformly to single updates as well as to se-
quences of updates (Sections 7.5–7.7). For some of the formalisms based on
the rewriting of the involved programs, iteration is technically possible, i.e.,
the program resulting after an update can be directly sub ject to update again.
(Sections 7.1, 7.2, 7.4–7.8).
Using Methods of Declarative LP for Intel ligent Information Agents
31
In what follows, we discuss several update and revision formalisms in more detail.
7.1 Revision Programs by Marek and Truszczy´nski
In (Marek & Truszczy´nski, 1994), a language for revision specification of knowledge
bases is presented, which is based on logic programming under the stable model
semantics. A knowledge base is in this context a set of atomic facts. Revision rules
describe which elements should be present (so-called in-rules) or absent (out-rules)
from the knowledge base, possibly under some conditions. Note that such a defi-
nition is self-referential, because the latter conditions must be evaluated over the
resulting knowledge base. A fixed-point operator, which satisfies some minimality
conditions, is introduced to compute the result of a revision program. As for sta-
ble models, there may be several or no knowledge base satisfying a given revision
program. While the approach was formulated for the propositional case, a possible
extension to the predicate case is also briefly discussed.
7.2 Update Rules as Logic Programs by Pereira et al.
In (Alferes & Pereira, 1997), transition rules for updating a logic program are
specified by means of another logic program. Update programs are defined as a
collection of update rules, specifying all the atoms which must be added or deleted
depending on some pre-conditions.
The notion of justified update of a total interpretation for an initial program is
defined, given an update program and a stability condition, on the basis of revision
definitions. The stability condition guarantees that the initial interpretation is pre-
served as much as possible in the final one, applying inertia rules to those literals
not directly affected by the update program.
In case the initial base is a normal logic program (NLP), a transformation is
given which produces an extended logic program (ELP) whose models enact the
required changes in the models of the initial program, as specified by update rules.
The transformation guarantees that, by inertia, rules in the initial program remain
unchanged unless they are affected by some update rules. This is realized by re-
naming all the atoms in the rules which are possibly affected by some update rule,
and introducing inertia rules stating that any possibly affected atom contributes to
the definition of its new version, unless it is overridden by the contrary conclusion
of a rule from the update.
A generalization of this transformation to the case where both the initial and the
update program are ELPs is also given. This generalization is necessary because
the result of the update is an ELP but not an NLP in general; thus, a sequence
of updates could not be processed by simple iteration of updates. The approach
is similar to the approach for NLPs and works as desired under the well-founded
semantics. With this extension, successive update transformations may take place.
In later papers (Leite & Pereira, 1997; Alferes et al., 1998; Alferes et al., 2000a),
the focus is shifted from models to programs. The idea is that a logic program
encodes more than a set of models: it encodes also implicit relationships between
32
Eiter et al.
the elements of those models. The principle of inertia should therefore be applied
to the rules in the program rather than to the atoms in the models. In order to
update a program, it must be checked whether the truth value of a body which
determines the truth value of a head has not changed, before concluding the truth
value of the head by inertia. In fact, the truth value of the body may change due to
an update rule. The truth of any element in the updated models should therefore
be supported by some rule (one with a true body) either of the update program or
of the given one, in the light of the new knowledge.
For illustration, we consider a simple example from (Leite & Pereira, 1997),
which shows the difference of shifting inertia from literals to rules, as in (Marek &
Truszczy´nski, 1994) and (Alferes & Pereira, 1997). Suppose the knowledge base is
given by the following program P :
go home ← not have money ;
go restaurant ← have money ;
have money ← ,
whose single model is {have money , go restaurant }. According to revision program-
ming, updating this model with the information out (have money ) ← in (robbed )
and in (robbed ) ← results in the model {robbed , go restaurant }, which is counter-
intuitive, as argued by Leite and Pereira (1997). On the other hand, updating P
by applying inertia to rules, as realized in Leite and Pereira’s approach, yields the
desired model {robbed , go home }.
In this sense, model updates are a special case of program updates, since models
can be encoded as programs containing only facts. The updating of NLPs and ELPs
is thus defined with the same approach, only shifting the application of inertia from
atoms to rules in the original program. The resulting updated program depends
only on the initial program and the update program, but not on any specific initial
interpretation. The rules of the initial program carry over to the updated one,
due to inertia, just in case they are not overruled by the update program. This
is achieved by defining a subprogram of the initial program containing those rules
which should persist due to inertia. This subprogram is then used together with
the update program to characterize the models of the resulting updated program.
A joint program transformation is then given, taking as input the initial and the
update program, and producing an updated program whose models are the required
updates, by introducing suitable inertia rules.
7.3 Abductive Updates by Inoue and Sakama
Inoue and Sakama (1995) present an approach to theory update which focuses
on nonmonotonic theories. They introduce an extended form of abduction and a
framework for modeling and characterizing nonmonotonic theory change through
abduction. This framework is shown to capture and extend both view updates in
relational databases and contradiction removal for nonmonotonic theories.
Intuitively, this is achieved by extending an ordinary abductive framework to au-
toepistemic theories (Moore, 1985), introducing the notions of negative explanations
Using Methods of Declarative LP for Intel ligent Information Agents
33
and anti-explanations (which make an observation invalid by adding hypotheses),
and then defining autoepistemic updates by means of this framework.
The aim of the extension is to provide abduction with the following features:
• both the background theory and the candidate hypotheses may be autoepis-
temic theories;
• hypotheses can not only be added to the theory but also discarded from it to
explain observations;
• observations can also be unexplained.
Inoue and Sakama introduce the notion of a minimal pair of positive and negative
(anti-)explanations, and define update as a function on autoepistemic theories,
which consists in the insertion and/or deletion of formulas in autoepistemic logic,
based on (anti-)explanations accounting for the new observations. A combination
of such an abductive and inductive framework is further elaborated in (Sakama,
2000).
The framework of extended abduction is then used in (Inoue & Sakama, 1999)
to model updates of nonmonotonic theories which are represented by ELPs. The
following example from (Alferes et al., 2000a) shows what the results of the approach
are. Consider the initial knowledge base given by the program P :
P = {watch tv ← tv on ;
sleep ← not tv on ;
tv on ← }
and the update expressed by the program U :
U = {power failure ← ;
¬tv on ← power failure }.
The update program resulting by the abductive representation of the problem has
the unique (under minimality of change) answer set {power failure , ¬tv on , sleep }.
Using appropriate update programs, i.e., appropriate ELPs specifying changes
on abductive hypotheses, the following problems are solved through abduction:
View updates: The problem of view update consists of incorporating new literals
in the variable part of a knowledge base in which variable and invariable knowl-
edge are distinct (as in a database made up of an intensional and an extensional
part, updates to the intensional part have to be translated into updates to the
extensional part, the variable one).
Theory updates: For theory updates, the whole knowledge base is sub ject to
change. New information in the form of an update program has to be added to
the knowledge base and, if conflicts arise, higher priority is given to the new
knowledge. The updated knowledge base is defined as the union Q ∪ U of the
new information U and a maximal subset Q ⊆ P of the original program that is
consistent with the new information (which is always assumed to be consistent).
The abductive framework is in this context used for specifying priorities between
current and new knowledge, by choosing as abducibles the difference between the
initial and the new logic program.
34
Eiter et al.
Inconsistency removal: The problem of inconsistency removal is seen as a special
case of theory update, where the initial program is inconsistent and the new
program is empty.
These three different problems are considered in a uniform abductive framework,
namely that of (Inoue & Sakama, 1995). Based on the necessary update rules, an
appropriate update program is defined for each case, which is an ELP specifying
the changes on the abductive hypotheses.
7.4 Updates by Means of PLPs by Foo and Zhang
In (Foo & Zhang, 1997b), the update of a knowledge base of ground literals by
means of a prioritized logic program (PLP, as defined in Section 6.4) over the
same language is described. An update is itself a PLP over an extended language,
consisting of the following elements:
• a program P containing initial knowledge rules, inertia rules and update rules;
• a naming function N for rules in the resulting program; and
• a partial order < containing a pair r < r′ for each inertia rule r and each
update rule r′ in the new program, stating higher priority of inertia rules.
The possible resulting knowledge base is defined on the basis of the answer sets of
the resulting program as follows: if it has no answer set, then the updated knowledge
base coincides with the initial one; if its answer set is the set of all literals in the
extended language (i.e., if we have inconsistency), then the updated knowledge base
is given by the set of ground literals of the original language; if it has a consistent
answer set, then the updated knowledge base is given by the set of ground literals of
the original language such that the corresponding literal in the extended language
belongs to the answer set of the resulting program. It is shown that the introduced
definitions satisfy the minimal change property with respect to set inclusion.
In (Foo & Zhang, 1998), the problem of updates is addressed if both old and new
knowledge is encoded as ELP. The idea in updating the initial program with respect
to the new one is to first eliminate contradictory rules from the initial program with
respect to the new one, and then to solve conflicts between the remaining rules by
means of a suitable PLP.
In the first step, each answer set of the initial program is updated with the
information in the new one, producing a set of ground literals which has minimal
difference to the answer sets of the initial program and which satisfies each rule
in the new one. If the resulting set of literals is consistent, a maximal subset of
the initial program is extracted such that the given set of literals is coherent with
the union of this maximal subset and the new program. Thus the given set of
literals is a subset of an answer set of the defined union. The resulting subset,
called transformed program, is guaranteed to maximally retain rules of the initial
program which are not contradictory to rules of the new program. If the given set
of literals is not consistent, then the transformed program is any maximal subset
of the initial program such that its union with the new program is consistent.
Using Methods of Declarative LP for Intel ligent Information Agents
35
In a second step, possible conflicts between the rules in the transformed program
and in the new program need to be resolved. To this purpose, a special PLP is
introduced as follows:
• the program itself is the union of the transformed program and of the update
program;
• a naming function for rules is given;
• the priority relation contains all the possible pairs consisting of one rule of
the transformed program and one rule of the new program, stating higher
priority of new rules.
The semantics of the update is thus given by the semantics of the corresponding
PLP.
7.5 Dynamic Logic Programming by Alferes et al.
Dynamic logic programming (DynLP) is introduced in (Alferes et al., 1998) as
a new paradigm, which extends the idea of an update of a logic program P by
means of a logic program U (denoted P ⊕ U ) to compositional sequences L P =
P1 ⊕ P2 ⊕ ... ⊕ Pn .
The approach is obtained by defining a new extended language and by building
up the program update L P over this new language by means of a (linear-time)
transformation of the set of rules contained in the original sequence to a gener-
alized logic program (GLP), which encodes the intended sequence of updates. An
extended interpretation is defined from interpretations over the original language
to interpretations over the new one. The stable models of the program update are
defined to be the pro jection of L P to the original language, and a characterization
of stable models in terms of rule rejection set is provided.
As an example, consider again the programs P and U from Subsection 7.3 (Alferes
et al., 2000a):
P =
{watch tv ← tv on ;
sleep ← not tv on ;
tv on ← }
U = {power failure ← ;
¬tv on ← power failure }.
The unique stable model of the update of P by U is, as intended,
{power failure , sleep }.
Performing a second update by the rule not power failure ← , the new stable model
becomes then {tv on , watch tv }.
In (Alferes et al., 2000a), some further features of DynLP are discussed. The first
one is the efficient representation of background knowledge. That is to say, some
kind of knowledge that is true in every program module or state. Background rules
are easy to represent and efficient to handle, as adding a rule to every state up to
the current one is equivalent to adding that rule only in the current state.
36
Eiter et al.
Another issue for which several possibilities are mentioned is dealing with contra-
diction and how to establish consistency of updated programs. The contradiction
may be explicit and may depend on several reasons: the current state is already
inconsistent or the update program is contradictory, or both. In the first case, an
approach to deal with this situation is to prevent the contradiction to be inherited,
by limiting the inheritance by inertia. In the second case, the solution is not so
intuitive, as the rules of the update program must be, by definition, true in the
updated program. A possible approach here is to require the update program to be
consistent. An alternative approach is to accept contradiction in the current update
and again prevent it from further proliferating by changing the inertia rules as in
the first case. In the third case, the same approach is possible as well, or it is possible
to establish priorities between different rules in order to prevent contradiction.
Yet another kind of contradiction is an implicit one: when the updated program
is normal, explicit contradiction cannot arise, but it may well be that the updated
program has no stable model. A possibility consists in shifting to three-valued stable
semantics or well-founded semantics.
The approach suggested suffers from problems with changing the inertia rules.
While it is easy to encode several conditions and applications to prevent or re-
move contradiction, to enact preferences, or to ensure compliance with integrity
constraints, such changes in the inertia rules require that the semantic characteri-
zation of the program updates is adjusted accordingly. This seems far from trivial.
DynLP is combined in (Dell’Acqua & Pereira, 1999b) with agents based on the
paradigm of Kowalski and Sadri (Kowalski & Sadri, 1996; Dell’Acqua et al., 1998),
for building rational reactive agents which can dynamically change their knowledge
bases and goals (directed by observations), and learn new rules. According to the
approach of Kowalski and Sadri, each agent is an abductive logic program executing
an observe-think-act cycle, where abducibles are given by actions to be executed
or explanations of observations. Inputs to each cycle are the required updates; for
each agent and for each cycle the knowledge base of that agent is identified with
the dynamic program update at the corresponding state.
7.6 Updates and Preferences by Alferes and Pereira
In (Alferes & Pereira, 2000), an integrated framework combining updates and pref-
erences is introduced, based on DynLP and the preference semantics due to Brewka
and Eiter (1999). In this approach, a new language is defined, modeling sequences
of programs P1 , . . . , Pn resulting from consecutive updates of an initial program,
together with a priority relation among the rules of all successive programs. The
priority relation is itself sub ject to update.
This integrated approach is based on the idea that both preferences and updates
eliminate rules: preferences eliminate less preferred rules, selecting among the avail-
able stable models, and updates eliminate rules overruled by other ones, in order
to generate new models. Preferences require a strict partial order on rules, while
updates require a linear temporal order, or other distinct linear structures, allowing
nevertheless the production of a tree of linear updating sequences.
Using Methods of Declarative LP for Intel ligent Information Agents
37
In this framework, preferences may be enacted on the results of updates, whereas
updates may be used for the purpose of changing preferences. Preferences are under
this view intended to select further rules after computing the results of updates.
Intuitively, starting from the semantics of updates (erasing rejected rules), the se-
mantics of preferences is defined (erasing unpreferred rules) according to the method
of (Brewka & Eiter, 1999) and aiming at a combination of the two. An integrated
semantics is then formulated for both of them.
The ob ject language of the approach is that of DynLP under the stable model
semantics; explicit negation can be expressed by corresponding new rules and atoms.
Formally, these prioritized dynamic programs are defined using a strict partial order
over rules. This preference information is used to prefer among stable models which
are a fixed-point of an equation guaranteeing that the rules are being applied in
observance of the partial order.
In order to facilitate the capturing of both preferences and updates in a single
framework, the semantics has to be based on the removal of less preferred rules.
The main issue is therefore to determine criteria for rules to be removed in order
to obtain the same result as in (Brewka & Eiter, 1999). This is achieved in the
following way: while in (Brewka & Eiter, 1999), the head of a rule is not added to
the construction of a preferred stable model if the rule is defeated by the previous
constructed set (formed by the heads of the more preferred rules), the same effect
can be obtained by removing all rules defeated by the head of a more preferred rule
which in turn has not been removed itself. In other words: remove all rules whose
body is in the model and whose head defeats a more preferred rule. If the body of a
less preferred rule is not actually true in the model, then its defeat is only potential
and the rule must not be eliminated. Formally, the notions of unsupported and
unpreferred rules are introduced, in order to obtain the appropriate definition of
the preferred stable models. The definition of preferred stable models is equivalent
to the one of preferred answer sets given in (Brewka & Eiter, 1999).
After defining the appropriate semantics of preferences, preference and update
semantics are combined. The fundamental question is where to define the priority
relation. Two possible levels for the definition of priorities are identified: (i) prior-
ities among rules of the same program, or (ii) priorities among rules in the union
of all programs in the sequence. According to (Alferes & Pereira, 2000), the second
possibility is more general, as it does not prevent limiting the priority relation to
rules in the same program. The chosen language for such a representation must be
able to cope at the same time with evolution by means of updates and evolution of
the priority relation. Moreover, the language is an extension of DynLP: instead of
sequences of GLPs P1 , . . . , Pn , sequences of pairs hP1 , Q1 i, . . . , hPn , Qn i are given,
where each pair hPi , Qi i consists of a program Pi representing rules and of a pro-
gram Qi describing a priority relation among rules. In general, as an update of the
priority relation may depend on some other predicates, rules in programs describ-
ing priorities are allowed to refer to predicates defined in the programs representing
knowledge.
Let us consider an example from (Alferes & Pereira, 2000). The initial knowledge
is given by the program P1 containing the facts safe (chevrolet ) , fast (chevrolet ),
38
Eiter et al.
expensive (chevrolet ), safe (volvo ) , fast (porsche ), together with the following rules:
r1 : not buy (X ) ← avoid (X );
avoid (X ) ← not buy (X ), expensive (X );
r2 :
buy (X ) ← not fast (X );
r3 :
avoid (Y ) ← fast (X ), buy (X ), Y 6= X ,
r4 :
where the preferences over rules are as follows: r2 < r3 , r2 < r4 , stating that rule r2
is preferred over both r3 and r4 . Let us then update this knowledge by the program
P2 , made up of the following rules:
buy (X ) ← not avoid (X ), safe (X ) ;
r5 :
r6 : avoid (Y ) ← safe (X ), buy (X ), Y 6= X ,
with associated priorities r5 < r3 , r5 < r4 , r6 < r3 , r6 < r4 , r2 < r5 , r2 < r6 . Then,
the only preferred stable model of the update with priorities is
{buy (volvo ), avoid (porsche ), avoid (chevrolet ) }.
Besides the extension of DynLP to consider preferences, the work in (Dell’Acqua
& Pereira, 1999b) about updating agents is extended for preference and updating in
multi-agent systems (Dell’Acqua & Pereira, 1999a). Preferences are thus included
among the rules of an agent and allow a selection among several models of the
corresponding (updated) knowledge base. Preferences can themselves be updated,
possibly depending on information coming from other agents.
7.7 Inheritance Programs and Updates
Concerning the approach discussed in Section 6.6, it is natural to view a sequence
of updates as an inheritance program where later updates are considered to be
more specific. More specifically, as argued in (Eiter et al., 2000a), we may view
the sequence P = P1 , . . . , Pn of programs as a knowledge base with inheritance
order Pn < Pn−1 < . . . < P2 < P1 , i.e., for each j = 1, . . . , i, Pj+1 is more specific
than Pj . It is shown in (Eiter et al., 2000b) that for certain classes of programs,
the update sequence P1 , . . . , Pn is equivalent to such a corresponding inheritance
program.
For illustration, consider again the example from (Alferes et al., 2000a), but in
a slightly adapted form. Let the initial knowledge base be given by the following
program P :
P = {watch tv ← tv on ;
sleep ← not tv on ;
night ← ;
tv on ← }
and the update defined by the program U :
U = {power failure ← ;
¬tv on ← power failure }.
Using Methods of Declarative LP for Intel ligent Information Agents
39
The update of P by U has, as intended, the unique answer set
S = {power failure , ¬tv on , sleep , night }.
If, again, new knowledge arrives in form of the program U ′ = {¬power failure ← },
then the unique answer set of the sequence P, U, U ′ is
S ′ = {¬power failure , tv on , watch tv , night }.
It is easy to see that in the formalism of (Buccafurri et al., 1999a), S is an answer
set of the inheritance program U < P , and, similarly, S ′ is an answer set of the
inheritance program U ′ < U < P .
As shown in (Alferes et al., 2000a), dynamic logic programs naturally gener-
alize the idea of updating interpretations, viewed as collections of positive and
negative facts, through Marek and Truszczy´nski’s revision programs (Marek &
Truszczy´nski, 1994) (which, in turn, have been further considered by Przymusin-
ski and Turner (1997)). Thus, inheritance programs also generalize interpretation
updates.
(Buccafurri et al., 1999a) contains also a transformation from disjunctive inheri-
tance programs to inheritance-free disjunctive programs, similar in spirit to update
programs (cf. Section 7.5).
7.8 Revision of Preference Default Theories by Brewka
In (Brewka, 2000), a nonmonotonic framework for belief revision is introduced which
allows reasoning about the reliability of information, based on meta-knowledge ex-
pressed in the language of the information itself. In this language, revision strategies
can be declaratively specified as well. The idea is to revise nonmonotonic theories by
adding new information to the current theory, and to use an appropriate nonmono-
tonic inference relation to compute the accepted conclusions of the new theory. The
approach is described in Section 6.7.
The main advantage of this approach is that it allows the explicit representation of
preference information, which is commonly used by agents in the process of revising
their beliefs. Once preference relations are represented in the language itself, it is
possible to represent revision strategies declaratively as well, and to revise them
too.
In this context, epistemic states are identified with preference default theories,
and belief sets with their accepted conclusions, since preference default theories un-
der the accepted conclusion semantics always yield consistent belief sets. The belief
sets cannot be revised directly, but only through the revision of the corresponding
epistemic state. Given an epistemic state, revising it with new information simply
means generating a name for the new information, and adding it to the theory.
The notion of preferred extensions of the revised theory serves to compute the
revised belief set, taking the newly added information into account, possibly con-
taining preference information as well.
The suggested notion of revision is evaluated with respect to the AGM postu-
lates (Alchourr´on et al., 1985) (suitably reformulated in terms of belief sets). Fur-
40
Eiter et al.
thermore, it is explained why it is meaningful that some postulates are not satisfied
by the approach.
Contraction, i.e., making a formula underivable, can be handled in the case that
the agent receives explicit information of the form “believe ¬p”. However, if the
form of the incoming information is “do not believe p ”, the introduction of an extra
mechanism seems to be necessary in order to make the formalism be able to deal
with it as well. To this purpose, the idea of distinguishing constraints from premises
can be useful: constraints are viewed as formulas used in the construction of maxi-
mal consistent subsets of premises, but are not used in derivations to compute the
deductive closure of a set of formulas. The compatibility of the extensions with the
preference relation has then to be checked against both constraints and premises,
whereas extensions are generated only from the premises.
In this framework, knowledge bases grow monotonically. However, for agents with
limited resources, the expansion of the knowledge base cannot go on forever. The
problem consists in determining how and when some pieces of information should
be forgotten, activating a sort of garbage-collection strategy. This requires some
notion of utility over information.
7.9 Arbitration
A problem which is closely related to revision and update is arbitration between two
or more knowledge bases, which is typically performed when they are to be merged
into a single knowledge base. Under arbitration, one understands combining the
knowledge of different knowledge bases having the same priority.
Approaches from logic programming, such as inheritance programs (Buccafurri
et al., 1999b; Buccafurri et al., 1999a) and extensions of dynamic logic program-
ming (Alferes et al., 2000a; Alferes et al., 2000b), may be used to technically address
this problem, even though they have not been specifically designed for arbitration
between knowledge bases. Thus, they may not lead to a satisfactory behavior, and,
in particular, simple inconsistencies between knowledge bases of which each one is
consistent (like, e.g., {a ←} and {¬a ←}) may lead to inconsistency of the com-
bined knowledge base. On the other hand, ordered logic programming (Laenens
& Vermeir, 1990; Gabbay et al., 1991; Buccafurri et al., 1996; Buccafurri et al.,
1999c) can avoid this problem. The notion of a stable model (Buccafurri et al.,
1996; Buccafurri et al., 1999c) is in that context similar to the notion of an answer
set, once conflicts between rules have been eliminated. Thus, it can be seen as a
kind of answer set with implicit contradiction removal between knowledge bases.
While ordered logic does not provide weak negation, it can be simulated in the
language (Buccafurri et al., 1999c) (cf. also Section 6.8).
Observe that specific approaches for combining classical propositional and first-
order knowledge bases have been defined in (Baral et al., 1991b; Revesz, 1993; Baral
et al., 1991a; Liberatore & Schaerf, 1998), while (Baral et al., 1994) has addressed
the problem of combining default logic knowledge bases which must satisfy some
constraints. ELPs inherit this method by the correspondence between ELPs and
default logic theories. Roughly speaking, in terms of ELPs, the approach tries to
Using Methods of Declarative LP for Intel ligent Information Agents
41
retain as much as possible from the union of the sets of rules in the given programs
P1 , . . . , Pn such that consistency is retained and all given constraints are satisfied.
At the heart of the approach lies a contradiction removal method for conflicts that
arise between consistent default theories of equal priority. The authors also extend
their method to prioritized merging of default theories.
Axioms for arbitration operators in the context of classical propositional knowl-
edge bases have been addressed in (Revesz, 1993) and (Liberatore & Schaerf, 1998).
In (Revesz, 1993), arbitration is defined as a model-fitting operator, and it is shown
that the axiomatic description of arbitration operators is incompatible with revision
operators as in the AGM theory (Alchourr´on et al., 1985) and with update opera-
tors according to the theory of Katsuno and Mendelzon (1991). On the other hand,
in (Liberatore & Schaerf, 1998), axioms for arbitration operators are considered
such that, as argued, belief revision can be easily formulated through arbitration
but, conversely, a reformulation of arbitration in terms of belief revision is in general
much more complex.
We finally remark that, like for update and revision operators, no axiomatic
principles (except rather obvious ones) in the style of AGM revision theory have, to
the best of our knowledge, been discussed so far for arbitration operators between
logic programs. The discussion which properties such operators should enjoy is at
an early stage, and further research in this direction is needed.
8 Quantitative Information
Since van Emden’s pioneering paper (van Emden, 1986), the problem of expressing
quantitative information in logic programming has been addressed in several for-
malisms, depending on the meaning and the nature of the quantitative information
to be represented, and on the class of reasoning tasks to be performed.
Some formalisms allow the quantitative information to be associated with rules
(e.g., (Marek & Truszczy´nski, 1999)), others with rule heads or rule bodies (e.g.,
(Niemela et al., 1999; Lukasiewicz, 1998; Lukasiewicz, 2001)), or with single atoms
of the language (e.g., (Ng, 1997)), or with formulas (e.g., (Ng & Subrahmanian,
1992)). The information is sometimes expressed by means of a single value (e.g.,
(Marek & Truszczy´nski, 1999)), sometimes in the form of ranges or intervals of
values (e.g., (Ng & Subrahmanian, 1992; Ng & Subrahmanian, 1993; Niemela et al.,
1999; Lukasiewicz, 1998; Lukasiewicz, 2001)), maybe allowing variables to appear
in the definition of such intervals.
The point is that quantitative information has been considered for addressing dif-
ferent problems, which impose different requirements, and thus lead unsurprisingly
to different solutions and suggestions. Quantitative information can
• express preferences and priorities by means of an absolute value instead of an
order over rules;
• be related to the values and ranges assumed by numeric variables, or to prob-
ability distributions over domains;
• express costs and weights of performing some operations, thus implicitly en-
coding preference information;
42
Eiter et al.
• express uncertainty and probabilities of events or of beliefs;
• originate from statistical information, and be sub ject to further updates, and
so on.
We discuss here only a few approaches, which are quite different in their main
aims and theoretical foundations.
8.1 Disjunctive Programs with Weak Constraints by Buccafurri et al.
An enhancement of disjunctive logic programs by means of constraints is presented
in (Buccafurri et al., 2000). Two basic types of constraints are considered: strong
and weak constraints. Strong constraints must be satisfied like ordinary integrity
constraints in logic programming, whereas weak constraints are, if possible, to be
satisfied, and violations are taken into account when selecting the preferred models
of a program. Intuitively, the proposed semantics minimizes the number of instances
of weak constraints violated by a model. Moreover weak constraints may be assigned
different priorities, resulting in a ranking of the models of the program.
This extension of disjunctive programming is especially useful to encode planning
problems, combinatorial optimization problems, and abductive reasoning mecha-
nisms, since candidate solutions can be checked for mandatory properties through
strong constraints and then the best solutions can be chosen minimizing weak con-
straint violations.
8.2 Weight Constraint Rules by Niemela et al.
In (Niemela et al., 1999), logic programming is extended in order to allow quanti-
tative information in the body of rules. The extended rules are particularly useful
for representing combinatorial optimization problems, to encode constraints about
cardinality, costs and resources, which otherwise would be quite cumbersome or
impossible at all to encode in usual logic programming. The suggested semantics
is shown to generalize the stable model semantics for NLPs (Gelfond & Lifschitz,
1988).
In order to obtain the desired expressiveness, cardinality constraints are first in-
troduced, which are satisfied by any model such that the cardinality of the subset of
literals involved in the constraint and which are satisfied by the model lies between
the specified lower and upper bound for that constraint. This kind of constraint is
then generalized to the first-order case, and real-valued weights over variables are
introduced, giving the constraints a meaning of linear inequality. Finally, weight
rules involve in their body a set of such weight constraints over a set of variables.
8.3 Weighted Logic Programming by Marek and Truszczy´nski
In (Marek & Truszczy´nski, 1999), weighted logic programs are introduced. Each
rule in the program is assigned a non-negative real number called weight, which is
interpreted as the cost of applying that rule. Two different approaches for the com-
putation of costs are discussed, called reusability and no-reusability, respectively.
Using Methods of Declarative LP for Intel ligent Information Agents
43
In the reusability approach, the cost of applying rules for deriving an atom is paid
only once, and then the derived atom can be re-used as often as needed; in the
no-reusability approach, the derivation cost has to be paid each time an atom is
needed to be used in the body of a rule.
In this context, an interesting problem is the computation of the minimal cost
for deriving an atom or a set of atoms. In the no-reusability approach, for the
propositional case the minimal cost of a derivation can be computed by polynomial-
time shortest-path algorithms. In the reusability approach, computing the minimal
derivation cost is NP-hard.
8.4 Probabilistic Programs by Subrahmanian et al.
In a sequence of papers, Ng, Subrahmanian, and Dekhtyar develop a theory of
probabilistic logic programming (Ng & Subrahmanian, 1992; Ng & Subrahmanian,
1993; Ng & Subrahmanian, 1994; Dekhtyar & Subrahmanian, 2000), based on the
work about annotated logic (Subrahmanian, 1987; Kifer & Subrahmanian, 1992),
and on a possible-worlds semantics. Alternatively, probabilities can refer to the
elements in the domain, and represent statistical knowledge. We do not discuss this
approach here; for more information about this topic, cf., e.g., (Ng, 1997; Poole,
1993; Bacchus et al., 1993; Pearl, 1989).
In the possible-worlds approach, a probability range can be assigned to each
formula in the body and in the head of a rule, and can contain in its definition
variables ranging over [0, 1]. A suitable model theory and semantics is introduced
for negation-free probabilistic logic programs (Ng & Subrahmanian, 1993), which
is then extended to handle default reasoning under the stable model semantics (Ng
& Subrahmanian, 1994). This approach to probabilistic logic programming is used
in (Dix et al., 2000b) to extend the framework for agent programming presented
in (Subrahmanian et al., 2000), in order to express probabilistic information deriv-
ing from the lack of knowledge of the agent about its effective state. For example,
an agent can identify an ob ject in its environment with some degree of uncertainty,
which must be considered in its decision process. This idea is extended further
in (Dix et al., 2000a) in order to handle probabilistic beliefs: probabilities can be
assigned to agents’ beliefs about events, rather than to events themselves.
An alternative approach to probabilistic logic programming with possible-worlds
semantics is discussed in (Lukasiewicz, 1998; Lukasiewicz, 2001), where probabilistic
rules are interpreted as conditional probabilities. Program clauses are extended by
a subinterval of [0, 1], which describes the range for the conditional probability of
the head of the rule, given its body.
9 Temporal Reasoning
An important issue for information agents is temporal reasoning. Of particular
interest is reasoning about the temporal evolution of the agent’s knowledge base
and its behavior, depending on an underlying update policy or other principles.
This poses specific requirements both on the knowledge representation language,
44
Eiter et al.
which must allow the possibility to express temporal features, and on the reasoning
mechanism.
The topic may be considered from different perspectives. Temporal aspects are
relevant when considering agent actions and their effects, especially in conjunction
with beliefs, desires and intentions held by the agent. On the other hand, the same
temporal aspects need to be considered for reasoning about knowledge updates,
as described in Section 7. This approach has just been started with the definition
of LUPS (Alferes et al., 1999a), a language for declarative specifications of the
evolution of a knowledge base.
In the following subsections, we briefly discuss some works in the field of action
languages, temporal logics for BDI agents, the language LUPS, as well as providing
a pointer concerning how temporal logics and model checking techniques could be
useful in conjunction with LUPS for reasoning over updates. A full coverage of all
fields of research relevant to temporal reasoning and knowledge bases, however, is
beyond the scope of our paper. In particular, the following important research areas
are omitted in our review:
• dynamic logic (Harel, 1979);
• process logic, as described in (Harel et al., 1982), among others;
• event calculus (Kowalski & Sergot, 1986);
• situation calculus, tracing back to McCarthy and Hayes (Turner, 1997; Pinto
& Reiter, 1993; Pinto & Reiter, 1995), and its relationship with the event
calculus (Kowalski & Sadri, 1997);
• transaction logic and transaction logic programming (Bonner & Kifer, 1994);
• action logic (Pratt, 1990).
A comprehensive analysis of these and other related topics can be found in (Bon-
ner & Kifer, 1998).
9.1 Reasoning about Actions
The agent needs to reason about direct and indirect effects of its actions, possibly
involving several temporal units belonging to future time states. Effects may be
nondeterministic, requiring the agent to take concurrency (Baral & Gelfond, 1997)
and ramifications (Giunchiglia et al., 1997) into account. In this field, much work
has been devoted to describing and discussing the features of alternative action lan-
guages (Gelfond & Lifschitz, 1998; Lifschitz, 1999), together with the corresponding
reasoning mechanisms, combining them with temporal action logics or with the sit-
uation calculus (Giunchiglia & Lifschitz, 1999). A review of the relevant tasks for
reasoning agents in such dynamic domains can be found in (Baral & Gelfond, 2000).
Actions may be related to the update of a knowledge base in different ways.
An update process can be modeled as an action, which takes as input the update
information and, as an effect, incorporates this information into the knowledge
base. The difference, however, is that this action changes the intensional part of the
epistemic state, while usual actions only change the extensional part. Vice versa, a
proposal for expressing actions as updates was given in (Alferes et al., 1999b).
Using Methods of Declarative LP for Intel ligent Information Agents
45
9.2 Temporal Logics for BDI agents
Temporal aspects are addressed in BDI agents (belief, desire, intention ; see, e.g.,
(Georgeff & Rao, 1991)) by means of an extension of the branching-time temporal
logic CTL* (Emerson & Srinivasan, 1988). This approach is based on a possible-
worlds model, represented by a time tree with a single past and a branching-time
future. Branches in the time tree represent the different choices available to the
agent at each given point of time, as to which action is to be performed. A set of
belief-accessible worlds is associated with each state, namely the worlds that the
agent believes to be possible. Each belief-accessible world is in turn itself a time tree.
Multiple worlds result from the agent’s lack of knowledge about the actual state
of the world, and within each of these worlds the different futures represent the
choices available to the agent in selecting an action to perform. The same approach
is used for goal-accessible and intention-accessible worlds as well. In order to handle
these possible-worlds, CTL* is extended by introducing modal operators for beliefs,
goals, and intentions, and by adding first-order features.
For an extensive treatment of this sub ject, we refer the reader to Wooldridge’s
excellent monograph (Wooldridge, 2000).
9.3 LUPS, a Language for Specifying Updates
In (Alferes et al., 1999a), the language LUPS for specifying update sequences of
logic programs has been introduced. The need for such a language originates from
the consideration that dynamic logic programming describes how updates are to
be performed, but offers no possibility for making each update program in a given
sequence of programs depend on some specifications and conditions. Furthermore,
given the current epistemic state, how to (declaratively) specify the appropriate
updates to be performed? How to describe rules and laws for the update behavior
of an agent? LUPS statements specify how a logic program should be updated, de-
scribing which rules under what conditions should be incorporated into or retracted
from the knowledge base. In the class of all epistemic states of an agent, a LUPS
program constitutes a transition function from one state to another state.
The semantics of LUPS is formally defined by means of the semantics of a cor-
responding DynLP. A translation of a sets of LUPS statements into a single GLP,
written in a meta-language, is provided, so that the stable models of the resulting
program correspond to the intended semantics. Knowledge can be queried at any
epistemic state.
10 Evaluation
In the light of the general features of information agents, as identified in Section 2,
and of the problems and challenges discussed in Section 3, in the following we
evaluate the declarative methods reviewed in Sections 6 and 7. Criteria for the
evaluation are:
46
Eiter et al.
• what kind of problems and challenges for information agents a method can
address;
• how difficult it is to modify a method for slightly different purposes or policies,
to generalize or extend it, or to couple it with other standard methodologies;
• what limitations and bounds a method poses on the ob jects it processes;
• how near or similar to other formalisms a method is;
• which of the possible features identified for the tasks in Section 3 can be con-
sidered as desiderata over corresponding formalisms, and which are satisfied
by the method in question.
As far as temporal reasoning over evolution of updates is concerned, a possible
approach to this task is discussed in Section 11.
To keep track about the different approaches evaluated below, Tables 1 and 2
summarize the main features of each analyzed formalism.
10.1 Preference Hand ling
Some requirements or desiderata can be identified with respect to formalisms for
preference handling, which allow some evaluation of the methods we reviewed.
First of all, it must be possible to express priorities over single rules and facts as
well as over sets of them. This is required in order to select the information sources
to answer a given query, to account for the reliability of the sources, or to update
the knowledge base of the agent with some information coming from an external
source. The update can be coupled with priority information, e.g., recording the
accuracy or reliability of the new knowledge. Moreover, priority information makes
it possible to merge answers from different sources into a single query answer which
should be forwarded to the user, and gives a powerful instrument to remove possible
inconsistencies and to select the intended answers.
From the point of view of priority level (over rules or over sets of rules), the
approach of Section 6.3 is the only one which considers priorities over literals. This
seems not necessarily the best solution for information agents. It allows, however,
to indirectly express also preferences over rules by means of additional atoms and
corresponding rules.
If possible, preferences should be expressed in the ob ject language of the knowl-
edge base, by means of a naming function for rules and of a special symbol in the
language, so that priorities can be dynamically handled.
Dynamic priorities could be of interest for addressing several of the problems for
information agents. E.g., the reliability of an information source may not be fixed in
advance, but depend on the field of interest, or a query answer may depend on the
user profile. Consider, for example, the adaptive Web site of a clothing company,
which determines what items have to be shown depending not only on the explicit
input from the user (who is looking for brown shoes, say), but also on the average
value of the orders of the customer, and then shows the selected items ranked after
their price. By means of dynamic priorities, the items more likely to be interesting
for each customer can be better selected, and different ranking criteria may be
Using Methods of Declarative LP for Intel ligent Information Agents
47
Table 1. Approaches for preference handling
Approach
Section Reference
Features
Prioritized Defaults
6.1
(Gelfond & Son, 1998)
Preferred Answer Sets
6.2
(Brewka & Eiter, 1999)
Prioritized Programs
6.3
(Inoue & Sakama, 1996)
(Inoue & Sakama, 1999)
Prioritized Programs
6.4
(Foo & Zhang, 1997a)
Compiled Preferences
6.5
(Delgrande et al., 2000c)
Inheritance Programs
6.6
(Buccafurri et al., 1999a)
(Buccafurri et al., 1999b)
Preference Theories
6.7
(Brewka, 2000)
Dynamic priorities
over rules at the
ob ject level, plus
axiom rules
Static order over
rules, with ad hoc
semantics
Static priority
relation over literals,
with ad hoc
semantics
Static or dynamic
order over rules,
with ad hoc
semantics
Dynamic priorities
over rules at the
ob ject level, with
program rewriting
Static order over
rules, with program
rewriting
Dynamic priorities
over rules at the
ob ject level, with
ad hoc semantics
Ordered LPs
6.8
(Laenens & Vermeir, 1990) Dynamic priorities
over rules at the
(Gabbay et al., 1991)
(Buccafurri et al., 1999c)
ob ject level
adopted for different customers: e.g., customers usually buying expensive articles
may be presented with expensive items first (in order to maximize the profit, in case
the customer decides to buy something), while customers who in the past preferred
cheaper items may be presented a list in inverse order, to encourage them to buy.
Concerning the possibility of addressing preferences over preferences, methods
with explicit priorities in the knowledge base itself are to be preferred, like the ones
in Sections 6.1, 6.5, 6.7, and 6.8.
A mechanism is then needed to compute the intended set (or sets) of conclusions,
starting from the logic program expressing the epistemic state of the agent. Prior-
ities in the program can be thought of as constraints on the construction of this
48
Eiter et al.
set, and determine which of the conclusion sets (like stable models or answer sets)
is to be preferred. This is obviously useful to merge answers coming from different
sources, and to resolve possible inconsistencies. Observe, however, that preferences
must be used with care in the respective approaches. While the approach in Sec-
tion 6.8 implicitly supports removal of conflicts between contradicting information
from different information sources (viewed as sets of rules) of equal priority, such a
mechanism must in the other approaches be achieved through proper representation
of the knowledge, in order to avoid inconsistency.
Approaches for handling preferences that are based on standard semantics rather
than on ad hoc semantics offer some advantages:
1. they are easier to be compared with other approaches;
2. their behavior can be better understood;
3. they are easier to be integrated or extended to other frameworks.
Particularly useful are the methods based on some rewriting of the (prioritized)
logic program, since changes to the rewriting algorithm can implement other kinds
of preference handling in the same framework, see Sections 6.1, 6.5, and 6.6.
The policy for priority handling (determining, e.g., which rules are to be con-
sidered conflicting and which ones are to be given precedence) should preferably
be encoded explicitly in the ob ject language of the knowledge base, in the form of
rules about preferences. This is the case in the approaches based on a rewriting of
the program, as well as in the method of Section 6.7. This makes it possible to im-
plement in the same program more than one of such policies, to be used in different
situations. For instance, the way in which priorities are handled for determining a
user’s profile may be different from the way in which they are handled for choosing
the most promising information source.
It would also be useful to have the possibility of expressing several levels of
priorities. Consider for example the situation in which new information (in the
form of a single rule or a set of rules) comes from another agent and should be
incorporated in the current knowledge base: a priority value expresses the general
reliability value for that agent, and single values express, e.g., the reliability of each
rule with respect to the domain of interest.
This feature is also interesting for the application of learning procedures on the
basis of the history experienced by the agent. Different levels of priorities could un-
dergo different learning and updating policies, making lower levels easier to modify,
and requiring more striking evidence before changing priorities for the higher levels.
None of the methods we discussed offers such a possibility, but some of them could
be more easily extended than others, in particular the approach in Section 6.5, and
perhaps the one in Section 6.6.
As regards the flexibility of the approaches, some of them are interesting since
they allow to capture other formalisms as well, e.g., the approaches discussed in
Sections 6.3 and 6.5. An interesting aspect is the possibility of unifying preferences
and updates in a single framework. A tentative approach is reviewed in Section 7.6
and discussed in Section 10.3. Another approach consists in adding the update
knowledge to the current one in the form of a set of rules with higher priority.
Using Methods of Declarative LP for Intel ligent Information Agents
49
This is possible, e.g., in the approach reviewed in Section 6.1, as well as in the
one of Section 6.5. A first comparison of the latter one with the update policy of
Section 7.5 and (Eiter et al., 2000a) reveals that the semantics is different, but
further investigation is needed to elucidate the relation.
From this point of view, an interesting issue is addressed in (Inoue & Sakama,
1999), concerning a form of reasoning about preferences in which “sufficient” pref-
erences to derive intended conclusions are abductively selected. In the other for-
malisms we considered, no similar form of reasoning is suggested. This feature could
be particularly interesting from the point of view of learning.
Further improvements should also provide the possibility of associating quantita-
tive information with rules, in contrast to relative preference information. General
principles for preference handling (as in (Brewka & Eiter, 1999)) and particular
properties of the approaches as to the behavior of the corresponding agent should
be better analyzed. Interesting is also the problem of self-reference of the preference
relation, as pointed out in (Brewka, 2000), which has not been discussed in other
approaches.
10.2 Logic Programs with Quantitative Information
If quantitative information is to be incorporated into logic programs, the first ques-
tion to be answered is what kind of information it is, and where this information
should be added. Several possibilities have been suggested, each of them addressing
a different purpose and usage of quantitative information:
Rule level: Quantitative information may be associated with rules in the form of
a single number, expressing preference or priority by means of absolute values
instead of a relative order among rules. For example, this is the case in weighted
logic programming (Marek & Truszczy´nski, 1999) and also, to some extent, in
disjunctive logic programming with weak constraints (Buccafurri et al., 2000),
where a ranking of the constraints is possible. This feature can be useful if the
agent has to solve combinatorial problems (e.g., resource allocation).
Rule bodies: Another approach useful for combinatorial optimization consists in
associating quantitative information with the bodies of the rules, in the form of
a range identified by lower and upper bounds (Niemela et al., 1999).
Atom level: The quantitative information can be associated with atoms in the
form of a probability range, or with formulas. In case of non-atomic formulas,
the question arises as to how the probability ranges associated with the underly-
ing atomic formulas have to be combined, or, vice versa, decomposed. In differ-
ent application domains, there may be different dependency relationship between
atomic formulas. An interesting solution is discussed in (Dekhtyar & Subrahma-
nian, 2000), where the user is allowed defining and using different operators in the
same probabilistic logic program, to combine ranges, corresponding to the dif-
ferent dependency relationships which may exist between formulas (e.g., mutual
exclusion, independence, and so on). Probability information can be successfully
used, e.g., in problems of user profiling, in the tentative integration of incomplete
50
Eiter et al.
information, and in learning procedures for agents. An implementation of this
proposal is already possible in the IMPACT environment, as described in (Dix
et al., 2000b), whereas the behavior in practical domains of such probabilistic
agents has to be further investigated.
Conditional probability: If the quantitative information expresses a conditional
probability between the head and the body of a rule, the probability range is
to be associated with the head of rules, as suggested in (Lukasiewicz, 1998).
Applications of this approach in real-world domains should be studied, since it
could be interesting to solve the problem of guessing missing pieces of information
(e.g., in a user profile or in a query answer), or how to choose among conflicting
conclusions.
Most of these approaches require extensions of some standard semantics (often
the stable model semantics), and the integration of these features in other logic
programming frameworks seems to be difficult. For example, how should a sequence
of probabilistic logic programs or of weighted logic programs be updated? How can
strategies of inconsistency removal be defined, when the inconsistency between rules
resides in the quantitative information associated with them or with part of them?
Such questions need to be addressed, before information agents can take advantage
of these formalisms.
10.3 Revision and Update
The aspect of revision and update has to be addressed in order to make informa-
tion agents adapting their behavior over time, as external conditions change. For
example, when an information source is no longer available, when the contents of
a database or a user profile have been updated, or when a message from another
agent in the same system has been received, and so on.
A correct knowledge update can be relevant in order to decide which information
source has to be queried, or to modify an existing query plan to adapt it to run-time
information. It has a great impact also when implementing learning procedures for
the agent. In this case, changes and updates of the knowledge base are not only
determined by external evolution over time, but also by internal processing.
The problem for information agents is the following: new information in the form
of a single rule or fact, or in the form of a logic program, should be incorporated
in the current knowledge base. The new information may represent a temporal
evolution of the epistemic state, possibly originating from other agents who hold
more reliable beliefs, or coming from the environment and communicated changes
of content or availability of some database, or the information may be the result
of some learning procedure over past experience, and so on. When incorporating
this new knowledge into the current one, the consistency of knowledge and the
correctness of the agent’s behavior have to be further ensured.
The approaches which provide the possibility of updates by means of logic pro-
grams (cf. Sections 7.2, 7.5, and 7.7) are in our opinion to be preferred, since they
can be applied in case of a single rule or fact as well. Furthermore, assuming that
Using Methods of Declarative LP for Intel ligent Information Agents
51
Table 2. Approaches for revision and update
Approach
Section Reference
Features
Revision Programs
7.1
(Marek & Truszczy´nski, 1994) Revision of a set
of facts by a set
of rules
Update Programs
7.2
(Alferes & Pereira, 1997)
(Leite & Pereira, 1997)
Revision of a LP
by a LP, through
rewriting
Abductive Updates
7.3
(Inoue & Sakama, 1995)
(Inoue & Sakama, 1999)
Theory update
by a set of facts
Updates by
Priorities
7.4
(Foo & Zhang, 1997b)
(Foo & Zhang, 1998)
Dynamic Programs
7.5
(Alferes et al., 1998)
(Alferes et al., 2000a)
Updates plus
Preferences
Updates by
Inheritance
Revision of
Preference
Theories
7.6
(Alferes & Pereira, 2000)
7.7
(Buccafurri et al., 1999a)
7.8
(Brewka, 2000)
Arbitration
7.9
Different proposals
Update of a set
of facts or a LP
by a LP, through
rewriting
Update by a
(sequence of ) LPs,
through rewriting
Update by a
(sequence of ) LPs
plus preferences
Update by a
(sequence of ) LPs,
through rewriting
Revision by a
single rule/fact,
ad hoc semantics
Merging of two
knowledge bases
the knowledge base is represented by a logic program, no special language is re-
quired to specify updates, and the two logic programs can be used directly in the
update process.
In the update process, it should be possible to merge the current with the new
knowledge, usually assigning higher priority to the new one, thus solving possible
conflicts arising during the merge process. This is possible in all the approaches we
analyzed except arbitration (Section 7.9), which considers knowledge bases having
the same priority level.
It would also be useful to have the possibility of explicitly specifying which of the
knowledge bases to be merged has higher priority, since there may be application
domains in which old information is to be preserved against new one. Moreover, this
would permit to model not only temporal evolution, but also reciprocity of update
52
Eiter et al.
among agents. For example, information agents working on the same databases can
exchange information as to the current availability of the sources, their average
response-time, possible changes in their contents or available services, and so on.
This is not addressed by the methods we discussed, as they focus on sequential
updates instead of general priorities, and the common assumption is that the new
information must be preserved by the update, maybe rejecting some old beliefs. For
further considerations about possible handling of priorities, and about multi-level
priorities, cf. Section 10.1.
Like for priority handling, those revision approaches are to be preferred whose
update strategy is realized under a standard semantics, e.g. by means of a rewrit-
ing of the two logic programs (e.g. Section 7.5 and 7.7, together with the proposal
in (Eiter et al., 2000a)). The rewriting technique makes the update policy explicit
in the resulting knowledge base, and can be usually modified or extended in or-
der to encode other strategies or update mechanisms. This could be interesting in
order to differentiate the update process on the basis of the source of the update
information, which could be, e.g., another information agent, or one of the data
sources communicating changes in its content, or an internal learning procedure.
The approach would also permit, as outlined in Section 10.1, a uniform handling
of preferences and updates as a particular form of preference mechanism.
A tentative approach to such a uniform framework is the one described in Sec-
tion 7.6, which, however, suffers from some drawbacks, e.g., preference and update
information are treated separately in pairs of programs. It is evident that in many
application domains such a distinction is not possible. Consider an agent which
obtains from an information source some data, together with their reliability, and
the agent has to incorporate this new knowledge in its current epistemic state. In
this case, the update information cannot be simply separated from its “priority”,
and the update process has to consider both aspects (in the current knowledge base
this information may already be present, but with different reliability). Moreover,
in the proposed framework of Alferes and Pereira (2000), new information is always
supposed to have higher priority, and only in a second step the preference informa-
tion contained in the second program of each pair is considered to select among the
rules who “survived” the update. A more general approach would be useful.
Another method which could be used for a uniform handling of updates (as
explained in Section 7.7) and preferences is disjunctive logic programming with
inheritance (Buccafurri et al., 1999a), although it is originally conceived as a mech-
anism for representing inheritance hierarchies, thus not being appropriate for every
sort of preference relation (cf. the considerations in Section 10.1).
Other approaches addressing both updates and preferences have hitherto not
been proposed. A possibility would be to extend an approach such as the one in
Section 6.5, in order to handle updates as a special case of priority (as it actually is),
but the current version of the rewriting mechanism produces for updates not always
the intended results, compared with the approaches in Sections 7.5, 7.7 and (Eiter
et al., 2000a). Further investigation on this topic is required.
Another interesting issue lies in the possibility of updating the update (or revi-
sion) policy itself. This is only possible in the approach discussed in Section 7.8,
Using Methods of Declarative LP for Intel ligent Information Agents
53
in which the revision policy is declaratively encoded in the knowledge base. This
feature would be very interesting for learning agents.
Other interesting features which are relevant to the process of updating are the
following:
Inconsistency removal. Update methods are implicitly faced with the problem
of inconsistency removal, in order to resolve conflicts, usually in favor of the new
knowledge. Some methods address the topic explicitly, like the abductive updates
of Section 7.3. Others can result in an inconsistent knowledge base (like the
formalism of Section 7.5), and additional mechanisms or changes to the update
process are required to at least ensure that inconsistency is not propagated to
future updates. Ideally, an inconsistency removal strategy should be declaratively
encoded in the knowledge base, as it is realized for preference and update policies.
This allows of course different treatments of inconsistencies depending on their
level and relevance. For example, a conflict in the “core” part of the knowledge
base must always be resolved, while a conflict in a query answer for the user can
be communicated to the user as a form of incomplete information, if the agent is
unable to decide which of two data items is more reliable.
Retraction. As in theory revision, the problem of retracting a fact, a rule, or a set
of rules should be addressed too. In the approach of Section 7.8, this is partially
possible, if the retraction can be expressed in a “positive” form, while it is not
possible for general retraction. Explicit retraction of atoms is possible only in
the approaches of Sections 7.1, 7.2, and 7.3, whose languages allow the explicit
definition of the atoms to be inserted or to be true in the resulting belief set and
of the atoms to be deleted or false. For the other methods, this is not possible at
all. It is evident that this aspect has to be addressed for, e.g., learning procedures,
as well as for some forms of communication and update exchange between agents.
Minimality of change. The agent should be forced to give up as little as possible
of its past epistemic state when incorporating new knowledge. Some approaches
include in their basic definitions some notion of minimality (as the one in Sec-
tion 7.3), for others specializations of the update definition are proposed, which
express minimality requirements and are encoded in the update process (cf. (Eiter
et al., 2000a)). A common approach is to express minimality over rules instead
of atoms, establishing a sort of causal dependency among rules.
Rule overriding. Another feature on which further investigation is required is the
handling of overridden rules. In some approaches these rules are simply deleted,
while in others rules which are rejected by other rules having higher priority are
only “deactivated”. The rule is still present in the knowledge base but is inhib-
ited (Sections 7.5, 7.7, and (Eiter et al., 2000a)). The deactivated rule can also
be reactivated in future updates, if some newly added rule can in turn override
the rule that caused the initial rejection. A problem which should be addressed,
as suggested in (Brewka, 2000), is forgetting of rules. Since the knowledge base
of an actual agent has size constraints, rules cannot be simply added and further
deactivated or reactivated infinitely. For space reasons, some rules must sooner
or later be physically deleted, and suitable deletion criteria are needed, like the
54
Eiter et al.
number of updates since the last deactivation without further reactivation, or
some utility factor. Observe that this feature is related to the problem of forget-
ting data in temporal relational databases. Different methods have been proposed
for vacuuming relational databases, which is a non-trivial task.
Other improvements or extensions of the formalisms are clearly significant for
information agents. An interesting feature is the one suggested in (Alferes et al.,
2000b), namely the capability of performing updates along several dimensions (e.g.,
different time periods, different hierarchies) at the same time. This would especially
be interesting for combining it with the issue of nondeterminism in temporal evo-
lution, as indicated in Section 9.
A further interesting issue is the investigation of properties of update mechanisms.
The work in (Eiter et al., 2000a) is a first attempt in this direction.
10.4 Temporal Reasoning
Temporal reasoning for information agents can be viewed from different perspec-
tives, originating from different requirements, respectively. On the one hand, this
has been addressed in conjunction with actions and planning (Lifschitz, 1999;
Cimatti & Roveri, 1999). Reasoning tasks involve here the evolution of the worlds,
and have to face various aspects, including:
• the action choices available to the agent itself;
• their (possibly nondeterministic (Giunchiglia et al., 1997)) effects;
• the possible results of concurrent actions (Baral & Gelfond, 1997);
• uncertainties about the effective state of the world (Georgeff & Rao, 1991);
• events in the world not known to the agent in advance and possibly affecting
its decisions.
A number of action languages have been developed (cf. (Gelfond & Lifschitz,
1998; Lifschitz, 1999)), and their expressivity and the possibility of querying knowl-
edge at different stages of the evolution has been discussed.
On the other hand, LUPS (Alferes et al., 1999a), as the first language for spec-
ifying update transitions of logic programs, has opened a new interesting field of
temporal reasoning, namely reasoning over the evolution of the knowledge base.
Reasoning about LUPS specifications needs to consider two kinds of updates: up-
dates coming from the world (i.e., new information about a changing world), and
updates driven from the specifications of the agent, defining how they are to be inte-
grated. Both of them can be affected by nondeterminism and uncertainty, whereas
the current possibilities of LUPS do not include nondeterministic specifications of
updates, and the only uncertainties come from events in the world which are not
under the control of the agent. This and other issues need to be better investigated
in order to understand what the desiderata for a specification language for logic
program updates are, and whether LUPS needs to be extended for some special
purposes.
While approaches along the first point of view of temporal reasoning often em-
ploy the situation calculus, an interesting tool for the second approach could be
Using Methods of Declarative LP for Intel ligent Information Agents
55
found in temporal reasoning and in model checking (cf. (Cimatti et al., 1998;
Cimatti & Roveri, 1999)), choosing a branching-time logic similar to the BDI frame-
work (Georgeff & Rao, 1991). Advantages could also be drawn from further compar-
isons between different action languages and LUPS (or other specification languages
yet to come).
Another interesting feature is the possibility of reasoning over the past, not only
over the future. In BDI logic and in other approaches, several possible futures
are considered but only one past exists. While each logic program, in an update
sequence, is the result of a precise sequence of successive updates, reasoning over
the past could nevertheless be an interesting feature, e.g. when introducing learning
functionalities or utility measures over rules, or as a further element in decision
making.
11 Conclusion
In this paper, we tried to identify how declarative methods developed in the field
of logic programming could be successfully employed for developing advanced rea-
soning capabilities for information agents. We described the role such agents could
have in a multi-agent system for intelligent information access, together with the
tasks an information agent or a set of information agents is required to perform.
We shortly reviewed some existing systems and platforms addressing the topic
of information integration, mostly based on procedural rather than declarative ap-
proaches, and we discussed some interesting ongoing pro jects as well, in order to
have an overview of the most common problems and of some successful solutions
concerning the implementation of the required features.
As our main focus is on declarative methods in the field of logic programming,
we then discussed several declarative approaches for specific reasoning tasks for
information agents, namely tasks for preference handling, revision and update, ex-
pressing quantitative information, and reasoning over temporal evolution.
We then provided for each of these tasks a tentative evaluation of the declarative
formalisms we reviewed, trying to identify which of them are suitable to address
which particular problem information agents are faced with. On the basis of this
evaluation, some open problems and paths for future research can now be identified,
which we briefly discuss in the following.
Decision Making Based on Quantitative Information. The decision making capa-
bilities of information agents could be improved by encoding some form of quanti-
tative information into a logic programming representation, expressing either costs,
resource bounds, or uncertainty and probabilities. Many different formalisms and
semantics have been proposed in the field.
A first step should consist in a thorough analysis of each of these approaches, or of
a class of similar ones, in order to identify for which information problems they are
more suitable, in terms of reasoning capabilities for an agent, which requirements
they satisfy, and which drawbacks they present. In a second step, the theoretical
and practical issues concerning the integration of quantitative information with
56
Eiter et al.
other reasoning capabilities required for information agents (e.g., revision or update
management) should be addressed, investigating possible integrating frameworks for
a both qualitative and quantitative decision making process. A proposal, though
not strictly concerned with logic programming, is outlined in (Dix et al., 2000b).
Declarative Strategies for Information Merging. A promising application domain for
testing the decision making capabilities of information agents is the classification
and merging of the content of XML documents. This application domain, which is
relevant for many Web systems, would profit from the declarative specification of
strategies for information merging or classification of semi-structured data, and in
particular data stored in XML, which offers promising structural properties. This
issue has recently been addressed using techniques based on default logic (Hunter,
2000a; Hunter, 2000b), description logics (Calvanese et al., 1998; Calvanese et al.,
1999), ob ject oriented formalisms (Abiteboul et al., 1994), and automated theorem
proving and machine learning techniques (Thomas, 1998; Thomas, 1999).
The target is to express such strategies declaratively in the knowledge base of
the agent, depending on the available knowledge about the particular application
domain and on the data structure of the available sources, and employing the
declarative strategies for preference handling, which are already available in some
of the formalisms we discussed.
Unified Framework for Preference and Update Hand ling. In the light of the work in
the field of updates of logic programs, a homogeneous framework could be defined,
allowing to handle both updates and preferences in a uniform way. This would be
interesting for information agents, as far as it allows to address the problem of
updates with information coming from different information sources, together with
meta-information such as priority or reliability of the source itself. From this point
of view, the agent is faced with several updates, having to consider on the one hand
the temporal sequence (since new updates have priority over old information), and
on the other hand the relative priorities of the sources or of the information they
provide. Work on this issue has already been done in (Foo & Zhang, 1998; Alferes
& Pereira, 2000; Alferes et al., 2000b), partly based on the approach suggested in
(Brewka & Eiter, 1999), but further research is needed.
Temporal Reasoning over Know ledge Base Evolution. An interesting open issue is
reasoning about the evolution of the knowledge base, drawing conclusions over se-
quences of updates. Based on the declarative specifications guiding the evolution,
e.g., in the form of LUPS programs (Alferes et al., 1999a) or some other spec-
ification language for logic program updates, whose desired features have to be
carefully defined, the target is to perform temporal reasoning over the evolution
of the knowledge base, answering questions about the possibility of reaching some
desired condition or violating some intended constraint.
These reasoning capabilities naturally rely on temporal logics (Emerson, 1996),
and could take advantage of techniques and tools developed in the field of model
checking (Clarke et al., 2000) or planning (Cimatti et al., 1998; Cimatti & Roveri,
Using Methods of Declarative LP for Intel ligent Information Agents
57
1999). Features allowing nondeterminism in the specifications, e.g., if the evolution
depends on external events not known to the agent in advance or the language
itself allows nondeterministic updates, could enrich the representation, also taking
advantage of similar work in the field of action languages.
The goal of reasoning tasks over the knowledge base, and the specification of its
evolution, is to know whether the agent behaves correctly, and how the epistemic
state of the agent changes over time if uncertain classes of updates (e.g., possibly
leading to states violating some constraint) are going to be performed.
Acknowledgments
We would like to thank the anonymous referees for their constructive comments
which helped to improve this paper. This work was partially supported by the
Austrian Science Fund (FWF) under grants P13871-INF, P14781-INF, and Z29-
INF.
References
Abiteboul, S., Christophides, V., Cluet, S., & Scholl, M. (1994). From Structured Doc-
uments to Novel Query Facilities. Pages 313–324 of: Snodgrass, R., & Winslett, M.
(eds), Proc. 1994 ACM SIGMOD Int. Conf. on Management of Data. ACM Press.
Abiteboul, S., McHugh, J., Quass, D., Widom, J., & Goldman, R. (1997a). Lore: A
Database Management System for Semistructured Data. SIGMOD Record, 26(3), 54–
66.
Abiteboul, S., McHugh, J., Quass, D., Widom, J., & Wiener, J. (1997b). The Lorel Query
Language for Semistructured Data. International Journal on Digital Libraries, 1(1),
68–88.
Adali, S., & Emery, R. (1995). A Uniform Framework for Integrating Knowledge in
Heterogeneous Knowledge Systems. Pages 513–521 of: Chen, L., & Yu, P. (eds), Proc.
11th Int. Conf. on Data Engineering (ICDE’95 ). IEEE Computer Society.
Adali, S., & Subrahmanian, V.S. (1996). Amalgamating Knowledge Bases, III: Algorithms,
Data Structures, and Query Processing. J. of Logic Programming, 28(1), 45–88.
Adali, S., Brink, A., Emery, R., Lu, J., Ra jput, A., Rogers, T., Ross, R., Subrahmanian,
V.S., & Ward, C. (1995). HERMES: A Heterogeneous Reasoning and Mediator System.
To appear.
Alchourr´on, C., Gardenfors, P., & Makinson, D. (1985). On the Logic of Theory Change:
Partial Meet Functions for Contraction and Revision. J. of Symbolic Logic, 50, 510–530.
Alferes, J., & Pereira, L. (1997). Update-Programs Can Update Programs. Pages 110–131
of: Dix, J., Pereira, L., & Przymusinski, T. (eds), Proc. of Non-Monotonic Extensions
of Logic Programming (NMELP’96 ). Lecture Notes in Computer Science, vol. 1216.
Springer.
Alferes, J., & Pereira, L. (2000). Updates plus Preferences. Pages 345–360 of: Aciego,
M., de Guzm´an, I., Brewka, G., & Pereira, L. (eds), Proc. 7th European Workshop on
Logics in Artificial Intel ligence (JELIA 2000 ). Lecture Notes in Computer Science, vol.
1919. Springer.
Alferes, J., Leite, J., Pereira, L., Przymusinska, H., & Przymusinski, T. (1998). Dynamic
Logic Programming. Pages 98–111 of: Cohn, A., Schubert, L., & Shapiro, S. (eds),
Proc. 6th Int. Conf. on Principles of Know ledge Representation and Reasoning (KR’98 ).
Morgan Kaufmann.
58
Eiter et al.
Alferes, J., Pereira, L., Przymusinska, H., & Przymusinski, T. (1999a). LUPS - A Language
for Updating Logic Programs. Pages 162–176 of: Gelfond, M., Leone, N., & Pfeifer,
G. (eds), Proc. 5th Int. Conf. on Logic Programming and Nonmonotonic Reasoning
(LPNMR’99 ). Lecture Notes in Computer Science, vol. 1730. Springer.
Alferes, J., Pereira, L., Przymusinska, H., Przymusinski, T., & Quaresma, P. (1999b).
Preliminary Exploration on Actions as Updates. Meo, M., & Vilares-Ferro, M. (eds),
Proc. 1999 Joint Conf. on Declarative Programming (AGP’99 ).
Alferes, J., Leite, J., Pereira, L., Przymusinska, H., & Przymusinski, T. (2000a). Dynamic
Updates of Non-Monotonic Knowledge Bases. J. of Logic Programming, 45(1–3), 43–70.
Alferes, J., Leite, J., & Pereira, L. (2000b). Multi-dimensional Dynamic Logic Program-
ming. Pages 17–26 of: Sadri, F., & Satoh, K. (eds), Proc. CL-2000 Workshop on Com-
putational Logic in Multi-Agent Systems (CLIMA 2000 ).
Analyti, A., & Pramanik, S. (1995). Reliable Semantics for Extended Logic Programs
with Rule Prioritization. J. of Logic and Computation, 5(3), 303–324.
Arens, Y. (1990). Services and Information Management for Decision Support. Proc.
Annual AI Systems in Government Conf. (AISIG’90 ).
Arens, Y., & Knoblock, C. (1992). Planning and Reformulating Queries for Semantically-
Modeled Multidatabase Systems. Proc. 1st Int. Conf. on Information and Know ledge
Managements.
Arens, Y., Chee, C., Hsu, C., & Knoblock, C. (1993). Retrieving and Integrating Data
from Multiple Information Sources. Int. J. of Cooperative Information Systems, 2(2),
127–158.
Arens, Y., Knoblock, C., & Shen, W. (1996). Query Reformulation for Dynamic Informa-
tion Integration. J. of Intel ligent Information Systems, 6(2–3), 99–130.
Arisha, K., Eiter, T., Kraus, S., Ozcan, F., Ross, R., & Subrahmanian, V.S. (1999). IM-
PACT: A Platform for Collaborating Agents. IEEE Intel ligent Systems, 14(2), 64–72.
Baader, F., & Hollunder, B. (1995). Priorities on Defaults with Prerequisites, and their
Application in Treating Specificity in Terminological Default Logic. J. of Automated
Reasoning, 15(1), 41–68.
Bacchus, F., Grove, A.J., Halpern, J.Y., & Koller, D. (1993). Statistical Foundations for
Default Reasoning. Pages 563–569 of: Ba jcsy, R. (ed), Proc. 13th Int. Joint Conf. on
Artificial Intel ligence (IJCAI’93 ). Morgan Kaufmann.
Baral, C., & Gelfond, M. (1997). Reasoning About Effects of Concurrent Actions. J. of
Logic Programming, 31(1–3), 85–117.
Baral, C. & Gelfond, M. (2000). Reasoning Agents in Dynamic Domains. Pages 257–279
of: Minker, J. (ed), Logic-Based Artificial Intel ligence. Kluwer Academic Publishers.
Baral, C., Kraus, S., Minker, J., & V.S. Subrahmanian. (1991a). Combining Knowledge
Bases Consisting of First Order Theories. Pages 92–101 of: Ras, Z., & Zemankova,
M. (eds), Proc. 6th Int. Symp. on Methodologies for Intel ligent Systems (ISMIS’91 ).
Lecture Notes in Computer Science, vol. 542. Springer.
Baral, C., Kraus, S., & Minker, J. (1991b). Combining Multiple Knowledge Bases. IEEE
Trans. on Know ledge and Data Engineering, 3(2), 208–220.
Baral, C., Kraus, S., & Minker, J. (1994). Combining Default Logic Databases. Int. J. of
Intel ligent and Cooperative Information Systems, 3(3), 319–348.
Bargmeyer, B., Fowler, J., Nodine, M., & Perry, B. (1999). Agent-Based Semantic Inter-
operability in InfoSleuth. SIGMOD Record, 28(1), 60–67.
Bayardo, R., Bohrer, B., Brice, R., Cichocki, A., Fowler, J., Helal, A., Kashyap, V.,
Ksiezyk, T., Martin, G., Nodine, M., Rashid, M., Rusinkiewicz, M., Shea, R., Un-
nikrishnan, C., Unruh, A., & Woelk, D. (1997).
InfoSleuth: Semantic Integration of
Using Methods of Declarative LP for Intel ligent Information Agents
59
Information in Open and Dynamic Environments (Experience Paper). Pages 195–206
of: Peckham, J. (ed), Proc. ACM SIGMOD Int. Conf. on Management of Data. ACM
Press.
Bohrer, W., Ngu, A., & Nodine, M. (1999). Semantic Brokering over Dynamic Hetero-
geneous Data Sources in InfoSleuth. Pages 358–365 of: Proc. 15th Int. Conf. on Data
Engineering (ICDE’99 ). IEEE Computer Society Press.
Bondarenko, A., Dung, P.M., Kowalski, R., & Toni, F. (1997). An Abstract Argumen-
tation-Theoretic Approach to Default Reasoning. Artificial Intel ligence, 93, 63–101.
Bonner, A., & Kifer, M. (1994). An Overview of Transaction Logic. Theoretical Computer
Science, 133(2), 205–265.
Bonner, A., & Kifer, M. (1998). The State of Change: A Survey. Pages 1–36 of: Freitag,
B., Decker, H., Kifer, M., & Voronkov, A. (eds), Transactions and Change in Logic
Databases, International Seminar on Logic Databases and the Meaning of Change. Lec-
ture Notes in Computer Science, vol. 1472. Springer.
Bressan, S., & Goh, C. (1998). Answering Queries in Context. Pages 68–82 of: Andreasen,
T., Christiansen, H., & Legind Larsen, H. (eds), Proc. 3rd Int. Conf. on Flexible Query
Answering Systems (FQAS’98 ). Lecture Notes in Computer Science, vol. 1495. Springer.
Bressan, S., Goh, C., Fynn, K., Jakobisiak, M., Hussein, K., Kon, H., Lee, T., Madnick, S.,
Pena, T., Qu, J., , Shum, A., & Siegel, M. (1997). The COntext INterchange Mediator
Prototype. Pages 525–527 of: Peckham, J. (ed), Proc. ACM SIGMOD Int. Conf. on
Management of Data. ACM Press.
Brewka, G. (1994). Adding Priorities and Specificity to Default Logic. Pages 247–260
of: MacNish, C., Pearce, D., & Pereira, L. (eds), Proc. Europ. Workshop on Logics
in Artificial Intel ligence (JELIA’94 ). Lecture Notes in Computer Science, vol. 838.
Springer.
Brewka, G. (1996). Well-Founded Semantics for Extended Logic Programs with Dynamic
Preferences. J. of Artificial Intel ligence Research, 4, 19–36.
Brewka, G. (2000). Declarative Representation of Revision Strategies. Pages 18–22 of:
Proc. 14th European Conf. on Artificial Intel ligence (ECAI 2000 ).
Brewka, G., & Eiter, T. (1999). Preferred Answer Sets for Extended Logic Programs.
Artificial Intel ligence, 109(1–2), 297–356.
Brewka, G., & Eiter, T. (2000). Prioritizing Default Logic. Pages 27–45 of: Holldobler, S.
(ed), Intel lectics and Computational Logic – Papers in Honor of Wolfgang Bibel. Kluwer
Academic.
Buccafurri, F., Leone, N., & Rullo, P. (1996). Stable Models and their Computation for
Logic Programming with Inheritance and True Negation. J. of Logic Programming,
27(1), 5–43.
Buccafurri, F., Faber, W., & Leone, N. (1999a). Disjunctive Logic Programs with In-
heritance. Tech. rept. DBAI-TR-99-30. Institut fur Informationssysteme, Technische
Universitat Wien, Austria. To appear in Theory and Practice of Logic Programming.
Buccafurri, F., Faber, W., & Leone, N. (1999b). Disjunctive Logic Programs with Inheri-
tance. Pages 79–93 of: Schreye, D. De (ed), Proc. 16th Int. Conf. on Logic Programming
(ICLP’99 ). MIT Press.
Buccafurri, F., Leone, N., & Rullo, P. (1999c). Semantics and Expressiveness of Disjunctive
Ordered Logic. Ann. of Math. and Artificial Intel ligence, 25(3–4), 311–337.
Buccafurri, F., Leone, N., & Rullo, P. (2000). Enhancing Disjunctive Datalog by Con-
straints. IEEE Trans. on Know ledge and Data Engineering, 12(5), 845–860.
Calvanese, D., De Giacomo, G., & Lenzerini, M. (1998). What can Knowledge Repre-
sentation do for Semi-Structured Data? Pages 205–210 of: Proc. 15th Nat. Conf. on
60
Eiter et al.
Artificial Intel ligence and 10th Innovative Applications of Artificial Intel ligence Conf.
(AAAI/IAAI 1998 ). AAAI Press/MIT Press.
Calvanese, D., De Giacomo, G., & Lenzerini, M. (1999). Representing and Reasoning on
XML Documents: A Description Logic Approach. J. of Logic and Computation, 9(3),
295–318.
Cannata, P., Huhns, M., Jacobs, N., Ksiezyk, T., Ong, K., Sheth, A., Singh, M., Tomlinson,
C., & Woelk, D. (1997). The Carnot Heterogeneous Database Pro ject: Implemented
Applications. Distributed and Paral lel Databases, 5(2), 207–225.
Cimatti, A., & Roveri, M. (1999). Conformant Planning via Model Checking. Pages 21–
34 of: Biundo, S., & Fox, M. (eds), Recent Advances in AI Planning, Proc. 5th Europ.
Conf. on Planning (ECP’99 ). Lecture Notes in Computer Science, vol. 1809. Springer.
Cimatti, A., Roveri, M., & Traverso, P. (1998). Strong Planning in Non-Deterministic
Domains via Model Checking. Pages 36–43 of: Simmons, R., Veloso, M., & Smith, S.
(eds), Proc. 4th Int. Conf. on Artificial Intel ligence Planning Systems (AIPS’98 ). AAAI
Press.
Clarke, E.M., Grumberg, O., & Peled, D. (2000). Model Checking. MIT Press.
Cohen, W. (1998a). Integration of Heterogeneous Databases Without Common Domains
Using Queries Based on Textual Similarity. Pages 201–212 of: Haas, L., & Tiwary,
A. (eds), Proc. ACM SIGMOD Int. Conf. on Management of Data (SIGMOD 1998 ).
ACM Press.
Cohen, W.W. (1998b). A Web-Based Information System that Reasons with Structured
Collections of Text. Pages 400–407 of: Sycara, K., & Wooldridge, M. (eds), Proc. 2nd
Int. Conf. on Autonomous Agents. ACM Press.
Decker, K., Sycara, K., & Williamson, M. (1997). Middle-Agents for the Internet. Pages
578–583 of: Proc. 15th Int. Joint Conf. on Artificial Intel ligence (IJCAI’97 ), vol. 1.
Morgan Kaufmann.
Dekhtyar, A., & Subrahmanian, V.S. (2000). Hybrid Probabilistic Programs. J. of Logic
Programming, 43(3), 187–250.
Delgrande, J., & Schaub, T. (2000). Expressing Preferences in Default Logic. Artificial
Intel ligence, 123(1–2), 41–87.
Delgrande, J., Schaub, T., & Tompits, H. (2000a). A Compilation of Brewka and Eiter’s
Approach to Prioritization. Pages 376–390 of: Aciego, M., de Guzm`an, I., Brewka,
G., & Pereira, L. (eds), Proc. 7th Europ. Workshop on Logics in Artificial Intel ligence
(JELIA 2000 ). Lecture Notes in Computer Science, vol. 1919. Springer.
Delgrande, J., Schaub, T., & Tompits, H. (2000b). A Compiler for Ordered Logic Pro-
grams. Baral, C., & Truszczynski, M. (eds), Proc. 8th Int. Workshop on Non-Monotonic
Reasoning (NMR 2000 ).
Delgrande, J., Schaub, T., & Tompits, H. (2000c). Logic Programs with Compiled Prefer-
ence. Pages 392–398 of: Horn, W. (ed), Proc. 14th Europ. Conf. on Artificial Intel ligence
(ECAI 2000 ). IOS Press.
Dell’Acqua, P., & Pereira, L. (1999a). Preferring and Updating with Multi-Agents. Draft.
Dell’Acqua, P., & Pereira, L. (1999b). Updating Agents. Rochefort, S., Sadri, F., & Toni,
F. (eds), Proc. ICLP’99 Workshop on Multi-Agent Systems in Logic (MASL’99 ).
Dell’Acqua, P., Sadri, F., & Toni, F. (1998). Combining Introspection and Communica-
tion with Rationality and Reactivity in Agents. Pages 17–32 of: Dix, J., del Cerro,
L., & Furbach, U. (eds), Proc. European Workshop on Logics in Artificial Intel ligence
(JELIA’98 ). Lecture Notes in Computer Science, vol. 1489. Springer.
Dimopoulos, Y., & Kakas, A. (2000). Information Integration and Computational Logic.
Using Methods of Declarative LP for Intel ligent Information Agents
61
Computational Logic, Special Issue on the Future Technological Roadmap of Compulog-
Net. URL: http://www.compulog.org/net/Forum/Supportdocs.html.
Dix, J. (1996). Semantics of Logic Programs: Their Intuitions and Formal Properties.
An Overview. Pages 241–327 of: Fuhrmann, A., & Rott, H. (eds), Logic, Action and
Information - Essays on Logic in Philosophy and Artificial Intel ligence. DeGruyter.
Dix, J., Pick, G., & Subrahmanian, V.S. (2000a). Meta Agent Programs. J. of Logic
Programming, 46(1–2), 1–60.
Dix, J., Nanni, M., & Subrahmanian, V.S. (2000b). Probabilistic Agent Programs. ACM
Trans. on Computational Logic, 1(2), 208–246.
Dung, P.M. (1993). On the Acceptability of Arguments and its Fundamental Role in
Nonmonotonic Reasoning and Logic Programming. Pages 852–859 of: Ba jcsy, R. (ed),
Proc. 13th Int. Joint Conf. on Artificial Intel ligence (IJCAI’93 ). Morgan Kaufmann.
Duschka, O., & Genesereth, M. (1997). Query Planning in Infomaster. Pages 109–111 of:
Proc. 12th Annual ACM Symposium on Applied Computing (SAC’97 ).
Eiter, T., & Subrahmanian, V.S. (1999). Heterogeneous Active Agents, II: Algorithms
and Complexity. Artificial Intel ligence, 108(1–2), 257–307.
Eiter, T., Pick, G., & Subrahmanian, V.S. (1999). Heterogeneous Active Agents, I: Se-
mantics. Artificial Intel ligence, 108(1–2), 179–255.
Eiter, T., Fink, M., Sabbatini, G., & Tompits, H. (2000a). Considerations on Updates
of Logic Programs. Pages 2–20 of: Aciego, M., de Guzm´an, I., Brewka, G., & Pereira,
L. (eds), Proc. 7th Europ. Workshop on Logic in Artificial Intel ligence (JELIA 2000 ).
Lecture Notes in Artificial Intelligence, vol. 1919. Springer.
Eiter, T., Fink, M., Sabbatini, G., & Tompits, H. (2000b). On Updates of Logic Programs:
Semantics and Properties. Tech. rept. INFSYS RR-1843-00-08. Technische Universitat
Wien, Inst. fur Informationssysteme, Abt. Wissensbasierte Systeme.
Emerson, E. (1996). Automated Temporal Reasoning about Reactive Systems. Pages
41–101 of: Moller, F. & Birtwistle, G. (eds), Logics for Concurrency: Structure versus
Automata. Lecture Notes in Computer Science, vol. 1043. Springer.
Emerson, E., & Srinivasan, J. (1988). Branching Time Temporal Logic. Pages 123–172 of:
de Bakker, J., de Roever, W., & Rozenberg, G. (eds), Linear Time, Branching Time and
Partial Order in Logics and Models for Concurrency (REX Workshop 1988 ). Lecture
Notes in Computer Science, vol. 354. Springer.
Finin, T., Fritzson, R., McKay, D., & McEntire, R. (1994). KQML as an Agent Communi-
cation Language. Pages 456–463 of: Proc. 3rd Int. Conf. on Information and Know ledge
Management (CIKM’94 ). ACM.
Flach, P., & Kakas, A. (2000). Abduction and Induction - Essays on Their Relation and
Integration. Kluwer Academic.
Flores-Mendez, R. (1999). Towards a Standardization of Multi-Agent System Frameworks.
ACM Crossroads, 5(4).
Foo, N., & Zhang, Y. (1997a). Answer Sets for Prioritized Logic Programs. Pages 69–83
of: Maluszynski, J. (ed), Proc. Int. Logic Programming Symp. (ILPS’97 ). MIT Press.
Foo, N., & Zhang, Y. (1997b). Towards Generalized Rule-based Updates. Pages 82–88
of: Proc. 15th Int. Joint Conf. on Artificial Intel ligence (IJCAI’97 ), vol. 1. Morgan
Kaufmann.
Foo, N., & Zhang, Y. (1998). Updating Logic Programs. Pages 403–407 of: Prade, H.
(ed), Proc. 13th Europ. Conf. on Artificial Intel ligence (ECAI’98 ). Wiley.
Frohlich, P., Nejdl, W., Schroeder, M., Dam´asio, C., & Pereira, L. (1999). Using Extended
Logic Programming for Alarm-Correlation in Cellular Phone Networks. Pages 343–352
of: Imam, I., Kodratoff, Y., El-Dessouki, A., & Ali, M. (eds), Multiple Approaches to
62
Eiter et al.
Intel ligent Systems, Proc. 12th Int. Conf. on Industrial and Engineering Applications of
Artificial Intel ligence and Expert Systems (IEA/AIE’99 ). Lecture Notes in Computer
Science, vol. 1611. Springer.
Gabbay, D., Laenens, E., & Vermeir, D. (1991). Credulous vs. Sceptical Semantics for
Ordered Logic Programs. Pages 208–217 of: Allen, J., Fikes, R., & Sandewall, E.
(eds), Proc. 2nd Int. Conf. on Principles of Know ledge Representation and Reasoning
(KR’91 ). Morgan Kaufmann.
Garcia-Molina, H., Hammer, J., Ireland, K., Papakonstantinou, Y., Ullman, J., & Widom,
J. (1995). Integrating and Accessing Heterogeneous Information Sources in TSIMMIS.
Pages 61–64 of: Proc. AAAI Symp. on Information Gathering.
Garcia-Molina, H., Papakonstantinou, Y., Quass, D., Ra jaraman, A., Sagiv, Y., Ullman,
J., Vassalos, V., & Widom, J. (1997). The TSIMMIS Approach to Mediation: Data
Models and Languages. J. of Intel ligent Information Systems, 8(2), 117–132.
Gardenfors, P., & Rott, H. (1995). Belief Revision. Pages 35–132 of: Gabbay, D., Hogger,
C., & Robinson, J. (eds), Handbook of Logic in Artificial Intel ligence and Logic Program-
ming, Volume IV: Epistemic and Temporal Reasoning. Oxford Science Publications.
Gelfond, M., & Lifschitz, V. (1988). The Stable Model Semantics for Logic Programming.
Pages 1070–1080 of: Kowalski, R., & Bowen, K. (eds), Logic Programming, Proc. 5th
Int. Conf. and Symp. (ICLP/SLP’88 ). MIT Press.
Gelfond, M., & Lifschitz, V. (1991). Classical Negation in Logic Programs and Disjunctive
Databases. New Generation Computing, 9(3–4), 365–386.
Gelfond, M., & Lifschitz, V. (1998). Action Languages. Electronic Trans. on AI, 3(16).
Gelfond, M., & Son, T. (1998). Reasoning with Prioritized Defaults. Pages 164–223 of:
Dix, J., Pereira, L., & Przymusinski, T. (eds), Logic Programming and Know ledge Repre-
sentation, 3rd Int. Workshop, Selected Papers (LPKR’97 ). Lecture Notes in Computer
Science, vol. 1471. Springer.
Genesereth, M. (1991). Knowledge Interchange Format. Pages 599–600 of: Allen, J.,
Fikes, R., & Sandewall, E. (eds), Proc. 2nd Int. Conf. on Principles of Know ledge
Representation and Reasoning (KR’91 ). Morgan Kaufmann.
Genesereth, M., Keller, A., & Duschka, O. (1997). Infomaster: An Information Integration
System. Pages 539–542 of: Peckham, J. (ed), Proc. ACM SIGMOD Int. Conf. on
Management of Data. ACM Press.
Georgeff, M., & Rao, A. (1991). Modeling Rational Agents within a BDI-architecture.
Pages 473–484 of: Allen, J., Fikes, R., & Sandewall, E. (eds), Proc. 2nd Int. Conf. on
Know ledge Representation and Reasoning (KR’91 ). Morgan Kaufmann.
Giunchiglia, E., & Lifschitz, V. (1999). Action Languages, Temporal Action Logics and the
Situation Calculus. Working Notes IJCAI’99 Workshop on Nonmonotonic Reasoning,
Action, and Change.
Giunchiglia, E., Kartha, G., & Lifschitz, V. (1997). Representing Action: Indeterminacy
and Ramifications. Artificial Intel ligence, 95(2), 409–438.
Harel, D. (1979). First-Order Dynamic Logic. Lecture Notes in Computer Science, vol.
68. Springer.
Harel, D., Kozen, D., & Parikh, R. (1982). Process Logic: Expressiveness, Decidability,
Completeness. J. of Computer and System Sciences, 25(2), 144–170.
Huhns, M., & Singh, M. (1992). The Semantic Integration of Information Models. AAAI
Workshop on Cooperation among Heterogeneous Intel ligent Agents.
Huhns, M., & Singh, M. (1994). Automating Workflows for Service Provisioning: Inte-
grating AI and Database Technologies. IEEE Expert, 9(5).
Using Methods of Declarative LP for Intel ligent Information Agents
63
Hunter, A. (2000a). Merging Potentially Inconsistent Items of Structured Text. IEEE
Trans. on Know ledge and Data Engineering, 34(3), 305–332.
Hunter, A. (2000b). Reasoning with Inconsistency in Structured Text. Know ledge Engi-
neering Review, 15(2), 1–21.
Inoue, K., & Sakama, C. (1994). On the Equivalence between Disjunctive and Abductive
Logic Programs. Pages 489–503 of: Van Hentenryck, P. (ed), Proc. 11th Int. Conf. on
Logic Programming (ICLP’94 ). MIT Press.
Inoue, K., & Sakama, C. (1995). Abductive Framework for Nonmonotonic Theory Change.
Pages 204–210 of: Proc. 14th Int. Joint Conf. on Artificial Intel ligence (IJCAI’95 ).
Morgan Kaufmann.
Inoue, K., & Sakama, C. (1996). Representing Priorities in Logic Programs. Pages
82–96 of: Maher, M. (ed), Proc. Joint Int. Conf. and Symp. on Logic Programming
(JICSLP’96 ). MIT Press.
Inoue, K., & Sakama, C. (1999). Updating Extended Logic Programs through Abduction.
Pages 147–161 of: Gelfond, M., Leone, N., & Pfeifer, G. (eds), Proc. 5th Int. Conf.
on Logic Programming and Nonmonotonic Reasoning (LPNMR’99 ). Lecture Notes in
Artificial Intelligence, vol. 1730. Springer.
Inoue, K., & Sakama, C. (2000a). Abductive Logic Programming and Disjunctive Logic
Programming: Their Relationship and Transferability. J. of Logic Programming, 44(1–
3), 75–100.
Inoue, K., & Sakama, C. (2000b). Prioritized Logic Programming and Its Applications to
Commonsense Reasoning. Artificial Intel ligence, 123(1–2), 185–222.
Jennings, N., Noriega, P., Parsons, S., & Sierra, C. (1997). A Framework for Argumen-
tation-Based Negotiation. Pages 177–192 of: Singh, M., Rao, A., & Wooldridge, M.
(eds), Intel ligent Agents IV, Proc. 4th Int. Workshop on Agent Theories, Architectures,
and Languages (ATAL’97 ). Lecture Notes in Artificial Intelligence, vol. 1365. Springer.
Kakas, A., & Riguzzi, F. (1997). Learning with Abduction. Pages 181–188 of: Lavrac,
N., & Dzeroski, S. (eds), Proc. 7th Int. Workshop on Inductive Logic Programming
(ILP’97 ). Lecture Notes in Computer Science, vol. 1297. Springer.
Kakas, A., Kowalski, R., & Toni, F. (1992). Abductive Logic Programming. J. of Logic
and Computation, 2(6), 719–770.
Kakas, A., Kowalski, R., & F.Toni. (1998). The Role of Abduction in Logic Programming.
Pages 235–324 of: Gabbay, D., Hogger, C., & Robinson, J. (eds), Handbook of Logic in
Artificial Intel ligence and Logic Programming, Volume V: Logic Programming. Oxford
University Press.
Kakas, A., Miller, R., & F.Toni. (1999). An Argumentation Framework of Reasoning
about Actions and Change. Pages 78–91 of: Gelfond, M., Leone, N., & Pfeifer, G. (eds),
Proc. 5th Int. Conf. on Logic Programming and Nonmonotonic Reasoning (LPNMR’99 ).
Lecture Notes in Computer Science, vol. 1730. Springer.
Katsuno, H., & Mendelzon, A. (1991). On the Difference between Updating a Knowledge
Base and Revising It. Pages 387–394 of: Allen, J., Fikes, R., & Sandewall, E. (eds), Proc.
2nd Int. Conf. on Principles of Know ledge Representation and Reasoning (KR’91 ).
Morgan Kaufmann.
Kifer, M., & Subrahmanian, V.S. (1992). Theory of Generalized Annotated Logic Pro-
gramming and its Applications. J. of Logic Programming, 12(3–4), 335–367.
Knoblock, C. (1995). Planning, Executing, Sensing and Replanning for Information
Gathering. Pages 1686–1693 of: Proc. 14th Int. Joint Conf. on Artificial Intel ligence
(IJCAI’95 ). Morgan Kaufmann.
Kowalski, R., & Sadri, F. (1991). Logic Programs with Exceptions. New Generation
Computing, 9(3–4), 387–400.
64
Eiter et al.
Kowalski, R., & Sadri, F. (1996). Towards a Unified Agent Architecture that Combines
Rationality with Reactivity. Pages 137–149 of: Pedreschi, D., & Zaniolo, C. (eds), Logic
in Databases, Proc. Int. Workshop (LID’96 ). Lecture Notes in Computer Science, vol.
1154. Springer.
Kowalski, R., & Sadri, F. (1997). Reconciling the Event Calculus with the Situation
Calculus. J. of Logic Programming, 31(1–3), 39–58.
Kowalski, R., & Sadri, F. (1999). From Logic Programming Towards Multi-Agent Systems.
Ann. of Math. and Artificial Intel ligence, 25(3–4), 391–419.
Kowalski, R., & Sergot, M. (1986). A Logic-based Calculus of Events. New Generation
Computing, 4(1), 67–95.
Kowalski, R., & Toni, F. (1995). Reduction of Abductive Logic Programs to Normal
Logic Programs. Pages 367–381 of: Sterling, L. (ed), Proc. 12th Int. Conf. on Logic
Programming (ICLP’95 ). MIT Press.
Kowalski, R., & Toni, F. (1996). Abstract Argumentation. Artificial Intel ligence and Law,
4(3–4), 275–296.
Laenens, E., & Vermeir, D. (1990). A Fixpoint Semantics for Ordered Logic. J. of Logic
and Computation, 1(2), 159–185.
Lamma, E., Mello, P., Milano, M., & Riguzzi, F. (1999). Integrating Induction and Ab-
duction in Logic Programming. Information Sciences, 116(1), 25–54.
Lamma, E., Riguzzi, F., & Pereira, L. (2000). Strategies in Combined Learning via Logic
Programs. Machine Learning, 38(1–2), 63–87.
Leite, J., & Pereira, L. (1997). Generalizing Updates: from Models to Programs. Pages
224–246 of: Dix, J., Pereira, L., & Przymusinski, T. (eds), Logic Programming and
Know ledge Representation, Selected Papers 3rd Int. Workshop (LPKR’97 ). Lecture
Notes in Artificial Intelligence, vol. 1471. Springer.
Levy, A., & Rousset, M. (1996). CARIN: A Representation Language Combining Horn
Rules and Description Logics. Pages 323–327 of: Wahlster, W. (ed), Proc. 12th Europ.
Conf. on Artificial Intel ligence (ECAI’96 ). Wiley.
Levy, A., & Weld, D. (2000). Intelligent Internet Systems. Artificial Intel ligence, 118(1–2),
1–14.
Levy, A., Srivastava, D., & Kirk, T. (1995). Data Model and Query Evaluation in Global
Information Systems. J. of Intel ligent Information Systems, 5(2), 121–143.
Levy, A., Ra jaraman, A., & Ordille, J. (1996a). Query-Answering Algorithms for Infor-
mation Agents. Pages 40–47 of: Proc. 13th Nat. Conf. on Artificial Intel ligence and 8th
Innovative Applications of Artificial Intel ligence Conf. (AAAI/IAAI’96 ), vol. 1. AAAI
Press/MIT Press.
Levy, A., Ra jaraman, A., & Ordille, J. (1996b). Querying Heterogeneous Information
Sources Using Source Descriptions. Pages 251–262 of: Vijayaraman, T., Buchmann,
A., Mohan, C., & Sarda, N. (eds), Proc. 22th Int. Conf. on Very Large Data Bases
(VLDB’96 ). Morgan Kaufmann.
Liberatore, P., & Schaerf, M. (1998). Arbitration (or How to Merge Knowledge Bases).
IEEE Trans. on Know ledge and Data Engineering, 10(1), 76–90.
Lifschitz, V. (1999). Action Languages, Answer Sets and Planning. Pages 357–373 of:
Apt, K., Marek, V., Truszczy´nski, M., & Warren, D. (eds), The Logic Programming
Paradigm: A 25-Year Perspective. Springer.
Lifschitz, V., & Woo, T. (1992). Answer Sets in General Nonmonotonic Reasoning (Prelim-
inary Report). Pages 603–614 of: Nebel, B., Rich, C., & Swartout, W. (eds), Proc. 3rd
Int. Conf. on Principles of Know ledge Representation and Reasoning (KR’92 ). Morgan
Kaufmann.
Using Methods of Declarative LP for Intel ligent Information Agents
65
Lukasiewicz, T. (1998). Probabilistic Logic Programming. Pages 388–392 of: Prade, H.
(ed), Proc. 13th Europ. Conf. on Artificial Intel ligence (ECAI’98 ). Wiley.
Lukasiewicz, T. (2001). Probabilistic Logic Programming with Conditional Constraints.
ACM Trans. on Computational Logic. To appear. Also available as: Tech. rept. INFSYS-
TR 1843-00-01, Institute of Information Systems, TU Vienna, 2000.
MacGregor, R. (1991). Inside the LOOM Description Classifier. SIGART Bul letin, 2(3),
88–92.
Marek, V., & Truszczy´nski, M. (1993). Nonmonotonic Logics - Context-Dependent Rea-
soning. Springer.
Marek, V., & Truszczy´nski, M. (1994). Revision Specifications by Means of Programs.
Pages 122–136 of: MacNish, C., Pearce, D., & Pereira, L. (eds), Proc. Europ. Workshop
on Logics in Artificial Intel ligence (JELIA’94 ). Lecture Notes in Artificial Intelligence,
vol. 838. Springer.
Marek, V., & Truszczy´nski, M. (1999). Logic Programming with Costs. Unpublished
manuscript.
Moore, R. (1985). Semantical Considerations on Nonmonotonic Logics. Artificial Intel li-
gence, 25(1), 75–94.
Nebel, B. (1998). How Hard is it to Revise a Belief Base? Pages 77–145 of: Dubois,
D., & Prade, H. (eds), Handbook on defeasible reasoning and uncertainty management
systems, volume iii: Belief change. Kluwer Academic.
Ng, R. (1997). Semantics, Consistency, and Query Processing of Empirical Deductive
Databases. IEEE Trans. on Know ledge and Data Engineering, 9(1), 32–49.
Ng, R., & Subrahmanian, V.S. (1992). Probabilistic Logic Programming. Information and
Computation, 101(2), 150–201.
Ng, R., & Subrahmanian, V.S. (1993). A Semantical Framework for Supporting Sub jective
and Conditional Probabilities in Deductive Databases. J. of Automated Reasoning,
10(2), 191–235.
Ng, R., & Subrahmanian, V.S. (1994). Stable Semantics for Probabilistic Deductive
Databases. Information and Computation, 110(1), 42–83.
Niemela, I., Simons, P., & Soininen, T. (1999). Stable Model Semantics of Weight Con-
straint Rules. Pages 317–331 of: Gelfond, M., Leone, N., & Pfeifer, G. (eds), Proc. 5th
Int. Conf. on Logic Programming and Nonmonotonic Reasoning (LPNMR’99 ). Lecture
Notes in Computer Science, vol. 1730. Springer.
Nute, D. (1994). Defeasible Logic. Pages 353–395 of: Gabbay, D., Hogger, C., & Robinson,
J. (eds), Handbook of Logic in Artificial Intel ligence and Logic Programming, Volume
III. Clarendon Press.
Pearl, J. (1989). Probabilistic Semantics for Nonmonotonic Reasoning: A Survey. Pages
505–516 of: Brachman, R., Levesque, H., & Reiter, R. (eds), Proc. 1st Int. Conf. on
Principles of Know ledge Representation and Reasoning (KR’89 ). Morgan Kaufmann.
Pinto, J., & Reiter, R. (1993). Temporal Reasoning in Logic Programming: A Case Study
for the Situation Calculus. Pages 203–221 of: Warren, D. (ed), Proc. 10th Int. Conf.
on Logic Programming (ICLP’93 ). MIT Press.
Pinto, J., & Reiter, R. (1995). Reasoning about Time in the Situation Calculus. Ann. of
Math. and Artificial Intel ligence, 14(2–4), 251–268.
Poole, D. (1993). Logic Programming, Abduction and Probability - A Top-Down Anytime
Algorithm for Estimating Prior and Posterior Probabilities. New Generation Computing,
11(3), 377–400.
Pradhan, S., & Minker, J. (1996). Using Priorities to Combine Knowledge Bases. Int. J.
of Cooperative Intel ligent Systems, 5(2–3), 333–364.
66
Eiter et al.
Prakken, H., & Sartor, G. (1997). Argument-based Extended Logic Programming with
Defeasible Priorities. J. of Applied Non-Classical Logics, 7, 25–75.
Prakken, H., & Sartor, G. (1999). A System for Defeasible Argumentation, with Defeasible
Priorities. Pages 365–379 of: Wooldridge, M., & Veloso, M. (eds), Artificial Intel ligence
Today: Recent Trend and Developments. Lecture Notes in Computer Science, vol. 1600.
Springer.
Pratt, V. (1990). Action Logic and Pure Induction. Pages 97–120 of: van Eijck, J. (ed),
Proc. Europ. Workshop of Logics in Artificial Intel ligence (JELIA’90 ). Lecture Notes
in Computer Science, vol. 478. Springer.
Przymusinski, T., & Turner, H. (1997). Update by Means of Inference Rules. J. of Logic
Programming, 30(2), 125–143.
Quass, D., Ra jaraman, A., Sagiv, Y., Ullman, J., & Widom, J. (1995). Querying Semistruc-
tured Heterogeneous Information. Pages 319–344 of: Ling, T., Mendelzon, A., & Vieille,
L. (eds), Proc. 4th Int. Conf. on Deductive and Object-Oriented Databases (DOOD’95 ).
Lecture Notes in Computer Science, vol. 1013. Springer.
Ramakrishnan, I., Rao, P., Sagonas, K., Swift, T., & Warren, D. (1995). Efficient Tabling
Mechanisms for Logic Programs. Pages 697–711 of: Sterling, L. (ed), Proc. 12th Int.
Conf. on Logic Programming (ICLP’95 ). MIT Press.
Revesz, P. (1993). On the Semantics of Theory Change: Arbitration between Old and New
Information. Pages 71–82 of: Proc. 12th ACM SIGACT SIGMOD-SIGART Symp. on
Principles of Database Systems (PODS’93 ). ACM Press.
Rintanen, J. (1998). Lexicographic Priorities in Default Logic. Artificial Intel ligence,
106(2), 221–265.
Sadri, F., & Toni, F. (2000). Computational Logic and Multi-Agent Systems: a Roadmap.
Computational Logic, Special Issue on the Future Technological Roadmap of Compulog-
Net. URL: http://www.compulog.org/net/Forum/Supportdocs.html.
Sagonas, K., Swift, T., & Warren, D. (1994). XSB as an Efficient Deductive Database
Engine. Pages 442–453 of: Snodgrass, R., & Winslett, M. (eds), Proc. 1994 ACM
SIGMOD Int. Conf. on Management of Data. ACM Press.
Sakama, C. (2000). Abductive Generalization and Specialization. Pages 253–265 of:
Flach, P., & Kakas, A. (eds), Abduction and Induction - Essays on their Relation and
Integration. Kluwer Academic.
Subrahmanian, V.S. (1987). On the Semantics of Quantitative Logic Programs. Pages
173–182 of: Proc. 4th IEEE Symp. on Logic Programming (SLP’87 ). Computer Society
Press.
Subrahmanian, V.S. (1994). Amalgamating Knowledge Bases. ACM Trans. on Database
Systems, 19(2), 291–331.
Subrahmanian, V.S., Dix, J., Eiter, T., Kraus, S., Ozcan, F., Ross, R., & Bonatti, P.
(2000). Heterogeneous Agent Systems. MIT Press.
Thomas, B. (1998).
Intelligent Web Querying with Logic Programs. Proc. Workshop
on Inference Mechanisms in Know ledge-based Systems at the Nat. German AI Conf.
(KI’98 ).
Thomas, B. (1999). Logic Programs for Intelligent Web Search. Pages 190–198 of: Ras,
Z., & Skowron, A. (eds), Proc. 11th Int. Symp. on Foundations of Intel ligent Systems
(ISMIS’99 ). Lecture Notes in Computer Science, vol. 1609. Springer.
Turner, H. (1997). Representing Actions in Logic Programs and Default Theories: A
Situation Calculus Approach. J. of Logic Programming, 31(1–3), 245–298.
van Emden, M. (1986). Quantitative Deduction and its Fixpoint Theory. J. of Logic
Programming, 3(1), 37–53.
Using Methods of Declarative LP for Intel ligent Information Agents
67
Wiederhold, G. (1992). Mediators in the Architecture of Future Information Systems.
IEEE Computer, 25(3), 38–49.
Wooldridge, M. (2000). Reasoning about Rational Agents. MIT Press.
|
1707.07399 | 2 | 1707 | 2017-08-18T01:44:18 | Learning for Multi-robot Cooperation in Partially Observable Stochastic Environments with Macro-actions | [
"cs.MA",
"cs.LG",
"cs.RO"
] | This paper presents a data-driven approach for multi-robot coordination in partially-observable domains based on Decentralized Partially Observable Markov Decision Processes (Dec-POMDPs) and macro-actions (MAs). Dec-POMDPs provide a general framework for cooperative sequential decision making under uncertainty and MAs allow temporally extended and asynchronous action execution. To date, most methods assume the underlying Dec-POMDP model is known a priori or a full simulator is available during planning time. Previous methods which aim to address these issues suffer from local optimality and sensitivity to initial conditions. Additionally, few hardware demonstrations involving a large team of heterogeneous robots and with long planning horizons exist. This work addresses these gaps by proposing an iterative sampling based Expectation-Maximization algorithm (iSEM) to learn polices using only trajectory data containing observations, MAs, and rewards. Our experiments show the algorithm is able to achieve better solution quality than the state-of-the-art learning-based methods. We implement two variants of multi-robot Search and Rescue (SAR) domains (with and without obstacles) on hardware to demonstrate the learned policies can effectively control a team of distributed robots to cooperate in a partially observable stochastic environment. | cs.MA | cs | Learning for Multi-robot Cooperation in Partially Observable
Stochastic Environments with Macro-actions
Miao Liu1, Kavinayan Sivakumar2, Shayegan Omidshafiei3, Christopher Amato4 and Jonathan P. How3
7
1
0
2
g
u
A
8
1
]
A
M
.
s
c
[
2
v
9
9
3
7
0
.
7
0
7
1
:
v
i
X
r
a
Abstract- This paper presents a data-driven approach for
multi-robot coordination in partially-observable domains based
on Decentralized Partially Observable Markov Decision Pro-
cesses (Dec-POMDPs) and macro-actions (MAs). Dec-POMDPs
provide a general framework for cooperative sequential decision
making under uncertainty and MAs allow temporally extended
and asynchronous action execution. To date, most methods
assume the underlying Dec-POMDP model is known a priori
or a full simulator is available during planning time. Previous
methods which aim to address these issues suffer from local
optimality and sensitivity to initial conditions. Additionally, few
hardware demonstrations involving a large team of heteroge-
neous robots and with long planning horizons exist. This work
addresses these gaps by proposing an iterative sampling based
Expectation-Maximization algorithm (iSEM) to learn polices
using only trajectory data containing observations, MAs, and
rewards. Our experiments show the algorithm is able to achieve
better solution quality than the state-of-the-art learning-based
methods. We implement two variants of multi-robot Search
and Rescue (SAR) domains (with and without obstacles) on
hardware to demonstrate the learned policies can effectively
control a team of distributed robots to cooperate in a partially
observable stochastic environment.
I. INTRODUCTION
There has been significant progress in recent years on
developing cooperative multi-robot systems that can operate
in real-world environments with uncertainty. Example ap-
plications of social and economical interest include search
and rescue (SAR) [1], traffic management for smart cities
[2], planetary navigation [3], robot soccer [4], e-commerce
and transport logistic processes [5]. Planning in such en-
vironments must address numerous challenges,
including
imperfect models and knowledge of the environment, re-
stricted communications between robots, noisy and limited
sensors, different viewpoints by each robot, asynchronous
calculations, and computational limitations.
These planning problems, in the most general form, can be
formulated as a Decentralized Partially Observable Markov
Decision Process (Dec-POMDP) [6], a general framework
for cooperative sequential decision making under uncertainty.
In Dec-POMDPs, robots make decisions based on local
streams of information (i.e., observations), such that
the
expected value of the team (e.g., number of victims rescued,
1Miao Liu is with IBM T. J. Watson Research Center, Yorktown Heights,
NY, USA [email protected]
2Kavinayan Sivakumar is with the Department of Electrical Engineering,
Princeton University, Princeton, NJ, USA [email protected]
3 Shayegan Omidshafiei and Jonathan P. How are with Laboratory of
Information and Decision Systems, Massachusetts Institute of Technology,
Cambridge, MA, USA {shayegan, jhow}@mit.edu
4Christopher Amato
and
Information Science, Northeastern University, Boston, MA, USA
[email protected]
of Computer
the College
is with
average customer satisfaction) is maximized. However, rep-
resenting and solving Dec-POMDPs is often intractable for
large domains, because finding the optimal (even epsilon-
approximate) solution of a Dec-POMDP (even for finite
horizon) is NEXP-complete [6]. To combat this issue, re-
cent research has addressed the more scalable macro-action
based Dec-POMDP (MacDec-POMDP), where each agent
has temporally-extended actions, which may require different
amounts of time to complete [7]. Moreover, significant
progress has been made on demonstrating the usefulness
of MacDec-POMDPs via a range of challenging robotics
problems, such as a warehouse domain [8], bartending and
beverage service [9], and package delivery [10], [11]. How-
ever, current MacDec-POMDP methods require knowing
domain models a priori. Unfortunately, for many real-world
problems, such as SAR,
the domain model may not be
completely available. Recently, researchers started to address
this issue via reinforcement learning and proposed a policy-
based EM algorithm (PoEM) [12], which can learn valid
controllers via only trajectory data containing observations,
macro-actions (MAs), and rewards.
Although PoEM has convergence guarantees for the batch
learning setting and can recover optimal policies for bench-
mark problems with sufficient data, it suffers from local
optimality and sensitivity to initial conditions for compli-
cated real-word problems. Inevitably, as an EM type algo-
rithm, the results of PoEM can be arbitrarily poor given
bad initialization. Additionally, few hardware demonstrations
based on challenging tasks such as SAR, which involves a
large team of heterogeneous robots (both ground vehicles
and aerial vehicles) and with MacDec-POMDP formulation
exists. This paper addresses these gaps by proposing an it-
erative sampling-based Expectation-Maximization algorithm
(iSEM) to learn polices. Specifically,
this paper extends
previous approaches by using concurrent (multi-threaded)
EM iterations providing feedback to one another to enable
re-sampling of parameters and reallocation of computational
resources for threads which are clearly converging to poor
values.
The algorithm is tested in batch learning settings, which
is commonly used in learning from demonstration. Through
theoretical analysis and numerical comparisons on a large
multi-robot SAR domain, we demonstrate the new algorithm
can better explore the policy space. As a result,
iSEM
is able to achieve better expected values compared to the
state-of-the-art learning-based method, PoEM. Finally, we
present an implementation of two variants of multi-robot
SAR domains (with and without obstacles) on hardware
to demonstrate the learned policies can effectively control
a team of distributed robots to cooperate in a partially
observable stochastic environment.
II. BACKGROUND
We first discuss the background on Dec-POMDPs and
MacDec-POMDPs and then describe the PoEM algorithm.
A. Dec-POMDPs and MacDec-POMDPs
POMDPs
Decentralized
(Dec-POMDPs)
generalize
POMDPs to the multiagent, decentralized setting [6], [13].
Multiple agents operate under uncertainty based on partial
views of the world, with execution unfolding over a bounded
or unbounded number of steps. At each step, every agent
chooses an action (in parallel) based on locally observable
information and then receives a new observation. The agents
share a joint reward based on their joint concurrent actions,
making the problem cooperative. However, agents' local
views mean that execution is decentralized.
Formally, a Dec-POMDP is represented as an octuple
(cid:104)N, A, S, Z, T, Ω, R, γ(cid:105), where N is a finite set of agent
indices; A = ⊗nAn and Z = ⊗nZn respectively are sets of
joint actions and observations, with An and Zn available to
agent n. At each step, a joint action (cid:126)a = (a1,··· , aN) ∈ A
is selected and a joint observation (cid:126)z = (z1,··· , zN) is
received; S is a set of finite world states; T : S × A × S →
[0, 1] is the state transition function with T (s(cid:48)s, (cid:126)a) denoting
the probability of transitioning to s(cid:48) after taking joint action (cid:126)a
in s; Ω : S × A× Z → [0, 1] is the observation function with
Ω((cid:126)zs(cid:48), (cid:126)a) the probability of observing (cid:126)o after taking joint
action (cid:126)a and arriving in state s(cid:48); R : S×A → R is the reward
function with r(s, (cid:126)a) the immediate reward received after
taking joint action (cid:126)a in s; γ ∈ [0, 1) is a discount factor. Be-
cause each agent lacks access to other agents' observations,
each agent maintains a local policy πn, defined as a mapping
from local observation histories to actions. A joint policy
consists of the local policies of all agents. For an infinite-
horizon Dec-POMDP with initial state s0, the objective is
to find a joint policy π = ⊗nπn, such that the value of π
maximized. Specifically, given ht = {a0:t−1, z0:t} ∈ Hn,
the history of actions and observations up to t, the policy πn
probabilistically maps ht to at: Hn × An → [0, 1].
starting from s0, V π(s0) = E(cid:2)(cid:80)∞
t=0 γtr(st, (cid:126)at)s0, π(cid:3), is
A MacDec-POMDP with (local) macro-actions extends
the MDP-based options [14] framework to Dec-POMDPs.
Formally, a MacDec-POMDP is defined as a tuple
(cid:104)N, A, M, S, Z, O, T, Ω, R, γ(cid:105), where N, A, S, Z, T, Ω, R
and γ are the same as defined in the Dec-POMDP; O =
⊗On are sets of joint macro-action observations which are
functions of the state; M = ⊗Mn are sets of joint macro-
n ⊂ H M
actions, with Mn = (cid:104)I m
is
the initiation set that depends on macro-action observation
n} ∈ H M
n ,
histories, defined as hM
n : S → [0, 1] is a stochastic termination condition that
βm
n : Hn×Mn → [0, 1]
depends on the underlying states, and πm
is an option policy for macro-action m (Hn is the space of
history of primitive-action and observation). Macro-actions
are natural representations for robot or human operation for
completing a task (e.g., navigating to a way point or placing
n (cid:105), where I m
n,··· , ot−1
n , βm
n,t = {o0
n , mt
n , πm
n, m1
n
an object on a robot). MacDec-POMDPs can be thought of
as decentralized partially observable semi-Markov decision
processes (Dec-POSMDPs) [9], [10], because it is important
to consider the amount of time that may pass before a macro-
action is completed. The high level policy for each agent
Ψn, can be defined for choosing macro-actions that depends
on macro-action observation histories. Given a joint policy,
the primitive action at each step is determined by the high-
level policy that chooses the MA, and the MA policy that
chooses the primitive action. The joint high level policies
and macro-action policies can be evaluated as: V Ψ(s0) =
t=0 γtr(st, (cid:126)at)s0, π, Ψ(cid:3) 1.
E(cid:2)(cid:80)∞
B. Solution Representation
A Finite State Controller (FSC) is a compact way to
represent a policy as a mapping from histories to actions.
Formally, a stochastic FSC for agent n is defined as a tuple
Θn = (cid:104)Qn, Mn, On, δn, λn, µn(cid:105), where, Qn is the set of
nodes2; Mn and On are the output and input alphabets
(i.e., the macro-action chosen and the observation seen);
δn : Qn × On × Qn → [0, 1]
is the node transition
n : Qn × Mn →
probability, i.e., δn(q, o, q(cid:48)) = Pr(q(cid:48)q, o); λ0
is the output probability for node qn,0, such that
[0, 1]
n(qn,0, mn,0) = Pr(mn,0qn,0); λn : Qn × On ×
mn,0 ∼ λ0
Mn → [0, 1] is the output probability for nodes (cid:54)= qn,0
that associates output symbols with transitions, i.e. mn,τ ∼
λn(qn,τ , on,τ , mn,τ ) = Pr(mn,τqn,τ , on,τ ); µ : Qn →
[0, 1] is the initial node distribution qn,0 ∼ µn = Pr(qn,0).
This type of FSC is called a Mealy machine [16], where an
agent's local policy for action selection λn(q, o, m) depends
on both current controller node (an abstraction of history)
and immediate observation. By conditioning action selections
on immediate observations, a Mealy machine can use this
observable information to help ensure a valid macro-action
controller is constructed [12].
C. Policy Learning Through EM
A Dec-POMDP problem can be transformed into an
inference problem and then efficiently solved by an EM
algorithm. Previous EM methods [17], [18] have achieved
success in scaling to larger problems, but these methods
require a Dec-POMDP model both to construct a Bayes net
and to evaluate policies. When the exact model parameters
T , Ω and R are unknown, a Reinforcement Learning (RL)
problem must be solved instead. To this end, EM has been
adapted to model-free RL settings to optimize FSCs for Dec-
POMDPs [19], [20] and MacDec-POMDPs [12].
For both purposes of self-containment and ease of an-
alyzing new algorithm, we first review the policy based
EM algorithm (PoEM) developed for the MacDec-POMDP
case [12].
1Note that MacDec-POMDPs allows asynchronous decision making, so
synchronization issues must be dealt with by the solver as part of the
optimization. Some temporal constraints (e.g., timeouts) can be encoded
into the termination condition of a macro-action.
2A controller node can be understood as a decision state (summary of
history). They are commonly used for policy representation when solving
infinite horizon POMDPs [15] and Dec-POMDPs [6].
0, rk
0, (cid:126)mk
0 ,··· (cid:126)ok
Definition 1: (Global
empirical value
)}K
function) Let
D(K) = {((cid:126)ok
k=1 be a set of
episodes resulting from N agents who choose macro-
actions according to Ψ=⊗nΨn, a set of arbitrary stochastic
policies with pΨn (mh) > 0, ∀ action m, ∀ history h. The
global empirical value function is defined as
, (cid:126)mk
Tk
, rk
Tk
Tk
V(cid:0)D(K); Θ(cid:1)def.
=
K(cid:88)
1
K
N(cid:89)
Tk(cid:88)
n,1:t), 0≤ γ < 1 is the discount.
p(mk
pΨn (mk
n,t, Θn)
n,t)
n,0:thk
n,0:thk
γtrk
t
n=1
(1)
k=1
t=0
n,0:t−1, ok
D. PoEM
n,τhk
n,t = (mk
quences {(cid:126)q k
by Jensen's inequality):
where hk
Definition 1 provides an off-policy learning objective:
given data D(K) generated from a set of behavior policies Ψ,
find a set of parameters Θ ={Θi}N
is maximized. Here, we assume a factorized policy rep-
resentation p( (cid:126)mk
n,τ , Θn) to
n=1 p(mk
accommodate decentralized policy execution.
i=1 such that V(cid:0)D(K);Θ(cid:1)
0:τ(cid:126)h1:τ , Θ) = (cid:81)N
0:t : k = 1 . . . , K, t = 1 : Tk} and maximize
Direct maximization of V(cid:0)D(K); Θ(cid:1) is difficult; instead,
V(cid:0)D(K); Θ(cid:1) can be augmented with controller node se-
the lower bound of the logarithm of V(cid:0)D(K); Θ(cid:1) (obtained
ln V(cid:0)D(K); Θ(cid:1) = ln (cid:80)
0:t(cid:101)Θ)rk
0:t(cid:101)Θ)
≥ (cid:80)
t p( (cid:126)mk
0:t(cid:101)Θ) ln r k
= lb(Θ(cid:101)Θ),
f k
t ((cid:126)q k
0:t(cid:101)Θ)
0:t(cid:126)o k
0:t(cid:101)Θ)
1:t,(cid:101)Θ)/ V (D(K);(cid:101)Θ),
0:t(cid:101)Θ) ≥ 0} satisfy the normalization con-
0:t(cid:126)o k
(cid:80)
straint (cid:80)K
(cid:80)Tk
0:t(cid:101)Θ) = K with (cid:101)Θ the
rmin)/(cid:81)t
t −
τ ),∀t, k are reweighted rewards with
τhk
rmin denoting the minimum reward, leading to the following
0:t;(cid:101)Θ(cid:1)(cid:9),Θ
max(cid:8)f k
(cid:0)(cid:126)q k
lb(Θ(cid:101)Θ)
constrained optimization problem
(cid:80)Tk
subject to:(cid:80)K
(cid:80)Q1:N
0:t;(cid:101)Θ) =(cid:81)N
0:t;(cid:101)Θ) = K,
n,0:t,(cid:101)Θn). (3)
t ((cid:126)q k
n,0:to k
recent estimate of Θ, and rk
t
where f k
t ((cid:126)q k
and {f ((cid:126)q k
n,0:t=1 f k
n,0:t, q k
t=0
q k
n=1 p(m k
t p( (cid:126)mk
0:t,(cid:126)q k
f k
t ((cid:126)q k
def.
= γt(rk
τ =0 pΨ( (cid:126)mk
def.
= rk
t p( (cid:126)mk
t ((cid:126)q k
f k
t ((cid:126)q k
f k
0:t, (cid:126)q k
0:t(cid:126)o k
0:t(cid:126)q k
p( (cid:126)m k
most
k=1
t=0
1:t,Θ)
def.
k,t,(cid:126)q k
0:t
f k
t ((cid:126)q k
0:t,(cid:126)q k
1:t,Θ)
k,t,(cid:126)q k
0:t
(2)
(cid:126)q k
0:t
t
k=1
Based on the problem formulation (3), an EM algorithm
can be derived to learn the macro-action FSCs. Algorithmi-
cally, the main steps involve alternating between computing
the lower bound of the log empirical value function (2) (E-
step) and parameter estimation (M-step). This optimization
algorithm is called policy based expectation maximization
(PoEM), the details of which is referred to [12].
III. RELATED WORK
The use of multi-robot teams has recently become viable
for large-scale operations due to ever-decreasing cost and
increasing accessibility of robotics platforms, allowing robots
to replace humans in team-based decision-making settings
including, but not limited to, search and rescue [1]. Use
of multiple robots allows dissemination of heterogeneous
capabilities across the team, increasing fault-tolerance and
decreasing risk associated with losing or damaging a single
all-encompassing vehicle [21].
The large body of work on multi-robot
task alloca-
tion (MRTA) comes in decentralized, centralized, and dis-
tributed/hybrid flavors. Centralized architectures [22], [23]
rely on full information sharing between all robots. However,
in settings such as SAR, communication infrastructure may
be unavailable, requiring the use of alternative frameworks.
Distributed frameworks, such as those used in auction-
based algorithms [24], use local communication for con-
sensus on robot policies. This enables robustness against
communication failures in hazardous, real-world settings.
However, in settings such as SAR, it can be unreasonable
or impossible for robots to communicate with one another
during task execution. Decentralized frameworks, such as
Dec-POMDPs [13] and the approach proposed in this paper,
target this setting, allowing a spectrum of policies rang-
ing from communication-free to explicitly communication-
enabled. The flexibility offered by decentralized planners
makes them suitable candidates for multi-robot operation in
hazardous or uncertain domains, such as SAR.
Finally, note that unlike the majority of the existing MRTA
literature, the work presented here exploits the strengths of
the MacDec-POMDP framework [8] to develop a unifying
framework which considers sources of uncertainty,
task-
level learning and planning, temporal constraints, and non-
deterministic action durations.
IV. ITERATIVE SAMPLING BASED EXPECTATION
MAXIMIZATION ALGORITHM
The PoEM algorithm [12] is the first attempt to address
policy learning for MacDec-POMDPs with batch data. How-
ever, one of the biggest challenge for PoEM is that
it
only grantees convergence to a local solution, a problem
often encountered when optimizing mixture models, such
as the empirical value function (1) 3. Moreover, PoEM
is a deterministic algorithm for approximate optimization,
meaning that it converges to the same stationary point if
initialized repeatedly from the same starting value. Hence,
PoEM can be prone to poor local solution for more com-
plicated real-world problems (as it will be shown in a later
numerical experiment). To address these issues, we propose
a concurrently (multi-threaded) randomized method called
iterative sampling based Expectation Maximization (iSEM).
The iSEM algorithm is designed to run multiple instances of
PoEM with randomly initialized FSC parameters in parallel
to minimize the probability of converging to a sub-optimal
solution due to poor initialization. Furthermore, to exploit
information and computational efforts on runs of PoEM
which are clearly converging to poor values, iSEM allows
re-sampling of parameters once convergence of V (Dtest) is
detected, increasing the chance of overcoming poor local
optima. Because of the re-sampling step, which involves
random reinitialization for threads converging to poor local
value, iSEM can be deemed as a randomized version of
3Note that
the empirical value function (1) can be interpreted as a
likelihood function for FSCs with the number of mixture components equal
to the total number of subepisodes(cid:80)K
k=1 Tk [25].
Algorithm 1 ISEM
Require: Episodes D(K)
train,D(K)
eval, number of MC samples
M, maximum iteration number Tmax, threshold , J =
∅, Iter = 0
i ) using (1)
I = {1,··· , M} \ J, Iter = Iter + 1
for i ∈ I do
Sample {Θi} ∼ Dirichlet(1)
Θ∞
i = PoEM(Θi, Dtrain)
Compute V (Deval, Θ∞
1: while I (cid:54)= ∅ or Iter ≤ Tmax do
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
end for
14:
15: end while
16: return Controller parameters Θ∗.
end for
Compute Θ∗ = arg maxi∈{1,··· ,M} V (Deval, Θ∞
i )
J = ∅
for i = 1 to M do
J = J ∪ {i}
if V (Deval, Θ∗) − V (Deval, Θ∞
i ) < then
end if
train
i }i∈I
the PoEM algorithm. This is essential for convergence to
well-performing policies, since it widely known that global
optimization paradigms are often based on the principal of
stochasticity [26].
iSEM is outlined in Algorithm 1. Domain experience data
is first partitioned into training and evaluation sets, D(K)
and D(K)
eval. iSEM takes the partitioned data, the number of
Monte Carlo samples (threads) M and parameters controlling
convergence as input, and maintains two sets, I and J: I
records the indices of threads whose evaluation values are
lower than the best value, and J records the remaining
thread indices (and is initialized as empty). iSEM iteratively
applies four steps: 1) update I (line 2); 2) for the threads in I,
randomly initialize FSC parameters by drawing samples from
Dirichlet distributions with concentration parameter 1, run
the PoEM algorithm [12] and evaluate the resulting policy
{Θ∞
4 (line 4-6); 3) update the best policy and its evalu-
ation value obtained in current iteration (line 8); 4) update J
by recording the indices of threads whose converged policy
values are close to the best policy (line 9-13). Critically,
the final step (update of J) enables distinguishing threads
that clearly converge to poor local solutions and "good"
local solutions. In the subsequent iteration, threads with poor
local solutions are reinitialized and re-executed until the
policy values from all the threads are close to the best
solution learned so far. The iSEM algorithm is guaranteed
to monotonically increase the lower bound of empirical
value function over successive iterations and the convergence
property is summarized by the following theorem.
V(cid:0)D(K); Θ(cid:1), until convergence to a maximum.
increases
Proof: Assume that Θ∗(t) is a policy with the highest
evaluation value among the policies learned by all the threads
at iteration t, and the set Jt records the thread indices with
Theorem 2: Algorithm 1 monotonically
4∞ sign indicates run the PoEM algorithm until convergence.
corresponding policy value close to V (Deval, Θ∗(t)). In
the iteration t + 1, the set It+1 contains the thread indices
with corresponding policy values satisfy V (Deval, Θ∗(t)) −
V (Deval, Θi(t)) > ,∀i ∈ It+1 = {1,··· , M} \ Jt.
Starting from t = 0, we have V (Deval, Θ∗(0)) ≥
i (0)),∀i ∈ I0 = {1,··· , M}. In the next
V (Deval, Θ∞
iteration (i.e., t = 1), we have I1 = {1,··· , M} \
J0 ≤ I0. The steps 4-6 allow the threads in I1 to
rerun with randomly reinitialized parameters. According to
step 8 (Algorithm 1), we can obtain V (Deval, Θ∗(1)) ≥
V (Deval, Θ∗(0)). Following the same analysis for t > 1, we
can obtain V (Deval, Θ∗(t)) ≥ V (Deval, Θ∗(t − 1)). Since
{V (Deval, Θ∗(t))}∞
t=0 is a monotone sequence and it is up-
per bounded by Rmax
1−γ , according to Monotone convergence
theorem, V (Deval, Θ∗(t)) has a finite limit, which completes
the proof.
Note that the convergence of iSEM is different from that
iSEM updates a global pa-
of PoEM in the sense that
rameter estimate based on feedbacks from several
local
optima (obtained from random initialization). It
is also
worth mentioning that with finite number of threads, iSEM
might still converge to a local maximum. However, we
can show that on average,
iSEM has higher probability
of convergence to better solutions than PoEM. Moreover,
the iSEM algorithm can be considered a special case of
evolutionary programming (EP) [27], which maintains a
population of solutions (i.e., the set of policy parameters in
J). Yet, there are obvious differences between iSEM and PE.
Notably, instead of mutating from existing solutions, iSEM
resamples completely new initializations for parameters and
optimizes them using PoEM. In additional, iSEM is highly
parallelizable due to its use of concurrent threads.
V. EXPERIMENTS
This section presents simulation and hardware experiments
for evaluating the proposed policy learning algorithm. First
a simulator for a large problem motivated by SAR is
introduced. Then, the performance of iSEM is compared
to previous work based on the simulated SAR problem.
Finally, a multi-robot hardware implementation is presented
to demonstrate a working real-world system.
A. Search and Rescue Problem
The SAR problem involves a heterogeneous set of robots
searching for victims and rescuing survivors after a disaster
(e.g., bringing them to a location where medical atten-
tion can be provided). Each robot has to make decisions
using information gathered from observations and limited
communications with teammates. Robots must decide how
to explore the environment and how to prioritize rescue
operations for the various victims discovered.
The scenario begins after a natural disaster strikes the
simulated world. The search and rescue domain considered is
a 20 × 10 unit grid with s = 6 designated sites: 1 muster site
and 5 victim sites. All robots are initialized at the muster site.
Victim sites are randomly populated with victims (6 victims
total). Each victim has a randomly-initialized health state.
While the locations of the sites are known, the number of
victims and their health at each site is unknown to the robots.
The maximum victim capacity of each site also varies based
on the site size. Each victim's health degrades with time.
An unmanned aerial vehicle (UAV) surveys the disaster
from above. A set of 3 unmanned ground vehicles (UGVs)
can search the space or retrieve victims and deliver them
to the muster site, where medical attention is provided.
The objective of the team is to maximize the number of
victims returned to the muster site while they are still alive.
This is a challenging domain due to its sequential decision-
making nature, large size (4 agents), and both transition and
observation process uncertainty, including stochasticity in
communication. Moreover, as communication only happens
within a limited radius, synchronization and sharing of global
information are prohibited, making this a highly-realistic and
challenging domain.
B. Simulator Description
All simulation is conducted within the Robot Operating
System (ROS) [28]. The simulator executes a time-stepped
model of the scenario, where scenario parameters define the
map of the world, number of each type of robot, and locations
and initial states of victims.
Each robot's macro-controller policy is executed by a
lower-level controller which checks the initiation and ter-
mination conditions for the macro-action and generates se-
quences of primitive actions.
1) Primitive Actions: The simulator models primitive
actions, each of which take one time-step to execute. The
primitive actions for the robots include: (a) move vehicle,
(b) pick-up victim (UGVs only), (c) drop-off victim (UGVs
only) and (d) do nothing. Observations and communication
occur automatically whenever possible and do not take any
additional time to execute.
Macro-action policies, built from these primitive actions,
may take any arbitrary amount of time in multiples of the
time-steps of the simulator. Macro-action durations are also
non-deterministic, as they are a function of the scenario
parameters, world state, and inter-robot interactions (e.g.,
collision avoidance).
2) The World: While the underlying robotics simulators
utilized are three-dimensional, the world representation is
in two dimensions. This allows increased computational
efficiency while not detracting from policy fidelity, as the
sites for ground vehicles are ultimately located on a 2D
plane. The world is modeled as a 2D plane divided into an
evenly-spaced grid within a rectangular boundary of arbitrary
size. Each rescue site is a discrete rectangle of grid spaces
of arbitrary size within the world.
Some number of victims are initially located in each
rescue site. Victim health is represented as a value from 0
to 1, with 1 being perfectly healthy and 0 being deceased.
Each victim may start at any level of health, and its health
degrades linearly with time. If a victim is brought to the
muster location, its health goes to 1 and no longer degrades.
One victim at a time may be transported by a UGV to the
muster, although this can be generalized to larger settings by
allowing the vehicle to carry multiple victims simultaneously.
3) Movement: Simulated dynamical models are used to
represent the motion of the air and ground vehicles within
ROS. The vehicles can move within the rectangular bound-
aries of the world defined in the scenario.
UGV motion is modeled using a Dubins car model. Real-
time multi-robot collision avoidance is conducted using the
reciprocal velocity obstacles (RVO) formulation [29]. State
estimates are obtained using a motion capture system, and
processed within RVO to compute safe velocity trajectories
for the vehicles.
UAV dynamics are modeled using a linearization of a
quadrotor around hover state, as detailed in [30]. Since the
UAV operates at a higher altitude than UGVs and obstacles,
there are no restrictions to the air vehicle's movement.These
dynamics correspond to the transition model T specified
in the (Mac)Dec-POMDP frameworks discussed in the sec-
tion II-A.
4) Communication: Communication is
range-limited.
When robots are within range (which is larger for UAV-
UGV communication than for UGV-UGV communication),
they will automatically share their observations with two-
way communication. Communication is imperfect, and has
a .05 probability of failing to occur even when robots are
in range. For the scenarios used to generate the results in
this study, a UGV can communicate its observation with any
other UGV within 3 grid spaces in any direction; the UAV
can communicate with any UGV within 6 grid spaces in any
direction.
C. MacDec-POMDP Representation
We now describe the MacDec-POMDP represention that
is used for learning. Note that the reprentation in Section
V-B is not observable to the robots and is only used for
constructing the simulator.
1) Rewards: The joint reward is +1 for each victim
brought back to muster alive and −1 for each victim who
dies.
2) Observations: In the SAR domain, a UAV can observe
victim locations when over a rescue site. However, victim
health status is not observable by air. A UGV that is in a
rescue site can observe all victims (location and health status)
within that site. Robots are always able to observe their own
location and whether they are holding a victim at a given
moment.
The observation vector O on which the macro-controller
makes decisions is a subset of the raw observations each
robot may have accumulated through the execution of the
prior macro-action. The robots report the state of their current
location and one other location (which could be directly
observed or received via communication while completing
the macro action). The second location reported is the most
urgent state with the most recent new observation. If there are
no new observations other than the robot's own location, the
second location observation is equivalent to the self location.
The observation vector is as follows,
O = [self state, self location, location state,
second location, second location state]
(4)
where, self state ∈ {1/0= is/not holding victim}, self location
∈ {site 1, site 2,
..., site s}, location state ∈ {0 = no
victims needing help, 1 = victims needing help (not critical),
2 = victims needing help (critical)}, second location ∈ {site
..., site s}, and second location state ∈ {0 =
1, site 2,
no victims needing help, 1 = victims needing help (not
critical), 2 = victims needing help (critical)}. There are
18s2 possible observation vectors, making the observation
space substantially larger than previous macro-action based
domains [10], [8].
3) Macro-Actions: The macro-actions utilized in this
problem are as follows:
• Go to Muster (available to both UAV and UGV): Robot
attempts to go to the muster point from anywhere else,
but only if it is holding a live victim. If a victim is
on-board, victim will always disembark at the muster.
• Pick up Victim (available only to UGV): Robot (UGV
only) attempts to go to a victim's location from a
starting point within the site. Terminates when the robot
reaches the victim; also may terminate if there is no
longer a victim needing help at the site (i.e., another
robot picked the victim up first or the victim died). If
victim and robot are located in the same grid cell, the
victim can be "picked up".
• Go to Site i ∈ {1, . . . , s} (available to both UAV
and UGV): Robot goes to a specified disaster site i.
Terminates when the robot is in the site. Robot can
receive observations of the victims at the site.
D. Simulations and Numerical Results
The SAR domain extends previous benchmarks for
MacDec-POMDPs both in terms of the number of robots
and the number of states/actions/observations. Notably, due
to the very large observation space cardinality of the SAR
domain, it is difficult to generate an optimal solution with
existing solvers such as [10], [8] in a reasonable amount
of time. Hence, due to the lack of a known global optima,
the RL algorithms (iSEM and PoEM) are compared over the
same datasets. The dataset is collected through the simulator
by using a behavior policy combining a hand-coded expert
policy (the same used in [12]) and a random policy, with ρ
denoting the percentage of expert policy.
To compare iSEM and PoEM on the SAR domain,
experiments are conducted with ρ = [50, 75, 85] and
Qn = [1, 3, 10] (varying controller sizes)5. Corresponding
test (holdout) set results are plotted in Figure 1. Several
conclusions can be drawn from the results. First, as the
amount of training data (K) increases, the cumulative reward
increases for both PoEM and iSEM (under the same η,
as shown in Fig.1b). Second, with the same K,
iSEM
achieves better performance than PoEM, which validates that
iSEM is better at overcoming the local optimality limitation
suffered by PoEM. In addition, as the number of threads
M increases, iSEM converges to higher average values and
smaller variance (as indicated by the error-bar, compared to
PoEM), according to Figure 1c, which empirically justifies
5Qn = 1 corresponds to reactive policies (based only on current
observations).
the discussion under Theorem 2. Moreover, as shown in
Fig.1d, under three settings of Qn, the FSCs learned by
iSEM render higher value than the PoEM policy. As Qn
increases, the difference between PoEM and iSEM (with
fixed M) tends to decrease, which indicates we should in-
crease M as iSEM is exploring higher dimensional parameter
spaces. Finally, even in cases where the mean of iSEM is
only slightly higher than PoEM, the variance of iSEM is
is consistently lower than PoEM – a critical performance
difference given the uncertainty involved in the underlying
domain tested.
RVIZ [31] was used in conjunction with ROS to visualize
the simulations. Fig.1a shows the start of one trial with
the different colored circles being sites, the stacked cubes
positioned at sites as victims with colors indicating their
health values, and the 4 green cylinders indicating the 3
UGVs and the UAV. The sites are as follows (from furthest
to closest): site 1 (red circle), site 2 (green), site 3 (sky blue),
site 4 (pink), site 5 (turquoise), site 6 (orange). Note that the
normal gridworld model used in the POMDP formulation
usually assumes discrete state and discrete primitive actions,
whereas the simulation models are based on macro-actions
which comprised low-level controllers that can deal with both
discrete and continuous primitive action and states.
E. Hardware Implementation
While simulation results validate that
the proposed
MacDec-POMDP search algorithm achieves better perfor-
mance than state-of-the-art solvers, we also verify the ap-
proach on a SAR mission with real robots. This allows
further learning from realworld experiences. A video demo
is made available online6. Learning from simulation allows
robots operate in a reasonable (safe) way, whereas real robots
experiments can potentially provide "realworld" experiences
that are not fully captured by the simulators, hence allowing
the robots to improve their baseline policy (learned from sim-
ulators). The video essentially demonstrates this potential,
assuming the training data is collected from the "realworld".
A DJI F330 quadrotor is used as the UAV for hardware
experiments, with a custom autopilot for low-level control
and an NVIDIA Jetson TX1 for high-level planning and task
allocation (Fig. 2a). The UGVs are Duckiebots [32], which
are custom-made ground robots with an onboard Raspberry
Pi 2 for computation (Fig. 2b). Experiments were conducted
in a 40 ft. × 20 ft. flight space with a ceiling-mounted pro-
jection system [33] used to visualize site locations, obstacles,
and victims. As discussed earlier, limited communication
occurs between robots, with a motion capture system used
to ensure adherence to maximal inter-robot communication
distances.
The hardware experiments conducted demonstrated that
the policy generated from iSEM (with ρ > 75, K >
100, Qn > 3) was able to save all victims consistently
well, despite robots having to adhere to collision avoidance
constraints. In some instances, the robots were not able to
save all 6 victims. However, in these scenarios, only 1 victim
was lost, with the cause of loss due to an extremely low
6Video URL: https://youtu.be/B3b60VqWMIE
(a)
(b)
(c)
(d)
Fig. 1: (a) RVIZ simulation of the experiment; Testing performance using (b) different number of training samples (with
Qn=10), (c) threads (with Qn=10, K=500), and (d) controller sizes (with M = 8, K = 100, ρ = 85).
(a) UAV is a DJI F330
quadrotor with onboard Jet-
son TX1 flight controller.
(b) UGVs are custom-build
ground robots with onboard
Raspberry Pi 2.
Fig. 2: Robots in used Hardware Implementation.
starting health for multiple victims. In such cases, an early
victim death would occur before any robot could respond.
Fig. 4 shows the progression of one hardware trial. Sites
are randomly populated with 6 victims total. All robots
initiate at the muster site (Fig. 4a). As the UGVs navigate
towards sites (dictated by their policy), they simultaneously
begin observing their surroundings. When they do, the outer
ring surrounding them turns into the color of the latest victim
observed (Fig. 4b). A UGV can only pick up a new victim
if it is not currently carrying a victim. Its inner circle then
indicates the health of the victim it is carrying, while its outer
ring indicates the health of a randomly-selected victim still
present at the site (if any). Fig. 4c illustrates a situation where
no more victims are present at site 6, thereby causing the
UGV's outer ring to turn black (no victims to save at latest
encountered site). Note that an observed deceased victim also
falls under this category. After a UGV picks up a victim, it
drops it off at the muster (Fig. 4d). The victim returns to full
health, indicating a successful rescue. When a UAV visits a
site, its outer ring also turns into the color of the victim
observed at the site (Fig. 4e). The UAV has no inner circle
because it cannot pick up victims. Fig. 4f and 4g show
two more instances of a UGV picking up a victim from
site 5. As mentioned before, a deceased victim results in a
observation color of black in Fig. 4g. Fig. 4h shows the end
of the hardware trial, where all healthy victims have been
rescued.
VI. CONCLUSION
This paper presents iSEM, an efficient algorithm which
improves the state-of-the-art learning-based methods for co-
ordinating multiple robots operating in partially observed
environments. iSEM enables cooperative sequential decision
making under uncertainty by modeling the problem as a
MacDec-POMDP and using iterative sampling based Ex-
pectation Maximization trials to automatically learn macro-
action FSCs. The proposed algorithm is demonstrated to ad-
dress local convergence issues of the state-of-the-art macro-
action based reinforcement learning approach, PoEM. More-
over, simulation results showed that iSEM is able to gener-
ate higher-quality solutions with fewer demonstrations than
PoEM. The iSEM policy is then applied to a hardware-
based multi-robot search and rescue domain, where we
demonstrate effective control of a team of distributed robots
to cooperate in this partially observable stochastic environ-
ment. In the future, we will make our demonstration even
closer to real world scenarios by modeling observations and
communications as actions and assigning costs. We will also
experiment with other methods other than random sampling,
such as active sampling for the resampling step in iSEM, to
accommodate restrictions of computational resources (i.e.,
number of threads).
REFERENCES
[1] S. Grayson, "Search & Rescue using Multi-Robot Systems," http:
//www.maths.tcd.ie/∼graysons/documents/COMP47130 SurveyPaper.
pdf, 2014.
[2] K. Dresner and P. Stone, "Multiagent traffic management: Opportuni-
ties for multiagent learning," in Learning and Adaption in Multi-Agent
Systems. Springer, 2006, pp. 129–138.
[3] D. Bernstein, S. Zilberstein, R. Washington, and J. Bresina, "Planetary
rover control as a markov decision process," in Sixth Int'l Symposium
on Artificial Intelligence, Robotics, and Automation in Space, 2001.
[4] K. Jolly, K. Ravindran, R. Vijayakumar, and R. S. Kumar, "Intel-
ligent decision making in multi-agent robot soccer system through
compounded artificial neural networks," Robotics and Autonomous
Systems, vol. 55, no. 7, pp. 589–596, 2007.
[5] M. Gath, Optimizing transport logistics processes with multiagent
planning and control. Springer, 2016.
[6] F. A. Oliehoek and C. Amato, A Concise Introduction to Decentralized
POMDPs. Springer, 2016.
[7] C. Amato, G. D. Konidaris, and L. P. Kaelbling, "Planning with macro-
actions in decentralized POMDPs," in Proc. of the int'l Conf. on
Autonomous agents and multi-agent systems (AAMAS-14), 2014.
[8] C. Amato, G. Konidaris, G. Cruz, C. Maynor, J. How, and L. Kael-
bling, "Planning for decentralized control of multiple robots under
uncertainty," in 2015 IEEE Int'l Conf. on Robotics and Automation
(IRCA), 2015.
[9] C. Amato, G. Konidaris, A. Anders, G. Cruz, J. P. How, and L. P.
Kaelbling, "Policy Search for Multi-Robot Coordination under Uncer-
tainty," in Proc. of the 2015 Robotics: Science and Systems Conference
(RSS-15), 2015.
[10] S. Omidshafiei, A. akbar Agha-mohammadi, C. Amato, and J. P. How,
"Decentralized control of partially observable Markov decision pro-
cesses using belief space macro-actions." in 2015 IEEE International
Conference on Robotics and Automation (ICRA).
[11] S. Omidshafiei, A.-a. Agha-mohammadi, C. Amato, S.-Y. Liu, J. P.
How, and J. Vian, "Graph-based cross entropy method for solving
multi-robot decentralized pomdps," in 2016 IEEE International Con-
ference on Robotics and Automation (ICRA), pp. 5395–5402.
; (percentage of hand-coded policy) 507585Cumulative reward-2024data collection policyPoEMiSEM M=4iSEM M=8Qn (controller size) 1310Cumulative reward01234PoEMiSEM (a) Zoomed in view of 3 vic-
tims (indicated as squares) at
a particular site.
(b) UGV observes the victim
at the site with high health.
(c) UGV outer ring color is
black, as there are no other
victims at this site.
(d) UAV can only observe,
but not carry, victims. Thus, it
only has an outer ring indicat-
ing observations.
Fig. 3: Overview of hardware domain with 1 UAV and 3 UGVs. Projection system used to visualize sites and victim
locations/health state. Victims shown as squares with colors representing health (green: high health, yellow: low health, red:
critical health, black: deceased). For all robots, outer ring color indicates its noisy observation of the health of one of the
victims present. For UGVs, inner circle color indicates health of the victim it is carrying.
(a) Start of experiment
(b) UGV observes victim at
site 6
(c) UGV picks up a victim,
observes no others at site 6
(d) UGV drops off victim at
muster
(e) UAV observes victim at
site
(f) UGV picks up victim
from site 5, observes another
healthy victim at site
(g) UGV picks up a victim
from site 5, observes no more
healthy victims at site 5
(h) All healthy victims have
been saved. End of experiment
Fig. 4: Overview of hardware domain, with 1 UAV and 3 UGVs. Ceiling-mounted projection system used to visualize sites
and victim locations/health state.
[12] M. Liu, C. Amato, E. Anesta, J. Griffith, and J. How, "Learning
for decentralized control of multiagent systems in large, partially-
observable stochastic environments," in AAAI Conf. on Artificial
Intelligence, 2016.
allocation with moving targets," in Decision and Control, 2004. CDC.
43rd IEEE Conference on, vol. 4, 2004, pp. 4280–4285.
[24] H.-L. Choi, L. Brunet, and J. P. How, "Consensus-based decentralized
auctions for robust task allocation," IEEE transactions on robotics,
vol. 25, no. 4, pp. 912–926, 2009.
[13] D. S. Bernstein, R. Givan, N. Immerman, and S. Zilberstein, "The
complexity of decentralized control of Markov decision processes,"
Mathematics of Operations Research, vol. 27, no. 4, pp. 819–840,
2002.
[14] R. S. Sutton, D. Precup, and S. Singh, "Between MDPs and semi-
MDPs: A framework for temporal abstraction in reinforcement learn-
ing," Artificial intelligence, vol. 112, no. 1, pp. 181–211, 1999.
[15] L. P. Kaelbling, M. L. Littman, and A. R. Cassandra, "Planning and
acting in partially observable stochastic domains," Artificial intelli-
gence, vol. 101, no. 1, pp. 99–134, 1998.
[16] C. Amato, B. Bonet, and S. Zilberstein, "Finite-state controllers based
on Mealy machines for centralized and decentralized POMDPs." 2010.
[17] A. Kumar, S. Zilberstein, and M. Toussaint, "Probabilistic inference
techniques for scalable multiagent decision making," Journal of Arti-
ficial Intelligence Research, vol. 53, no. 1, pp. 223–270, 2015.
[18] Z. Song, X. Liao, and L. Carin, "Solving DEC-POMDPs by expecta-
tion maximization of value functions," 2016.
[19] F. Wu, S. Zilberstein, and N. R. Jennings, "Monte-Carlo expectation
maximization for decentralized POMDPs." in Proc. of the 23rd Int'l
Joint Conf. on Artificial Intelligence (IJCAI-13), 2013.
[20] M. Liu, C. Amato, X. Liao, J. P. How, and L. Carin, "Stick-Breaking
Policy Learning in DEC-POMDPs," in Proc. of the 24th Int´l Joint
Conf. on Artificial Intelligence (IJCAI-15), 2015.
[21] C. Y. Wong, G. Seet, and S. K. Sim, "Multiple-robot systems for
USAR: key design attributes and deployment issues," International
Journal of Advanced Robotic Systems, vol. 8, no. 1, pp. 85–101, 2011.
[22] Y. Jin, A. A. Minai, and M. M. Polycarpou, "Cooperative real-time
search and task allocation in UAV teams," in Decision and Control,
2003. Proceedings. 42nd IEEE Conference on, vol. 1, 2003, pp. 7–12.
[23] D. Turra, L. Pollini, and M. Innocenti, "Fast unmanned vehicles task
[25] A. Kumar and S. Zilberstein, "Anytime planning for decentralized
POMDPs using expectation maximization," in Proc. of the 26th Conf.
on Uncertainty in Artificial Intelligence (UAI-10), 2010.
[26] R. Horst and P. M. Pardalos, Handbook of global optimization.
Springer Science & Business Media, 2013, vol. 2.
[27] D. Simon, "Evolutionary optimization algorithms: biologically-
inspired and population-based approaches to computer intelligence.
hoboken," 2013.
[28] M. Quigley, K. Conley, B. P. Gerkey, J. Faust, T. Foote, J. Leibs,
R. Wheeler, and A. Y. Ng, "Ros: an open-source robot operating
system," in ICRA Workshop on Open Source Software, 2009.
[29] J. Van den Berg, M. Lin, and D. Manocha, "Reciprocal velocity
obstacles for real-time multi-agent navigation," in 2008 IEEE Int'l
Conf. on Robotics and Automation (ICRA), pp. 1928–1935.
[30] D. Mellinger, N. Michael, and V. Kumar, "Trajectory generation
and control for precise aggressive maneuvers with quadrotors," The
International Journal of Robotics Research, vol. 32, no. 5, pp. 664–
674, 2012.
[31] D. Gossow and W. Woodall. (2016, nov) RVIZ. http://wiki.ros.org/rviz.
[32] L. Paull, J. Tani, H. Ahn, J. Alonso-Mora, L. Carlone, M. Cap,
Y. F. Chen, C. Choi, J. Dusek, Y. Fang, et al., "Duckietown: an
open, inexpensive and flexible platform for autonomy education and
research," in 2017 IEEE Int'l Conf. on Robotics and Automation
(ICRA).
[33] S. Omidshafiei, A.-A. Agha-Mohammadi, Y. F. Chen, N. K. Ure, S.-
Y. Liu, B. T. Lopez, R. Surati, J. P. How, and J. Vian, "Measurable
augmented reality for prototyping cyberphysical systems: A robotics
platform to aid the hardware prototyping and performance testing of
algorithms," IEEE Control Systems, vol. 36, no. 6, pp. 65–87, 2016.
|
1810.01719 | 2 | 1810 | 2019-01-02T16:59:06 | A Puff of Steem: Security Analysis of Decentralized Content Curation | [
"cs.MA"
] | Decentralized content curation is the process through which uploaded posts are ranked and filtered based exclusively on users' feedback. Platforms such as the blockchain-based Steemit employ this type of curation while providing monetary incentives to promote the visibility of high quality posts according to the perception of the participants. Despite the wide adoption of the platform very little is known regarding its performance and resilience characteristics. In this work, we provide a formal model for decentralized content curation that identifies salient complexity and game-theoretic measures of performance and resilience to selfish participants. Armed with our model, we provide a first analysis of Steemit identifying the conditions under which the system can be expected to correctly converge to curation while we demonstrate its susceptibility to selfish participant behaviour. We validate our theoretical results with system simulations in various scenarios. | cs.MA | cs |
A Puff of Steem: Security Analysis of
Decentralized Content Curation
Aggelos Kiayias1,2, Benjamin Livshits3,4, Andrés Monteoliva
Mosteiro1,5, and Orfeas Stefanos Thyfronitis Litos1
1 University of Edinburgh
2 IOHK
3 Imperial College of London
4 Brave Software
5 Clearmatics
[email protected], [email protected], [email protected],
[email protected]
Abstract. Decentralized content curation is the process through which
uploaded posts are ranked and filtered based exclusively on users' feed-
back. Platforms such as the blockchain-based Steemit6 employ this type
of curation while providing monetary incentives to promote the visibil-
ity of high quality posts according to the perception of the participants.
Despite the wide adoption of the platform very little is known regarding
its performance and resilience characteristics. In this work, we provide
a formal model for decentralized content curation that identifies salient
complexity and game-theoretic measures of performance and resilience
to selfish participants. Armed with our model, we provide a first analy-
sis of Steemit identifying the conditions under which the system can be
expected to correctly converge to curation while we demonstrate its sus-
ceptibility to selfish participant behaviour. We validate our theoretical
results with system simulations in various scenarios.
1
Introduction
The modern Internet contains an immense amount of data; a single user
can only consume a tiny fraction in a reasonable amount of time. There-
fore, any widely used platform that hosts user-generated content (UGC)
must employ a content curation mechanism. Content curation can be un-
derstood as the set of mechanisms which rank, aggregate and filter rele-
vant information. In recent years, popular news aggregation sites like Red-
dit7 or Hacker News8 have established crowdsourced curation as the pri-
mary way to filter content for their users. Crowdsourced content curation,
6 https://steemit.com/ Accessed: 2019-01-02
7 https://www.reddit.com/ Accessed: 2019-01-02
8 https://news.ycombinator.com/ Accessed: 2019-01-02
1
as opposed to more traditional techniques such as expert- or algorithmic-
based curation, orders and filters content based on the ratings and feed-
back of the users themselves, obviating the need for a central moderator
by leveraging the "wisdom of the crowd" [1].
The decentralized nature of crowdsourced curation makes it a suit-
able solution for ranking user-generated content in blockchain-based con-
tent hosting systems. The aggregation and filtering of user-generated
content emerges as a particularly challenging problem in permissionless
blockchains, as any solution that requires a concrete moderator implies
that there exists a privileged party, which is incompatible with a permis-
sionless blockchain. Moreover, public blockchains are easy targets for Sybil
attacks, as any user can create new accounts at any time for a marginal
cost. Therefore, on-chain mechanisms to resist the effect of Sybil users are
necessary for a healthy and well-functioning platform; traditional counter-
Sybil mechanisms [2] are much harder to apply in the case of blockchains
due to the decentralized nature of the latter. The functions performed
by moderators in traditional content platforms need to be replaced by
incentive mechanisms that ensure self-regulation. Having the impact of
a vote depend on the number of coins the voter holds is an intuitively
appealing strategy to achieve a proper alignment of incentives for users
in decentralized content platforms; specifically, it can render Sybil attacks
impossible.
However, the correct design of such systems is still an unsolved prob-
lem. Blockchains have created a new economic paradigm where users are
at the same time equity holders in the system, and leveraging this prop-
erty in a robust manner constitutes an interesting challenge. A variety
of projects have designed decentralized content curation systems [3,4,5].
Nevertheless, a deep understanding of the properties of such systems is
still lacking. Among them, Steemit has a long track record, having been
in operation since 2016 and attaining a user base of more than 1.08 M9
registered accounts10. Steemit is a social media platform which lets users
earn money (in the form of the STEEM cryptocurrency) by both creating
and curating content in the network. Steemit is the front-end of the social
network, a graphical web interface which allows users to see the content of
the platform. On the other hand, all the back-end information is stored on
a distributed ledger, the Steem blockchain. Steem can be understood as
9 https://steemdb.com/accounts Accessed: 2019-01-02
10 The number of accounts should not be understood as the number of active users, as
one user can create multiple accounts.
2
an "app-chain", a blockchain with a specific application purpose: serving
as a distributed database for social media applications [4].
Our Contributions. In this work we study the foundations of decen-
tralized content curation from a computational perspective. We develop
an abstract model of a post-voting system which aims to sort the posts
created by users in a distributed and crowdsourced manner. Our model
is constituted by a functionality which executes a protocol performed by
N players. The model includes an honest participant behaviour while it
allows deviations to be modeled for a subset of the participants. The N
players contribute votes in a round-based curation process. The impact of
each vote depends on the number of coins held by the player. The posts
are arranged in a list, sorted by the value of votes received, resembling
the front-page model of Reddit or Hacker News. In the model, players
vote according to their subjective opinion on the quality of the posts and
have a limited attention span.
Following previous related work [6,1], we represent each player's opin-
ion on each post (i.e. likability) with a numerical value l ∈ [0, 1]. The
objective quality of a post is calculated as the simple summation of all
players' likabilities for the post in question. To measure the effectiveness
of a post-voting system, we introduce the property of convergence under
honesty which is parameterised by a number of values including a metric
t, that demands the first t articles to be ordered according to the objective
quality of the posts at the end of the execution assuming all participants
signal honestly to the system their personal preferences. Armed with our
post-voting system abstraction, we proceed to particularize it to model
Steemit and provide the following results.
i) We characterise the conditions under which the Steemit algorithm
converges under honesty. Our results highlight some fundamental lim-
itations of the actual Steemit parameterization. Specifically, for cu-
rated lists of length bigger than 70 the algorithm may not achieve
even 1-convergence.
ii) We validate our results with a simulation testing different metrics
based on correlation that have been proposed in previous works [7,8]
and relating them to our notion of convergence.
iii) We demonstrate that "selfish" deviation from honest behavior results
to substantial gains in terms of boosting the ranking of specific posts
in the resulting list of the post-voting system.
3
2 Related Work
User-generated content (UGC) has been identified as a fundamental com-
ponent of social media platforms and Web 2.0 in general [9]. The content
created by users needs to be curated, and crowdsourced content cura-
tion [1] has emerged as an alternative to expert-based [10] or algorithmic-
based [11] curation techniques. Motivated by the widespread adoption of
crowdsourced aggregation sites such as Reddit or Digg11, several research
efforts [12,6,13] have aimed to model the mechanics and incentives for
users in UGC platforms. This surge of interest is accompanied by studies
which have shown how social media users behave strategically when they
publish and consume content [14]. As an example, in the case of Reddit,
users try to maximize their 'karma' [15], the social badge of the social
media platform [16].
Previous works have analyzed content curation from an incentives and
game-theoretic standpoint [6,12,17,14,13] . Our formalisation is based on
these models and inherits features such as the quality distribution of the
articles and the users' attention span [1,6]. In terms of the analysis of our
results, the analysis of our t-convergence metric is similar to the top-k
posts in [1]. We also leverage the rank correlation coefficients Kendall's
Tau [7] and Spearman's Rho [8] to measure content curation efficiency.
Our approach describes the mechanics of post-voting systems from a com-
putational perspective, something that departs from the approach of all
previous works, drawing inspiration from the real-ideal world paradigm
of cryptography [18,19] as employed in our definition of t-convergence.
Post-voting systems constitute a special case of voting mechanisms,
as studied within social choice theory, belonging to the subcategory of
cardinal voting systems [20]. In this context, it follows from Gibbard's
theorem [21] that no decentralised non-trivial post-voting mechanism can
be strategy-proof. This is consistent with our results that demonstrate
how selfish behaviour is beneficial to the participants. Our system shares
the property of spanning multiple voting rounds with previous work [22].
Other related literature in social choice [23,24,25] is centered on political
elections and as a result attempts to resolve a variation of the problem
with quite different constraints and assumptions. In more detail, in the
case of political elections, voter communication in many rounds is costly
while navigating the ballot is not subject to any constraints as voters are
assumed to have plenty of time to parse all the options available to them.
As a result, voters can express their preferences for any candidate, irre-
11 http://digg.com/ Accessed: 2019-01-02
4
spective of the order in which the latter appear on the ballot paper. On
the other hand, the online and interactive nature of post-voting systems
make multi-round voting a natural feature to be taken advantage of. At
the same time, the fairness requirements are more lax and it is accept-
able (even desirable) for participants to act reactively on the outcome of
each others' evaluations. On the other hand, in the post-voting case, the
"ballot" is only partially available given the high number of posts to be
ranked that may very well exceed the time available to a (human) user
to participate in the process. As a result a user will be unable to vote for
posts that she has not viewed, for instance, because they are placed in
the bottom of the list. This is captured in our model by introducing the
concept of "attention span."
Content curation is also related to the concept of online governance.
The governance of online communities such as Wikipedia has been thor-
oughly studied in previous academic work [26,27]. However, the financially
incentivized governance processes in blockchain systems, where the vot-
ers are at the same time equity-holders, have still many open research
questions [28,29]. This shared ownership property has triggered interest
in building social media platforms backed by distributed ledgers, where
users are rewarded for generated content and variants of coin-holder vot-
ing are used to decide how these rewards are distributed. The effects of
explicit financial incentives on the quality of content in Steemit has been
analyzed in [30]. Beyond the Steemit's whitepaper [4], a series of blog
posts [31,32] effectively extend the economic analysis of the system. In
parallel with Steemit, other projects such as Synereo [3] and Akasha12 are
exploring the convergence of social media and decentralized content cura-
tion. Beyond blockchain-based social media platforms, coin-holder voting
systems are present in decentralized platforms such as DAOs [33] and
in different blockchain protocols [34,35]. However, most of these systems
use coin-holder voting processes to agree on a value or take a consensual
decision.
3 Model
We first introduce some useful notation:
-- We denote an ordered list of elements with A = [e1, . . . , en] and the
i-th element of the list with A [i] = ei.
-- Let n ∈ N∗. [n] denotes {1, 2, . . . , n}.
12 https://akasha.world/ Accessed: 2019-01-02
5
3.1 Post list
Definition 1 (Post). Let N ∈ N∗. A post is defined as p = (m, l), with
m ∈ [N] , l ∈ [0, 1]N.
-- Author. The first element of a post is the id of its creator, m.
-- Likability. The likability of a post is defined as l ∈ [0, 1]N.
N represents the number of voters (a.k.a. players). A post has a distinct
likability in [0, 1] for each player.
Definition 2 (Ideal Score of a post). Let post p = (m, l). We define
the ideal score of p as idealSc (p) =
lP
li.
i=1
The ideal score of a post is a single number that represents its overall
worth to the community. By using simple summation, we assume that
the opinions of all players have the same weight.
Definition 3 (Post List). Let M ∈ N∗. A post list P = [p1, . . . , pM]
is an ordered list containing posts. It may be the case that two posts are
identical.
In the case of many UGC platforms, e.g. Steemit, there exists a feed
(commonly named "Trending") that displays the same ordered posts for
all users. In such an ordered list, posts placed closer to the top are more
visible, since users typically consume content from top to bottom. We can
thus measure the quality of an ordered list of posts by comparing it with
a list that contains the same posts in decreasing order of ideal score.
Definition 4 (t-Ideal Post Order). Let P a list of posts, t ∈ [M]. The
property Idealt (P) holds if
∀i < j ∈ [t], idealSc (P [i]) ≥ idealSc (P [j]) .
We say that P has a t-ideal rank if Idealt (P) holds and t is the maximum
integer less or equal to M with this property.
3.2 Post Voting System
We now define an abstract post-voting system. Such a system is defined
through two Interactive Turing Machines (ITMs), GFeed and Πhonest. The
first controls the list of posts and aggregates votes, whereas one copy
of the second ITM is instantiated for each player. GFeed sends the post
6
list to one player at a time, receives her vote and reorders the post list
accordingly. The process is possibly repeated for many rounds.
A measure of the quality of a post-voting system is the t-ideal rank
of the post list at the end of the process.
In a more general setting, some of the honest protocol instantiations
may be replaced with an arbitrary ITM. A robust post-voting system
should still produce a post list of high quality.
Definition 5 (Post-Voting System). Consider four PPT algorithms
Init, Aux, HandleVote and Vote. The tuple S consisting of the four
algorithms is a Post-Voting System. S parametrizes the following two
ITMs:
GFeed is a global functionality that accepts two messages: read, which
responds with the current list of posts and vote, which can take various
arguments and does whatever is defined in HandleVote.
Πhonest is a protocol that sends read and vote messages to GFeed when-
ever it receives (activate) from E.
U ← ∅
Init (initArgs)
Algorithm 1 GFeed (Init, Aux, HandleVote) (P, initArgs)
1: Initialization:
2:
3:
4:
5: Upon receiving (read) from upid:
6:
7:
8:
9:
10: Upon receiving (vote, ballot) from upid:
11:
U ← U ∪ {upid}
aux ← Aux (upid)
Send (posts, P, aux) to upid
HandleVote(ballot)
. Set of players
Algorithm 2 Πhonest (Vote)
1: Upon receiving (activate) from E:
2:
3: Wait for response (posts, P, aux)
4:
5:
Send (read) to GFeed
ballot ← Vote (P, aux)
Send (vote, ballot) to GFeed
7
Players are activated by an Environment ITM that sends activation mes-
sages (Algorithm 2, line 1).
Definition 6 (Post-Voting System Activation Message). We de-
fine actpid as the message (activate, pid), sent to upid.
Definition 7 (Execution Pattern). Let N, R ∈ N∗, N ≥ 2.
o
(cid:1) :
∀i ∈ [R] ,∀k ∈ [N] ,∃j ∈ [N] : pid(i−1)N+j = k
ExecPatN,R =(cid:8)(cid:0)actpid1, . . . , actpidN R
,
Let Environment E that sends messages msgs =(cid:0)actpid1, . . . , actpidn
i.e. activation messages are grouped in R rounds and within each round
each player is activated exactly once. The order of activations is not
(cid:1)
fixed.
sequentially. We say that E respects ExecPatN,R if msgs ∈ ExecPatN,R.
(Note: this implies that n = N R.)
Definition 8 ((N, R, M, t)-convergence under honesty). We say that
a post-voting system S = (Init, Aux, HandleVote, Vote) (N, R, M, t)-
converges under honesty (or t-converges under honesty for N players, R
rounds and M posts) if, for every input P such that P = M, for every
E that respects ExecPatN,R and given that all protocols execute Πhonest,
it holds that after E completes its execution pattern, GFeed contains a post
list P0 such that Idealt (P0) is true.
Note that concrete post voting systems may or may not give information
such as the total number of rounds R to the players. This is decided in
algorithm Aux.
We now give a high-level description of a concrete post voting system,
based on the Steemit platform. According to this mechanism, each player
is assigned a number of coins known as "Steem Power" (SP) that remains
constant throughout the execution and another number called "Voting
Power" (VP) in [0, 1], initialized to 1. A vote is a pair containing a post
and a weight w ∈ [0, 1]. Upon receiving a list of posts, the honest player
chooses to vote her most liked post amongst the top attSpan posts of
the list. The weight is chosen to be equal to the respective likability.
The functionality increases the score of the post by SP (aVPw + b) and
subsequently decreases the player's Voting Power by the same amount
(but keeping it within the aforementioned bounds).
8
3
3
regen
m
regen
regen
l a+b
5·24·60·60 = 0.0000069¯4, thus l a+b
her Voting Power. The constraintl a+b
m
Definition 9 (Steemit system). The Steemit system is the post vot-
ing system S with parameters a, b, regen ∈ [0, 1] : a + b < 1,
>
1, attSpan ∈ N∗, SP ∈ RN+. The four parametrizing procedures can be
found in Appendix B.
Remark 1. The constraint a + b < 1 ensures that a single vote of full
weight cast by a player with full Voting Power does not completely deplete
> 1 excludes the degenerate case
in which the regeneration of a single round is enough to fully replenish
the Voting Power in all cases; in this case the purpose of Voting Power
would be defeated.
Remark 2. The Steem blockchain protocol defines a = 0.02, b = 0.0001
and regen =
voted for 7 days from its creation and at most one vote can be cast every
3 seconds, thus R = 7·24·60·60
Remark 3. Note (Algorithm 6, lines 24-40) that an honest player attempts
to vote for as many posts as possible and spreads her votes with the max-
imum distance between them. The purpose of this is to efficiently utilize
the available Voting Power to "make her voice heard". Also, efficiently
using Voting Power on the Steemit website increases the voter's curation
reward [31].
Theorem 1.
1. If ∃i 6= j ∈ [N] : SPi 6= SPj (i.e. if not all players have the same
2. If ∀i 6= j ∈ [N] , SPi = SPj (i.e. if all players have the same Steem
m = 2895. A post can be
Steem Power) then Steemit does not (N, R, M, 1)-converge.
= 201600.
Power) and
(a) R − 1 ≥ (M − 1)l a+b
(b) R−1 < (M − 1)l a+b
at the top. For a constant SP, when R − 1 ≥ (M − 1)l a+b
m then Steemit (N, R, M, M)-converges.
m then Steemit does not (N, R, M, 1)-converge.
m, there are
Proof Sketch. When SP is not constant, we build a post list where the
most liked post is not preferred by rich players and thus is not placed
regen
regen
enough rounds to ensure full regeneration of every player's Voting Power
between two votes and thus the resulting post list reflects the true pref-
erences of the players. In the opposite case, we can always craft a post
list that exploits the fact that some votes are cast with reduced Voting
Power in order to trick the system into placing a wrong post in the top
regen
9
position.
See Appendix A for full proof.
Corollary 1. The Steemit system parametrised according to Remark 2,
for any number of players N ≥ 2, constant SP and M ≤ 70 posts
(N, R, M, M)-converges. If M > 70 or SP is not constant, then there
exists a list of posts such that the system does not (N, R, M, 1)-converge.
4 Simulation
The previous outcomes are here complemented with experiments that
verify our findings. We have implemented a simulation framework that
realizes the execution of Steemit's post-voting system as defined above.
In particular, we consider two separate scenarios: First, we simulate
the case when all players follow the prescribed honest strategy of Steemit,
investigating how the curation quality of the system varies with the num-
ber of voting rounds. We successfully reproduce the result of Theorem 1,
which implies that the system converges perfectly when a sufficient num-
ber of voting rounds is permitted, but otherwise the resulting list of posts
may have a 0-ideal rank, i.e. the top post may not have the best ideal
score. Moreover, we compare our t-convergence metric with previously
used metrics of convergence based on correlation demonstrating that they
are very closely aligned.
The second case measures how resilient is the curation quality of
Steemit against dishonest agents. Since a creator is financially rewarded
when her content is upvoted, she has incentive to promote her own posts.
A combination of in-band methods (apart from striving to produce posts
of higher quality) can help her to that end. Voting for one's own posts,
refraining from voting posts created by others and obtaining Sybil [36]
accounts that only vote for her posts are only an indicative subset. We
thus examine the quality of the resulting list when certain users do not
follow the honest protocol, but apply the aforementioned self-promoting
methods. We observe that there exists a cutoff point above which a small
increase in the number of selfish players has a detrimental effect to the
t-ideal rank of the post voting system. Furthermore, we measure the num-
ber of positions on the list that the selfish post gains with respect to the
number of selfish players.
10
4.1 Methodology
We leverage three metrics to compare the curated list with the ideal list:
Kendall's Tau [7], Spearman's Rho [8], and t-ideal rank.
In addition to the t-ideal rank and the rank correlation coefficients
used in the first scenario, in the case of dishonest participants we include
a metric that measures the gains of the selfish players. In particular, the
metric is defined as the difference between the real position of the "selfish"
post after the execution and its ranking according to the ideal order. We
are thus able to measure how advantageous is for users to behave selfishly.
Furthermore, t-ideal rank informs us how this behavior affects the overall
quality of curation of the platform.
4.2 Execution
In all simulations, the likabilities of all "honest" posts have been drawn
from the [0, 1]-uniform distribution and all players have Steem Power
equal to 1; we leave the case of variable Steem Power as future work.
Scenario A As already mentioned, the results closely follow Theorem 1.
Figures 1 and 2 show the t-ideal rank and Kendall's Tau coefficient re-
spectively when the number of rounds is enough for all votes to be cast
with full Voting Power. In particular, the parameters used are a = 1
50 , b =
10−4, regen =
5·24·60·60 , R = 200000, attSpan = 10, N = 270 and M = 70.
(Observe that R − 1 > (M − 1)l a+b
m.)
3
regen
As we can see, all three measures show that the real list converges
rapidly to the ideal order at the very end of the execution; meanwhile,
the quality of the list improves very slowly.
Figures 3 and 4 depict what happens when the rounds are not suf-
ficient for all votes to be cast with full Voting Power. In particular, the
corresponding simulation was executed with the same parameters, except
for M = 100 and N = 300. (Observe that R − 1 < (M − 1)l a+b
m.)
Here we see that at the end of the execution, only the first three
posts are correctly ordered. Regarding the rest of the list, both Kendall's
Tau and Spearman's Rho coefficients show that the order of the posts
improves only slightly throughout the execution of the simulation.
regen
Scenario B: Selfish users. In order to understand how the presence of
voting rings/Sybil accounts affects the curation quality, we simulate the
execution of the game for various ring sizes, where ring members vote only
11
Fig. 1. t-ideal rank evolution with 270 honest players, 70 posts and 200.000 rounds
Fig. 2. Kendall's Tau and Spearman's Rho evolution with 270 honest players, 70 posts
and 200.000 rounds
12
Fig. 3. t-ideal rank evolution with 300 honest players, 100 posts and 200.000 rounds
Fig. 4. Kendall's Tau and Spearman's Rho evolution with 300 honest players, 100 posts
and 200.000 rounds
13
for a particular, "selfish" post. We fix the rest of the system parameters to
handicap the selfish post. In particular, the voting rounds are sufficient for
all votes to be cast with full Voting Power, the likability of the selfish post
is 0 for all players and it is initially placed at the bottom of the post list.
Define the gain of the post of the selfish players as its ideal position minus
its final position. Figure 5 shows the gain of the selfish post for a varying
number of selfish players, from 1 to 100. Figure 6 depicts the t-ideal rank
of the resulting list at the same executions. The system parameters are
N = 101..200, a = 1
5·24·60 , attSpan = 10, R = 5000.
50 , b = 10−4, regen = 3
Fig. 5. Positions gained by selfish post with 100 honest players, 100 posts and 1 to 100
selfish players
5 Summary and Future Work
We have defined an abstract post-voting system, along with a particular-
ization inspired by the Steemit platform. We proved the exact conditions
on the Steemit system parameters under which it successfully curates
arbitrary lists of posts. We provided the results of simulations of the exe-
cution of the voting procedure under various conditions. Both cases with
14
Fig. 6. t-ideal rank with 100 honest players, 100 posts and 1 to 100 selfish players
only honest and mixed honest and selfish players were simulated. We
conclude that the Voting Power mechanism of Steem and the fact that
self-voting is a profitable strategy may hurt curation quality.
We have studied the curation properties of decentralized content cura-
tion platforms such as Steemit, obtaining new insights on the resilience of
these systems. Some assumptions have been made in the presented model.
Various relaxations of these assumptions constitute fertile ground for fu-
ture work. First of all, the selfish strategy can be extended and refined
in various ways. For example, voting rings can be allowed to create more
than one posts in order to increase their rewards. Optimizing the number
of posts and the vote allocation in this case would contribute towards a
robust attack against the Steemit platform.
Selfish behavior is considered only in the simulation. Our analysis can
be augmented with a review of games with selfish players and voting rings.
The addition of the economic factor invites the definition of utility
functions and strategic behavior for the players. Its inclusion would imply
the need for an expansion of our theorems and definitions to the strategic
case, along with a full game-theoretic analysis. Furthermore, several pos-
15
sible refinements could be introduced; for example, the process of creating
Sybil accounts could be associated with a monetary cost.
Last but not least, in our model, posts are created only at the begin-
ning of the execution. A dynamic model in which posts can be created
at any time and the execution continues indefinitely (as is the case in a
real-world UGC system) is also interesting as a future direction.
Appendix A Proof of Theorem 1: Steem convergence
Proof. -- Statement 1: Reorder the players such that SP1 ≥ SP2 ≥ ··· ≥
{SPj 6= SPj+1}. We first cover the case when
SPN. Let k = min
j∈[N−1]
attSpan ≥ 2.
Let13
strongPost = (0, . . . , 0
,
weakPost = (0, . . . , 0
{z }
k−1
2SPk
nullPost = (0, . . . , 0
)
, 1, 0, . . . , 0
{z }
{z }
k−1
N−k
SPk − SPk+1
{z }
)
N
{z }
, 1, 0, . . . , 0
N−k−1
)
}
P = [weakPost, strongPost, nullPost, . . . , nullPost
We first note that SPk > SPk+1 ≥ 0 ⇒ 0 ≤ SPk−SPk+1
strongPost is a valid post. We then observe that
{z
M−2
2SPk
] .
≤ 1, thus
∀i ∈ {3, . . . , M} , idealSc (P [i]) = 0 <
< idealSc (P [1]) = 1 < 1 + SPk − SPk+1
= idealSc (P [2]) ,
2SPk
thus ∀P0 that contain the same posts as P and Ideal1 (P0) holds, it
is P0 [1] = P [2].
Since attSpan ≥ 2, all players apart from uk+1 vote for P [1] in the
first round and for P [2] in the second, whereas uk+1 votes for P [2]
in the first round and for P [1] in the second. Thus the two first posts
will have been voted by all players by the end of the second round and
their score will not change until the execution completes. We have:
13 We thank Heng Guo from the University of Edinburgh for this counterexample.
16
SPjb + SPk (a + b) + SPk+1 minn
sc2 (P [1]) = scR (P [1]) =
b, VPregk+1,r2
k−1X
j=1
o +
MX
j=k+2
SPjb and
sc2 (P [2]) = scR (P [2]) =
k−1X
SPj min {b, VPregj,r2}+
j=1
SPk − SPk+1
MX
2SPk
VPregk,r2 + b, VPregk,r2} + SPk+1 (a + b) +
SPj min {b, VPregj,r2} ⇒
j=k+2
SPk min {a
k−1X
j=1
SPjb + SPk(a
SPk − SPk+1
2SPk
scR(P [2]) ≤
+ b) + SPk+1 (a + b) +
MX
j=k+2
SPjb .
In the case that VPregk+1,r2 ≥ b, it is
k−1X
j=1
SPk − SPk+1
scR (P [1]) =
k−1X
SPjb + SPk(a
j=1
MX
MX
j=k+2
j=k+2
SPjb ≥
SPjb + SPk (a + b) + SPk+1b +
SPjb >
2SPk
+ b) + SPk+1 (a + b) +
scR (P [2]) ⇒ scR (P [1]) > scR (P [2]) ,
thus Ideal1 (P0) does not hold.
Since uk+1 does not vote in any round between r1 and r2, and r2 ≥ 2, it
is VPregk+1,r2 ≥ 1−a−b+regen. Thus the case when VPregk+1,r2 < b
can happen only when b > 1− a− b + regen ⇔ b > 1−a+regen
. We now
provide a counterexample for the case when b > 1−a+regen
.
Once more we order the players in descending Steem Power, like in
{SPj 6= SPj+1} and we only
the previous case. Once again k = min
j∈[N−1]
2
2
17
care for the case when attSpan ≥ 2. Let 0 < γ < 1 and
{z }
2 , 0, . . . , 0
N−k−1
{z }
, γ, 1, 0, . . . , 0
N−k−1
{z }
{z }
strongPost = (0, . . . , 0
weakPost = (0, . . . , 0
k−1
k−1
γ
, 1,
)
)
nullPost = (0, . . . , 0
)
}
P = [weakPost, strongPost, nullPost, . . . , nullPost
{z
] .
M−2
{z }
N
We observe that ∀i ∈ {3, . . . , M} , idealSc (P [i]) = 0 < idealSc (P [1]) =
2 < 1 + γ = idealSc (P [2]), thus ∀P0 that contain the same posts
1 + γ
as P and Ideal1 (P0) holds, it is P0 [1] = P [2].
Since attSpan ≥ 2, all players apart from uk+1 vote for P [1] in the
first round and for P [2] in the second, whereas uk+1 votes for P [2]
in the first round and for P [1] in the second. Thus the two first posts
will have been voted by all players by the end of the second round and
their score will not change until the execution completes. We have:
sc2 (P [1]) = scR (P [1]) =
SPjb + SPk (a + b) + SPk+1VPregk+1,r2 +
sc2 (P [2]) = scR (P [2]) =
MX
j=k+2
SPjb and
SPj min {b, VPregj,r2} + SPkVPregk,r2 + SPk+1 (a + b) +
k−1X
k−1X
j=1
j=1
k−1X
j=1
MX
j=k+2
SPj min {b, VPregj,r2} ≤
SPjb + SPkVPregk,r2 + SPk+1 (a + b) +
MX
j=k+2
SPjb .
We note that VPregk,r2 = VPregk+1,r2 because both uk and uk+1
vote with full Voting Power in the first round. Let VP = VPregk,r2.
18
We have
SPk (a + b) + SPk+1VP > SPkVP + SPk+1 (a + b) ⇔
SPk (a + b) + SPk+1VP − SPkVP − SPk+1 (a + b) > 0 ⇔
(a + b) (SPk − SPk+1) − VP (SPk − SPk+1) > 0 ⇔
(SPk − SPk+1) (a + b − VP) > 0
The last expression is true because SPk > SPk+1 and VP < b, thus the
first expression is true as well. We can then deduce that scR (P [1]) >
scR (P [2]), thus Ideal1 (P0) does not hold.
We conclude the treatment of Statement 1 by covering the case when
(cid:19)
attSpan = 1. Consider the following list of posts:
(cid:18)SP1 − SPN
weakPost =
SP1 − SPN
(cid:18)SPN − SPN
(
SP1 − SPN
min
1,
, . . . ,
strongPost =
SPN−i+1 − SPN
SPi − SPN
SP1 − SPN
, . . . ,
SPN − SPN
SP1 − SPN
SPk+2 − SPN
)
SP1 − SPN
(SPk − SPk+1)2
(cid:19)
, . . . ,
2SPN−k max {1, SP1 − SPN}
,
,
SPk+1 − SPN
SP1 − SPN
, . . . ,
SP1 − SPN
+
SPk − SPN
SP1 − SPN
, . . . ,
nullPost =
weakPost, strongPost, nullPost, . . . , nullPost
.
}
{z
P =
SP1 − SPN
0, . . . , 0
SP1 − SPN
{z }
N
M−2
All the aforementioned posts are valid, since ∀i ∈ [N] , SP1 ≥ SPi ≥
SPN ⇒ 1 ≥ SPi−SPN
2SPN−k max {1,SP1−SPN} > 0, which
SP1−SPN
ensures that strongPostN−k > 0. We also observe that SPk+1 < SPk ≤
SP1 ⇒ SPk+1−SPN
SP1−SPN
Regarding the ideal scores, we have
< 1, thus strongPostN−k >
SPk+1−SPN
SP1−SPN
≥ 0. Also
(SPk−SPk+1)2
.
∀i ∈ {3, . . . , M} , idealSc (P [i]) = 0 ,
idealSc (P [1]) =
NX
SPi − SPN
SP1 − SPN
NX
=
i=1
i=1
SPN−i+1 − SPN
SP1 − SPN
NX
i=1
,
SPi − SPN
SP1 − SPN
,
idealSc (P [2]) >
19
Thus P [2] has the highest ideal score and ∀P0 that contain the same
posts as P and Ideal1 (P0) holds, it is P0 [1] = P [2].
Since attSpan = 1, all players vote for all posts in the order they
appear in the list of posts, thus P [1] is voted first, with full Voting
Power by all players. P [2] will be voted next by all players with at
most full Voting Power, thus sc1 (P [1]) = scR (P [1]) and sc2 (P [2]) =
scR (P [2]). We will prove that scR (P [2]) < scR (P [1]). It is
scR (P [1]) =
(cid:19)
+ b
,
(cid:18)
NX
SPi (aP [1]i + b) =
i=1
NX
scR (P [2]) ≤ NX
NX
(cid:18)
a
i=1
SPi
SPi − SPN
SP1 − SPN
SPi (aP [2]i + b) =
(cid:19)
+ b
+
i=1
i6=N−k
SPN−k
i=1
SPN−i+1 − SPN
SP1 − SPN
(cid:26)
SPi
a
1,
a min
(cid:18)
(SPk − SPk+1)2
(cid:18)
2SPN−k max {1, SP1 − SPN}
NX
SPN−i+1 − SPN
SPi
SP1 − SPN
a
(SPk − SPk+1)2
i=1
SPN−ka
SPk+1 − SPN
)
SP1 − SPN
+
!
≤
+ b
(cid:19)
+ b
+
2SPN−k max {1, SP1 − SPN} = A .
20
Since scR (P [2]) ≤ A, it is sufficient to prove A < scR (P [1]).
A < scR (P [1]) ⇔ SPN−ka
(SPk − SPk+1)2
2SPN−k max {1, SP1 − SPN}+
(cid:18)
(cid:18)
NX
NX
i=1
i=1
SPi
SPi
a
SPN−i+1 − SPN
(cid:19)
SP1 − SPN
SPi − SPN
+ b
SP1 − SPN
a
(SPk − SPk+1)2
(cid:19)
+ b
−
< 0 ⇔
2 max {1, SP1 − SPN}+
NX
SPi
i=1
(SPk − SPk+1)2
(cid:18)SPN−i+1 − SPN
− SPi − SPN
SP1 − SPN
NX
SP1 − SPN
2 max {1, SP1 − SPN} +
SPi
SPN−i+1 − SPi
SP1 − SPN
(cid:19)
< 0 ⇔
i=1
It is
(SPk − SPk+1)2
2 max {1, SP1 − SPN} ≤ (SPk − SPk+1)2
2 (SP1 − SPN)
.
< 0
(1)
(2)
Furthermore,
2 cX
b N
i=1
NX
i=1
SPi (SPN−i+1 − SPi) =
(SPi (SPN−i+1 − SPi) + SPN−i+1 (SPi − SPN−i+1)) .
From (1) and (2), it suffices to prove that
(SPk − SPk+1)2
2 (SP1 − SPN) +
SPi (SPN−i+1 − SPi) + SPN−i+1 (SPi − SPN−i+1)
SP1 − SPN
< 0 .
2 cX
b N
i=1
21
(SPk − SPk+1)2
2 (SP1 − SPN) +
SPi (SPN−i+1 − SPi) + SPN−i+1 (SPi − SPN−i+1)
SP1 − SPN
(SPk − SPk+1)2
+
2
< 0 ⇔
(SPi (SPN−i+1 − SPi) − SPN−i+1 (SPN−i+1 − SPi)) < 0 ⇔
2 cX
b N
i=1
2 cX
b N
i=1
(SPi − SPN−i+1)2
< 0
(SPk − SPk+1)2
If k ≤j N
2
k, then
2
2 cX
b N
i=1
−
2 cX
b N
i=1
−
(SPk − SPk+1)2
2
(SPi − SPN−i+1)2 =
(SPk − SPk+1)2
2
− (SPk − SPN−k+1)2 −
(SPi − SPN−i+1)2
2 cX
b N
i=1
i6=k
It is
2 cX
b N
−
(SPi − SPN−i+1)2 ≤ 0 and
(cid:23)
i=1
i6=k
(cid:22) N
(SPk − SPk+1)2 ≥ (SPk − SPN−k+1)2 SPk>SPN−k+1⇒
⇒ N − k + 1 ≥ k + 1 ⇒ SPN−k+1 ≤ SPk+1 ⇒
2
k ≤
(SPk − SPk+1)2
>
(SPk − SPN−k+1)2
2
We deduce that
(SPk − SPk+1)2
2
−
⇒
< 0
(SPk − SPN−k+1)2
2
− (SPk − SPk+1)2
2 cX
b N
(SPi − SPN−i+1)2
< 0
i=1
22
j N
2
k, then
Else if k >
(SPk − SPk+1)2
2
2 cX
b N
i=1
−
(SPk − SPk+1)2
2
− (SPN−k − SPk+1)2 −
(SPi − SPN−i+1)2
(SPi − SPN−i+1)2 =
2 cX
b N
i=1
i6=N−k
It is
2 cX
b N
(cid:23)
i=1
i6=N−k
−
(cid:22) N
(SPi − SPN−i+1)2 ≤ 0 and
⇒ N − k < k ⇒ SPN−k ≥ SPk ⇒
k >
(SPN−k − SPk+1)2 ≥ (SPk − SPk+1)2 SPk>SPk+1⇒
2
(SPN−k − SPk+1)2
(SPk − SPk+1)2
>
2
(SPk − SPk+1)2
2
− (SPN−k − SPk+1)2
⇒
< 0
We deduce that
(SPk − SPk+1)2
2
2 cX
b N
i=1
−
(SPi − SPN−i+1)2
< 0
We have concluded that in every case scR (P [2]) < scR (P [1]), thus
Ideal1 (P0) does not hold.
-- Statement 2a: Suppose that
R − 1 ≥ (M − 1)
Observe that
(3) ⇒ R − 1
M − 1 ≥
(cid:24) a + b
regen
(cid:25) rhs⇒
integer
(cid:24) a + b
(cid:25)
(cid:22) R − 1
regen
M − 1
.
(cid:23)
(cid:25)
(cid:24) a + b
regen
≥
(3)
.
(4)
Let pid ∈ [N]. From (3) we deduce that R ≥ M and according
to VoteThisRound in Algorithm 6, upid votes non-null in rounds
23
(r1, . . . , rM) with ri =j(i − 1) R−1
M−1
k + 1. We define the following:
k ∈ N, w ∈ R ,
n ∈ Z, p ∈ [0, 1) : (k − 1) w = n + p ,
m ∈ Z, q ∈ [0, 1) : w = m + q .
We have
(
bkwc =
b(k − 1) wc = n ,
n + m,
n + m + 1, p + q ≥ 1 (impossible if p = 0)
p + q < 1
(
bwc = m
p = 0
m,
m + 1, p > 0
dwe =
(5), (6), (7), (8), p + q < 2 ⇒
From (9) we deduce that
bkwc ∈ {b(k − 1) wc + bwc,b(k − 1) wc + dwe}
(cid:25)(cid:27)
From (4) and (10) we have that ∀i ∈ [M − 1] , ri+1 − ri ≥l a+b
∀i ∈ [M] \ {1} , ri ∈
(cid:22) R − 1
(cid:24) R − 1
M − 1
M − 1
, ri−1 +
ri−1 +
(cid:26)
(cid:23)
regen
will now prove by induction that ∀i ∈ [M] , VPpid,ri = 1.
• For i = 1, VPpid,1 = 1 (Algorithm 3, line 4).
• Let VPpid,ri = 1. Until ri+1, a single non-null vote is cast by upid,
which reduces VPpid by at most a + b (Algorithm 5, line 7) and
m regenerations, each of which replenishes VPpid by
at leastl a+b
.
(10)
m. We
regen
regen. Thus
(cid:26)
(5)
(6)
(7)
(8)
(9)
(cid:27)
(cid:24) a + b
(cid:25)
regen
VPpid,ri+1 ≥ min
VPpid,ri − a − b + regen
, 1
≥ 1 .
But VPpid cannot exceed 1 (line 4), thus VPpid,ri+1 = 1.
Since the above holds for every pid ∈ [N], it holds that at the end of
the execution, all votes have been cast with full Voting Power, thus
24
∀i ∈ [M] , scR (P [i]) = N b + a
sorted by decreasing score (Algorithm 5, line 20). We observe that
P [i]pid and the posts in PR are
pid=1
NP
∀i 6= j ∈ [M] , idealSc (P [i]) > idealSc (P [j]) ⇒
P [i]pid >
NX
pid=1
NX
NX
pid=1
pid=1
P [j]pid ⇒
NX
pid=1
N b + a
P [i]pid > N b + a
P [j]pid .
Thus all posts will be ordered according to their ideal scores; put
otherwise, IdealScoreM (PR) holds.
-- Statement 2b: Suppose that
R − 1 < (M − 1)
(cid:25)
(cid:24) a + b
regen
.
(11)
Several lists of posts will be defined in the rest of the proof. Given that,
when all players are honest, the creator of a post is irrelevant, we omit
the creator from the definition of posts to facilitate the exposition.
Thus every post will be defined as a tuple of likabilities.
First, we consider the case when
attSpan + R ≤ M .
(12)
In this case, no player can ever vote for the last post, as we will show
now. First of all, (12) ⇒ R < M, thus all players cast R votes in total.
Let pid ∈ N, i ∈ [R] and vpid,i the index of the last post that has ever
been in upid's attention span until the end of round i, according to
the ordering of P. It is vpid,1 = attSpan and ∀i ∈ [R] \ {1} , vpid,i =
vpid,i−1 + 1, since in every round upid votes for a single post and
the first unvoted post of the list is added to their attention span.
Note that, since this mechanism is the same for all players, the same
unvoted post is added to all players' attention span at every round.
Thus ∀pid ∈ N, vpid,R = attSpan + R − 1 (12)
< M. We deduce that no
player has ever the chance to vote for the last post.
25
The above observation naturally leads us to the following counterex-
ample: Let
strongPost =
1, . . . , 1
{z }
0, . . . , 0
{z }
nullPost, . . . , nullPost
}
nullPost =
{z
N
N
M−1
P =
, strongPost
∀i ∈ [M − 1] , it is idealSc (P [M]) > idealSc (P [i]), thus ∀P0 that
contain the same posts as P and Ideal1 (P0) holds, it is P0 [1] =
P [M]. However, since the last post is not voted by any player and the
first post is voted by at least one player, it is scR (P [1]) > scR (P [M]),
thus Ideal1 (PR) does not hold.
We now move on to the case when attSpan + R > M. Let V =
min {R, M}. Each player casts exactly V votes. Consider P1 = 1M×N
and pid ∈ [N]. Let
i ∈ [V ] :(cid:16)VPregpid,ri < 1 ∧ (cid:64)i0 < i : VPregpid,ri0 < 1(cid:17)
,
i.e. i is the first round in which upid votes with less than full Voting
Power. Such a round exists in every case as we will show now. Note
that, since the first round is a voting round and the Voting Power of
all players is full at the beginning, if i exists it is i ≥ 2.
• If R ≥ M, it is V = M.
If (cid:64)i ∈ [M] : (cid:16)VPregpid,ri < 1 ∧ (cid:64)i0 < i : VPregpid,ri0 < 1(cid:17), then
∀i ∈ [M] , VPregpid,ri = 1 ⇒ ∀i ∈ [M] \ {1} , ri ≥ ri−1 +l a+b
m
weight, full-Voting Power vote. Thus rM ≥ 1 + (M − 1)l a+b
m
regen
to have enough rounds to replenish the Voting Power after a full-
>
• If R < M, every player votes on all rounds, thus r2 = 2. Note that
R, contradiction.
regen
≥ 2 ⇒ a + b
regen > 1 ⇒ a + b > regen .
(13)
(cid:24) a + b
(cid:25)
regen
Thus ∀pid ∈ [N] , VPregpid,r2 = 1− a− b+regen (13)
< 1, thus i = 2.
26
We proved that i exists. Since all players follow the same voting pat-
tern, the Voting Power of all players in each round is the same. Let
rVP = VPreg1,ri. Assume that attSpan < i ∨ i > 2. We cover
the case where attSpan ≥ i ∧ i = 2 later. In case N is even, let
0 < γ < 0, 0 < < γ (1 − rVP),
N/2
N/2
1, . . . , 1
,
{z }
{z
}
, γ − , . . . , γ −
γ, . . . , γ
,
{z }
{z
}
0, . . . , 0
,
{z }
, 1, . . . , 1
{z
N/2
N/2
N
M−i
, strongPost, nullPost, . . . , nullPost
.
}
weakPost =
strongPost =
nullPost =
weakPost, . . . , weakPost
}
{z
i−1
P =
First of all, it is
∀j ∈ [i − 1] , idealSc (P [j]) = N
2 (1 + γ − ) <
<
N
2 (1 + γ) = idealSc (P [i])
and ∀j ∈ {i + 1, . . . , M} , idealSc (P [j]) = 0 < idealSc (P [i]), thus the
strong post has strictly the highest ideal score of all posts and as a
result, ∀P0 that contains the same posts as P and Ideal1 (P0) holds,
it is P0 [1] = P [i].
We observe that all players like both weak and strong posts more
than null posts, thus no player will vote for a null post unless her
attention span contains only null posts. This can happen in two cases:
First, if the player has not yet voted for all non-null posts, but the
first attSpan posts of the list, excluding already voted posts, are null
posts. Second, if the player has already voted for all non-null posts.
For a null post to rank higher than a non-null one, it must be true
that there exists one player that has cast the first vote for the null
post. However, since the null posts are initially at the bottom of the
list and it is impossible for a post to improve its ranking before it is
27
voted, we deduce that this first vote can be cast only after the voter
has voted for all non-null posts. We deduce that all players vote for
all non-null posts before voting for any null post.
We will now see that the first N
2 players vote first for all weak posts
and then for the strong post. These players like the weak posts more
than the strong post. As we saw, they will not vote any null post before
voting for all non-null ones. If attSpan > 1 they vote for the strong
post only when all other posts in their attention span are null ones and
thus they will have voted for all weak posts already. If attSpan = 1
and since no post can increase its position before being voted, the
strong post will become "visible" for all players only once they have
voted for all weak posts. Thus in both cases the first N
2 players vote
for the strong post only after they have voted for all weak posts first.
The two previous results combined prove that the first N
2 players vote
for the strong post in round ri exactly. We also observe that these
players have experienced the exact same Voting Power reduction and
regeneration as in the case of P1 since they voted only for posts with
likability 1, thus in round ri their Voting Power after regeneration is
, VPregpid,ri =
exactly the same as in the case of P1 : ∀pid ∈ h N
i
rVP.
We observe that the first N
2 players vote for all weak posts with full
Voting Power. As for the last N
2 players, we observe that, if attSpan <
i, they all vote for the first weak post of the list in the first round,
and thus with full Voting Power. If attSpan ≥ i and i > 2, they vote
for the strong post in the first round and for the first weak post in
r2 with full Voting Power. Thus in all cases the last N
2 players vote
for the first weak post with full Voting Power. Therefore, the score
of the first weak post at the end of the execution is scR (P [1]) =
2 (a + b) + N
N
On the other hand, at the end of the execution the strong post has
2 players with rVP Voting Power and by the
been voted by the first N
2 players with at most full Voting Power, thus its final score will
last N
2 (rVP · γa + b) + N
be at most scR (P [i]) ≤ N
< γ (1 − rVP) ⇒
2 ((γ − ) a + b).
2 (a + b). It is
2
N
2 (rVP · γa + b) + N
N
2 (a + b) + N
2 (a + b) <
scR (P [i]) < scR (P [1]) .
2 ((γ − ) a + b) ⇒
Thus PR [1] 6= P [i] and Ideal1 (PR) does not hold.
28
N−1 (1 − rVP). In
As for the case when N is odd, let 0 < < γ N−3
this case, we assume that the likability of the first i posts (weak and
strong) for the additional player is γ, whereas the likability of the last
M −i posts (the null posts) is 0. This means that the additional player
votes first for the weak and strong posts and then for the null posts.
The rest of the likabilities remain as in the case when N is even. We
observe that the ideal score of the strong post is still strictly higher
than the rest. Furthermore, since the additional player votes for the
first weak post within the first i voting rounds, her Voting Power at
the time of this vote will be at least rVP. We thus have the following
bounds for the scores:
scR (P [i]) ≤ N − 1
2
scR (P [1]) ≥ N − 1
((γ − ) a + b) + rVP · γa + b .
2
Given the bounds of , it is scR (P [i]) < scR (P [1]), thus Ideal1 (PR)
does not hold.
We finally cover the previously untreated edge case where attSpan ≥
i ∧ i = 2. rVP is defined like before. We first consider the case when
N is even and greater than 2: ∃k ∈ N \ {0, 1} : N = 2k. Let 0 < γ <
1, 0 < < 2γ 1−rVP
(rVP · γa + b) + N − 1
2
(a + b) + N − 1
2
(a + b) + γa + b ,
P =
We first observe that
∀j ∈ {3, . . . , M} , idealSc (P [j]) = 0 <
< idealSc (P [1]) = k − 1 + (k − 1) (γ − ) + 2γ =
= k − 1 + (k + 1) γ − (k − 1) <
< k − 1 + (k + 1) γ = idealSc (P [2]) ,
29
weakPost =
k−1
k−1
(k−1)rVP,
,
,
weakPost, strongPost, nullPost, . . . , nullPost
.
}
1, . . . , 1
{z }
}
, γ − , . . . , γ −
γ, . . . , γ
}
{z
{z
{z }
{z
strongPost =
, 1, . . . , 1
, γ, γ
k−1
k−1
, γ, γ
M−2
thus the strong post has strictly the highest ideal score of all posts and
as a result, ∀P0 that contains the same posts as P and Ideal1 (P0)
holds, it is P0 [1] = P [2].
The first k − 1 and the last two players vote first for P [1] and then
for P [2], whereas players k, . . . , 2k − 2 vote first for P [2] and then for
P [1], thus at the end of the execution,
scR (P [1]) = (k − 1) (a + b) + 2 (γa + b) + (k − 1) ((γ − ) rVPa + b) ,
scR (P [2]) = (k − 1) (a + b) + (k + 1) (γrVPa + b) .
Given the bound on , it is scR (P [1]) > scR (P [2]), thus Ideal1 (PR)
does not hold.
Second, we consider the case when N is odd: ∃k ∈ N : N = 2k + 1.
Let 0 < γ < 1, 0 < < γ 1−rVP
,
1, . . . , 1
krVP ,
}
{z
{z }
, γ − , . . . , γ −
γ, . . . , γ
,
{z }
}
{z
weakPost, strongPost, nullPost, . . . , nullPost
.
}
{z
strongPost =
weakPost =
, 1, . . . , 1
, γ
P =
k
k
, γ
k
k
M−2
We first observe that
∀j ∈ {3, . . . , M} , idealSc (P [j]) = 0 <
< idealSc (P [1]) = k + k (γ − ) + γ =
= k + (k + 1) γ − k < k + (k + 1) γ = idealSc (P [2]) ,
thus the strong post has strictly the highest ideal score of all posts and
as a result, ∀P0 that contains the same posts as P and Ideal1 (P0)
holds, it is P0 [1] = P [2].
The first k and the last player vote first for P [1] and then for P [2],
whereas players k + 1, . . . , 2k vote first for P [2] and then for P [1],
thus at the end of the execution,
scR (P [1]) = k (a + b) + γa + b + k ((γ − ) rVPa + b) ,
scR (P [2]) = k (a + b) + (k + 1) (γrVPa + b) .
30
Given the bound on , it is scR (P [1]) > scR (P [2]), thus Ideal1 (PR)
does not hold.
Last but not least, we consider the case when N = 2. In this case, let
0 < γ < 1 and
(1, 0) ,
(cid:18)
P =
γ, 1 − γ
1 + rVP
2
, nullPost, . . . , nullPost
(cid:19)
{z
M−2
.
}
2 = γ + 1 − γ 1+rVP
It is ∀j ∈ {3, . . . , M} , idealSc (P [j]) = 0 < idealSc (P [1]) = 1 rVP<1
<
2 = idealSc (P [2]), thus P [2] has strictly
1 + γ 1−rVP
the highest ideal score of all posts and as a result, ∀P0 that contains
(cid:17)
the same posts as P and Ideal1 (P0) holds, it is P0 [1] = P [2].
2
On the other hand, scR (P [1]) = a+2b > γrVPa+b+(cid:16)1 − γ 1+rVP
a+
b = scR (P [2]), thus Ideal1 (PR) does not hold.
31
Appendix B Steem post voting system procedures
Algorithm 3 Init (attSpan, a, b, regen, R, SP)
1: Store input parameters as constants
2: r ← 1
3: lastVoted ← (0, . . . , 0) ∈ (N∗)N
4: VP ← (1, . . . , 1) ∈ [0, 1]N
5: scores ← (0, . . . , 0) ∈(cid:0)R+(cid:1)M
Algorithm 4 Aux
1: return (attSpan, a, b, r, regen, R, SP)
VPpid,r ← VPpid
VPpid ← max {VPpid + regen, 1}
VPregpid,r ← VPpid
if ballot 6= null then
Parse ballot as (p, weight)
cost ← a · VPpid · weight + b
if VPpid − cost ≥ 0 then
score ← cost · SPpid
VPpid ← VPpid − cost
score ← VPpid · SPpid
VPpid ← 0
Algorithm 5 HandleVote (ballot, upid)
1: if lastVotedpid 6= r then
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18: end if
19: if ∀i ∈ [N] , lastVotedi = r then
20:
21:
22:
23: end if
else
. One vote per player per round
. For proofs
. For proofs
. round over
. order posts by votes
. For proofs
end if
scoresp ← scoresp + score
end if
lastVotedpid ← r
P ← Order (P, scores)
Pr ← P
r ← r + 1
32
. TODO: count rounds? simplify with set of voted and check of length?
{lpid}[1]
return null
end if
idx ← idx + 1
if P [idx] /∈ votedPosts then
res ← res ∪ {P [idx]}
res ← ∅
idx ← 1
while res < attSpan & idx ≤ P do
top ← ChooseTopPosts (attSpan,P, votedPosts)
(i, l) ← argmax
(i,l)∈top
votedPosts ← votedPosts ∪ (i, l)
return ((i, l) , lpid)
Algorithm 6 Vote (P, aux)
1: Store aux contents as constants
2: voteRounds ← VoteRounds (R,P)
3: if VoteThisRound (r,P) = yes then
4:
5:
6:
7:
8: else
9:
10: end if
11:
12: function ChooseTopPosts(attSpan,P, votedPosts)
13:
14:
15:
16:
17:
18:
19:
end while
20:
return res
21:
22: end function
23:
24: function VoteThisRound(r, M)
25:
26:
27:
28:
29:
30:
31:
32: end function
33:
34: function VoteRounds(R, M)
35:
36:
37:
38:
39:
40: end function
voteRounds ← ∅
for i = 1 to M do
voteRounds ← voteRounds ∪(cid:8)1 +(cid:4)(i − 1) R−1
if R < M then
return yes
else if r ∈ voteRounds then
end for
return voteRounds
return yes
return no
else
end if
. One vote per post per player
(cid:5)(cid:9)
M−1
References
1. Askalidis G., Stoddard G.: A theoretical analysis of crowdsourced content curation.
In The 3rd Workshop on Social Computing and User Generated Content: vol. 16
(2013)
33
2. Levine B. N., Shields C., Margolin N. B.: A survey of solutions to the sybil attack.
University of Massachusetts Amherst, Amherst, MA: vol. 7, p. 224 (2006)
3. Konforty D., Adam Y., Estrada D., Meredith L. G.: Synereo: The Decentralized
and Distributed Social Network. Self-published: https://pdfs.semanticscholar.
org/253c/c4744e6b2b87f88e46188fe527982b19542e.pdf Accessed: 2019-01-02
(2015)
4. Steem Whitepaper. https://steem.io/steem-whitepaper.pdf Accessed: 2019-
5. Goldin M.: Token-Curated Registries 1.0. https://medium.com/@ilovebagels/
token-curated-registries-1-0-61a232f8dac7 Accessed: 2019-01-02 (2017)
6. Ghosh A., McAfee P.: Incentivizing high-quality user-generated content. In Pro-
ceedings of the 20th international conference on World wide web: pp. 137 -- 146:
ACM (2011)
7. Kendall M. G.: Rank Correlation Methods: vol. 9: p. 68. Stuart, Alan: 2 ed.: doi:
01-02 (2018)
10.1111/j.2044-8317.1956.tb00172.x
8. Spearman C.: The proof and measurement of association between two things. The
American journal of psychology: vol. 15(1), pp. 72 -- 101 (1904)
9. Kaplan A. M., Haenlein M.: Users of the world, unite! The challenges and oppor-
tunities of Social Media. Business horizons: vol. 53(1), pp. 59 -- 68 (2010)
10. Stanoevska-Slabeva K., Sacco V., Giardina M.: Content Curation: a new form
of gatewatching for social media? Documento electrónico. Recuperado el: vol. 16
(2012)
11. Rader E., Gray R.: Understanding user beliefs about algorithmic curation in the
Facebook news feed. In Proceedings of the 33rd annual ACM conference on human
factors in computing systems: pp. 173 -- 182: ACM (2015)
12. Das Sarma A., Das Sarma A., Gollapudi S., Panigrahy R.: Ranking mechanisms
in twitter-like forums. In Proceedings of the third ACM international conference
on Web search and data mining: pp. 21 -- 30: ACM (2010)
13. Abbassi Z., Hegde N., Massoulié L.: Distributed content curation on the Web.
ACM Transactions on Internet Technology (TOIT): vol. 14(2-3), p. 9 (2014)
14. May A., Chaintreau A., Korula N., Lattanzi S.: Filter & follow: How social media
foster content curation. In ACM SIGMETRICS Performance Evaluation Review:
vol. 42: pp. 43 -- 55: ACM (2014)
15. Bergstrom K.: "Don't feed the troll": Shutting down debate about community
expectations on Reddit. com. First Monday: vol. 16(8) (2011)
16. Anderson A., Huttenlocher D., Kleinberg J., Leskovec J.: Steering user behavior
with badges. In Proceedings of the 22nd international conference on World Wide
Web: pp. 95 -- 106: ACM (2013)
17. Gupte M., Hajiaghayi M., Han L., Iftode L., Shankar P., Ursu R. M.: News posting
by strategic users in a social network. In International Workshop on Internet and
Network Economics: pp. 632 -- 639: Springer (2009)
18. Goldreich O.: The foundations of modern cryptography. In Modern Cryptography,
Probabilistic Proofs and Pseudorandomness: pp. 1 -- 37: Springer (1999)
19. Lindell Y., Katz J.: Introduction to modern cryptography. Chapman and Hal-
l/CRC (2014)
20. Hillinger C.: The case for utilitarian voting. Homo Oeconomicus: vol. 22(3)
21. Gibbard A.: Manipulation of voting schemes: a general result. Econometrica: jour-
nal of the Econometric Society: pp. 587 -- 601 (1973)
22. Kalech M., Kraus S., Kaminka G. A., Goldman C. V.: Practical voting rules with
partial information. Autonomous Agents and Multi-Agent Systems: vol. 22(1), pp.
151 -- 182 (2011)
34
23. Lu T., Boutilier C.: Robust approximation and incremental elicitation in voting
protocols. In IJCAI: vol. 1: pp. 287 -- 293 (2011)
24. Conitzer V., Sandholm T.: Communication complexity of common voting rules. In
Proceedings of the 6th ACM conference on Electronic commerce: pp. 78 -- 87: ACM
(2005)
25. Xia L., Conitzer V.: Compilation Complexity of Common Voting Rules. In AAAI
(2010)
29. Ehrsam
26. Leskovec J., Huttenlocher D. P., Kleinberg J. M.: Governance in social media: A
case study of the wikipedia promotion process. In ICWSM: pp. 98 -- 105 (2010)
27. Forte A., Bruckman A.: Scaling consensus:
Increasing decentralization in
Wikipedia governance. In Hawaii International Conference on System Sciences,
Proceedings of the 41st Annual: pp. 157 -- 157: IEEE (2008)
28. Buterin V.: Notes on Blockchain Governance. https://vitalik.ca/general/
2017/12/17/voting.html Accessed: 2019-01-02 (2017)
Blockchain
F.:
https://www.medium.com:
Governance:
ture.
blockchain-governance-programming-our-future-c3bfe30f2d74
2019-01-02 (2017)
Programming
Fu-
https://www.medium.com/@FEhrsam/
Accessed:
Our
30. Thelwall M.: Can social news websites pay for content and curation? The SteemIt
cryptocurrency model. Journal of Information Science: p. 0165551517748290 (2017)
in HF19. https://steemit.
com/steemit/@jga/author-and-curator-rewards-in-hf19. Accessed: 2019-01-
02 (2018)
31. González J.: Author and Curator
32. González J.: Self-voters can achieve an interest of 248% APR!! https://steemit.
rewards
com/utopian-io/@jga/self-voters-can-achieve-an-interest-of-248-apr.
Accessed: 2019-01-02 (2018)
33. Philip Daian Tyler Kell I. M., Juels A.: On-Chain Vote Buying and the Rise of Dark
DAOs. http://hackingdistributed.com/2018/07/02/on-chain-vote-buying/
Accessed: 2019-01-02 (2018)
34. Duffield E., Diaz D.: Dash: A PrivacyCentric CryptoCurrency. Self-published
35. Goodman L.: Tezos -- a self-amending crypto-ledger White paper. https://www.
tezos.com/static/papers/white_paper.pdf (2014)
36. Douceur J. R.: The Sybil Attack. International workshop on Peer-To-Peer Systems
(2015)
(2002)
35
|
1803.06725 | 3 | 1803 | 2019-05-30T16:26:40 | Detection under One-Bit Messaging over Adaptive Networks | [
"cs.MA"
] | This work studies the operation of multi-agent networks engaged in binary decision tasks, and derives performance expressions and performance operating curves under challenging conditions with some revealing insights. One of the main challenges in the analysis is that agents are only allowed to exchange one-bit messages, and the information at each agent therefore consists of both continuous and discrete components. Due to this mixed nature, the steady-state distribution of the state of each agent cannot be inferred from direct application of central limit arguments. Instead, the behavior of the continuous component is characterized in integral form by using a log-characteristic function, while the behavior of the discrete component is characterized by means of an asymmetric Bernoulli convolution. By exploiting these results, the article derives reliable approximate performance expressions for the network nodes that match well with the simulated results for a wide range of system parameters. The results also reveal an important interplay between continuous adaptation under constant step-size learning and the binary nature of the messages exchanged with neighbors. | cs.MA | cs |
Detection under One-Bit Messaging
over Adaptive Networks
Stefano Marano and Ali H. Sayed, Fellow, IEEE
1
Abstract
This work studies the operation of multi-agent networks engaged in binary decision tasks, and derives performance expressions
and performance operating curves under challenging conditions with some revealing insights. One of the main challenges in the
analysis is that agents are only allowed to exchange one-bit messages, and the information at each agent therefore consists of both
continuous and discrete components. Due to this mixed nature, the steady-state distribution of the state of each agent cannot be
inferred from direct application of central limit arguments. Instead, the behavior of the continuous component is characterized in
integral form by using a log-characteristic function, while the behavior of the discrete component is characterized by means of an
asymmetric Bernoulli convolution. By exploiting these results, the article derives reliable approximate performance expressions
for the network nodes that match well with the simulated results for a wide range of system parameters. The results also reveal
an important interplay between continuous adaptation under constant step-size learning and the binary nature of the messages
exchanged with neighbors.
Distributed detection; adaptive networks; one-bit messaging; diffusion schemes; ATC rule.
Index Terms
I. INTRODUCTION
T HE theory of adaptive decision systems lies at the intersection of the fields of decision theory [2], [3] and adaptive learning
and control [4], [5]. In many instances, the qualification "adaptive" in adaptive decision systems refers to the ability of
the system to select the best action based on the observed data [6] as in cognitive radar [7], active hypothesis testing [8], or
controlled sensing [9], [10]. The qualification "adaptive" can also refer to the ability of the decision system to track changes
in the underlying state of nature and to monitor its drifts over time and deliver reliable decisions in real time. The theme of
this article falls into this second type of decision systems.
The classical implementation of adaptive decision systems has often relied on the use of centralized (fusion) processing
units. In more recent years, there has been a shift from centralized architectures to sensor network architectures [11] -- [17]
where data are monitored/collected in a distributed fashion by a collection of individually simple devices but the processing
continues to be centralized. Energy efficiency, robustness, and security issues become a challenge over such implementations.
Besides, the presence of a single central unit makes the system vulnerable to failures and external attacks [12], [13], [18], [19].
One approach to remedy these difficulties is the SENMA (sensor networks with mobile agents) paradigm in which several
mobile central units travel across the network to query the remote nodes from close proximity [20] -- [23].
A more prominent and flexible solution is to avoid the presence of central units. A fully-flat or fully-decentralized sensor
network refers to a network in which all the information processing takes place at the nodes in a fully distributed fashion,
and no data storage or processing at centralized devices is allowed. The evident mitigation of security and failure issues, and
the added robustness, come at the expense of the need for local information processing capabilities at the nodes, which will
now need to interact with each other, perform processing tasks with groups of nearby agents, and arrive at local decisions.
Statistical signal processing over fully-decentralized networks or graphs has become an active area of research (e.g., see the
overviews in [24], [25] and the many references therein). The theme of this work is to design and analyze an adaptive decision
system over such networks.
A. Related Work
Inference problems over fully-flat sensor networks have received considerable attention in the last years in connection with
estimation problems first and, more recently, in connection with detection/decision problems by employing either consen-
sus [26] -- [32] or diffusion [25], [33] -- [42] strategies.
Consensus solutions employ diminshing step-sizes to enhance the memory properties of the decision system, which leads
to asymptotically optimal performance [26] -- [31]. Unfortunately, decaying step-size parameters limit the adaptation ability of
the resulting network because learning comes to a halt as the step-size parameter approaches zero. Switching to constant
step-size adaptation poses a challenge for consensus-based solutions because of an inherent asymmetry in the update equations
S. Marano is with DIEM, University of Salerno, via Giovanni Paolo II 132, I-84084, Fisciano (SA), Italy (e-mail: [email protected]). A. H. Sayed is with
the Ecole Polytechnique Federale de Lausanne EPFL, School of Engineering, CH-1015 Lausanne, Switzerland (e-mail: [email protected]).
The work of A. H. Sayed was also supported in part by NSF grant CCF-1524250.
Part of this work has been presented at EUSIPCO 2018 [1].
2
of consensus implementations. This asymmetry has been studied in some detail and shown earlier in the works [24], [35] to
be a source of instability when constant step-sizes are used for adaptation purposes. In other words, consensus strategies under
constant step-sizes can be problematic for applications that necessitate continuous learning due to potential instability. This fact
motivates us to focus on diffusion implementations since these strategies do not suffer from the aforementioned asymmetry and
have been shown to deliver superior performance under both constant and decaying step-size learning scenarios [24], [43]. There
have been a series of works that develop the theory of diffusion strategies with constant step-size for inference purposes and
explore their capabilities for learning and adaptation in dynamic environments [25], [33] -- [42]. For example, references [33] --
[35] deal with estimation problems, and the latter also addresses a comparison between consensus and diffusion protocols.
In [36] the adaptive diffusion scheme for detection is studied, and [25], [37], [38] present extensive overviews of these detection
algorithms, as well as many access points to the related literature. The learning behavior of the network is investigated in [39],
[40], while a large deviation analysis and the so-called exact asymptotic framework are the focus of [41], [42].
B. Contribution & Preview of the Main Results
The diffusion scheme considered in this paper employs a modified form of the adapt-then-combine (ATC) diffusion rule,
which has some advantages with respect to alternative schemes [37]. According to the ATC rule, each node updates (adapts)
its state by incorporating the fresh information coming from new measurements, and then makes its current state available to
its neighbors for the combination stage. In the combination stage each node weighs its state with those of its neighbors. In
all the articles mentioned so far in this introduction it is assumed that the communication among nearby nodes is essentially
unconstrained. This means that the nodes can share their state with the neighbors with full precision.
In many practical scenarios, an unconstrained inter-node communication capability cannot be guaranteed, and the system
designer is faced with the problem of revisiting the signal/information processing of the network in order to take into account
this limitation. Thus, the basic consideration that motivates our work is that the nodes of the network cannot exchange their
states as they are, because the communication links do not support messages with arbitrary precision. Taking this viewpoint
to one extreme, we assume severe communication constraints which impose that only one bit can be reliably delivered, per
link usage, over the inter-node links. Accordingly, in the combination stage of the ATC rule, the neighbors of node k cannot
be informed about the value of the state of node k, but they can only be informed about a one-bit quantized quantity. In the
binary detection problem addressed here, this quantized quantity can be regarded as a local decision made by node k at the
current time.
Of course, decentralized inference using quantized messages in networks equipped with a central unit (or having other
classical structures, such as the tandem architecture or some variation thereof) has a long history, see e.g., [44] -- [48] and
the references therein. Also, consensus implementations with quantized messages has been widely investigated [49] -- [57].
Apparently there are not similar studies on distributed strategies ensuring continuous learning and adaptation even under
drifting or non-stationary conditions.
The one-bit diffusion messaging scheme addressed in this paper poses new challenges. The combination stage of the ATC
scheme will now fuse discrete and continuous variables and the analysis of the steady-state distribution becomes more complex
than that developed, e.g., in [41]. In contrast to the results of [41], our analysis shows that (a version of) the central limit
theorem (CLT) only applies in the special circumstance that the step size is small and the weight assigned to local decisions
gathered by neighbor nodes is vanishing. In general cases, for arbitrary step sizes and combination weights, deriving the
steady-state statistical distributions requires separate analysis of the continuous and discrete components. Neither of these can
be, in general, approximated by a Gaussian distribution via some version of the CLT, and different analysis tools are required.
The main results of this work can be summarized as follows. By exploiting a key distributional structural property [see (17)],
the steady-state distribution of the continuous component [see (15)] is obtained in an integral form that involves the log-
characteristic function. A series approximation for the continuous component distribution, particularly suitable for numerical
analysis, is also provided. These results are collected in Theorem 1.
As to the discrete component, we show that this reduces to a combination of geometric series with random signs -- the
so-called asymmetric Bernoulli convolution. The Bernoulli convolution has been widely studied in the literature for its measure-
theoretic implications and it is known that, aside from some special cases, its distribution does not reduce to simple forms. This
notwithstanding, exploiting the fact that the node state is the sum of the discrete component with the continuous one, we derive
simple approximations in the regime of highly reliable local decisions [1 − pd, pf (cid:28) 1, see (7)]. In principle, approximations
of any degree can be developed, but the second-order approximation detailed in Sec. IV-B2 gives accurate results even for
moderately large values of 1 − pd and pf . The combination of several Bernoulli convolutions requires numerical convolutions
and we develop careful approximations for the individual contributions so that these convolutions can be easily computed over
discrete, low-cardinality, sets.
Exploiting the above results we finally get reliable approximate expressions for the steady-state distributions of the network
nodes. Our analysis highlights the role of the system parameters on the these distributions, and the inherent system tradeoffs.
The examples of computer experiments (reported in Figs. 3-6) show that the shape of the steady-state distributions is by no
means obvious. It is rewarding that the developed theory is able to closely follow those shapes for a wide range of values of
3
the relevant system parameters. The main performance figures are the system-level detection and false alarm probabilities Pd
and Pf , which are straightforwardly related to the distributions of the network nodes. Expressing Pd in function of Pf , the
receiver operating characteristic (ROC) curve is obtained (Figs. 7-8).
The analysis developed in this paper allows us to easily derive the decision performance of the system for a wide range
of the parameters under the control of the system designer -- step size µ and combination weights {ak}. A critical scenario
is when the self-combination weights {ak} are very large and µ is very small, because the developed numerical procedures
can be time-consuming. Furthermore, for ak → 1 and µ → 0, both the continuous and the discrete components may become
indeterminate and a joint analysis of these components is necessary. For this scenario we develop a tailored version of the
CLT for triangular arrays and continuous parameters, which is the subject of Theorem 2.
From a practical perspective, our main results are a formula (shown just after Theorem 1) providing a simple numerical
recipe for the steady-state distributions of the continuous component, and an approximation for the discrete component, which
are simply combined to yield the final distributions of the agents at the steady-state. Using these results the system designer
can tune the agents' thresholds in order to achieve a desired value of system-level Pf , and compute the corresponding Pd.
The remainder of this paper is organized as follows. Section II introduces the classical adaptive diffusion scheme for detection.
The one-bit-message version of these detection systems is designed in Sec. III, and the steady-state analysis is conducted in
Sec. IV. Examples of applications of the developed theory and results of computer experiments are presented in Sec. V.
Extensions of the proposed approach are briefly discussed in Sec. VI, while Sec. VII concludes the paper with final remarks.
Some technical material is postponed to Appendices A-D.
II. ADAPTIVE DIFFUSION FOR DETECTION
We consider a multi-agent network consisting of S nodes running an adaptive diffusion scheme to solve a binary hypothesis
test problem in which the state of nature is represented by H0 or H1. Using the same notation from [41], the update rule for
the diffusion strategy is given by
vk(n) = yk(n − 1) + µ[xk(n) − yk(n − 1)],
(1a)
yk(n) =
(1b)
where 0 < µ (cid:28) 1 is the step-size parameter, usually much smaller than one. Moreover, the symbol xk(n) denotes the
data received by agent k at time n, while yk(n) represents a local state variable that is updated regularly by the same agent
through (1b). This latter expression combines the intermediate values v(cid:96)(n) from the neighbors of agent k using the nonnegative
convex combination weights {ak(cid:96)}. The weights are required to satisfy
ak(cid:96)v(cid:96)(n),
(cid:96)=1
n ≥ 1,
(2)
where Nk denotes the set of neighbors connected to agent k, including k itself. In the above notation, the scalar ak(cid:96) denotes the
weight by which information flowing from (cid:96) to k is scaled. If agents k and (cid:96) are not neighbors, then ak(cid:96) = 0. Expressions (1a)-
(1b) can be grouped together across all agents in vector form, say, as:
ak(cid:96) = 1,
(cid:96)∈Nk
(3a)
(3b)
where the combination matrix A = [ak(cid:96)] is S × S, while yn = col{y1(n), y2(n), . . . , yS(n)}. Similarly for vn and xn.
Iterating (3a)-(3b) gives
vn = (1 − µ)yn−1 + µxn,
yn = Avn,
n ≥ 1,
(cid:88)
ak(cid:96) ≥ 0,
S(cid:88)
yn = (1 − µ)nAny0 + µ
(1 − µ)iAi+1xn−i.
(4)
In this work we make the assumption that the incoming data xk(n) is a statistic computed from some observed variable, say,
rk(n), namely, xk(n) is a prescribed function of rk(n). Under both hypotheses H0 and H1, the observations rk(n) are i.i.d.
(independent and identically distributed) across all sensors k = 1, . . . , S, and over time. It follows that the same i.i.d. property
holds for {xk(n)}. We further assume that each xk(n) is an absolutely continuous random variable having a probability density
function (PDF) with respect to the Lebesgue measure [58], under both hypotheses. This assumption is mainly because the case
of continuous random variables is the most interesting in the presence of data quantization.
We refer to rk(n) as the local observation, and to xk(n) as the marginal decision statistic, where the adjective "marginal"
is meant to indicate that xk(n) is based on the single sample rk(n). The variable yk(n) is referred to as the state of the node.
The detection problem consists of comparing the state yk(n) against a threshold level, say γ ∈ (cid:60), and deciding on the state
of nature H0 or H1, namely,
yk(n)
γ.
(5)
n−1(cid:88)
i=0
H1
>
<H0
While our formulation is general enough to address different types of marginal statistics, special attention will be given to
the case in which xk(n) is selected as the log-likelihood ratio of rk(n):
where fr,h(rk(n)) is the PDF of rk(n) under Hh, h = 0, 1.
xk(n) = log
fr,1(rk(n))
fr,0(rk(n))
,
(6)
4
A. Some Technical Conditions
that
We introduce the following technical conditions. First, we let Eh denote the expectation under hypothesis Hh, h = 0, 1,
and assume that −∞ < E0x < E1x < ∞. The assumption E0x < E1x is automatically verified when the marginal statistic
is the log-likelihood ratio (6) because, in that case, the quantities −E0x and E1x are two Kullback-Leibler divergences and,
therefore, they are strictly positive for distinct hypotheses [59]. We also assume that the variance Vhx exists and is finite for
h = 0, 1. Note that, for simplicity, we are using the short-hand notation x instead of xk(n).
Second, we let Ph denote the probability operator under hypothesis Hh, h = 0, 1, and assume, for all agents (cid:96) = 1, 2, . . . , S,
0 < pf (cid:44) P0(x ≥ γloc) < P1(x ≥ γloc) (cid:44) pd < 1,
(7)
where γloc is a local threshold level, and where pd and pf represent the marginal detection and false alarm probabilities,
namely, the probabilities that would be obtained by making decisions based on the marginal statistic and the local threshold:
x(cid:96)(n) ≥ γloc ⇒ decide locally in favor of H1, otherwise decide for H0. Note that γloc is the same for all sensors, which
is justified by the assumption of i.i.d. observations under both hypotheses. As a consequence, all sensors have the same
marginal performance pd and pf . The assumption pf , pd (cid:54)= 0, 1, in (7) rules out trivialities. The condition pf < pd is known
as unbiasedness of the marginal decisions [2].
III. ONE-BIT DIFFUSION MESSAGING
The classical diffusion rule is described by equations (1a)-(1b), and its detection properties are studied in detail in [41], [42].
In the system described by (1a)-(1b), the data exchanged among the nodes are uncompressed and non-quantized. However, in
most sensor network scenarios, a more realistic assumption is that the messages exchanged among the nodes are quantized.
Accordingly, we will consider an update rule in which the information sent at time n by node (cid:96) to its neighbors is the marginal
statistic x(cid:96)(n) quantized to one bit, as follows:(cid:101)x(cid:96)(n) (cid:44)
(cid:26) E1x,
if x(cid:96)(n) ≥ γloc,
if x(cid:96)(n) < γloc,
E0x,
(8)
where γloc is the local threshold level from (7). Thus, if the state of nature is H1, node (cid:96) sends to its neighbors the message
E1x with probability pd, and the message E0x with probability (1 − pd). Similarly, under H0, E1x is sent with probability
pf , and E0x with probability (1− pf ). Needless to say, given that the nodes are aware of the detection problem they are faced
with, there is no need to deliver the actual values E0,1x, but simply a binary flag. We can interpret(cid:101)x(cid:96)(n) as representative of
the marginal decision about the hypothesis, made by node (cid:96) by exploiting only its current observation r(cid:96)(n).
When the nodes compute the log-likelihood ratio of the observations using (6), we can set γloc = 0 in (7) and (8). Comparing
the log-likelihood ratio of r(cid:96)(n) to zero corresponds to the optimal ML (Maximum Likelihood) marginal decision about the
underlying hypothesis [60].
Formally, instead of (1a)-(1b), we consider the following update rule with the one-bit messages:
vk(n) = yk(n − 1) + µ[xk(n) − yk(n − 1)],
yk(n) = akkvk(n) +
ak(cid:96)(cid:101)x(cid:96)(n),
n ≥ 1.
(cid:88)
(cid:96)(cid:54)=k
(9a)
(9b)
Note that in the scheme of (9a)-(9b) sensor (cid:96) sends to its neighbors the quantized version (cid:101)x(cid:96)(n) of the marginal decision
statistic. One alternative would be a system in which the sensor sends to its neighbors the quantized version(cid:101)v(cid:96)(n) of its state.
x(cid:96)(n) available at node (cid:96) will be used [through its quantized version (cid:101)x(cid:96)(n)] by the set of neighbors of node (cid:96) (even though
In this case, however, it would not be possible to derive a simple analytical relationship between yk(n) and yk(n − 1) that
yields an explicit expression for the state yk(n) [as in (11) below], and the analytical tractability would be compromised. More
importantly, the scheme (9a)-(9b) ensures improved adaptation performance properties. Indeed, by using (9a)-(9b), the quantity
it is never made available to non-neighboring nodes). Changes in the state of nature are immediately reflected in the value of
x(cid:96)(n), while, with µ (cid:28) 1, the state v(cid:96)(n) of the node incorporates these changes only slowly, as shown in (1a) and (9a). In
dynamic environments, where the state of nature changes with time, this means that the system described by (9a)-(9b) will be
able to react more rapidly to these changes, especially when the self-combination coefficient akk is small.
5
(cid:101)v(cid:96)(n) (blue), and to a system that exchanges the unquantized state v(cid:96)(n) (red). The state of nature is H0 from n = 1 to n = 1000, then switches to H1
Fig. 1. Example of the adaptive properties of the diffusion scheme (9a)-(9b) (curve in black), as compared to a system that exchanges the quantized state
up to n = 2000, and finally switches back to H0. The curves show the evolution of the expected value E[yk(n)] for node k = 3 for the network shown in
Fig. 2, when the x's are Gaussian random variables distributed as detailed in the example of Sec. V-A, with ρ = 2. The combination matrix is as in (57),
with ak = 0.75, and the step size is µ = 0.1. The expectations are computed by averaging 100 Monte Carlo runs. The inset zooms on the region around
n = 1000.
Fig. 2. Topology of the network used in the computer experiments.
Figure 1 illustrates the improved adaptive properties of the scheme (9a)-(9b) by showing the expected value E[yk(n)] for
node k = 3 of the network shown in Fig. 2, when the x's are Gaussian random variables distributed as detailed in the example
of Sec. V-A. The state of nature is initially H0, then switches to H1 at n = 1000, and finally switches down to H0 at
n = 2000. The solid curve in black represents E[yk(n)] for the system defined by recursion (9a)-(9b), where nodes exchange
the quantized version (cid:101)x(cid:96)(n) of their marginal statistic, while the curve in blue refers to a one-bit message scheme in which
the nodes exchange the quantized version (cid:101)v(cid:96)(n) of their state. It is evident that this latter system does not react promptly to
changes in the underlying state of nature and therefore is less suitable to operate in dynamic environments, with respect to the
system defined by (9a)-(9b).
For comparison purposes, Fig. 1 also shows the expected value E[yk(n)] of a diffusion scheme in which no restrictions are
imposed on the messages and therefore the nodes are allowed to exchange the unquantized state v(cid:96)(n). This is the diffusion
system (1a)-(1b) studied in [41], [42]. The inset of Fig. 1 makes it evident that the scheme of (9a)-(9b) exhibits faster reaction
even in comparison to the message-unconstrained diffusion scheme (1a)-(1b). Therefore, investigating the steady-state detection
performance of the system defined by (9a)-(9b) is of great importance, and is the main theme of this work.
A. Explicit Form of the State yk(n)
Let us introduce the coefficients
(cid:26) ak(cid:96), k (cid:54)= (cid:96),
(10)
and, for notational simplicity, let us write ak in place of akk. The {ck(cid:96)} correspond to the off-diagonal elements of A. By
iterating the expressions in (9a)-(9b), we arrive at
k = (cid:96),
ck(cid:96) =
0,
yk(n) = [(1 − µ)ak]nyk(0)
n−1(cid:88)
S(cid:88)
i=0
+ µ ak
n−1(cid:88)
+
(1 − µ)iai
(1 − µ)iai
kxk(n − i)
kck(cid:96)(cid:101)x(cid:96)(n − i).
i=0
For notational convenience, we also introduce the scalar:
(cid:96)=1
ηk (cid:44) (1 − µ)ak,
ηk ∈ (0, 1),
(11)
(12)
05001000150020002500300035004000−2−1012nE[y3(n)]95010001050−202nE[y3(n)] exchange!xℓ(n)exchange!vℓ(n)exchangevℓ(n)6 4 10 7 1 9 5 3 2 8 which combines the step size µ with the self-combination coefficient ak. It is useful to bear in mind that ηk is defined as a
combination of µ and ak, although our notation does not emphasize that.
Consider the quantity yk(n) in (11). After a change of variable i ← i + 1 we have:
k xk(n − i + 1)
ηi−1
(cid:123)(cid:122)
(cid:125)
(cid:124)
ck(cid:96)(cid:101)x(cid:96)(n − i + 1)
(cid:123)(cid:122)
(cid:125)
(cid:124)
n(cid:88)
(cid:124)
(cid:123)(cid:122)
S(cid:88)
k yk(0)
transient
yk(n) = ηn
n(cid:88)
(cid:44)uk(n)
ηi−1
k
+ µ ak
i=1
(cid:96)=1
(cid:125)
i=1
+
.
(cid:44)zk(n)
6
(13)
(15)
(16)
For n → ∞, the transient part converges exponentially to zero with probability one so that, by Slutsky theorem [2, Th.
11.2.11], limn→∞ yk(n) converges in distribution to limn→∞[uk(n) + zk(n)]. In the following, we will investigate the steady-
state properties of the system described by Eqs. (9a)-(9b) and, accordingly, we can set without loss in generality,
so that the transient part ηn
assumed.
(14)
k yk(0) of (13) is eliminated. Henceforth, if not stated explicitly otherwise, condition (14) is always
yk(0) = 0,
In the next section we analyze separately the two components uk(n) (referred to as the continuous component) and zk(n) (the
discrete component), in the asymptotic regime of large number of iterations. This allows us to provide a suitable approximation
for the statistical distribution of limn→∞ yk(n), which is our goal.
Consider the quantity uk(n) defined in (13). Since the random variables {xk(n)} are i.i.d., the following equality in
A. Continuous Component uk(n)
distribution holds:
and, therefore,
lim
n→∞
n(cid:88)
IV. STEADY-STATE ANALYSIS
∞(cid:88)
k xk(n − i + 1) d=
ηi−1
ηi−1
k xk(i) (cid:44) w(cid:63)
k,
i=1
i=1
uk(∞) (cid:44) lim
n→∞ uk(n) d= ak µ w(cid:63)
k.
Let us introduce a new random variable xk(0), which is an independent copy of the random variable xk(n). From definition (15)
the following structural property of w(cid:63)
k immediately follows:
ηk w(cid:63)
k + xk(0) d= w(cid:63)
k.
(17)
k and therefore the distribution
The forthcoming Theorem 1 exploits (17) to derive the distribution of the random variable w(cid:63)
of the steady-state variable uk(∞) defined in (16), whose mean and variance are easily computed from (16) and (17):
akµ
1 − ηk
√−1 be the imaginary unit and, for t ∈ (cid:60), let
Eh[uk(∞)] =
Ehx, Vh[uk(∞)] =
Let j =
a2
kµ2
1 − η2
k
Vhx.
(18)
(19)
(20)
k, respectively, under hypothesis Hh, h = 0, 1. Also, denote by
be the log-characteristic functions [58] of xk(n) and w(cid:63)
Fu,h(u), u ∈ (cid:60), and Φu,h(t) = Φw,h(akµ t) the cumulative distribution function (CDF) and the log-characteristic function of
the steady-state continuous component uk(∞), respectively, under Hh. Denote by (cid:98)z(cid:99) the largest integer ≤ z, and by Im{z}
the imaginary part of z. Finally, to simplify the notation, let
Φx,h(t) (cid:44) log Ehej t x,
Φw,h(t) (cid:44) log Ehej t w(cid:63)
k ,
(cid:110)
(cid:104)
(cid:0) 2n+1
2 δ(cid:1) − j u
µak
2n + 1
(cid:105)(cid:111)
2n+1
2 δ
Im
Ωn(u, δ) =
exp
Φw,h
.
(21)
∞(cid:88)
n=1
∞(cid:88)
(cid:90) ∞
n=1
ϕn,h
1 − ηn
tn,
t < τx,h.
Φw,h(t) =
ii) If τx,h = ∞, we have the representation:
1
2
Fu,h(u) =
− 1
2πj
k
(cid:40) ∞(cid:88)
n=1
exp
−∞
and the series appearing in (24) is (absolutely and uniformly) convergent for t ∈ (cid:60).
iii) For 0 < τx,h ≤ ∞, and δ > 0:
(cid:41)
dt
t
,
− jut
k
Im
exp
ϕn,h
− j
(µ akt)n
1 − ηn
(cid:40)
(cid:34)
(cid:19)m(cid:35)(cid:41)
(cid:18) 2n + 1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ (cid:48)
(cid:0)u + 2πakµ
2
2
δ
,
δ
u δ
µak
2n + 1
2
+ ∆u(u, δ).
(cid:1)(cid:9) ≤ (cid:48)
2 ,
(cid:98) τx,h
ηk δ −1(cid:99)(cid:88)
∞(cid:88)
n=0
− 2
π
(cid:19)
δ
+
m=1
ηm
k ϕm,h
1 − ηm
k
1
2n + 1
Ωn(u, δ) −
(cid:98) τx,h
ηk δ −1(cid:99)(cid:88)
(cid:0)u − 2πakµ
(cid:1), 1−Fu,h
n=0
δ
Ωn(u, δ)
Fu,h(u) =
+ Φx,h
1
2
2
(cid:18) 2n + 1
2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ∞(cid:88)
max(cid:8)Fu,h
n=0
π
Fix, (cid:48) > 0. If δ at the right-hand side of (25) verifies
7
(22)
(23)
(24)
(25)
(26)
Theorem 1 (Distribution of uk(∞)): The continuous component uk(n) converges in distribution for n → ∞, and the CDF
Fu,h(u) of its limit uk(∞) can be characterized as follows. Suppose that Fu,h(u) admits a density. Suppose also that Φx,h(t)
can be expanded in a power series with radius of convergence 0 < τx,h ≤ ∞, namely:
Φx,h(t) =
ϕn,h tn,
t < τx,h.
Then we have the following results.
i) The log-characteristic function Φw,h(t) can be uniquely expanded in a power series with radius of convergence τx,h:
where Ωn(u, δ) is defined in (21), then ∆u(u, δ) ≤ (cid:48).
Proof: See Appendix A.
(cid:3)
The assumed existence of the density in Theorem 1 is known as absolute continuity of Fu,h(u) [58], which holds for most
distributions of practical interest. The series at the right-hand side of (22) is well defined provided that Φx,h(t) is infinitely
differentiable at the origin -- these derivatives are related to the cumulants of the random variable xk(n) [61]. Equality (22)
holds if Φx,h(t) is analytic in t < τx,h [62].
Expression (25) is used in practice by neglecting the term ∆u(u, δ) and truncating the series over m, yielding the approxi-
mation
¯n(cid:88)
(cid:19)
¯m(cid:88)
n=0
(cid:40)
(cid:34)
(cid:18) 2n + 1
exp
− 2
π
Fu,h(u) ≈ 1
2
(cid:19)m(cid:35)(cid:41)
To control the error, the first condition in (26) consists in truncating the series(cid:80)∞
(cid:18) 2n + 1
ηm
k ϕm,h
1 − ηm
u δ
µak
+ Φx,h
2n + 1
− j
for some ¯n, ¯m, and δ.
m=1
Im
+
2
1
2
δ
δ
k
2n + 1
n=0 Ωn(u, δ) to a sufficiently large integer
ηkδ − 1(cid:99) appearing in (25)
¯n. In the proof of Theorem 1 it is shown that ¯n and δ must verify (¯n + 1
is accordingly chosen. The second condition in (26) can be enforced by exploiting some estimate of the tails of Fu,h(u), for
which many methods are available [63]. An example is given in Sec. V, see Appendix D.
ηkδ . The value ¯n = (cid:98) τx,h
2 ) < τx,h
Since (25) is not in simple analytical form, it is not immediate to understand how Fu,h(u) depends on the system parameters,
without resorting to numerical investigations as we shall do in Sec. V. However, from (18), the dispersion index can be computed:
(28)
which reveals that the random variable uk(∞) becomes more concentrated as ηk grows. Note that the second factor at the
right-hand side of (28) is not larger than unity, and depends on the parameters ak and µ only combined into ηk.
Eh[uk(∞)] =
1 + ηk
,
(cid:112)Vh[uk(∞)]
(cid:114) 1 − ηk
√Vhx
Ehx
2
,
(27)
One important remark is in order. In the special case that x is a Gaussian random variable with mean Ehx and variance
Vhx, it is easily found that Φx,h(t) = j t Ehx − 1
Φu,h(t) = j t
2 t2 Vhx, see e.g. [61]. From (23) one immediately gets
akµ
1 − ηk
Ehx and variance a2
Ehx − 1
2
kµ2
1 − η2
kµ2
1−η2
t ∈ (cid:60),
t2 a2
Vhx,
k
revealing that uk(∞) is Gaussian with mean akµ
Vhx, see (18). This appears to be an obvious result
1−ηk
because, for all n, uk(n) is a linear transformation of the variables {xk(n)}, and linear transformations preserve Gaussianity,
see e.g. [61]. For general (non-Gaussian) x, one might wonder if some form of CLT can be applied to the series appearing
in (15), to infer that the distribution of uk(∞) is Gaussian. This is not the case: One usual assumption for the CLT is that
Vhx converges to a finite value for
n → ∞.
the sum of the variances of the summands diverges [61, Eq. (8-123)], whereas(cid:80)n
i=1 η2i−2
k
k
B. Discrete Component zk(n)
We now derive an approximate distribution for the steady-state component limn→∞ zk(n), where zk(n) is defined in (13).
The approximation is valid for large pd and (1 − pf ) [see (7) for the definitions of these quantities], namely in the regime
where the marginal decisions of the nodes are reliable enough.
Let us start by an obvious equality in distribution:
(cid:101)x(cid:96)(n − i + 1)
zk(∞) (cid:44) lim
k
(cid:96)=1
ck(cid:96)
ηi−1
∞(cid:88)
n→∞ zk(n) =
S(cid:88)
∞(cid:88)
S(cid:88)
(cid:101)x(cid:96)(i) (cid:44) z(cid:63)
b(cid:96)(i) (cid:44) 2(cid:101)x(cid:96)(i) − (E1x + E0x)
ηi−1
ck(cid:96)
(cid:96)=1
k.
i=1
k
i=1
d=
We introduce the normalized binary variables
8
(29)
(30)
(31)
(32)
whose alphabet is {−1, +1}. When xk(n) is the log-likelihood ratio, as in (6), then b(cid:96)(i) is simply the result of the signum
k in (31) can be rewritten, after
function applied to (cid:101)x(cid:96)(i), because E1x > 0 and E0x < 0. Using (32), the quantity z(cid:63)
E1x − E0x
straightforward algebra, as
S(cid:88)
z(cid:63)
k =
ck(cid:96)
1 − ηk
(E1x + E0x) + (E1x − E0x) zk(cid:96)
zk(cid:96) (cid:44) (1 − ηk)
∞(cid:88)
2
ηi−1
k b(cid:96)(i).
i=1
,
(33)
(34)
where
(cid:96)=1
From (8), note that under H1 we have P1(b(cid:96)(i) = 1) = pd, while, under H0, P0(b(cid:96)(i) = 1) = pf .
Let us summarize some known properties of the series (34). Suppose that H1 is in force. If we had pd = 1/2, then
expression (34) would represent a geometric series with equally likely random signs. This is known as the Bernoulli convolution,
and attracted considerable interest since the pioneering works by Erdos [64], [65]. It is known that the Bernoulli convolution
is absolutely continuous with a finite-energy density for almost every ηk ∈ (1/2, 1), and is purely singular for ηk ∈ (0, 1/2),
being in that case supported on a Cantor set of zero Lebesgue measure. It is also easily verified that for ηk = 1/2 the random
variable zk(cid:96) is uniform in (−1, 1). We refer to [66] and the references therein for details.
Likewise, the asymmetric Bernoulli convolution with pd (cid:54)= 1/2, which is of interest to us, has been extensively studied. For
pd ∈ (1/2, 2/3), it is known that zk(cid:96) is absolutely continuous for almost all ηk > 2−Hb(pd), and singular for ηk < 2−Hb(pd) [67],
where Hb(p) (cid:44) −p log2(p) − (1 − p) log2(1 − p) is the binary entropy function [59]. Obviously, the same considerations hold
under H0, with pd replaced by pf .
Returning to (34), let us assume first that H1 is in force and consider the following approximation. For any positive integer ωk:
zk(cid:96) = (1 − ηk)
≈ (1 − ηk)
ωk(cid:88)
ωk(cid:88)
i=1
i=1
k b(cid:96)(i) + (1 − ηk)
ηi−1
ηi−1
k b(cid:96)(i) + ηωk
k
∞(cid:88)
(cid:44) (cid:98)zk(cid:96),
i=ωk+1
ηi−1
k b(cid:96)(i)
(35)
value +1, yielding (1 − ηk)(cid:80)∞
where the approximation consists of assuming that all the binary digits b(cid:96)(ωk + 1), b(cid:96)(ωk + 2), . . . are equal to the most likely
k . To the other extreme, when they are all equal to the most unlikely
k b(cid:96)(i) = ηωk
i=ωk+1 ηi−1
bounded (with probability one) by
value −1, we have (1 − ηk)(cid:80)∞
0 ≤(cid:98)zk(cid:96) − zk(cid:96) ≤ 2 ηωk
random variable zk(cid:96) by a discrete random variable(cid:98)zk(cid:96) taking on 2ωk values.
k b(cid:96)(i) = −ηωk
i=ωk+1 ηi−1
To choose the value of ωk, we need to see the effect of the approximation on the variable z(cid:63)
k, obtained when zk(cid:96) is replaced by(cid:98)zk(cid:96) in (33). Using (36) in (33), we have
version of z(cid:63)
(36)
with the upper bound achieved when b(cid:96)(ωk + 1) = b(cid:96)(ωk + 2) = ··· = −1. Note, from (35), that we have approximated the
k ,
k. Let(cid:98)z(cid:63)
k be the approximate
k , which shows that the error involved in the approximation (35) is
9
(37)
(38)
0 ≤(cid:98)z(cid:63)
k − z(cid:63)
k ≤ (E1x − E0x)(1 − ak)
≤ (E1x − E0x)
.
ηωk
k
1 − ηk
ηωk
k
1 − ηk
To control the error in the approximation, we enforce the condition
k ≤ (E1x − E0x)
(cid:98)z(cid:63)
k − z(cid:63)
ηωk
k
1 − ηk
≤ k,h
for some "small" k,h > 0, whose choice will be discussed later. Therefore, the index ωk introduced in (35) is selected to
comply with (38) as follows:
ωk =
,
(39)
(cid:38) log
(cid:39)
E1x−E0x
k,h(1−ηk)
log 1
ηk
where (cid:100)z(cid:101) is the smallest integer ≥ z. Note that the right-hand side of (39) is a decreasing function of k,h and an increasing
function of ηk, when k,h < (E1x−E0x)/(1− ηk). Note also, from (39), that ωk depends also on the hypothesis Hh, h = 0, 1.
This dependence is not made explicit for simplicity of notation.
1) First-Order Approximation of(cid:98)zk(cid:96): We now exploit the assumption pd ≈ 1 to get an approximation for(cid:98)zk(cid:96). The value
taken by the random variable(cid:98)zk(cid:96) in (35) depends on the value of the binary variables b(cid:96)(1), . . . , b(cid:96)(ωk). For large values of
variable(cid:98)zk(cid:96) taking on 2ωk values, into a random variable that takes on only the (ωk + 1) values corresponding to strings with
at most one "−1". In Table I we report, arranged in ascending order, the (ωk + 1) values taken by the quantized(cid:98)zk(cid:96), followed
pd, such string of ωk binary variables typically consists of many "+1" and a few "−1". This suggests to quantize the random
by the string of the ωk binary digits that generates that value, which is referred to as the pattern ("+" means "+1", and "−"
means "−1"). In Table I the subindex k to ηk and ωk is omitted.
RANDOM VARIABLE(cid:98)zk(cid:96) FOR THE FIRST-ORDER APPROXIMATION.
TABLE I
value
1 − 2(1 − η)
1 − 2η(1 − η)
1 − 2η2(1 − η)
pattern
− + + + · · · + +
+ − + + · · · + +
+ + − + · · · + +
...
probability
(1 − pd)
(1 − pd)pd
(1 − pd)p2
d
...
...
1
1 − 2ηω−2(1 − η) + + + + · · · − + (1 − pd)pω−2
1 − 2ηω−1(1 − η) + + + + · · · + − (1 − pd)pω−1
d
+ + + + · · · + +
d
pω
d
We do not implement a standard quantizer. Instead, the probabilities assigned to the (ωk + 1) values are shown in the last
column of Table I, and are computed as follows. The probability assigned to the first pattern "− + + + +··· +" is the sum of
the probabilities of all patterns in the form "− (cid:63) (cid:63)··· (cid:63)", where "(cid:63)" can be either "+" or "−". The probability of the second
pattern "+ − + + +··· +" is the sum of the probabilities of all patterns in the form "+ − (cid:63)··· (cid:63)", and so forth. The general
rule is to replace by stars all the symbols in the pattern following the symbol "−", if any.
Other approximations can be conceived. First, to the values shown in the first column of Table I, one could assign a probability
proportional to that of the corresponding pattern. But this approach amounts to neglect the values of(cid:98)zk(cid:96) having small probability,
which leads to a poor approximation. More appealing would be to implement a regular quantization of(cid:98)zk(cid:96), with quantization
for any value of ηk and, for ηk ≤ 1/2, just amounts to a regular quantization of the random variable(cid:98)zk(cid:96), as it can be shown
regions that are intervals. However, for ηk > 1/2, it may not be easy to identify the sequences b(cid:96)(1), . . . , b(cid:96)(ωk) that, inserted
in (35), yield a value belonging to a prescribed interval. The approach followed in Table I, instead, is analytically straightforward
by simple algebra.
RANDOM VARIABLE(cid:98)zk(cid:96) FOR THE SECOND-ORDER APPROXIMATION.
TABLE II
value
1 − 2(1 − η) − 2η(1 − η)
1 − 2(1 − η) − 2η2(1 − η)
pattern
− − + + · · · + ++
− + − + · · · + ++
probability
(1 − pd)2
(1 − pd)2pd
10
− + + + · · · + +− (1 − pd)2pω−2
(1 − pd)pω−1
− + + + · · · + ++
+ − − + · · · + ++
(1 − pd)2pd
+ − + − · · · + ++
(1 − pd)2p2
d
d
d
+ − + + · · · + +− (1 − pd)2pω−2
(1 − pd)pω−1
+ − + + · · · + ++
+ + − − · · · + ++
(1 − pd)2p2
d
d
d
...
...
...
...
...
...
1 − 2(1 − η) − 2ηω−1(1 − η)
1 − 2η(1 − η) − 2η2(1 − η)
1 − 2η(1 − η) − 2η3(1 − η)
1 − 2(1 − η)
...
1 − 2η(1 − η) − 2ηω−1(1 − η)
1 − 2η2(1 − η) − 2η3(1 − η)
1 − 2η(1 − η)
...
...
...
...
...
...
...
...
1 − 2η2(1 − η) − 2ηω−1(1 − η) + + − + · · · + +− (1 − pd)2pω−2
(1 − pd)pω−1
+ + − + · · · + ++
1 − 2η2(1 − η)
d
d
1 − 2ηω−3(1 − η) − 2ηω−2(1 − η) + + + + · · · − −+ (1 − pd)2pω−1
1 − 2ηω−3(1 − η) − 2ηω−1(1 − η) + + + + · · · − +− (1 − pd)2pω−2
(1 − pd)pω−1
1 − 2ηω−2(1 − η) − 2ηω−1(1 − η) + + + + · · · + −− (1 − pd)2pω−2
(1 − pd)pω−1
(1 − pd)pω−1
1 − 2ηω−3(1 − η)
1 − 2ηω−2(1 − η)
1 − 2ηω−1(1 − η)
+ + + + · · · − ++
+ + + + · · · + −+
+ + + + · · · + +−
+ + + + · · · + ++
1
d
d
d
d
d
d
pω
d
√
2) Second-Order Approximation of(cid:98)zk(cid:96): Along the same lines of the approximation just developed, one can assume that the
variable(cid:98)zk(cid:96) is approximated by a random variable that takes on the 1 + ωk + ωk(ωk − 1)/2 values shown in Table II, with
sequence b(cid:96)(1), . . . , b(cid:96)(ωk) in (35) contains at most two occurrences of the unlikely digit "−1", in which case the random
associated patterns and probabilities. In Table II the subindex k to ηk and ωk is omitted.
The probabilities in the last column of Table II are computed as follows. Consider a generic pattern with two occurrences
of "−", and let us replace with the symbol "(cid:63)" all the "+" appearing to the right of the rightmost "−". Then, the probability
assigned to that pattern is the sum of the probabilities corresponding to all the distinct patterns obtained by assigning to the
stars either the symbol "+" or the symbol "−". For the ωk patterns with only one "−", the probabilities are exactly those of the
pattern, without modification. To understand why we use this convention, consider for instance the pattern "+ +− +··· + ++",
corresponding to the value 1 − 2η2
k(1 − ηk). All probabilities of patterns of the form "+ + − (cid:63) ··· (cid:63) (cid:63)(cid:63)" are included in the
probability of some pattern with two "−", except the single pattern "++−+···+++", whose probability is just (1−pd)pωk−1
.
Straightforward algebra shows that the values in the first column of Table II are arranged in ascending order only if1
2 ≈ 0.618, but the order of the elements does not matter for the final approximation. Furthermore, depending on
5−1
it may be convenient to reduce the cardinality of the random variable(cid:98)zk(cid:96) by aggregating these values in a single value and
k , see (36). In this case,
ηk <
the values of ηk, it may happen that several entries in the first column of Table II are closer than 2ηωk
associating with it the sum of the probabilities of the merged entries.
So far, we have developed the approximation under the assumption that hypothesis H1 is in force. Exploiting the problem
symmetry, it can be seen that the approximation under H0 can be obtained by applying the procedure described for the
hypothesis H1 to the random variable −zk(cid:96), and by replacing pd with (1 − pf ).
3) Approximate Distribution for zk(∞): As shown in (35), we replace zk(cid:96) by the discrete random variable(cid:98)zk(cid:96) with 2ωk
values, where ωk is given in (39). In turn, (cid:98)zk(cid:96) is approximated by the lower-cardinality discrete random variable shown in
Table I, if the first-order approximation is used, and shown in Table II for the more accurate second-order approximation. In
the latter case, the alphabet of the random variable is further reduced by grouping together values that are closer than 2ηωk
k .
This way, an approximate probability mass function (PMF) of zk(cid:96) is obtained.
As indicated in (33), the random variable zk(cid:96) must be multiplied by (E1x−E0x)/2, then added to (E1x+E0x)/2, and finally
multiplied by ck(cid:96)/(1 − ηk). The resulting random variables represent the individual summands of the sum over (cid:96) = 1, . . . , S,
appearing in (31). Note that these random variables are independent. Summing over (cid:96) = 1, . . . , S, corresponds to convolving
the PMFs of the summands, which finally provides the desired PMF [equivalently, the CDF Fz,h(u)] of the steady-state discrete
contribution zk(∞). The number of convolutions is Nk − 2, and these convolutions can be easily implemented numerically,
but the computation might become cumbersome when the alphabet of the individual PMFs becomes excessively large. Thus,
d
1Note that
√
5−1
2
is the golden ratio
√
5+1
2 minus one. The golden ratio is the ratio of successive terms in the Fibonacci series [68].
11
after computing each convolution, similarly to what we have done before, the values of the resulting variable that are close to
k /(1 − ηk) [see (38)] are merged into a single value, and the corresponding probabilities
each other more than (E1x − E0x)ηωk
(cid:101)x(cid:96)(i), (cid:96) =
are summed up.
(cid:101)x(cid:96)(i), and is obtained by computing
analytically. The desired PMF for zk(∞) is the PMF of the finite sum(cid:80)S
1, . . . , S, in (31). Each random variable has been approximated by a discrete random variable whose PMF has been characterized
Summarizing, we have derived simple expressions for the distribution of the random variables ck(cid:96)
(cid:80)∞
i=1 ηi−1
(cid:80)∞
i=1 ηi−1
numerically the convolutions that correspond to the sum over (cid:96). For later use, let us denote this PMF as follows:
(cid:96)=1 ck(cid:96)
k
k
Ph [zk(∞) = zi,h] = νi,h ,
i = 1, . . . , L,
unity, let us assume µ = 0.1. For ak very small so that ηk (cid:28) 1, we see from Tables I and II that the random variable(cid:98)zk(cid:96) takes
(40)
where L (cid:28) 2ωk. In other words, the CDF Fz,h(z), z ∈ (cid:60), of zk(∞) is approximated by a staircase function with steps of
size {νi,h} at the L points {zi,h}.
An informal discussion is useful to better understand the role of the system parameters in determining the distribution of
the discrete component zk(∞). Consider first hypothesis H1. Since the typical values of the step-size µ are much less than
value 1 with probability close to one, and value ≈ −1 with probability close to zero. If node k has a single neighbor, i.e.,
Nk = 2, the sum in (33) disappears and the discrete component zk(∞) takes only two values with non-negligible probability:
E0x ≈ E0x. In this situation, moderate variations of µ have little effect
≈ 1−ak
1−ηk
on ηk, and hence a little effect at all. When ak grows, so does ηk, and from Tables I and II we see that the PMF of zk(∞)
is enriched by additional contributions at points between the extremes, and these extremes become closer to each other. In
this case, diminishing the step-size µ makes ηk larger, which tends to further increasing the number of points of the PMF
(the effect on the position of the extreme values is modest, unless ak is very large). For nodes with larger connectivity Nk,
because of the sum in (33), in each of the above situations, the PMF of the discrete component is enriched by the contributions
due to the convolutions. Similar arguments apply under H0.
E1x ≈ E1x with large probability, and ≈ 1−ak
1−ηk
C. Distribution of the Steady-State Statistic
The CDF of uk(∞) defined in (16) is Fu,h(u), and the CDF of zk(∞) defined in (30) is Fz,h(z). We have seen in the
previous sections how to compute approximately these two distributions, and we now discuss how to obtain the CDF Fy,h(y),
y ∈ (cid:60), of the variable
(41)
which characterizes the state of node k in the steady-state regime of n → ∞. We assume, as already done in Theorem 1, that
Fu,h(u) admits a density, denoted by fu,h(u). This implies that Fy,h(y) also admits a density, and the CDF Fy,h(y) is given
by the convolution [69, pp. 144 -- 146]
n→∞ yk(n) = uk(∞) + zk(∞)
(cid:90) ∞
yk(∞) (cid:44) lim
(cid:90) ∞
−∞
(cid:90) ∞
(cid:90) ∞
(cid:90) ∞
−∞
−∞
−∞
=
≤
≤
Fz,h(ξ − k,h) fu,h(y − ξ) dξ
(cid:98)Fz,h(ξ) fu,h(y − ξ) dξ =
(cid:88)
(cid:98)νi,hFu,h(y −(cid:98)zi,h)
Fz,h(ξ) fu,h(y − ξ) dξ = Fy,h(y).
i
(45)
Fz,h(ξ) fu,h(y − ξ) dξ.
Fy,h(y) =
Using approximation (35) in Sec. IV-B, the random variable z(cid:63)
k is replaced by(cid:98)z(cid:63)
(cid:98)Fz,h(z), z ∈ (cid:60). From condition (38) the following relationship between events follows
k ≤ z − k,h} ⊆ {(cid:98)z(cid:63)
k ≤ z} ⊆ {z(cid:63)
k ≤ z},
Fz,h(z − k,h) ≤ (cid:98)Fz,h(z) ≤ Fz,h(z).
implying that
{z(cid:63)
−∞
Before applying the approximations described in Secs. IV-B1 and IV-B2, the random variable (cid:98)z(cid:63)
on (at most) 2ωk(Nk−1) values. For i = 1, . . . , 2ωk(Nk−1), let {(cid:98)zi,h} denote these values and {(cid:98)νi,h} be the corresponding
probabilities, so that (cid:98)Fz,h(z) is a staircase function with steps of size {(cid:98)νi,h} at the 2ωk(Nk−1) points {(cid:98)zi,h}. From (42) and
(44)
k is discrete and takes
(44):
Fy,h(y − k,h) =
Fz,h(ξ) fu,h(y − k,h − ξ) dξ
k [see (37)], whose CDF will be denoted by
(42)
(43)
This proves that
provided that
Fy,h(y) ≈ 2ωk (Nk−1)(cid:88)
i=1
(cid:98)νi,hFu,h(y −(cid:98)zi,h)
(cid:90) ∞
−∞
Fz,h(ξ) fu,h(y − k,h − ξ) dξ ≈
Fz,h(ξ) fu,h(y − ξ) dξ,
(cid:90) ∞
−∞
12
(46)
(47)
Fy,h(y) ≈ L(cid:88)
The final approximation of the steady-state variable yk(∞) is obtained from (46) after manipulating the random variable(cid:98)z(cid:63)
which is true when fu,h(u) is sufficiently smooth over intervals of length k,h. The value of k,h will be chosen just to ensure
this condition, see Sec. V.
as detailed in Secs. IV-B1, IV-B2, and IV-B3, which reduces its cardinality from 2ωk(Nk−1) to the much smaller value L.
This yields
k
νi,h Fu,h(y − zi,h),
y ∈ (cid:60),
i=1
(48)
where {νi,h} and {zi,h} are defined in (40) and the sum involves L (cid:28) 2ωk(Nk−1) terms. Our final approximation (48) is
only semi-analytical, in the sense that we have an analytical expression for Fu,h(u) but this expression involves the truncation
of series to finite sums, see (27). Similarly, after analytical derivations, computing the sequences {νi,h} and {zi,h} requires
numerical convolutions (over discrete, low-cardinality, sets). In both cases, the numerical procedures are very simple.
The role of the system parameters on the steady-state distribution can be understood by recalling the comments made at the
end of Sec. IV-B3, and exploiting expression (48). From (48) we see that the CDF of yk(∞) consists of the superposition of L
copies of the CDF of the continuous component uk(∞), scaled by νi,h and centered at the points zi,h, i = 1, . . . , L. Assuming
that fu,h(u) is unimodal, the interplay between its width and the spacing between the zi,h's determines the smoothness of
the CDF of yk(∞). Quantitative predictions require a case-by-case analysis, but for many scenarios of practical interest, as
those presented in Sec. V, the general trend is as follows. Suppose µ = 0.1, Nk = 2, and ak (cid:28) 1. As Nk and/or ak grow,
the CDF Fy,h(y) under H0 (resp. H1) moves progressively from a jump-wise shape with larger jumps for small (resp. large)
values of y, to a jumpless smoother shape. This behavior will be confirmed in Sec. V.
The theory developed so far allows us to approximate the distribution Fy,h(y) of the steady-state statistic yk(∞) for most
values of µ and ak of interest. However, for ηk → 1 the approach may become problematic, for several reasons. From a
numerical viewpoint, when ηk is too close to unity the index ωk appearing in (39) may be excessively large. More important,
for ηk → 1 we see from (28) that the dispersion index of the continuous component uk(∞) converges to zero. In addition,
ηk → 1 is equivalent to µ → 0 and ak → 1, and the expected value Eh[uk(∞)] of the continuous component takes different
values depending on the order of the two limits, see the first equation in (18). The same indeterminateness may affect the
appearing in (33) when Nk = 2. Thus, ηk → 1 is a challenging
discrete component zk(∞), because of the factor 1−ak
1−ηk
scenario in which a separate analysis of the discrete and continuous components may not be sufficient. A different approach
is required, which is presented in the next section.
D. Asymptotic Normality
Ignoring the transient term, the random process yk(n) in (13) can be expressed as the sequence of partial sums
n(cid:88)
i=1
yk(n) =
wk(n, i)
(cid:104)
µakxk(n − i + 1) +
of a triangular array {wk(n, i)}, with i ≤ n, and n = 1, 2, . . . , where2
wk(n, i) (cid:44) ηi−1
k
Clearly, it holds that:
(cid:105)
.
(49)
(50)
(51)
(52)
S(cid:88)
(cid:96)=1
ck(cid:96)(cid:101)x(cid:96)(n − i + 1)
(cid:33)
a2
k(cid:96)
,
(cid:96)(cid:54)=k
Eh[wk(n, i)] = ηi−1
Vh[wk(n, i)] = η2(i−1)
k
k
[µakEhx + (1 − ak) Eh(cid:101)x] ,
(cid:32)
(cid:88)
Vhx + Vh(cid:101)x
µ2a2
k
2The quantity wk(n, i) is not to be confused with w(cid:63)
k defined in (15).
where (52) follows by using ckk = 0 in (50) -- see (10). It is worth noting that the two sequences Eh[wk(n, i)] and Vh[wk(n, i)]
do not depend on the index n: they are standard sequences, not arrays [the same is true for any moment of wk(n, i), because
the x's are i.i.d.]. From (51) and (52):
13
mk,h(n) (cid:44) Eh[yk(n)] =
Eh[wk(n, i)]
n(cid:88)
1 − ηn
1 − ηk
k
=
k,h(n) (cid:44) Vh[yk(n)] =
s2
(cid:32)
=
µ2a2
k
i=1
i=1
Vh[wk(n, i)]
[µakEhx + (1 − ak) Eh(cid:101)x] ,
n(cid:88)
Vhx + Vh(cid:101)x
k,h(n) through the quantities S, {ak}S
≤(cid:88)
1 − η2n
1 − η2
k(cid:96) ≤ (1 − ak).
a2
(cid:88)
(cid:33)
a2
k(cid:96)
(cid:96)(cid:54)=k
.
k
k
(53)
(54)
(cid:96)(cid:54)=k a2
k(cid:96). As to the latter,
(55)
k=1, and(cid:80)
The network topology plays its role in mk,h(n) and s2
we have
(1 − ak)2
S − 1
To prove (55), note that ak ≤ 1 implies(cid:80)
which yields ((cid:80)
(cid:96)(cid:54)=k ak(cid:96))2 ≤ (S − 1)(cid:80)
impose that the off-diagonal entries on each row of matrix A are equal. Then, the lower bound(cid:80)
(cid:96)(cid:54)=k ak(cid:96) = (1 − ak), and the upper bound follows. The lower bound
can be derived by applying the Cauchy-Schwarz inequality to the sequence {ak(cid:96)} and to the constant sequence of all ones,
k(cid:96). Of special interest is the situation in which all the neighbors of a given node
are equally reliable, and there is no reason to assign different weights to different neighbors. In this case it makes sense to
in (55)
k(cid:96) = (1−ak)2
S−1
is achieved, and the only topological parameters are the number of nodes S and the self-combination coefficients ak.
k(cid:96) ≤(cid:80)
(cid:96)(cid:54)=k a2
(cid:96)(cid:54)=k a2
(cid:96)(cid:54)=k a2
(cid:96)(cid:54)=k
The limiting quantities mk,h(∞) (cid:44) limn→∞ mk,h(n) and s2
k,h(∞) (cid:44) limn→∞ s2
k,h(n) are immediately obtained from (53)
and (54) by noting that limn→∞ ηn
Theorem 2 (Asymptotic normality): Suppose that there exists β > 0 such that Ex2+β < ∞. Then, under hypothesis Hh,
h = 0, 1, we have:
k = 0, and we have the following result.
yk(∞) − mk,h(∞)
sk,h(∞)
lim
ηk→1
= g,
(56)
where g is a standard Gaussian random variable, and the convergence is in distribution.
Proof. The result essentially follows from the CLT for triangular arrays under the Lyapunov condition, see e.g., [58, Th. 27.3].
One key modification in the argument is that the limit involves also the continuous parameter ηk. The detailed proof is given
(cid:3)
in Appendices B and C.
Note that, for the theorem to hold, it suffices that Ex3 is finite, a condition that is usually easily checked. It is also worth
noting that µ → 0 is not sufficient to ensure the asymptotic normality of (a normalized version of) yk(∞). Technically, the
reason is that the Lyapunov condition (C.1) in Appendix C does not hold if one replaces the limit ηk → 1 with the limit
µ → 0 -- see the arguments at the end of Appendix C. The discussion below elaborates on this issue from a more intuitive
perspective.
There are several substantial differences between the classical diffusion scheme (1a)-(1b) studied in [41], [42], and the
version (9a)-(9b) proposed here for one-bit messaging. One critical difference is just the regime in which the steady-state
statistic is normally distributed. For (1a)-(1b), the steady-state statistic is the rightmost sum in (4). From that expression we see
that the step size µ appears as a common factor for the n elements of the sum. For n → ∞, this implies that the steady-state
statistic is the sum of infinitely many infinitesimal contributions, provided that µ → 0. The CLT applies just to that kind of
situations where a large number of independent contributions are added together and none of them dominates the other terms,
namely the individual variances of the components are small in comparison to their sum.
Contrast this situation with the case of one-bit messaging in (11). When n → ∞, the infinite sum appearing in the second
line of equation (11) consists of infinitesimal contributions when µ → 0. However, this is not the case for the infinite sum
appearing in the third line of (11), which takes into account the decisions of neighboring nodes. In order to invoke the CLT,
we must impose that the contributions of this latter infinite sum become vanishing, and this requires ck(cid:96) → 0, for all k and (cid:96),
which can be obtained by ak → 1. As a matter of fact, in (1a)-(1b) making the step size vanishing ensures the Gaussianity of
the statistic, while in (9a)-(9b) we need ηk = (1 − µ)ak → 1, i.e., we need a vanishing step size and a combination matrix A
with highly dominant diagonal entries.
So far, we have derived the approximate CDF Fy,h(y), y ∈ (cid:60), of the steady-state value yk(∞) of node k = 1, . . . , S, under
both hypotheses H0 and H1. This has been done for general values of the parameters µ and ak in the previous subsections,
14
Fig. 3. Gaussian example under hypothesis H0, with ρ = 1. It is shown the CDF Fy,0(y) of the steady-state variable yk(∞) for node k = 3 (solid curves)
and k = 9 (dashed), obtained by the approximations developed in this paper. The symbols "o" show the results of computer experiments.
and for the case ηk → 1 in the present subsection. In the next section we apply the developed theory to a couple of examples
of practical relevance, and investigate the detection properties of the network, which is our final goal.
V. EXAMPLES
We now consider two examples in which the distributions of the observations are, respectively, Gaussian and exponential.
The role of the system parameters in determining the distributions of the steady-state variable yk(∞), and hence the detection
capabilities of the network, is investigated. We also report the results of computer experiments to check the accuracy of the
approximated formulas provided in the previous sections.
In the examples that follow, we consider the network made of S = 10 nodes depicted in Fig. 2 (this is the same network
considered in [41]), and we refer to the right-stochastic combination matrix A with entries:
ak,
1−ak
Nk−1 ,
0,
ak(cid:96) =
(cid:96) = k,
(cid:96) (cid:54)= k are neighbors,
otherwise,
(57)
where, we recall, Nk is the set of neighbors of node k, including k itself. In addition, to simplify the presentation of the
results, we also assume that the self-combination coefficients ak are all equal.
As to the computer experiments designed to simulate the network evolution, the simulation procedure is as follows. For
the given network topology, at each time instant i = 1, . . . n, the observations xk(i) are simulated for all nodes k = 1, . . . S,
and the update rule (9a)-(9b) is implemented. This is made for both hypotheses H0 and H1, and the resulting state yk(n) is
stored. The value of n is large enough to make negligible the transient part of (13), and to make yk(n) a good approximation
for yk(∞). The computation is repeated many times according to the Monte Carlo principle, thus obtaining the empirical
distribution of yk(∞) under H0 and under H1. These empirical CDFs are then compared with the theoretical CDFs obtained
in the previous sections. The number of Monte Carlo trials for obtaining each point of the empirical CDFs is 104, and the
number of iterations is n = 100, if not specified otherwise.
A. Gaussian Observations
Let r ∼ N (a, b) denote that the random variable r is normally distributed with mean a and variance b. Suppose that the
i.i.d. observations are so distributed: r ∼ N (0, b) under H0, and r ∼ N (a, b) under H1, with a, b > 0. Suppose also that the
nodes compute the log-likelihood ratios of these observations, according to (6), and the local threshold is γloc = 0, see (8).
Straightforward computation shows that the log-likelihood ratios are normally distributed, as follows:
H0∼ N (−ρ, 2ρ),
H1∼ N (ρ, 2ρ),
x
(58)
where ρ = a2/2b is the Kullback-Leibler divergence from H0 to H1 (or vice versa: in this case the divergence is symmetric) [59].
Figure 3 shows the CDF of yk(∞) under H0, for the highly-connected node k = 3 with 5 neighbors, and for the weakly
connected node k = 9 with only one neighbor. In the figure we set ρ = 1, µ = 0.1, and consider three different values of
ak = 0.1, 0.25, 0.5. The curves represent the theoretical CDFs corresponding to the second-order approximation, with solid
lines for k = 3, and dashed lines for k = 9. The symbol "o" denotes values obtained by computer experiments. The curve in
green, which refers to µ = 0.01 and ak = 0.99, shows the normal distribution predicted by Theorem 2 compared to computer
simulations. In this case, 100 iterations are not sufficient to reach the steady-state, and we use n = 1000.
x
-1.5-1-0.500.511.500.20.40.60.81 15
Fig. 4. Gaussian example under hypothesis H1, with ρ = 1. It is shown the CDF Fy,1(y) of the steady-state variable yk(∞) for node k = 3 (solid curves)
and k = 9 (dashed), obtained by the approximations developed in this paper. The symbols "o" show the results of computer experiments.
(cid:18) akµ
(cid:19)
Figure 4 addresses the same analysis under H1. It is worth mentioning that, under both hypotheses, the theoretical CDFs
computed by the approximations developed in the previous section perfectly match the simulation points.
In the present example, from the remark following Theorem 1, we know that
uk(∞) ∼ N
Ehx,
kµ2
a2
1 − η2
Vhx
,
k
1 − ηk
a reasonable choice is k,h = 10−1(cid:112)Vh[uk(∞)], which is used in all the experiments.
(59)
and therefore the distribution of uk(∞) is known without any approximation. The accuracy of the theoretical curves shown
in Figs. 3 and 4 means that the (second-order) approximation developed in Sec. IV-B for the discrete component zk(∞) is
accurate. Note that with ρ = 1 we have pd = (1 − pf ) ≈ 0.76, implying that the approximation for the discrete component
works quite well already for values of pd and (1 − pf ) only moderately close to unity.
As stated below Eq. (47), to compute the theoretical approximation of Fy,h(y) we need to set the value of k,h appearing
in condition (38). Since the PDF of uk(∞) is Gaussian, a meaningful index of its variability is the standard deviation. Thus,
It is worth stressing that the distribution of the steady-state variable yk(∞) is, in general, far from being a Gaussian-shaped
function. Due to the structure of Eq. (48), the CDF of yk(∞) exhibits Gaussian-like increasing regions interleaved by regions
where the CDF is constant. When ak gets large, however, the relative weight of the discrete component zk(∞) decreases,
and the distribution of the yk(∞) tends to approach the Gaussian distribution of the continuous component uk(∞). This is in
agreement with the expected impact of the system parameters, as discussed in Sec. IV-C: for larger ak and/or Nk a smoother
CDF is obtained. The curves shown in Figs. 3 and 4 confirm this behavior.
B. Exponential Observations
Let r ∼ E(λ) denote an exponential random variable of parameter λ, with PDF f (r) = λ exp(−λr), for r ≥ 0. Suppose
that the i.i.d. observations at the various nodes are so distributed: r ∼ E(λ0) under H0, and r ∼ E(λ1) under H1, with
0 < λ1 < λ0. Assume also that the marginal statistics of the nodes are the log-likelihood ratios and that γloc = 0, see (6)
and (8). Straightforward algebra shows that x = (λ0 − λ1)r − log(λ0/λ1).
It is convenient to introduce the normalized parameter λe (cid:44) λ0/λ1 > 1. Under H0, the PDF of x is
(cid:40) λe
e − log λe, and V0x =(cid:0)1 − λ−1
λe−1 (x+log λe), x ≥ − log λe,
− λe
otherwise.
e
(cid:1)2, while the log-characteristic function,
(60)
λe−1 e
0,
The mean and the variance of x are E0x = 1 − λ−1
see (20), is
fx,0(x) =
This function admits the series expansion Φx,0(t) =(cid:80)∞
(cid:26) (cid:0)1 − λ−1
(cid:2)(cid:0)1 − λ−1
Φx,0(t) = log
ϕn,0 =
λe − j t (λe − 1)
.
i=1 ϕn,0tn, with coefficients
e − log λe
(cid:1) j, n = 1,
,
n > 1,
(cid:1) j(cid:3)n
e
1
n
λ1−j t
e
and radius of convergence τx,0 = λe
λe−1.
(61)
(62)
-1.5-1-0.500.511.500.20.40.60.81 16
Fig. 5. Exponential example under hypothesis H0, with λe = 5. It is shown the CDF Fy,0(y) of the steady-state variable yk(∞) for node k = 3 (solid
curves) and k = 9 (dashed), obtained by the approximations developed in this paper. The symbols "o" show the results of computer experiments.
Fig. 6. Exponential example under hypothesis H1, with λe = 5. It is shown the CDF Fy,1(y) of the steady-state variable yk(∞) for node k = 3 (solid
curves) and k = 9 (dashed), obtained by the approximations developed in this paper. The symbols "o" show the results of computer experiments.
Using (61)-(62) in (27) gives the CDF Fu,0(u), u ∈ (cid:60), of the continuous component uk(∞), which according to (48) can
be used in combination with the second-order approximation developed for zk(∞) in Sec. IV-B, to yield the CDF Fy,0(y),
y ∈ (cid:60), of the steady-state variable yk(∞).
k,h [see condition (38)]. Clearly, ¯n = (cid:98) τx,0
to k,h, since the PDF of uk(∞) is unimodal and smooth, a reasonable choice is k,h = 10−1(cid:112)Vh[uk(∞)], as done in the
Actually, to get the approximate CDF Fu,0(u) and the final Fy,0(y) we need to set the values of δ, ¯n, ¯m [see (27)], and
ηkδ − 1(cid:99) from (26), while the choice of δ and ¯m is discussed in Appendix D. As
(cid:40) 1
Consider now hypothesis H1. In this case the PDF of x is
λe−1 (x+log λe), x ≥ − log λe,
− 1
Gaussian case.
fx,1(x) =
λe−1 e
0,
otherwise.
The mean and the variance of this random variable are E1x = λe − 1− log λe, and V1x = 5− 2 λe + λ2
function corresponding to PDF (63) can be easily computed, yielding:3
λ−j t
which again admits a series expansion in the form Φx,1(t) =(cid:80)∞
Φx,1(t) = log
e
1 − j t (λe − 1)
,
i=1 ϕn,1tn, with coefficients
(cid:26) (λe − 1 − log λe) j, n = 1,
1
n [(λe − 1) j ]n ,
n > 1,
ϕn,1 =
e. The log-characteristic
(63)
(64)
(65)
3Alternatively, a known property of the log-likelihood ratio can be exploited [70, Eq. (90), p. 44]: For x ∈ (cid:60), let fx,h(x) be the PDF of the log-likelihood
ratio x. Then, one gets fx,1(x) = fx,0(x)ex, which in terms of log-characteristic functions becomes Φx,1(t) = Φx,0(t − j).
-1-0.500.5100.20.40.60.81-1012300.20.40.60.81 17
Fig. 7. ROCs of agents k = 3 and k = 9 for the Gaussian example of Sec.V-A, with µ = 0.1 and different combinations of ρ and ak. Solid and dashed
curves refer to the theoretical expressions derived in Sec. IV while the small circles denote simulation points.
λe−1. As before, exploiting these formulas in Theorem 1 gives the desired CDF Fu,1(u)
and radius of convergence τx,1 = 1
of the steady-state continuous component uk(∞), which combined as shown in (48) with the approximation developed for
zk(∞), yields the CDF Fy,1(y). The quantities δ, ¯n, ¯m, and k,h, are computed as under H0.
The theoretical CDFs Fy,0(y) and Fy,1(y) are compared to the results of Monte Carlo experiments in Figs. 5 and 6. Figure 5
shows the CDF of yk(∞) under H0, for the highly-connected node k = 3 with 5 neighbors, and for the weakly connected
node k = 9 with only one neighbor. In these experiments we set λe = 5, yielding (1− pf ) ≈ 0.87, µ = 0.1, and consider three
different values of ak = 0.1, 0.25, 0.5. The curves represent the theoretical CDFs, with solid lines for k = 3, and dashed lines
for k = 9. The symbol "o" denotes values obtained by computer experiments. The prescriptions of Theorem 2 are confirmed
by the curve in green, which refers to µ = 0.01 and ak = 0.99, and by the correspondent simulation points obtained with
n = 1000 iterations. The matching between theory and simulation points is practically perfect.
The same comparison between theoretical predictions and results of computer experiments is conducted under H1, in which
case pd ≈ 0.67. The results are reported in Fig. 6. Again, we see that the matching between theory and simulation is very
satisfying. We also note that the effect of the network topology, encoded in Nk, and the system parameters ak and µ, is as
predicted, see discussion in Sec. IV-C.
C. Detection Performance
Consider the statistical test (5) at steady-state. We define the system-level false alarm probability of node k as Pf (cid:44)
P0(yk(∞) > γ) = 1 − Fy,0(γ), and the system-level detection probability of node k as Pd (cid:44) P1(yk(∞) > γ) = 1 − Fy,1(γ).
The quantities Pf and Pd depend on the node index k and on the threshold γ ∈ (cid:60), but for notational convenience these
dependencies are not made explicit. By computing Fy,0(γ) for a wide range of values of γ, the system designer may choose
the threshold value, say it γ∗, such that a prescribed value of false alarm P ∗
f is achieved. Then, the corresponding value of
d = 1 − Fy,1(γ∗) can be computed. The function relating Pd to Pf can be obtained by varying γ∗,
detection probability P ∗
and is usually referred to as the ROC (receiver operating characteristic) [70]. We see that the ROCs of the diffusion network
under one-bit messaging are easily derived by exploiting the theoretical CDFs Fy,h(y), h = 0, 1, obtained in this paper. The
effect on the ROCs of the system parameters -- node connectivity Nk, self-combination coefficient ak, and step-size µ --
can be inferred by the effect of these parameters on the CDFs, which was discussed in the previous sections.
Throughout this section the step size µ is maintained constant to the typical value 0.1, see [41], [42]. Figures 7 and 8
show the system-level ROC of node k for the Gaussian and for the exponential cases, respectively. Consider first the Gaussian
case in Fig. 7, and let us focus on node k = 3 (solid curves) and ρ = 0.1. We see that the ROC curve is rather nontrivial
when ak = 0.1, and tends to be smoother, closer to a concave curve, and closer to the ideal curve (Pd = 1 for all Pf > 0)
for ak = 0.25. The irregular non-concave behavior4 of the ROC is an immediate consequence of the irregular behavior of
the CDFs Fy,h(y), h = 0, 1, namely of the highly multimodal densities arising from the interaction between the continuous
and the discrete components. For small values of ak much credit is given to information coming from the neighbor nodes.
Since the information coming from the neighbor agents is in the form of marginal decisions, the more irregular shape of the
ROCs can be explained by the stronger effect of the discrete component with respect to the continuous one. When ak grows
the distributions of the steady-state yk(∞) under the two hypotheses tend to be more separated and indeed the performance
improves. Recall that the smaller ak is, the better the adaptive properties of the system become. Thus, the improved detection
performance obtained when ak grows is paid in terms of a slower adaptivity to changes in the state of nature, which is a
4A concave ROC can be obtained by randomization, see [3].
00.20.40.60.8100.20.40.60.81 18
Fig. 8. ROCs of agents k = 3 and k = 9 for the exponential example of Sec. V-B, with µ = 0.1 and different combinations of λe and ak. Solid and dashed
curves refer to the theoretical expressions derived in Sec. IV while the small circles denote simulation points.
manifestation of an inherent system tradeoff. The same behavior is observed for ρ = 0.5. Clearly, the larger ρ is, the larger
the detection performance becomes and indeed the ROC curves for ρ = 0.5 stand above the curves for ρ = 0.1.
Consider next the ROC curves of the less-connected node k = 9 (dashed lines). These curves show the same qualitative
behavior as node 3. However, the lower connectivity of node 9 is paid in terms of a worse detection performance, which shows
the effect of the network topology and the benefit of a stronger system connectivity.
It is worth noting from Fig. 7 that the simulation points closely fit the theoretical curves, and some discrepancy can be only
observed for node k = 9 at small ρ and large ak. In general, the accuracy of the theoretical predictions improves with a larger
number of agent neighbors, larger values of ρ, and smaller values of ak.
The same analysis conducted for the Gaussian example is repeated in Fig. 8 for the exponential example of Sec. V-B, and
similar considerations apply.
VI. EXTENSIONS
In this section we briefly discuss two possible extensions of our work. First, in many practical scenarios, especially when
the network covers wide geographical areas, the assumption of identical distribution of the data collected by the remote agents
may not be valid. The diffusion rule (9a)-(9b), however, remains unaltered because the one-bit messages received by agent k
are binary flags denoting the local decisions of its neighbors. These flags can be incorporated in (9b) as E1x or E0x, where
the expectations are taken with respect to the statistical distributions of agent k's observations, regardless of the distributions
characterizing the observations of the agents that originate the messages. One difference with respect to the analysis developed
in the previous sections is that the marginal detection and false alarm probabilities, see (7), can differ from agent to agent. Thus,
the characterization of the discrete component in Sec. IV-B must be revisited in order to take into account the non-identical
distributions of the normalized binary variables b(cid:96)(i) appearing in (32).
Another case of practical relevance is that of multihypothesis tests, in which the task of the network is to decide among H > 2
possible states of nature H0, . . . ,HH−1. To see how the tools developed in this article can be exploited in this scenario, recall
from (8) that the messages exchanged among agents can be interpreted as the agents' marginal decisions. Accordingly, let us
k (n) = log fr,h(rk(n)), h = 0, . . . , H−1, in place of the single variable
assume that there are H marginal decision statistics x(h)
shown in (6), and the message delivered by agent (cid:96) at time n consists of its local decision κ(cid:96)(n) = arg maxh=0,...,H−1 x(h)
k (n),
k (n), h = 0, . . . , H − 1, can be defined, and we have the
with κ(cid:96)(n) ∈ {0, . . . , H − 1}. Then, H local state variables y(h)
following straightforward extension of (9a)-(9b):
(cid:88)
k (n − 1) + µ[x(h)
k (n) − y(h)
k (n − 1)],
ak(cid:96) Eκ(cid:96)(n)x(h), n ≥ 1,
v(h)
k (n) = y(h)
y(h)
k (n) = akkv(h)
(66a)
(66b)
k (n) +
(cid:96)(cid:54)=k
for h = 0, . . . , H − 1. The decision of agent k at steady-state is arg maxh=0,...,H−1 y(h)
k (∞). Note that in the diffusion
scheme (66a)-(66b), each of the H local state variables {y(h)
k (n)} evolves exactly as does the single variable yk(n) dealt with
in the previous sections, and the analysis is similar. In particular, the analysis of the continuous component closely follows
the one developed in this article, while the analysis of the discrete component requires further elaboration to generalize the
methods of Sec. IV-B to the presence of H > 2 local decisions. The details are left for future investigations.
Pf00.20.40.60.81Pd00.20.40.60.81λe=2,ak=.1(node3)λe=2,ak=.25(node3)λe=5,ak=.1(node3)λe=5,ak=.25(node3)λe=2,ak=.1(node9)λe=2,ak=.25(node9)λe=5,ak=.1(node9)λe=5,ak=.25(node9) 19
VII. SUMMARY & FUTURE WORK
Inter-agent communication constraints is an important aspect of multi-agent systems for on-line adaptation and learning. In
this work we consider a suitably modified version of the classical ATC diffusion rule, under the constraint that the inter-agent
links support only one-bit messages. For the proposed one-bit messaging network, a careful investigation of the steady-state
distribution of the agents' state is carried on, and the resulting performance curves reveal sophisticated/nontrivial behavior
which is predicted well by the derived theoretical expressions. The role played by the main system parameters -- step-size µ,
self-combination coefficients {ak}, and network topology Nk -- is investigated.
The results provided in this paper can be exploited for the design and analysis of more complex decision systems, some of
which are mentioned in Sect. VI. Further examples of these possible future extensions include the analysis of networks where
the messages can be delivered to farther agents by multi-hop communications, managing the tradeoff between the number of
delivered bits per link usage and the number of channel accesses, and addressing more advanced hypothesis testing problems
(composite hypotheses, sequential detection, robust techniques). Incorporating security issues is another important direction for
future studies such as addressing the presence of adversary agents (e.g., Byzantine sensors [71]) that conspire for impairing
the network by delivering malicious messages.
Consider Eq. (15). Before letting n → ∞, the following equality in distribution holds
APPENDIX A
PROOF OF THEOREM 1
n(cid:88)
n(cid:88)
k
ηi−1
k xk(i).
k xk(n − i + 1) d=
ηi−1
i=1
i=1
i=1 ηi−1
of (A.1) for the linearity of the almost everywhere limit [72].
k) < ∞, and this proves the convergence with probability one of(cid:80)n
(A.1)
k xk(i)] → Ehx/(1 − ηk), for n → ∞. Thus, if the sum
(xk(i) − Ehx) of zero-mean independent variables converges with probability one, so does the right-hand side
Recalling from (12) that ηk ∈ (0, 1), we have Eh[(cid:80)n
(cid:80)n
i=1 ηi−1
fact, it converges with probability one if the sum of the variances converges [58, Th. 22.6]. In our case limn→∞(cid:80)n
The sum of zero-mean independent random variables converges either with probability one or with probability zero and, in
shows that(cid:80)n
(cid:12)(cid:12)(cid:12) ϕn,h
To prove (23), suppose that (22) holds. The series on the right-hand side of (22) converges absolutely and uniformly for
t < τx,h [73, Th. 93] [74, 10.5, p. 198]. A necessary and sufficient condition for the absolute and uniform convergence
in the interval t < τx,h of this series, is that its coefficients {ϕn,h} satisfy lim supn→∞ ϕn,h 1
x,h, see e.g., [73, Th.
94]. The quantity τ−1
k )− 1
n = 1, yielding
Ehx)] = Vhx/(1− η2
k xk(i) to the random variable
k, defined in (15). Convergence with probability one implies convergence in distribution [72] and, again from (15), this
w(cid:63)
i=1 ηi−1
k xk(n − i + 1), hence uk(n), converges in distribution.
i=1 ηi−1
n = τ−1
x,h is interpreted as 0 when τx,h = ∞. For all 0 < ηk < 1 we have limn→∞(1 − ηn
x,h. Thus, the radius of convergence of the series
lim supn→∞
Vh[ηi−1
= τ−1
(cid:12)(cid:12)(cid:12) 1
1−ηn
i=1
k
n
k
(xk(i)−
∞(cid:88)
n=1
ϕn,h
1 − ηn
k
tn
(A.2)
(A.3)
(A.4)
is τx,h, namely, (A.2) converges absolutely and uniformly for t < τx,h.
To show that the series (A.2) represents the log-characteristic function of w(cid:63)
k, note that, from (17),
Therefore, for all t ∈ (cid:60), the following linear functional equation in the unknown Φw,h(t) holds:
Φw,h(t) = log Ehejtw(cid:63)
= log Ehejtηkw(cid:63)
k = log Ehejt(ηkw(cid:63)
k + log Ehejtx.
k+x)
Φw,h(t) − Φw,h(ηk t) = Φx,h(t).
We now substitute the function Φw,h(t) appearing in (A.4) by the power series (A.2). Before, note that:
(cid:12)(cid:12)(cid:12)(cid:12) 1
(cid:12)(cid:12)(cid:12)(cid:12) ηn
which shows that the radius of convergence of(cid:80)∞
∞(cid:88)
∞(cid:88)
1 − ηn
lim
n→∞
n=1
k
k
n
(cid:12)(cid:12)(cid:12)(cid:12) 1
(cid:12)(cid:12)(cid:12)(cid:12) ηn
∞(cid:88)
be added term by term [73, Th. 3, p. 134], using (A.2) in place of Φw,h(t), from (A.4) one gets:
= ηk ⇒ lim sup
n→∞
k tn is τx,h
ηn
ηk
ϕn,h
1−ηn
k
n
=
k ϕn,h
1 − ηn
> τx,h. Since absolutely convergent power series can
ηk
τx,h
(A.5)
k
,
tn −
ϕn,h
1 − ηn
k
ηn
k tn =
n=1
n=1
ϕn,htn = Φx,h(t),
ϕn,h
1 − ηn
k
∞(cid:88)
=
n=1
n=1
ϕn,h
1 − ηn
t < τx,h,
k
(1 − ηn
k )tn
(A.6)
20
k
n=1
ϕn,h
1−ηn
which shows that (A.4) is verified in the interval t < τx,h, thus proving (23).
The solution Φw,h(t) =(cid:80)∞
power series in a neighbor of the origin t < t0, for some t0 > 0. Indeed, suppose that g(t) =(cid:80)∞
in that class. Then, g(ηk t) =(cid:80)∞
obtaining g(t) − g(ηk t) =(cid:80)∞
tn to (A.4) is unique among the class of functions that admit an absolutely convergent
n=1 dntn is another solution
k tn is absolutely convergent for t < t0/ηk. We can add term by term these series,
n=1 ϕn,htn, where the last equality holds because, by assumption, g(t)
k ), for
solves (A.4) in t < t0. The uniqueness of the power series expansion [73, Th. 97, p. 172] implies dn = ϕn,h/(1 − ηn
all n. Because the coefficients are the same, we conclude that g(t) = Φw,h(t) for t < τx,h.
k )tn =(cid:80)∞
For τx,h = ∞, Eq. (24) immediately follows by a known formula for computing the CDF of w(cid:63)
k, assumed everywhere
continuous, from its characteristic function exp[Φw,h(t)] [75], [76, Eq. (2)], upon recalling from (16) that the distribution of
uk(∞) is simply a scaled version of the distribution of w(cid:63)
k.
To arrive at (25), let 0 < τx,h ≤ ∞. From [63], [75, Eq. (7)] we have a formula for obtaining the CDF Fw,h(u), u ∈ (cid:60),
k from the corresponding log-characteristic function Φw,h(t), t ∈ (cid:60). Expressed in terms of Fu,h(u) = Fw,h
n=1 dn(1 − ηn
(cid:1), the
(cid:0) u
n=1 dnηn
µak
of w(cid:63)
formula is:
Fu,h(u) =
1
2
− 2
π
Ωn(u, δ) + χu(u, δ),
(A.7)
n=0
where Ωn(u, δ) is given in (21), and the "error" term verifies:
(cid:17)(cid:111)
(A.8)
(cid:80)∞
Fix (cid:48) > 0, and suppose that δ is such that the second condition in (26) is verified. Then, from (A.7)-(A.8) we see that the
n=0 Ωn(u, δ) is ≤ (cid:48)/2. Next, suppose
absolute value of the error incurred when computing Fu,h(u) by the expression 1
that δ also verifies the first condition in (26), namely truncating the series implies an error not larger than (cid:48)/2. Then:
χu(u, δ) ≤ max
u − 2πakµ
u − 2πakµ
, 1 − Fw,h
2 − 2
(cid:110)
(cid:16)
(cid:17)
(cid:16)
Fu,h
δ
δ
π
.
∞(cid:88)
¯n(cid:88)
n=0
where we have denoted (cid:98) τx,h
Consider now the function Φw,h(·) appearing at the exponent of (21). We have:
ηkδ − 1(cid:99) by ¯n. In (A.9), by triangular inequality, we have ∆u(u, δ) < (cid:48).
Fu,h(u) =
1
2
− 2
π
Ωn(u, δ) + ∆u(u, δ),
Φw,h(t) = Φx,h(t) + Φw,h(ηk t)
ϕm,h
1 − ηm
= Φx,h(t) +
m=1
k
∞(cid:88)
ηm
k tm,
where (A.10a) is the same of (A.4), and the series expansion in (A.10b) has radius of convergence τx,h/ηk because of (A.5).
Noting that Ωn(u, δ) in (21) involves Φw,h((2n+1)δ/2), whose argument lies in the region of convergence of expansion (A.10b)
for all n ≤ (cid:98) τx,h
ηkδ − 1(cid:99), we can use (A.10b) in the expression of Ωn(u, δ) appearing in (A.9), yielding (25), which proves claim
iii).
The material in this section is adapted from [58, Theorems 27.2 and 27.3]. From (13), let us denote by
APPENDIX B
PROOF OF THEOREM 2
the zero-mean versions of xk(n) and (cid:101)xk(n), respectively, and consider the normalized (zero-mean unit-variance) process
(A.9)
(A.10a)
(A.10b)
(B.1)
(B.2)
where [see (50), (53) and (54)]:
¯yk(n) =
yk(n) − Eh[yk(n)]
(cid:101)x(cid:48)
k(n) = (cid:101)xk(n) − Eh(cid:101)x,
k(n) = xk(n) − Ehx,
x(cid:48)
n(cid:88)
(cid:112)Vh[yk(n)]
¯wk(n, i) (cid:44) wk(n, i) − Eh[wk(n, i)]
(cid:88)
sk(n)
k(n − i + 1) +
µakx(cid:48)
i=1
=
=
ηi−1
sk(n)
k
ak(cid:96)(cid:101)x(cid:48)
(cid:96)(cid:54)=k
¯wk(n, i),
(cid:96)(n − i + 1)
.
(B.3)
(B.4)
For t ∈ (cid:60), let φ(t; n, i) (cid:44) Ehejt ¯wk(n,i) be the characteristic function of ¯wk(n, i), where j =
function of condition C. For sufficiently small > 0, and each fixed t ∈ (cid:60), one gets
¯w2
(cid:18)
1 − t2
2
≤ Eh
≤ Eh
≤ Eh
Vh[ ¯wk(n, i)]
1 + jt ¯wk(n, i) − t2
2
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)φ(t; n, i) −
(cid:18)
(cid:12)(cid:12)(cid:12)(cid:12)ejt ¯wk(n,i) −
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:2)min{t ¯wk(n, i)2,t ¯wk(n, i)3/6}(cid:3)
(cid:3) /6
(cid:2)t ¯wk(n, i)3I ¯wk(n,i)<
(cid:2)t ¯wk(n, i)2I ¯wk(n,i)≥
(cid:3)
(cid:3)
(cid:2) ¯w2
k(n, i)I ¯wk(n,i)≥
(cid:12)(cid:12)(cid:12)(cid:12) ≤ min{tz2,tz3/6},
+ Eh
t3Vh[ ¯wk(n, i)] + t2 Eh
≤ 1
6
k(n, i)
2
t2z2)
k(n, i) , which gives (B.9).
(cid:12)(cid:12)(cid:12)(cid:12)ejtz − (1 + jtz − 1
(cid:12)(cid:12)(cid:12)(cid:12)φ(t; n, i) −
(cid:18)
n(cid:88)
n(cid:88)
(cid:2) ¯w2
(cid:12)(cid:12)(cid:12)(cid:12)φ(t; n, i) −
(cid:18)
n(cid:88)
(cid:2) ¯w2
t3 + t2
1 − t2
2
n(cid:88)
≤ 1
6
lim
n→∞
Eh
i=1
i=1
lim
ηk→1
lim
n→∞
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)
Vh[ ¯wk(n, i)]
k(n, i)I ¯wk(n,i)≥
(cid:3) .
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) = 0,
(cid:3) = 0,
(cid:2) ¯w2
(cid:34)
k(n, i)I ¯wk(n,i)≥
Eh
≤ Eh
¯w2
(cid:35)
i=1
Eh
k(n, i)
(cid:19)β I ¯wk(n,i)≥
k(n, i)I ¯wk(n,i)≥
(cid:3)
(cid:18) ¯wk(n, i)
(cid:2) ¯wk(n, i)2+βI ¯wk(n,i)≥
(cid:2) ¯wk(n, i)2+β(cid:3) .
n(cid:88)
(cid:2) ¯wk(n, i)2+β(cid:3) = 0.
Eh
(cid:3)
1
=
β
≤ 1
β
Eh
Eh
lim
ηk→1
lim
n→∞
i=1
i=1
Since > 0 can be made as small as desired, this means
for all t ∈ (cid:60), provided that the Lindeberg condition: ∀ > 0,
lim
ηk→1
1 − t2
2
Vh[ ¯wk(n, i)]
is verified.
Now, whatever is, for any β > 0
21
√−1. Denote by IC the indicator
(B.5)
(B.6)
(B.7)
(B.8)
(B.9)
(B.11)
(B.12)
(B.13)
(B.14)
(B.15)
where (B.6) follows by Jensens' inequality [59]; inequality (B.7) follows by the relationship
(B.10)
which holds for any z ∈ (cid:60), see, e.g. [58, 26.42]; inequality (B.8) is valid for any > 0; finally, for ¯wk(n, i) < we have
¯wk(n, i)3 = ¯w2
k(n, i) ¯wk(n, i) < ¯w2
Summing over i = 1, . . . , n, inequality (B.5)-(B.9) yields
This shows that Lindeberg condition (B.13) is implied by the following Lyapunov condition: there exists β > 0 such that
Lemma 1 in Appendix C shows that (B.15) holds true. Therefore, we now proceed to conclude the proof of the theorem,
assuming (B.15), which implies (B.13), which in turn implies (B.12).
The characteristic function of ¯yk(n) is Ehejt(cid:80)n
note that e−t2/2 =(cid:81)n
i=1 φ(t; n, i), and we want to show that such characteristic
function converges to the characteristic function Eh[ejt g] = e−t2/2 of a standard Gaussian random variable g. To show this,
i=1 e−Vh[ ¯wk(n,i)]t2/2, and consider the following:
22
e−Vh[ ¯wk(n,i)]t2/2
i=1
i=1
=
=
i=1
i=1
1 − t2
2
i=1 ¯wk(n,i) =(cid:81)n
i=1 ¯wk(n,i) − e−t2/2(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)Ehejt(cid:80)n
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n(cid:89)
φ(t; n, i) − n(cid:89)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n(cid:89)
(cid:18)
φ(t; n, i) − n(cid:89)
(cid:18)
n(cid:89)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n(cid:89)
φ(t; n, i) − n(cid:89)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n(cid:89)
(cid:18)
(cid:2) ¯w2
(cid:2) ¯w2
Vh[ ¯wk(n, i)] = Eh[ ¯w2
= Eh
≤ Eh
k(n, i)]
k(n, i)I ¯wk(n,i)≥
k(n, i)I ¯wk(n,i)≥
Vh[ ¯wk(n, i)]
Vh[ ¯wk(n, i)]
1 − t2
2
1 − t2
2
1 − t2
2
(cid:18)
(cid:19)
≤
i=1
i=1
i=1
+
i=1
+
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:19)
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Vh[ ¯wk(n, i)]
i=1
− n(cid:89)
(cid:19)
− n(cid:89)
i=1
Vh[ ¯wk(n, i)]
e−Vh[ ¯wk(n,i)]t2/2
e−Vh[ ¯wk(n,i)]t2/2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) .
(cid:3)
k(n, i)I ¯wk(n,i)<
(cid:2) ¯w2
(cid:3) + Eh
(cid:3) + 2.
(cid:2) ¯w2
k(n, i)I ¯wk(n,i)≥
(cid:3) = 0,
(B.16)
(B.17)
(B.18)
(B.19)
(B.21)
(B.22)
Thus, we have to show that both moduli appearing in (B.16) converge to zero. Consider the first. We know that φ(t; n, i) ≤ 1,
because the modulus of a characteristic function is never larger than unity. In addition, ∀ > 0,
From (B.13) we see that
n→∞ max
lim
which, used in (B.17), for the arbitrariness of , gives
lim
ηk→1
i=1,...,n
Eh
lim
ηk→1
lim
n→∞ max
i=1,...,n
Vh [ ¯wk(n, i)] = 0.
As a consequence of (B.19), there exist a value of n sufficiently large and a value of ηk sufficiently close to one such that,
for all larger n and ηk closest to one, we have:
(B.20)
Then, we can apply the following inequality, valid for complex numbers ξ1, . . . , ξn, ζ1, . . . , ζn, satisfying ξi ≤ 1, ζi ≤ 1,
i = 1, . . . , n:
i = 1, . . . , n.
0 ≤ 1 − t2
2
Vh[ ¯wk(n, i)] ≤ 1,
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n(cid:89)
ξi − n(cid:89)
see [58, Lemma 1, p. 358]. This yields (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n(cid:89)
φ(t; n, i) − n(cid:89)
(cid:12)(cid:12)(cid:12)(cid:12)φ(t; n, i) −
≤ n(cid:88)
i=1
i=1
i=1
i=1
ζi
i=1
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ n(cid:88)
(cid:18)
(cid:18)
1 − t2
2
1 − t2
2
i=1
ξi − ζi,
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) ,
Vh[ ¯wk(n, i)]
Vh[ ¯wk(n, i)]
which converges to zero for n → ∞ followed by ηk → 1, in view of (B.12).
Consider next the second modulus appearing in (B.16). Fix an arbitrary > 0. We have:
i=1
1 − t2
2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n(cid:89)
(cid:18)
(cid:12)(cid:12)(cid:12)(cid:12)1 − t2
≤ n(cid:88)
n(cid:88)
n(cid:88)
≤ t4
8
≤ t4
8
V2
i=1
i=1
2
i=1
(cid:19)
− n(cid:89)
Vh[ ¯wk(n, i)]
e−Vh[ ¯wk(n,i)]t2/2
i=1
Vh[ ¯wk(n, i)] − e−Vh[ ¯wk(n,i)]t2/2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)
h[ ¯wk(n, i)]
Vh[ ¯wk(n, i)] =
t4
8
.
23
(B.23)
(B.24)
(B.25)
(B.26)
(C.1)
(cid:3)
(C.2)
(C.3)
i=1
In the above, (B.24) follows by using again (B.20) and (B.21); inequality (B.25) is a consequence of the bound 1 − z − e−z ≤
Vh[¯yk(n)] =(cid:80)n
z2/2, valid for all z ≥ 0; for any > 0, the inequality in (B.26) follows by (B.19); the equality in (B.26) follows from
So far, we have shown that the characteristic function Ehejt(cid:80)n
Vh[ ¯wk(n, i)] = 1. Since is arbitrary, we see that (B.23) goes to zero when n → ∞ followed by ηk → 1.
i=1 ¯wk(n,i) of the random variable ¯yk(n) converges to the
characteristic function E[ejt g] = e−t2/2 of a standard Gaussian random variable g. Continuity theorem [58, p. 349] states that
¯yk(n) converges in distribution to g if, and only if, the characteristic function of ¯yk(n) converges to the characteristic function
of g, for each t ∈ (cid:60). This shows that the distribution of yk(n)−mk,h(n)
converges to the standard Gaussian distribution, when
n → ∞ followed by ηk → 1, which concludes the proof.
sk,h(n)
APPENDIX C
LYAPUNOV CONDITION
The proof of Theorem 2 developed in Appendix B is founded on the following result.
Lemma 1 (Lyapunov condition): Suppose that there exists β > 0 such that Eh[x2+β] < ∞. Then, we have
(cid:2) ¯wk(n, i)2+β(cid:3) = 0.
Eh
n(cid:88)
i=1
lim
ηk→1
lim
n→∞
Proof. For z, ν ∈ (cid:60) with ν ≥ 1, the function zν is convex. This implies, for z1, z2 ∈ (cid:60),
z1 + z2ν ≤ 2ν−1 (z1ν + z2ν) .
From (B.4), using the notations (B.1) and (B.2), and exploiting (C.2) with ν = 2 + β, one gets
¯wk(n, i)2+β =
η(i−1)(2+β)
k
s2+β
(n)
k
(cid:40)
≤ 21+β η(i−1)(2+β)
k
s2+β
k
(n)
+
(cid:96)(cid:54)=k
(cid:96)(n − i + 1)
k(n − i + 1)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) µ ak x(cid:48)
(cid:88)
ak(cid:96)(cid:101)x(cid:48)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:88)
+ M 2+β(1 − ak)2+β(cid:9),
k(n − i + 1)2+β
x(cid:48)
ak(cid:96)(cid:101)x(cid:48)
x(cid:48)
k(n − i + 1)2+β
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2+β
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2+β
(cid:96)(n − i + 1)
(cid:96)(cid:54)=k
+
k
k
µ2+βa2+β
(cid:110)
≤ 21+β η(i−1)(2+β)
k
s2+β
k
(n)
µ2+βa2+β
where M/2 (cid:44) max(E0x,E1x). In (C.3) we have exploited (cid:80)
(cid:96) ≤ (cid:80)
(1 − p)E0x, with p equal to pd under H1, and equal to pf under H0. This gives (cid:101)x(cid:48)
therefore (cid:80)
(cid:96)(cid:54)=k ak(cid:96)(cid:101)x(cid:48)
(cid:96) ≤ M (1 − ak). From inequality (C.3) it follows
n(cid:88)
(cid:96)(cid:54)=k ak(cid:96)(cid:101)x(cid:48)
(cid:96)(cid:54)=k ak(cid:96)(cid:101)x(cid:48)
(cid:96) ≤(cid:80)
24
(cid:96)(cid:54)=k ak(cid:96)(cid:101)x(cid:48)
(cid:96) ≤ (cid:101)x(cid:96) + pE1x +
(cid:96) ≤ (cid:101)x(cid:96) + max(E0x,E1x) ≤ M, and
(cid:96), and (cid:101)x(cid:48)
i=1
s2+β
k
k
k
k
k
i=1
(n)
(n)
21+β
k
s2+β
k
µ2+βa2+β
µ2+βa2+β
Eh[x(cid:48)2+β]
1 − ηn(2+β)
1 − η2+β
21+β η(i−1)(2+β)
Eh ¯wk(n, i)2+β ≤ n(cid:88)
Eh[x(cid:48)2+β] + M 2+β(1 − ak)2+β(cid:111)
×(cid:110)
(cid:110)
+M 2+β(1 − ak)2+β(cid:9) .
1+ β
2(cid:18) 1 − η2n
(cid:88)
(cid:96) to the real sequences ξ(cid:96) = ak(cid:96) and ζ(cid:96) = 1, gives(cid:80)
(cid:19)1+ β
2(cid:18) 1 − η2n
µ2a2
(cid:96) ξ(cid:96)ζ(cid:96))2 ≤(cid:80)
(cid:18)
Vhx + Vh(cid:101)x
(cid:80)
Vhx + Vh(cid:101)x
(cid:19)1+ β
(cid:19)1+ β
1 − η2
(1 − ak)2
S − 1
1 − η2
µ2a2
k
a2
k(cid:96)
(cid:96)(cid:54)=k
k
k
k
k
.
,
2
2
(n) ≥
k
(C.4)
(C.5)
(cid:96)(cid:54)=k a2
k(cid:96) ≥
=
Next, from (54):
s2+β
k
Applying Cauchy-Schwarz inequality ((cid:80)
S−1 ((cid:80)
(cid:96)(cid:54)=k ak(cid:96))2 = (1−ak)2
S−1
1
(n) =
(cid:96) ζ 2
. Using this latter inequality in (C.5):
(cid:96) ξ2
(cid:96)
which, inserted in (C.4), finally gives
s2+β
k
n(cid:88)
i=1
Eh ¯wk(n, i)2+β ≤ 21+βΨ(µ, ak) Γ(ηk, n),
where
and
(cid:16)
Ψ(µ, ak) (cid:44) µ2+βa2+β
µ2a2
k
k
Eh[x(cid:48)2+β] + M 2+β(1 − ak)2+β
(cid:17)1+ β
Vhx + Vh(cid:101)x (1−ak)2
(cid:1)1+ β
(cid:0)1 − η2
S−1
2
2
,
k
1 − η2+β
k
.
Γ(ηk, n) (cid:44) 1 − ηn(2+β)
k )1+ β
The function Ψ(µ, ak) in (C.8) can be upper bounded as follows:
(1 − η2n
k
2
(C.6)
(C.7)
(C.8)
(C.9)
(C.10)
(C.11)
Ψ(µ, ak) =
(cid:16)
k
Eh[x(cid:48)2+β]
µ2+βa2+β
Vhx + Vh(cid:101)x (1−ak)2
µ2a2
k
(cid:17)1+ β
2
(cid:16)
+
µ2a2
k
2
+
S−1
M 2+β(1 − ak)2+β
S−1
M 2+β(1 − ak)2+β
(cid:17)1+ β
Vhx + Vh(cid:101)x (1−ak)2
(cid:17)1+ β
(cid:16)Vh(cid:101)x (1−ak)2
(cid:18) M 2(S − 1)
(cid:19)1+ β
Vh(cid:101)x
2(cid:35)
(cid:18) M 2(S − 1)
(cid:19)1+ β
Vh(cid:101)x
(cid:19)1+ β
S−1
+
+
.
,
2
2
2
≤ µ2+βa2+β
(µ2a2
k
k
Eh[x(cid:48)2+β]
Vhx)1+ β
2
Eh[(x(cid:48)2)1+ β
2 ]
(Eh[x(cid:48)2])1+ β
2
(cid:34)(cid:18) x(cid:48)2
Eh[x(cid:48)2]
Ψmax (cid:44) Eh
yielding 0 ≤ Ψ(µ, ak) ≤ Ψmax, where
=
Now, it is easily seen that Eh[x(cid:48)2] > 0 and Vh(cid:101)x > 0. The former is because Eh[x(cid:48)2] = 0 would imply x = Ehx with
probability one, and we have instead assumed that x has a density. The latter follows by the assumptions E1x (cid:54)= E0x and
pd, pf (cid:54)= 0, 1. In addition, the condition Eh[x2+β] < ∞ ensures the finiteness of Eh[x(cid:48)2+β] [69]. Therefore, Ψmax in (C.11)
is finite: Ψmax < ∞.
As to Γ(ηk, n) in (C.9), simple algebra shows that 0 ≤ Γ(ηk, n) ≤ 1, yielding, from (C.7),
Eh ¯wk(n, i)2+β ≤ 21+βΨmax.
(C.12)
Thus, the left-hand side of (C.12) is a bounded series of nonnegative terms, and therefore it converges to a finite limit. Now,
using Ψ(µ, ak) ≤ Ψmax in (C.7), we see that
n(cid:88)
i=1
lim
n→∞
n(cid:88)
Therefore, from
we get (C.1):
Eh ¯wk(n, i)2+β ≤ 21+βΨmax Γ(ηk, n).
i=1
lim
ηk→1
n→∞ Γ(ηk, n) = lim
lim
ηk→1
(cid:0)1 − η2
k
(cid:1)1+ β
2
1 − η2+β
k
= 0,
n(cid:88)
(C.15)
It is worth noting that if one replaces the limit ηk → 1 with the limit µ → 0, Lyapunov condition (C.15) does not hold
lim
ηk→1
lim
n→∞
i=1
Eh ¯wk(n, i)2+β = 0.
anymore. This can be seen from (C.14) by observing that
(1 − a2
k)1+ β
1 − a2+β
which does not go to zero, unless the further condition ak → 1 is enforced.
lim
n→∞ Γ(ηk, n) =
lim
µ→0
k
2
,
(C.16)
25
(C.13)
(C.14)
(D.3)
(D.4)
(D.5)
.
(D.1)
(D.2)
log λe. Then, a sufficient
log λe [otherwise (D.1) is automatically
Let us start by the second condition in (26): for some (cid:48) > 0,
CHOICE OF δ AND ¯m FOR THE EXPONENTIAL EXAMPLE
APPENDIX D
(cid:18)
u − 2πakµ
δ
(cid:18)
2πakµ
u +
Fu,h
1 − Fu,h
≤ (cid:48)
(cid:19)
2
,
≤ (cid:48)
2
(cid:19)
δ
Since, with probability one, xk(∞) ≥ − log λe, from (15)-(16) we see that uk(∞) ≥ − akµ
1−ηk
condition ensuring (D.1) is u − 2πakµ/δ ≤ − akµ
1−ηk
verified for all δ > 0], this yields:
log λe. Assuming u ≥ − akµ
1−ηk
δ ≤
2π
.
u
akµ + log λe
1−ηk
As to (D.2), exploiting Chebyshev inequality [61], we have
= Ph
1 − Fu,h
2πakµ
u +
uk(∞) > u +
(cid:18)
(cid:19)
(cid:18)
δ
(cid:18)
(cid:18)
= Ph
≤ Ph
≤
uk(∞) − Eh[uk(∞)] > u − Eh[uk(∞)] +
uk(∞) − Eh[uk(∞)] > u − Eh[uk(∞)] +
Vh[uk(∞)]
(u − Eh[uk(∞)] + 2πakµ/δ)2 ,
(cid:19)
2πakµ
(cid:19)
(cid:19)
δ
2πakµ
2πakµ
δ
δ
where we have assumed u − Eh[uk(∞)] + 2πakµ
δ > 0. Now, the last line of (D.4) is not larger than (cid:48)/2 if
δ ≤
,
(cid:113) 2Vh[uk(∞)]
(cid:48)
2πakµ
+ Eh[uk(∞)] − u
assuming that right-hand side of (D.5) is positive.
Admittedly, bounds (D.3) and (D.5) are loose, but sufficient for our purposes. The criterion used in the numerical experiments
is to fix (cid:48), and then: if u ≥ − akµ
log λe we choose δ as the minimum between the right-hand sides of (D.3) and (D.5),
1−ηk
otherwise we choose δ equal to the right-hand side of (D.5). In our experiments (cid:48) = 2 10−5. For this value we have verified
numerically that the error on Fu,h(u) due to the series truncation [see first condition in (26)] is negligible. Finally, the value
of ¯m is selected using the empirical criterion η ¯m ≤ (cid:48)(cid:48), for some sufficiently small (cid:48)(cid:48). Namely, we set ¯m = (cid:100) log (cid:48)(cid:48)
(cid:101), and in
the numerical experiments we use (cid:48)(cid:48) = (cid:48) = 2 10−5.
log ηk
REFERENCES
26
[1] S. Marano and A. H. Sayed, "Decisions Under Binary Messaging over Adaptive Networks," in Proceedings of the 26th European Signal Processing
Conference (EUSIPCO 2018), Rome, Italy, Sept. 3-7 2018, pp. 415-419.
[2] E. L. Lehmann and J. P. Romano, Testing Statistical Hypotheses, 3rd ed. New York, NY, USA: Springer, 2005.
[3] H. V. Poor, An Introduction to Signal Detection and Estimation. New York: Springer-Verlag, 1988.
[4] A. H. Sayed, Adaptive Filters. NY: Wiley, 2008.
[5] K. J. Astrom and B. Wittenmark, Adaptive Control, 2nd ed. NY: Dover, 2008.
[6] H. Chernoff, "Sequential design of experiments," Annals Math. Statist, vol. 30, 1959.
[7] S. Haykin, "Cognitive radar: a way of the future," IEEE Signal Processing Magazine, vol. 23, no. 1, pp. 30 -- 40, Jan. 2006.
[8] M. Naghshvar and T. Javidi, "Active sequential hypothesis testing," The Annals of Statistics, vol. 6, no. 41, pp. 2703 -- 2738, 2013.
[9] S. Nitinawarat, G. K. Atia, and V. V. Veeravalli, "Controlled sensing for multihypothesis testing," IEEE Transactions on Automatic Control, vol. 58,
no. 10, pp. 2451 -- 2464, Oct. 2013.
[10] M. Franceschetti, S. Marano, and V. Matta, "Chernoff test for strong-or-weak radar models," IEEE Transactions on Signal Processing, vol. 65, no. 2,
[11] S. S. Pradhan, J. Kusuma, and K. Ramchandran, "Distributed compression in a dense microsensor network," IEEE Signal Processing Magazine, vol. 19,
[12] J. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, "A survey on sensor networks," IEEE Communications Magazine, vol. 40, no. 8, pp.
[13] C. Chong and S. Kumar, "Sensor networks: Evolution, opportunities, and challenges," Proceedings of the IEEE, vol. 91, no. 8, pp. 1247 -- 1256, Aug.
[14] Z.-Q. Luo, M. Gastpar, J. Liu, and A. Swami, "Distributed signal processing in sensor networks," IEEE Signal Processing Magazine, vol. 23, no. 4, pp.
pp. 289 -- 302, Jan. 2017.
pp. 51 -- 60, Mar. 2002.
102 -- 114, Aug. 2002.
2003.
14 -- 15, Jul. 2006.
56 -- 69, Jul. 2006.
407 -- 416, Feb. 2003.
pp. 16 -- 26, Jul. 2006.
[15] J. B. Predd, S. R. Kulkarni, and H. V. Poor, "Distributed learning in wireless sensor networks," IEEE Signal Processing Magazine, vol. 23, no. 4, pp.
[16] J.-F. Chamberland and V. V. Veeravalli, "Decentralized detection in sensor networks," IEEE Transactions on Signal Processing, vol. 51, no. 2, pp.
[17] B. Chen, L. Tong, and P. Varshney, "Channel-aware distributed detection in wireless sensor networks," IEEE Signal Processing Magazine, vol. 23, no. 4,
[18] S. Appadwedula, V. V. Veeravalli, and D. Jones, "Energy efficient detection in sensor networks," IEEE Journal on Selected Areas in Communications,
[19] R. S. Blum and B. M. Sadler, "Energy efficient signal detection in sensor networks using ordered transmissions," IEEE Transactions on Signal Processing,
[20] L. Tong, Q. Zhao, and S. Adireddy, "Sensor networks with mobile agents," in Proceedings of MILCOM 2003, vol. 1, Boston MA, Oct. 2003, pp.
vol. 23, pp. 639 -- 702, Apr. 2005.
vol. 56, no. 7, pp. 3229 -- 3235, Jul. 2008.
688 -- 693.
[21] Z. Yang and L. Tong, "On the error exponent and the use of LDPC codes for cooperative sensor networks with misinformed nodes," IEEE Transactions
[22] S. Marano, V. Matta, P. Willett, and L. Tong, "DOA estimation via a network of dumb sensors under the SENMA paradigm," IEEE Signal Processing
on Information Theory, vol. 53, no. 9, pp. 3265 -- 3274, Sep. 2007.
Letters, vol. 12, no. 10, pp. 709 -- 712, Oct. 2005.
[23] S. Marano, V. Matta, and L. Tong, "Secrecy in cooperative SENMA with unauthorized intrusions," in Proceedings of IEEE Workshop on Signal Processing
Advances in Wireless Communications (SPAWC), Cannes, France, July 2-5 2006.
[24] A. H. Sayed, "Adaptation, learning, and optimization over networks," in Foundations and Trends in Machine Learning. Boston-Delft: NOW Publishers,
2014, vol. 7, no. 4 -- 5, pp. 311 -- 801.
[25] -- -- , "Adaptive networks," Proc. IEEE, vol. 102, no. 4, pp. 460 -- 497, Apr. 2014.
[26] S. Kar and J. M. F. Moura, "Convergence rate analysis of distributed gossip (linear parameter) estimation: Fundamental limits and tradeoffs," IEEE
Journal of Selected Topics in Signal Processing, vol. 5, no. 4, pp. 674 -- 690, Aug. 2011.
[27] P. Braca, S. Marano, and V. Matta, "Enforcing consensus while monitoring the environment in wireless sensor networks," IEEE Transactions on Signal
[28] P. Braca, S. Marano, V. Matta, and P. Willett, "Asymptotic optimality of running consensus in testing binary hypotheses," IEEE Transactions on Signal
Processing, vol. 56, no. 7, pp. 3375 -- 3380, 2008.
Processing, vol. 58, no. 2, pp. 814 -- 825, 2010.
[29] D. Bajovic, D. Jakoveti´c, J. Xavier, B. Sinopoli, and J. M. F. Moura, "Distributed detection via Gaussian running consensus: Large deviations asymptotic
analysis," IEEE Transactions on Signal Processing, vol. 59, no. 9, pp. 4381 -- 4396, Sep. 2011.
[30] D. Jakoveti´c, J. M. F. Moura, and J. Xavier, "Distributed detection over noisy networks: Large deviations analysis," IEEE Transactions on Signal
Processing, vol. 60, no. 8, pp. 4306 -- 4320, Aug. 2012.
[31] D. Bajovic, D. Jakoveti´c, J. M. F. Moura, J. Xavier, and B. Sinopoli, "Large deviations performance of consensus+innovations distributed detection with
non-Gaussian observations," IEEE Transactions on Signal Processing, vol. 60, no. 11, pp. 5987 -- 6002, Nov. 2012.
[32] S. S. Boyd, A. Ghosh, and B. S. D. Prabhakar, "Gossip algorithms: Design, analysis and applications," in Proc. of INFOCOM, Miami, USA, March,
13-17 2005, pp. 1653 -- 1664.
1035 -- 1048, Mar. 2010.
no. 10, pp. 5107 -- 5124, 2012.
[33] F. S. Cattivelli and A. H. Sayed, "Diffusion LMS strategies for distributed estimation," IEEE Transactions on Signal Processing, vol. 58, no. 3, pp.
[34] X. Zhao and A. H. Sayed, "Performance limits for distributed estimation over LMS adaptive networks," IEEE Transactions on Signal Processing, vol. 60,
[35] S.-Y. Tu and A. H. Sayed, "Diffusion strategies outperform consensus strategies for distributed estimation over adaptive networks," IEEE Transactions
on Signal Processing, vol. 60, no. 12, pp. 6217 -- 6234, Dec. 2012.
[36] F. S. Cattivelli and A. H. Sayed, "Distributed detection over adaptive networks using diffusion adaptation," IEEE Transactions on Signal Processing,
vol. 59, no. 5, pp. 1917 -- 1932, 2011.
27
[37] A. H. Sayed, S.-Y. Tu, J. Chen, X. Zhao, and Z. J. Towfic, "Diffusion strategies for adaptation and learning over networks: an examination of distributed
strategies and network behavior," IEEE Signal Processing Magazine, vol. 30, no. 3, pp. 155 -- 171, 2013.
[38] A. H. Sayed, "Diffusion adaptation over networks," in Academic Press Library in Signal Processing, R. Chellapa and S. Theodoridis., Eds. North-Holland:
Academic Press, Elsevier, 2014, vol. 3, pp. 323 -- 454. Also available as arXiv:1205.4220[cs.MA]May2012.
[39] J. Chen and A. H. Sayed, "On the learning behavior of adaptive networks -- Part I: Transient analysis," IEEE Transactions on Information Theory, vol. 61,
[40] -- -- , "On the learning behavior of adaptive networks -- Part II: Performance analysis," IEEE Transactions on Information Theory, vol. 61, no. 6, pp.
no. 6, pp. 3487 -- 3517, Jun. 2015.
3518 -- 3548, Jun. 2015.
[41] V. Matta, P. Braca, S. Marano, and A. H. Sayed, "Diffusion-based adaptive distributed detection: Steady-state performance in the slow adaptation regime,"
IEEE Transactions on Information Theory, vol. 62, no. 8, pp. 4710 -- 4732, Aug. 2016.
[42] -- -- , "Distributed detection over adaptive networks: Refined asymptotics and the role of connectivity," IEEE Trans. Signal and Inf. Process. over
[43] Z. Towfic, J. Chen, and A. H. Sayed, "Excess-risk of distributed stochastic learners," IEEE Transactions on Information Theory, vol. 62, no. 10, pp.
Networks, vol. 2, no. 4, pp. 442 -- 460, Dec. 2016.
5753 -- 5785, Oct. 2016.
[44] J. N. Tsitsiklis, "Decentralized detection by a large number of sensors," Math. Contr., Signals, Syst., vol. 1, pp. 167 -- 182, 1988.
[45] P. K. Varshney, Distributed Detection and Data Fusion. New York, NY: Springer, 1997.
[46] J.-J. Xiao and Z.-Q. Luo, "Universal decentralized detection in a bandwidth-constrained sensor network," IEEE Transactions on Signal Processing,
[47] S. Marano, V. Matta, and F. Mazzarella, "Refining decisions after losing data: The unlucky broker problem," IEEE Transactions on Signal Processing,
[48] S. Kar, H. Chen, and P. K. Varshney, "Optimal identical binary quantizer design for distributed estimation," IEEE Transactions on Signal Processing,
vol. 53, no. 8, pp. 2617 -- 2624, Aug. 2005.
vol. 58, no. 4, pp. 1980 -- 1990, Apr. 2010.
vol. 60, no. 7, pp. 3896 -- 3901, Jul. 2012.
[49] S. Zhu and B. Chen, "Distributed detection over connected networks via one-bit quantizer," in Proc. of the IEEE International Symposium on Information
Theory (ISIT), Barcelona, Spain, 10 - 15 Jul 2016, pp. 1526 -- 1530.
[50] -- -- , "Quantized consensus by the ADMM: probabilistic versus deterministic quantizers," IEEE Transactions on Signal Processing, vol. 64, no. 7, pp.
[51] -- -- , "Corrections to "Quantized consensus by the ADMM: probabilistic versus deterministic quantizers"," IEEE Transactions on Signal Processing,
[52] M. E. Chamie, J. Liu, and T. Basar, "Design and analysis of distributed averaging with quantized communication," IEEE Transactions on Automatic
[53] S. Zhu, Y. C. Soh, and L. Xie, "Distributed parameter estimation with quantized communication via running average," IEEE Transactions on Signal
[54] T. Li, M. Fu, L. Xie, and J. F. Zhang, "Distributed consensus with limited communication data rate," IEEE Transactions on Automatic Control, vol. 56,
1700 -- 1713, Apr. 2016.
vol. 65, no. 6, pp. 1638 -- 1639, Jan. 2017.
Control, vol. 61, no. 12, pp. 3870 -- 3884, Dec. 2016.
Processing, vol. 63, no. 17, pp. 4634 -- 4646, Sep. 2015.
no. 2, pp. 272 -- 292, Feb. 2011.
Processing, vol. 58, no. 3, pp. 1383 -- 1400, Mar. 2010.
Jan. 2010.
vol. 56, no. 10, pp. 4905 -- 4918, Oct. 2008.
[55] S. Kar and J. M. Moura, "Distributed consensus algorithms in sensor networks: quantized data and random link failures," IEEE Transactions on Signal
[56] R. Carli, F. Fagnani, P. Frasca, and S. Zampieri, "Gossip consensus algorithms via quantized communication," Automatica, vol. 46, no. 1, pp. 70 -- 80,
[57] T. C. Aysal, M. J. Coates, and M. G. Rabbat, "Distributed average consensus with dithered quantization," IEEE Transactions on Signal Processing,
[58] P. Billingsley, Probability and Measure, 3rd ed. New York: Wiley-Interscience, 1995.
[59] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed. New Jersey, USA: Wiley-Interscience, 2006.
[60] S. M. Kay, Fundamentals of Statistical Signal Processing, Volume II: Detection Theory. Englewood Cliffs, New Jersey: Prentice Hall, 1998.
[61] A. Papoulis, Probability, Random Variables, and Stochastic Processes, 3rd ed. New York: McGraw-Hill, 1991.
[62] V. A. Zorich, Mathematical Analysis I. Berlin: Springer, 2004.
[63] R. B. Davies, "Numerical inversion of a characteristic function," Biometrika, vol. 60, no. 2, pp. 415 -- 417, 1973.
[64] P. Erdos, "On a family of symmetric Bernoulli convolutions," Amer. J. Math, vol. 61, pp. 974 -- 975, 1939.
[65] -- -- , "On the smoothness properties of Bernoulli convolutions," Amer. J. Math, vol. 62, pp. 180 -- 186, 1940.
[66] Y. Peres and B. Solomyak, "Absolute continuity of Bernoulli convolutions, a simple proof," Mathematical Research Letters, vol. 3, pp. 231 -- 239, 1996.
[67] -- -- , "Self-similar measures and intersections of Cantor sets," Transactions of the American Mathematical Society, vol. 350, no. 10, pp. 4065 -- 4087,
[68] M. Schroered, Fractal, Chaos, Power Laws. Minutes from an infinite paradise. New York: W. H. Freeman and Company, 1991.
[69] W. Feller, An Introduction to Probability and Its Applications, Volume 2. New York: John Wiley & Sons, 1971.
[70] H. L. Van Trees, Detection, Estimation, and Modulation Theory. Part I. New York: John Wiley & Sons, Inc., 1968 (reprinted, 2001).
[71] S. Marano, V. Matta, and L. Tong, "Distributed detection in the presence of Byzantine attacks," IEEE Transactions on Signal Processing, vol. 57, no. 1,
Oct. 1998.
pp. 16 -- 29, Jan. 2009.
[72] H. Shao, Mathematical Statistics, 2nd ed. New York, NY, USA: Springer, 2003.
[73] K. Knopp, Theory and Application of Infinite Series. London and Glasgow: Blakie & Son Limited, 1954.
[74] W. Rudin, Real and Complex Analysis, 3rd ed. McGraw-Hill International Editions, 1987.
[75] N. G. Shephard, "Inversion formulae for the distribution of ratios," J. Statist. Comput. Simul., vol. 39, pp. 37 -- 46, 1991.
[76] J. Gurland, "Inversion formulae for the distribution of ratios," The Annals of Statistics, vol. 19, pp. 228 -- 237, 1948.
|
cs/0603127 | 1 | 0603 | 2006-03-30T22:58:12 | Complex Systems + Systems Engineering = Complex Systems Engineeri | [
"cs.MA"
] | One may define a complex system as a system in which phenomena emerge as a consequence of multiscale interaction among the system's components and their environments. The field of Complex Systems is the study of such systems--usually naturally occurring, either bio-logical or social. Systems Engineering may be understood to include the conceptualising and building of systems that consist of a large number of concurrently operating and interacting components--usually including both human and non-human elements. It has become increasingly apparent that the kinds of systems that systems engineers build have many of the same multiscale characteristics as those of naturally occurring complex systems. In other words, systems engineering is the engineering of complex systems. This paper and the associated panel will explore some of the connections between the fields of complex systems and systems engineering. | cs.MA | cs | Position paper for 1.1 Panel on Complex Systems Engineering
Complex Systems + Systems Engineering
= Complex Systems Engineering
Russ Abbott
California State University, Los Angels
and The Aerospace Corporation
Los Angeles, Ca
[email protected]
Abstract
One may define a complex system as a system in which phenomena emerge as a conse-
quence of multiscale interaction among the system’s components and their environments. The
field of Complex Systems is the study of such systems—usually naturally occurring, either bio-
logical or social. Systems Engineering may be understood to include the conceptualising and
building of systems that consist of a large number of concurrently operating and interacting com-
ponents—usually including both human and non-human elements. It has become increasingly
apparent that the kinds of systems that systems engineers build have many of the same multiscale
characteristics as those of naturally occurring complex systems. In other words, systems engi-
neering is the engineering of complex systems. This paper and the associated panel will explore
some of the connections between the fields of complex systems and systems engineering.
What do complex systems and systems engineering have in common?
A complex system—in the sense of the word as used to identify the field of study by that
name—typically consist of elements that operate asynchronously. These elements operate ac-
cording to their own schedules and are not all entrained by a single overall timing driver, i.e., a
driver whose rate is transmitted throughout the system by mechanisms such as gears, drive
shafts, belts, electronic timing signals, etc. Complex systems tend to exhibit what are often re-
ferred to as multiscale or emergent phenomena: phenomena at an aggregate level that (a) result
from and (b) are a consequence of but (c) may be described independently of how the interac-
tions among the individual elements bring them about. See (Abbott 2005).
Large-scale engineered systems typically consist of many concurrently operating and inter-
acting elements—human and non-human—which as an ensemble either produce a result or pro-
vide a service or set of services. Such systems exhibit similar emergent and multiscalar proper-
ties. It has become increasingly clear that systems engineering is the engineering of complex sys-
tems. Although this realization is by now fairly widely shared, at least implicitly, and although
complex systems ideas, tools, and techniques have been applied to systems engineering problems
for some time, there has been little effort to date to bring the two fields—and the communities of
people who work in these fields—together in a more formal and explicit way. The current CSER
session is the first in a series of events whose goal is to introduce these communities to each
other and to embark on what we expect to be an extended dialog.
© 2006 The Aerospace Corporation
Following this CSER session, sessions will be held at the IEEE System of Systems Engineer-
ing Conference (April 2006) and at the INCOSE Symposium (July 2006). A stand-alone work-
shop on Complex Systems Engineering is being planned for January 2007 to be followed by a
full conference at which revised papers from that workshop (along with other submitted papers)
will be presented to a larger audience. The follow-on conference will be sited and scheduled to
coordinate with the INCOSE 2007 Symposium.
What issues are of interest to both complex systems and systems engineering?
In forging a connection between systems engineering and complex systems, the questions to
be addressed may be summarized as follows.
• What does systems engineering need from complex systems? What difficulties have large (and complex) engi-
neered systems encountered that seem to derive from the fact that they are complex in the sense described
above?
• What can complex systems offer to systems engineering? What conceptual frameworks from the field of com-
plex systems (and the tools that result from those frameworks) can be applied to systems engineering prob-
lems?
• What can systems engineering offer to complex systems? Systems engineers have successfully built some
very large and very complex systems. What intuitions have systems engineers developed about how to go
about building such systems that would be of interest to the complex systems community? To what extent can
systems engineering serve as a laboratory for the study of complex systems in ways that are not possible
when one is studying pre-existing and naturally occurring systems?
The remainder of this paper contains an initial (alphabetical) catalogue (preliminary and ten-
tative in many cases) of areas that we expect will be worth exploring.
Adaptability
See flexibility and evolvability.
Agents
Because the elements of complex systems operate asynchronously, these elements are fre-
quently referred to as agents. Generically, an agent is an asynchronously operating element
whose actions—either as a part of a system or as an actor that interacts with the system—must be
taken into account when the overall operation of a complex system is being analysed.1
The use of an agent-based framework for analyzing or modelling a system does not in itself
imply a particular level of agent autonomy or intelligence. Agents may be minimally autono-
mous and may simply follow orders—although being agents whose schedule is not entrained by
whatever is generating the orders, they follow their orders according to their own scheduling
mechanism. Alternatively, so-called intelligent agents may be understood as having significant
autonomy and intelligence and as capable of being given a goal and then being set loose to
achieve that goal using some combination of allocated and found resources.
In both complex systems and systems engineering agent-based modelling allows one to ex-
plore how behaviors defined at the level of system elements (the agents) give rise to phenomena
at the system level. (See emergence.)
An agent-based perspective is important in a second way for systems engineering: it defines
what is potentially a common conceptual framework for system architectures. Just as the lan-
guage of object-oriented programming has become the de facto vocabulary for large scale soft-
1 See,
2006
31,
January
of
as
available
site,
Tesfatsion’s web
example,
for
http://www.econ.iastate.edu/tesfatsi/ace.htm, for a fairly comprehensive list of agent-based modeling resources.
at
ware systems, the language of agent-based architectures are likely to become the de facto vo-
cabulary for engineered complex systems.
See also distributed control, market mechanisms, and stigmergy.
Architectures
For the most part, members of the complex systems community do not think in terms of sys-
tem architectures in the same way that systems engineers do.2 This is not to say that architectures
are not important in understanding naturally occurring complex systems. Scientists who study
biological, ecological, and social systems work hard to understand their underpinnings and inter-
nal organizations. But systems engineering-style architectures sense seem not to be central.
Are there systems engineering architectural concepts that can be helpful in explicating the
operation of naturally occurring complex systems? If so, this would be an area in which intui-
tions developed by systems engineers can enrich the field of complex systems.
On the other hand, if systems engineering-style architectures turn out not to be useful for un-
derstanding naturally occurring complex systems, one must wonder why these frameworks have
value for systems engineering and not for complex systems. One possible reason might be that
systems engineers are responsible for building systems whereas complex systems scientists study
systems that already exist. Does development responsibility require a level of architectural sup-
port that isn’t needed when simply studying a system?
Bricolage
Bricolage is the use of whatever is at hand to satisfy an immediate need—as when a rock is
used as a hammer. Reflecting a bricolage-style design, many naturally occurring systems include
components that support a wide range of uses. Hands can be used to type, to assess temperature
and roughness, to click a mouse, to grasp, to sew, to strum, to point, to caress, to scratch, and to
strike. Faces can express emotions; they can also ingest air, food, and light waves. In naturally
occurring systems, once a second use for a system component has established its value, evolution
tends to enhance its usability for that purpose. To what extent is bricolage acceptable/desirable as
a systems engineering design approach?
See also system of systems.
CONOPS (Concept of Operations)
A system’s CONOPS is the plan for how it will be used once deployed. What is lacking from
many CONOPS considerations is a realization that the user is outside the system and is embed-
ded in a larger system, which may put pressure on him/her to use the system in unexpected ways.
See systems of systems.
Distributed control
Because complex systems consists of agents, the functionality produced by a complex system
is a by-product of the interaction among the agents in the system. Desirable as it may be to be
able to point to a single place in a system as the ultimate determiner of how the system will be-
have, the actual behavior of a system depends on all its components, which, being agents, are not
2 See, for example, the various architecture frameworks (such as DODAF and Zachman) described on the Systems
and Software Consortium’s Architecture web
page,
available
as
of
January
31,
2006:
http://www.software.org/pub/architecture/fwhome.asp. Consider also the various specification languages and
meta-languages (such as UML) produced by the Object Management Group. See their web page, available as of
January 31, 2006: http://www.uml.org/.
under the entrained control of a single system authority. Systems with distributed control are
common in both naturally occurring and engineered complex systems.
It is possible, of course, to design a system with agents that have absolutely minimal auton-
omy—or to design a system in which there are no agents at all and in which the entire system is
understood as a single immensely complicated device. Systems are not designed in this manner.
1. In a large system there are so many micro-level decisions to be made that to make them
all at a central decision making authority would turn that central authority into an enor-
mous bottleneck.
2. Were there a central decision making authority, it simply would not have the expertise
and detailed information to make all the required decisions intelligently.
As experience with command-driven economies illustrate, such rigidly hierarchical and mi-
cro-managed systems are both inefficient and ineffective.
A system that is rigidly micro-managed has an additional significant weakness: its decision
making process is vulnerable to manipulation. Two sorts of vulnerabilities occur, depending on
the style with which decisions are made.
1. If decisions tend to be based on clearly specified rules and regulations, entities that need
to have decisions made learn how to manipulate requests so that they will get the results
they want. This is frequently called gaming the system: submit a request formulated in a
particular way because one knows how the rules will require the system to respond.
2. If decisions are made by human beings with broad discretionary powers, manipulate the
person—rather than the rules—either through a bribe or through some less overt form of
pressure, entreaty, or temptation. This reflects an alternative gambit: if one doesn’t suc-
ceed in accomplishing one’s end within a particular system, retreat to the larger system
within which the subject system is embedded. This is discussed further in (Abbott 2006).
The fact that a centralized decision making authority is vulnerable to manipulation confirms
the inevitable agent-based nature of any truly complex system. As we have seen, even in a sys-
tem in which one attempts to centralize decision making, many of the real decisions are in fact
made elsewhere. The primary difference is that an additional burden of manipulative overhead is
imposed on the agents—and hence on the system as a whole.
See also agents, and systems of systems.
Energy
Both naturally occurring and engineered complex systems must accommodate themselves to
the fact that neither energy nor mass is emergent. Although they can be interchanged, neither can
be created from interactions among more primitive elements. In most naturally occurring and
engineered complex systems mass tends to be fixed.3 Thus energy is critical.
Both naturally occurring and engineered complex systems persist and function only in so far
as they are powered by external sources of energy. (Hence the frequent reference to “far from
equilibrium” systems.) Naturally occurring complex systems are typically powered by sunlight
or its stored equivalents. Social complex systems are typically powered by energy proxies such
as money. Engineered complex systems also require sources of energy. We hypothesize that
analyses of energy flows will reveal fundamental similarities among naturally occurring and en-
gineered complex systems. “Follow the money” really means “follow the energy.”
3 Any release of energy involves the conversion of some mass to energy. But other than in the release of nuclear
energy, the amount of mass involved is negligible.
Emergence
See requirements.
Engineering
See science.
Environments
See systems of systems
Evolutionary computation
Evolutionary computational techniques (genetic algorithms and genetic programming) have
been applied to engineering problems since they were first developed. Recent GECCO confer-
ences feature what are called human competitive (Koza 2003) designs produced through evolu-
tionary means.4 A somewhat more recent development involves work in a relatively new sub-
field of complex systems known as developmental systems (Hornby 2007).
We are awaiting the time when computational power and sophistication will be sufficient to
enable agents in agent-based models (or even in agent-oriented system implementations) to
model the worlds in which they are functioning and to perform their own on-the-fly genetic pro-
gramming computations in order to determine their most effective courses of action.
Evolvability
Evolution is the first principle of biology. It is desirable for complex systems to have the
structural capacity to evolve. Yet a fundamental difference between systems that are engineered
and systems that result from natural processes is that engineered systems are often optimized for
one or more particular properties with little thought given to evolvability. When the environment
in which an engineered system changes or when the uses to which one wants to put an engi-
neered system changes, it is often quite difficult to change the system to accommodate those new
needs. Natural systems tend to be more adaptable. Can we understand how this difference comes
about and what, if anything we can do about it?
It seems (a) intuitive, (b) well established theoretically (Adami 2000) (Heylighen 1996), and
(c) obvious just by looking around that evolution leads to an increase in complexity. Yet the in-
creasingly complex systems that evolve are not re-designed from scratch with every evolutionary
step. Each step is an evolutionary neighbor of something a bit less complex. The only way such
an evolutionary process can occur is if the evolutionary sequence provides both (a) the specific
features needed at each step and (b) a framework (or architecture) that can support the evolution-
ary process itself. In other words, the only architectures that survive an evolutionary process are
those that support and facilitate evolutionary change. Thus the evolutionary process produces not
only fitness for changing environments but evolvability itself.
Successive versions of most engineered systems are not forced to run this sort of evolution-
ary gauntlet. New versions of many engineered systems are often built from scratch; they are not
evolved as extensions of previous versions. And those that are extensions of previous versions
are often quite clumsy. Of course many evolved systems are also quite messy. But in the long
run, evolution prunes away structures that impede evolution leaving evolvable core frameworks.
4 The “Human-Competitive Results” page of the genetic-programming.com web site (available as of January 31,
2006: http://www.genetic-programming.com/humancompetitive.html) lists 36 human-competitive results as of
December 31, 2003. See also http://www.genetic-programming.org/hc2005/cfe2005.html (available as of January
31, 2006) for the 2005 GECCO “Humie” awards.
We fail to build evolvable systems because we often don’t try. Evolvability is not required of
most engineered systems. One reason for this is that designing for evolvability is easier said than
done. A second is that we don’t have a satisfactory way either of specifying or measuring
evolvability. Given the current state of our knowledge about evolvability, the most effective way
of requiring evolvability is economic: require developers to absorb some of the cost of post-
delivery enhancements. It’s difficult to see, however, how one would structure a contract to that
effect that was attractive to developers and fair to both developers and customers.
On the other hand, we as system builders do learn from previously built systems. Although
we don’t build systems that themselves are evolvable, we do evolve our own knowledge of how
to build better and better systems. The analogy is that an individual’s DNA is comparable to the
initial conceptualization of a system; a fully grown and functioning individual is comparable to a
functioning system; and the developmental process, in biology known as ontogeny, which we
now know is controlled by gene switches (Carroll, 2005), is comparable to our retained and
evolved knowledge of how to build systems. Just as we now know that as much of biological
evolution is in the gene switches (and not in the genes), much of the evolution (retained knowl-
edge) of systems engineering is in the theory, tools, and practices of development.
Failures
See power laws.
Far from equilibrium
See energy.
Genetic algorithms and genetic programming
See evolutionary computation.
Market mechanisms
Market-based mechanisms allow agents to interact within a defined framework to exchange
goods, services, and resource proxies (such as money). An important advantage of a market
mechanism is that the decision making process is completely transparent and (mainly) immune
to manipulation. Market mechanisms frequently occur in both social and engineered complex
systems and are important to both complex systems and systems engineering.
A market is the visible part of an iceberg whose hidden part consists the rest of the economic
system within which the market functions. A market optimizes the use of resources by setting
resource rates of exchange that reflects the relative values of the resources in the larger context.
Markets work only because the resources traded have value to the market participants outside a
market. Otherwise, it’s not clear what benefit the participants would gain from participating in
the market. If a market were an isolated (and sterile) mechanism for performing exchanges, each
participant would be exactly as well off after a trade at the market price as he or she was before.
Markets work only when each participant is better off after a transaction than before. Better
off means that the resources they have after the transaction are worth more to them than the re-
sources they had before the transaction. Why would a resource be worth more to one market par-
ticipant than to another? The answer is that it is the differences in the situations of the market
participants outside the market that makes markets work. A simple example is the sale of an
electronic component. The seller manufactures the component for less than he receives in the
market. That’s why he participates. The buyer uses the component in a product that he sells for
more than the sum of the costs of the components and the cost of assembling them. That’s why
he participates. It is only because each participant is situated differently in the larger system that
they value the component differently and are each better off after trading in the market.
When a collection of agents, each of which is individually and differently situated in a larger
context, can each place a value on various resources from the perspective of their own situations,
then a market provides a mechanism to reconcile those various values and to optimize the use of
those resources. Markets don’t set a price. They allow exchange rates to adjust as perceived ex-
ternal values change. Other extra-market properties (e.g., inventories, financial resources, market
impact) may also affect buy/sell decisions and thereby the prices at which resources are traded.
See also agents.
Modeling
See agents.
Non-algorithmic programming
See requirements.
Non-linearity
Non-linearity does not just mean more sophisticated mathematics.
See phase transitions
Phase transitions
The operation of many systems varies relatively smoothly over a wide range of the system’s
parameters. A phase transition is a point at which a small change in a system parameter causes a
very large change in the performance, nature, or operation of the system. Many engineered and
naturally occurring complex systems exhibit phase transitions—often unanticipated.
Power laws
Failures in the electrical power grid follow a power law when intensity is plotted against fre-
quency. (Talukdar 2003) Similar power laws are observed in a wide range of phenomena—
intensity vs. frequency of earthquakes (Christensen 2002) being probably the best known.
It is hypothesized that failures in engineered systems—when the cost or significance of the
failure is plotted against the frequency—will also be found to follow a power law. It is also hy-
pothesized that this is an optimisation effect for which (Doyne 2005) coined the terms robust yet
fragile and highly optimized tolerance. As we understand this phenomenon better, the insights
we gain may be applied when assessing risk, fault tolerance, and other failure-related issues.
Requirements
(Abbott 2005) defines a property or phenomenon as emergent if it can be specified independ-
ently of its implementation. Specifying system properties and phenomena in this way is funda-
mental to systems engineering. Emergence seems somewhat more mysterious in naturally occur-
ring complex systems because it occurs without apparent intention.
We find it mundane to engineer a satellite in geosynchronous orbit because we know how to
do it. Yet the specification of geosynchronicity—that a satellite remain motionless with respect
to earth as a reference frame—is specified independently of its implementation. One might na-
ively imagine achieving such an effect by tying a long cable to the satellite, as if it were a bal-
loon. Of course that’s not how it works. Geosynchronicity is achieved by putting the satellite in
an orbit whose period is the same as the period of the earth’s rotation. There is nothing about this
(or any particular orbit) that holds a satellite motionless with respect to earth as a reference
frame. Geosynchronicity emergences as a consequence of the fact that the two periods are the
same. One might refer to this as planned emergence or non-algorithmic programming.
When ants forage or birds flock as a result of relatively simple rules we may (initially) find
this amazing—because the emergent phenomena seem to appear as if by magic from a system of
components which are following rules that on their face seem to have nothing to do with the end
result. But once we understand (Spector 2003) how these phenomena result from the underlying
rules, this sort of emergence is no more (and perhaps no less) mysterious than geosynchronicity.
Two other everyday examples of naturally occurring emergence are the 24-hour day-night
cycle and the yearly seasonal cycle. The first is an emergent consequence of the earth's rotation
with respect to the sun as a source of light. The second is an emergent consequence of the tilt of
the earth's axis and its revolution about the sun as a source of heat. In systems engineering one
often starts with requirements, which are emergent properties of a hypothetical system. One then
attempts to reverse engineer such a hypothetical system to determine how such emergence might
actually be accomplished. If one didn't know about the geometric relationships between the earth
and the sun, how might one engineer a system that exhibited day-night and seasonal cycles?
Reuse
See architectures and evolvability.
Risk
See failures.
Science
Engineering is the application of scientific and mathematical principles to practical ends. Yet
there is no science to which systems engineering looks for its principles. Its body of knowledge
consists mainly of best practices. Complex systems is the science that studies, among other
things, the sorts of systems that systems engineers build. To the extent that complex systems es-
tablishes itself as a science, it can serve as one of the sources of scientific principles that systems
engineering, as an engineering discipline, can apply.
Service Oriented Architecture (SOA)
An architecture that organizes a system as a framework within which capabilities are made
available as services. This is much the same model as the economy in which products and ser-
vices are made available to those who might (buy and) use them. See system of systems.
Stigmergy
(Bonabeau 1999) borrowed the term stigmergy from entomology to refer to the use of the en-
vironment for communication among agents in an agent-based systems. Probably the most famil-
iar example is ant foraging in which ants leave pheromone trails (in the environment) for other
ants to follow. Most engineered systems tend not to use this kind of mechanism and to favor in-
stead communication mechanisms over which the system has more direct control.
See systems of systems.
Systems of systems
A system that never changes—even if its components are themselves systems—is just an-
other system. In (Abbott, 2006) we characterize the nature of systems of systems in terms of the
following properties. A system of systems is (a) open at the top (to new applications), (b) open at
the bottom (to new implementation of primitives), and (c) continually but slowly changing.
An additional and essential aspect of a system of systems perspective is the recognition that
every system, other than nature itself, is embedded in an environment—which is a larger system.
A fundamental insight—and one that ties work in systems of systems to work in complex
systems—is that whenever something exists in an environment, other elements of that environ-
ment may use it in ways that were not originally anticipated.
A fundamental example is the predator-prey relationship. It may sound trivial to say so, but
predators are able to exploit their prey only because the prey exist in the predator’s environment.
But the prey typically5 don’t place themselves in the environment for the benefit of the predators.
Exploitation of an existing element in an environment isn’t always a one-way street. Plants
and bees are the standard example of mutual (and mutually beneficial) exploitation.6, 7 Each
makes use of the existence of the other for its own purposes. The relationship between them has
become so intimate that each now depends for its survival on the use the other makes of it.
The fundamental principle is that once a system (or any mechanism) exists in the world, it
may be used for purposes that are different from those intended or anticipated when the system
or mechanism was created and deployed. This is a source of both great creativity and unintended
consequences. Anti-bacterial cleaners have led to the evolution of bacteria that are resistant to
the anti-bacterial agents in the cleaners. Tax laws have led to an accounting specialty that finds
ways to avoid the effects of those laws. In both cases, the creation of a mechanism in the world
that is noxious to a population has led to the development of counter-mechanisms that allow
elements of the population to escape the most noxious aspects of those mechanisms.8
To build a successful system of systems, one must adopt an environmental—one might call it
an outward facing—perspective. Such a perspective leads to the following questions which must
be addressed about any system—and especially about systems of systems.
• On what environmental presumptions/capabilities does the system depend? In what ways does the system's
functioning depend on environment-based (i.e., stigmergic) interactions? Can the system make more use of
stigmergic mechanisms?
• How does the system change the environment within which it exists simply by existing? How do/might non-
system elements interact with/exploit/foil it once it exists? What symbiotic relationships exist among the sys-
tem and other elements of the environment? Are these desirable/inevitable?
• How does the system serve as an environment for its components?
•
If the system is to exist within an environment that is undergoing rapid evolution, how can one build it so that
it won't be obsolete by the time it's finished? How can it be designed so that it can evolve as the environments
within which it exists evolves around it?
Human civil society is a system of systems. We add new systems to it all the time, and it
continues to function—more or less. What is it about civil society that enables it to integrate
more and more systems into itself without apparent limit—at least so far?
See also agents, distributed control, market mechanisms, and stigmergy.
References
Abbott, R, “Emergence Explained,” Submitted for publication, October 2005. Available as of
January 31, 2006: http://cs.calstatela.edu/~wiki/images/9/90/Emergence_Explained.pdf.
5 An interesting example of prey actually inserting themselves into an environment for the benefit of predators is
the strategy adopted by a team from the University of Southampton in a Prisoner’s Dilemma tournament
(Grossman, 2006). Some Southampton entries sacrificed themselves for the benefit of other Southampton entries,
which won the tournament—beating Tit-for-Tat for the first time.
6 Here exploitation is intended to be understood as “to make use of” and not necessarily to disadvantage another.
7 An environment that includes an interacting “plants system” and “bees system” is a system of systems.
8 The “Museum of unintended consequences” is an informal web collection of such unintended consequences. It is
available
as
of
January
31,
2006:
http://cs.calstatela.edu/~wiki/index.php/Courses/CS_461/Museum_of_unintended_consequences
Abbott, R., "Open at the Top, Open at the Bottom, and Continually (but slowly) Changing,"
Submitted for publication, April, 2006.
Adami, C., C. Ofria, and T. C. Collier, “Evolution of biological complexity,” Proceedings of the
National Academy of Sciences, Vol. 97, Issue 9, 4463-4468, April 25, 2000. Available as of
January 31, 2006: http://www.pnas.org/cgi/content/full/97/9/4463.
Bonabeau, E., M. Dorigo, and G. Theraulaz, Swarm Intelligence: From Natural to Artificial Sys-
tems, Oxford University Press, 1999.
Carroll, S. B., Endless Forms Most Beautiful: The New Science of Evo Devo and the Making of
the Animal Kingdom, W. W. Norton & Company, April 11, 2005.
Christensen, K., L. Danon, T. Scanlon, and P. Bak, “Unified scaling law for earthquakes,” Pro-
ceedings of the National Academy of Sciences, February 19, 2002. Available as of January
31, 2006: http://www.pnas.org/cgi/reprint/99/suppl_1/2509.pdf.
Doyle, J. “The ‘robust yet fragile’ nature of the Internet,” Proceedings of the National Academy
of Sciences, vol. 102, no. 41, pp 14497-14502. October 11, 2005. Available as of January 31,
2006: http://www.pnas.org/cgi/reprint/102/41/14497.pdf.
Grossman, w., “New Tack Wins Prisoner's Dilemma,” wired.com, October, 13, 2004. Available
as of January 31, 2006: http://www.wired.com/news/culture/0,1284,65317,00.html.
Heylighen, F., “The Growth of Structural and Functional Complexity during Evolution,” The
Evolution of Complexity" (Kluwer Academic Publishers), 1996. Available as of January 31,
2006: http://pespmc1.vub.ac.be/Papers/ComplexityGrowth.html.
Hornby, G., et. al., Developmental Systems, planned special issue of Genetic Programming and
Evolvable Machines, to appear 2007. Announcement available as of January 31, 2006:
http://ic.arc.nasa.gov/people/hornby/sids06/gpem_sids06.htm.
Koza, J. et. al., Genetic Programming IV: Routine Human-Competitive Machine Intelligence,
Springer, 2003.
Spector, L., J. Klein, C. Perry, and M. Feinstein, “Emergence of Collective Behavior in Evolving
Populations of Flying Agents,” Proceedings of the Genetic and Evolutionary Computation
Conference (GECCO-2003), pp. 61–73. Berlin: Springer-Verlag. Available as of January 31,
2006: http://hampshire.edu/lspector/pubs/spector-gecco2003.pdf.
Talukdar, S.N., J. Apt, M. Ilic, L. B. Lave, and M. G. Morgan, “Cascading Failures: Survival vs.
Prevention,” The Electricity Journal, Dec. 2003. Available as of January 31, 2006:
http://www.pserc.wisc.edu/ecow/get/publicatio/2003public/talukdar_cascading_failures.pdf.
Biography
Russ Abbott is a Professor of Computer Science at California State University, Los Angeles
and a member of the staff at The Aerospace Corporation.
|
1803.05028 | 1 | 1803 | 2018-03-13T20:07:26 | Decentralised Learning in Systems with Many, Many Strategic Agents | [
"cs.MA"
] | Although multi-agent reinforcement learning can tackle systems of strategically interacting entities, it currently fails in scalability and lacks rigorous convergence guarantees. Crucially, learning in multi-agent systems can become intractable due to the explosion in the size of the state-action space as the number of agents increases. In this paper, we propose a method for computing closed-loop optimal policies in multi-agent systems that scales independently of the number of agents. This allows us to show, for the first time, successful convergence to optimal behaviour in systems with an unbounded number of interacting adaptive learners. Studying the asymptotic regime of N-player stochastic games, we devise a learning protocol that is guaranteed to converge to equilibrium policies even when the number of agents is extremely large. Our method is model-free and completely decentralised so that each agent need only observe its local state information and its realised rewards. We validate these theoretical results by showing convergence to Nash-equilibrium policies in applications from economics and control theory with thousands of strategically interacting agents. | cs.MA | cs |
Decentralised Learning in Systems with Many, Many
Strategic Agents
David Mguni1,
Joel Jennings1
Enrique Munoz de Cote1,2
1PROWLER.io, Cambridge, UK
2Department of Computer Science, INAOE, Mexico
March 15, 2018
Abstract
Although multi-agent reinforcement learning can tackle systems of strategically interacting entities, it
currently fails in scalability and lacks rigorous convergence guarantees. Crucially, learning in multi-agent
systems can become intractable due to the explosion in the size of the state-action space as the number of
agents increases. In this paper, we propose a method for computing closed-loop optimal policies in multi-
agent systems that scales independently of the number of agents. This allows us to show, for the first time,
successful convergence to optimal behaviour in systems with an unbounded number of interacting adaptive
learners. Studying the asymptotic regime of N−player stochastic games, we devise a learning protocol
that is guaranteed to converge to equilibrium policies even when the number of agents is extremely large.
Our method is model-free and completely decentralised so that each agent need only observe its local state
information and its realised rewards. We validate these theoretical results by showing convergence to Nash-
equilibrium policies in applications from economics and control theory with thousands of strategically
interacting agents.
Introduction
Multi-agent reinforcement learning (MARL) provides the potential to systematically analyse environments
with strategically interacting agents. Despite the fundamental relevance of multi-agent systems (MASs) with
appreciably large populations, learning stable, best-response policies in MASs with more than a few agents
remains a significant challenge due to growth in complexity as the number of agents increases (Shoham and
Leyton-Brown 2008). Consequently, the task of understanding agent behaviour in many systems of interest
has been left unaddressed.
This paper seeks to address the problem of learning stable, best-response policies within non-cooperative1
MASs when the size of the population is large, therefore expanding the range of applications of multi-agent
technology.
In a non-cooperative MAS, selfish agents compete to obtain a sequence of rewards within an unknown
environment. A stochastic (dynamic) game (SG) is a mathematical framework that analyses the behaviour of
strategically interacting entities in non-cooperative settings. Stochastic games enable stable policy outcomes
1In game theory, the term non-cooperative implies that each agent seeks to pursue its own objectives and agreements between agents
over their actions cannot be enforced.
1
in which agents respond optimally to one another (know as equilibria), to be fully described. In SGs, it is
assumed that agents have either fixed knowledge of their environment or can acquire knowledge of any miss-
ing data by simply observing other agents with which agents can compute best-response actions. However,
in many multi-agent systems, agents do not have full information of the environment from the outset and
direct computation of optimal behaviour is often prohibitively complex.
Naturally, integrating stochastic game theory with reinforcement learning (RL) - a framework that en-
ables agents to learn optimal behaviour within an unknown environment through direct interaction and ex-
ploration - suggests the potential to learn stable policies in multi-agent systems.
Although this approach has led to fruitful analysis in multi-agent systems with few interacting agents (Leibo
et al. 2017), current methods of computing multi-agent equilibria using RL (e.g. Nash Q-Learning (Hu and
Wellman 2004), Friend-or-Foe Q-learning (Littman 2001), minimax-Q (Littman 1994)) have computational
complexity that increases exponentially with the number of agents (Busoniu, Babuska, and De Schutter 2008;
Tuyls and Weiss 2012). This renders the task of using RL to learn equilibrium policies intractable for many
systems of interest.
In this paper, we introduce an approach that enables equilibrium policies of multi-agent systems to be
computed even when the size of the population is extremely large. Unlike current multi-agent learning pro-
cedures, our method scales independently of the number of interacting agents. In contrast to approaches that
compute equilibria in large population games (Cardaliaguet and Hadikhanloo 2017), our method is a model-
free, fully decentralised learning procedure that only requires agents to observe local state information and
their realised rewards.
Our main result demonstrates that the equilibria of N− player SGs can be computed by solving an op-
timal control problem (OCP) using a model-free learning procedure under very mild assumptions. To do
this, we prove a series of theoretical results: first, we establish a novel link between reinforcement learning
in MASs and a class of games known as discrete-time mean field games - N−player SGs in an asymptotic
regime as N → ∞. Secondly, we demonstrate that in the asymptotic regime, the resultant game belongs
to a class of games known as potential games. These are reducible to a single objective OCP leading to a
vast reduction in the problem complexity. Our last result proves that the equilibria of SGs in an asymptotic
regime are in fact approximate equilibria of the N− player SG with an approximation error that vanishes as
N increases. Finally, we validate our theoretical results by application to a series of problems within eco-
nomics and optimal control theory. Our approach is based on a variant of the fictitious play - a belief-based
learning rule for static games introduced by Brown (Brown 1951) and generalised to adaptive play in (Leslie
and Collins 2006) for games with finite action sets.
After formulating the problem as an SG, the paper is organised as follows: first, we provide a formal
description of a discrete-time mean field game and show that the game is a potential game (Theorem 1). We
show that given the potentiality property, the problem is reducible to a single objective OCP. We then intro-
duce the learning protocol and show that under this protocol, the game has strong convergence guarantees
to equilibrium policies (Theorem 2). We lastly show that the equilibria generated for the mean field game
are approximate equilibria of the N−player SG with an approximation error that vanishes as N increases
(Theorem 3). Taking benchmark examples from economics and the multi-agent literature, we demonstrate
our method within a number of examples with large populations of interacting agents.
Background
We now give the background for SG theory by introducing the N−player SG formalism. In order to handle
multi-agent systems with large populations, we consider the N−player SG formalism when the number
2
of agents tends to infinity - we therefore introduce the notion of mean field games - SGs studied in the
asymptotic regime in the number of agents.
Problem formulation: N−Player Stochastic Games
The canonical framework to describe multi-agent systems in which agents behave rationally and non-
cooperatively is a stochastic (dynamic) game (SG). Let us therefore introduce a formal description of an
SG:
Let N (cid:44) {1, . . . , N} denote the set of agents where N ∈ N. At each time step k ∈ 1, 2, . . . T ∈ N,2 the
state of agent i ∈ N is xi
k ∈ S i where S i ⊂ Rd is a d−dimensional state space. The state of the system at
time k ≤ T is given by xk (cid:44) (xi
k)i∈N where xk ∈ S (cid:44) ×j∈NS j, which is the Cartesian product of space
of states for each agent. Let Πi be a non-empty compact set of stochastic closed-loop policies3 for agent i
where πi : S → ∆Ai where Ai ⊆ Rd is a compact, non-empty action set for each agent i ∈ N . Denote
by Π the set of policies for all agents i.e. Π (cid:44) ×j∈N Πj. We denote by Π−i (cid:44) ×j∈N\{i}Πj, the Cartesian
product of the policy sets for all agents except agent i ∈ N .
At each time step, each agent i ∈ N exercises its policy πi, the agent's state then transitions according to
the following4:
xi
k+1 = f (xi
(1)
where {ζk}0≤k≤T is a collection of i.i.d. random variables that introduce randomness in the agent's state
transition.
Each agent i has a cumulative reward function J i : S × Πi × Π−i → R that it seeks to maximise given by
the following:
k ∼ πi, k = 0, 1, . . . , T
k, ai
k, ζk), ai
J i[xt, πi, π−i] = Ex∼f
L(xi
k, x−i
k , ai
k)
k ∼ πi
,
(2)
(cid:34) T(cid:88)
k=t
(cid:35)
(cid:12)(cid:12)(cid:12)ai
k ∈ S i and x−i
k ∈ S−i are the state for agent i and the collection of states for agents j ∈ N\{i} at
where xi
time k ≤ T respectively and πi
k is the policy for agent i. The function L is the instantaneous reward function
which measures the reward received by the agent at each time step. We refer to the system of equations (1) -
(2) as game (A).
We now formalise the notion of optimality within an SG, in particular, a Markov-Nash-equilibrium for
the game (A) is the solution concept when every agent plays their best response to the policies of other
agents. Formally, we define the notion of equilibrium for this game by the following:
Definition 1. The strategy profile π = (πi, π−i) ∈ Π is said to be a Markov-Nash equilibrium (M-NE)
strategy if, for any policy for agent i, π(cid:48)i ∈ Πi and ∀y ∈ S, we have:
J i[y, πi, π−i] ≥ J i[y, π(cid:48)i, π−i], ∀i ∈ N .
(3)
The M-NE condition identifies strategic configurations in which no agent can improve their rewards by
a unilateral deviation from their current strategy.
We will later consider approximate solutions to the game (A). In order to formalise the notion of an
2The formalism can be straightforwardly extended to infinite horizon cases by appropriate adjustment of the reward function.
3Closed-loop policies are maps from states to actions and are likely to be the only policies that produce optimal behaviour in
stochastic systems. Open-loop policies simply specify pre-computed (state-independent) sequences of actions.
4With this specification, agents do not influence each others' transition dynamics directly - this is a natural depiction of various
systems e.g. a portfolio manager's modification to their own market position. This does not limit generality since prohibited state
transitions (e.g. collisions) can be disallowed with a reward function that heavily penalises such joint action behaviour.
3
approximate solution, we introduce −Markov-Nash equilibria (−M-NE) which extends the concept of M-
NE to strategy profiles in which the incentive to deviate never exceeds some fixed constant. The notion of an
−M-NE can be described using an analogous condition to (3). Formally, the strategy profile (πi, π−i) ∈ Π
is an −Markov-Nash equilibrium strategy profile if for a given > 0 and for any individual strategy for
agent i, π(cid:48)i ∈ Πi we have that ∀y ∈ S:
J i[y, πi, π−i] ≥ J i[y, π(cid:48)i, π−i] − , ∀i ∈ N .
(4)
Although in principle, methods within RL such as TD learning can be used to compute the equilibrium
policies (Sutton and Barto 1998), learning takes place in the product space of the state space and the set of
actions across agents, so the problem complexity grows exponentially with the number of agents.
A second issue facing RL within MASs is the appearance of non-stationarity produced by other adaptive
agents. During the learning phase, agents update their policies and thus the way they influence the system. In
a non-cooperative MAS with even just a few agents learning independently, the presence of other adaptive
agents induces the appearance of a non-stationary environment from the perspective of an individual agent.
This in turn may severely impair the agent's own reinforcement learning process and lead to complex and
non-convergent dynamics (Tuyls and Weiss 2012).
With these concerns, we present an alternative approach which involves studying the game (A) in an
asymptotic regime as N → ∞. This results in a mean field game - an SG with an infinite population which,
as we shall show is both reducible to a single OCP and has M-NE that are −M-NE for N−player SGs
where ∼ O( 1√
).
N
In order to show the discrete-time mean field game is reducible to single OCP, we demonstrate that they
belong to a class of games known as dynamic potential games (PGs). Before considering the mean-field
game case, let us formally define a PG in the context of an N−player SG:
Definition 2. An SG is called a (dynamic) potential game (PG) if for each agent i ∈ N and for any given
strategy profile π ∈ Π there exists a potential function Ω : · × Πi × Π−i → R that satisfies the following
condition ∀π(cid:48)i ∈ Πi :
J i[·, (πi, π−i)] − J i[·, (π(cid:48)i, π−i)] = Ω[·, (πi, π−i)] − Ω[·, (π(cid:48)i, π−i)].
(5)
A PG has the property that any agent's change in reward produced by a unilateral deviation in their
strategy is exactly expressible through a single global function. In PGs - the (Nash) equilibria can be found
by solving an OCP (Monderer and Shapley 1996). This is a striking result since obtaining the solution to
an OCP is, in general, an easier task than standard methods to obtain equilibria which rely on fixed point
arguments.
Mean Field Games
In this section, we introduce a mean field game (MFG) which is a central framework to our approach. Mean
field game theory is a mathematical formalism that handles large-population systems of non-cooperative ra-
tional agents. MFGs are formulated as SGs in the form (A) analysed at the asymptotic limit as the number of
agents tends to infinity. This formulation enables the collective behaviour of agents to be jointly represented
by a probability distribution over the state space (Lasry and Lions 2007).
The MFG formulation results in a description of the agents' optimal behaviour that is compactly char-
acterised by a coupled system of partial differential equations. However, obtaining closed analytic solutions
(or even approximations by tractable numerical methods) for the system of equations is often unachievable
but for specific cases.
4
This work offers a solution to this problem; in particular we introduce a learning procedure by which the
equilibria of MFG can be learned by adaptive agents. Beginning with the case in which the number of agents
is finite, we introduce the following empirical measure which describes the N agents' joint state at time k:
N(cid:88)
i=1
mxk
(cid:44) 1
N
δxi
k
,
(6)
k ∈ S and δx is the Dirac-delta distribution evaluated at the point x ∈ S.
where xi
We now study a game with an infinite number of agents by considering the formalism the N−player
game (A) in the asymptotic regime as N → ∞ which allows us to treat the ensemble in (6) as being
continuously distributed over S. We call this limiting behaviour the mean field limit which is an application
of the law of large numbers for first order (strategic) interactions in the game (A). We observe that by taking
the limit as N → ∞ and using de Finetti's theorem5, we can replace the empirical measure (6) with a
probability distribution m ∈ P(H) where P(H) is a space of probability measures. The distribution m
describes the joint locations of all agents in terms of a distribution.
With this structure, instead of agents responding to the actions of other agents individually, each agent
now performs its actions in response to the mass which jointly represents the collection of states for all
agents.
As is standard within the MFG framework (Lasry and Lions 2007), we assume that the MFG satisfies
the indistinguishably property - that is the game is invariant under permutation of the agents' indices.
The following concept will allow us to restrict our attention to games with a single M-NE:
Definition 3. The function v : P(H) × · → R is said to be strictly monotone in the L2− norm given
m1, m2 ∈ P(H) if the following is satisfied:
(v(m1,·) − v(m2,·))(m1 − m2)dx ≥ 0 =⇒ m1 ≡ m2.
(cid:90)
S
The strict monotonicity condition means that in any given state, agents prefer a lower concentration
of neighbouring agents. This property is a natural feature within many practical applications in which the
presence of others reduces the available rewards for a given agent e.g. spectrum sharing (Ahmad et al. 2010).
We make use of the following result which is proved in (Lasry and Lions 2007)6:
Proposition 1 (Lasry & Lions, 2007). If the instantaneous reward function of a MFG is strictly monotone
in m ∈ P(H), then there exists a unique M-NE for the MFG.
MAS with Infinite Agents
To develop a learning procedure that scales with the number of agents, we now consider the game (A) in the
mean-field limit. We shall demonstrate that this procedure allows us to reduce the game (A) to a strategic
interaction between an agent and an entity that represents the collection of other agents. This plays a key
role in reducing the problem complexity and collapsing it to a single OCP.
We shall firstly define the N−player stochastic game (A) in the asymptotic regime. We note that in light
of the indistingushability criterion, we can drop the agent indices:
5Given a sequence of indexed random variables x1, x2, . . . which are invariant under permutations of the index, De Finetti's theorem
(de Finetti 1931) ensures the existence of the random variable mxk in (6) in the limit as N → ∞.
6In (Lasry and Lions 2007) the result is proven for mean field games with continuous action and state spaces in continuous time.
The corresponding results for discrete games (discrete state space, time and action set) have also been proven (see theorem 2, pg 6 in
(Gomes, Mohr, and Souza 2010)).
5
Definition 4 (Discrete-Time Mean Field Game). We call the following system a discrete-time mean field
game if its dynamics can be represented by the following system:
xk+1 = f (xk, ak, ζk)
mxk+1 = g(mxk , ak),
(7)
(8)
where ak ∼ π for some π ∈ Π and ak = (ak)i∈N , k ∈ 0, 1, . . . , T , for a given time horizon T < ∞,
mxk ∈ P(H) is the agent density corresponding to the asymptotic distribution (6) evaluated at xk ∈ S, πk is
a policy exercised at each time step k ≤ T and ζk is an i.i.d. variable which captures the system stochasticity.
We refer to the system (7) - (8) as game (B).
Given mxt and π ∈ Π with the agent index removed, we consider games where each agent has the following
reward function:
J[xt, π, mxt] = E(cid:104) T(cid:88)
k=t
(cid:105)
(cid:12)(cid:12)(cid:12)ak ∼ π
L(xk, mxk , ak)
,
(9)
where xt ∈ S is some initial state and ak is the action taken by the agent.
response to the actions of other agents.
We are now in a position to describe a MFG system at equilibrium i.e. when each agent plays a best-
Given some initial state xt ∈ S, the joint solution (π, m) to the game (B) is described by the following
triplet of equations which describes the M-NE:
π ∈ argmax
π∈Π
J[xt, π, mxt],
(10)
(11)
(12)
where as before, {ζk}0≤k≤T is a collection of i.i.d random variables and, mx is the agent density induced
when the policy π is exercised by each agent.
xk+1 = f (xk, ak, ζk), ak ∼ π ∈ Π
mxk+1 = g( mxk , ak), ak = (ak)i∈N ,
An important feature of the system (10) - (12) is that the agent's problem is reduced to a strategic interac-
tion between itself and a single entity mx. This property serves a crucial role in overcoming the appearance
of non-stationarity in an environment with many adaptive learners since the influence of all other agents on
the system is now fully captured a single entity mx which, influences the system dynamics in a way that an
adaptive agent can learn its optimal policy.
Existing methods of computing equilibria in MFGs however rely on the agents having full knowledge
of the environment to compute best responses and involve solving to non-linear partial differential equa-
tions (Cardaliaguet and Hadikhanloo 2017; Cardaliaguet et al. 2015) which, in a number of cases leads
to intractability of the framework. MFGs are closely related to anonymous games - games in which the
agents' rewards do not depend on the identity of the agents they interact with (but do depend on the inter-
acting agents' strategies). Multi-agent learning has been studied for anonymous games (Kash, Friedman,
and Halpern 2011) however, this approach requires agents to fix their policies over stages and to explicitly
compute approximate best-responses. In the following sections of the paper, we develop a technique which
enables equilibrium policies of MFGs to be computed by adaptive learners in an unknown environment
without solving partial differential equations.
6
Theoretical Contribution
Mean Field Games are Potential Games
We now demonstrate that the discrete-time MFG problem (B) is reducible to an objective maximisation
problem. By proving that the discrete-time mean field game is a PG, the following theorem enables us to
reduce the problem to a single OCP:
Theorem 1. The discrete-time mean field game (B) is a PG.
We defer the proof of the theorem to the appendix. The key insight of Theorem 1 is that the M-NE
of MFGs can be computed by considering a general form of a team game in which each agent seeks to
maximise the potential function. Crucially, thanks to Theorem 1, the problem of computing the equilibrium
policy is reduced to solving a control problem for the potential function.
Learning in Large Population MAS
We now develop a model-free decentralised learning procedure based on a variant of fictitious play using
the potentiality property. This generates a sequence of polices that converges to the M-NE of the discrete-
time MFG.
Firstly, it is necessary to introduce some concepts relating to convergence to equilibria:
Definition 5. Let {πi,n}n≥1 be a set of policies for agent i ∈ N . We define a path by a sequence of strategies
(cid:44) (πi,n)n≥1 ∈ ×n≥1Πi, where πi,n+1 is obtained from an update of πi,n using some given learning
ρi
π
rule.
π ∈ ×n≥1Πi is called an improvement path for agent i if after
Definition 6. Given π−i ∈ Π−i, the path ρi
every update the agent's expected reward increases, formally an improvement path satisfies the following
condition:
J i[·, πi,n+1, π−i] ≥ J i[·, πi,n, π−i], ∀i ∈ N .
(13)
Definition 7. A path converges to equilibrium if each limit point is an equilibrium.
We now describe a 'belief-based' learning rule known as fictitious play (Brown 1951) of which our
π ∈ ×n≥1Πi be a path, then the learning rule is a fictitious play process (FPP) if
method is a variant: Let ρi
the update in the sequence {πi,n}n≥1 is performed ∀x ∈ S, ∀i ∈ N as follows:
J i[x, π(cid:48), π−i,n] = J i[x, πi,n+1, π−i,n],
sup
π(cid:48)∈Π
(14)
so that πi,n+1 is a best-response policy against π−i,n.
If the FPP converges to equilibrium then we say that the game has the fictitious play property.
We now apply these definitions to the case of MFGs. We note that by the indistinguishability assumption
for MFGs, we can drop the agent indices in each of the above definitions. We define the FPP for the MFG
(B) by the following learning procedure ∀x ∈ S:
n(cid:88)
j=1
mj
x
J[x, π(cid:48), ¯mn
x] = J[x, πn+1, ¯mn
x], ¯mn
x
(cid:44) 1
n
sup
π(cid:48)∈Π
7
so that πn+1 is a best-response policy against ¯mn
x which summarises each agent's belief of the joint state of
all agents after the nth update.
In order to solve game (B) we therefore seek a learning process that produces a sequence {(πn, mn)}n≥1
for which {πn}n≥1 is an improvement path for the policy πn.
For the discrete-time MFG, we shall seek a pair (πn, mn) ∈ Π × P(H) that converges to equilibrium as
n → ∞ so that the sequence {(πn, mn)}n∈N converges to a cluster point (π, m) which is a solution to (B).
Our convergence result is constructed using results that we now establish. Before proving the result we
report an important result in the model-based setting:
Proposition 2 (Cardaliaguet, Hadikhanloo; 2017). Mean Field Games have the fictitious play property.
This result was established in (Cardaliaguet and Hadikhanloo 2017) within a continuous-time and model-
based setting. Here, given some initial belief about the distribution mn and some initial value function
vn associated to each agent's problem, the agents update the pair (vn, mn) according to a (model-based)
fictitious play procedure. This produces a paired sequence {(vn, mn)}n≥1 for which limn→∞(vn, mn) =
(v, m) where (v, m) is joint solution to the continuous-time MFG.
In order to compute the best responses at each step, the FPP discussed in (Cardaliaguet and Hadikhanloo
2017) requires agents to use knowledge of their reward functions. Moreover, the agents' update procedure
involves solving a system of partial differential equations at each time step. Obtaining closed solutions to
this system of equations is generally an extremely difficult task and often no method of obtaining closed
solutions exists.
We are therefore interested in procedures for which agents can achieve their M-NE policies by simple
adaptive play with no prior knowledge of the environment. We are now in position to state our main result:
Theorem 2. There exists a fictitious-play improvement path process such that the sequence {(πn, mn)}n≥1 ∈
Π × P(H) converges to an −M-NE of the game (A).
The following corollary demonstrates that we can construct a learning procedure that leads to an im-
provement path, the limit point of which is a solution of the MFG (B).
Corollary 1. Let {(un, mn)}n≥1 be a mean field improvement path generated by an actor-critic fictitious
play method, then {(un, mn)} converges to a cluster point {(u, m)}. Moreover, the cluster point {(u, m)}
is a solution to (B).
Corollary 1 immediately leads to our method which computes the optimal policies for the MFG (B).
The method uses an actor-critic framework with TD learning on the critic and policy gradient on the learner.
An episode is simulated using some initial belief over the distribution m over S. The agent then updates its
policy using an actor-critic and updates the distribution mk fictitiously.
The Approximation Error
In this section, we show that the Nash equilibria generated by the game (B) are approximate equilibria
for the N−player stochastic game (A).
Theorem 3. Let ¯π, π ∈ Π be the NE strategy profile for the game (A) and game (B) respectively and let be
a NE strategy profile for the MFG (B). Let mx and ¯mx be the distributions generated by the agents in the
mean field game and the N−player SG respectively; then there exists a constant c > 0 s.th ∀x ∈ S:
√
J(x, π, mx) − J(x, ¯π, ¯mx) < c\
N .
(15)
8
over, the approximation error from using a MFG to approximate the N−player game is O(cid:0) 1√
Theorem 3 says that the solution to the MFG (B) is in fact an −M-NE to the N−agent SG (A). More-
(cid:1).
As a direct consequence of theorem 3, we can use the actor-critic fictitious play method on the MFG
formulation (B) to compute near-optimal policy solutions of the stochastic game (A). Moreover, the error
produced by the mean field approximation vanishes asymptotically as we consider systems with increasing
numbers of agents.
N
Experiments
To investigate convergence of our method, we present three experiments drawn from benchmark prob-
lems within economics and control theory that involve large populations of strategically interacting agents.
In each case, we show that our method converges to an M-NE policy.
We firstly demonstrate the use of our technique in a (stochastic) congestion game, testing the conver-
gence to a stationary policy in a large-population system with a complicated reward structure. The second
problem is a supply and demand problem that we formulate as an SG allowing us to test convergence to
optimal policies in dynamic problems requiring long-term strategic planning in the presence of other learn-
ing agents. This demonstrates that our method is able to overcome the non-stationary interference of other
adaptive agents. Lastly, we apply our method to study a multi-agent generalisation of a fundamental problem
within optimal control theory, namely the linear quadratic control (LQC) problem. The analytic solution of
the LQC problem allows us to verify that our method converges to a known M-NE policy.
Experiment 1: Spatial Congestion Game
In the spatial congestion game the rewards are dependent on the agents' use of a shared resource (a sub-
region of S ⊂ R2) and the number agents using that resource. Games of this type are known as congestion
games and represent a large class of interactions e.g. spectrum sharing problems.
In the spatial congestion (SC) game there are N agents, given some initial position x0 ∈ S, each agent
chooses an action in order to move to a desired location xT ∈ S which is a terminal state. Certain areas of S
are more desirable to occupy than others, however the agents are averse to occupying crowded areas - they
receive the largest rewards for occupying parts of S that are both desirable and have relatively low concen-
trations of agents. The agents simultaneously select a movement vector u ∈ R2×1 resulting in movement to
a terminal state xT . Each agent then receives its reward L which depends on the desirability of the location
and the concentration of agents mxT at xT .
Formally, we model the desirability of a region xt ∈ S at time t as7:
L(xt, mxt) =
2π(cid:112)Σ
1
e−(xt−µ)T Σ−1(xt−µ)
(1 + mxt)α
,
where mxt ∈ P(H) is the density of agents at the point xt and µ ∈ R2, Σ ∝ 12×2 are given parameters
representing the mean and spread of the distribution of the rewards over S. The map L : S × P(H) → R
measures the instantaneous reward for an agent at xt with a local agent density mxt. The parameter α > 0
is a measure of each agent's averseness to occupying the same region as other agents with higher values
representing greater averseness.
Given some initial position x0 ∈ S and u ∈ R2×1, the transition dynamics are given by the following
expression:
(16)
7We note that the function L is continuously differentiable in S so that assumption 2 is a fortiori satisfied, moreover, it can be easily
xT = f1(x0, u, ) (cid:44) A1x0 + B1u + σ1,
verified that assumption 3 holds.
9
Figure 1: a) Rewards over 2000 episodes of training for a Gaussian distribution of desirability. b) agent
distributions for different averseness parameter α. c) agent distribution for bimodal Gaussian function reward
function.
where ∼ N (0, σ); σ1, A1, B1 ∝ 1(2×2) and c, σ ∈ R+.
The reward function for an agent is then given by the following expression:
J[x0, π, mx0 ] = ExT ∼f1
(cid:104)
L(xT , mxT ) − 1
2
uT Ru(cid:12)(cid:12)u ∼ π
(cid:105)
,
where R = η1(2×2) is a control weight matrix and η ∈ R is the marginal control cost (cost of movement).
Using the indistingishability condition, we have omitted agent indices.
At equilibrium each agent optimally trades-off state-dependent rewards with its proximity to nearby agents.
The problem generalises the beach domain problem studied in (Devlin et al. 2014) since we now con-
sider a reward function with state dependency. In particular, the desirability over S is described by a Gaussian
function over S. Moreover the problem we now consider consists of a system with 1000 interacting agents.
The problem is also closely related to the spectrum sharing problem, see (Ahmad et al. 2010).
In accor-
dance with the theory, our method converges to a stable policy - after 2,000 episodes of training we find that
the agent's policy and rewards stabilise Figure 1.a). Figure 1.b) shows the terminal distribution of agents over
S for α ∈ {1.0, 2.0, 3.0}. We observe that the agents learn to optimally trade-off state-dependent rewards
with distance from neighbouring agents resulting in a fixed terminal distribution of agents. As expected, the
agents disperse themselves further as the value of α is increased - in all cases converging to a stable distri-
bution over S.
Figure 1.c) shows the distribution of agents for a more complicated reward structure specified by a mix-
ture of two Gaussians over S with peaks at (−1, 0) and (0, 0)). We initialise the agents at the point (1, 0). In
this case, an individual agent's reward is given by the following expression:
J[x0, π, mx0] = ExT ∼f1
Li(xT , mxT ) − 1
4
uT
i Rui
(cid:105)
(cid:104) 2(cid:88)
i=1
10
rewards α=1 α=2 rewards α=2 α=3 a)b)c) Figure 2: a) Paths of agents with different initial points (coordinates indicated) for cost of movement η = 2.
Shown in black is the path of the rewards (demand) and in purple is the path of agents with η = 0. b) The
intra-episode rewards for the paths show in a). Note incursion of negative rewards.
2 e−(xt−µi)T Σ
where Li(xt, mxt) (cid:44) [16π2Σi]− 1
· (1 + mxt)−α and ui ∼ π. Since learning is internal to each agent, a possible (suboptimal) outcome is for
the agents to cluster at the nearest peak of rewards. However, using our method, the agents learn to spread
themselves across the state space and distribute themselves across both peaks.
Experiment 2: Supply with Uncertain Demand
Optimally distributing goods and services according to demand is a fundamental problem within logistics
and industrial organisation. In order to maximise their revenue, firms must strategically locate their supplies
given some uncertain future demand whilst considering the actions of rival firms which may reduce the
firm's own prospects.
We now apply our method to a supply and demand problem in which individual firms seek to maximise
their revenue by strategically placing their goods when the demand process has future uncertainty. The de-
mand process, which quantifies the level of demand associated with each point in space, is a priori unknown
and is affected by the actions of thousands of rival firms. Each firm directs supply of its goods to regions in
time and space however, the firms face transport costs so that each firm seeks to optimally trade-off trans-
portation costs and tracking the demand. As firms begin to concentrate on a particular area of demand, the
sale opportunities diminish, reducing the rewards associated to that region of demand.
We model this problem as an episodic problem with a distribution of rewards traversing a path through
the state space S ∈ R2 (illustrated in Figure 2.a)). Agents seek to locate themselves in areas of high con-
centrations of rewards for a fixed number (30) time steps. The agents are penalised for both movement and
occupying areas with a high density of other agents. The experiment tests the ability of the method to avoid
convergence to suboptimal outcomes. In particular, movement costs are highly convex so traversing the path
of rewards leads to low overall rewards.
Figure 2.a) illustrates the path of the agents after training for several different initial positions. An inter-
−1
i
(xt−µi)
11
a)b) esting outcome is that with non-zero movement costs, the agents learn to move horizontally to intersect the
path of the rewards at a later time. This behaviour conforms with intuition - to maximise long-term rewards
the agents must choose a path that initially incurs higher costs, forgoing immediate rewards whilst they tra-
verse the regions with sparse rewards.
Figure 2.b) shows that the agents are able to learn to initially incur negative rewards to maximise their cumu-
lative payoffs. Without long-term planning each agent would attempt to trace the same path as the rewards.
Such a strategy would lead to reduced overall rewards since attempting to match the locations of the rewards
is prohibitively costly. Setting η = 0 we see that the agents trace the path of the rewards (represented by the
dashed purple line in Figure 2.a)).
Experiment 3: Mean-Field Linear Quadratic Regulator
The linear quadratic control (LQC) problem is a fundamental problem within optimal stochastic control the-
ory (OSCT). It concerns a system whose transition dynamics evolves according to a stochastic process that
is linearly controlled subject to quadratic costs. The LQC problem essentially captures the local problem of
a large class of problems in OSCT and can be solved analytically, the solution being given by the linear-
quadratic regulator (Bardi 2012; Xu 2007). LQC models have been extended to mean field interactions in
which a large population of agents affect the dynamics of a system using linear controls subject to quadratic
control costs and a cost term which depends on the actions of other agents.
The reward function for the LQC problem is given by:
(cid:104) Tn(cid:88)
t=0
t Rut}(cid:105)
,
J[x0, u, mx0 ] = ExT ∼f2
{C(xt, ¯mxt) − 1
2
uT
where C(xt, mxt) (cid:44) −(xt − α)T Qt(xt − α).
At time k ≤ Tn, given some position xk ∈ S, each agent then chooses an vector control parameter ut ∈
R2×1. The transition are given by the following expression:
(17)
where k ∼ N (0, σk ) ∀k < Tn, A1, B1 ∝ 1(2×2), σ1 = c1(2×2) where c ∈ R+ is some constant that
measures the magnitude of the stochasticity in each agent's transition.
xk+1 = f2(xk, uk, k) (cid:44) A1xk + B1uk + σ1k,
In (Bardi 2012) the distribution mx is reported after convergence to the stationary M-NE. We compare
our results (E3) with this stationary policy (B1) in the following table:
B1
E3
µ
σ2
(0.50000, -0.50000)
(0.50717, -0.50537)
0.14060
0.16100
Clearly, our results converge to values that closely replicate the analytic solution.
Conclusion
We develop an approach to MARL with large numbers of agents. This is the first paper to prove convergence
results to best-response policies in multi-agent systems with an unbounded number of agents. This allows
RL to be applied across a broader range of applications with large agent populations, in contrast to current
methods (Hu and Wellman 2004; Littman 2001; Littman 1994). Our approach advances existing work in
MFGs as in (Cardaliaguet and Hadikhanloo 2017) that require both full knowledge of the environment and
12
to perform involved analytic computation. In contrast, by developing a connection between RL in MASs and
MFGs, we demonstrate a procedure that is model-free, enabling agents to learn best-response policies solely
through adaptive play which overcomes the problem of non-stationarity. In our experiments we provide a
novel approach of analysing problems in control theory and economics.
Acknowledgements
We would like to thank Haitham Bou-Ammar, Sofia Ceppi and Sergio Valcarcel Macua for helpful
comments.
References
[Ahmad et al. 2010] Ahmad, S.; Tekin, C.; Liu, M.; Southwell, R.; and Huang, J. 2010. Spectrum sharing as
spatial congestion games. Preprint arXiv:1011.5384.
[Bardi 2012] Bardi, M. 2012. Explicit solutions of some Linear-Quadratic Mean Field Games. Networks
and Heterogeneous Media 7(2):243 – 261.
[Borkar 2008] Borkar, V. S. 2008. Stochastic Approximation: A Dynamical Systems Viewpoint. Hindustan
Book Agency.
[Brown 1951] Brown, G. W. 1951. Iterative solution of games by fictitious play. Activity analysis of produc-
tion and allocation 13(1):374 – 376.
[Busoniu, Babuska, and De Schutter 2008] Busoniu, L.; Babuska, R.; and De Schutter, B. 2008. A compre-
hensive survey of multiagent reinforcement learning. IEEE Transactions on Systems, Man, and Cybernetics-
Part C: Applications and Reviews, 38 (2), 2008.
[Cardaliaguet and Hadikhanloo 2017] Cardaliaguet, P., and Hadikhanloo, S. 2017. Learning in mean field
games: The fictitious play. ESAIM: Control, Optimisation and Calculus of Variations 23(2):569–591.
[Cardaliaguet et al. 2015] Cardaliaguet, P.; Delarue, F.; Lasry, J.-M.; and Lions, P.-L. 2015. The master
equation and the convergence problem in mean field games. Preprint arXiv:1509.02505.
[de Finetti 1931] de Finetti, B. 1931. Funzione caratteristica di un fenomeno aleatorio.
Accademia Nazionale dei Lincei, volume 4, 251 – 299.
[Devlin et al. 2014] Devlin, S.; Yliniemi, L.; Kudenko, D.; and Turner, K. 2014. Potential-based difference
rewards for multiagent reinforcement learning. In Int. Conf. on Autonomous Agents and Multiagent Systems
(AAMAS), volume 1, 165–172.
[Gomes, Mohr, and Souza 2010] Gomes, D. A.; Mohr, J.; and Souza, R. R. 2010. Discrete time, finite state
space mean field games. Journal de Math´ematiques Pures et Appliqu´ees 93(3):308 – 328.
[Hu and Wellman 2004] Hu, J., and Wellman, M. 2004. Nash Q-learning for general-sum stochastic games.
Journal of Machine Learning Research 4(6):1039 – 1069.
[Kash, Friedman, and Halpern 2011] Kash, I. A.; Friedman, E. J.; and Halpern, J. Y. 2011. Multiagent
learning in large anonymous games. Journal of Artificial Intelligence Research 40:571 – 598.
[Lasry and Lions 2007] Lasry, J.-M., and Lions, P.-L. 2007. Mean field games. Japanese Journal of Mathe-
matics 2(1):229 – 260.
In Atti della R.
13
[Leibo et al. 2017] Leibo, J. Z.; Zambaldi, V.; Lanctot, M.; Marecki, J.; and Graepel, T. 2017. Multi-agent
reinforcement learning in sequential social dilemmas. In Proc. Conf. on Autonomous Agents and MultiAgent
Systems (AAMAS), 464 – 473.
[Leslie and Collins 2006] Leslie, D. S., and Collins, E. J. 2006. Generalised weakened fictitious play. Games
and Economic Behavior 56:285 – 298.
[Littman 1994] Littman, M. L. 1994. Markov games as a framework for multi-agent reinforcement learning.
In Proc. Int. Conf. on Machine Learning (ICML), volume 157, 157 – 163.
[Littman 2001] Littman, M. L. 2001. Friend-or-foe Q-learning in general-sum games. In Proc. Int. Conf. on
Machine Learning (ICML), volume 1, 322 – 328.
[Monderer and Shapley 1996] Monderer, D., and Shapley, L. S. 1996. Potential games. Games and Economic
Behavior 14(1):124 – 143.
[Shoham and Leyton-Brown 2008] Shoham, Y., and Leyton-Brown, K. 2008. Multiagent Systems: Algorith-
mic, Game-Theoretic, and Logical Foundations. Cambridge University Press.
[Sutton and Barto 1998] Sutton, R. S., and Barto, A. G. 1998. Reinforcement Learning: An Introduction,
volume 1. MIT press Cambridge.
[Tuyls and Weiss 2012] Tuyls, K., and Weiss, G.
prospects. AI Magazine 33(3):41 – 52.
[Xu 2007] Xu, Y. 2007. Uniformly Lipschitz feedback optimal controls in a linear-quadratic framework.
Journal of Mathematical Analysis and Applications 327(1):665 – 678.
2012. Multiagent learning: Basics, challenges, and
14
Supplementary Material
Experimental Setup
In all three experiments, all learning was performed in TensorFlow using actor-critic method with the
Adam optimiser. The actor is represented by a two-layer fully connected network represented by a neural
network with a Gaussian output with variance 0.1. The critic is a two-layer fully connected neural network.
The experimental parameters are as follows: the learning rates in experiments 1, 2 and 3 were set to 10−4,
10−3 and 10−3 respectively. In all experiments we used a discount factor of 0.99. Experiment 1 is a one-shot
game, experiment 2 was run for 30 time steps. We ran experiment 1 with 1,000 agents and experiments 2
and 3 with 200 agents. In experiment 1, the agents' initial position was sampled from a Gaussian distribution
N ((1, 0), 0.1). For experiment 2 the agents' initial position was sampled from a Gaussian N ((x, 0.10) where
the mean x took the values x ∈ {(−0.20, 0.00), (−0.20, 0.30),
(−0.39, 0.16), (−0.60, 0.00), (−0.6, 0.30)}. In experiments 1 and 2, for each initial position we performed 6
runs. For experiment 1 we set the marginal cost parameter η = 0 and for experiment 2, η = 2. In experiment
1 we tested for α ∈ {1.0, 1.5, 2.0, 2.5, 3.0} and in experiment 2 we set α = 0.1.
Assumptions
The results within the paper are built under the following assumptions on the instantaneous reward function
L and the functions f and g.
Assumption 1. The function f (x, y,·) is Lipschitz continuous in (x, y).
Assumption 2. The function g(x, y) is Lipschitz continuous in (x, y).
Assumption 3. The instantaneous reward function L(s, x, y, u) is Holder-continuous in (s, x, y) and convex
in u.
Assumption 4. The instantaneous reward function L is monotone in m (in the sense of definition 3) and
both convex and separable u.
Assumption 5. The function L(·, m) : P(H) → R is C 1 in the sense of (Cardaliaguet et al. 2015), that is
there exists a continuous map δL
δm s.t.:
(cid:90) (cid:90) δL
δm
L(·, m(cid:48)) − L(·, m) =
(x, (1 − t)m + tm(cid:48))(m(cid:48) − m)(dx)dt ∀m, m(cid:48) ∈ P(H).
Assumption 6. The instantaneous reward function L is bounded.
Technical Proofs
Proof of Theorem 1
The following set of results are instructive for the proof of theorem 1:
Proposition 3. Identical interest games are PGs.
The following result allows us to make use of proposition 1 in the context of mean field games:
15
Lemma 1. Mean field games are identical interest games.
Proof. The result follows immediately from the indistinguishability condition - in particular we note that
the potentiality condition (c.f. in Definition 2)) is satisfied by the instantaneous function L itself.
To prove the theorem, we exploit directly assumption 4 from which we can deduce the existence of the
quantity δL
δm. Thus, in full analogy with lemma 4.4 in (Monderer and Shapley 1996), we conclude that the
game has a potential and that the equilibria of the mean field game (B) can be obtained by maximising the
potential function.
Proof of Theorem 3
Proof. The proof of theorem 3 exploits both the boundedness properties and continuity of the functions f
and J. We build the proof in two parts the main part of which is given by the following result which we shall
prove immediately:
Given NE strategy profile for the game (A), ¯π ∈ Π and a NE strategy profile for the mean field game (B)
π then the following inequalities hold:
1. f (x,·) − f (¯x,·) < c√
2. J(x, π, mx) − J(¯x, ¯π, ¯mx) < c supk∈[t,T ] xk − ¯xk
N
where c > 0 is an arbitrary constant (that may vary in each line) and as before m and ¯m are the distributions
generated by the agents in the mean field game and the N−player stochastic game respectively.
The first inequality bounds the difference in trajectories between the instance that the agents use the
control for the mean field game and the control for the N−player stochastic game. The second inequality
bounds the change in rewards received by the agents after perturbations in their location. Since the agents
use closed-loop policies, we will make use of the result to describe changes in their policies.
Let N be the number of agents in the N−player stochastic game (A), then given ¯π and π as in part I,
k) and xk+1 =
if xN and x are solutions to the processes (1) and (7) respectively, i.e. xN
f (xk, ak, k) where ai
k ∼ ¯π and ak ∼ π.
k+1 = f (xi
k − xk(cid:107) ≤ ck√
To prove (cid:107)xN
we firstly must modify the representation of the function f as a function
By (Xu 2007), the policy π can be expressed by a function of the form Γ : Si × S−i → R s.t.: π =
k ∈ S−i we can therefore express the transition function f
purely in terms of the spatial variables.
Γ(xi, x−i), moreover, the function Γ is Lipschitz continuous in each variable.
in the following way ∀k ≤ T :
We note that for any given pair xi
k ∈ S i, x−i
k, ai
k, i
N
k, ζk) ≡ Θ(xi
k, x−i
k ∼ π
ai
f (xi
k, ai
k , ζk),
(A18)
where {ζk} is a set of i.i.d. random variables and the function Θ is a bounded, Lipschitz continuous function.
By the same reasoning (and recalling (6) - (7)), given a NE policy for the discrete MFG (B), π ∈ Π and
the induced mean field distribution mx ∈ P(H), we can similarly express the transition dynamics for the
discrete-time MFG in terms of the function Θ as f (xk, ak, ζk) ≡ Θ(xk, mxk , ζk) where ak ∼ π. Moreover,
given the boundedness of the function Θ, we also make the following observation which bounds the variance
taken w.r.t x ∈ S, ∀k ≤ T :
Θ(xi
k, x−i
k ,·)
≤ k
n
(1 + max
l
(cid:107)Θl(cid:107)∞)2 ≤ ck
N
,
(A19)
d(cid:88)
l=1
var
(cid:32)
N(cid:88)
j(cid:54)=i
1
n
(cid:33)
16
where (cid:107)fl(cid:107) denotes the bound on the lth component of Θ.
We secondly note that using (A18) and by the state transition equations (1) and (7) we note that we can
express the difference in trajectories as:
k − xk(cid:107) = E(cid:104)(cid:13)(cid:13)(cid:13) 1
N
(cid:107)xN
N(cid:88)
j(cid:54)=i
Θ(¯xi
k, ¯xj
k, ζ) −
(cid:90)
Θ(xk, x(cid:48), ζ)m(dx(cid:48))
(cid:13)(cid:13)(cid:13)(cid:105)
.
We now prove the statement by induction on the time index k ∈ N.
In the base case, we seek to prove the following bound:
S
,
(cid:107)xN
1 − x1(cid:107) ≤ ck√
N
(cid:90)
which is equivalent to the following inequality:
This follows straightforwardly since
N
j(cid:54)=i
E(cid:104)(cid:13)(cid:13)(cid:13) 1
N(cid:88)
E(cid:104)(cid:13)(cid:13)(cid:13) 1
N
N
(cid:13)(cid:13)(cid:13)(cid:105) ≤ ck√
(cid:13)(cid:13)(cid:13)2(cid:105)
Θ(x0, xj
0, ζ) −
Θ(x0, x(cid:48),·)m(dx(cid:48))
S
(cid:90)
S
N(cid:88)
≤ d(cid:88)
j(cid:54)=i
l=1
Θ(x0, xj
(cid:32)
0, ζ) −
N(cid:88)
j(cid:54)=i
1
n
var
Θ(x0, x(cid:48),·)m(dx(cid:48))
(cid:33)
Θ(x0, x−i,·)
≤ ck
N
,
using the variance bound (A19) by reordering the expression, after applying the Cauchy-Schwarz inequality
we arrive at the required result.
For the general case, we firstly make the following inductive hypothesis:
(cid:107)xN
k − xk(cid:107) ≤ ck√
N
,
We therefore seek to show that the following bound is satisfied:
k+1 − xk+1(cid:107) ≤ ck√
N
(cid:107)xN
,
which is equivalent to:
.
(A20)
To achieve this, we consider the term:
N
j(cid:54)=i
E(cid:104)(cid:13)(cid:13)(cid:13) 1
N(cid:88)
E(cid:104)(cid:13)(cid:13)(cid:13) 1
N
(cid:90)
S
Θ(xk, x(cid:48), ζ)m(dx(cid:48))
Θ(¯xi
k, ¯xj
k, ζ) −
N
(cid:13)(cid:13)(cid:13)(cid:105) ≤ ck√
(cid:13)(cid:13)(cid:13)(cid:105)
.
Θ(xi
k, x(cid:48), ζ)m(dx(cid:48))
N(cid:88)
j(cid:54)=i
(cid:90)
S
Θ(xi
k, xj
k, ζ) −
17
Now by the triangle inequality and the Lipschitz continuity of Θ we have that:
N(cid:88)
j(cid:54)=i
N
E(cid:104)(cid:13)(cid:13)(cid:13) 1
(cid:90)
k, ζ) −
N(cid:88)
S
k(cid:107) + c2
1
N
j(cid:54)=i
(cid:107)xj
k − ¯xj
+E(cid:104)(cid:13)(cid:13)(cid:13) 1
N
N(cid:88)
j(cid:54)=i
k − ¯xi
≤ c1(cid:107)xi
N(cid:88)
N(cid:88)
j(cid:54)=i
j(cid:54)=i
Θ(xi
k, xj
k, ζ) − 1
N
+ E(cid:104)(cid:13)(cid:13)(cid:13) 1
(cid:13)(cid:13)(cid:13)(cid:105)
k(cid:107) + E(cid:104)(cid:13)(cid:13)(cid:13) 1
N(cid:88)
k, ζ)
N
N
j(cid:54)=i
Θ(¯xi
k, xj
k, ζ)
(cid:13)(cid:13)(cid:13)(cid:105)
k, ζ) −
(cid:90)
S
(cid:90)
S
Θ(¯xi
k, xj
Θ(¯xi
k, ¯xj
Θ(¯xi
k, ¯xj
Θ(xk, x(cid:48), ζ)m(dx(cid:48))
Θ(¯xi
k, ¯xj
k, ζ) −
Θ(xk, x(cid:48), ζ)m(dx(cid:48))
(cid:13)(cid:13)(cid:13)(cid:105)
(cid:13)(cid:13)(cid:13)(cid:105)
,
where c1, c2 > 0 are arbitrary (Lipschitz) constants. After summing over i and dividing by N and, using
(A20) we deduce the required result.
pressed as π = Γ(xi, x−i) we express the instantaneous function L as L : S i × S−i using (Xu 2007).
For part (ii) we exploit the Lipschitzianity of the function L. Moreover, since the policy π can be ex-
We now observe the following estimate:
(cid:90)
E(cid:104)(cid:13)(cid:13)(cid:13) 1
N
N(cid:88)
N(cid:88)
j(cid:54)=i
j(cid:54)=i
L(¯xi
k, xi
k) −
L(xk, x(cid:48))m(dx(cid:48))
S
(cid:34) d(cid:88)
l=1
(cid:32)
N(cid:88)
j(cid:54)=i
1
n
(cid:13)(cid:13)(cid:13)(cid:105)
≤ c1(cid:107)xi
k − ¯xi
k(cid:107) + c2
1
N
(cid:107)xj
k − ¯xj
k(cid:107) +
var
Θ(xi
k, x−i
k ,·)
(cid:33)(cid:35) 1
2
.
where c1, c2 > 0 are arbitrary constants.
As in part (i), after summing in i and dividing by N and, using the result in part (i) we deduce the
result.
Proof of Theorem 2
Proof. The theorem is proved by demonstrating that given some initial belief m0 of the distribution m,
each agent can generate an iterative sequence { mn}{n∈N} s.th. { mn}{n∈N} → m as n → ∞. Thereafter,
we deploy a two-timescales method (Borkar 2008; Leslie and Collins 2006) so that the the beliefs of the
distribution (at a given state) are updated slowly and thus are quasi-static from the perspective of the updating
procedure for the policy. In particular, in the following, we show that provided the updating procedure to the
agent's belief of the distribution m are performed at a sufficiently pace relative to the updates to the policy,
the problem is reduced to an optimisation problem for each agent where at each stage of the iteration each
agent plays an approximate best-response given some belief of the distribution of the agents (evaluated at
the agent's own location). The agents improve their policies using a stochastic gradient procedure which,
with an update to the distribution that appears static, produces enough traction to enable convergence to the
optimal policy.
Consider first, each agent's optimal value function which, given a belief of the distribution m is given by
the following:
vπ, m(x) (cid:44) sup
π(cid:48)∈Π
J[x, π(cid:48), mx],
(21)
Suppose also that each agent has a belief mn of the distribution m over n = 0, 1, 2, . . ., iterations which are
updated using a procedure that we will later specify (this can in fact be viewed as a parameter of the value
18
function at iteration n), then following (21) we can define the nth iterate of v by the following expression:
vπ, m
n
(x) (cid:44) sup
π(cid:48)∈Π
J[x, π(cid:48), mn
x],
∀x ∈ S
(22)
vπ, m
(x) = sup
n,π(cid:48)∈Π
J[x, π(cid:48), mn
k=t (cid:107)L(cid:107)∞ < ∞, so
x] ≤(cid:80)T
Using the boundedness of L we observe that sup
n
that the sequence {vπ, m
(x)}n≥1 consists of bounded terms.
n
n
Let us analogously define the nth fictitious best-response as:
πn ∈ argsup
π(cid:48)∈Π
vπ(cid:48), mn
n
(x),
∀x ∈ S
(23)
n
mn+1
n = 0, 1, . . . ,
∀x ∈ S,
x = mn
x) + M ( m)
n+1],
x + β(n)[B(an, mn
n β(n) = ∞ and(cid:80)
Suppose that each iteration each agent updates their belief over the distribution via the procedure:
an ∼ vn
where β(n) is a positive step-size function s.th.(cid:80)
(24)
n β(n)2 < ∞, B : A × R → R is
some well-behaved function (recall that mx ∈ Rd is the agent density evaluated at the point ∈ S ⊂ Rd) and
M ( m)
is a martingale difference sequence8. In particular, we assume that B is Lipschitz continuous in both
variables and satisfies a growth condition i.e (cid:107)B(x, y)(cid:107) ≤ c(1 + (cid:107)(x, y)(cid:107)) for any (x, y) ∈ A × P(H) and
(cid:107)B(x, y) − B(w, z)(cid:107) ≤ d(cid:107)(x, y) − (w, z)(cid:107) for any (w, z), (x, y) ∈ A × P(H) for some constants c, d > 0.
We note that since the update can be viewed as an updating procedure over a parameterisation of the
n (parame-
value function, the update in (24) can be viewed as a critic update of the value function v m
terised by m1, m2, . . .) within the context of a mean field game. It can be straightforwardly shown that
n < ∞ (see for example, Appendix C in (Borkar 2008)).
n β(n)M ( m)
Since the action an is sampled from the policy πn and, using (23), we can rewrite (24) as the following
(cid:80)
expression:
mn+1
x = mn
(25)
for some map B(cid:48) which maps to and from the same spaces and satisfies the same assumptions as B (we have
suppressed the super-indices on the function v).
n = 0, 1, . . .
x + β(n)[B(cid:48)(vn, mn) + M ( m)
n+1],
∀x ∈ S
Our goal is to show that the following expression holds ∀x ∈ S:
n→∞ v πn, mn
lim
n
(x) = vλ(mx),mx
a.s.
(26)
where (λ(m), m) is a fixed point solution to the MFG system (B) (i.e. λ(m) ∈ supπ(cid:48)∈Π vπ(cid:48)(mx),mx (x)).
Let us now consider the following update procedure for vn:
n
x) + M (π)
where M (π)
vn+1(x) = vn(x) + α(n)[A(πn, mn
(27)
is a martingale difference sequence and A : Π× R → R is a Lipschitz continuous function that
n α(n) = ∞
np ) for some p > 1 so that the belief mn
is updated slowly relative to π and α(n)\β(n) → 0 as n → ∞. In a similar way it can be shown that
satisfies a growth condition in both variables and where α is a positive step-size function s.th.(cid:80)
and(cid:80)
(cid:80)
n α(n)2 < ∞ and is chosen such that α(n)\β(n) ∼ O( 1
n = 0, 1, . . .
n+1],
∀x ∈ S
n < ∞.
n α(n)M (π)
8W.r.t. the increasing family of σ− algebras σ(xl, M ( m)
l
, l ≤ n).
19
We can now rewrite the update process over m as the following:
x + α(n)[ B(cid:48)(vn, mn
n = 0, 1, . . .
(cid:80)
x) and M ( m)
n+1
α(n) M ( m)
x = mn
mn+1
x) (cid:44) β(n)
α(n) B(cid:48)(vn, mn
x) + M ( m)
n+1],
n+1. We note that the sequence { mn
(cid:44) β(n)
(28)
x}n≥1
where B(cid:48)(vn, mn
is quasi-static w.r.t. the sequence {vn}n≥1, moreover the sequence {vn}n≥1 converges when mn
x is fixed
at a particular n ∈ N (the result follows since it can be shown that vn strictly increases whenever vn is
suboptimal).
n < ∞. Analogously, we have that
x is defined over a bounded domain and, us-
mn
x < ∞. We can therefore apply theorem 2, ch. 6 in (Borkar 2008)
, mn) → (vλ(m),m, m) a.s. after which, using (23), we deduce the
vπ, m
n < ∞ moreover, since mn
Let us firstly recall that sup
n
n α(n) M ( m)
n β(n)M ( m)
n = (cid:80)
(x) < ∞ and (cid:80)
n
n α(n)M (π)
ing Assumption 2 we deduce that sup
n
from which find that limn→∞(v πn, m
thesis.
n
20
|
1811.01688 | 1 | 1811 | 2018-11-05T13:55:25 | Collaboration and integration through information technologies in supply chains | [
"cs.MA"
] | Supply chain management encompasses various processes including various conventional logistics activities, and various other processes These processes are supported -- to a certain limit -- by coordination and integration mechanisms which are long-term strategies that give competitive advantage through overall supply chain efficiency. Information Technology, by the way of collecting, sharing and gathering data, exchanging information, optimising process through package software, is becoming one of the key developments and success of these collaboration strategies. This paper proposes a study to identify the methods used for collaborative works in the supply chain and focuses on some of its areas, as between a company and its suppliers (i.e., inventory sharing) and its customers (i.e., customer demand, forecasting), and also the integration of product information in the value chain. | cs.MA | cs | Collaboration and integration through Information Technologies
in supply chains
Gilles NEUBERT, Yacine OUZROUT, Abdelaziz BOURAS
[[email protected]]
Université Lumière Lyon 2
Laboratoire PRISMa -- CERRAL
IUT LUMIERE
160 bd de l'universite
69676 BRON FRANCE
Abstract: Supply chain management encompasses various processes including various
conventional logistics activities, and various other processes These processes are supported --
to a certain limit -- by coordination and integration mechanisms which are long-term strategies
that give competitive advantage through overall supply chain efficiency. Information
Technology, by the way of collecting, sharing and gathering data, exchanging information,
optimising process through package software, is becoming one of the key developments and
success of these collaboration strategies. This paper proposes a study to identify the methods
used for collaborative works in the supply chain and focuses on some of its areas, as between
a company and its suppliers (i.e., inventory sharing) and its customers (i.e., customer demand,
forecasting), and also the integration of product information in the value chain.
Keywords: IT; supply chain; integration; collaborative processes; product management.
1.
INTRODUCTION
Today business is dominated by an accelerated change in many domains: market,
technology, organization, Information, product design, and so on. To face a global economy
and rapidly changing markets companies and their suppliers are led to new kind of
organization based on value network. To face the globalization of markets the role of
manufacturing companies has changed from supplying domestic markets with products, via
supplying international markets through export, to supply international markets through local
manufacturing [Rudberg & Olhager, 2003].
The complexity of products and technologies is increasing while time to market and
product life cycle is decreasing. As mentioned by Momme [Momme 2002], globalization and
technological innovation call for improved organizational adaptability and more flexible and
advanced systems relative to manufacturing, logistics, engineering, information and process
technology. This accelerated environment has driven companies to focus on their core
competencies and to outsource some activities or parts of their production. Responsive
delivery, without excessive inventory, requires short manufacturing cycle times, reliable
processes, and effective integration of disparate functional areas from an inter enterprise point
of view as well as from an intra enterprise one.
As highlighted by A.Y. Nahm et al [A.Y. Nahm et al, 2003], manufacturer operating in
this so called post-industrial environment focus on customers. Their organizational structure
has shift from functional multiple-level hierarchy organization to cross functional orientation,
with concurrent information flows and decision making, and relatively few layers in
hierarchy. These post industrial organizations focus on more autonomous work, learning, and
collaborative management rules.
From many years now, the concepts of Supply Chain (SC) and Supply Chain management
(SCM) have been developed in companies' environment, as well as in research papers.
The first one (SC) is relative to the physical organization and flows of the chain from
suppliers to end customer. A Supply Chain spans from the raw material suppliers, trough
manufacturing, storage, wholesalers and retail companies to the final consumers of the
product. It is important to notice that a Supply Chain does not end at the frontiers of an
enterprise; it is a chain that comes through different enterprises
The second one (SCM) concerns the organizational aspect and the optimization of the
various processes along the Supply Chain. It is mainly based on integration, coordination and
information sharing with the goal of achieving a competitive advantage. The important things
to notice are that Supply Chain Management has to be seen as a transversal process, and that
the objective of the Supply Chain Management is to satisfy the customer at the lowest
possible cost
As it is very difficult to have a global view of a Supply Chain and to manage it from
suppliers' supplier to customers' customer, Supply chain and Supply chain management can
bee seen from an extended point of view as well as from a restrictive point of view. In a broad
sense, a supply chain consists of two or more legally separated organizations also named
inter-organizational supply chain. The term of intra-organizational Supply chain is applied to
a large company with several sites in a narrow sense of the supply Chain (see [Stadtler 2000]).
To support these "intra enterprise" and "inter enterprises" (re)organizations new methods
and technologies have arisen. They have to support a better integration from suppliers to
customers, a quicker response to customer demand, a better adaptation to market needs and
new practices in manufacturing processes. As mentioned by Kobayashi et al [Kobayashi et al,
2003], much attention has been focused on the management methodology of Supply Chain
Management (SCM), which integrates business processes from suppliers to customers and
manage various tasks, such as sales, manufacturing, logistics, and finance.
In particular, Information Technology (IT), by the way of collecting, sharing and
gathering data, exchanging information, optimizing process through package software, has
been one of the key developments of SCM. Information Technology makes it possible to
process more information, more accurately, more frequently, from more sources, even from
all over the globe. And IT makes it possible to digest, to understand, and to act on this
growing abundance of information by using sophisticated analysis, modeling, and decision
support capabilities [Boyson et al, 2003]
Beyond the goal of improving the performances related to an existing chain for an existing
product, Supply chain management is also deployed to improved new product development
and launching. Companies are required to develop new product more quickly, less expansive
and of better quality. Traditional Research and Development supported by stand alone
specification based on product data management is not enough anymore. Customer
requirement and suppliers knowledge have to be integrated in the development of new
product. This highlights the importance of collaborative design and new kind of information
and knowledge are to be shared and exchange along the chain. Product development life cycle
naturally includes numerous people, operating in various departments, and typically from
multiple companies, each with locations in multiple countries around the world.
Considering all these dimensions of the Supply Chain management, this paper propose a
study of the literature to identify the methods used for collaborative works in the supply
chain. The selected items are:
Ø The relation between a company and its suppliers, such as inventory sharing, etc.
Ø The relation between a company and its customers, such as customer demand,
forecasting, etc.
Ø The relation all along the chain such as information sharing and PLM
Ø The optimization of a company processes, such as planning, etc
2.
INTEGRATION AND INFORMATION TECHNOLOGIES
As the current practice of business is increasingly networked and interconnected, more
than ever integration and information are going hand in hand. This last decade, one of the
most significant developments in Information Systems has been Enterprise Resource Planning
(ERP) systems.
For many companies, ERP is considered as a good answer to replace the fragmented back-
office systems and to manage and integrate cross functional business processes. As mentioned
by Umble et al (2003), ERP provides two major benefits that do not exist in non-integrated
departmental systems: (1) a unified enterprise view of business that encompasses all functions
and departments; and (2) an enterprise database where all business transactions are entered,
recorded, processed, monitored, and reported. This unified view increases the requirement for,
and the extend of, interdepartmental cooperation and coordination.
ERP is standardized software that attempt to put together all departments and function
across a company onto a single information system using one unique database. It gets through
the old standalone computer software in finance, HR, manufacturing, and so on, and replaces
them with a single unified software divided into modules that roughly approximates the old
standalone systems. Its role is to provide the connectivity and the common data models
needed to link and coordinate the disparate functional silos within the organization as
mentioned by White (2002). It aims at integrating key businesses and management processes
to provide a consistent view of what is going on in your organization. The great benefit of
ERP is integration; as it combines each department program and data all together into a
single, integrated software program that runs off a single database, all employees can use the
same information and business processes and get the same results when the system is
required. In this way the various departments can easily share information and communicate
with each other
Implementing an ERP increases drastically the transparency of the information system.
People now need to fill in real time the data bases so that each one in the process can use it for
its own job. As noticed by Koch (2002) "People in the warehouse who used to keep inventory
in their heads or on scraps of paper now need to put that information online. If they don't,
customer service reps will see low inventory levels on their screens and tell customers that
their requested item is not in stock".
For this reason, ERP systems are essentially considered as process-oriented IT tools for
improving business performance [Al-Mashari et al, 2003]. They are highly complex
information systems that work essentially at integrating data. The commercially available
software packages promise seamless integration of all information flows in the company-
financial and accounting information, human resource information, supply chain information,
and customer information [Umble et al, 2003].
But ERP projects are not only technical projects involving the information system or the
computers configuration. One of the major critical success factors highlighted during ERP
implementation is the organizational change management. ERP project have to be manage as
Business Process Reengineering project that shake up the whole organization. ERP software
embedded standardized solution so that it asks people to change the way they do their jobs. Of
course most ERP are build with a certain amount of "flexibility" in their solutions but it will
nether fit the way of doing business as well as the old specific in-house developments. That's
the reason why implementing an ERP in a company usually involves changing business
processes; and employee resistance to these changes can be a major brake on the project
success. As noticed by Koch (2002), "If you simply install the software without changing the
ways people do their jobs, you may not see any value at all -- indeed, the new software could
slow you down by simply replacing the old software that everyone knew with new software
that no one does".
As mentioned by Lamber and Cooper (2000), successful SCM requires a change from
managing individual functions to integrating activities into key supply chain processes. ERP
Systems are a good response (not the only one) to this organizational need of change inside a
company. They can now be considered as the core backbone to which other applications may
be linked as a mean of extending the functionality of enterprise software.
ERP packages are incomplete to fit an individual company's business processes. Further
on, it is common to see companies choose one ERP vendor for financials while choosing
another software vendor for human resources applications or production planning. As firms
moved toward business process rather than departmental applications, the need to integrate
functionalities is increasing: Shop floor, inventory, accounts receivable and advanced
planning application, HR, etc. need to communicate so that companies can make accurate
promises to customers, and executives can decide more quickly.
For some enterprises business processes and information are still maintained in
customised legacy systems based on dated technologies and stored in standalone databases.
Whether it is legacy systems or a packaged business solution, applications in most cases
could not effectively exchange information and the integration challenge resides now in
uniting the many different "island of automation" that exist within an enterprise. As written
by Erasala et al (2003), the need of the hour is to be able to share information and business
processes without having to make sweeping changes to existing applications and data
structures. A promising approach to achieve this is Enterprises Application Integration (EAI).
EAI represents an attractive proposition to firms, since it offers them the opportunity to
leverage their systems into a seamless chain of processes and present a unified view of their
business to customers.
EAI is not like Application Program Interfaces (API's) or 'point to point' solution but it
creates a common way for both business processes and data to speak to one another across
applications.
Irina et al. (2003) propose a taxonomy to help managers to identify technologies that can
be used for enterprise and cross-enterprise application, which can lead to the development of
an integrated infrastructure that support intra-and inter-organisational applications [Irani et al,
2003]. They distinguish: (1)
inter-
organisational application integration (cross enterprise), and (3) hybrid application integration
(B2C application)
intra-organisational application
integration, (2)
3. COLLABORATION AND IT
Historically, companies have focused only on their resources, constraints, and policies to
make decisions and reduce costs [Lambert & Cooper, 2000]. With intense competition and
reducing profit margins, this approach is no longer sufficient. They need to consider the
interactions with their suppliers and customers and incorporate them into their decision
making process. They also need to reformulate their business policies to enable them to
incorporate the information regarding their supply chain into their decisions.
As companies outsource more and more of their current in-house activities, they will have
to develop tools the software tools to control and collaborate with their outsourcers as well as
evaluate their performance in real time. These types of closely tracked relationships amongst
supply chain partners will put an increasing level of pressure on collaboration and execution
efficiencies [Wilson, 2002]. As mentioned by Akintoye et al., the various definitions which
have been proposed indicate that SCM prescribes organizational restructuring, extended to the
achievement of a company-wide collaborative culture [Akintoye et al., 2000]. In their paper
Thomas and Griffin defined three categories of operational coordination: (1) Buyer-Vendor
coordination,
Inventory-Distribution
coordination [Thomas and Griffin, 1996]. They show the breakdown of some models
reviewed in the literature according to their typology. From then, the rapid evolution of
Information Technology and the increasing use of internet technology have drastically
changed our way of doing business. As a consequence, the need for designing new business
processes with the assumption that there are no barriers between departments and enterprises
has increased. The internet can redefine how back-end operations -- product design and
development, procurement, production, inventory, distribution, after-sales service support,
and even marketing -- are conducted, and in the process alter the roles and relationships of
various parties, fostering new supply networks, services and business models [Lee and
Whang, 2001].
(2) Production-Distribution
coordination
and
In traditional organization, each company operates individually and manage its own
situation creating multiple decision points and minimal information sharing. Each actor is
responsible for his own inventory, production and distribution activities. On the other side,
Supply Chain Management stresses the need for process integration and new practices to
share responsibilities among supply chain partners. Effective supply chain management
requires coordination among the various channel members including retailers, manufacturers,
and intermediaries. Programs such as vendor managed inventory (VMI), continuous
replenishment planning (CRP), and Customer Relationship Management (CRM) have been
advocated by some as promising approaches to supply chain coordination. [Waller & al.,
1999].
3.1 Through existing collaborative processes: from VMI to CPFR
Analyzing the inventory decision in a distribution network has probably been the most
extensively research area. Traditionally, the focus has been on understanding the structure of
the optimal decision that minimizes a cost function ( in term of ordering quantities or review
levels, etc.) assuming some hypothesises (known demand, etc.) (see [Erenguç et al, 1999] ).
In order to reduce lead -- time and inventory cost, suppliers and customers have developed
cooperative models based on information sharing. As noticed by Huang et al [Huang et al,
2003] sharing of inventory level information can improve decision such as safety stock
placement, order replenishment and transhipment in a distribution supply chain. Knowing the
inventory level of downstream facilities can reduce the inventory holding cost of the whole
supply chain. Such a practice is called Vendor Management Inventory (VMI).
VMI is an operating model in which the supplier takes responsibility for the inventory of
its customer. In this kind of partnership, the supplier makes the main inventory replenishment
decisions for the customer. By this way, the supplier controls the buyer's inventory level, so
as to ensure that predetermined customer service levels are maintained. VMI systems achieve
its goal through more accurate sales forecasting methods and more effective distribution of
inventory in the supply chain
Achabal et al give an overview of the models developed for the VMI and develop a
decision support systems that combines inventory optimization methods an promotional
response models [Achabal et al, 2000]. It highlights the role of the collaboration between
customer (ie retailer) and supplier (ie vendor), the importance of information sharing between
the two partners, and the integration between inventory and forecast models.
Figure 1: Models used for VMI Decision Support System [Achabal et al, 2000]
Before any supply chain coordinating arrangement is entered into, the retailer and
manufacturer manage their inventories independently, with the manufacturer in a less
favourable position regarding information. The manufacturer relies on historical order data
from the retailer to predict both future ordering patterns of the retailer and true demand
patterns of the retailer's customers. Obviously, the manufacturer's understanding of the
retailer's actual demand pattern is distorted by the latter's demand forecasting and ordering
behaviour. Therefore, it is difficult to closely synchronize the manufacturer's delivery at the
retailer's location with demand. A mismatch between delivery and demand also causes
fluctuations in order releases. VMI reduces information distortion, which is one cause of the
bullwhip effect.
In addition, VMI provides the supplier with the opportunity to better manage its own
production, inventory, and transportation costs. For suppliers, the major attraction of VMI is
in smoothing demand [Kaipia and Tanskanen, 2003]. Large, infrequent orders from customers
force suppliers to maintain inventories that enable them to respond to the uneven demand. In
VMI, the supplier is able to smooth the peaks and valleys in the flow of goods, and therefore
to keep smaller buffers of capacity and inventory. The impact of demand amplification is
dampened as the manufacturer now receives a direct view of end customer demand patterns
and can use this in forecasting [Disney et al, 2003].
VMI can increase customer service levels and reduce stock-outs through better
understanding of demand and more sophisticated inventory policies. For retailers, benefits
come from increasing service levels with lower inventory costs.
The notion underlying such programs as VMI and continuous replenishment planning
(CRP is a move from pushing products from inventory holding areas to pulling goods onto
grocery shelves based on consumer demand) has been around for a long time, but recent
advances in technology, including electronic data interchange and the Internet, are now
making these programs feasible. For instance, two major barriers associated with managing
another organization's inventory at remote locations -- knowing the exact inventory status and
predicting the demand activity -- are being overcome as a result of the real-time capture and
transmission of transaction information (Stank et al., 1999).
Companies are now evolving to collaborative replenishment and co-managed inventory
solutions which are more balanced approach
initiatives. In collaborative
replenishment, both the customer and the supplier have the visibility to demand, and longer-
term forecast are used to highlight upcoming surges and lulls that call for advanced planning.
This is one issue of Collaborative Planning, Forecasting and Replenishment (CPFR),
which is a triptych of collaborative planning, collaborative forecasting and collaborative
replenishment. This model was proposed by the Voluntary Inter-Industry Commerce
Standards Organisation (VICS -- see http://www.cpfr.org ). CPFR is a process model, shared
by a buyer and supplier, through which inventory-status, forecast-, and promotion oriented
information are shared and replenishment decisions are made. It helps trading partners
generate the most accurate forecasts possible and set highly effective replenishment plans.
CPFR is centred on open business relationships that share information, data and
communication; it requires information technology to build, share, and adjust on-line
forecasts and plans.
to CRP
CPFR is the first step toward developing a broader collaborative relationship between
retailers and suppliers. Collaboration is a long-term strategy that give competitive advantage
through overall supply chain efficiency. CPFR streamlines the flow of physical goods to
reduce distribution costs and supply chain inventory and improve in-stock positions and sales.
It also facilitates a number of intangible benefits between trading partners and suppliers
including improved trust, increased collaboration in areas outside the scope of CPFR, and idea
sharing.
The two tables below show the typical CPFR Benefits for retailers and manufacturers:
Typical Improvement
2% to 8%
10% to 40%
5% to 20%
3% to 4%
Retailer Benefits
Better Store Shelf Stock Rates
Lower Inventory Levels
Higher Sales
Lower Logistics Costs
Manufacturer Benefits
Lower Inventory Levels
Faster Replenishment Cycles
Higher Sales
Better Customer service
Source: AMR Research April 2001- [AMR 2001]
3.2 Beyond industrial application: AI and Collaborative Decision Making
Typical Improvement
10% to 40%
12% to 30%
2% to 10%
5% to 10%
As previously mentioned in this paper SCM encompasses various processes and is mainly
supported by coordination and integration mechanism. These processes include of course
conventional logistics activities, such as inventory management, transportation, order
processing, and other processes such as customer and supplier relationship management,
product development, and so on.
Decisions are usually based on department's own constraints and are optimized locally
within the departments but do not assure a global optimum for the enterprise. Of course,
decision support tools exist for local decision-making, e.g. planning and scheduling systems,
inventory management systems, market trading optimization systems, etc. An integration of
these tools would not solve the problem and there is a need for a unified approach for
modelling and analysis of supply chains, which explicitly captures the interactions among
enterprises and within the departments (Julka & al. 2002).
Many information technologies have been developed for handling transactions in supply
chains such as electronic document interchange (EDI) and enterprise resource planning
(ERP). Lately, Internet-based technologies such as the ebXML and Web Service (Walsh,
2002) are emerging. However, despite the merits of these technologies, there are limitations in
the flexibility and dynamic coordination of distributed participants in supply chains. The
agent system is an alternative technology for supply chain management because of the
features such as distributed collaboration, autonomy, and intelligence (Nissen, 2001; Yuan et
al., 2001)
3.2.1 Multi-Agent Systems technology
Agents are considered as entities with goals, capable of actions endowed with domain
knowledge and situated in an environment to achieve internal goals, an agent must reason
about its environment, to generate plans and execute them. Multi Agent System (MAS) is an
organization of agents, interacting together to collectively achieve their goals (Chehbi & al.
2003). Software agents can be defined in different ways depending on the way they are
implemented and the tasks they perform. (Wooldridge & al.,1998).
Many researchers in distributed artificial intelligence have dealt with developing theories
and techniques to study and apply multi agent systems in organizations management where
agents collaborating will play a major role. A good summary of some applications based on
SMA can be found in (Wu, & al., 2000).
Research literature on intelligent agent system architectures has proven that problems that
are inherently distributed or require the synergy of a number of distributed elements for their
solution can be efficiently implemented as a multi-agent system (MAS) (Jennings, Sycara, &
Woolridge, 1998; Ferber, 1999). All agents participating in MAS communicate with each
other by exchanging messages, encoded in a specific Agent Communication Language
(ACL). Each agent in the MAS is designated to manipulate the content of the incoming
messages and take specific actions/decisions that conform to a particular reasoning
mechanism designed by the agent programmer.
Previous researches on agent-based supply chain management can be divided into three
types (Ahn 2003):
§ Agent-based architecture for coordination : in this first type of research, functional
units of the chain are modelled as independent and collaborating agent systems, such as
logistics agent, transportation agent, scheduling agent, etc. (Fox et al., 2000; Ito & Abadi,
2002). The aim of this type of research is to show how to coordinate various activities of
the agents in different conditions. For example, Reis et al. (2001) showed how distributed
agents can be coordinated for effective production scheduling.
§ Agent-based simulation of supply chains: the second type of research shows how agent-
based supply chains can gain visibility and efficiency through simulation under various
strategies (Fox et al., 2000). Fox et al. (2000) showed how agent systems can be modelled
for supply chain simulation when there are unexpected events in some of the companies
within the supply chain.
§ Dynamic formation of supply chains by agents: the last type of research shows how
supply chains can be formed dynamically meeting various environmental constraints that
may change over time. In this type of research, agent systems that represent each company
in a supply chain perform negotiations with other agents to form a virtual supply chain,
which lasts for a certain period to achieve a common objective. The major concern in this
type of research is to show how agent technology enables more flexible organization of
supply chains using autonomous, collaborative, and intelligent features. For example,
Chen et al. (2000) show how virtual supply chains can be formed by solving distributed
constraint satisfaction problems. In Walsh and Wellman's research (Walsh & Wellman,
1999), supply chains are formed by subcontract auctions of participating agents.
Ahn & All (2003) explain that one of the main benefits of using agent technology for
supply chain management is the dynamic formation of supply chains using negotiations or
contracts by agents (Chen et al., 1999). However, the limitation of existing agent systems is
that it is difficult to make agent based supply chains adapt to new products or new trading
partners because agent systems usually use a fixed set of transaction sequences, which is a
critical barrier to adaptability (Freire & Botelho, 2002). So, the purpose of their research is to
propose a flexible agent system, which is adaptable to the dynamic changes of transactions in
supply chains.
3.2.2 Collaboration and Multi-Agent Systems
In the last few years, multi-agent systems have been a preferred tool for solving supply
chain problems:
§
§ Goodwin, Keskinocak, Murthy, Wu, and Akkiraju (1999) present a framework for
providing decision support for an online exchange. They use a multi-agent system to find
matches of demand and supply on the exchange and provide the user with the best set of
transactions..
§ Sauter, Parunak, and Goic (1999) present an architecture called Agent Network for Task
Scheduling (ANTS), inspired by insect colonies and humans.Agents represent elements in
the supply chain and within a factory. Each firm in turn is viewed as a small supply chain,
thus the interface between agents within a
Julka & al. (2002) present a framework which helps to analyze the business policies with
respect to different situations arising in the supply chain. This architecture is based on
amulti-agent system.
§ Swaminathan, Smith, and Sadeh (1998) present a modelling and simulation framework for
developing customized decision support tools for supply chain reengineering. Agents
represent supply chain entities, e.g. customers, manufacturers, and transportation.
§ Another work presents the idea of open tender concept with blackboard-based negotiation
to develop a collaborative supply chain system (CSCS) (Ito & Salleh 2000). Blackboard-
based negotiation is based on the mechanism of negotiation among intelligent agents (IAs)
using a blackboard as a media where each agent exchange information for collaboration.
The key members communicate and interact each other in order to generate the solution
for efficient material supply process.
§ Symeonidis & all (2003) present a MAS for facilitating Supply Chain and Customer
Relation (SC-CR systems These system can be viewed as networks of collaborative units
that regulate, control and organize all distributed activities involved in procurement,
manufacturing, order processing, order transaction and product distribution.
The main objective of their work is the optimization of the quality of services provided by
the existing ERP, which provides a robust means for storing and manipulating a large amount
of data on company transactions. The choice of developing their architecture as a MAS,
provides the advantage of untroubled modification and extension of the system, according to
altering company requirements.
4. PRODUCT MANAGEMENT AND TECHNOLOGIES
The reduction of product development cycle time and the improvement of product quality
has been supported in the last two decades through the implementation of various computer
aided technologies. The introduction of the related tools, however, was often only task instead
of process oriented. Departments tried to automate their tasks as good as possible and created
an environment which best helps to meet their targets. In practice, only little integration took
place. Each tool produced its own data format which led to so called 'islands of automation.
This problem could only partially be resolved by various interfaces and data exchange
formats.
Romano (2001) in this context states, that during the engineering supply chain,
manufactures and suppliers need to work as design partners, however their design and
manufacturing systems are frequently incompatible. This introduces very significant time and
cost penalties as they attempt to share design information. Moreover, the large amount of data
generated by the various tools was often not centrally organised and therefore inaccessible to
others. New developments instead of using already designed parts were often the costly
consequences.
Realizing this problem and the associated business opportunity several software
companies started in the mid 1980s to develop Product Data Management (PDM) systems that
initially provided vaulting and file management capabilities for engineering documents. In the
late 1980s and early 1990s engineering change management to control and track the changes
made to engineering data was added to the functionality together with configuration and
classification management capabilities. As a result of design systems and teamwork
approaches entering the market PDM systems started to support the management of complex
relationships between parts, assemblies, drawings, metadata, people and groups of people.
Since the main functionality of PDM has been design centric data and workflow
management it was mainly used in the engineering organisation, sometimes in the
manufacturing organisation. This was due to the existing technology which made the
accessibility for other than design engineering departments difficult. The software was not
usable without extensive training and the files could only be viewed in the native design
system.
In the late 1990s a new class of software, Product Lifecycle Management (PLM) systems
evolved from PDM systems. Besides the term PLM there are other terms existing that were
established by different Research Firms1. With the advances in user interfaces and database,
viewing as well as the Internet technology the technological prerequisites were provided to
share data more easily. Providing integrated visualisation and Digital Mock-Up (DMU) Tools,
PLM systems make it possible to view, mark-up and redline native design data without the
need for having access to the native design system. The Internet serves as highly effective
platform to communicate product data information far beyond the engineering organisation.
Nadamuni (1999) therefore states that the vision for PLM is to do more than just design
centric data and workflow management and become an enterprise application that ties
together all the information sources in a corporation.
A PLM system can be described as an enterprise-wide Information Technology (IT)
infrastructure to support management of product definition throughout its complete lifecycle
(from initial concept to product obsolescence). Including workflow management, PLM
systems, as a single source of product information, ensure that up to date information are
available and accessible for the right people in the right format at the right time. It is also
viewed as an effective tool in managing the product definition supply chain by serving as an
informational bridge connecting OEM's, partners, subcontractors, vendors, consultants and
customers (Miller, 1998).
PLM systems provide a consistent view on product related information in the extended
enterprise whereas the easily sharing of product data facilitates real-time collaboration across
individuals and groups.
departmental barriers and among geographically dispersed
1 CPC (Collaborative Product Commerce), cPDM (collaborative Product Definition Management, ePLM
(electronic Product Lifecycle Management), PDC (product definition and commercialisation), PIM (Product
Information Management).
(i.e. with suppliers and customers). PLM means
To cope with the industry as well as social trends and to gain competitive advantage a new
integrated approach for developing products with respect to the whole product lifecycle has to
be taken. Integration has to take place internally (between different departments) as well as
externally
the management of
comprehensive, accurate and timely information over the entire product lifecycle in order to
realise collaborative product development. PLM systems are gaining acceptance for managing
all information about a corporation's products throughout their full lifecycle, from
conceptualization to operations and disposal (including requirements management, design,
engineering, manufacturing, procurement, sourcing, maintenance and service, and more.
These sub-processes are linked by a business process workflow). The PLM systems also form
the apex of the corporate software hierarchy and as such depend on subsidiary systems for
detailed information capture and dissemination. The tools that support PLM philosophy can
be broadly categorized as (1) Authoring tools that create product content, and (2)
Optimization and decision support system tools that manage and optimize the processes
involved in the lifecycle of the product.
It is therefore important not only to take the technological but also the organizational
perspective of PLM into account. Collaborating and sharing of data within a company and
especially in the extended enterprise challenges existing processes and culture. For that
reason, the success of PLM heavily depends on the willingness of the organization to accept
change and especially on the people that must form inter-organizational and cross-functional
teams to collaboratively develop and manage products.
Consequently the Aberdeen Group (1999) defines CPC respectively PLM as a class of
software and services that uses Internet technologies to permit individuals -- no matter what
role they have in the commercialisation of a product, no matter what computer based tool they
use, no matter where they are located geographically or within the supply net - to
collaboratively develop, build and manage products throughout the entire lifecycle.
In a context of product development process that is distributed among many value chain
participants, PLM also delivers application support to neatly connect to and accommodate the
workings of other processes, including ERPs, CRM, Component Supplier Management
(CSM), etc. The benefits of PLM are realized once previously disparate systems such as
CAD, PDM, CRM, ERP, SCM.., are integrated (ie. CAD and PDM can share existing design
data with the CRM sales configurator). These configurations become the preferred options or
option packages available to the customer. The CRM sales configurator may even be
powerful enough to present configurations beyond the pre-engineered ones, but within
engineered constraints, allowing customers more freedom to choose exactly the combination
of features they want. When giving customers this degree of flexibility, SCM also plays a role
at the beginning of the development process, enabling a company to "build to order". The
customer-selected preference is passed to the SCM requirement planning module. SCM
generates and manages the requirements and then passes them to PLM, where they enter the
design stage of product development. Parts are sourced using a Component Supplier
Management system, also part of SCM. This allows leverage of referred suppliers and design
re-use. PLM portal technologies facilitate supplier involvement during the design stage.
Once the initial customer-specific design is proposed, the next step is to simulate the
product performance on the computer, saving tremendous time and money over creation and
destruction of physical prototypes. And if the simulation shows that the virtual prototype isn't
optimized, there's still enough time and money left to go back and try again. Next,
manufacturability is verified through advanced simulation, and a final Manufacturing Bill of
Material (MBOM) is generated. The MBOM, process plans and other technical data are
passed to ERP and SCM for product production.
In fact PLM plays a significant role even while a product is in service. For example, PLM
can help companies reduce service and maintenance costs by providing product content
information to call centers and field service centers. Call centers can dynamically share a
three dimensional model of the product while interacting with a customer. While conducting
field service, maintenance engineers can use three dimensional product information to
understand maintenance procedures, shortening maintenance times and ensuring higher
product quality. As parts are repaired or replaced, the product model can even be updated to
reflect the product as maintained, versus the product as built.
4.1 Life-cycle standardisation technologies : the PLCS initiative…
PLCS (Product Life-Cycle Systems) is the first and only programme under the auspices of
ISO/TC184/SC4 to build upon ISO 10303 (STEP) in extending the capability of STEP to
enable the exchange, sharing and archiving of product data in all aspects of though-life
support for an extensive range of industry sectors. PLCS initiative aims to accelerate
development of new standards for product data in the area of support during operation and
disposal of complex engineering assets. The initiative is international and produces draft ISO
standards as the result of an initial three-year work programme. These standards are the
mechanism by which to ensure support information is aligned with the evolving product
definition over the entire life-cycle.
The first year was spent in capturing through-life business process models that represent
the essential facets of life-cycle support. The four specific areas for the modelling were:
• Support Engineering -- specifying, providing and sustaining the support infrastructure
(the support system);
• Configuration Management -- managing changes to significant product items within the
engineered asset and associated support system, including tracking of serial-numbered items;
• Resource Management -- buying, storing, packing, moving, issuing and disposing of
product items and elements of the support system;
• Maintenance and Feedback -- maintaining, testing, diagnosing, calibrating, repairing and
modifying product items and support elements, including the roles of schedules, resources and
feedback.
5. CONCLUSIONS
In the new economic context, companies have focused on their core competency and externalised
many sorts of activities. This outsourcing of some parts of their production, has led to a wholly
distributed environment where performances are reached through highly interconnected organisations
and processes. It has also put the stress on the role of integration and coordination to get a lot out of
this new kind of business.
It is obvious that the availability of information technology has played a role to this issue in
supporting these interactions to implement relationship strategies. IT permits databases consistency,
real-time data exchange and, information sharing that are the bases for integrated and collaborative
work.
In this paper, we highlighted the importance of software application integration, such as ERP and
EAI, as a base for integration and development of collaborative tools and methods. We examined
some industrial cross-functional integration through co-managed processes that integrate both
suppliers and customers (VMI and CPFR), and show how Multi Agent Systems could help in
developing future research dealing with these distributed organisations. Finally, we have drawn
attention to the significant role of product information as a vector of integration all along the supply
chain.
The future generation of applications will adapt to the internet-based world of today and
tomorrow, through changes in functionality, technology and architecture:
• functionality changes will become deeper and more specific to industry domain requirements
• technology will evolve to leverage the internet for both inter-enterprise connectivity and a
unification of the end-user experience
• architectural changes will allow for more integration and interoperability.
These changes will also cause vendors and user enterprises to evolve:
• vendors will focus more specifically on domain requirements, providing feature-rich applications
that emphasise complete process integration
• enterprises, freed from the monolithic architectures of existing systems will create system
deployment strategies that best fit enterprise needs -- without relying on a single vendor to deliver all
application components
• standards for data exchange will become of increasing importance in facilitating integration and
interoperability.
However, it is important to notice that despite the increasing capabilities of the technology,
managing and developing collaborative attitude need to be discussed within and between the partners.
New supply chain systems process data as they are programmed to do, but the technology cannot
absorb a company's history so that cultural impact of Business Process Reengineering for example, is
in fact more challenging than the technical issues. The difficulties in organisational transformation
should not be underestimated; changing people and the way they do business is, in fact much more
difficult than implementing new technologies.
6. REFERENCES
• Rudberg, M. and Olhager, J. (2003) 'Manufacturing networks and supply chains: an
operations strategy perspective', Omega, Vol. 31, pp.29 -- 39.
• Momme, J. (2002) 'Framework for outsourcing manufacturing: strategic and operational
implications', Computers in Industry, Vol. 49, pp.59 -- 75.
• Nahm, A.Y., Vonderembse, M.A. and Koutfteros, X. (2003) 'The
impact of
organizational structure on time-based manufacturing and plant performance', Journal of
Operations Management, Vol. 21, pp.281 -- 306.
• Stadtler, H. (2000) 'Supply chain management: an overview', in Kilger, C. (Eds.): Supply
Chain Management and Advanced Planning -- Concepts, Models, Software and Case
Studies, Springer-Verlag, pp.7 -- 28.
• Kobayashi, T., Tamki, M. and Komoda, N. (2003) 'Business process integration as a
solution to the implementation of supply chain management systems', Information &
Management, Vol. 40, pp.769 -- 780.
• Boyson, S., Corsi, T. and Verbraeck, A. (2003) 'The e-supply chain portal: a core business
model', Transportation Research, Part E39, pp.175 -- 192.
• Umble, E.J., Haft, R.R. and Umble, M.M. (2003) 'Enterprise resource planning:
implementation procedures and critical success factors', European Journal of Operational
Research, Vol. 146, pp.241 -- 257.
• White, J.W. (2002) 'Making ERP work the way your business works', CIO.com Research
Center Digest, White Paper.
• Koch
(2002)
'The ABCs of ERP', CIO.com Research Center Digest,
http://www.cio.com/research/erp/edit/erpbasics.html.
• Al-Mashari, M., Al-Mudimigh, A. and Zairi, M. (2003) 'Enterprise resource planning: a
taxonomy of critical factors', European Journal of Operational Research,Vol.146, pp.352-
364.
• Lambert, D.M. and Cooper, M.C. (2000) 'Issues in supply chain management', Industrial
Marketing Management, Vol. 29, pp.65 -- 83.
• Erasala, N., Yen, D.C. and Rajkumar, T.M. (2003) 'Enterprise application integration in
the electronic commerce world', Computer Standard & Interfaces, Vol. 25, pp.69 -- 82.
Irani, Z., Themistocleous, M. and Love, P.E.D. (2003) 'The impact of enterprise
application integration on information system lifecycles', Information & Management,
Vol.41, pp.177-187.
•
• Wilson, G.D. (2002) 'Worldwide logistics software forecast and analysis, 2002 -- 2006',
IDC Bulletin, No. 27390, June.
• Akintoye, A., McIntosh, G. and Fitzgerald, E. (2000) 'A survey of supply chain
collaboration an management in the UK construction industry', European Journal of
Purchasing & Supply Management, Vol. 6, pp.159 -- 168.
• Thomas, G.J. and Griffin, P.M. (1996) 'Coordinated supply chain management', European
Journal of Operational Research, Vol. 94, pp.1 -- 15.
• Lee, H.L. and Whang, S. (2001) 'E-business and supply chain integration', Standford
Global Supply Chain Management Forum, SGSCM-W2, p.20.
• Waller, M., Johnson, M.E. and Davis, T. (1999) 'Vendor-managed inventory in the retail
supply chain', Journal of Business Logistics, Vol. 20, No. 1, pp.183 -- 203.
• Erenguç,
S.S.,
Simpson, N.C.
'Integrated
production/distribution planning in supply chains: an invited review', European Journal of
Operational Research, Vol.115, pp.219 -- 236.
and Vakharia, A.J.
(1999)
• Achabal, D.D., Mcintyre, S.H., Smith, S.A. and Kalyanam, K. (2000) 'A decision support
system for vendor managed inventory', Journal of Retailing, Vol. 76, No. 4, pp.430 -- 454.
• Kaipia, R. and Tanskanen, K. (2003) 'Vendor managed category management -- an
outsourcing solution in retailing', Journal of Purchasing & Supply Management, Vol. 9,
pp.165 -- 175.
• Disney, S.M. Potter, A.T. and Gardner, B.M. (2003) 'The impact of vendor managed
inventory on transport operations', Transportation Research Part E: Logistics and
Transportation Review, Vol. 39, No. 5, pp.363 -- 380.
• Stank, T., Crum, M. and Arango, M. (1999) 'Benefits of interfirm coordination in food
industry supply chains', Journal of Business Logistics, Vol. 20, No. 2, pp.21 -- 41.
• AMR (2001) in [http://www.syncra.com/newsroom/CPFR_ValueProp.pdf].
•
Julka, N., Sirinivasan, R. and Karimi, I. (2002) 'Agent -- based supply chain management:
framework', Computers and Chemical Engineering, Vol. 26, pp.1755 -- 1769.
• Nissen, M.E. (2001) 'Agent-based supply chain integration', Information Technology and
Management, Vol. 2, pp.289 -- 312.
• Yuan, Y., Liang, T.P. and Zhang, J.J. (2001) 'Using agent technology to support supply
chain management: potentials and challenges', Michael G. DeGroote School of Business
Working Paper No. 453, http://merc.mcmaster.ca/workingpapers.html.
• Chehbi, S., Derrouiche, R., Ouzrout, Y. and Bouras (2003) 'Multi-agent supply chain
architecture to optimize distributed decision making', SCI'2003 Multi-conference on
Informatics and Cybernetics, Orlando FL, Vol. 16, pp.306 -- 311.
• Wooldridge, M.J, Jennings, N.R. and Sycara, K. (1998) 'A roadmap of agent research and
development',Autonomous Agents and Multi-Agent Systems, Kluwer Academic, Boston,
pp.7-38
• Ahn, H.J., Lee, H. and Park, S.J. (2003) 'A flexible agent system for change adaptation in
supply chains', Expert Systems with Applications, Vol. 25, pp.603 -- 618.
• Fox, M.S., Barbuceanu, M. and Teigen, R. (2000) 'Agent-oriented supply chain
management', Int. Journal of Flexible Manufacturing Systems, Vol.12, No.2/3, p.165-188.
Ito, T. and Abadi, S. (2002) 'Agent-based material handling and inventory planning in
warehouse', Journal of Intelligent Manufacturing, Vol. 13, pp.201 -- 210.
•
• Symeonidis, A.L., Kehagias, D.D. and Mitkas, P.A. (2003) 'Intelligent policy
recommendations on enterprise resource planning by the use of agent technology and data
mining techniques', Expert Systems with Applications, Vol. 25, pp.589 -- 602.
• Swaminathan, J.M. (1997) 'Modeling supply chain dynamics: a multi-agent approach',
Decision Sciences, Vol. 29, No. 3, pp.607 -- 632.
• Kabachi, N. and Ouzrout, Y. (2003) 'A society of agents for decision making in SCM',
SCI'2003 Multi-conference on Informatics and Cybernetics, Orlando, Vol. 16, p.317-321.
• Sauter, J.A., Parunak, H.V.D. and Goic, J. (1999) 'ANTS in the supply chain', Working
Notes of the agents'99 Workshop on Agents for Electronic Commerce and Managing the
Internet-Enabled Supply Chain, Seattle, WA.
• Chen, Y., Peng, Y., Labrou, Y., Cost, S., Chu, B., Yao, J., Sun, R. and Willhelm, B.
(1999) 'A negotiation-based multi-agent system for supply chain management', Working
Notes of the Agents'99 Workshop on Agents for Electronic Commerce and Managing the
Internet- Enabled Supply Chain, Seattle, WA, pp.15 -- 20.
• Walsh, W.E. and Wellman, M.P. (1999) 'Modeling supply chain formation in multi-agent
systems', Proceedings of the Agent Mediated Electronic Commerce Workshop (IJCAI-
99), pp.94 -- 101.
• Wu, D.J. (2001) 'Software agents for knowledge management: coordination in multi-
agent supply chains and auctions', Journal of Expert Systems with Applications, Vol. 20,
pp.51 -- 64.
• Freire, J. and Botelho, L. (2002) 'Executing explicitly represented protocols', Proceedings
of the AAMAS Workshop on Agent cities: Challenges in Open Agent Systems, pp.37 -- 40.
round',
http://www.mcadcafe.com/NEWS/MCAD_News/20010611_proficiency, Press Release,
June.
$27 million
• Romano,
investment
J.
(2001)
'Proficiency
closes
• Nadamuni, D. (1999) 'New directions in the PDM market', Gartner Group-Dataquest
Report, May.
• Miller, E. (1998) 'Managing the design supply chain', Computer-Aided Engineering
Magazine, Nov., http://www.cimdata.com/article9811.html.
• STEP (ISO10303) http://www.nist.gov/sc4.
|
1807.00376 | 1 | 1807 | 2018-07-01T19:18:09 | Human Satisfaction as the Ultimate Goal in Ridesharing | [
"cs.MA"
] | Transportation services play a crucial part in the development of modern smart cities. In particular, on-demand ridesharing services, which group together passengers with similar itineraries, are already operating in several metropolitan areas. These services can be of significant social and environmental benefit, by reducing travel costs, road congestion and co2 emissions. The deployment of autonomous cars in the near future will surely change the way people are traveling. It is even more promising for a ridesharing service, since it will be easier and cheaper for a company to handle a fleet of autonomous cars that can serve the demands of different passengers.
We argue that user satisfaction should be the main objective when trying to find the best assignment of passengers to vehicles and the determination of their routes. Moreover, the model of user satisfaction should be rich enough to capture the traveling distance, cost, and other factors as well. We show that it is more important to capture a rich model of human satisfaction than peruse an optimal performance. That is, we developed a practical algorithm for assigning passengers to vehicles, which outperforms assignment algorithms that are optimal, but use a simpler satisfaction model.
To the best of our knowledge, this is the first paper to exclusively concentrate on a rich and realistic function of user satisfaction as the objective, which is (arguably) the most important aspect to consider for achieving a widespread adaption of ridesharing services. | cs.MA | cs | Human Satisfaction as the Ultimate Goal in Ridesharing
Chaya Levinger, Amos Azaria, Noam Hazon
Department of Computer Science, Ariel University
[email protected], [email protected], [email protected]
8
1
0
2
l
u
J
1
]
A
M
.
s
c
[
1
v
6
7
3
0
0
.
7
0
8
1
:
v
i
X
r
a
Abstract
Transportation services play a crucial part in the de-
velopment of modern smart cities. In particular, on-
demand ridesharing services, which group together
passengers with similar itineraries, are already op-
erating in several metropolitan areas. These ser-
vices can be of significant social and environmental
benefit, by reducing travel costs, road congestion
and co2 emissions. The deployment of autonomous
cars in the near future will surely change the way
people are traveling.
It is even more promising
for a ridesharing service, since it will be easier
and cheaper for a company to handle a fleet of au-
tonomous cars that can serve the demands of differ-
ent passengers.
We argue that user satisfaction should be the main
objective when trying to find the best assignment
of passengers to vehicles and the determination of
their routes. Moreover, the model of user satisfac-
tion should be rich enough to capture the traveling
distance, cost, and other factors as well. We show
that it is more important to capture a rich model of
human satisfaction than peruse an optimal perfor-
mance. That is, we developed a practical algorithm
for assigning passengers to vehicles, which outper-
forms assignment algorithms that are optimal, but
use a simpler satisfaction model.
To the best of our knowledge, this is the first pa-
per to exclusively concentrate on a rich and realistic
function of user satisfaction as the objective, which
is (arguably) the most important aspect to consider
for achieving a widespread adaption of ridesharing
services.
1 Introduction
The National Household Travel Survey performed in the U.S.
in 2009 [Santos et al., 2011] revealed that approximately
83.4% of all trips in the U.S. were in a private vehicle (other
options being public transportation, walking, etc.). The av-
erage vehicle occupancy was only 1.67 when compensating
for the number of passengers (i.e., if two people travel in
the same vehicale, their travel distance is multiplied by two).
This extremely low average vehicle occupancy entails a very
large number of vehicles on the road that collectively con-
tribute to carbon dioxide emissions, fuel consumption, air
pollution and an increase in traffic load, which in turn requires
additional investment in enlarging the road infrastructure. In
recent years, ride hailing services such as Uber and Lyft have
gained popularity and an increasing number of passengers use
these services as one of their main means of transportation
[Wallsten, 2015]. Both Uber and Lyft are now also offer-
ing ride-sharing options, and other companies, such as Super-
Shuttle and Via, are explicitly targeted at customers who want
to share their ride.
The deployment of autonomous vehicles in the near future
will have a significant impact on the way people are travel-
ing. The implication of this revolutionary way of transporta-
tion is not fully known nowadays [Guerra, 2016], but it is
safe to claim that autonomous vehicles will have a positive
effect on the development of ridesharing services. Indeed, it
will be easier and cheaper for a company to handle a fleet of
autonomous vehicles that can serve the demands of different
passengers. It can also rule-out some negative human-driver
factors, such as driver's fatigue from the long travels and the
driver's inconvenience from having multiple pick-up and drop
stops along his route.
The basic challenge of a ridesharing service is how to as-
sign the passengers' requests for a ride to vehicles and de-
fine the routes for the fleet of vehicles in an optimal manner.
This problem belongs to the generic class of Vehicle Rout-
ing and scheduling Problems (VRPs), which have been ex-
tensively studied over the past 50 years, mainly in the oper-
ation research and transportation science communities. Sev-
eral variants with different characteristics have been devel-
oped. For example, the initial formulation of the VRP as-
sumes that the environment is static, i.e., all requests are
known before-hand and do not change thereafter [Dantzig
and Ramser, 1959]. The more complex variants, includ-
ing the rideharing problem, are dynamic, where real-time
requests are gradually revealed along the service operating
time.
In this setting the assignment of passengers to vehi-
cles and the determination of vehicles' routes may be adjusted
when they are already in transit [Psaraftis et al., 2016; Shen
et al., 2016]. Arguably, a major criterion that characterizes
each variant of the VRP is the objective function. It is very
common to consider objectives from the service provider's
perspective, for example, minimizing the total distance trav-
elled [Secomandi, 2000], minimizing the fleet size [Diana et
al., 2006; Secomandi and Margot, 2009], or maximizing the
service provider's profit [Campbell and Savelsbergh, 2005;
Parragh et al., 2014]. However, as noted by Cordeau and
Laporte [Cordeau and Laporte, 2003], one should be inter-
ested not only in minimizing the operating costs for the ser-
vice provider but also in maximizing the quality of the service
and the user satisfaction.
Many works integrate quality of service and user satisfac-
tion considerations as additional constraints of the problem.
For example, a time window restricts the waiting time a pas-
senger is willing to face before being picked up [Jaw et al.,
1986; El-Sherbeny, 2010], and it is usually combined with a
bound on the maximum user ride time [Paquette et al., 2009].
In addition, there are several works that combine the afore-
mentioned operational objectives with the objective of max-
imizing the user satisfaction (or its antonym, minimizing the
user inconvenience). The common interpretation for user sat-
isfaction is the minimization of the total user on-board (ride)
time and the total user waiting time [Psaraftis, 1980], the
extra riding time due to ride-sharing [Lin et al., 2012], or
the amount of deviations from desired departure and arrival
times [Fu and Teply, 1999; Yang et al., 2013]. However, to
the best of our knowledge, there are no works that exclusively
focus on maximizing a complex user satisfaction function,
which captures the traveling distance, cost, and other factors
as well.
Modeling human behaviour is not a easy task, and a theo-
retical model might fail to accurately capture real human be-
haviour. We therefore ran experiments with actual humans
and build a deep learning based function to estimate user
satisfaction. We introduce Simsat, an algorithm for assign-
ing passengers to vehicles while maximizing a complex user
satisfaction function as the objective, for the ridesharing last
mile variant [Cheng et al., 2014] setting. We show that Sim-
We investigate a comprehensive human-centric approach
for the ridesharing problem. Our basic claim is that the user
satisfaction should be the main objective of the ridesharing
service. Moreover, the model of user satisfaction should be
rich enough to capture the complex interdependencies among
several factors. Therefore, we develop a method for maxi-
mizing a complex user satisfaction function.
Indeed,
One approach for handling a rich objective function is
to treat its factors as multiple objectives.
there
are several methods in the literature on VRP for handling
multiple objectives. The most common approach is to ag-
gregate the objectives into a single weighted-sum objective
function [Molenbruch et al., 2017a], and advanced utility
model may be used for modeling the interactions between
the objectives [Lehu´ed´e et al., 2014]. Additional strate-
gies define hierarchical objective function [Schilde et al.,
2011], or return a set of non-comparable solutions which
do not weakly dominate each other [Parragh et al., 2009;
Molenbruch et al., 2017b]. Since our rich objective func-
tion models user satisfaction, we propose a different, human-
centric, approach. Specifically, we investigate machine learn-
ing methods for modeling the rich satisfaction function from
real humans.
sat outperforms optimal assignment methods that use a sim-
pler objective function, indicating that it is more important to
obtain a richer model of user satisfaction, than improving the
performance of the assignment algorithm.
2 Related Work
We will briefly review the current literature on the broad class
of Vehicle Routing and scheduling Problems (VRPs), to place
our ridesharing problem in an appropriate context. The VRP
was first introduced by [Dantzig and Ramser, 1959]. The
growing body of research on routing problem over the past 50
years has led to the development of several research commu-
nities, which sometimes denote the same problem types by
various names. In particular, the traditional VRP and some
of its extensions deal with finding an optimal set of routes
for a fleet of vehicles to traverse in order deliver or pickup
some goods to a given set of costumers. We refer to the com-
prehensive survey of [Parragh et al., 2008a] on this class of
problems, which they denote by Vehicle Routing Problems
with Backhauls (VRPB). A more recent survey, that also de-
fines a taxonomy to classify the various variants of VRP by
11 criteria, is given by [Psaraftis et al., 2016]. A second
class of problems, that is denoted by Parragh et al. as Vehi-
cle Routing Problems with Pickups and Deliveries (VRPPD),
deal with all those problems where goods are transported be-
tween pickup and delivery customers. We refer to the sur-
vey of [Parragh et al., 2008b] on this class of problems.
One subclass of VRPPD compromises the dial-a-ride prob-
lem (DARP), where the goods that are transported are pas-
sengers with associated pickup and delivery points. As noted
by [Cordeau and Laporte, 2003], the DARP is distinguished
from other problems in vehicle routing since transportation
cost and user inconvenience must be weighed against each
other in order to provide an appropriate solution. Therefore,
the DARP typically includes more quality constraints that aim
at capturing the user's inconvenience. We refer to a recent
survey on DARP by [Molenbruch et al., 2017a], which also
makes this distinction.
A domain closely related to ride-sharing is car-pooling. In
this domain, ordinary drivers, may opt to take an additional
passenger on their way to a shared destination. The common
setting of car-pooling is within a long-term commitment be-
tween people to travel together for a particular purpose, where
ridesharing is focused on single, non-recurring trips. Indeed,
several works investigated car-pooling that can be established
on a short-notice, and they refer to this problem as rideshar-
ing [Agatz et al., 2012]. We stress that in our ridesharing
problem, similar to the DARP setting, there is a central or-
ganization that owns the vehicles, and they thus do not have
their own travel plans.
3 Basic Model and Notation
Informally, the ridesharing problem consists of a weighted
graph, requests given by passengers, each with an origin and
a destination that are both nodes in the graph, and a set of
vehicles, each with a given capacity. All the vehicles are as-
sumed to be operated by a central entity. In this paper we
focus on the last mile variant [Cheng et al., 2014] setting.
In this variant it is assumed that all the passengers are po-
sitioned at the same origin location (e.g. airport), where all
the vehicles are also located, and must be taken to their fi-
nal destination. The problem requires assigning travel routes
(on the graph) to vehicles, in order to satisfy these passenger
requests while optimizing a given objective function. In our
work, we concentrate on the objective of maximizing the user
satisfaction function.
Let n be the number of passengers (and thus the number
of requests). We assume that the service provider incurs a
fixed cost per minute of travel, M, that encapsulates the full
operation cost including any desired revenue. For example,
if the fuel, tolls and any maintenance costs are estimated at x
dollars per minute of travel, and the service provider commits
to receiving only a certain percentage of the user payment (as
its revenue), ρ, M equals x
1−ρ. The service provider is free to
determine how to distribute this cost among the passengers,
and in section 3.2 we discuss the properties of this payment
function.
Every user, u ∈ U, is assumed to have a primary travel
time to(u) and distance do(u), which are the time and dis-
tance, respectively, it would take the user to reach her desti-
nation had she received a direct ride. Consequently, we define
for each user a primary travel cost co(u) = M · to(u). We
will also add a fixed constant to to(u) that represents waiting
time of the passenger had she received a direct ride. Given
an assignment, P , and a user u, the actual travel time of the
user is denoted by tP (u), the actual travel distance is denoted
by dP (u), the actual amount paid by the user is denoted by
cP (u), and the user satisfaction is denoted by sP (u).
satisfaction, i.e.,(cid:80)
(cid:80)
3.1 The Human Satisfaction Function
Our definition of the objective of the ridesharing problem is
to find an assignment P that maximizes the sum on all user
u∈U sP (u) (or, equivalently, the average
u∈U sP (u)). For simplicity, in the last mile
satisfaction 1
n
problem we assume that a passenger traveling alone has some
baseline satisfaction level ("neither satisfied nor disatisfied")
from the trip. Satisfaction factors may include any or all of
the following:
• Travel cost. cP (u).
• Actual travel time. tP (u). The travel time may also
depend on other parameters, for example, a user may
care more about travel time during the weekend.
• Extra travel time. tP (u) − to(u).
• Actual travel distance. dP (u).
• Extra travel distance. dP (u) − do(u)
• Total number of passengers. Users may rather travel
alone. The more passengers a ride may have the more
inconvenient it may become to each of the passengers.
• User's seat. In every vehicle with a constant capacity
some seats may be preferred over others. For example,
if a vehicle with 5 seats reaches its full capacity, most
people prefer siting on the front seat rather than the mid-
dle back seat.
• Working status / Occupation. Unemployed passengers
may be willing to travel longer in exchange for a lower
cost.
• User demographic information. Depending on the
user demographic group (e.g. age, gender, annual in-
come), users may care more about the other factors. For
example, young people may not mind traveling longer
if they save a few dollars, but people in their 40's may
be more concerned about their time. Some of this in-
formation may be extracted from an image of the user.
Note that this factor allows to define a more personal-
ized function, since different users might end up with
different satisfaction values for identical rides.
3.2 Payment Function for the Last Mile Problem
Given n passengers and their destinations, who travel in a sin-
gle vechile, an assignment P determining the drop-off path
(i.e. the order of which the passengers are dropped-off), the
time it takes to reach each destination under this assignment
(tP (u)), the total cost of the ride-shared trip, and the cost
(co(u)) and travel time (to(u)) that each of the passengers
would have encountered had they traveled alone, the pay-
ment function determines how much each passenger must pay
(CP (u)) when all passengers share the ride. We define the
following axioms on the payment function:
1. The aggregated payment from all passengers should ex-
actly cover the cost of the trip.
2. n passengers traveling to the same destination split the
trip cost equally.
3. Given two passengers with the same distance from the
source, the passenger who is dropped off second should
pay less than the passenger who was dropped off first.
4. Given a passenger that is "on the way" to another pas-
senger, both passengers should pay strictly less than
what they would pay had they traveled alone (even
though, both passengers do not travel any longer).
The following payment schedule satisfies all the axioms
above: Let α, β ∈ R+, we define the user's gain (given an
assignment P ) as:
gP (u) = (αto(u) + βco(u)) − (αtP (u) + βcP (u))
(1)
If we define the inconvenience of a user, iP (u), as αtP (u) +
βcP (u) (and similarly, io(u) is αto(u) + βco(u)), Equation
1 can be simplified to gP (u) = io(u) − iP (u). The payment
schedule sets cP (u) such that all passengers traveling in the
same ride have the same gP (u), and the sum of cP (u) equals
the cost of the whole ride. Equation 1 implies a simplified
view of human behavior with the single concept of "time is
money", and it further assumes a linear relation between the
two.
4 Satisfaction Model Learned from Humans
In order to develop a more realistic human satisfaction model,
we use machine learning techniques based upon data col-
lected from humans. To this end, we solicited 414 human
subjects from Mechanical Turk to obtain satisfaction level
data. Based on this data, we use deep learning to build a
satisfaction model.
4.1 Data Collection
The subjects were first asked to provide the following per-
sonal details: year of birth, gender and whether they were
employed or unemployed.
Our satisfaction model tries to predict the relative satisfac-
tion, that is, how much a passenger traveling by shared-ride
is more or less satisfied than the same passenger traveling in
a private ride. However, asking users to provide their rela-
tive satisfaction is unrealistic, and we thus split every travel
scenario into two sub-parts.
In the first part we asked the
subjects to determine their satisfaction level from a direct pri-
vate ride to some destination.
In the second part the sub-
jects were asked to determine their satisfaction level from a
shared ride to the same destination. Specifically, in the first
part of each scenario we described a direct private ride with
a given time (random number between 5 minutes and one
hour) and price (dollar per minute).
In the second part of
each scenario we described a shared ride to the same des-
tination, where we varied the travel time and cost. Travel
time of a shred ride can never be shorter than a direct pri-
vate ride, and we thus uniformly sampled a number from
{1, 1.1, 1.2, 1.3, 1.4, 1.5, 1.75, 2, 3, 4} and multiplied it by
the direct private travel time. The cost of a shared ride should
be lower than the cost of a direct private ride. In the optimal
sharing scenario, assuming a 4 passenger vehicle (excluding
the driver), there could be 4 passengers traveling to the same
destination; in this case the cost will be reduced by a factor of
4. We thus divided the direct private ride's cost by a number
uniformly sampled from {1.1, 1.2, 1.3, 1.4, 1.5, 1.75, 2, 3, 4}.
In addition, we randomly sampled the number of additional
passengers from {1, 2, 3}, and we randomly sampled the
user's seat from {front passenger, middle back, right back,
left back}. The subjects could choose one of seven satisfac-
tion levels on a Likert scale, between very satisfied (7) to very
dissatisfied (1) with the middle being 'neither satisfied nor
dissatisfied' (4).
Each subject was asked a total of six travel scenarios (each
with a private and a shared ride). In order to eliminate sub-
jects that may be selecting satisfaction levels at random, we
added two sanity check scenarios. In these scenarios the cost
of the shared ride was more expensive than the private ride.
Since it is unreasonable for a person to be more satisfied with
a shared ride, being both longer and more expensive, we have
disqualified any subject who expressed her satisfaction in this
question to be higher than her satisfaction from the private
ride of that scenario. 131 subjects failed one of these sanity
tests, and were removed from our analysis.
26 subjects refused to answer the personal questions and
were eliminated from our analysis as well. Of the remaining
257 subjects 147 were females and 110 were males. Their age
ranged from 19 to 67 with an average of 32.3 and a median
of 31. 195 were employed while 62 were unemployed. Each
of the 257 subjects had 4 scenarios, resulting with 1028 data-
points. The average satisfaction for a private trip was 5.01
and a shared ride was 4.41.
4.2 Deep-Learning Based Model
Using the collected data we consider deep learning based
models with a varying depth to find a good satisfaction model,
Model
Train-error Validation error
Linear regression
1 hidden layer
2 hidden layers
3 hidden layers
1.680
1.621
1.592
1.480
1.734
1.711
1.696
1.785
Table 1: Train and validation error for the different model depths.
Values indicate the root mean squared error (RMSE) of each model.
that is, a model that will accurately predict user satisfaction
levels of a new user, based on different features of the user
and the user trip. We split the data into train, validation (dev)
and test sets. We use mean squared error to measure the per-
formance of each model. The neural network depth varied
from 1 (linear regression) to 4 (3 hidden layers). Each hid-
den layer consisted of 100 neurons. We used early stopping
[Prechelt, 1998], i.e., we used the validation set to determine
when to stop training. Table 1 presents the results obtained by
each of these models. Since the 2-hidden layers model per-
formed best, we use it as our satisfaction function, this model
achieved a root mean squared error of 1.75 on the test set.
We set the satisfaction of a private ride to 'neither satisfied
nor dissatisfied' (4).
5 Experimental Evaluation
In this section we present a stochastic algorithm for the last
mile variant. We assume that there are a sufficiently large
number of vehicles so that any request could be satisfied, and
that the capacity of each vehicle is 4 passengers. We compare
the performance of our algorithm, in terms of user satisfac-
tion, to an optimal algorithm that uses a simpler satisfaction
function. We show that the algorithm outperforms the optimal
algorithm, and that emphasizes the importance of capturing a
rich model of human satisfaction.
5.1 Stochastic-Based Satisfaction Algorithm
(Simsat)
We now present, a practical algorithm for assigning passen-
gers to vehicles with the objective of maximizing the sum on
user satisfaction: Simsat (Algorithm 1). Simsat runs Floyd-
Warshell on the graph at the initialization, to obtain the mini-
mal travel time between every two vertices. Simsat then runs
its main procedure as follows. Simsat shuffles all passengers
and assigns every passenger to the vehicle that maximizes the
current satisfaction sum. For computing the total satisfac-
tion of a single vehicle (the SatFunc function in Algorithm 1),
we use the nearest neighbor algorithm (the greedy approach)
for ordering the passengers drop-offs (based upon the Floyd-
Warsell matrix). The main procedure is repeated multiple
times (n2) and the assignment that yields the maximal total
satisfaction is selected. The complexity of Simsat is clearly
O(n4). The number of times the main procedure is repeated
can vary; the more times it is repeated the higher the expected
performance. Therefore, Simsat is an any-time algorithm.
5.2 Optimal Algorithm
We use the following method to obtain the optimal assign-
ment. First, the algorithm runs Floyd-Warshell on the graph.
Algorithm 1: The Simsat algorithm
Input: A graph (Graph), with source vertex
Passenger destinations list (Passengers),
A satisfaction function that returns the total satisfaction
of all passengers in a vehicle (SatFunc),
Result: An assignment of all passengers to vehicles.
Compute Floyd-Warshell on Graph;
MaxSum := 0;
for i:=0 to n2 do
Shuffle Passengers;
SatSum := 0;
Clear CabList;
for CurrentPassenger : Passengers do
MAX := -1;
for CurrentCab : CabList do
if CurrentCab not full then
CurrentSat := SatFunc(CurrentCab);
Add CurrentPassenger to CurrentCab;
SatWithCurPass = SatFunc(CurrentCab);
Remove CurrentPassenger from
CurrentCab;
if (SatWithCurPass - CurrentSat) larger
than MAX then
OptimalCab := CurrentCab;
MAX := SatWithCurPass -
CurrentSat
end
end
end
if MAX < 0 then
Add CurrentPassenger to newCab;
Add newCab to CabList;
Add CurrentPassenger to OptimalCab;
SatSum += MAX;
end
if (SatSum > MaxSum) then
MaxSum := SatSum;
OptimalFullAssignment := CabList;
else
end
end
end
Figure 1: A graph created from a map of the city of Toulouse,
France.
4
The algorithm then solves a coin-change problem ([Harris et
al., 2008, p. 171]) to obtain all possible ways to split the num-
ber of passengers into vehicles. For example, when n = 10,
we get {3, 3, 3, 1}, {2, 2, 2, 2, 2}, {4, 4, 2} etc. For each split-
ting option the algorithm iterates over all possible assign-
ments (we explicitly handle multiple vehicles with the same
number of passengers, since it does not matter which group
of passengers travels in which vehicle if they are in the same
size). For example, for a group of {4, 3, 3, 2, 2, 2}, it first iter-
ates over all assignments of 4 passengers (there are(cid:0)n
(cid:1) such
assignments), then, recursively calls the assignment function
with {3, 3, 2, 2, 2} and the remaining passengers. The recur-
sive call iterates over all possible assignments of three people
to two vehicles and preforms a recursive call with the remain-
ing vehicles and passengers. For each vehicle, the algorithm
computes all possible options for dropping off its passengers
(this is done once for each set of users), and, based upon the
Floyd-Warshell matrix and the satisfaction model, selects the
most efficient travel order.
5.3 Data
We considered two different types of graphs, a randomly gen-
erated graph and a more realistic graph, the city of Toulouse,
France1. The random graph was created by placing 35 ver-
tices uniformly on the plane. We then randomly chose a pair
of vertices, and connected them with an edge with a probabil-
ity that is proportional to their distance. The weight of each
edge was determined by the air-distance multiplied by a ran-
dom number (uniformly sampled) between 1 and 2, to model
topological variance. The graph of the city of Toulouse is
presented in Figure 1. This graph includes the actual dis-
tances between the different vertices. We cropped the graph
to 40, 000 vertices, by running Dijkstra algorithm starting at
the airport, sorting all vertices by their distance from the air-
port, and removing all farther away vertices (including those
that are unreachable).
1obtained from https://www.geofabrik.de/data/
shapefiles_toulouse.zip
Figure 2: Average satisfaction for 6, 7, 8, 9 and 10 passengers when
using a random graph.
Figure 3: Average satisfaction for 6, 7, 8, 9 and 10 passengers when
using a map of the city of Toulouse, France.
5.4 Experimental Settings
Being a last mile problem, we set the origin vertex to be the
same for all passengers. That is, in the random graph we ran-
domly generated an origin vertex. In the city of Toulouse the
graph includes the Toulouse-Blagnac airport, and it was used
as the origin vertex. The destination vertices were randomly
sampled for every passenger using a uniform distribution over
all vertices.
In the payment schedule we set α to 0.3 and β to 1 for
the gain function. α and β were set according to the average
U.S. wage. That is, average annual hours worked per worker
in U.S. at 2016 was 1783, and the average annual income in
the U.S. per worker in 2016 was $31,099. Dividing the two
we get $17.5 per hour, or approximately $0.3 per minute. We
set the average speed to 60 kph, and the cost per km travel
distance to $1. We tested 5 assignment algorithms:
1. The optimal algorithm with the full satisfaction function
(developed in section 4).
2. Simsat with the full satisfaction function.
3. The optimal algorithm with simpler satisfaction func-
tions:
(a) Travel cost only.
(b) Travel time only.
(c) Time and cost according to the payment function
(that is, the gain function is used as a substitute for
user satisfaction).
All the algorithms were evaluated with the complete satisfac-
tion function developed in section 4, regardless of the func-
tion actually used by the assignment algorithm.
5.5 Results
Figure 2 presents the average satisfaction for 6,7,8,9 and 10
passengers when using a random graph, and Figure 3 presents
the results for the city of Toulouse. The results were obtained
by averaging over 100 samples of passenger destinations. The
results for the optimal method using travel time only were
omitted, as it constantly yields an average user satisfaction of
4 (since it assigns a private vehicle to each and every passen-
ger). As depicted in both figures, our satisfaction oriented as-
signment method (Simsat) obtains results that are quite close
to the optimal assignment. Simsat's average satisfaction level
is much closer to the optimal assignment than that of the op-
timal assignments using a simpler user-satisfaction model.
These results indicate that it is more important to obtain a
richer model of user satisfaction, than improving the perfor-
mance of the assignment algorithm. That being said, we do
not disregard the importance of improving the performance of
the assignment algorithm and do intend to pursue additional
algorithms that may perform better.
6 Conclusions & Future Work
Ridesharing has a true potential for improving the qual-
ity of life for many people [Cici et al., 2013], and it is
part of the general concept of sharing economy that is be-
ing evolved nowadays. However, despite both Uber and
Lyft offering ride-sharing options, not many users elect to
share their rides with additional passengers [Koebler, 2016;
RSG, 2017]. Following the statement by Carnegie [Carnegie,
1936, p. 37], "There is only one way to get anybody to do any-
thing. And that is by making the other person want to do it.",
we believe that the key ingredient required for a widespread
adaptation of ridesharing is to focus on user satisfaction.
The importance of the paper lies in its being the first to ex-
clusively concentrate on a rich and realistic function of user
satisfaction as the objective, which is (arguably) the most im-
portant aspect to consider for achieving a widespread adap-
tion of ridesharing services. We use deep learning to model
user satisfaction based upon data collected from actual hu-
man subjects. We present a satisfaction oriented assignment
method (Simsat), and show that it outperforms optimal as-
signments using a simpler user-satisfaction model. These re-
sults indicate that it is more important to obtain a richer model
of user satisfaction, than improving the performance of the
assignment algorithm.
In future work we intend to extend our model to the more
general ridesharing schenario, where people may have differ-
ent origins. We also intend to build a game that will simulate
an actual ride for the subjects; this should allow us to obtain
more exact satisfaction levels. This game could include addi-
tional travel information such as the other passengers in the
trip, and allow the subject to select her seat when entering
a vehicle. Since users will be playing the game more than
3.544.555.566.56 7 8 9 10 Average User-Satisfaction SimsatOptimalGainCost3.544.555.566.56 7 8 9 10 Average User-Satisfaction SimsatOptimalGainCost once, the satisfaction model can be further improved by per-
sonalization, taking into account user's feedback on previous
rounds.
References
[Agatz et al., 2012] Niels Agatz, Alan Erera, Martin Savelsbergh,
and Xing Wang. Optimization for dynamic ride-sharing: A re-
view. European Journal of Operational Research, 223(2):295 --
303, 2012.
[Campbell and Savelsbergh, 2005] Ann Melissa Campbell and
Martin WP Savelsbergh. Decision support for consumer di-
rect grocery initiatives. Transportation Science, 39(3):313 -- 327,
2005.
[Carnegie, 1936] Dale Carnegie. How to win friends and influence
people. Simon and Schuster, 1936.
[Cheng et al., 2014] Shih-Fen Cheng, Duc Thien Nguyen, and
Hoong Chuin Lau. Mechanisms for arranging ride sharing and
In AAMAS, pages
fare splitting for last-mile travel demands.
1505 -- 1506, 2014.
[Cici et al., 2013] Blerim Cici, Athina Markopoulou, Enrique
Fr´ıas-Mart´ınez, and Nikolaos Laoutaris. Quantifying the poten-
tial of ride-sharing using call description records. In Proceedings
of the 14th Workshop on Mobile Computing Systems and Appli-
cations, page 17, 2013.
[Cordeau and Laporte, 2003] Jean-Franc¸ois Cordeau and Gilbert
Laporte. A tabu search heuristic for the static multi-vehicle dial-
a-ride problem. Transportation Research Part B: Methodologi-
cal, 37(6):579 -- 594, 2003.
[Dantzig and Ramser, 1959] G. B. Dantzig and J. H. Ramser. The
truck dispatching problem. Management Science, 6(1):80 -- 91,
1959.
[Diana et al., 2006] Marco Diana, Maged M Dessouky, and Nan
Xia. A model for the fleet sizing of demand responsive trans-
portation services with time windows. Transportation Research
Part B: Methodological, 40(8):651 -- 666, 2006.
[El-Sherbeny, 2010] Nasser A El-Sherbeny. Vehicle routing with
time windows: An overview of exact, heuristic and metaheuristic
methods. Journal of King Saud University-Science, 22(3):123 --
131, 2010.
[Fu and Teply, 1999] Liping Fu and Stan Teply. On-line and
off-line routing and scheduling of dial-a-ride paratransit vehi-
cles. Computer-Aided Civil and Infrastructure Engineering,
14(5):309 -- 319, 1999.
[Guerra, 2016] Erick Guerra. Planning for cars that drive them-
selves: Metropolitan planning organizations, regional transporta-
tion plans, and autonomous vehicles. Journal of Planning Edu-
cation and Research, 36(2):210 -- 224, 2016.
[Harris et al., 2008] John Michael Harris, Jeffry L Hirst, and
Michael J Mossinghoff. Combinatorics and graph theory, vol-
ume 2. Springer, 2008.
[Jaw et al., 1986] Jang-Jei Jaw, Amedeo R Odoni, Harilaos N
Psaraftis, and Nigel HM Wilson. A heuristic algorithm for
the multi-vehicle advance request dial-a-ride problem with time
Transportation Research Part B: Methodological,
windows.
20(3):243 -- 257, 1986.
[Koebler, 2016] Jason Koebler. Why everyone hates uberpool?,
2016.
[Lehu´ed´e et al., 2014] Fabien Lehu´ed´e, Renaud Masson, Sophie N
Parragh, Olivier P´eton, and Fabien Tricoire. A multi-criteria large
neighbourhood search for the transportation of disabled people.
Journal of the Operational Research Society, 65(7):983 -- 1000,
2014.
[Lin et al., 2012] Yeqian Lin, Wenquan Li, Feng Qiu, and He Xu.
Research on optimization of vehicle routing problem for ride-
sharing taxi. Procedia-Social and Behavioral Sciences, 43:494 --
502, 2012.
[Molenbruch et al., 2017a] Yves Molenbruch, Kris Braekers, and
An Caris. Typology and literature review for dial-a-ride prob-
lems. Annals of Operations Research, 2017.
[Molenbruch et al., 2017b] Yves Molenbruch, Kris Braekers,
An Caris, and Greet Vanden Berghe. Multi-directional lo-
cal search for a bi-objective dial-a-ride problem in patient
transportation. Computers & Operations Research, 77:58 -- 71,
2017.
[Paquette et al., 2009] Julie Paquette, Jean-Franc¸ois Cordeau, and
Gilbert Laporte. Quality of service in dial-a-ride operations.
Computers & Industrial Engineering, 56(4):1721 -- 1734, 2009.
[Parragh et al., 2008a] Sophie N. Parragh, Karl F. Doerner, and
Richard F. Hartl. A survey on pickup and delivery problems.
part I: Transportation between customers and depot. Journal fur
Betriebswirtschaft, 58(1):21 -- 51, 2008.
[Parragh et al., 2008b] Sophie N. Parragh, Karl F. Doerner, and
Richard F. Hartl. A survey on pickup and delivery problems. part
II: Transportation between pickup and delivery locations. Jour-
nal fur Betriebswirtschaft, 58(1):81 -- 117, 2008.
[Parragh et al., 2009] Sophie N Parragh, Karl F Doerner, Richard F
Hartl, and Xavier Gandibleux. A heuristic two-phase solution
approach for the multi-objective dial-a-ride problem. Networks,
54(4):227 -- 242, 2009.
[Parragh et al., 2014] Sophie N Parragh, Jorge Pinho de Sousa, and
Bernardo Almada-Lobo. The dial-a-ride problem with split re-
quests and profits. Transportation Science, 49(2):311 -- 334, 2014.
[Prechelt, 1998] Lutz Prechelt. Automatic early stopping using
cross validation: quantifying the criteria. Neural Networks,
11(4):761 -- 767, 1998.
[Psaraftis et al., 2016] Harilaos N Psaraftis, Min Wen, and Chris-
tos A Kontovas. Dynamic vehicle routing problems: Three
decades and counting. Networks, 67(1):3 -- 31, 2016.
[Psaraftis, 1980] Harilaos N Psaraftis. A dynamic programming
solution to the single vehicle many-to-many immediate request
Transportation Science, 14(2):130 -- 154,
dial-a-ride problem.
1980.
[RSG, 2017] Rsg 2017 survey results: Driver earnings, satisfaction
and demographics, 2017.
[Santos et al., 2011] Adella
Nancy
McGuckin,
Hikari Yukiko Nakamoto, Danielle Gray, and Susan Liss.
Summary of travel
2009 national household travel
survey. Technical report, 2011.
Santos,
trends:
[Schilde et al., 2011] Michael Schilde, Karl F Doerner,
and
Richard F Hartl. Metaheuristics for the dynamic stochastic dial-
a-ride problem with expected return transports. Computers &
operations research, 38(12):1719 -- 1730, 2011.
[Secomandi and Margot, 2009] Nicola Secomandi and Franc¸ois
Margot. Reoptimization approaches for the vehicle-routing prob-
lem with stochastic demands. Operations research, 57(1):214 --
230, 2009.
[Secomandi, 2000] Nicola Secomandi. Comparing neuro-dynamic
programming algorithms for the vehicle routing problem with
Computers & Operations Research,
stochastic demands.
27(11):1201 -- 1225, 2000.
[Shen et al., 2016] Bilong Shen, Yan Huang, and Ying Zhao. Dy-
namic ridesharing. SIGSPATIAL Special, 7(3):3 -- 10, 2016.
[Wallsten, 2015] Scott Wallsten. The competitive effects of the
sharing economy: how is uber changing taxis? Technology Policy
Institute, 22, 2015.
[Yang et al., 2013] Jianjie Yang, Jin Li, Yun Chen, Xintong Liu,
et al. Multi-objective distribution model and algorithm for on-
line shopping express logistics. JCP, 8(10):2558 -- 2564, 2013.
|
1904.12066 | 1 | 1904 | 2019-04-26T22:41:03 | ABIDES: Towards High-Fidelity Market Simulation for AI Research | [
"cs.MA"
] | We introduce ABIDES, an Agent-Based Interactive Discrete Event Simulation environment. ABIDES is designed from the ground up to support AI agent research in market applications. While simulations are certainly available within trading firms for their own internal use, there are no broadly available high-fidelity market simulation environments. We hope that the availability of such a platform will facilitate AI research in this important area. ABIDES currently enables the simulation of tens of thousands of trading agents interacting with an exchange agent to facilitate transactions. It supports configurable pairwise network latencies between each individual agent as well as the exchange. Our simulator's message-based design is modeled after NASDAQ's published equity trading protocols ITCH and OUCH. We introduce the design of the simulator and illustrate its use and configuration with sample code, validating the environment with example trading scenarios. The utility of ABIDES is illustrated through experiments to develop a market impact model. We close with discussion of future experimental problems it can be used to explore, such as the development of ML-based trading algorithms. | cs.MA | cs |
ABIDES: TOWARDS HIGH-FIDELITY MARKET SIMULATION
FOR AI RESEARCH
A PREPRINT
David Byrd
Maria Hybinette
School of Interactive Computing
Georgia Institute of Technology
Department of Computer Science
The University of Georgia
Atlanta, GA 30308
[email protected]
Athens, GA 30303
[email protected]
Tucker Hybinette Balch
School of Interactive Computing
Georgia Institute of Technology
Atlanta, GA 30308
[email protected]
ABSTRACT
We introduce ABIDES, an Agent-Based Interactive Discrete Event Simulation environment. ABIDES
is designed from the ground up to support AI agent research in market applications. While simulations
are certainly available within trading firms for their own internal use, there are no broadly available
high-fidelity market simulation environments. We hope that the availability of such a platform will
facilitate AI research in this important area. ABIDES currently enables the simulation of tens of
thousands of trading agents interacting with an exchange agent to facilitate transactions. It supports
configurable pairwise network latencies between each individual agent as well as the exchange. Our
simulator's message-based design is modeled after NASDAQ's published equity trading protocols
ITCH and OUCH. We introduce the design of the simulator and illustrate its use and configuration
with sample code, validating the environment with example trading scenarios. The utility of ABIDES
is illustrated through experiments to develop a market impact model. We close with discussion of
future experimental problems it can be used to explore, such as the development of ML-based trading
algorithms.
1 Background
We have developed ABIDES, an agent-based interactive discrete event simulation, to facilitate the creation, deployment,
and study of strategic agents in a highly configurable market environment. We were inspired by Daniel Freidman's view
that simulation provides a powerful tool to analyze individual participant behavior as well as overall market outcomes
that emerge from the interaction of the individual agents. In Freidman's review of empirical approaches to the analysis
of continuous double auction (CDA) markets such as NASDAQ and the New York Stock Exchange, he outlines the
strengths and weaknesses of three major approaches:
1. Field studies of actual operating markets,
2. Laboratory studies of small controlled markets,
3. Computer simulation of markets.
Freidman concludes that field studies are clearly relevant, but do not provide experimental access to all relevant
information; laboratory studies improve control and observation, but are of necessity small and expensive; and computer
simulations feature perfect control and observation. However "trader's strategies are not endogenously chosen, but
rather must be specified exogenously" [2].
As Freidman observed, market simulations provide an attractive platform for research in equity trading questions. This
has led to the development of a number of simulation platforms such as those on which X. Wang and Wellman [9]
and J. Wang et al. [8] have reported their results. We developed ABIDES as a fresh implementation to incorporate
lessons learned from the deployment of prior platforms. With ABIDES, we aim to address Freidman's primary concern
regarding computerized market simulations -- that strategies must be exogenously specified -- with a platform enabling
powerful learning agents to easily participate in a realistically structured market via a common framework. We believe
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
this is necessary to properly investigate the behavior and impact of intelligent agents interacting in a complex market
environment.
ABIDES is intended to be a curated, collaborative open-source project that provides researchers with tools that support
the rapid prototyping and evaluation of complex market agents. With it, we hope to to further empower researchers
of financial markets to undertake studies which would be difficult or impossible in the field, due to the absence of
fine-grained data identifiable to individual traders (see Figure 1), a lack of knowledge concerning participant motivation,
and an inability to run controlled "what if" studies against particular historical dates.
Figure 1: Simulation allows agent-identifiable data which is lost in the flow of real-world orders.
We acknowledge existing high-quality academically-targeted multi-agent market simulators such as that used by
X. Wang and Wellman. In a recent study they used their simulation platform to study spoofing agents in a market
environment populated by zero intelligence (ZI) and heuristic belief learning (HBL) traders [9]. Their approach analyzes
the results from an empirical game-theoretic view [10]. We believe ABIDES makes a complementary contribution
through its experimental focus on the "market physics" of the real world including:
• Support for continuous double-auction trading at the same nanosecond time resolution as real markets such as
NASDAQ;
• Ability to simulate specific dates in market history with gated access to historical data;
• Variable electronic network latency and agent computation delays;
• Requirement that all agents intercommunicate solely by means of standardized message protocols;
• Easy implementation of complex agents through a full-featured hierarchy of base agent classes.
The focus on these features should enable an expanded range of experimental studies. We believe ABIDES is also the
first full-featured, modern market simulator to be shared with the community as an open source project.
2
Important Questions Simulation Can Help Us Address
ABIDES can support a number of different kinds of investigations into market behavior that are not easily conducted
using historical data or live experiments.
• The benefits of co-location: In the past 20 years hedge funds and other market participants have invested
in the deployment of computing resources co-located at major exchanges [11]. This so-called "co-location"
enables quicker access to market information than if the trading server were located further away. It is not
feasible to investigate the value of the advantage co-location provides with available historical data, because it
does not include information about the geographic location, network latency, or network reliability of each
actor. With a platform that does not require formal arms-length messaging using a realistic network model,
we cannot simulate the effects of these factors even if they are known. ABIDES provides a network model
and mandatory messaging protocol that enables detailed experiments in this area: Creating a population of
2
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
agents with known distribution of network latency and reliability, conducting trials in which one agent is
incrementally shifted from a co-location facility out to a great distance, and evaluating the impact of this shift
on each agent's profitability while otherwise pursuing the same strategies.
• The impact of large orders on price: The very act of trading, and even placing orders in a market may affect
the price. For instance, if there is significant selling pressure evidenced by a large volume of sell orders, it is
generally expected that the price will go down. The extent to which the price moves because of an order is
referred to as market impact. Market participants of course want to minimize such impact, because the market
usually moves contrary to their profit incentives. In a market field study, it is not feasible to perform controlled
A/B tests. One cannot place a market buy at the NYSE for one million shares of IBM at 10 AM on Oct 22,
2018, and then also not place that order, and compare the difference. Without the "control", any observed
result from the large order could be attributable to some other factor. A key feature of ABIDES is the ability
to re-simulate the same historical market day with known, limited changes while holding all other factors
constant, thus enabling the desired experimental control population.
• Cost-benefit analysis of AI: When analyzing historical market data, we cannot know the logic behind
individual trader actions. In a simulation without a model for computational time delays that directly impact
time-to-market for the resulting orders, we cannot readily study the trade-off between simpler, faster predictors
and slower, more powerful predictors. ABIDES introduces a flexible, integrated model for computation delay
that permits the "speed" of each agent's thought process to be represented, and to have that representation
affect the timing of all of outbound messages as well as the next time at which the agent can be roused for
participation. Thus heavier thinkers will take longer to deliver a resulting order to the exchange and will be
unable to act as frequently.
• Explanation of learning agent behavior: A current key area of AI research across all application fields is
explainability -- once taken for granted in classic knowledge-based AI, but now increasingly difficult with
"black box" ML algorithms. By providing a platform with centralized, time-synced event logging for all agents,
we envision a "clear box" in which each agent's decision, intent, behavior, and result for every action are fully
visible. We hope to use this ability to dive deeply into the why of learned policy functions.
3 ABIDES Architecture
The ABIDES framework includes a customizable configuration system, a simulation kernel, and a rich hierarchy of
agent functionality as illustrated in Figure 2.
Figure 2: Class relations within the ABIDES simulation framework.
3
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
3.1 Functions and Features of the ABIDES Kernel
The simulation is built around a discrete event-based kernel [1]. The kernel resides in the Kernel class in the default
package and is required in all simulations. All agent messages must pass through the kernel's event queue. The
kernel supports simulation of geography, computation time, and network latency. It also acts as enforcer of simulation
"physics", maintaining the current simulation time, tracking a separate "current time" for each agent, and ensuring there
is no inappropriate time travel. Key features of the ABIDES kernel include:
• Historical date simulation
All simulation occurs on a configurable historical date or sequence of dates.
This permits "real" historical information to be seamlessly injected into the simulation at appropriate times
when required for a particular experiment.
• Nanosecond resolution: Because we seek to emulate real markets, we simulate time at the same resolution as
an example exchange: the NASDAQ exchange. All simulation times are represented as Pandas Timestamp
objects with nanosecond resolution. This allows a mixture of agents to participate in the simulation on
very different time scales with minimal developer overhead. Events that occur simultaneously (in the same
nanosecond) will be executed in arbitrary order.
• Global Virtual Time (GVT): GVT is the latest simulated time for which all messages are guaranteed to have
been processed. The kernel tracks GVT as the simulation progresses. It is usually the case that GVT advances
much more quickly than wall clock time, but for very complex scenarios, it may not. The GVT value is not
available to the agents.
• Current time per agent: The kernel tracks a "current time" per individual participating agent which is
incremented upon return from any call to Agent.receiveMessage() or Agent.wakeup(). In situations
where the current time for the agent is "in the future" (i.e., larger than GVT), the kernel will delay delivery of
messages or wakeup calls to this agent until GVT catches up.
• Computation delay: The kernel stores a computation delay per agent which is added to the agent's "current
time" after each activity. The delay is also added to the sent time and delivery time of any outbound message
from an agent to account for the agent's computation effort. Agents may alter this computation delay to
account for different sorts of computation events.
• Configurable network latency: The kernel maintains a pairwise agent latency matrix and a latency noise
model which are applied to all messages between agents. This permits simulation of network conditions and
agent location, including co-location.
• Deterministic but random execution: The kernel accepts a single pseudo-random number generator (PRNG)
seed at initialization. This PRNG is then used to generate seeds for an individual PRNG object per agent,
which must rely solely on that object for stochastic methods. Since our system is single-threaded, this allows
the entire simulation to be guaranteed identical when the same seed is initialized within the same experimental
configuration. This would not ordinarily permit the desired A/B testing, because the "agent of change" might
consume an additional pseudo-random number from the sequence and thus change the stochastic source for
all subsequent agents. Because of our careful use of the primary PRNG only to generate subsidiary PRNGs
per agent, the "agent of change" in an ABIDES A/B experiment will not alter the set of pseudo-random
numbers given to any other agent throughout the simulation, even if it uses more or fewer such inputs for its
changed activity. In this way, changes in the behavior of other agents will be caused by a changed simulation
environment (e.g. stock prices) and not simple stochastic perturbation.
3.2 ABIDES Kernel Lifecycle Phases
During a simulation, the kernel follows a series of life cycle phases. All except the event queue processing phase consist
entirely of sending the relevant event notification to all agents, and are described in the Agent subsection below. The
event queue processing phase is elaborated upon here:
1. Kernel Initializing
2. Kernel Starting
3. Repeat until the event queue is empty or currentTime > stopTime :
− Extract next scheduled event and set currentTime = event.deliveryTime
− If agentTimes[event.target] > currentTime:
· event.deliveryTime = agentTimes[event.target]
· Place event back in queue and goto 3
4
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
− agentTimes[event.target] = event.deliveryTime
− Call target.wakeup() or target.receiveMessage()
− agentTimes[event.target] += computationDelay[event.target]
4. Kernel Stopping
5. Kernel Terminating
The kernel additionally supports a few critical methods upon which agents depend:
• sendMessage(sender, recipient, message, delay) - Schedules message to be transmitted from
sender to recipient with an (optional) non-negative additional delay. The "sent time" will be the sender's
current time, plus its computation delay, plus any requested extra delay. The "delivery time" will be the sent
time plus network latency plus jitter, as determined by configured parameters for the experiment.
• setWakeup(sender, requestedTime) - Schedules a wakeup call for the sender at the requested future
time.
• findAgentByType(type) - Returns the numeric identifier of an agent of the requested type if one can be
found. If multiple agents of the type exist, one is selected arbitrarily. It is not possible for an agent to obtain a
reference to another agent (and thus bypass the kernel in the future).
• writeLog(sender, dfLog) - Called by an agent to request that its log be archived to disk for analysis. The
log is expected to be a Pandas DataFrame with index type DatetimeIndex.
3.3 The Agent Class
All simulator agents are defined in the agent package. All participants in a simulation must inherit from the base
agent.Agent class, which implements a number of required methods that allow basic participation in the full life
cycle of the simulation.
The following methods must be supported by all simulation agents and will be called exactly one time per agent by the
kernel. The order in which agents are activated in each life cycle phase is arbitrary.
• kernelInitializing(kernel) - The kernel has just started running. The existence of other agents should
not be assumed. There is no "current time". The base Agent simply retains the given kernel reference.
• kernelStarting(startTime) - Event queue processing is about to begin. All other agents are now
guaranteed to exist. There is no "current time". startTime contains what will be the initial simulation
timestamp. The base Agent requests a wakeup call for this initial timestamp.
• kernelStopping() - Event queue processing has just ended. All other agents are still guaranteed to exist.
There is no longer a "current time". The base Agent takes no action.
• kernelTerminating() - The kernel is about to shut down. The existence of other agents should not be
assumed. There is no longer a "current time". Agents are expected to log any final data and clean up. The base
Agent passes off its individual event log, if there are entries, to the kernel for archival.
The following methods must be supported by all simulation agents. They will be called by the kernel in order of
increasing delivery timestamp of queued messages and wakeup calls. In both cases, the base Agent simply updates its
internal current time and displays an informative message.
• receiveMessage(currentTime, msg) - The kernel
is delivering a message from another agent.
currentTime is the current simulation time as a Pandas Timestamp (nanosecond resolution). msg is an
instance of class message.Message which the agent must interpret.
• wakeup(currentTime) - The kernel is delivering a previously-scheduled "wakeup call" to the agent.
currentTime is the current simulation time. No message is delivered, thus the agent must use internal
state and logic to determine what it should do next.
While not required by the simulation kernel, the base Agent class also provides logEvent(eventType, event),
which can be called by any agent to append to an individual timestamped log of events. As noted above, by default this
log is reported to the kernel for archival during the kernelTerminating life cycle phase.
5
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
3.4 The Exchange Agent Class
The agent.ExchangeAgent class inherits from agent.Agent and represents a stock exchange such as NASDAQ.
The message protocols supported by this agent are based on NASDAQ's published ITCH and OUCH protocols. [5, 6]
The exchange is initialized with market opening and closing times, which it will enforce. These are not required to
match the simulation start and stop times. The exchange agent is not privileged in any way; it must participate in the
simulation just as any other agent. The ExchangeAgent understands how to respond to these types of messages:
• Market Open Time: Returns the timestamp at which the exchange will begin processing order-related
messages.
• Market Close Time: Returns the timestamp at which the exchange will stop processing order-related mes-
sages.
• Query Last Trade: Returns the last trade price for a requested symbol. Until the first trade of the day, the
exchange reports the oracle open price (historical or generated data) as the "last trade price". The exchange
does not yet implement the opening cross auction.
• Query Spread / Depth: Returns a list of the N best bid and best ask prices for a requested symbol and the
aggregate volume available at each price point. With a requested depth of one, this is equivalent to querying
"the spread".
• Limit Order: Forwards the attached limit order to the requested symbol's order book for matching or
acceptance. Agents currently simulate market orders using a limit order with an arbitrarily high or low limit
price.
• Cancel Order: Forwards the attached order to the requested symbol's order book to attempt cancellation.
Outside of market hours, the exchange will only honor messages relating to market hour inquiries and final trade
prices (after the close). The exchange sends a "market closed" message to any agent which contacts it with disallowed
messages outside of market hours.
The exchange agent demonstrates one use of the inbuilt Kernel logging facility, recording either the full order stream or
snapshots of its order books at a requested frequency, enabling extremely detailed visualization and analysis of the
order book at any time during simulation. For example, Figure 3 shows a time window surrounding one "high impact"
market buy order, which drives prices upward immediately and has a follow-on effect on other agents' value beliefs.
3.5 The Order Book
Within an Exchange Agent, an order book tracks all open orders, plus the last trade price, for a single stock symbol. All
order book activity is logged through the exchange agent. The order book implements the following functionality:
• Order Matching Attempts to match the incoming order against the appropriate side of the order book. The
best price match is selected. In the case of multiple orders at the same price, the oldest order is selected.
• Partial Execution Either the incoming order or the matched limit order may be partially executed. When the
matched limit order is partially executed, the order is left in the book with its quantity reduced. When the
incoming order is partially executed, its quantity is reduced and a new round of matching begins. Participants
receive one "order executed" message, sent via the exchange, per partial execution noting the fill price of each.
When the incoming order is executed in multiple parts, the average price per share is recorded as the last trade
price for the symbol.
• Order Acceptance When the incoming limit order has remaining quantity after all possible matches have
been executed, it will be added to the order book for later fulfillment, and an "order accepted" message will be
sent via the exchange.
• Order Cancellation The order book locates the requested order by unique order id, removes any remaining
unfilled quantity from the order book, and sends an "order cancelled" message via the exchange.
One might reasonably expect the order book in a market simulation to include a model for slippage. We assert that our
platform produces realistic slippage naturally, without the need for such a model. Orders directed to the exchange suffer
dynamic computation and network delays, during which time other orders are being executed.
6
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
Figure 3: Example of order book visualization around the time of a high impact trade.
3.6 The Trading Agent Class
The agent.TradingAgent class inherits from agent.Agent and represents the base class for a financial trading agent.
It implements a number of additional features beyond the basic simulator Agent, upon which subclassed strategy agents
may rely:
• Portfolio The base trading agent maintains an equity portfolio including a cash position. It automatically
updates this portfolio in response to "order executed" messages.
• Open Orders The trading agent keeps a list of unfilled orders that is automatically updated upon receipt of
"order executed" and "order cancelled" messages, and when new orders are originated.
• Last Known Symbol Info The trading agent tracks known information about all symbols in its awareness,
including the most recent trade prices, daily close prices (after the close), and order book spread or depth.
These are automatically updated when receiving related messages.
• Market Status Upon initially waking at simulation start, the trading agent automatically locates an exchange
agent, requests market open and close times, and schedules a second wakeup call for the time of market open.
It also maintains and provides a simple "market closed" flag for the benefit of subclassing agents.
• Mark to Market The trading agent understands how to mark its portfolio to market at any time, using its
most current knowledge of equity pricing. It automatically marks to market at the end of the day.
• Messages The trading agent knows how to originate all of the messages the exchange understands, and to
usefully interpret and store all of the possible responses from the exchange.
• Logging The trading agent logs all significant activity: when it places orders; receives notification of order
acceptance, execution, or cancellation; when its holdings change for any reason; or when it marks to market at
the end of the day.
7
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
4 ABIDES Implementation
The ABIDES simulator is implemented using Python, currently 3.6, and the data analytical libraries NumPy [7], and
Pandas [4]. It makes use of a virtual environment to provide platform independence and provides a straightforward
deployment. It is seamlessly built to facilitate quick reconfiguration of varying agent populations, market conditions,
exchange rules, and agent hyperparameters.
Basic execution of the simulation can be as simple as: python abides.py -c config, where config is the name
of an experimental configuration file. Additional command line parameters are forwarded to the configuration code for
processing, so each experimental configuration can add its own required parameters to a standard interface. Complex
experimental configuration can be performed directly within the config file since it is simply Python code, however the
inclusion of command line arguments is beneficial for coarse grain parallelization of multiple experiments of the same
type, but with varied simulation parameters.
A typical configuration file will specify a historical date to simulate and a simulation start and stop time as a nanosecond-
precision pandas.Timestamp objects. It will then initialize a population of agents for the experiment, configuring each
as desired. For example, an experiment could involve 1,000 background agents (perhaps Zero Intelligence agents or
Heuristic Belief Learning agents), 100 high-frequency trading agents, and one impact agent with various initialization
parameters to control their behavior. Each agent will at least be given a unique identifier and name. The configuration
file will also construct a latency matrix (pairwise between all agents at nanosecond precision) and latency noise model
which will be applied to all inter-agent communications. If a "data oracle", a utility with access to a data source outside
the simulation, is required for the experiment, the configuration file will initialize one. Finally a simulation kernel will
be initialized and run, passing it the agent population, oracle, and other simulation parameters.
Note that there is nothing finance-specific about the bootstrapper, configuration template, simulation kernel, or the base
Agent class. All are appropriate for use in any continuous-time discrete event simulation.
4.1 Example: A Momentum Trading Agent
To highlight the simplicity of creating a functional trading agent in our simulated environment, we present the code for
a basic momentum trader. It wakes each minute during the day, queries the last trade price, projects a future price using
linear regression over a configurable last N data points, and places a market order based on this projection. Following
is the complete source, excluding import statements:
class MomentumAgent ( TradingAgent ):
def __init__ ( self , id , name , symbol , startingCash , lookback ):
super (). __init__ (id , name , startingCash )
self . symbol = symbol
self . lookback = lookback
self . state = " AWAITING_WAKEUP "
self . trades = []
def wakeup ( self , currentTime ):
can_trade = super (). wakeup ( currentTime )
if not can_trade : return
self . getLastTrade ( self . symbol )
self . state = " AWAITING_LAST_TRADE "
def receiveMessage ( self , currentTime , msg ):
super (). receiveMessage ( currentTime , msg )
if self . state == " AWAITING_LAST_TRADE " and \
msg . type == " QUERY_LAST_TRADE ":
last = self . last_trade [ self . symbol ]
self . trades = ( self . trades + [ last ])[: self . lookback ]
8
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
if len ( self . trades ) >= self . lookback :
m , b = np . polyfit ( range ( len ( self . trades )) , self . trades , 1)
pred = self . lookback * m + b
holdings = self . getHoldings ( self . symbol )
if pred > last :
else :
self . placeLimitOrder ( self . symbol , 100 - holdings ,
True , self . MKT_BUY )
self . placeLimitOrder ( self . symbol , 100+ holdings ,
False , self . MKT_SELL )
self . setWakeup ( currentTime + pd . Timedelta ("1m" ))
self . state = " AWAITING_WAKEUP "
5 Case Study: Background Agents
A long-term goal is to produce realistic but possibly noisy re-simulations of particular days in history to play out various
"what if" scenarios. The idea is to populate the simulation with a large number of trading agents that provide a realistic
environment into which experimental agents can be injected.
(a) IBM: September 30, 2008
(b) MSFT: June 24, 2016
Figure 4: Simulated trades versus historical trades on two days.
Our initial effort towards this goal involves the introduction of a data oracle with access to fine-resolution historical
trade information, and the creation of a "background" agent which is able to request a noisy observation of the most
recent historical trade as of the agent's current simulated time. The approach is meant to reproduce the behavior of
a trader whose beliefs regarding the fundamental value of a stock are informed by interpretations of news and other
incoming information. It was inspired by the concept of a stock's "fundamental value" as used in the work of Wang and
Wellman. [9] Our approach is similar, but it uses historical data as a baseline rather than a mean-reverting stochastic
process.
A common baseline agent in the continuous double auction literature is the Zero Intelligence (ZI) trader [3] which
submits random bids and offers to the market, usually drawn from some stochastic distribution around a central value
belief for the underlying instrument.
Our agent.BackgroundAgent class follows the general spirit of the ZI trader, but with two important distinctions:
The central value belief at any time is a mixture of the prior belief with a noisy observation of a historical trade; and the
agent implements an extremely basic arbitrage strategy between the last simulated trade price and its internal belief.
Thus the valuation is influenced by random factors, but the direction of limit orders placed is then rational, with the
agent assuming the simulated price will converge to its value belief over time. Each background agent trades only a
single symbol on a single exchange.
In our current configuration, a background agent typically follows the following basic logic, given some wake frequency
F in some unit of time (microseconds, seconds, etc):
9
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
1. Request an initial wakeup time selected randomly from a uniform distribution across the first F interval after
market open
2. On wakeup, cancel any unfilled orders and wait for confirmation.
3. Query the exchange for the last trade price of this agent's symbol of interest and wait for the response.
4. Request a new noisy historical observation from the data oracle, and mix this observation with any prior belief
to obtain a new posterior value belief.
5. Determine the direction from the simulated last trade to this agent's value belief. Place a limit order to bring
the agent's holdings in line with a presumed profitable position: entering, exiting, or reversing position as
necessary.
6. Request a new wakeup call for the current time plus approximately F .
Figure 4 compares the behavior of 100 background agents interacting in ABIDES with the actual intra-day price on
two separate days in history. Ideally, we will see a price history that closely resembles the day in history, with similar
statistical properties.
6 Case Study: Market Impact
One area in which we believe simulation can add significant value to the current state of knowledge in finance is more
accurate models of the market impact of large trades. Each order placed at the exchange potentially "moves the market"
due to the nature of the market microstructure within the order book: arriving orders can add liquidity at a better price,
altering the spread; or can match existing orders and remove liquidity from the market. See Figure 5 for an example of
mechanical market impact.
Figure 5: Example of mechanical market impact.
Models that rely on historical data encounter limitations stemming from the inability to repeat history while introducing
an experimental change and allowing subsequent events to be altered by that change. Models can attempt to compare
"similar" days in history, but no two market days are ever the same.
If one could instead create a multi-agent simulation of a particular date in history such that a near approximation of
historical trades emerged in the absence of any significant change, but the trading agents would realistically react to any
such changes, a more accurate understanding of large trade impact could be attained. Here we present a preliminary
investigation of this idea.
We begin each simulation with a population of background agents and at least one exchange agent. For this experiment,
we add a single experimental agent, agent.ImpactAgent, which simply places a single large market order at a
predetermined time of day. The experimental parameter for the agent is its "greed"; that is, the proportion of available
10
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
order book liquidity near the spread it consumes at the time of trade. For example, a long impact agent with greed = 0.1
will place a market buy order for 10% of the shares on offer.
Our experiment includes 100 background agents and one exchange agent handling an order book for a set of symbols
including IBM. In Figure 6, the blue line represents each trade made by our population of background agents in
the absence of an impact trader. The orange line shows each trade made by the simulated trading agents given the
introduction of a single impact agent with varying "greed", acting one time with one trade at 10:00 AM on September
30, 2008. Both series are smoothed to improve visibility of the differences.
(a) MARKET BUY 1232 IBM
(b) MARKET BUY 2874 IBM
(c) MARKET BUY 5338 IBM
(d) MARKET BUY 7801 IBM
Figure 6: Market impact of trades at 10:00 AM.
(a) Impact agent with greed 0.5
(b) Impact agent with greed 0.1
Figure 7: Market impact event studies.
11
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
The impact trader has a clear effect on the market, despite the background agents' central tendency to arbitrage the price
toward historical levels, and the impact grows larger proportionally with its market bid size. The change is particularly
noticeable in the cyclical peaks of the auction. Due to the price elevation it caused, the impact trader's total profit
increased with the size of its bid from an average of $2,633 with greed = 0.3 to $12,502 with greed = 1.9. However
its profit per share declined from $2.14 to $1.60. We found a correlation between profit per share and trade size of
r = −0.31 across sixty experimental trials.
It is useful to consider these market impacts in aggregate across multiple experimental examples. ABIDES makes it
easy to produce study plots from logged simulation data. Figure 7 shows a time-aligned event study of many impact
trades at different times, on different days, to illustrate the range of likely price effects after the time of impact.
7 Conclusion and Future Challenges
We presented the design and implementation of ABIDES, a high-fidelity equity market simulator. ABIDES provides an
environment within which complex research questions regarding trading agents and market behavior can be investigated.
The simulation is demonstrated in two case studies. The first case study shows how previous intra-day transaction histo-
ries are closely reproduced by a population of interacting background trading agents communicating with an exchange
agent. These background agents are designed to provide a realistic market environment into which experimental agents
can be injected. The second case study illustrates how large market orders impact simulated prices not just immediately,
but for a significant period after the order arrives at the exchange. It is also intended to demonstrate the experimental
potential of the ABIDES platform.
We now have a robust simulation environment in which to develop and experiment with more complex trading agents,
including those based on approaches in machine learning and artificial intelligence.
8 Open Source Access and License
ABIDES is available through GitHub at https://github.com/abides-sim/abides under the BSD 3-clause license.
9 Acknowledgements
This material is based upon research supported by the National Science Foundation under Grant No. 1741026.
12
ABIDES: Towards High-Fidelity Market Simulation for AI Research
A PREPRINT
References
[1] BANKS, J. Handbook of simulation: principles, methodology, advances, applications, and practice. John
Wiley & Sons, 1998.
[2] FREIDMAN, D. The double auction market institution: A survey. The Double Auction Market Institutions,
Theories and Evidence, Addison Wesley (1993).
[3] GODE, D. K., AND SUNDER, S. Allocative efficiency of markets with zero-intelligence traders: Market as
a partial substitute for individual rationality. Journal of political economy 101, 1 (1993), 119 -- 137.
[4] MCKINNEY, W., ET AL. Data structures for statistical computing in python. In Proceedings of the 9th
Python in Science Conference (2010), vol. 445, Austin, TX, pp. 51 -- 56.
[5] NASDAQ OMX GROUP. NASDAQ TotalView - ITCH 5.0. http://www.nasdaqtrader.com/
content/technicalsupport/specifications/dataproducts/NQTVITCHSpecification.pdf.
Accessed: 2018-10-25.
[6] NASDAQ OMX GROUP. O*U*C*H Version 4.2. http://www.nasdaqtrader.com/content/
technicalsupport/specifications/TradingProducts/OUCH4.2.pdf. Accessed: 2018-10-25.
[7] OLIPHANT, T. E. A guide to NumPy, vol. 1. Trelgol Publishing USA, 2006.
[8] WANG, J., GEORGE, V., BALCH, T., AND HYBINETTE, M. Stockyard: A discrete event-based stock
market exchange simulator. In Simulation Conference (WSC), 2017 Winter (2017), IEEE, pp. 1193 -- 1203.
[9] WANG, X., AND WELLMAN, M. P. Spoofing the limit order book: An agent-based model. In Proceedings
of the 16th Conference on Autonomous Agents and MultiAgent Systems (2017), International Foundation
for Autonomous Agents and Multiagent Systems, pp. 651 -- 659.
[10] WELLMAN, M. P. Methods for empirical game-theoretic analysis. In AAAI (2006), pp. 1552 -- 1556.
[11] ZOOK, M., AND GROTE, M. H. The microgeographies of global finance: High-frequency trading and the
construction of information inequality. Environment and Planning A: Economy and Space 49, 1 (2017),
121 -- 140.
13
|
1612.00947 | 1 | 1612 | 2016-12-03T09:48:55 | Disruptive innovations in RoboCup 2D Soccer Simulation League: from Cyberoos'98 to Gliders2016 | [
"cs.MA"
] | We review disruptive innovations introduced in the RoboCup 2D Soccer Simulation League over the twenty years since its inception, and trace the progress of our champion team (Gliders). We conjecture that the League has been developing as an ecosystem shaped by diverse approaches taken by participating teams, increasing in its overall complexity. A common feature is that different champion teams succeeded in finding a way to decompose the enormous search-space of possible single- and multi-agent behaviours, by automating the exploration of the problem space with various techniques which accelerated the software development efforts. These methods included interactive debugging, machine learning, automated planning, and opponent modelling. The winning approach developed by Gliders is centred on human-based evolutionary computation which optimised several components such as an action-dependent evaluation function, dynamic tactics with Voronoi diagrams, information dynamics, and bio-inspired collective behaviour. | cs.MA | cs |
Disruptive innovations in RoboCup 2D Soccer
Simulation League: from Cyberoos'98 to
Gliders2016
Mikhail Prokopenko1 and Peter Wang2
1 Complex Systems Research Group, Faculty of Engineering and IT
The University of Sydney, NSW 2006, Australia
[email protected]
2 Data Mining, CSIRO Data61
PO Box 76, Epping, NSW 1710, Australia
Abstract. We review disruptive innovations introduced in the RoboCup
2D Soccer Simulation League over the twenty years since its inception,
and trace the progress of our champion team (Gliders). We conjecture
that the League has been developing as an ecosystem shaped by diverse
approaches taken by participating teams, increasing in its overall com-
plexity. A common feature is that different champion teams succeeded
in finding a way to decompose the enormous search-space of possible
single- and multi-agent behaviours, by automating the exploration of
the problem space with various techniques which accelerated the soft-
ware development efforts. These methods included interactive debug-
ging, machine learning, automated planning, and opponent modelling.
The winning approach developed by Gliders is centred on human-based
evolutionary computation which optimised several components such as
an action-dependent evaluation function, dynamic tactics with Voronoi
diagrams, information dynamics, and bio-inspired collective behaviour.
1
Introduction
Agent Smith: "You can't win, it's pointless to keep fighting!
Why, Mr. Anderson? Why do you persist?"
Neo: "Because I choose to."
The Matrix Revolutions.
The first official RoboCup was held in 1997, proposing a new benchmark
for Artificial Intelligence (AI) and robotics. Incidentally, another classical AI
challenge was successfully met in May 1997 when IBM Deep Blue defeated the
human world champion in chess. By design, RoboCup and chess differ in a few
key elements: environment (static vs dynamic), state change (turn-taking vs
real-time), information accessibility (complete vs incomplete), sensor readings
(symbolic vs non-symbolic), and control (central vs distributed) [1]. These dif-
ferences are emphasised in the RoboCup 2D Soccer Simulation League [2], which
quickly gained prominence, becoming one of the largest RoboCup leagues.
In this league, two teams of 12 fully autonomous software programs (called
"agents") play soccer in a two-dimensional virtual soccer stadium (11 player
2
Mikhail Prokopenko and Peter Wang
agents and 1 coach agent in each team), with no remote control. Each player
agent receives relative and noisy input from its virtual sensors (visual, acoustic
and physical) and may perform some basic actions in order to influence its en-
vironment, e.g., running, turning and kicking the ball. The coach agent receives
perfect input but can communicate with the player agents only infrequently and
through a fairly limited channel. The ability to simulate soccer matches with-
out physical robots abstracts away low-level issues such as image processing and
motor breakages, allowing teams to focus on the development of complex team
behaviours and strategies for a larger number of autonomous agents [3,4].
A simulated game lasts just over 10 minutes on average, and is played over
a small network of computer workstations which execute the code in parallel.
Each simulation step takes merely a tenth of a second, during which the entire
sensory-motor cycle takes place within an agent: starting with receiving new
sensory inputs from the simulator, proceeding to updating the internal memory,
to evaluating possible choices, to sending the chosen action back to the simulator.
The main challenge for each agent is to derive the best possible action to execute
at any specific time, while facing unexpected actions of the opposing agents.
Over 20 years, the RoboCup community has developed the open-source 2D
simulator and visualisation software which currently, with various packaged util-
ities and basic agent libraries, contains nearly a million lines of code. During this
period, the League and the participating teams have undergone several transi-
tions each of which eventually expanded the level of agents' intelligence and their
behavioral complexity. In this paper we attempt to trace not only the ten-year
long progress of our own team from its first implementation (Cyberoos; partic-
ipated between 1998 and 2003) to the RoboCup-2016 champion team (Gliders;
competed first in 2012), but also put this trace in the context of the twenty-year
long evolution of the Sim2D League itself.
The conjecture we put forward is that the League has been developing as an
ecosystem with an increasing complexity shaped by different approaches taken
by participating teams. Furthermore, this evolving ecosystem has experienced a
series of salient transitions leading to emergence of qualitatively new properties
in the intelligence exhibited by the agents. By a transition we do not mean a mere
extension of some simulated capabilities, such as the introduction of goalkeepers,
heterogeneous player types, or a coach language. Instead, we associate a transi-
tion with a specific methodological advance which played the role of a disruptive
innovation, with wide-spread consequences affecting the entire "ecosystem", for
example, a release of standard libraries, and so on. We use the term "disruptive
innovation" in a broad sense to indicate an innovation that creates a new ecosys-
tem (by analogy with a new market or value network), eventually disrupting an
existing system, displacing established structures and relationships.
2 A simulated world
The foundation supporting the evolution of the League is undoubtedly the con-
struction of the soccer server itself, providing a centralised world model with
several key features, enhanced over the following years:
From Cyberoos'98 to Gliders2016
3
-- distributed client/server system running on a network and producing frag-
mented, localised and imprecise (noisy and latent) information about the
environment (virtual soccer field) [5,6];
-- concurrent communication with a number of autonomous agents [7];
-- heterogeneous sensory data (visual, auditory, kinetic) without a global vision,
and limited range of basic commands/effectors (turn, kick, dash, . . .) [8];
-- asynchronous perception-action activity and limited window of opportunity
to perform an action [9];
-- autonomous decision-making under constraints enforced by teamwork (col-
laboration) and opponent (competition) [10];
-- conflicts between reactivity and deliberation [11].
The only restriction that was imposed from the outset is that participants should
"never use central control mechanisms to control a team of agents" [12].
A crucial feature making this simulated world an evolving "ecosystem" is
the availability of binaries (and sometimes the source code) of participating
teams, contained within an online team repository. The repository is updated
after each annual RoboCup competition, allowing the participants to improve
their teams with respect to the top teams of the previous championships. These
improvements diversify the teams' functionality and explore the immense search-
space of possible behaviours in the quest for optimal solutions. This process
results in a co-evolution of the teams, raising the overall competition level.
3 Partial automation of development efforts
"AT Humboldt" from Humboldt University, Germany became the first champion
of the League at RoboCup-1997 (Nagoya, Japan). The team used a combination
of reactive and planning systems, successfully deploying its agents within the
simulated world.
The following couple of years passed under the domination of "CMUnited"
team from Carnegie Mellon University (USA) which took the championship in
1998 (Paris, France) and 1999 (Stockholm, Sweden). One of the key reasons
for this success was the development of several tools partially automating the
overall effort, such as an offline agent training module, and layered disclosure: a
technique for disclosing to a human designer the specific detailed reasons for an
agent actions (in run-time or retroactively). Layered disclosure made it possible
to inspect the details of an individual player's decision-making process at any
point [13], becoming in our view the first disruptive innovation in the League.
Together with the offline agent training module, it clearly exemplified the power
of automation in accelerating the development effort -- precisely because it en-
abled the design effort to reach into a larger part of the search-space by encoding
more diverse behaviours.
It is important to point out that there were other novelties introduced by
CMUnited-98 and CMUnited-99, such as "single-channel, low-bandwidth com-
munication", "predictive, locally optimal skills (PLOS)", "strategic positioning
using attraction and repulsion (SPAR)", etc. [13], but we believe that it is the
partial automation of the software development that became the disruptive in-
novation. It has led to a wide-spread adoption of several debugging, visualising,
log-playing, log-analysing, and machine learning tools.
4
Mikhail Prokopenko and Peter Wang
4 Configurational space
A number of new teams in 2000 utilised the code base of the 1999 champi-
ons, CMUnited-99: it provided code for interaction with the soccer server, skills,
strategies, and debugging tools in a variety of programming languages [14]. The
champion of RoboCup-2000 held in Melbourne, Australia, "FC Portugal" from
University of Aveiro and University of Porto, extended this code base with a
systematic approach to describing team strategy, the concepts of tactics, forma-
tions and player types, as well as the situation based strategic positioning, the
dynamic positioning and role exchange mechanisms [15,11].
The generic innovation underlying these mechanisms comprised the ability
to configure diverse single- and multi-agent behaviours. The range of these be-
haviours span from active (ball possession) to strategic (ball recovery), from
formations to tactics, and from individual skills to team strategies. Such diver-
sity resulted in a considerable configurational flexibility displayed by the winning
team, significantly increasing the software development productivity, and more
importantly, expanding the extent of the available behavioural search-space.
Not surprisingly, the expansion brought about by the larger configurational
capacity was further exploited by the introduction of a standard coach language
[16] enabling high-level coaching with explicit definition of formations, situations,
player types and time periods, and resulting in a high-level coordination of team
behaviour. In other words, a disruptive innovation again was delivered by a
method which allowed to access deeper regions of the available search-space.
Team "TsinghuAeolus" from Tsinghua University, China, which won the
next two championships (RoboCup-2001 in Seattle, USA, and RoboCup-2002
in Fukuoka), focussed specifically on increasing the agents' adaptability via a
novel online advice-taking mechanism [17]. The configurational space was ex-
tended by a task-decomposition mechanism that assigned different parts of the
task to different agents.
A major boost to the League was provided by the partial release of the source
code of the next champion, team "UvA Trilearn" from University of Amsterdam,
The Netherlands, which won RoboCup-2003 in Padua, Italy [18]. This release
resulted in a standardisation of many low-level behaviours and world model,
effectively "locking in" the configurational space attained by that time, and
motivating several teams to switch their code base to UvA Trilearn base.
5 Cyberoos: 1998 -- 2003
At this stage we take a brief look at our first team, Cyberoos, which partici-
pated in RoboCup competitions between 1998 and 2003. The Cyberoos'98 team
took 3rd place in the 1998 Pacific Rim RoboCup competition [19], while Cy-
beroos'2000 were 4th in the Open European RoboCup-2000 [9]. Despite these
regional successes, the team's best result at the world stage was a shared 9th
place which Cyberoos repeatedly took at the RoboCup competitions in 2000,
2001, 2002 and 2003, never reaching the quarter-finals [20,21,22,23]. In hind-
sight, the main reason for this lack of progress was an oversight of the main
From Cyberoos'98 to Gliders2016
5
tendency driving the innovations in the League: the exploration of the search-
space due to the automation of the development efforts and the standardisation
of the configurational space.
Instead, the approach taken by Cyberoos focussed on self-organisation of
emergent behaviour within a purely reactive agent architecture [21]. Only dur-
ing the later years the Cyberoos architecture diversified, and included semi-
automated methods that quantified the team performance in generic information-
theoretic terms [22,23]. This approach focussed on measuring the behavioural
and belief dynamics in multi-agent systems, offering a possibility to evolve the
team behaviour, optimised under a universal objective function, within the
framework of information-driven self-organisation [24,25,26]. However, this frame-
work has started to take a functional shape only a few years later, after the time
when the Cyberoos team effort stopped in 2003.
6 Search-space decomposition
The next decade of RoboCup championships witnessed an intense competition
between three teams: "Brainstormers" from University of Osnabruck, Germany,
"WrightEagle" from University of Science and Technology of China, and "HE-
LIOS" from Fukuoka University and Osaka Prefecture University, Japan. Brain-
stormers became champions three times: in 2005 (Osaka, Japan), 2007 (Atlanta,
USA), and 2008 (Suzhou, China); WrightEagle came first an incredible six times:
in 2006 (Bremen, Germany), 2009 (Graz, Austria), 2011 (Istanbul, Turkey), 2013
(Eindhoven, The Netherlands), 2014 (Joao Pessoa, Brazil) and 2015 (Hefei,
China); and HELIOS succeeded twice: in 2010 (Singapore) and 2012 (Mexico
City, Mexico).
6.1 Machine learning
Brainstormers' effort focussed on reinforcement learning methods aiming at a
universal machine learning system, where the agents learn to generate the ap-
propriate behaviors to satisfy the most general objective of "winning the match".
Unfortunately, as has been acknowledged [27], "even from very optimistic com-
plexity estimations it becomes obvious, that in the soccer simulation domain,
both conventional solution methods and also advanced today's reinforcement
learning techniques come to their limit -- there are more than (108 × 50)23 dif-
ferent states and more than (1000)300 different policies per agent per half time".
The high dimensionality of the search space motivated Brainstormers to
use a multilayer perceptron neural network [27]: a feedforward artificial neural
network which utilises a supervised learning technique called backpropagation
for training the network. Rather than developing a universal learning system,
Brainstormers succeeded in decomposing the problem into a number of individ-
ual behaviours (e.g., NeuroKick, NeuroIntercept, NeuroHassle) and tactics (e.g.,
NeuroAttack2vs2, NeuroAttack3vs4, NeuroAttack7vs8), learned with supervised
learning techniques.
Recently, there has been some renewed interest in backpropagation networks
due to the successes of deep learning. In our view, the potential of reinforcement
learning methods in RoboCup has not yet been fully realised, and deep learning
may yet to become a disruptive innovation for the Simulation league.
6
Mikhail Prokopenko and Peter Wang
6.2 Automated planning
WrightEagle team addressed the challenges of (i) high dimensionality of the
search space and (ii) the limited computation time available in each decision
cycle, by using Markov Decision Processes (MDPs). The developed framework
decomposes a given MDP into a set of sub-MDPs arranged over a hierarchical
structure, and includes heuristics approximating online planning techniques [28].
WrightEagle approach abandoned "the pursuit of absolute accuracy" and divided
the continuous soccer field into the discrete space, further subdividing it into the
players' control areas according to geometric reachability. The resultant structure
enables automated planning, accelerating the search process and extending the
search depth [28].
6.3 Opponent modelling
"HELIOS" team [29,30] followed a similar path, targeting a decomposition of
the problem space in developing an unsupervised learning method based on
Constrained Delaunay Triangulation (CDT) [31]. A Delaunay triangulation for
a set P of points in a plane is a triangulation D(P) such that no point in P
is inside the circumcircle of any triangle in D(P) (in CDT the circumcircle of
some triangles contains other triangles' vertices). The method divides the soccer
field into a set of triangles, which provide an input plane region for Neural Gas
(NG) and Growing Neural Gas (GNG) methods. Specifically, the set Pb of N
points represents specifically chosen positions of the ball on the field, while sets
Pi describe the sets of coordinates of each player 1 ≤ i ≤ 11, so that there is
a bijective correspondence between Pb and each of Pi. Moreover, when the ball
takes any position within a triangle of D(Pb), each player's position is computed
in a congruent way within D(Pi). During offline experiments or even during a
game, the behaviour of the opponent, for example, the players' motion, directions
of the passes, and the overall team formations, can be mapped, analysed and
categorised [29,30].
It is evident that the main reason behind the recurrent successes of all three
champion approaches is a dynamic decomposition of the problem space and its
subsequent efficient exploration. This innovation goes beyond a simple standard-
isation of low-level behaviours within a rich but static configurational space, by
employing automated learning and planning methods in a dynamic search.
7 Standardisation of "hardware"
An influential disruptive innovation arrived in 2010, when HELIOS team released
a major update of their well-developed code base [32]:
-- librcsc-4.0.0 : a base library for the RoboCup Soccer Simulator (RCSS);
-- agent2d-3.0.0 : a base source code for a team;
-- soccerwindow2-5.0.0 : a viewer and a visual debugger program for RCSS;
-- fedit2-2.0.0 : a team formation editor for agent2d.
This resulted in nearly 80% of the League's teams switching their code base
to agent2d over the next few years. One may think of this phenomenon as a
standardisation of the simulated hardware, freeing the effort to improving the
higher-level tactical behaviours.
From Cyberoos'98 to Gliders2016
7
8 Gliders (2012 -- 2016): Fusing human innovation and
artificial evolution
We turn our attention to our champion team which won RoboCup-2016 (Leipzig,
Germany): Gliders [33,34,35,36,37]. Gliders2012 and Gliders2013 reached the
semi-finals of RoboCup in 2012 and 2013; Gliders2014 were runner-ups in 2014;
Gliders2015 finished third in RoboCup-2015, and Gliders2016 (a joint effort of
the University of Sydney and CSIRO) became world champions in 2016.
RoboCup-2016 competition included 18 teams from 9 countries: Australia,
Brazil, China, Egypt, Germany, Iran, Japan, Portugal and Romania. Gliders2016
played 23 games during several rounds, winning 19 times, losing twice and draw-
ing twice, with the total score of 62:13, or 2.70 : 0.57 on average. In the two-game
semi-final round, Gliders2016 defeated team CSU Yunlu from Central South Uni-
versity (China), winning both games with the same score 2:1. The single-game
final against team HELIOS2016 (Japan) went into the extra time, and ended
with Gliders2016 winning 2:1. The third place was taken by team Ri-one from
Ritsumeikan University (Japan).
The 2016 competition also included an evaluation round, where all 18 par-
ticipating teams played one game each against the champion of RoboCup-2015,
team WrightEagle (China). Only two teams, the eventual finalists Gliders2016
and HELIOS2016, managed to win against the previous year champion, with
Gliders defeating WrightEagle 1:0, and HELIOS producing the top score 2:1.
The Gliders team code is written in C++ using agent2d-3.1.1 [32], and frag-
ments of source code of team MarliK released in 2012 [38].
In order to optimise the code, the Gliders development effort over the last
five years involved human-based evolutionary computation (HBEC): a set of evo-
lutionary computation techniques that rely on human innovation [39,40].
In general, evolutionary algorithms search a large space of possible solutions
that together form a population. Each solution is a "genotype": a complex data
structure representing the entire team behaviour encoded through a set of "de-
sign points". A design point can be as simple as a single parameter (e.g., risk
tolerance in making a pass), or as complicated as a multi-agent tactical behaviour
(e.g., a conditional statement describing the situation when a defender moves
forward to produce an offside trap).
Some design points are easy to vary. For instance, a formation defined via
Delaunay Triangulations D(Pb) and D(Pi), 1 ≤ i ≤ 11, is an ordered list of
coordinates, and varying and recombining such a list can be relatively easily
automated. Other design points have an internal structure and are harder to
permute. For example, a conditional statement describing a tactic has a condition
and an action, encoded by numerous parameters such as positional coordinates,
state information, and action details. Once such a statement (a design point) is
created by human designers, its encoding can be used by evolutionary algorithms.
However, the inception of the tactic needs creative innovation in the first place,
justifying the hybrid HBEC approach.
The HBEC solutions representing team behaviours are evaluated with re-
spect to their fitness, implemented as the average team performance, estimated
over thousands of games for each generation played against a specific opponent.
8
Mikhail Prokopenko and Peter Wang
Some solutions are retained and recombined (i.e, the members of the popula-
tion live) and some are removed (i.e., die) through selection. Importantly, the
evolutionary process is carried out within different landscapes (one per known
opponent), and typically results in different solutions evolved to outperform spe-
cific opponents. In order to maintain coherence of the resultant code, each design
point is implemented with a logical mask switching the corresponding part of the
genotype on and off for specific opponents (determined by their team names).
This is loosely analogous to epigenetic programming [41].
The approach is aimed at constantly improving performance from one arti-
ficial "generation" to another, with team designers innovating and recombining
behaviours while the fitness landscape and the mutations are for the most part
automated. The performance of Gliders was evaluated on several supercomputer
clusters, executing on some days tens of thousands of the experimental runs
with different behaviour versions. It would be a fair estimate that the number of
such trials is approaching 10 million. The overall search-space explored by the
HBEC includes variations in both Gliders behaviour and opponent modelling.
The approach incorporates disruptive innovations of the past years, including
the standardisation of simulated "hardware" and several effective search-space
decompositions.
Specific variations included (i) action-dependent evaluation function, (ii) dy-
namic tactics with Voronoi diagrams, (iii) information dynamics, and (iv) bio-
inspired collective behaviour.
The approach introduced in Gliders2012 [33] retained the advantages of a
single evaluation metric (implemented in agent2d [32]), but diversified the eval-
uation by considering multiple points as desirable states. These desirable states
for action-dependent evaluation are computed using Voronoi diagrams which
underlie many tactical schemes of Gliders.
Starting from 2013, Gliders utilised information dynamics [42,43,44,45,46,47]
for tactical analysis and opponent modelling. This analysis involves computation
of information transfer and storage, relating the information transfer to respon-
siveness of the players, and the information storage within a team to the team's
rigidity and lack of tactical richness.
The constraints on mobility, identified by the information dynamics, were
investigated and partially overcome with bio-inspired collective behaviour [36].
Gliders2015 utilise several elements of swarm behavior, attempting to keep each
player's position as close as possible to that suggested by a specific tactical
scheme, while incorporating slight variations in order to maximise the chances
of receiving the pass and/or shooting at the opponent's goal. This behaviour
increased the degree of coherent mobility: on the one hand, the players are
constantly refining their positions in response to opponent players, but on the
other hand, the repositioning is not erratic and the players move in coordinated
ways.
These directions were unified within a single development and evaluation
framework which allowed to explore the search-space in two ways: translating
human expertise into new behaviours and tactics, and exhaustively recombining
them with an artificial evolution, leveraging the power of modern supercom-
puting. This fusion, we believe, produced a disruptive innovation on its own,
providing the winning edge for Gliders.
From Cyberoos'98 to Gliders2016
9
9 Conclusion
In this paper we reviewed disruptive innovations which affected advancement
of the RoboCup 2D Soccer Simulation League over the twenty years since its
inception, and placed the progress of our champion team in this context. It is
important to realise that the neither of these processes has been linear, and many
ideas have been developing along a spiral-shaped trajectory, resurfacing over the
years in a different implementation. For example, the utility of evolutionary
computation supported by supercomputing has been suggested as early as 1997,
when a simulated team was developed with the agents whose high-level decision
making behaviors had been entirely evolved using genetic programming [48]. Yet
the complexity of the domain proved to be too challenging for this approach to
gain a widespread adoption at that time.
Without an exception, all the winning approaches combined elements of some
automation (debugging, machine learning, planning, opponent modelling, evolu-
tionary computation) with human-based innovation in terms of a decomposition
of the search-space, providing various configurations, templates and structures.
Is there still a way toward a fully automated solution, when the agents learn
or evolve to play a competitive game without a detailed guidance from human
designers, but rather by trying to satisfy a universal objective ("win a game")?
On the one hand, the ability to run a massive number of simulated games
on supercomputing clusters producing replicable results will only strengthen in
time [4], and so may lend some hope in meeting this challenge positively. On the
other hand, the enormous size and dimensionality of the search-space would defy
any unstructured exploration strategy. A methodology successfully resolving this
dilemma may not only provide an ultimate disruptive innovation in the League,
but also provide a major breakthrough in the general AI research.
Acknowledgments
Several people contributed to Cyberoos and Gliders development over the years.
Marc Butler, Thomas Howard and Ryszard Kowalczyk made exceptionally valu-
able contributions to Cyberoos' effort during 1998 -- 2002 [19,9,20,21]. We are
grateful to Gliders team members Oliver Obst, particularly for establishing
the tournament infrastructure supporting the team's performance evaluation on
CSIRO Accelerator Cluster (Bragg), and Victor Jauregui, for several important
insights on soccer tactics used in Gliders2016 [37]. We thank David Budden for
developing a new self-localisation method introduced in Gliders2013 [49,34] as
well as contributing to the analysis of competition formats [4], and Oliver Cliff
for developing a new communication scheme adopted by Gliders from 2014 [35].
The overall effort has also benefited from the study quantifying tactical interac-
tion networks, carried out in collaboration with Oliver Cliff, Joseph T. Lizier, X.
Rosalind Wang and Oliver Obst [46]. We are thankful to Ivan Duong, Edward
Moore and Jason Held for their contribution to Gliders2012 [33]. Gliders team
logo was created by Matthew Chadwick.
References
1. Asada, M., Kitano, H., Noda, I., Veloso, M.: RoboCup: Today and tomorrow --
What we have have learned. Artificial Intelligence 110 (1999) 193 -- 214
10
Mikhail Prokopenko and Peter Wang
2. Kitano, H., Tambe, M., Stone, P., Veloso, M.M., Coradeschi, S., Osawa, E., Mat-
subara, H., Noda, I., Asada, M.: The RoboCup Synthetic Agent Challenge 97. In:
RoboCup-97: Robot Soccer World Cup I. Springer, London (1998) 62 -- 73
3. Budden, D., Wang, P., Obst, O., Prokopenko, M.: Simulation leagues: Analysis of
competition formats. In Bianchi, R.A.C., Akin, H.L., Ramamoorthy, S., Sugiura,
K., eds.: RoboCup 2014: Robot Soccer World Cup XVIII. Volume 8992 of Lecture
Notes in Computer Science., Springer (2014) 183 -- 194
4. Budden, D.M., Wang, P., Obst, O., Prokopenko, M.: Robocup simulation leagues:
IEEE
Enabling replicable and robust investigation of complex robotic systems.
Robotics and Automation Magazine 22(3) (2015) 140 -- 146
5. Noda, I., Stone, P.: The RoboCup Soccer Server and CMUnited Clients: Imple-
mented Infrastructure for MAS Research. Autonomous Agents and Multi-Agent
Systems 7(1 -- 2) (July -- September 2003) 101 -- 120
6. Haker, M., Meyer, A., Polani, D., Martinetz, T.: A method for incorporation of new
evidence to improve world state estimation. In Birk, A., Coradeschi, S., Tadokoro,
S., eds.: RoboCup 2001: Robot Soccer World Cup V. Volume 2377 of Lecture Notes
in Computer Science., Springer (2002) 362 -- 367
7. Stone, P., Veloso, M.: Task decomposition, dynamic role assignment, and low-
bandwidth communication for real-time strategic teamwork. Artificial Intelligence
110(2) (1999) 241 -- 273
8. Riley, P., Stone, P., Veloso, M.: Layered disclosure: Revealing agents' internals.
In Castelfranchi, C., Lesperance, Y., eds.: Intelligent Agents VII. Agent Theories,
Architectures, and Languages -- 7th International Workshop, ATAL-2000, Boston,
MA, USA, July 7 -- 9, 2000. Lecture Notes in Artificial Intelligence. Springer (2001)
9. Butler, M., Prokopenko, M., Howard, T.: Flexible synchronisation within RoboCup
environment: A comparative analysis. In Stone, P., Balch, T., Kraetzschmar, G.,
eds.: RoboCup 2000: Robot Soccer World Cup IV. Volume 2019 of Lecture Notes
in Artificial Intelligence. Springer, Berlin, Heidelberg (2001) 119 -- 128
10. Stone, P., Riley, P., Veloso, M.: Defining and using ideal teammate and oppo-
In: Proceedings of the Twelfth Annual Conference on Innovative
nent models.
Applications of Artificial Intelligence. (2000)
11. Reis, L.P., Lau, N., Oliveira, E.: Situation based strategic positioning for co-
In: Balancing Reactivity and Social
ordinating a team of homogeneous agents.
Deliberation in Multi-Agent Systems, From RoboCup to Real-World Applications.
Springer, London, UK (2001) 175 -- 197
12. Noda, I., Suzuki, S., Matsubara, H., Asada, M., Kitano, H.: Robocup-97: The first
robot world cup soccer games and conferences. AI Magazine 19(3) (1998) 49 -- 59
13. Stone, P., Riley, P., Veloso, M.: The CMUnited-99 champion simulator team. In
Veloso, M., Pagello, E., Kitano, H., eds.: RoboCup-99: Robot Soccer World Cup
III. Volume 1856 of Lecture Notes in Artificial Intelligence. Springer (2000) 35 -- 48
14. Stone, P., Asada, M., Balch, T.R., Fujita, M., Kraetzschmar, G.K., Lund, H.H.,
Scerri, P., Tadokoro, S., Wyeth, G.: Overview of RoboCup-2000.
In Stone, P.,
Balch, T.R., Kraetzschmar, G.K., eds.: RoboCup 2000: Robot Soccer World Cup
IV. Volume 2019 of Lecture Notes in Computer Science., Springer (2000) 1 -- 28
15. Reis, L.P., Lau, N.: FC Portugal Team Description: RoboCup 2000 Simulation
League Champion. In Stone, P., Balch, T.R., Kraetzschmar, G.K., eds.: RoboCup
2000: Robot Soccer World Cup IV. Volume 2019 of Lecture Notes in Computer
Science., Springer (2001) 29 -- 40
16. Reis, L., Lau, N.: COACH UNILANG - A Standard Language for Coaching a
(Robo)Soccer Team.
In Birk, A., Coradeschi, S., Tadokoro, S., eds.: Robocup
2001: Robot Soccer World Cup V. Volume 2377 of Lecture Notes in Computer
Science., Springer (2001) 183 -- 192
From Cyberoos'98 to Gliders2016
11
17. Jinyi, Y., Ni, L., Fan, Y., Yunpeng, C., Zengqi, S.: Technical Solutions of Ts-
inghuAeolus for Robotic Soccer. In Polani, D., Browning, B., Bonarini, A., Yoshida,
K., eds.: RoboCup 2003: Robot Soccer World Cup VII, Springer (2004) 205 -- 213
18. Kok, J.R., Vlassis, N., Groen, F.: UvA Trilearn 2003 team description. In Polani,
D., Browning, B., Bonarini, A., Yoshida, K., eds.: Proceedings CD RoboCup 2003,
Padua, Italy, Springer (July 2003)
19. Prokopenko, M., Kowalczyk, R., Lee, M., Wong, W.Y.: Designing and modelling
In: Proceedings of the PRICAI-98
situated agents systematically: Cyberoos98.
Workshop on RoboCup. (1998) 75 -- 89
20. Prokopenko, M., Butler, M., Howard, T.: On emergence of scalable tactical and
strategic behaviour.
In Stone, P., Balch, T., Kraetzschmar, G., eds.: RoboCup
2000: Robot Soccer World Cup IV. Volume 2019 of Lecture Notes in Artificial
Intelligence. Springer, Berlin, Heidelberg (2001) 357 -- 366
21. Prokopenko, M., Wang, P., Howard, T.: Cyberoos'2001: "Deep Behaviour Projec-
tion" Agent Architecture. In Birk, A., Coradeschi, S., Tadokoro, S., eds.: RoboCup
2001: Robot Soccer World Cup V. Volume 2377 of Lecture Notes in Computer Sci-
ence., Springer (2002) 507 -- 510
22. Prokopenko, M., Wang, P.: Relating the entropy of joint beliefs to multi-agent
coordination.
In Kaminka, G.A., Lima, P.U., Rojas, R., eds.: RoboCup 2002:
Robot Soccer World Cup VI. Volume 2752 of Lecture Notes in Computer Science.,
Springer (2003) 367 -- 374
23. Prokopenko, M., Wang, P.: Evaluating team performance at the edge of chaos.
In Polani, D., Browning, B., Bonarini, A., Yoshida, K., eds.: RoboCup 2003:
Robot Soccer World Cup VII. Volume 3020 of Lecture Notes in Computer Sci-
ence., Springer (2004) 89 -- 101
24. Nehaniv, C., Polani, D., Olsson, L., Klyubin, A.: Evolutionary information-
theoretic foundations of sensory ecology: Channels of organism-specific meaningful
information. Modeling Biology: Structures, Behaviour, Evolution (2005) 9 -- 11
25. Prokopenko, M., Gerasimov, V., Tanev, I.: Measuring spatiotemporal coordina-
tion in a modular robotic system.
In Rocha, L., Yaeger, L., Bedau, M., Flore-
ano, D., Goldstone, R., Vespignani, A., eds.: Artificial Life X: Proceedings of The
10th International Conference on the Simulation and Synthesis of Living Systems,
Bloomington IN, USA (2006) 185 -- 191
26. Prokopenko, M., Gerasimov, V., Tanev, I.: Evolving spatiotemporal coordination
in a modular robotic system. In Nolfi, S., Baldassarre, G., Calabretta, R., Hallam,
J.C.T., Marocco, D., Meyer, J.A., Miglino, O., Parisi, D., eds.: From Animals to
Animats 9: 9th International Conference on the Simulation of Adaptive Behavior
(SAB 2006), Rome, Italy, September 25-29 2006. Volume 4095 of Lecture notes in
computer science. (2006) 558 -- 569
27. Riedmiller, M., Gabel, T., Trost, F., Schwegmann, T.: Brainstormers 2D -- Team
Description 2008. In: RoboCup 2008: Robot Soccer World Cup XII; CD. (2008)
28. Zhang, H., Chen, X.: The decision-making framework of WrightEagle, the
In: Robot Soccer
RoboCup 2013 soccer simulation 2D league champion team.
World Cup, Springer (2013) 114 -- 124
29. Akiyama, H., Noda, I.: Multi-agent positioning mechanism in the dynamic envi-
ronment. In Visser, U., Ribeiro, F., Ohashi, T., Dellaert, F., eds.: RoboCup 2007:
Robot Soccer World Cup XI. Springer, Berlin, Heidelberg (2008) 377 -- 384
30. Akiyama, H., Shimora, H.: Helios2010 team description. In: RoboCup 2010: Robot
Soccer World Cup XIV; CD. (2010)
31. Chew, L.P.: Constrained Delaunay Triangulations. Algorithmica 4(1-4) (1989)
97 -- 108
32. Akiyama, H.: Agent2D Base Code. http://www.rctools.sourceforge.jp (2010)
12
Mikhail Prokopenko and Peter Wang
33. Prokopenko, M., Obst, O., Wang, P., Held, J.: Gliders2012: Tactics with action-
dependent evaluation functions. In: RoboCup 2012 Symposium and Competitions:
Team Description Papers, Mexico City, Mexico, June 2012. (2012)
34. Prokopenko, M., Obst, O., Wang, P., Budden, D., Cliff, O.: Gliders2013: Tactical
analysis with information dynamics. In: RoboCup 2013 Symposium and Compe-
titions: Team Description Papers, Eindhoven, The Netherlands, June 2013. (2013)
35. Prokopenko, M., Obst, O., Wang, P.: Gliders2014: Dynamic Tactics with Voronoi
In: RoboCup 2014 Symposium and Competitions: Team Description
Diagrams.
Papers, Joao Pessoa, Brazil, July 2014. (2014)
36. Prokopenko, M., Wang, P., Obst, O.: Gliders2015: Opponent avoidance with bio-
In: RoboCup 2015 Symposium and Competitions:
inspired flocking behaviour.
Team Description Papers, Hefei, China, July 2015. (2015)
37. Prokopenko, M., Wang, P., Obst, O., Jaurgeui, V.: Gliders2016: Integrating multi-
agent approaches to tactical diversity. In: RoboCup 2016 Symposium and Compe-
titions: Team Description Papers, Leipzig, Germany, July 2016. (2016)
38. Tavafi, A., Nozari, N., Vatani, R., Yousefi, M.R., Rahmatinia, S., Pirdir, P.: MarliK
2012 Soccer 2D Simulation Team Description Paper. In: RoboCup 2012 Symposium
and Competitions: Team Description Papers, Mexico City, Mexico. (2012)
39. Kosorukoff, A.: Human based genetic algorithm. In: Systems, Man, and Cybernet-
ics, 2001 IEEE International Conference on. Volume 5., IEEE (2001) 3464 -- 3469
40. Cheng, C.D., Kosorukoff, A.: Interactive one-max problem allows to compare the
performance of interactive and human-based genetic algorithms. In Deb, K., ed.:
Genetic and Evolutionary Computation -- GECCO 2004: Genetic and Evolutionary
Computation Conference, Seattle, USA, June 26-30, 2004. Springer (2004) 983 -- 993
41. Tanev, I., Yuta, K.: Epigenetic programming: Genetic programming incorporating
epigenetic learning through modification of histones. Information Sciences 178(23)
(2008) 4469 -- 4481
42. Lizier, J.T., Prokopenko, M., Zomaya, A.Y.: Information modification and particle
collisions in distributed computation. Chaos 20(3) (2010) 037109
43. Wang, X.R., Miller, J.M., Lizier, J.T., Prokopenko, M., Rossi, L.F.: Quantifying
and tracing information cascades in swarms. PLoS ONE 7(7) (2012) e40084
44. Ay, N., Bernigau, H., Der, R., Prokopenko, M.:
Information-driven self-
organization: the dynamical system approach to autonomous robot behavior. The-
ory in Biosciences 131 (2012) 161 -- 179
45. Lizier, J.T., Prokopenko, M., Zomaya, A.Y.: Coherent information structure in
complex computation. Theory in Biosciences 131 (2012) 193 -- 203
46. Cliff, O., Lizier, J., Wang, R., Wang, P., Obst, O., Prokopenko, M.: Towards
quantifying interaction networks in a football match. In Behnke, S., Veloso, M.,
Visser, A., Xiong, R., eds.: RoboCup 2013: Robot Soccer World Cup XVII, Springer
(2013) 1 -- 12
47. Lizier, J.T., Prokopenko, M., Zomaya, A.Y.: A framework for the local information
dynamics of distributed computation in complex systems. In Prokopenko, M., ed.:
Guided Self-Organization: Inception. Springer (2014) 115 -- 158
48. Luke, S.: Genetic Programming Produced Competitive Soccer Softbot Teams for
RoboCup97.
In Koza, J.R., Banzhaf, W., Chellapilla, K., Deb, K., Dorigo, M.,
Fogel, D.B., Garzon, M.H., Goldberg, D.E., Iba, H., Riolo, R.L., eds.: Proceedings
of the 3rd Annual Genetic Programming Conf., Morgan Kaufmann (1998) 214 -- 222
49. Budden, D., Prokopenko, M.: Improved particle filtering for pseudo-uniform belief
distributions in robot localisation. In Behnke, S., Veloso, M.M., Visser, A., Xiong,
R., eds.: RoboCup 2013: Robot Soccer World Cup XVII. Volume 8371 of Lecture
Notes in Artificial Intelligence., Springer (2013)
|
1806.05542 | 1 | 1806 | 2018-06-14T13:45:25 | Status maximization as a source of fairness in a networked dictator game | [
"cs.MA",
"cs.GT",
"cs.SI",
"q-fin.GN"
] | Human behavioural patterns exhibit selfish or competitive, as well as selfless or altruistic tendencies, both of which have demonstrable effects on human social and economic activity. In behavioural economics, such effects have traditionally been illustrated experimentally via simple games like the dictator and ultimatum games. Experiments with these games suggest that, beyond rational economic thinking, human decision-making processes are influenced by social preferences, such as an inclination to fairness. In this study we suggest that the apparent gap between competitive and altruistic human tendencies can be bridged by assuming that people are primarily maximising their status, i.e., a utility function different from simple profit maximisation. To this end we analyse a simple agent-based model, where individuals play the repeated dictator game in a social network they can modify. As model parameters we consider the living costs and the rate at which agents forget infractions by others. We find that individual strategies used in the game vary greatly, from selfish to selfless, and that both of the above parameters determine when individuals form complex and cohesive social networks. | cs.MA | cs |
Status maximization as a source of fairness in a networked dictator game
Jan E. Snellman1, Gerardo Iniguez4,2,1,3, J´anos Kert´esz4,1, R. A. Barrio5, and Kimmo K. Kaski1
1Department of Computer Science, Aalto University School of Science, FI-00076 AALTO, Finland
2Next Games, FI-00100 Helsinki, Finland
3Instituto de Investigaciones en Matem´aticas Aplicadas y en Sistemas,
Universidad Nacional Aut´onoma de M´exico, 01000 M´exico D.F., Mexico
4Department of Network and Data Science, Central European University, H-1051 Budapest, Hungary and
5Instituto de F´ısica, Universidad Nacional Aut´onoma de M´exico, 01000 M´exico D.F., Mexico
(Dated: June 15, 2018)
Human behavioural patterns exhibit selfish or competitive, as well as selfless or altruistic tenden-
cies, both of which have demonstrable effects on human social and economic activity. In behavioural
economics, such effects have traditionally been illustrated experimentally via simple games like the
dictator and ultimatum games. Experiments with these games suggest that, beyond rational eco-
nomic thinking, human decision-making processes are influenced by social preferences, such as an
inclination to fairness. In this study we suggest that the apparent gap between competitive and
altruistic human tendencies can be bridged by assuming that people are primarily maximising their
status, i.e., a utility function different from simple profit maximisation. To this end we analyse a sim-
ple agent-based model, where individuals play the repeated dictator game in a social network they
can modify. As model parameters we consider the living costs and the rate at which agents forget
infractions by others. We find that individual strategies used in the game vary greatly, from selfish
to selfless, and that both of the above parameters determine when individuals form complex and
cohesive social networks. Dictator game, agent-based social simulation, superiority maximization
PACS numbers:
I.
INTRODUCTION
People's behaviour towards others lies in a broad range
from selfish to selfless: In the former, individuals are con-
cerned with their own narrow minded benefit like profit,
thus acting competitively against others, while in the lat-
ter they are concerned with the needs of the rest, that is,
they behave in an altruistic and even self-sacrificing way.
How people behave in real-life situations depends on
the social or economic context and on individual char-
acteristics. In order to get a deeper insight into human
behavioural patterns in an economic context, Kahneman
at al. designed the dictator game [1], a two-player game
where one of the players, the dictator, is tasked to di-
vide a given sum of money between both players. The
dictator can divide the money in any way, even keeping
it all. The game was designed to test some of the as-
sumptions of modern economic theories [2], mainly the
assumption of rationality, under which the dictator will
always keep all the money, not giving anything to the
other player.
In experiments, however, dictators tend
to give at least a minor fraction of money to the other
player, thus challenging the assumption of rationality of
economic actors [3–5]. It should be emphasised at this
point that the classical theory assumes that the utility
function of the dictator is just personal profit.
The mismatch between theory and observation in the
dictator game is largely caused by social factors, such as
appreciation for fairness and equality, caring for others,
reputation, etc. In other words, the empirical behaviour
of the dictators may be understood as renouncing eco-
nomic advantages to gain social status, such as esteem
of the other player, or more generally a good reputa-
tion. While social status has been studied extensively in
the social sciences, the analysis of its effects in economic
theories (via, say, the concept of bounded rationality [2]
has been more limited). A way to bridge this gap be-
tween human social and economic behaviour is to con-
sider how the wish to maximise social status may affect
social network structure and thus the economic decisions
of individuals. Social relations are a significant part of
the overall social standing or status of people, with both
negative and positive effects depending on the individ-
ual behaviour. Positive effects include increased political
and economic opportunities and social support in times
of need, while negative effects can range from social ex-
clusion to outright hostility. Hence humans usually have
to take into account how their actions affect their re-
lationships with others.
In the context of the dictator
game, we could see dictators weighing the worth of the
monetary reward against the penalty incurred to their
social relations, should they be seen to act too selfishly.
The idea of status maximisation as a driving motiva-
tion of humans can be traced back to Adler's school of
thought of individual psychology [6]. Adler explored the
scenario in which many psychological problems are the
result of feelings of inferiority and, consequently, applied
this idea to develop therapeutic techniques for what he
termed inferiority complex. More recently, we used status
maximisation (referred here as 'better than hypothesis'
[BTH]) as a key mechanism to model the co-evolution of
opinions and groups in social networks [7]. The aim of
this study is to understand how a simulated society of in-
dividuals, driven by status maximisation, behaves when
agents are allowed to freely form economic relationships
but with a utility, where the profit aspect is only one com-
ponent of status maximisation. To that end we use the
BTH to model the social interactions of agents and their
strategies in the dictator game. Previous agent-based
models have considered the effect of social preferences
such as inequality aversion and a tendency to fairness [8]
but not status maximisation. Here we show how the
drive of individuals to increase their status with respect
to others determines both their economic strategies and
the structural properties of the social network in which
they reside.
Our approach to create a new utility function based
on social status could be considered as an attempt to
restore the idea of rational decision making in the eco-
nomic interactions. However, we should emphasize that
this interpretation has its limits. First, quantifying social
status is highly non-trivial, but it could be circumvented
by making simple assumptions. Second, while profit is
simply measured in dollars (though the value of 1$ is
quite different for a beggar or for a millionaire), estimat-
ing social status has always a subjective component and
it changes depending on the circumstances. Our purpose
with the agent based approach is to test to what extent
people are willing to invest in enhancing their social cap-
ital at the expense of their profits if their comparative
position in the competition becomes better as a result.
The original dictator game is a two player game, where
one of the players, the dictator, is tasked with dividing
a fixed reward between the players at will [1, 3, 4].
It
is different from the ultimatum game introduced in [9],
where the second player gets to either accept or reject
the offer, such that the rejection would result in neither
player getting anything. Both games were constructed
to demonstrate the limitations of rational economic be-
haviour, but in this study we focus on the dictator game.
Assuming economic rationality (or profit maximisa-
tion) on the part of the dictators one would expect them
to keep the whole reward, and not give anything to the
other player. However, in the hundreds of dictator game
experiments in the past few decades it has been shown
that many players, in the role of the dictator, actually
do give out non-zero proportions of the reward to the
other player (see e.g. the meta-analyses in [4, 10, 11]),
which naturally challenges the notion of humans as ra-
tional economic actors. While there is a lot of variation
in the results, the average offers to other player can reach
as high proportions as almost 40%, as can be seen in Ta-
ble I for various studies, The effects of different social
influences on the behaviour of the dictators have been
studied extensively, (see for example, [12] and references
therein), and especially those of social norms (see e.g.
[13–15]). More recently, the effects of social networks,
the players are embedded in, has been studied in [16] for
the dictator game and in [17] for the ultimatum game.
From the theoretical point of view, the experimen-
tal results obtained with the dictator game and similar
games have been interpreted in terms of altruism [10, 18],
fairness, in the sense of inequity aversion [19–21], bio-
logical evolution [22–24], or even experimental artefacts
2
[12, 25]. Here we study the dictator game by means of
agent based modelling in the context of coevolving social
networks populated by agents, who are driven by max-
imising their social status.
The paper is organized as follows:
In the next sec-
tion we introduce the networked dictator game model
and explain the utility function we use for the evolution
dynamics of the game, as well as the model parameters.
Section III contains the results, including the analysis of
the network geometries as a function of the parameters.
In section IV we draw conclusions.
II. THE NETWORK DICTATOR GAME MODEL
Our dictator game model consists of a network of N
agents that form and break social connections to each
other and play the dictator game repeatedly with the
connected agents. The simulation proceeds in cycles, in
which each connected agent plays the dictator game with
all the agents connected to it. Note that for each pair of
linked agents the game is played twice in each game cycle
with the agents exchanging the roles of the dictator and
the supplicant between these instances of the game.
Each agent i is characterised by the accumulated "win-
nings", or wealth, denoted by vi. For every transaction
of the game an amount M of money is given to the agent
acting as the dictator with its own dictatorial division
strategy, denoted by xi (0 ≤ xi ≤ 1) which is the pro-
portion of M that it gives to the other players. After
each cycle is completed, an amount cM is subtracted
from every agent's wealth as living cost. For the sake of
simplicity we here assume that the proportion c is the
same for all the agents and for all times. If the wealth
of an agent should go below zero, it is set to zero, mean-
ing that the agent is still in the game.
In addition to
their social connections, the agents also adjust their divi-
sion strategies during the simulation with a hill climbing
algorithm,explained in more detail below.
When the agent i plays with agent j, with i being
the dictator, the accumulated monetary reserves of the
agents i and j change by
vi(t1) = vi(t0) + (1 − xi)M,
vj(t1) = vj(t0) + xiM,
(1)
(2)
where t0 denotes the moment of time before the transac-
tion and t1 moment of time after the transaction. When
one takes into account the reduction by cM , the full
amount of accumulated wealth of the agent i at cycle
T1 is as follows
vi(T1) = max(cid:16)vi(T0)+M(cid:0)ki(1−xi)+ Xj∈mi (1)
xj −c(cid:1), 0(cid:17),
(3)
where ki is the number of neighbours of the agent i and
T0 is the cycle preceding T1.
Focus
Average offers
3
Type
Study
Engels 2011
Cardenas & Carpenter 2008 meta-analysis
Camerer 2003
meta-analysis
Zhao 2015
Forsythe et al 1994
experiment
experiment
development
game theory
politeness
real rewards
meta-analysis general overview
28%
34%
∽ 20%
39%
∽ 20%
TABLE I: Some of the more signicant studies of the dictator game summarized.
The social network of agents was initially set to be ran-
domly connected the average degree hki, but in the course
of simulations there are no limitations on the degree of
the agents. We deal with an adaptive network [26, 27]:
At the end of each cycle the network is let to reconfigure
through rewiring the connections in such a way that the
agents keep track of how the other agents affect their sta-
tus. Here we assume that the agents compare themselves
against their neighbours, which means that every agent
stores information not only what the other agents have
given to it, but also what was given to its neighbours
and how well they have accumulated wealth in compari-
son to them. For the sake of encouraging agents to renew
their connections after negative experiences, we let their
memories fade over time.
The way the agent i determines the influence of agent
j on its overall status is ultimately derived according to
the BTH change of utility ∆i of agent i. We assume that
the agents only compare themselves to their neighbours
(or, in other words, that the agents only know the accu-
mulated wealth of their neighbours). Thus we can write
the utility as follows
∆i(T1) = vi(T1) − vi(T0) + Xl∈m1(i)
(vi(T1) − vl(T1))
− Xl∈m1(i)
(vi(T0) − vl(T0)).
(4)
To see how the actions of agent j affect ∆i it is instructive
to rearrange the terms of ∆i in the following way:
∆i(T1) = (ki(T0) + 1)(vi(T1) − vi(T0))
− Xl∈m1(i)
(vl(T1) − vl(T0))
= (ki(T0) + 1)(vi(T1) − vi(T0))
− Xl∈m1(i)/{j}
(vl(T1) − vl(T0))
−(vj(T1) − vj(T0)),
(5)
where ki is the degree of agent i. From Eq. 5 one can
see that agent j can influence ∆i in three ways: first
by giving money to i, second by giving money to the
other neighbours of i and third by accumulating money
itself. The amount of money given by agent j to all
connected agents is xjM per dictator game cycle. Thus,
it is possible to define a cumulative utility matrix Uij to
describe how the agent i has benefited from the actions
of agent j at cycle T1 as
Uij(T1) = Uij (T0) + aij U ′
ij(T1) + γij(T0),
(6)
where aij is an element of the adjacency matrix:
aij = (cid:26) 1,
0, otherwise,
if agents i and j are linked
(7)
γij is the matrix of memory parameters, nI is the amount
of agents in I = m1(i) ∩ m1(j) and
U ′
ij(T1) = (ki(T0) − nI + 1)xj(T0)M
−(vj(T1) − vj (T0))
(8)
The memory parameter matrix γij measures the speed
at which the agents forget how they were treated, and it
is designed so as to reduce Uij to zero in time. Thus, it
can be written in the form
γij (T0) =
γ0,
if U T
ij (T0) ≥ γ0,
− γ0,
U T
ij (T0),
− U T
ij (T0),
if U T
ij (T0) ≤ −γ0
if − γ0 ≤ U T
if 0 ≤ U T
ij (T0) ≤ 0
ij (T0) ≤ γ0
(9)
ij (T0) = Uij (T0) + aijU ′
where the memory parameter γ0 (assumed to be constant
for the sake of simplicity) is the maximum pace of forget-
ting and U T
ij (T1). For the sake of
simplicity we also assume that the agents have full knowl-
edge where the accumulated wealth of the other agents
is coming from, so, for instance, the agent i can adjust
Uij even if the agent j is not connected to it.
After each cycle of the game, the agents form connec-
tions with the agents that have benefited them and cut
connections with agents that have not, in other words the
agent i will form a connection with the agent j if such
a link is not already present and Uij ≥ 0, and cut an
existing link with the agent j if Uij < 0.
The agents adjust their division strategies xi using a
simple hill climbing algorithm. At first, the xi's are ran-
domized, and changed by dx at every step. In the second
step, dx is randomly chosen to be either −0.1 or 0.1.
According to BTH, the change of status of the agent i
in between game cycles is given by Eq. (4), which deter-
mines the further evolution of xi: if ∆i ≥ 0, the direction
of dx is the same as before, but if ∆i < 0, the direction
of dx is changed.
Inspired by the simulated annealing
techniques, we reduce dx linearly during the first 1000
time steps to a minimum of 0.01.
A. Motivation of the model parameters c and γ0
The main motivation for including the cost parame-
ter c in the dictator game model is to test the effect of
mutual dependence on the social systems of the agents.
The cost parameter is important from the BTH (Better
Than Hypothesis) perspective, because it can be inter-
preted to represent outside pressure to the agents. Under
the BTH assumption it is not immediately clear whether
the status maximizing agents would form social bonds of
any type, let alone for the purpose of playing the dicta-
tor game. However, it is conceivable that some common
needs might force the agents to interact socially. Thus,
reducing the wealths of the agents by an amount con-
trolled by the parameter c, introduces into the model
an effect that requires agents to cooperate in order to
gain anything in the long run. This then allows us test-
ing whether mutual need enhances social interactions be-
tween the agents.
The potential range of the cost parameter c can in
principle extend to be positive or negative without lim-
its, but will be limited for the purposes of this study by
considering the effects of the parameter to the wealths of
the agents. In order for an agent to make profit in the
model it needs to have a sufficient number of neighbours
that are willing to play the dictator game with them on
good enough terms. When c = 0, the agents retain all
the wealth they manage to acquire from their dictator
game interactions with others for all time, while if c > 0,
their wealths slowly decline if not replenished through
the dictator game. As a direct consequence of these facts
the agents need more and more neighbours to be able to
cover their expenses as the parameter c is increased. If
the agent i has only one neighbour, j, it can generate
profit from their relationship as long as c < 2, if it uses
totally selfish strategy (xi = 0) and its partner is totally
generous (xj = 1). Of course, this arrangement is dis-
advantageous to the agent j, and therefore not likely to
happen, unless agent j happens to have a multitude of
other more generous neighbours. A pair of agents using
similar division strategies can only make profit if c < 1,
but when c is increased, an agent needs at least ⌈c⌉ neigh-
bours with similar division strategy to cover the costs of
its living standard.
As can be seen, the theoretical maximum profit an
agent can make from one relationship per simulation cy-
cle is 2M , while more realistically it can be expected to
amount to about M . In any case, when c is high enough,
the agents cannot cover their costs anymore even if they
form social links with every other agent in the simula-
tion.
In a simulation with N agents this point can be
expected to be reached at the latest somewhere between
c = N − 1 and c = 2(N − 1), depending on the configura-
tion of the social network of the agents and their division
strategies. In this study we do not look into the effects
of "universal basic income", i.e. the c < 0 case, and we
limit our scrutiny of the cost parameter well below the
upper limit of 2(N − 1).
4
The function of the memory parameter γ0 in the dic-
tator game model is to allow the agents to reform links
that have once been broken, ensuring the continuation of
the social dynamics. Without the moderating influence
of the memory mechanics in the model, the space of pos-
sible social connections would steadily diminish during
the model simulations, resulting in a very limited social
network.
The interesting range of the memory parameter γ0 can
be estimated using the same procedure as the one used
for the cost parameter c, i.e. by calculating the point at
which the parameter's influence overwhelms everything
else. Negative values for γ0 would make no sense in our
context, so the lower limit can be set to 0. The maximum
limit can be estimated using Eq. (6), from which it can
be seen that in the case of the memory parameter find-
ing this limit amounts to finding the maximum value of
U ′
ij(T1), which γ0 would need to exceed. From the defi-
nition given in Eq. (8), we can see that U ′
ij(T1) depends
in a rather complicated way on both the structure of the
social network of the agents and their division strategies,
but thankfully there are only two terms to consider. Let
us denote these terms as
a = (ki(T0) − nI + 1)xj(T0)M
b = vj(T1) − vj(T0),
(10)
(11)
so that U ′
follows that a ≥ 0 always.
ij(T1) = a − b. Since necessarily ki(T0) ≥ nI , it
The term a attains its minimum value of 0 when
xj(T0) = 0, and its maximum value of N M when
ki(T0) = N − 1, nI = 0 and xj (T0) = 1, i.e. when
the agent i is connected to all other agents and the agent
j has no other connections and uses the most generous
strategy possible in the dictator game. Similarly, the
term b has a minimum value of −cM when the agent j
receives nothing from the other agents, and a maximum
value of (2(N −1)−c)M when xj (T0) = 0 and xk(T0) = 1
for all k 6= j. As can be seen, the maximum value for a
can occur simultaneously with the minimum value of b
and vice versa, which means that the maximum value of
U ′
ij(T1) can be found either in the case where the term
a is at maximum and b at minimum or in the case where
a is at minimum and b at maximum. The latter of these
cases yields the greater value for U ′
ij(T1), amounting to
a total of (2(N − 1) − c)M . This is therefore a sensible
upper limit to γ0, since beyond that one would expect the
social dynamics to settle. As in the case of the cost pa-
rameter, we limit our study to relatively small values of
γ0, so we do not approach the upper limit (2(N −1)−c)M .
III. RESULTS
We initialise the dictator game model with N agents
each having a random dictatorial strategy or proportion
of the total amount, xi, the agent gives to the other
player.
In the simulation run at each time step each
one of the N agents in turn acts as a dictator and we let
the system to run for a fixed period of 10000 time steps.
For M we use the value of 1. In this work we focused on
investigating the following characteristic network quan-
tities, i.e. the average degree hki, the average shortest
path hLi, the local and average clustering coefficients Ci
and hCi , the mean number of second neighbours hn(2)i,
the average assortativity coefficient hrai, and the aver-
age homophily coefficient hrhi. In addition, we measure
the susceptibility hsi, which is the second moment of the
number of s sized clusters, ns:
hsi = Ps nss2
Ps nss
.
(12)
As in percolation theory, the contribution of the largest
connected component of the network is neglected when
calculating the susceptibility (12). Furthermore, we in-
vestigated the assortativity and homophily coefficients
defined as the Pearson correlation coefficients of the de-
grees k and accumulated wealth v of linked agents, re-
spectively, as discussed in [28] and [7].
It should be
noted that the assortativity and homophily coefficients
are ill defined if agents all have exactly the same amount
of neighbours and if they are all connected to agents
with exactly the same amount of wealth, respectively.
These situations do rise in our simulations occasionally,
and when they do, the results for the assortativity and
homophily coefficients are excluded from averages calcu-
lated. We also calculated the Gini coefficient as a well
known measure of inequality, first proposed by Gini in
1912 [29], using the following definition
G = PN
j vi − vj
i PN
2N Pi vi
,
(13)
which basically measures the total difference between the
accumulated wealths v of agents.
A. Time-evolution of the model
In order to obtain sufficiently good statistics for de-
termining the averages of the quantities listed above the
simulations of the model were run for 100 realisations,
and time averages over the latter half of the time series
were also taken. The reason for taking the time averages
from the realisations was the very fluctuating nature of
the time-evolution of the model. At times, the entire
social network may cease to exist temporarily, although
these moments only occur within certain ranges of the
model parameters, especially when large values of γ0 are
involved.
In order to study the time-evolution of the properties
of the agents and their social networks in our model we
performed two singular simulations with N = 100 and
two different sets of parameter values, the first set being
γ0 = 0 and c = 5 (case A), and the second γ0 = 5 and
5
c = 0 (case B). In addition to determining the proper
measure for the averages of simulation results, the main
motivations for these experiments were to test converge
on the other hand, and to see how the model parameters
influence the temporal behaviour of the model on the
other. For example, one could surmise from the very
definition that γ0 could potentially have major effects on
the time-evolution of the social networks of the modeled
agents.
The results are illustrated in Fig. 1. The network prop-
erties seem to generally converge to some constant values
around which they fluctuate, but in case B these fluctu-
ations are very strong. The assortativity coefficient es-
pecially becomes almost meaningless as it can have both
negative and positive values in a very short period of time
due to the fluctuations, implying that the agents have
no clear preference on whether to seek connections with
those of same or different degree. In case A, in contrast,
while there are still relatively large fluctuations in the
value of the assortativity coefficient, the overall value of
the coefficient is clearly positive. While the fluctuations
of most network properties are very rapid, the homophily
coefficient in case A exhibits slowly varying behaviour.
y
t
r
e
p
o
r
p
300
200
100
0
40000
20000
y
t
r
e
p
o
r
p
0
0
N = 100, c = 5, γ0 = 0
1.0
wealth v
min wealth vm⟩n
max wealth vm⟨x
0.5
0.0
1.0
0.5
0.0
−0.5
assortativity ra
homophily rh
division strategy x
clustering C
hypergenerosity gh
Gini coefficient G
5000
time t
10000
0
N = 100, c = 0, γ0 = 5
5000
time t
10000
FIG. 1: The time-evolution of the agent and network prop-
erties when N = 100, with γ0 = 0 and c = 5 (above), or
γ0 = 5 and c = 0 (below). The minimum, average and max-
imum wealths of agents are displayed in the panels on the
left, while the right panels show the hypergenerosity preva-
lence along with homophily, assortativity, clustering and gini
coefficients.
In contrast to the network properties, the time evolu-
tion of the minimum, maximum and average monetary
reserves of agents in the simulations show no signs of
fast fluctuations. Also, it turns out that in this case it
is the parameter c that has greater impact: While in
case B the growth of all of the reserves is almost linear,
in case A the growth of the maximum and average re-
serves stall eventually and finally start fluctuating slowly
around constant values, while the minimum reserve stays
stubbornly at 0. There is thus a substantial difference in
behaviour between cases A and B when it comes to rel-
ative differences between the minimum, maximum and
average monetary reserves of the agents. All the reserves
reach tens of thousands in value in case B, while in case
A they do not rise above 300. Also, in case A the max-
imum reserves are in the final stages of the simulation
about four times the size of average reserves, while in
case B all the reserves are on relative terms very close to
each other, only diverging very slowly. These behavioural
differences are reflected in the gini coefficient, as it tends
to 0 in case A, and to a value of little over 0.6 in case
B, owing to the more unequal wealth distribution in the
latter case.
B. General characteristics of the social networks
produced by the model
The model parameters c and γ0 have a strong influ-
ence on the structure of the social network produced by
the model and the division strategies of the agents. The
social networks produced by the agents in the model can
vary from very simple to very complex depending on the
values chosen for these parameters. For a reference, a
final state of the social network with N = 100 agents,
c = 5 and γ0 = 2 is shown in Fig. 2. With these parame-
ter values the social networks of the agents produced by
the model acquire their most complex form and exhibit
clearly their most interesting features. Next we explain
what kinds of simpler forms the network may take and
with what parameter values, and how the complex net-
work shown in Fig. 2 emerges from these simpler forms.
When the model parameters are c = γ0 = 0, the net-
works formed consist only of collections of pairs or short
chains of linked agents. The more agents a chain has,
the rarer that chain is in the network. The total amount
of agents in a simulation also determines how long the
chains can get: chains longer than four agents seem to
never occur in simulations of 100 agents, but the chains of
even nine agents can manifest when the total population
in the simulation is increased to 300 agents. The strate-
gies xi employed by the linked agents are invariably most
generous possible, that is xi = 1. This is most likely due
to lack of a reason for the agents to tolerate unfairness
when c = 0, while the disconnected and linear nature
of the social networks formed by the agents is probably
due to the unforgiving nature of relation formation when
γ0 = 0.
Increasing the parameter γ0 from the value of 0 causes
the network structure to become gradually more compli-
cated.While the social networks of agents are generally
disjointed at relatively small values of γ0, the relatively
short chains of agents lengthen as γ0 increases, and at
large values of γ0 these chains become tangled. As γ0
is increased further, this tangledness only increases, un-
til the networks closely resemble that shown in Fig. 2.
At some point the agents start forming densely connected
hubs within the larger networks, which we name as "trade
associations". The agents in these associations seem to
have relatively coordinated strategies in comparison to
the agents outside these associations, which may be a re-
6
1.0
1.0
0.8
i
C
t
i
n
e
c
i
f
f
0.6
e
o
c
0.4
g
n
i
r
e
t
s
u
c
l
0.2
0.0
N = 100, c = 5, γ0 = 2
0.8
j
x
−
x
i
0.6
e
c
n
e
r
e
f
f
i
d
y
g
e
a
r
t
s
t
0.4
0.2
0.0
FIG. 2: An example of the final state of the social network
produced by our model, when N = 100, c = 5 and γ0 = 2.
The colours of the nodes show the local clustering coefficients,
while the colour of the edges show the proximity of the divi-
sion strategies employed by the agents linked by the edge, as
indicated. The sizes of the nodes correspond to the accumu-
lated wealths vi of the agents, while the widths of the edges
correspond to the strength of the connection between linked
agents, i.e min({Uij , Uji}).
sult of different "grand strategies" utilized by these dif-
ferent types of agents. The agents within the trade asso-
ciations seek fair exchange with other agents in the same
association, which in the context of our model means
having the same division strategies while the agents out-
side these associations generally fall into two categories:
those using relatively generous offerings to attract many
less generous partners, and those agents that in turn
take advantage of the more generous agents, but have
few other social connections themselves. In a way, this
arrangement is reminiscent of the patron-client relation-
ships, and as such we call the more loosely connected part
of the main component "patron-client network". As for
the offering proportions xi, the agents generally adopt
less generous strategies as the memory parameter γ0 is
increased, except for the trade associations, whose strate-
gies may be more flexible. In Fig. 2 the network has a
trade association in its lower part, while the rest is com-
posed of a patron client network.
Features that emerge only occasionally in our simu-
lations, but often enough to be noticeable, are totally
connected components that are completely disconnected
from the main network, and whose agents have totally
convergent division strategies. Obviously, these forma-
tions are extreme cases of trade associations, and as such
we call them "cartels". These cartels can be unstable in
the sense that they may periodically disband and reform,
but they may also be very robust at times.
When the parameter c is increased when γ0 = 0, the
strategies xi gradually become less generous, while the
general structure of the social network remains initially
the same as in the c = γ0 = 0 case, i.e. disjointed col-
lections of small chains of agents. However, when c is
increased sufficiently, the networks finally become more
complex. The network structure in the case of large c
and no γ0 shows some similarities to the one shown in
Fig. 2, such as clusters of densely connected agents rem-
iniscent of trade associations, but the distinction between
these and the patron-client network is weaker. Further-
more, in the case of large c the division strategies most
often decline to zero for all linked agents. Thus, one
cannot say for sure if any strategy coordination is tak-
ing place. One needs to remember that, while γ0 has a
direct effect on the relation formation behaviour of the
agents, c only has an indirect effect through the term
f = −(vj(T1) − vj (T0)) in Eq. 8, and that while the
former plays a role both in the forming and breaking of
relations, the effect of the latter turns out to be to purely
discourage the breaking of the relations. This is simply
because c always contributes positively to the Uij, since
xa)M,
vj (T0),
(c − kj(xj − 1) − Xa∈m1(j)
f =
ct = vj(T0) + (kj(xj − 1) + Xa∈m1(j)
if c ≤ ct,
where
if c ≥ ct,
(14)
xa)M.
(15)
7
while the average giving rate is about 28%), and on the
other hand the fact that the distribution of giving rates
is strongly skewed in favour of the dictator, with only
about 12% of dictators giving more than 50% of the re-
ward to the other player according to [11]. In this study
we use the term hypergenerosity for the tendency of the
agents to give more than 50%.
memory γ0
⟩.0
7.⟩
0.0
2.⟩
⟨0.0
cost c
⟩.0
7.⟩
0.0
2.⟩
⟨0.0
x
y
g
e
t
a
r
t
s
n
o
i
s
i
v
d
i
h
g
y
t
i
s
o
r
e
n
e
g
r
e
p
y
h
0.4
0.2
0.4
0.2
0.0
0.0
Thus, with sufficiently large c and small average degrees
Uij will always remain positive, and no existing relations
are ever broken. Most likely the typical course of a simu-
lation in γ0 = 0 case is that first the all the agents in the
simulation form a fully connected community at the first
time step, and subsequently most of the links between the
agents will be cut until the degrees of all agents with any
connections left are below what are allowed by c, after
which the network remains unchanged for the remainder
of the simulation.
The likely reason for the formation of chains of linked
agents in the low c and γ0 cases is the fact that any offer-
ings by an agent to the neighbours of its its neighbours
weaken the standing of the said neighbours. The for-
mation of the trade associations is probably connected
to these associations becoming socially acceptable when
forgiveness (γ0) of agents allows, or when there is enough
outside pressure (c) to the agents, or a suitable combi-
nation of these effects. Generally it seems as though γ0
on its own has a stronger effect on the network structure
than c, which on its own seems to have a stronger ef-
fect on the division strategies of agents than the network
structure.
The most relevant parameter values are those that re-
sult in the agents having similar division strategies to
those found in the real world experiments on the dicta-
tor game. Since our model is too simple to reproduce the
results of the experiments one-to-one, we focus only on
the most features of these results that are most relevant
in the context of our model. These features are the fact
that, on one hand, surprisingly large proportion of dic-
tators give something to the other player (according to
[11], only about 36% of dictators choose to give nothing,
0.0
2.⟩
⟩.0
cost c
7.⟩
⟨0.0
0.0
2.⟩
⟩.0
memory γ0
7.⟩
⟨0.0
FIG. 3: The average division strategies (upper panels) and
the prevalence of hypergenerosity (lower panels) as functions
of c with five constant values of γ0 (panels on the left) and γ0
with five constant values of c (panels on the right).
In order to find some values for the parameters with
which our model is able to at least some extent match
the average division strategies and the prevalence of hy-
pergenerosity, latter of which is denoted here by gh, we
performed simulations in which either c or γ0 was kept
constant, and the other was varied. The values tested
were 0, 2.5, 5, 7.5 and 10 for both parameters. As for a
definition for hypergenerosity prevalence we simply adopt
the proportion of agents with xi > 0.5 of all agents.
The results for average division strategies and hyper-
generosity are shown in Fig. 3. Considering the fact that
our agents are driven by the motive of superiority maxi-
mization, hypergenerosity is surprisingly common in our
simulations. Most notably, when one looks the lower left
panel of Fig. 3, one sees that for fixed values of 5.0, 7.5
and 10.0 for γ0, gh stays between values of approximately
0.2 and 0.35 for all values of c, which is considerably
above the 12% proportion reported in [11]. Also, when
γ0 is given value 2.5, gh only drops below 0.12 when either
c . 1.5 or c & 8.0. The highest proportion of overgener-
ous agents, or about 42%, occurs when γ0 = c = 0, but if
c is increased while γ0 is kept constant, this proportion
declines fast to a value only little over zero, as could be
expected from the social network behaviours discussed
N = 100, c = 7.5, γ0 = 2.0
N = 400, c = 7.5, γ0 = 2.0
8
1.0
0.8
0.6
0.4
0.2
0.0
i
C
t
i
n
e
c
i
f
f
e
o
c
g
n
i
r
e
t
s
u
c
l
1.0
0.8
0.6
0.4
0.2
0.0
j
x
−
x
i
e
c
n
e
r
e
f
f
i
d
y
g
e
t
a
r
t
s
1.0
0.8
0.6
0.4
0.2
0.0
i
C
t
i
n
e
c
i
f
f
e
o
c
g
n
i
r
e
t
s
u
c
l
1.0
0.8
0.6
0.4
0.2
0.0
j
x
−
x
i
e
c
n
e
r
e
f
f
i
d
y
g
e
t
a
r
t
s
3
2
i
)
x
(
P
n
o
i
t
u
b
i
r
t
s
d
i
1
0
0.0
0.1
0.2
0.3
0.4
division strategy xi
j
C
t
i
n
e
c
i
f
f
e
o
c
g
n
i
r
e
t
s
u
c
l
1.0
0.8
0.6
0.4
0.2
0.0
0.5
0.6
0.0
0.2
0.4
0.6
0.8
1.0
clustering coefficient Ci
FIG. 4: An example of the social network, when N = 100
and c = 7.5 and γ0 = 2.0, shown in the upper panel. Colours,
line widths, etc as indicated in Fig. 2. A histogram of the
resulting division strategies is shown in the lower panel.
FIG. 5: A social network resulting from a run of 400 agents
with c = 7.5 and γ0 = 2.0. A scatterplot of local clustering
coefficients of linked agents are shown in the lower panel.
above. While there is some rise with larger values of c,
gh only rises to about 5% at most in the very largest val-
ues of c tested. In general, from the lower left panel of
Fig. 3 one could draw the conclusion that the parameter
γ0 for the most part enhances gh, an observation that is
broadly confirmed by the lower right panel of Fig. 3: For
all the constant c values shown the trends are growing,
at least when γ0 . 1.5
As for the division strategies, it is interesting to note
that when γ0 > 0 the parameter c has relatively little
influence to the strategies, which generally hover around
or above values of 0.25, 0.30 or 0.35, depending on the
value of γ0 chosen. The value 0.30 happens to be rela-
tively close to the value found in [11], while 0.35 is closer
to findings of [10]. Only in the γ0 = 0 case we can see
the strong drop in the generosity of the agents, a phe-
nomenon already identified above in the context of the
social networks. As a function of γ0 the average divi-
sion strategy tends to be generally increasing, except in
the c = 0 case, in which it first declines from a high
of almost 0.5 to about 0.25, stagnates there and then
starts increasing. In all other tested cases the increasing
trends are rapid at first, but gradually slow down as γ0
increases. The values of the average division strategies
approximately stay between the 28% of [11] and 34% of
[10] when γ0 & 2 for the c = 5 case. The same holds for
the c = 10 case when γ0 & 5, and for the c = 2.5 and
c = 7.5 cases when γ0 & 4.
Choosing the most realistic values for the model pa-
rameters, i.e. those that yield the average hypergenos-
ity prevalence and division strategies closest to those ob-
served, involves a careful consideration of the effects of
the parameter γ0. On the one hand, this parameter in-
creases hypergenerosity prevalence, so it should not be
too high, but on the other hand it also increases aver-
age division strategies, so it should not be too low either.
While this dilemma places tight restraints on realistic pa-
rameter values, Fig. 3 shows that with γ0 = 2 and c = 7.5
gh is close to 12% and an average division strategy is
about 0.25, which are very close to the ones observed in
[11], although the division strategy is a little bit lower.
Better match might be found with a more through sweep
of the parameter space, but for our purposes this result
is close enough. Besides, as shown in Table I there is a
great deal of variance between results of different stud-
ies, with some of the older results being as low as 20%.
So, it could be argued that the 25% is clearly within the
spectrum of acceptable results. Fig. 4 shows an example
of the social network generated with γ0 = 2 and c = 7.5,
along with a histogram showing the frequency distribu-
tion of the division strategies. The network is somewhat
to similar to the one shown in Fig. 2, with a trade asso-
ciation and a patron-client network, and relatively dense
interconnections between these two components.
The histogram in Fig 4 shows a distribution of division
strategies of a peculiar shape with at least three local
peaks. The peak at about xi = 0.5 corresponds to fair
division, and such a peak has also been observed in ex-
periments [11]. The other two peaks at about xi = 0.15
and 0.3, however, have no counterparts in the experi-
ments, and there is no peak at xi = 0 (i.e. the point the
agent playing as dictator keeps everything to itself) as
one would expect from the experimental studies. That
our simulated experiments do not correspond one to one
to experimental observations is on one hand not surpris-
ing, since our setup differs greatly from typical dictator
game experiments, but on the other hand the distribution
of division strategies indicates that the special cases of to-
tally fair (xi = 0.5) and unfair (xi = 0) divisions have no
intrinsic meaning to our agents, a situation which might
change if the agents were made to compete for social
goodwill as well as accumulated wealths.
While the social networks of the agents belonging to
trade associations tend to be fully connected, those of
patron-client networks are characterised by aversion to
forming triangles. This can be seen by comparing the
local clustering coefficients of linked agents, as in Fig.
5, in which we show an example of a larger run with
400 agents, and in which c = 7.5 and γ0 = 2.0. The links
shown in the figure tend to concentrate in the upper right
and lower left portions of the plot, which provides a vi-
sual presentation of the different parts of the full social
network: In the trade association the agents tend to be
all connected to each other, thus they all have high local
clustering coefficients, and so the connections between
them show up in the upper right portion of Fig. 5. The
agents of the patron-client network have sparser connec-
tions and, therefore, lower local clustering coefficients,
and since they also mostly connect with each other, their
connections populate the lower left proportion of the plot.
The social relations formed by the agents that connect
these two communities are rarer, and show up outside
these areas. Especially strong are the sets of points run-
ning vertically and horizontally in straight lines in the
middle of the plot, which correspond to the links of the
peripheral members of the trade association seen in the
network. Since this particular network contains many
agents with only a single neighbour, for whom local clus-
tering coefficient is zero, and these particular agents tend
to connect the more sparsely connected part of the net-
work, there are especially prominent concentrations of
points on the lower parts of the axes.
With the parameter values c = 7.5 and γ0 = 2.0, our
model yields results for average hypergenerosity preva-
lence and division strategies that very roughly correspond
to those found experimentally. While parameters pro-
ducing a better match might be found by combing the
parameter space more carefully, it is also interesting to
ask, what do these (or any other) values tell about the
simulated society of the agents, and what are their rele-
vance to the real human societies?
In the case of the parameter c the answers to these
questions are relatively straightforward: As noted above,
⌈c⌉ is the minimum number of social connections one
needs in order to make profit in the game, when similar
division strategies are used. Thus c = 7.5 implies that
maintenance of the life-style of a single person requires
co-operation with at least eight other people, which sug-
gests a significant degree of interdependence in the social
system. The implications of the parameter γ0 are harder
to quantify, but in the simple system of only two agents i
and j linked to each other γ0 = 2.0 would allow the agents
to continue interacting even in an extremely unfair set-
ting, that is, for example, when xi = 1 and xj = 0. This
may seem to indicate rather radical levels of tolerance of
9
unfairness on part of the agents, but in the context of
our simulations we had 100 agents instead of only two,
and the average degree of the agents was well in excess
of eight. Since the components of the utility matrix Uij
of the agents change much more rapidly in this situation,
γ0 = 2.0 is really not that high. Take, for example, a
fully connected subgroup of nine members, where agent
1 has a selfish division strategy x1 = 0, while all the
others have hypergenerous strategy xi = 1. In this case,
U ′
1j = −16 after one round, which cannot be canceled by
such a small γ0. Thus we can say that the effect of γ0 pa-
rameter is rather subtle in the context of our illustration
case c = 7.5 and γ0 = 2.0.
C. The properties of the network and wealth
distribution
γ0 = 0.0
2.5
5.0
7.5
10.0
number of clusters nc
100maximum cluster size smax
degree k
0
0.5
clustering C
50
0
10
0
susceptibility s
0.0
20
10
0
number of second neighbours n⟨2⟩
10
0
25
0
0.5
0.0
assortativity ra
50
0
0.5
0.0
shortest path L
cluster size sc
homophily rh
0
5
cost c
10
0
5
cost c
10
FIG. 6: The network properties as functions of c. The
linestyles indicate, which of the five constant values of γ0 is
being shown.
In order to get more detailed picture on how our model
behaves as functions of its parameters, we performed
simulations to determine the averaged network proper-
ties mentioned above. Fig.
6 shows these averaged
network properties as functions of the parameter c, for
γ0 = 0, 2.5, 5, 7.5, 10. One of the most noticeable features
seen in the Figure is that in the γ0 = 0 case the behaviour
of the network properties is in most cases very different
to those of other γ0 values, which tend to behave sim-
ilarly to each other. Only the average numbers of first
and second neighbours and clustering coefficient exhibit
somewhat similar behaviours for all γ0 values.
The behaviours discussed in the context of Fig. 2 are
readily apparent in the Fig. 6, especially in the γ0 = 0
case. The steep decline in the number of clusters and the
increases in the average and maximum cluster sizes in
this case are the most obvious indication of the transition
from the collections of short chains of agents to the more
connected networks as c increases, which is also evident
in the steadily increasing average clustering coefficient.
Average susceptibility and path length give interesting
insight into the intermediate stages of this transition, as
they both spike at c = 2, which is due to the main com-
ponent of the social network becoming a one long chain
or a loop. It would seem that at this point the amount of
agents belonging to clusters outside the main component
reaches a maximum, as does the length of the chains of
agents in the main component.
As stated above, the results for the network properties
generally follow the same trends for all the other values of
γ0 tested. However, for all γ0 the average numbers of first
and second neighbours and the average clustering coeffi-
cient rise as c increases, implying that the parameter c
has a universally enhancing effect on the connectedness
of the agents. Especially the rising average clustering
coefficient might be an indication of growing trade asso-
ciations. As for the number of clusters and maximum
cluster size, the results obtained using γ0 values other
than zero follow the trends expected from earlier analy-
sis, as the former declines and latter increases, although
these trends are not nearly as clear as in the γ0 = 0 case,
and for γ0 = 7.5 and γ0 = 10 the maximum cluster size
is almost constant. Also, when γ0 6= 0 the average clus-
ter size increases monotonically as a function of c, while
for c . 3.5 and γ0 = 0 it actually starts decreasing from
its maximum value after a sharp increase, suggesting re-
emergence of clusters outside the main component of the
network. Unlike the γ0 = 0 case, the susceptibility and
the average path length do not exhibit drastic changes
as functions of c for other values of γ0, although there
may be slight increasing trend in the case of the former
and decreasing trend in the case of the latter. This near
constancy may possibly be due to the fact that the social
networks of agents can already be quite complex at c = 0
when γ0 is large enough, meaning that there is no clear
transition from one type of a network to another when c
is increased in this case, at least no transition that shows
up in the susceptibility and path lengths.
As noted above, the assortativity and homophily co-
efficients tend to fluctuate very strongly in time as the
simulations progress, often changing even signs. Never-
theless, taking averages over these quantities reveals some
interesting details on their behaviour as functions of the
model parameters. For example, both the homophily and
assortativity coefficients have increasing trends as func-
tions of c when γ0 = 0 and generally decreasing trends
otherwise, with only slight exceptions. The values of as-
sortativity coefficient are limited approximately to the
interval between −0.2 and 0.2 when γ0 6= 0, and rise to
little over 0.5 when γ0 = 0. In contrast, the values of the
homophily coefficient stay mostly positive for all values
of c and γ0, except for a point at c = 0.5, when γ0 = 0.
Thus one can draw the conclusion that, while the agents
do not have a clear preference on whether to connect to
10
similarly connected agents or not in the most cases stud-
ied, they do slightly favour forming connections to agents
with similar accumulated monetary reserves.
Other than the γ0 = 0 case, Fig. 6 contains rela-
tively little information on the effects of the parameter
γ0. While quantitative differences between results ob-
tained using different nonzero values for γ0 exist, they
tend to be small. Some systematic trends can be dis-
cerned, however. For example, the results for maximum
cluster size and average numbers of clusters start at sim-
ilar levels for all nonzero values of γ0, but they tend to
drift apart as c increases. The most striking effects, how-
ever, are seen in the average numbers of neighbours and
clustering coefficient, both of which clearly increase as
functions of both γ0 and c, a trend that is almost linear
for the former quantity. It is remarkable that for c = 0,
average number of neighbours seems to follow the value
of γ0 almost precisely.
c = 0.0
2.5
number of clusters nc
5.0
7.5
10.0
100maximum cluster si e smax
degree k
susceptibility s
0
0.5
0.0
clustering C
shortest path L
50
0
10
0
3
2
1
5.0
2.5
number of second neighbours n⟨2⟩
50
assortativity ra
0
0.5
0.0
0
5
memory γ0
10
50
0
0.5
0.0
cluster si e sc
homophily rh
0
5
memory γ0
10
FIG. 7: The network properties as functions of γ0. The
linestyles indicate, which of the five constant values of c is
being shown.
In order to study the effects of the γ0 parameter
more closely, we repeated the same exercise as we did
for parameter c, and calculated the averaged network
properties as functions of the parameter γ0, for c =
0, 2.5, 5, 7.5, 10. The results are shown in Fig. 7. Just
by looking at the figure we can see similar general issues
as noted in the case of Fig. 6, i.e. the c = 0 case often be-
haves differently to the others, while the results obtained
with other values of c often resemble each other qualita-
tively, with relatively small quantitative differences. In
the cases of the average numbers of clusters, maximum
cluster size and path length, for instance, the results tend
to converge to very similar values and trends for all c as
γ0 is sufficiently high, about 6 in the case of the aver-
age numbers of clusters and maximum cluster size, and
about 2 in the case of the average path length. Even be-
low these thresholds the results tend to match for c ≥ 5,
while the c = 0 and c = 2.5 cases tend to deviate from
the others.
The general trends are slowly decreasing for the path
length and the maximum cluster size, while the aver-
age number of clusters has a generally slowly increasing
trend. The results for the c = 0 case show trends oppo-
site to those of the other cases for low values of γ0, and
the changes tend to be more drastic. It is interesting to
note that, while the numbers of clusters creeps up and
the maximum cluster size creeps down, the average clus-
ter sizes become almost constant for all c values tested,
when γ0 is sufficiently large. Overall, it seems that the
parameter γ0 encourages the formation of small splinter
groups outside the main component of the social net-
work, thus the results on the cluster numbers and their
maximum sizes.
The results for the susceptibility and path length
shown in Fig. 7 exhibit a similar feature identified in
Fig. 6, that is, a spiking behaviour at γ0 = 1 when
c = 0, which is related to the formation of the long chains
of agents and their linking together, as discussed earlier.
A key difference is that the spikes in the susceptibility
and path length are not nearly as prominent as func-
tions of γ0 as they are as functions of c. Other than the
spike, the results for susceptibility tend to acquire rather
similar values for all c and γ0, although the they drift
apart somewhat as c becomes large. The general trend is
increasing, reflecting the increasing numbers of separate
clusters as γ0 rises.
The average numbers of neighbours and the clustering
coefficient behave rather similarly as functions of γ0 as
they do as functions of c. The general trends are increas-
ing, in the case of the former almost linearly so, with the
exception of very low values of γ0 and high values of c.
The reason for similar behaviour is again almost surely
related to growing trade associations. In the results on
the average numbers of second neighbours, however, one
can see a clear difference of behaviour between the two
cases. While the number of second neighbours tends to
increase as a function of c, as a function of γ0 it is only
rising almost monotonically when c = 0. For all the other
values of c tested it first rises to some maximum point
and then starts decreasing ever more slowly, eventually
becoming essentially constant. The steepness of the in-
crease and the decrease, along with the point where the
maximum are obtained, depend on the value of c tested:
For example, for c = 7.5 and c = 10 the rise is very steep,
the maximum occurs at γ0 = 0.5 and the decline is also
relatively rapid. For c = 2.5 and c = 5.0, in contrast, the
rise is rather slow, the maximums occur at γ0 = 6 and
at γ0 = 3, respectively, and the decline is almost imper-
ceptible. The reason limiting the rise of the number of
second neighbours may be related to the proliferation of
splinter clusters, especially of cartels, since they tend to
be fully connected.
The behaviours of the average assortativity and ho-
mophily coefficients as functions of γ0 is characterised by
11
relatively slow changes, with the exception of the c = 0
case. While the total change over the full range of γ0
can be significant in some cases, minimal changes from
a neighbouring value of γ0 are the rule. Overall, the
assortativity and homophily coefficients acquire similar
values as functions of γ0 as they do as functions of c, and
therefore similar conclusions apply.
memory γ0
5.0
7.5
0.0
2.5
10.0
100
cost c
5.0
7.5
0.0
2.5
10.0
10−1
G
t
n
e
i
c
i
f
f
e
o
c
i
n
G
i
10−2
0
5
cost c
10
0
5
memory γ0
10
FIG. 8: Gini coefficient G as function of c (left panel) and
γ0 (right panel). The linestyles indicate, which of the five
constant values of γ0 or c is being shown.
The wealth distribution, as measured by the gini coef-
ficient, is shown in Fig. 8. Generally, the gini coefficient
decreases as function of γ0 and increases as a function
of c, except in the γ0 = 0 case, which is characterised
by sharp fluctuations as a function of c. Also, in when
c = 0 or c = 2.5, the gini coefficient starts slowly rising
after steep decrease as a function of γ0, contrary to the
general trend. When γ0 6= 0, one sees that smaller the
γ0, the greater the speed of the rise of the gini coefficient
is as a function of c. Conversely, the greater the value of
c, the slower the decline of the gini coefficient will be as
a function of γ0. It can be thus concluded that the pa-
rameter c generally increases wealth disparities between
the agents, while γ0 tends to decrease them for a most
part, at least to a point.
D. The effects of the population number
Most of the simulations in this paper have only had
100 agents due to time and computational constraints.
To test the behaviour of our model with different num-
bers of agents, we performed simulations with N =
50, 100, 150, 200, 250, 300, 350 and N = 400 agents, with
parameter values c = 7.5 and γ0 = 2, which were chosen
for their ability to bring the average division strategies
and hypergenerosity prevalence close to those observed
at N = 100, as shown above. The results are shown in
Fig. 9. One of the more striking revelations from this
Figure is that the population number seems to have rel-
atively little effect on many of the results of the model,
especially the average number of neighbours, path length
and susceptibility, which only show very slight increasing
number of clusters nc
maximum cluster size ⟨smax
number of second neighbours ⟨n⟩2)
cluster size ⟨sc
clustering ⟨C
assortativity ⟨ra
homophily ⟨rh
division strategy ⟨x
Gini coefficient G
hypergenerosity gh
degree ⟨k
susceptibility ⟨s
shortest path ⟨L
200
100
y
t
r
e
p
o
r
p
0
0.50
y
t
r
e
p
o
r
p
0.25
0.00
y
t
r
e
p
o
r
p
7.5
5.0
2.5
50
100
150
200
250
network size N
300
350
400
FIG. 9: Various network properties as functions of the total
population N , when c = 7.5 and γ0 = 2. The upper panel
shows the average numbers of clusters, second neighbours and
cluster size along with maximum cluster size on average. The
middle panel shows the average division strategies and hyper-
generosity prevalence along with average gini, assortativity,
homophily and clustering coefficients. The lowest panel shows
the average path length, susceptibility and average number of
neighbours.
trends. The gini coefficient also changes little as a func-
tion of the population number, although its general trend
is a very slowly decreasing.
Most important of the results not affected much by
the population numbers are the average division strate-
gies and hypergenerosity prevalence. The average di-
vision strategy show a very slight growing trend when
N . 150, after which there is a slow decreasing trend
until N = 400, at which point the division strategies
have fallen slightly below 0.2 on average. Hypergeneros-
ity prevalence follows generally the very same pattern,
peaking at a value of about 0.2 when N = 150 and
slowly decreasing afterwards to a value of less than 0.1
at N = 400. While the average division strategy and hy-
pergenerosity prevalence obtained from the simulations
clearly correspond to the observations best at N = 100,
which is the point at which the calibration of the model
parameters was made, the fact that they stay relatively
close to the observations raises hopes for the general ap-
plicability of the model. In this study we do not, how-
ever, venture beyond N = 400 in our investigations, so
we cannot say exactly how the system behaves at very
large population numbers.
Of the network properties shown Fig. 9 the maximum
cluster size, average number of clusters, and assortativity
coefficient exhibit the greatest changes and most system-
atic trends, all of which happen to be increasing, while
the average clustering coefficient has a general decreasing
trend, which is rapid at first but slows down considerably
when N ≥ 100. In the upper panel of Fig. 9 we see an
12
interesting linkage between the average numbers of clus-
ters and maximum and average cluster sizes. While the
increase in maximum cluster size as a function of the pop-
ulation number is a matter of course, the simultaneous
strong rise in the average number of clusters drags the
cluster sizes down. This effect is seen both in the maxi-
mum and average cluster sizes: Although the maximum
cluster size is very close to N when N . 150, the relative
gap between N and the maximum cluster size gradually
widen as N increases, and so at N = 400 the maximum
cluster size is only about 250. The average cluster size,
however, shows a near consistent downward trend, which
is necessarily due to the large number of clusters gener-
ated by the model at ever larger population numbers.
The rising trend in the assortativity coefficient reveals
the changing preferences of the modeled agents in regard-
ing relation formation. While the social networks are dis-
sociative at N = 50, they become increasingly more as-
sociative at higher population numbers. The homophily
coefficient does not share such a straightforward trend,
as it is at times increasing and at other times decreas-
ing, but always positive, meaning that the agents will al-
ways favour forging or keeping ties to other agents with
similar accumulated wealths. It should be emphasized,
however, that both these coefficients, and especially the
assortativity coefficient, are subject to very strong tem-
poral fluctuations during the simulations, so these effects
are only present in the average sense.
In summary it could be said that the results on the
cluster sizes and numbers shown in Fig. 9 indicate that
the model produces ever greater amounts of ever smaller
clusters that splinter off the main component as popu-
lation numbers increase, while the decreasing clustering
coefficient and the increasing assortativity coefficient in-
dicate that the size of the patron-client networks grows
relative to the size of the trade associations. It should be
noted, however, that these results have been obtained
using only one set of model parameters calibrated at
N = 100 to replicate the observed results. We do not
delve deeper into the interaction of the model parame-
ters and the population number in this study.
IV. CONCLUSIONS AND DISCUSSION
In this study we have investigated the behaviour of
a network of agents seeking to maximize their relative
standings, according to better than hypothesis (BTH).
The agents are embedded in a co-evolving network, in
which the linked agents repeatedly play the dictator game
with each other for evolving their social relations, while
their status is measured by the amount of wealth they
thus acquire. The main motivation of this research is to
test, whether the agents driven by BTH would form any
connections or endow anything to their network partners.
In evolving their social relations the agents in this net-
work game keep track of how the other agents treat them
forming social relations but also cutting them as punish-
ment of selfish behaviour, and "forget" their treatment as
well as spend a fixed portion of their earnings for paying
their living costs once per game cycle. The cost of living
and the rate at which the agents "forget" their treat-
ment by the other agents are parameters of our model,
and we have studied their influence to the behaviour of
the system of agents.
Our simulation results show that agents acting along
the BTH do, indeed, form social connections for the pur-
pose of playing the dictator game and that the dictators
in these games often give non-zero amounts of money to
the other players. That the agents would give each other
anything at all in a game such as the dictator game is not
a self-evident conclusion, a priori. The agents can use ei-
ther very generous or very stingy strategies depending
on the model parameters and their position in the co-
evolving network.
Generally speaking, the structures of the social net-
works produced by the model vary strongly according to
the model parameters, and can be described as follows.
For small values of the model parameters the system
networks reduce to collections of short chains of agents,
which become longer as either model parameter is in-
creased. Ultimately these chains start fusing together at
high values of the parameters. The cost parameter alone
does not seem to have as dramatic effect on the structure
of the networks as the memory parameter, but it makes
the agents form more connections, especially in conjunc-
tion of non-zero memory parameter, thereby making the
network denser.
When the parameters are sufficiently large new sub-
structures, which we name "trade associations" and
"patron-client networks" start emerging. The former are
fully connected subgroups of agents that use relatively
similar division strategies, while the latter are composed
of agents that have diverse division strategies and rel-
atively sparse social connections, with generous agents
generally connecting to many stingier agents that in turn
do not form many connections, and virtually none with
each other. The emergence of these substructures demon-
strates on one hand that the agents driven by BTH are
capable of forming complex social structures using di-
verse strategies, and on the other hand that they can at
some level form social norms. Especially the fact that the
members of trade associations coordinate their division
13
strategies, indicates some appreciation of fairness in part
of the agents.
The results of our dictator game of networked agents
are in agreement with the empirically findings of altruis-
tic behaviour by humans in the role of the dictator, which
provides credence to the BTH (Better Than Hypothesis).
We find it very interesting that, based on rather simple
assumptions about the competition for superior social po-
sitions, the dynamics generates complex network struc-
tures indicating that this component of human behaviour
may have an important role in producing the empirically
observed structures in real societies.
It is also notable
that with suitable parameter values our model produces
average hyper-generosity prevalence and division strate-
gies that are reasonably in line with the ones observed
in earlier research. That BTH could mimic these ob-
servational facts, on top of being capable of facilitating
formation of complex social structures even at such a
simplistic level, is encouraging when considering possible
future uses of the BTH framework. However, it should
be emphasised that as such our model parameters do not
correspond to anything directly observable. In the future
work, parameters akin to the memory parameter could,
for example, be replaced with more detailed social mech-
anisms such as giving social relationships a value of their
own, and letting simulated agents compete over them.
Funding
This work was supported by Niilo Helander's founda-
tion grant No. 160095 (J.E.S.), the Academy of Fin-
land Research project (COSDYN) No. 276439 and EU
HORIZON 2020 FET Open RIA project (IBSEN) No.
662725 (K.K.), and Conacyt projects 799616 and 28327
(R.A.B.). R.A.B is also grateful for a sabbatical grant
from PASPA, DGAPA, UNAM, Mexico.
Acknowledgments
Computational
resources provided by the Aalto
Science-IT project have been utilised in this work. G.I.
and J.K. thank Aalto University for hospitality.
[1] D. Kahneman, J. L. Knetsch, and R. H. Thaler. Fair-
ness and the assumptions of economics. The Journal of
Business, 59(4):285–300, 1986.
[2] Herbert Simon. Bounded rationality and organizational
learning. Organization Science, 2(1):125–134, 1991.
[3] R. Forsythe, J. Horowitz, N. E. Savin, and M. Sefton.
Fairness in simple bargaining experiments. Games and
Economic Behavior, 6:347–369, 1994.
[4] C. Camerer. Behavioral Game Theory: Experiments in
Strategic Interaction. The Roundtable Series in Behav-
ioral Economics. Princeton University Press, 2003.
[5] Henrich J., R. Boyd, S. Bowles, C. Camerer, E. Fehr,
and H. Gintis. Foundations of Human Sociality: Eco-
nomic Experiments and Ethnographic Evidence from Fif-
teen Small-Scale Societies. Oxford University Press,
2004.
[6] A. Adler. The Practice and Theory of Individual Psychol-
ogy. Routledge, Trench and Trubner & Co, Ltd, 1924.
Reprint. Abingdon: Routledge (1999).
[7] Jan E Snellman, Gerardo Iniguez, Tzipe Govezensky, R A
14
Barrio, and Kimmo K Kaski. Modelling community for-
mation driven by the status of individual in a society.
Journal of Complex Networks, 5(6):817–838, 2017.
[8] B. Xianyu. Social preference, incomplete information,
and the evolution of ultimatum game in the small world
networks: An agent-based approach. Journal of Artificial
Societies and Social Simulation, 13(2):7, 2010.
[9] W. Guth, R. Schmittberger, and B. Schwarze. An ex-
perimental analysis of ultimatum bargaining. Journal of
Economic Behavior & Organization, 3(4):367–388, 1982.
[10] Juan Camilo Cardenas and Jeffrey Carpenter. Be-
havioural development economics: Lessons from field labs
in the developing world. The Journal of Development
Studies, 44(3):311–338, 2008.
entific Reports, 8(1222), 1 2018.
[18] David K. Levine. Modeling altruism and spitefulness in
experiments. Review of Economic Dynamics, 1(3):593 –
622, 1998.
[19] G. Bolton and A. Ockenfels. Strategy and equity: An
erc-analysis of the guth-van damme game. Journal of
Mathematical Psychology, 42:215–226, 1998.
[20] G. Bolton and A. Ockenfels. Erc: A theory of equity,
reciprocity and competition. The American Economic
Review, 90:166–193, 2000.
[21] E. Fehr and K. Schmidt. A theory of fairness, competi-
tion and cooperation. Quarterly Journal of Economics,
114(3):817–868, 1999.
[22] R. Axelrod and W. D. Hamilton. The evolution of coop-
[11] Christoph Engel. Dictator games: a meta study. Exper-
eration. Science, 211:1390–1396, 1981.
imental Economics, 14(4):583–610, Nov 2011.
[12] Nicholas Bardsley. Dictator game giving: altruism or
artefact? Experimental Economics, 11(2):122–133, Jun
2008.
[13] Nichola J. Raihani and Katherine McAuliffe. Dictator
game giving: The importance of descriptive versus in-
junctive norms. PLOS ONE, 9(12):1–17, 12 2014.
[14] Joseph Henrich, Jean Ensminger, Richard McElreath,
Abigail Barr, Clark Barrett, Alexander Bolyanatz,
Juan Camilo Cardenas, Michael Gurven, Edwins Gwako,
Natalie Henrich, Carolyn Lesorogol, Frank Marlowe,
David Tracer, and John Ziker. Markets, religion, commu-
nity size, and the evolution of fairness and punishment.
Science, 327(5972):1480–1484, 2010.
[15] Kun Zhao, Eamonn Ferguson, and Luke D. Smillie. In-
dividual differences in good manners rather than com-
passion predict fair allocations of wealth in the dictator
game. Journal of Personality, 85(2):244–256, 2017.
[16] Delia Baldassarri and Guy Grossman. The effect of group
attachment and social position on prosocial behavior. ev-
idence from lab-in-the-field experiments. PLOS ONE,
8(3):1–9, 03 2013.
[17] Xiao Han, Shinan Cao, Jian-Zhang Bao, Wen-Xu Wang,
Boyu Zhang, Zi-You Gao, and Angel Snchez. Equal sta-
tus in ultimatum games promotes rational sharing. Sci-
[23] Jeffrey C. Schank, Paul E. Smaldino, and Matt L. Miller.
Evolution of fairness in the dictator game by multilevel
selection. Journal of Theoretical Biology, 382:64 – 73,
2015.
[24] Jean-Baptiste Andr and Nicolas Baumard. The evolution
of fairness in a biological market. Evolution, 65(5):1447–
1456, 2011.
[25] J. A. List. On the interpretation of giving in dictator
games. Journal of Political Economy, 115(3):482–493,
2007.
[26] B Kozma and Allain Barrat. Consensus formation on
adaptive networks. Physical Review E, 77:016102, 2008.
[27] Gerardo Iniguez, J´anos Kert´esz, Kimmo Kaski, and
Rafael Barrio. Opinion and community formation in co-
evolving networks. Physical Review E, 80:066119, 2009.
[28] M. Piraveenan, M. Prokopenko, and A. Y. Zomaya. Lo-
cal assortativeness in scale-free networks. Europhysics
Letters, 84(2):28002, 2008.
[29] C. Gini. Variabilit`a e mutabilit`a: contributo allo stu-
dio delle distribuzioni e delle relazioni statistiche. [-.].
Number Part 1 in Studi economico-giuridici pubblicati
per cura della facolt`a di Giurisprudenza della R. Univer-
sit`a di Cagliari. Tipogr. di P. Cuppini, 1912.
|
1801.05479 | 2 | 1801 | 2018-11-05T20:57:21 | Belief Control Strategies for Interactions over Weakly-Connected Graphs | [
"cs.MA",
"cs.SI"
] | In diffusion social learning over weakly-connected graphs, it has been shown recently that influential agents shape the beliefs of non-influential agents. This paper analyzes this mechanism more closely and addresses two main questions. First, the article examines how much freedom influential agents have in controlling the beliefs of the receiving agents, namely, whether receiving agents can be driven to arbitrary beliefs and whether the network structure limits the scope of control by the influential agents. Second, even if there is a limit to what influential agents can accomplish, this article develops mechanisms by which they can lead receiving agents to adopt certain beliefs. These questions raise interesting possibilities about belief control over networked agents. Once addressed, one ends up with design procedures that allow influential agents to drive other agents to endorse particular beliefs regardless of their local observations or convictions. The theoretical findings are illustrated by means of examples. | cs.MA | cs | Belief Control Strategies for Interactions
over Weakly-Connected Graphs
Hawraa Salami, Student Member, IEEE, Bicheng Ying, Student Member, IEEE,
and Ali H. Sayed, Fellow, IEEE
1
8
1
0
2
v
o
N
5
]
A
M
.
s
c
[
2
v
9
7
4
5
0
.
1
0
8
1
:
v
i
X
r
a
Abstract -- In diffusion social learning over weakly-connected
graphs, it has been shown recently that influential agents shape
the beliefs of non-influential agents. This paper analyzes this
mechanism more closely and addresses two main questions. First,
the article examines how much freedom influential agents have
in controlling the beliefs of the receiving agents, namely, whether
receiving agents can be driven to arbitrary beliefs and whether
the network structure limits the scope of control by the influential
agents. Second, even if there is a limit to what influential agents
can accomplish, this article develops mechanisms by which they
can lead receiving agents to adopt certain beliefs. These questions
raise interesting possibilities about belief control over networked
agents. Once addressed, one ends up with design procedures that
allow influential agents to drive other agents to endorse particular
beliefs regardless of their local observations or convictions. The
theoretical findings are illustrated by means of examples.
Index Terms -- Social networks, diffusion learning, influential
agents, leader-follower relation, belief control, weak graph.
I. INTRODUCTION AND MOTIVATION
Several studies have examined the propagation of information over
social networks and the influence of the graph topology on this
dynamics [2] -- [28]. In recent works [27] -- [29], an intriguing phe-
nomenon was revealed whereby it was shown that weakly-connected
graphs enable certain agents to control the opinion of other agents to
great degree, irrespective of the observations sensed by these latter
agents. For example, agents can be made to believe that it is "raining"
while they happen to be observing "sunny conditions". Weak graphs
arise in many contexts, including in popular social platforms like
Twitter and similar online tools. In these graphs, the topology consists
of multiple sub-networks where at least one sub-network (called a
sending sub-network) feeds information in one direction to other
network components without receiving back (or being interested in)
any information from them. For example, a celebrity user in Twitter
may have a large number of followers (running into the millions),
while the user himself may not be tracking or following any (or
only a small fraction) of these users. For such networks with weak
graphs, it was shown in [28], [29] that, irrespective of the local
observations sensed by the receiving agents, a sending sub-network
plays a domineering role and influences the beliefs of the other groups
in a significant manner. In particular, receiving agents can be made
to arrive at incorrect inference decisions; they can also be made to
disagree on their inferences among themselves.
The purpose of this article is to examine this dynamics more
closely and to reveal new critical properties, including the develop-
ment of control mechanisms. We have three main contributions. First,
we show that the internal graph structure connecting the receiving
A short version of this work appears in the conference publication [1].
This work was supported in part by NSF grants CCF-1524250 and ECCS-
1407712. A. H. Sayed is with the ´Ecole Polytechnique F´ed´erale de Lausanne,
EPFL, School of Engineering, CH-1015 Lausanne, Switzerland. H. Salami
and B. Ying are with the Department of Electrical Engineering, University of
California, Los Angeles, CA 90025. Emails: {hsalami, ybc}@ucla.edu and
[email protected]
agents imposes a form of resistance to manipulation, but only to a
certain degree. Second, we characterize the set of states that can be
imposed on receiving networks; while this set is large, it turns out that
it is not unlimited. And, third, for any attainable state, we develop a
control mechanism that allows sending agents to force the receiving
agents to reach that state and behave in that manner.
A. Weakly-Connected Graphs
We start the exposition by reviewing the structure of weak graphs
from [27] -- [29] and by introducing the relevant notation. As explained
in [27], a weakly-connected network consists of two types of sub-
networks: S (sending) sub-networks and R (receiving) sub-networks.
Each individual sub-network is a connected graph where any two
agents are connected by a path. In addition, every sending sub-
network is strongly-connected, meaning that at least one of its agents
has a self-loop. The flow of information between S and R sub-
networks is asymmetric, as it only happens in one direction from S
to R. Figure 1 shows one example of a weakly-connected network.
The two top sub-networks are sending sub-networks and the two
bottom sub-networks are receiving sub-networks. The weights on the
connections from S to R networks are positive but can be arbitrarily
small. Observe how links from S−subnetworks to R−subnetworks
flow in one direction only, while all other links can be bi-directional.
Fig. 1. An example of a weakly connected network. The two sub-networks
on top are S−type, while the two sub-networks in the bottom are R−type.
Observe how links from S−networks to R−networks flow in one direction
only, while all other links can be bi-directional.
We
index the
strongly-connected sub-networks by s =
{1, 2,··· , S}, and the receiving sub-networks by r = {S +
1, . . . , S + R}. Each sub-network s has Ns agents, and the total
number of agents in the S sub-networks is denoted by NgS. Similarly,
each sub-network r has Nr agents, and the total number of agents
in the R sub-networks is denoted by NgR. We let N denote the total
number of agents across all sub-networks, i.e., N = NgS + NgR,
2
1) The first source consists of observational signals {ξk,i} stream-
ing in locally at agent k. These signals are generated according
to some known likelihood function parametrized by the true
state of agent k. We denote the likelihood function by Lk(.θ◦
r )
if agent k belongs to receiving sub-network r or Lk(.θ◦
s ) if
agent k belongs to sending sub-network s.
2) The second source consists of information received from
the neighbors of agent k, denoted by Nk. Agent k and
its neighbors are connected by edges and they continuously
communicate and share their opinions.
Using these two pieces of information, each agent k then updates
its belief according to the following diffusion social learning rule [2]:
ψk,i(θ) =
µk,i(θ) =
(cid:80)
µk,i−1(θ)Lk(ξk,iθ)
(cid:88)
θ(cid:48)∈Θ µk,i−1(θ(cid:48))Lk(ξk,iθ(cid:48))
a(cid:96)k ψ(cid:96),i(θ)
(cid:96)∈Nk
(4)
In the first step of (4), agent k updates its belief, µk,i−1(θ), based
on its observed private signal ξk,i by means of the Bayesian rule and
obtains an intermediate belief ψk,i(θ). In the second step, agent k
learns from its social neighbors through cooperation.
A consensus-based strategy can also be employed in lieu of (4),
as was done in the insightful works [3], [30], although the latter
reference focuses mainly on the problem of pure averaging and not
on social learning and requires the existence of certain anchor nodes.
In this work, we assume all agents are homogeneous and focus on
the diffusion strategy (4) due to its enhanced performance and wider
stability range, as already proved in [2] and further explained in the
treatments [31], [32]. Other models for social learning can be found
in [4], [5], [7], [12], [18], [33], [34].
When agents of sending sub-networks follow this model, they can
learn their own true states. Specifically, it was shown in [2], [29] that
i→∞ µk,i(θ
lim
◦
s ) a.s.= 1
(5)
for any agent k that belongs to sending sub-network s. Result (5)
means that the probability measure concentrates at location θ◦
s, while
all other possibilities in Θ have zero probability. On the other hand,
agents of receiving sub-networks will not be able to find their true
states. Instead, their beliefs will converge to a fixed distribution
defined over the true states of the sending sub-networks as follows
[29]. First, let
µks(1),i(θ)
µks(2),i(θ)
...
µks(Ns),i(θ)
and use N = {1, 2,··· , N} to refer to the indexes of all agents.
We assign a pair of non-negative weights, {ak(cid:96), a(cid:96)k}, to the edge
connecting any two agents k and (cid:96). The scalar a(cid:96)k represents the
weight with which agent k scales data arriving from agent (cid:96) and,
similarly, for ak(cid:96). We let Nk denote the neighborhood of agent k,
which consists of all agents connected to k. Each agent k scales
data arriving from its neighbors in a convex manner, i.e., the weights
satisfy:
a(cid:96)k ≥ 0,
a(cid:96)k = 1,
a(cid:96)k = 0 if (cid:96) /∈ Nk
(1)
(cid:88)
(cid:96)∈Nk
Following [27], [29], and without loss in generality, we assume that
the agents are numbered such that the indexes of N represent first
the agents from the S sub-networks, followed by those from the
R sub-networks. In this way, if we collect the {a(cid:96)k} into a large
N × N combination matrix A, then this matrix will have an upper
block-triangular structure of the following form:
(cid:122)
(cid:123)
Sub−networks:S+1,S+2,...,S+R
Sub−networks:1,2,...,S
(cid:125)(cid:124)
(cid:125)(cid:124)
(cid:123)
(cid:122)
A1
0
...
0
0
0
...
0
0
A2
...
0
0
0
...
0
0
0
. . .
. . .
...
...
. . . AS
0
. . .
. . .
0
...
. . .
...
0
...
A1,S+2
A2,S+2
A1,S+1
A2,S+1
...
AS,S+1
AS+1 AS+1,S+2
...
AS,S+2
AS+2,S+1 AS+2
...
AS+R,S+1 AS+R,S+2
...
AS,S+R
A1,S+R
A2,S+R
. . .
. . .
...
. . .
. . . AS+1,S+R
. . . AS+2,S+R
...
. . .
AS+R
...
(2)
The matrices {A1,··· , AS} on the upper left corner are left-
stochastic primitive matrices corresponding to the S strongly-
connected sub-networks. Likewise, the matrices {AS+1,··· , AS+R}
in the lower right-most block correspond to the internal weights of
the R sub-networks. We denote the block structure of A in (2) by:
(cid:20) TSS
0
A ∆=
(cid:21)
TSR
TRR
Notation: We use lowercase letters to denote vectors, uppercase
letters for matrices, plain letters for deterministic variables, and
boldface for random variables. We also use (.)T for transposition,
(.)−1 for matrix inversion, and (cid:22) and (cid:23) for vector element-wise
comparisons.
(3)
In order to characterize the set of attainable states, and to design
mechanisms for belief control over weak graphs, we need to summa-
rize first the main finding from [29]. The work in that reference
revealed the limiting states that are reached by receiving agents
over weak-graphs. An expression was derived for these states. Once
we review that expression, we will then examine its implications
closely. In particular, we will conclude from it that not all states are
attainable and that receiving sub-networks have an inherent resistance
mechanism. We characterize this mechanism analytically. We then
show how sending sub-networks can exploit
this information to
control the beliefs of receiving agents and to sow discord among
the agents.
Thus, following [29], we assume that each sub-network is observ-
ing data that arise from a true state value, denoted generically by
θ◦, which may differ from one sub-network to another. We denote
by Θ the set of all possible states, by θ◦
s the true state of sending
sub-network s and by θ◦
r the true state of receiving sub-network r,
where both θ◦
r are in Θ. At each time i, each agent k will
possess a belief µk,i(θ), which represents a probability distribution
over θ ∈ Θ. Agent k continuously updates its belief according to
two information sources:
s and θ◦
II. DIFFUSION SOCIAL LEARNING
i (θ) ∆=
µs
(6)
collect all beliefs from agents that belong to sub-network s, where
the notation ks(n) denotes the index of the n-th agent within sub-
network s, i.e.,
s−1(cid:88)
v=1
ks(n) =
Nv + n
and n ∈ {1, 2,··· , Ns}. Likewise, let
(7)
(8)
collect all beliefs from agents that belong to sub-network r, where
the notation kr(n) denotes the index of the n-th agent within sub-
network r, i.e.,
kr(n) = NgS +
Nv + n
(9)
i (θ) ∆=
µr
µkr (1),i(θ)
µkr (2),i(θ)
...
µkr (Nr ),i(θ)
r−1(cid:88)
v=S+1
and n ∈ {1, 2,··· , Nr}. Furthermore, let
i (θ)
...
µS
i (θ)
µS,i(θ) ∆=
collect all beliefs from all S−type sub-networks. Likewise, let
µ1
µS+1
...
µS+R
i
i
(θ)
(θ)
µR,i(θ) ∆=
collect the beliefs from all R−type sub-networks. Note that these
belief vectors are evaluated at a specific θ ∈ Θ. Then, the main
result in [28], [29] shows that, under some reasonable technical
assumptions, it holds that
i→∞ µR,i(θ) = W T(cid:16)
lim
i→∞ µS,i(θ)
lim
(cid:17)
where W is the NgS × NgR matrix given by:
W ∆= TSR(I − TRR)
−1
and I is the identity matrix of size NgR. The matrix W has non-
negative entries and the sum of the entries in each of its columns
is equal to one [27]. Expression (12) shows how the beliefs of the
sending sub-networks determine the limiting beliefs of the receiving
sub-networks through the matrix W . We can expand (12) to reveal
the influence of the sending networks more explicitly as follows.
k denote the row in W T that corresponds to receiving agent
k and partition it into sub-vectors as follows1:
(14)
where the {N1, N2, . . . , NS} are the number of agents in each sub-
network s ∈ {1, 2, . . . , S}. Then, according to (12), we have
k,NS
wT
wT
k,N2
k,N1
. . .
wT
Let wT
k =(cid:2) wT
i→∞ µk,i(θ) =(cid:2) wT
lim
k,N1 wT
k,N2
. . . wT
k,NS
lim
i→∞ µS,i(θ)
(15)
Note that this relation is for a specific θ ∈ Θ. Let us focus on the
case when θ = θ◦
s, assuming it is the true state parameter of the s-th
sending network only. We know from [2] and (5) that each agent
in the sending sub-network s will learn its true state θ◦
s. Therefore,
from (10),
(cid:3)
(cid:3)(cid:16)
(cid:17)
lim
i→∞ µS,i(θ
◦
s ) =
(16)
(10)
(11)
(12)
(13)
where
eθ,θ◦
s
∆=
(cid:26) 1Ns ,
◦
θ = θ
s
0Ns , otherwise
if
3
(19)
Result (18) means that the belief of receiving agent k will converge to
a distribution defined over the true states of the sending sub-networks,
which we collect into the set:
• ∆= {θ
Θ
◦
1 , θ
◦
2 , . . . , θ
S}
◦
(20)
Expression (12) shows how the limiting distributions of the sending
sub-networks determine the limiting distributions of the receiving
sub-networks through the matrix W T. In other words, it indicates how
influential agents (from within the sending sub-networks) can control
the steady-state beliefs of receiving agents. Two critical questions
arise at this stage: (a) first, how much freedom do influential agents
have in controlling the beliefs of the receiving agents? That is, can
receiving agents be driven to arbitrary beliefs or does the network
structure limit the scope of control by the influential agents? and
(b) second, even if there is a limit to what influential agents can
accomplish, how can they ensure that receiving agents will end up
with particular beliefs?
Questions (a) and (b) raise interesting possibilities about belief
(or what we will sometimes refer to as "mind") control. In the next
sections, we will address these questions and we will end up with
the conditions that allow influential agents to drive other agents to
endorse particular beliefs regardless of their local observations (or
"convictions").
III. BELIEF CONTROL MECHANISM
Observe from expression (18) that the limiting beliefs of receiving
agents depend on the columns of W = TSR (I − TRR)
−1. Note
also that the entries of W are determined by the internal combi-
nation weights within the receiving networks (i.e., TRR), and the
combination weights from the S to the R sub-networks (i.e., TSR).
The question we would like to examine now is that given a set of
desired beliefs for the receiving agents, is this set always attainable?
Or does the internal structure of the receiving sub-networks impose
limitations on where their beliefs can be driven to? To answer this
useful question, we consider the following problem setting. Let qk(θ)
denote some desired limiting distribution for receiving agent k (i.e.,
qk(θ) denotes what we desire the limiting distribution µk,i(θ) in
(18) to become as i → ∞). We would like to examine whether it is
possible to force agent k to converge to any qk(θ), i.e., whether it is
possible to find a matrix TSR so that the belief of receiving agent k
converges to this specific qk(θ).
A. Motivation
In this first approach, we are interested in designing TSR while
TRR is assumed fixed and known. This scenario allows us to under-
stand in what ways the internal structure of the receiving networks
limits the effect of external influence by the sending sub-networks.
This approach also allows us to examine the range of belief control
over the receiving sub-networks (i.e., how much freedom the sending
sub-networks have in selecting these beliefs). Note that the entries
of TSR correspond to weights by which the receiving agents scale
information from the sending sub-networks. These weights are set by
the receiving agents and, therefore, are not under the direct control
of the sending sub-networks. As such, it is fair to question whether
it is useful to pursue a design procedure for selecting TSR since
its entries are not under the direct control of the designer or the
sending sub-networks. The useful point to note here, however, is that
the entries of TSR, although set by the receiving agents, can still be
interpreted as a measure of the level of trust that receiving agents have
in the sending agents they are connected to. The higher this level of
confidence is between two agents, the larger the value of the scaling
weight on the link connecting them. In many applications, these
levels of confidence (and, therefore, the resulting scaling weights)
can be influenced by external campaigns (e.g., through advertisement
0N1
0N2
...
1Ns
...
0NS
where 1Ns denotes a column vector of length Ns whose elements
are all one. Similarly, 0N s denotes a column vector of length Ns
whose elements are all zero. Combining (15) and (16) we get
lim
i→∞ µk,i(θ
◦
s ) = wT
k,Ns
1Ns
(17)
This means that the likelihood of state θ◦
s at the receiving agent
k is equal to the sum of the entries of the weight vector, wk,Ns,
corresponding to sub-network s. More generally, for any other state
parameter θ ∈ Θ, its likelihood is given from (12) by
lim
i→∞ µk,i(θ) =
wT
k,Ns eθ,θ◦
s
(18)
S(cid:88)
s=1
1The index of the row in W T that corresponds to agent k is k − NgS.
or by way of reputation). In this way, we can interpret the problem
of designing TSR as a way to guide the campaign that influences
receiving agents to set their scaling weights to desirable values. The
argument will show that by influencing and knowing TSR, sending
agents end up controlling the beliefs of receiving agents in desirable
ways. For the analysis in the sequel, note that by fixing TRR and
designing TSR, we are in effect fixing the sum of each column of
TSR and, accordingly, fixing the overall external influence on each
receiving agent. In this way, the problem of designing TSR amounts
to deciding on how much influence each individual sub-network
should have in driving the beliefs of the receiving sub-networks.
B. Conditions for Attainable Beliefs
1 , θ◦
2 , . . . , θ◦
Given these considerations,
let us now show how to design
TSR to attain certain beliefs. As is already evident from (18),
the desired belief qk(θ) at any agent k needs to be a probability
distribution defined over the true states of all sending sub-networks,
Θ• = {θ◦
loss of generality,
that the true states of the sending sub-networks are distinct, so that
Θ• = S. If two or more sending sub-networks have the same true
state, we can merge them together and treat them as corresponding
to one sending sub-network; although this enlarged component is not
necessarily connected, it nevertheless consists of strongly-connected
elements and the same arguments and conclusions will apply.
S}. We assume, without
We collect the desired limiting beliefs for all receiving agents into
the vector:
qR(θ) ∆=
(21)
(cid:16)
qNgS +1(θ)
qNgS +2(θ)
...
qN (θ)
(cid:17)T
which has length NgR. Then, from (12), we must have:
qTR(θ) =
lim
i→∞ µS,i(θ)
W
Evaluating this expression at the successive states {θ◦
1 , θ◦
we get
(cid:124)
(cid:125)
=
(cid:124)
qTR(θ◦
1 )
qTR(θ◦
2 )
...
(cid:123)(cid:122)
qTR(θ◦
S)
∆
= QS×NgR
1T
N1
0
...
0
1T
N2
...
0
0
0
0
. . .
. . .
...
...
. . . 1T
NS
(cid:123)(cid:122)
∆
= ES×NgS
(22)
S},
2 , . . . , θ◦
W
(23)
(cid:125)
where Q is the S × NgR matrix that collects the desired beliefs for
all receiving agents. Using (13), we rewrite (23) more compactly in
matrix form as:
E TSR = Q (I − TRR)
(24)
Therefore, given Q and TRR, the design problem becomes one of
finding a matrix TSR that satisfies (24) subject to the following
constraints:
1TTSR + 1TTRR = 1T
TSR (cid:60) 0
tSR,k(j) = 0, if receiving agent k is not
connected to sending agent j
(25)
(26)
(27)
The first condition (25) is because the entries on each column of A
defined in (3) add up to one. The second condition (26) ensures
that each element of TSR is a non-negative combination weight.
The third condition (27) takes into account the network structure,
where tSR,k represents the column of TSR that corresponds to
receiving agent k, and tSR,k(j) represents the jth entry of this
column (which corresponds to sending agent j -- see Fig. 2). In other
words, if receiving agent k is not connected to sending agent j, the
corresponding entry in TSR should be zero.
4
Fig. 2. An illustration of the k−th column of TSR and the j−th entry on
that column.
It is useful to note that condition (25) is actually unnecessary and
can be removed. This is because if we can find TSR that satisfies
(24), then condition (25) will be automatically satisfied. To see this,
we first sum the elements of the columns on the left-hand side of
(24) and observe that
1T
SETSR = 1TTSR
(28)
We then sum the elements of the columns on the right-hand side of
(24) to get
(29)
S (Q − QTRR) = 1T − 1TTRR
1T
SQ = 1T
NgR since the entries on each column of Q
This is because 1T
add up to one. Thus, equating (28) and (29), we find that (25) must
hold. The problem we are attempting to solve is then equivalent to
finding TSR that satisfies (24) subject to
TSR (cid:60) 0
tSR,k(j) = 0, if receiving agent k is not
connected to sending agent j
(30)
(31)
To find TSR that satisfies (24) under the constraints (30)-(31), we
can solve separately for each column of TSR. Let tRR,k and qk,
respectively, denote the columns of TRR and Q that correspond
to receiving agent k. Then, relations (24) and (30) -- (31) imply that
column tSR,k must satisfy:
EtSR,k = qk − QtRR,k
subject to
tSR,k (cid:60) 0
tSR,k(j) = 0, if receiving agent k is not
connected to sending agent j
(32)
(33)
(34)
The problem is then equivalent to finding tSR,k for each receiving
agent k such that tSR,k satisfies (32)-(34). For Q to be attainable
(i.e., for the beliefs of all receiving agents to converge to the
desired beliefs), finding such tSR,k should be possible for each
receiving agent k. However, finding tSR,k that satisfies (32) under
the constraints (33)-(34) may not be always possible. The desired
belief matrix Q will need to satisfy certain conditions so that it is
not possible to drive the receiving agents to any belief matrix Q.
Before stating these conditions, we introduce two auxiliary matrices.
We define first the following difference matrix, which appears on the
right-hand side of (24) -- this matrix is known:
V ∆= Q(I − TRR)
(35)
Note that V has dimensions S×NgR. The k−th column of V , which
we denote by vk appears on the right-hand side of (32), i.e.,
vk = qk − QtRR,k
(36)
The (s, k)−th entry of V is then:
NgR(cid:88)
vk(s) = qk(θ
s ) −
◦
tRR,k((cid:96))qNgS +(cid:96)(θ
◦
s )
(37)
(cid:96)=1
Each (s, k)−th entry of V represents the difference between the
desired limiting belief at θ◦
s of receiving agent k and a weighted
combination of the desired limiting beliefs of its neighboring receiv-
ing agents. We remark that this sum includes agent k if tRR,k(k)
is not zero. Similarly, it includes any receiving agent (cid:96) if tRR,k((cid:96))
is not zero. In this way, the sum runs only over the neighbors of
agent k, because any agent (cid:96) that is not a neighbor of agent k has
its corresponding entry in tRR,k as zero.
Let C denote an S × NgR binary matrix, with as many rows as
the number of sending sub-networks and as many columns as the
number of receiving agents. The matrix C is an indicator matrix that
specifies whether a receiving agent is connected or not to a sending
sub-network. The (s, k)−th entry of C is one if receiving agent k
is connected to sending sub-network s; otherwise, it is zero. We are
now ready to state when a given set of desired beliefs is attainable.
Theorem 1. (Attainable Beliefs) A given belief matrix Q is attain-
able if, and only if, the entries of V will be zero wherever the entries
of C are zero, and the entries of V will be positive wherever the
(cid:4)
entries of C are one.
Before proving theorem 1, we first clarify its statement. For Q to
be achievable, the matrices V and C must have the same structure
with the unit entries of C translated into positive entries in V . This
theorem reveals two possible cases for each receiving agent k and
gives, for each case, the condition required for the desired beliefs to
be attainable.
In the first case, receiving agent k is not connected to any agent
of sending sub-network s (the (s, k)−th entry of C is zero). Then,
according to Theorem 1, receiving agent k achieves its desired
limiting belief qk(θ◦
s ) if, and only if,
vk(s) = qk(θ
s ) −
◦
tRR,k((cid:96))qNgS +(cid:96)(θ
◦
s ) = 0
(38)
That is, the cumulative influence from the agent's neighbors must
match the desired limiting belief.
In the second case, receiving agent k is connected to at least one
agent of sending sub-network s (the (s, k)−th entry of C is one).
Now, according to Theorem 1 again, receiving agent k achieves its
desired limiting belief qk(θ◦
s ) if, and only if,
vk(s) = qk(θ
s ) −
◦
tRR,k((cid:96))qNgS +(cid:96)(θ
◦
s ) > 0
(39)
Proof of Theorem 1: We start by first proving that if Q is
attainable, then V and C have the same structure. If Q is attainable,
then there exists tSR,k for each receiving agent k that satisfies (32)-
(34). Using the definition of E in (23), the s−th row on the left-hand
side of (32) is:
tSR,k(j)
(40)
where Is represents the set of indexes of sending agents that belong
to sending sub-network s. Expression (40) represents the sum of
the elements of the block of tSR,k that correspond to sending sub-
network s. Therefore, if Q is attainable, then the s−th row of (32)
NgR(cid:88)
(cid:96)=1
NgR(cid:88)
(cid:96)=1
(cid:88)
j∈Is
tSR,k(j) = vk(s)
(41)
satisfies the following relation:(cid:88)
agent in sub-network s, then (cid:80)
j∈Is
From this relation, we see that if agent k is not connected to any
tSR,k(j) = 0 which implies
that vk(s) is zero. On the other hand, if agent k is connected to sub-
j∈Is
5
network s, then(cid:80)
j∈Is
tSR,k(j) > 0 which implies that vk(s) > 0.
In other words, C and V have the same structure.
Conversely, if C and V have the same structure, then it is possible
to find tSR,k for each receiving agent k that satisfies (32)-(34).
In particular, if agent k is not connected to sub-network s, then
the (s, k)-th entry of C is zero. Since C and V have the same
structure, then vk(s) = 0 . By setting to zero the entries of tSR,k that
correspond to sending sub-network s, relation (41) is satisfied. On the
other hand, if agent k is connected to sub-network s (connected to at
least one agent in sub-network s), then the (s, k)-th entry of C is one.
Since C and V have the same structure, we get vk(s) > 0. Therefore,
since the entries of tSR,k must be non-negative, we first set to zero
the entries of tSR,k that correspond to agents of sub-network s that
are not connected to agent k and the remaining entries can be set to
non-negative values such that relation (41) is satisfied. That is, if C
and V have the same structure, then Q is attainable.
We next move to characterize the set of solutions, i.e., how we
can design tSR,k assuming the conditions on V are met.
C. Characterizing the Set of Possible Solutions
In the sequel, we assume that the conditions on V from Theorem
1 are satisfied. That is, if receiving agent k is not connected to sub-
network s, then vk(s) = 0. Otherwise, vk(s) > 0. The desired beliefs
are then attainable. This means that for each receiving agent k, we
can find tSR,k that satisfies (32)-(34). Many solutions may exist. In
this section, we characterize the set of possible solutions.
First of all, to meet (31), we set the required entries of tSR,k to
zero. We then remove the corresponding columns of E, and label the
reduced E by Ek. Similarly, we remove the zero elements of tSR,k
and label the reduced tSR,k by t(cid:48)
SR,k. On the other hand, if agent k is
not connected to some sub-network s, then the corresponding row in
E will be removed and Ek will have fewer number of rows, denoted
by S(cid:48). Without loss of generality, we assume agent k is connected
to the first S(cid:48) sending sub-networks. We denote by N k
s the number
of agents of sending sub-network s that are connected to receiving
gS the total number of all sending agents connected
agent k and by N k
to agent k. The matrix Ek will then have the form (this matrix is
obtained from E by removing rows and columns with zero entries;
the resulting dimensions are now denoted by S(cid:48) and N k
gS):
Ek =
1T
N k
1
0
...
0
0
1T
N k
2
...
0
. . .
0
0
. . .
...
...
. . . 1T
N k
S(cid:48)
(42)
S(cid:48)×N k
gS
Note that
if receiving agent k is connected to all sending sub-
networks, then E and Ek will have the same number of rows, S(cid:48) = S.
In the case where agent k is not connected to some sub-network
s, condition (38) should be satisfied, and the corresponding row
in qk − QtRR,k should be removed to obtain the reduced vector
k − Q(cid:48)tRR,k. We are therefore reduced to determining t(cid:48)
q(cid:48)
SR,k by
solving a system of equations of the form:
k − Q
(cid:48)
(cid:48)
(cid:48)
Ekt
SR,k = q
tRR,k
(43)
subject to
(cid:48)
SR,k (cid:60) 0
t
(44)
We can still have some of the entries of the solution t(cid:48)
SR,k turn out
to be zero. Now note that the number of rows of Ek is S(cid:48) (number of
sending sub-networks connected to k), which is always smaller than
or equal to N k
gS. Moreover, the rows of Ek are linearly independent
and thus Ek is a right-invertible matrix. Its right-inverse is given by
[35]:
(cid:16)
(cid:17)−1
ER
k = ET
k
EkET
k
(45)
Therefore, if we ignore condition (44) for now, then equation (43)
has an infinite number of solutions parametrized by the expression
[35]:
(cid:48)
SR,k = ER
k
t
k − Q
(cid:48)
(cid:48)
tRR,k
k Ek)y
(46)
(cid:1) + (I − ER
(cid:0)q
where y is an arbitrary vector of length N k
condition (44). Let
gS. We still need to satisfy
(cid:48)
v
k
and note that
∆= q
k − Q
(cid:48)
(cid:48)
tRR,k
(47)
(48)
ER
k v
(cid:48)
k =
(cid:16)
v(cid:48)
k(1)
N k
v(cid:48)
1
k(2)
N k
2
1
1
N k
1
N k
2
1
...
k(S(cid:48))
v(cid:48)
N k
S(cid:48)
(cid:17)−1
N k
S(cid:48)
where v(cid:48)
k(i) represents the ith entry of vector v(cid:48)
k. Likewise,
I − ET
k
EkET
k
Ek =
(cid:26)
diag
IN k
1
− 1
N k
1
1
1T
N k
1
N k
1
, IN k
2
− 1
N k
2
S(cid:48) − 1
IN k
N k
S(cid:48)
1
1T
N k
2
N k
2
,···
(cid:27)
1
N k
S(cid:48)
1T
N k
S(cid:48)
and if we partition y into sub-vectors as
While if we sum the elements of the vector on the right-hand side
of (53), we obtain:
6
− v(cid:48)
k(s)
N k
s
This means that we cannot find t(cid:48)
of the entries of v(cid:48)
k or q(cid:48)
1
N k
s
N k
s
1T
= −v
(cid:48)
k(s) > 0
SR,k such that t(cid:48)
k − Q(cid:48)tRR,k is negative.
(55)
SR,k (cid:23) 0 when any
In summary, we have established the validity of the following
statement.
Theorem 2. Assume receiving agent k is connected to N k
s agents
in sending sub-network s. If vk(s) > 0, then all possible choices for
the weights from sending agents in network s to receiving agent k
are parameterized as:
v(cid:48)
k(s)
N k
s
1
N k
s
+
IN k
s
− 1
N k
s
1
1T
N k
s
N k
s
yN k
s
(56)
(cid:18)
(cid:19)
where yN k
stays non-negative.
s
is an arbitrary vector of length N k
s chosen so that (56)
(cid:4)
D. Enforcing Uniform Beliefs
(49)
In this section, we explore one special case of attainable beliefs,
which is driving all receiving agents towards the same belief. In this
case, Q is of the following form:
Q = q1T
NgR
(57)
(50)
for some column q that represents the desired limiting belief (the
entries of q are non-negative and add up to one). We verify that
the conditions that ensure that uniform beliefs are attainable by all
receiving agents. In this case, vk is of the following form:
y =
+
1
yN k
yN k
2...
yN k
S(cid:48)
(cid:16)
(cid:16)
(cid:18)
then expression (46) becomes:
1
1
N k
1
N k
2
v(cid:48)
k(1)
N k
v(cid:48)
1
k(2)
N k
2
...
v(cid:48)
k(S(cid:48))
N k
S(cid:48)
(cid:48)
SR,k =
t
IN k
1
IN k
2
− 1
− 1
N k
1
N k
2
1T
N k
1
1T
N k
2
1
1
N k
1
N k
2
...
(cid:17)
(cid:17)
(cid:19)
yN k
1
yN k
2
S(cid:48) − 1
IN k
1
1
1T
N k
S(cid:48)
yN k
S(cid:48)
(51)
This represents the general form of all possible solutions, but from
these solutions we want only those which are nonnegative in order
to satisfy condition (44). From (51), the vector t(cid:48)
SR,k is partitioned
into multiple blocks, where each block has the form:
N k
S(cid:48)
N k
S(cid:48)
N k
S(cid:48)
v(cid:48)
k(s)
N k
s
1
N k
s
+
IN k
s
− 1
N k
s
1
N k
s
1T
N k
s
yN k
s
(52)
(cid:18)
(cid:19)
We already have from the conditions of attainable beliefs (39) that
v(cid:48)
as zero or set it to arbitrary
k(s) > 0. Therefore, we can choose yN k
values as long as (52) stays non-negative. We also know that for the
beliefs to be attainable, we cannot have v(cid:48)
k(s) < 0. Otherwise, no
solution can be found. Indeed, if v(cid:48)
k(s) < 0, then to make (52) non-
negative, we would need to select yN k
s
IN k
s
− 1
N k
s
1
1T
N k
s
N k
s
1
N k
s
(53)
However, there is no yN k
that satisfies this relation because if we
sum the elements of the vector on the left-hand side of (53), we
obtain:
s
(cid:19)
such that:
(cid:23) − v(cid:48)
k(s)
N k
s
s
yN k
s
(cid:18)
(cid:19)
(cid:18)
vk = qk − QtRR,k = (1 − 1T
NgR tRR,k)q
and the (s, k)-th entry of V is:
vk(s) = (1 − 1T
NgR tRR,k)q(θ
◦
s )
(58)
(59)
Now we know that 1 − 1T
NgR tRR,k > 0 when agent k is connected
to at least one agent from any sending sub-network, and that 1 −
NgR tRR,k = 0 when it is not connected to any sending sub-network.
1T
In the second case where 1−1T
NgR tRR,k = 0, expression (59) implies
that vk(s) = 0 for any s. Therefore, in this case, we have agent k not
connected to any sending sub-network s and vk(s) = 0 for any s, and
condition (38) is satisfied. In the first case where 1−1T
NgR tRR,k > 0
(i.e., agent k is connected to some sending sub-networks but not
necessarily to all of them), expression (59) implies that vk(s) >
0 no matter whether agent k is connected or not to sending sub-
network s. However, when agent k is not connected to sending sub-
network s, condition (38) requires that vk(s) = 0 for agent k to
achieve its desired belief at θ◦
s. In summary, we arrive at the following
conclusion.
Lemma 1. For the scenario of uniform beliefs to be attainable, agent
k should be connected either to all sending sub-networks or to none
of them.
We provide in reference [arXiv] two numerical examples that illus-
trate this construction.
E. Example 1
1T
N k
s
IN k
s
− 1
N k
s
1
1T
N k
s
N k
s
yN k
s
= 0
(54)
Consider the network shown in Fig. 3. It consists of N = 8 agents,
two sending sub-networks and one receiving sub-network, with the
following combination matrix:
A =
0.2
0.5
0.3
0
0
0
0
0
0.2
0.4
0.4
0
0
0
0
0
0.8
0.1
0.1
0
0
0
0
0
0
0
0
0.4
0.6
0
0
0
0
0
×
0
×
0
0.3 ×
0
0.7
0.2
0
0
0.1
0.1
0
0
0
×
0
×
0.3
0.2
0.2
×
0
0
×
0
0.2
0.3
0.1
so that
(cid:21)
(cid:20) v7(1)
v7(2)
(cid:21)
(cid:20) 0.06
0.24
=
(cid:48)
SR,7 =
t
7
(67)
(60)
The entries t(cid:48)
and 5 that are linked to agent 7.
SR,7 represent respectively the coefficients of agents 3
Next for agent 8, we have
v8 = 0.4q =
so that
t
(cid:48)
SR,8 =
Therefore, one possible solution is
TSR =
0.32
(cid:21)
(cid:20) 0.08
(cid:20) 0.08
(cid:21)
=
0.32
0
0
0.06
0
0.24
0.08
0
0
0.32
0
v8(2)
(cid:20) v8(1)
0.06
0.06
0.48
0
0
(68)
(69)
(70)
(cid:21)
We determine the columns of TSR one at a time. Starting with agent
6, we focus on the first column of TSR. The vector v6 defined in
(47) is given by (58) for the case of uniform beliefs. Therefore,
lim
i→∞ µk,i(θ
◦
1 ) =
Fig. 3. A weakly connected network consisting of three sub-networks
in a broadband influence scenario.
We assume that there are 3 possible states Θ = {θ◦
1 , θ◦
θ◦
1 is the true event for the first sending sub-network, θ◦
event for the second sending sub-network, and θ◦
for the receiving sub-network.
converge to the same belief over {θo
2}, say:
1, θo
Let us first design TSR so that all receiving agents' beliefs
3}, where
2 , θ◦
2 is the true
3 is the true event
(cid:21)
(cid:20) 0.2
0.8
q =
(cid:20) 0.12
(cid:21)
(cid:20) I2 − 1
0.48
v6 = 0.6q =
(cid:20) v6(1)
(cid:21)
(cid:21)
y
(62)
(63)
Thus, according to (51),
(cid:48)
SR,6 =
t
12
2
v6(2)
+
121T
2
2
111T
1
1 − 1
1
where y is an arbitrary vector of length 3. Note that t(cid:48)
SR,6 represents
respectively the coefficients of agents 2, 3 and 4 that are linked to
agent 6. It follows that
0.06
0.06
0.48
(cid:48)
SR,6 =
t
2 y(1) − 1
− 1
2 y(1) + 1
2 y(2)
2 y(2)
+
1
0.06 + α6
0.06 − α6
0
(cid:48)
t
SR,6 =
Let α6
∆= 1
2 y(1) − 1
2 y(2) so that
To verify that the beliefs of the receiving agents converge in this case
to the desired belief, we compute the matrix W T from (13):
W T = (I − T T
RR)
0.0169
0.0322
0.1034
=
SR
−1T T
0.0839
0.0394
0.0318
0.0992
0.1284
0.0648
0.7390
0.4441
0.6678
0.0610
0.3559
0.1322
(71)
Then, according to (18), we can compute the belief at θ◦
1 for each
receiving agent at steady state, by taking the first block in the agent's
corresponding row and summing its elements:
Likewise, we can compute the belief at θ◦
2 for each receiving agent at
steady state, by taking the second block in the agent's corresponding
row and summing its elements:
k = 6
k = 7
k = 8
0.0322 + 0.0394 + 0.1284 = 0.2,
0.1034 + 0.0318 + 0.0648 = 0.2,
0.0169 + 0.0839 + 0.0992 = 0.2,
0.7390 + 0.0610 = 0.8,
(cid:21)
(cid:20) 0.8
0.4441 + 0.3559 = 0.8,
0.6678 + 0.1322 = 0.8,
◦
2 ) =
0.7
0.3
0.75
0.25
0.2
Q =
k = 6
k = 7
k = 8
(72)
Note that now the beliefs are different from an agent to another, but
they are still close. Computing,
vk = qk − QtRR,k
lim
i→∞ µk,i(θ
Let us now consider the case where we want to design TSR so that
the desired limiting beliefs are not necessarily uniform but rather
(61)
(64)
(65)
0.48
In order to have positive entries for t(cid:48)
0.06.
Now for agent 7, the vector v7 is given by
SR,6, we can choose α6 ≤
(cid:21)
(cid:20) 0.06
0.24
v7 = 0.3q =
for each receiving agent k, we obtain:
v6 = q6 − QtRR,6 =
(66)
v7 = q7 − QtRR,7 =
v8 = q8 − QtRR,8 =
(cid:21)
(cid:21)
(cid:21)
(73)
(74)
(75)
(76)
(cid:20) 0.495
(cid:20) 0.17
(cid:20) 0.305
0.105
0.13
0.195
Therefore, one possible TSR is
TSR =
0
0.495/2
0.495/2
0.105
0
0
0
0.17
0
0.13
0.305
0
0
0.195
0
(77)
Let us now consider the case where the desired limiting beliefs are
more dispersed, such as
(cid:20) 0.8
0.2
Q =
0.2
0.8
0.3
0.7
(cid:21)
(cid:20) −0.14
0.44
(cid:21)
(78)
(79)
In this case for agent 7, we have
v7 = q7 − QtRR,7 =
with a negative first entry. Therefore, the desired belief for agent 7
cannot be attained.
F. Example 2
Consider now the network shown in Fig. 4 with the following
combination matrix
A =
0.2
0.5
0.3
0
0
0
0
0
0.2
0.4
0.4
0
0
0
0
0
0.8
0.1
0.1
0
0
0
0
0
0
0
0
0.4
0.6
0
0
0
0
0
×
0
×
0
0.3 ×
0
0.7
0.2
0
0.1
0
0
0.1
0
0
×
0
×
0.3
0.2
0.2
×
×
0
0
0
0.1
0.6
0
(80)
Fig. 4. A weakly connected network consisting of three sub-networks.
Let us consider the case where we want to design TSR so that the
desired limiting beliefs are as follows:
Q =
Computing,
(cid:20) 0.8
0.2
(cid:21)
0.7
0.3
0.8
0.2
vk = qk − QtRR,k
(81)
(82)
(83)
(84)
(85)
(cid:21)
(cid:21)
(cid:21)
(cid:20) 0.49
(cid:20) 0.16
(cid:20) 0.3
0.14
0.11
0
for each receiving agent k, we obtain:
v6 = q6 − QtRR,6 =
v7 = q7 − QtRR,7 =
v8 = q8 − QtRR,8 =
Note that in this example, agent 8 is not connected to the second
sending sub-network, but
the controlling scheme can still work
because condition (38) is satisfied. Therefore, one possible choice
for TSR is the following:
TSR =
0
0.49/2
0.49/2
0.11
0
0
0
0.16
0
0.14
0.3/2
0.3/2
0
0
0
8
(86)
To verify that the beliefs of the agents converge in this case to
the desired belief, we compute W T from (13) and use (18) to
determine the limiting beliefs at θo
2 at the receiving agents.
This calculation gives
1 and θo
0.0586 + 0.2200 + 0.4214 = 0.7,
0.1883 + 0.3193 + 0.2924 = 0.8,
0.0309 + 0.3737 + 0.3954 = 0.8,
0.1539 + 0.046 = 0.2,
0.0724 + 0.2276 = 0.3,
0.0588 + 0.1412 = 0.2,
◦
2 ) =
k = 6
k = 7
k = 8
k = 6
k = 7
k = 8
lim
i→∞ µk,i(θ
◦
1 ) =
and
lim
i→∞ µk,i(θ
IV. JOINT DESIGN OF TRR AND TSR
In the previous sections, we analyzed the conditions that drive
receiving agents to desired beliefs. The approach relies on deter-
mining the entries of the weighting matrix TSR from knowledge of
Q (the desired beliefs) and TRR (the internal weighting structure
within the receiving sub-networks). We saw how there is limitation
to where the beliefs of receiving agents can converge. In particular,
the internal combination of receiving sub-networks contribute to this
limitation. We now examine the problem of designing TSR and
TRR jointly, to see whether by having more freedom in choosing
the coefficients of TRR, we still encounter limitations on how to
influence the receiving agents. We assume that we know the number
of receiving sub-networks and the number of agents in each of these
sub-networks. Using (24), we have
= Q
(87)
(cid:21)
(cid:20) TSR
TRR
(cid:2) E Q (cid:3)
(cid:125)
(cid:123)(cid:122)
(cid:124)
∆
= B
Therefore, given Q (the desired limiting beliefs of the receiving
agents), the design problem becomes one of finding matrices TSR
and TRR that satisfy (87) subject to the following constraints:
1TTSR+1TTRR = 1T
TSR,k(j) = 0, if sending agent j does not feed into k
TSR,k(j) ≥ 0, otherwise
TRR,k(j) = 0, if receiving agent j does not feed into k
TRR,k(j) > 0, otherwise
(88)
(89)
(90)
In the last condition (90), we are requiring TRR,k(j) to be strictly
positive if receiving agent j feeds into k. This is in order to avoid
solutions where the receiving sub-networks become unconnected. For
instance, consider the example shown in Fig. 5. This figure shows a
case where agent k is connected to all sending sub-networks, and it
depicts only the incoming links into agent k. Let us assume that the
desired limiting belief for agent k is
(cid:21)
(cid:20) qk(θ◦
1 )
qk(θ◦
2 )
(cid:21)
(cid:20) 0.1
0.9
=
(91)
Then a possible solution to (87) is to assign zero as weights for
the data originating from its receiving neighbors, 0.1 for the data
received from sending agent 1, and 0.9 for the data received from
sending agent 2. Then, for this example,
(cid:20) 1
(cid:20) 0.1
0
0.9
(cid:21)
0
1
(cid:21)
, Q =
, TRR =
E =
TSR =
(cid:20) 0.1
0
0.9
0
0
(cid:21)
,
q4(θ◦
1 )
q4(θ◦
2 )
q5(θ◦
1 )
q5(θ◦
2 )
(92)
so that (87) is satisfied. However, this solution affects the connected-
ness of the receiving sub-network of agent k, because there will be
no path that leads to this agent.
Fig. 5. An example where the receiving network of agent k ends up
being disconnected.
To find TSR and TRR satisfying (87)-(90), we can solve separately
for each of their columns. If it is possible to find a solution for
each column, then Q is attainable. We explore next the possibility of
finding solutions for each column. Similarly to the previous section,
tSR,k and tRR,k respectively represent the columns of TSR and TRR
that correspond to receiving agent k, and tSR,k(j) and tRR,k(j)
respectively represent the j−th entries of this tSR,k and tRR,k. Also
qk denote the column of Q that corresponds to receiving agent k.
Then, relations (87) and (89) -- (90) imply that the columns tSR,k and
tRR,k must satisfy:(cid:2) E Q (cid:3)
(cid:123)(cid:122)
(cid:125)
(cid:124)
=B
(cid:21)
(cid:20) tSR,k
tRR,k
= qk
subject to
1TtSR,k+1TtRR,k = 1
tSR,k(j) = 0, if j does not feed k
tSR,k(j) ≥ 0, otherwise
tRR,k(j) = 0, if j does not feed k
tRR,k(j) > 0, otherwise
(93)
(94)
(95)
(96)
Since the connections within the sending and receiving networks
are known, but not the combination weights TSR and TRR whose
values we are seeking, we can then set to zero the entries of tSR,k
and tRR,k that correspond to unlinked agents. We remove these zero
entries and relabel the vectors as t(cid:48)
RR,k. We also remove
the corresponding columns of E and Q, and label the modified E
and Q by Ek and Qk. We are therefore reduced to determining t(cid:48)
SR,k
and t(cid:48)
RR,k by solving a system of equations of the form:
SR,k and t(cid:48)
(cid:2) Ek Qk
(cid:124)
(cid:123)(cid:122)
∆
= Bk
(cid:20) t(cid:48)
t(cid:48)
(cid:3)
(cid:125)
(cid:21)
SR,k
RR,k
= qk
subject to
1Tt
(cid:48)
SR,k+1Tt
(cid:48)
RR,k = 1
(97)
(98)
Ek =
1T
N k
1
0
...
0
0
1T
N k
2
...
0
. . .
0
0
. . .
...
...
. . . 1T
N k
S
(104)
(cid:48)
SR,k (cid:60) 0
t
RR,k (cid:31) 0
(cid:48)
t
9
(99)
(100)
(cid:16)
a(cid:96)k
(cid:88)
(cid:96)∈Nk
(cid:88)
(cid:96)∈Nk
Formulation (97)-(100) has the following interpretation. After some
sufficient time i ≥ I, we know that the beliefs of all agents will
approach some limiting beliefs, and based on the results of the
previous work [29], the belief update (4) approaches for i ≥ I,
ψk,i+1(θ) = µk,i(θ)
(cid:88)
µk,i+1(θ) =
(cid:96)∈Nk
This means that:
lim
i→∞ µk,i(θ) =
(cid:88)
(cid:96)∈Nk
a(cid:96)k ψ(cid:96),i+1(θ) =
a(cid:96)k µ(cid:96),i(θ)
(101)
(cid:17)
(102)
lim
i→∞ µ(cid:96),i(θ)
In other words, if we want the beliefs of the receiving agents to
converge to some belief vector q, then we need to make sure that
these desired beliefs satisfy the relationship:
qk(θ) =
a(cid:96)kq(cid:96)(θ)
(103)
1 ,··· , θ◦
S} and for all receiving agents k. In other
for any θ ∈ {θ◦
words, given the set of desirable beliefs, we would like to know if
it is possible to express the desired limiting belief for each receiving
agent k as a convex combination of the limiting beliefs of its receiving
neighbors and the limiting beliefs of the sending agents to which
agent k is connected. If this is possible for each agent k, then
Q is attainable, i.e., all receiving agents can reach their desired
limiting beliefs. This is precisely what the formulation (97) -- (100) is
attempting to enforce, by finding suitable coefficients such that (103)
is satisfied. Finding t(cid:48)
RR,k that satisfy (97) and constraints
(98)-(100) might not be always possible. Since each agent k can be
connected to all sending sub-networks, or to some of them or to none
of them, the matrix Ek that appears in (97) will have a different form
for each of these cases, which will affect the possibility of finding
a solution. Before analyzing how the three possible cases affect the
possibility of finding a solution, we summarize first the results:
SR,k and t(cid:48)
2) Agent k is connected to some sending sub-networks:
1) Agent k is connected to all sending sub-networks: the problem
RR,k that satisfies (112a) and (112b, which
reduces to finding t(cid:48)
always has a solution;
problem reduces to finding t(cid:48)
(120a)-(120c), which may not always have a solution;
problem reduces to finding t(cid:48)
(127a)-(127c), which may not always have a solution.
3) Agent k is not connected to any sending sub-network: the
RR,k that satisfies conditions
the
RR,k that satisfies conditions
Note that relations (112a) and (120a) are what condition (39)
required when we wanted to design TSR, for the case where agent k is
connected to sending sub-network s, when TRR was given. Similarly,
relations (120b) and (127a) are what condition (38) required when we
wanted to design tSR,k, for the case where agent k is not connected
to sending sub-network s, when TRR was given. In the earlier section,
we had to make sure that the given TRR satisfies (38) and (39) for
Q to be attainable. Here, we are designing for TRR as well, and we
need to make sure that the entries we choose satisfy these conditions.
We now analyze each case in detail.
Case 1: Agent k is connected to all sending sub-networks
We discuss first the case where agent k is connected to at least
one agent from each sending sub-network. In this case, Ek will have
the following form:
and relation (97) is then:
1T
N k
1
0
...
0
0
1T
N k
2
...
0
. . .
0
0
. . .
...
...
. . . 1T
N k
S
qk(1)(θ◦
1 )
qk(1)(θ◦
2 )
...
qk(1)(θ◦
s )
. . .
. . .
...
. . .
gR)(θ◦
1 )
gR)(θ◦
2 )
(cid:21)
gR)(θ◦
s )
. . .
qk(N k
qk(N k
qk(N k
(cid:20) t(cid:48)
t(cid:48)
SR,k
RR,k
= qk
s ) represents the desired limiting belief at θ◦
(105)
where qk(j)(θ◦
s for the
jth receiving neighbor of agent k, and N k
gR is the total number of
receiving agents that are neighbors of agent k. The problem here is
to find t(cid:48)
RR,k that satisfy (105) subject to the constraints
(98)-(99). It is useful to note that if we can find t(cid:48)
RR,k
that satisfy (105), then condition (98) will be automatically satisfied.
To see this, we first sum the elements of the vector on the left-hand
side of (105) and observe that
SR,k and t(cid:48)
SR,k and t(cid:48)
(cid:20) t(cid:48)
(cid:21)
(cid:20) t(cid:48)
(cid:21)
1T
SBk
t(cid:48)
SR,k
RR,k
= 1T
t(cid:48)
SR,k
RR,k
(106)
The matrix B was introduced in (93). This is because 1T
SBk = 1T
since the entries on each column of Bk add up to one. We then sum
the elements of the vector on the right-hand side of (105) to get
1T
Sqk = 1
(107)
Thus, equating (106) and (107), we obtain (98). The problem we are
attempting to solve is then equivalent to finding t(cid:48)
RR,k that
satisfy (105) subject to
SR,k and t(cid:48)
(cid:48)
SR,k (cid:60) 0
t
RR,k (cid:31) 0
(cid:48)
t
(108)
(109)
Now, note that (105) consists of S equations and note that the number
of variables (i.e., the total number of entries of t(cid:48)
RR,k) is
greater than the number of equations. Each equation relates the entries
of t(cid:48)
SR,k that correspond to agents of one of the sending sub-networks
to all entries of t(cid:48)
RR,k. In particular, the equation that corresponds to
sending sub-network s has the following form:
SR,k and t(cid:48)
(cid:48)
t
SR,k((cid:96)) = qk(θ
s ) −
◦
qk(j)(θ
◦
s )t
(cid:48)
RR,k(j)
(110)
RR,k, we compute the entries of t(cid:48)
Equation (110) shows how the entries of t(cid:48)
SR,k that correspond to
agents of sending sub-network s, are related to the entries of t(cid:48)
through the values of the desired beliefs at θ◦
RR,k
s. Therefore, the set of
all possible solutions to (105) consist of vectors whose entries satisfy
(110) for each s. In other words, by arbitrarily fixing the entries
of t(cid:48)
SR,k using (110) for each s
to obtain a solution to (105). This is because (105) is made of S
equations that only indicate how the entries of t(cid:48)
SR,k that correspond
to each sending sub-network s are related to t(cid:48)
RR,k without having
any additional equation for the entries of t(cid:48)
RR,k. Note that it does not
matter how the individual entries of t(cid:48)
SR,k that correspond to sub-
network s are chosen as long as their sum satisfies (110). However,
in the problem we are trying to solve, we are not interested in the
entire set of solutions to (105). This is because we have two additional
constraints (108) and (109). Therefore, in our problem we cannot
arbitrarily fix the entries of t(cid:48)
RR,k to any values as we need to also
satisfy (108) and (109). Constraint (108) implies that (110) should
be non-negative for each sending sub-network s, i.e.,
(cid:88)
(cid:96)∈Is
gR(cid:88)
N k
j=1
s ) ≥
◦
qk(θ
qk(j)(θ
◦
s )t
(cid:48)
RR,k(j)
(111)
gR(cid:88)
N k
j=1
10
gR(cid:88)
N k
Therefore, the problem reduces to finding t(cid:48)
RR,k that satisfies:
t
j=1
◦
s )t
qk(j)(θ
(112a)
RR,k(j), ∀s
(cid:48)
s ) ≥
◦
qk(θ
RR,k (cid:31) 0
(cid:48)
If it possible to find t(cid:48)
SR,k
can de determined using (110) and therefore a solution for agent k
is found. Finding t(cid:48)
RR,k that satisfies (112a) and (112b) is always
possible. By appropriately attenuating the entries of t(cid:48)
RR,k, we can
have the right-hand side of (112a) smaller than qk(θ◦
s ). For instance,
one solution is to assign the same value k > 0 to all entries of
t(cid:48)
RR,k. Then from (112a), we have for each s:
RR,k that satisfies (112a) and (112b), then t(cid:48)
(112b)
gR(cid:88)
N k
j=1
s ) ≥ k
◦
qk(θ
qk(j)(θ
◦
s )
which means that k should be chosen so that:
0 < k ≤ min
s
qk(θ◦
s )
(cid:80)N k
gR
j=1 qk(j)(θ◦
s )
(113)
(114)
We mentioned that, after finding t(cid:48)
RR,k that satisfies (112a) and
(112b), t(cid:48)
SR,k can be determined using (110). We can alternatively
express the solutions of t(cid:48)
SR,k using the same approach of the
previous section. This is because after choosing the entries of t(cid:48)
RR,k,
the problem is now similar to the previous problem of finding tSR,k
while tRR,k is given. Therefore, the solutions for t(cid:48)
SR,k can be also
given by (51). Note that (51) is expressed in terms of v(cid:48)
k to take into
account that agent k may not be connected to some sending sub-
networks, in the earlier section. Since in this case we are focusing
on agent k connected to all sending sub-networks, the solution for
SR,k is given by (51 where vk is used instead of v(cid:48)
t(cid:48)
k.
In summary, when agent k is connected to all sending sub-
networks, the problem can have an infinite number of solutions. We
first find t(cid:48)
RR,k that satisfies (112a) and (112b). Then, the entries
of t(cid:48)
SR,k are nonnegative values chosen to satisfy (110). In other
words, when a receiving agent k is under the direct influence of
all sending sub-networks, it is relatively straightforward to affect its
beliefs, especially since the influence from its receiving neighbors
can be attenuated as much as needed through the choice k.
Case 2: Agent k is connected to some sending sub-networks
We now consider the case where agent k is influenced by only a
subset of the sending networks. Without loss of generality, we assume
it is connected to the first s(cid:48) sending sub-networks. In this case, Ek
will have the following form:
1T
N k
1
0
...
0
0
...
0
0
1T
N k
2
...
0
0
...
0
. . .
0
0
. . .
...
...
. . . 1T
N k
s(cid:48)
. . .
0
...
...
. . .
0
Ek =
(115)
and relation (97) becomes:
1T
N k
1
0
...
0
0
...
0
0
1T
N k
2
...
0
0
...
0
. . .
0
0
. . .
...
...
. . . 1T
N k
s(cid:48)
0
. . .
...
. . .
...
0
qk(1)(θ◦
1 )
qk(1)(θ◦
2 )
...
qk(1)(θ◦
s(cid:48) )
qk(1)(θ◦
s(cid:48)+1)
...
qk(1)(θ◦
S)
. . .
. . .
...
. . .
. . .
...
. . .
qk(N k
qk(N k
qk(N k
qk(N k
gR)(θ◦
1 )
gR)(θ◦
2 )
...
gR)(θ◦
s(cid:48) )
gR)(θ◦
s(cid:48)+1)
(cid:20) t(cid:48)
...
gR)(θ◦
(cid:21)
S)
qk(N k
SR,k
t(cid:48)
RR,k
= qk
(116)
11
SR,k and t(cid:48)
The problem now is to find t(cid:48)
RR,k that satisfy (116) subject
to constraints (98)-(99). As before, if we can find t(cid:48)
RR,k
that satisfy (116), then condition (98) will be automatically satisfied.
Note now that (116) consists of s(cid:48) equations that relate the entries of
RR,k, and S − s(cid:48) equations that involve the
t(cid:48)
SR,k to the entries of t(cid:48)
entries of t(cid:48)
RR,k. Therefore, any vector that satisfies (116) will have
the following property:
SR,k and t(cid:48)
(cid:48)
SR,k((cid:96)) = qk(θ
s ) −
◦
t
qk(j)(θ
(cid:48)
◦
s )t
RR,k(j)
(117)
(cid:96)=1
j=1
but only for s ≤ s(cid:48). In other words, the entries of t(cid:48)
SR,k that
correspond to sub-network s ≤ s(cid:48) are expressed in terms of t(cid:48)
RR,k
through (117). In addition, and differently from case 1, any solution
to (116) should also satisfy:
1(cid:88)
N k
gR(cid:88)
N k
qk(j)(θ
◦
s )t
(cid:48)
RR,k(j) = qk(θ
◦
s )
(118)
gR(cid:88)
N k
j=1
for any s > s(cid:48). Likewise, constraint (108) implies that (117) should
be non-negative for each sending sub-network s where s ≤ s(cid:48), i.e.,
s ) ≥
◦
qk(θ
qk(j)(θ
(cid:48)
◦
s )t
RR,k(j)
(119)
for any s ≤ s(cid:48). Therefore, the problem reduces to finding t(cid:48)
satisfies:
RR,k that
gR(cid:88)
N k
j=1
N k
gR(cid:88)
gR(cid:88)
j=1
N k
j=1
s ) ≥
◦
qk(θ
◦
qk(θ
s ) =
RR,k (cid:31) 0
(cid:48)
t
qk(j)(θ
◦
s )t
(cid:48)
RR,k(j),
s ≤ s
(cid:48)
qk(j)(θ
◦
s )t
(cid:48)
RR,k(j),
(cid:48)
s > s
(120a)
(120b)
(120c)
If it possible to find t(cid:48)
RR,k that satisfies (120a)-(120c), then t(cid:48)
SR,k
can de determined using (117) or alternatively using (51). However,
in contrast to the case studied in the previous case, finding t(cid:48)
RR,k
that satisfies conditions (120a)-(120c) may not be always possible.
For instance, consider agent k shown in Fig. 6, which is connected
to only the first sending sub-network but not to the other two sending
sub-networks. Let us consider its desired limiting belief as
qk(θ◦
1 )
qk(θ◦
2 )
qk(θ◦
3 )
=
0.1
0.45
0.45
(121)
Fig. 6. An example where receiving agent k is only connected to
one sending sub-network.
while the desired limiting beliefs for its neighbors are:
q2(θ◦
1 )
q2(θ◦
2 )
q2(θ◦
3 )
=
0.2
0.5
0.3
,
q3(θ◦
1 )
q3(θ◦
2 )
q3(θ◦
3 )
=
0.1
0.4
0.5
(122)
Then, from (120a), we should have:
qk(θ
1 ) ≥ α2q2(θ
◦
and from (120b),
◦
1 ) + α3q3(θ
1 ) =⇒ 0.1 ≥ 0.2α2 + 0.1α3 (123)
◦
◦
2 ) = α2q2(θ
◦
2 ) + α3q3(θ
qk(θ
◦
3 ) = α2q2(θ
◦
3 ) + α3q3(θ
qk(θ
2 ) =⇒ 0.45 = 0.5α2 + 0.4α3
◦
3 ) =⇒ 0.45 = 0.3α2 + 0.5α3
◦
(124)
(125)
Solving (124) and (125) gives the following solution: α2 = 0.3462
and α3 = 0.6923. However, 0.2α2 +0.1α3 = 0.1385, which violates
(123). Still, we can have cases where all conditions (120a)-(120c) can
be met (we are going to provide one example in a later section), then
in these cases, we choose t(cid:48)
SR,k according to (117).
We observe from this case that the fewer the sending networks
that influence agent k, the harder it is to affect its limiting belief.
This emphasizes again the idea that the structure of the receiving
sub-networks helps in limiting external manipulation.
Case 3: Agent k is not connected to any sending sub-networks
When agent k is not connected to any sending sub-network,
relation (97) reduces to:
Qkt
The problem is then to find t(cid:48)
(cid:48)
RR,k = qk
RR,k that satisfies:
(cid:48)
RR,k = qk
(cid:48)
RR,k = 1
RR,k (cid:31) 0
(cid:48)
Qkt
1Tt
t
(126)
(127a)
(127b)
(127c)
This problem might not have an exact solution. For instance, we
discuss two examples in Appendix A of [arXiv paper], where in
the second example, we have an agent that is not connected to any
sending sub-network and its desired belief cannot be expressed as a
convex combination of the desired beliefs of its neighbors.
Comment and analysis
Since the problem of finding TSR and TRR satisfying (87)-(90) is
separable, we studied the possibility of finding a solution for each
column of TSR and TRR. We analyzed the problem for 3 cases and
discovered that for the first case (when agent k is connected to at
least one agent from each sending sub-network), problem (97) -- (100)
always has a solution. That is, if an agent k is connected to all
sending sub-networks and given knowledge of the limiting beliefs of
its neighbors, we can always find the weight combination for agent
k such that (103) is satisfied. For the second case (when agent k is
connected to some sending sub-networks) and the third case (when
agent k is not connected to any sending sub-network), we found out
that problem (97) -- (100) might not always have a solution, i.e., it is
not always possible to satisfy (103). These scenarios reinforce again
the idea that the internal structure of receiving agents can resist some
of the external influence.
12
Fig. 7.
networks.
A weakly-connected network consisting of three sub-
However, for Q to be achievable (i.e., for the beliefs of all receiving
agents converge to the desired beliefs), a solution must exist for
each agent k. If the desired limiting belief of any receiving agent
cannot be written as a convex combination of the limiting beliefs
of its neighbors (i.e., a solution cannot be found for problem (97) --
(100)), the whole scenario is not achievable. Even if it is possible for
agent k to find its appropriate weights t(cid:48)
RR,k, finding this
solution is based on the knowledge of the desired limiting beliefs of
its neighbors. However, if one of the receiving neighbors cannot reach
its desired belief, agent k will not be able anymore to reach its desired
belief. Therefore, for Q to be attainable, a solution for problem (97) --
(100) must exist for each receiving agent k. If Q is not attainable,
then the desired scenario should be modified to an attainable scenario,
by taking into consideration the limitation provided by the internal
connection of the receiving sub-networks. Or an approximate least-
squares solution for the weights can be found. That is, we can instead
seek to solve
SR,k and t(cid:48)
(cid:13)(cid:13)(cid:13)(cid:13)Bk
(cid:20) t(cid:48)
t(cid:48)
SR,k
RR,k
(cid:21)
(cid:13)(cid:13)(cid:13)(cid:13)2
− qk
min
SR,k,t(cid:48)
t(cid:48)
RR,k
subject to
1Tt
(cid:48)
(cid:48)
SR,k + 1Tt
RR,k = 1
SR,k (cid:23) 0
(cid:48)
t
RR,k (cid:31) 0
(cid:48)
t
The last condition can be relaxed to the following:
RR,k (cid:23) k1
(cid:48)
t
(129)
(130)
(131)
(132)
where 0 < k < 1. Clearly, when we solve problem (128) -- (132), this
does not mean that the objective function (128) will be zero at this
solution. Note further that the optimization problem (128) -- (132) is
a quadratic convex problem: its objective function is quadratic, and
it has a convex equality constraint (129) and inequality constraints
(130) and (132). The inequality constraints are element-wise, i.e.,
RR,k(j) ≥ k for all j, which can be equivalently written as
t(cid:48)
RR,k ≥ k for all j where ej is a vector where all its elements are
j t(cid:48)
eT
zero expect for the jth element that is one. In this way, the problem
becomes a classic constrained convex optimization problem, which
can be solved numerically (using for instance interior point methods).
V. SIMULATION RESULTS
We illustrate the previous results with the following simulation
example. Consider the social network shown in Fig. 7 which consists
of N = 23 agents.
We assume that there are 3 possible events Θ = {θ◦
3}, where
2 , θ◦
1 , θ◦
θ◦
1 is the true event for the first sending sub-network, θ◦
2 is the true
event for the second sending sub-network, and θ◦
3 is the true event for
the receiving sub-network. We further assume that the observational
signals of each agent k are binary and belong to Zk = {H, T} where
H denotes head and T denotes tail.
Agents of the first sending sub-network are connected through the
following combination matrix:
(128)
A1 =
0
0.4
0
0
0
0
0
0.6
0.3
0
0.7
0
0
0
0
0
0
0.3
0
0.4
0.3
0
0
0
0
0
0.5
0
0
0.5
0
0
0
0
0.25
0
0
0.25
0.3
0.2
0
0
0
0.3
0.1
0
0.6
0
0
0
0
0
0.2
0.1
0
0.7
0.3
0
0
0
0.45
0
0.25
0
Agents of the second sending sub-network are connected through the
following combination matrix:
A2 =
0
0.1
0
0.1
0
0
0
0.8
0.35
0.25
0.4
0
0
0
0
0
0
0.5
0
0
0.5
0
0
0
0.3
0
0
0
0.3
0
0.4
0
0
0
0.8
0.1
0
0.1
0
0
0
0
0
0
0.45
0
0.55
0
0
0
0
0.6
0
0.3
0
0.1
0.25
0
0
0
0
0
0
0.75
(cid:20) 0.55
(134)
The matrices TSR and TRR are going to be designed so that the
desired limiting beliefs for receiving agents are as follows:
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.45
0.55
0.45
Q1 =
(135)
In other words, the weights are going to be designed so that θ◦
1 and θ◦
2
are almost equally probable for the receiving agents. This illustrates
the case when the receiving agents listen to two different perspectives
from two media sources that are trustworthy for them, which leaves
them undecided regarding which true state to choose.
(cid:21)
The likelihood of the head signals for each receiving agent k is
selected as the following matrix:
LR(H) =
3/4
3/4
3/4
1/6
1/6
1/6
7/8
7/8
7/8
2/3
2/3
2/3
1/3
1/3
1/3
1/4
1/4
1/4
where each (j, k)-th element of this matrix corresponds to Lk(H/θj),
i.e., each column corresponds to one agent and each row to one
network state. The likelihood of the tail signal is L(T ) = 13×7 −
L(H). The likelihood of the head signals for each sending agent k
5/8
5/8
5/8
(133)
of the first sending sub-network is selected as the following matrix:
L1(H) =
3/4
3/4
3/4
1/6
1/6
1/3
1/2
2/3
1/2
1/3
1/2
1/4
1/5
1/5
1/5
4/5
2/3
4/5
1/2
1/2
1/3
(136)
and the likelihood of the head signals of agents of the second sending
sub-network is:
L2(H) =
5/8
2/3
2/3
1/4
5/8
5/8
1/2
1/3
1/4
1/2
1/2
1/2
1/2
1/2
1/5
6/7
8/9
8/9
1/4
1/4
1/4
5/8
1/4
5/8
7/8
7/8
1/3
(137)
Design and Result Simulation
To achieve Q1, we design TSR and TRR using the results in the
previous section. The details of the numerical derivation are omitted
for brevity. The non-zero weights in TSR are shown in Fig. 8, and
TRR is given as follows:
TRR =
0
0.3
0.3
0.3
0
0
0
0.1
0
0.1
0.1
0.1
0
0
0.25
0.25
0.25
0.25
0
0
0.25
0.25
0
0
0
0.25
0
0.25
0
0.3
0.3
0.3
0
0
0
0
0
0.5
0
0
0
0.5
0
0
0
0.5
0
0.5
0
13
Fig. 9. Evolution of the beliefs of the receiving agents at θ◦
time
1 over
(138)
Fig. 10. Evolution of the beliefs of the receiving agents at θ◦
time
2 over
Fig. 8.
Illustration of the limiting beliefs of receiving agents
We run this example for 7000 time iterations. We assigned to
3}. Figures
each agent an initial belief that is uniform over {θ◦
2 , θ◦
9 and 10 show the evolution of µk,i(θ◦
2 ) of agents in
the receiving sub-network. These figures show the convergence of
the beliefs of the agents in the receiving sub-networks to the desired
beliefs in Q1. Figure 8 illustrates with color the limiting beliefs of
receiving agents.
1 , θ◦
1 ) and µk,i(θ◦
VI. CONCLUSION
In this work, we characterized the set of beliefs that can be imposed
on non-influential agents and clarified how the graph topology of
these latter agents helps resist manipulation but only to a certain
degree. We also derived design procedures that allow influential
agents to drive the beliefs of non-influential agents to desirable
attainable states.
APPENDIX A
TWO REVEALING EXAMPLES FOR THE DESIGN
PROCEDURE (97)-(100)
Example I: Cases 1 and 2 (k is influenced by sending net-
works)
Consider the network shown in Fig. 11. It consists of N = 8
agents, two sending sub-networks and one receiving sub-network,
with the following combination matrix:
A =
0
0
0
0.4
0.6
0
0
0
0.2
0.5
0.3
0
0
0
0
0
0.8
0.1
0.1
0
0
0
0
0
0.2
0.4
0.4
0
0
0
0
0
0 × 0
0
0 × 0
0
0 × 0 ×
0.3 × × 0
0.7
0
0
0 × ×
0
0 × 0 ×
0 × × ×
We assume that there are 3 possible states Θ = {θ◦
3}, where
1 , θ◦
2 , θ◦
θ◦
1 is the true event for the first sending sub-network, θ◦
2 is the true
event for the second sending sub-network, and θ◦
3 is the true event
for the receiving sub-network. Let us consider the case where we
(139)
0
01000200030004000500060007000Number of iterations00.20.40.60.81Probability of Belief in θ1oBeliefs of Receiving AgentsAgent 17Agent 18Agent 19Agent 20Agent 21Agent 22Agent 2301000200030004000500060007000Number of iterations00.20.40.60.81Probability of Belief in θ2oBeliefs of Receiving AgentsAgent 17Agent 18Agent 19Agent 20Agent 21Agent 22Agent 23 where
α2 =
1
2
(0.2 − 0.3α4 − 0.5α5) − β
β ≤ 1
2
(0.2 − 0.3α4 − 0.5α5)
14
(149)
(150)
This choice of β ensures that α1 and α2 are non-negative and less
than 0.2 − 0.3α4 − 0.5α5. Moreover, we can check from (148) and
(149) that their sum satisfies (144). Therefore, the solution has the
following form:
(cid:20) t(cid:48)
(cid:21)
t(cid:48)
SR,6
RR,6
=
1
2 (0.2 − 0.3α4 − 0.5α5) + β
2 (0.2 − 0.3α4 − 0.5α5) − β
0.8 − 0.7α4 − 0.5α5
1
α4
α5
0.3α4 + 0.5α5 ≤ 0.2
0.7α4 + 0.5α5 ≤ 0.8
α4 > 0, α5 > 0
(0.2 − 0.3α4 − 0.5α5)
β ≤ 1
2
(cid:26) 0.2
(cid:27)
For example, one solution is to assign the same value 6 for α4 and
α5. Then, from (152), (153) and (114), we have:
0 < 6 ≤ min
0.5 + 0.3
,
0.8
0.7 + 0.5
= 0.25
(156)
Let 6 = 0.1 = α4 = α5, then
α1 + α2 = 0.2 − 0.3α4 − 0.5α5 = 0.12
α3 = 0.8 − 0.7α4 − 0.5α5 = 0.68
(157)
(158)
We can choose α1 = 0.1 and α2 = 0.02. Therefore, a possible
solution for tSR,6 is:
We follow a similar procedure for agent 7 and obtain:
tSR,6 =(cid:2) 0.1
(cid:21)
(cid:20) t(cid:48)
SR,7
t(cid:48)
RR,7
∆=
0
0
0.68
0.1
0.1 (cid:3)T
0.3 − 0.2β3 − 0.5β4
0.7 − 0.8β3 − 0.5β4
β3
β4
0.02
=
0
β1
β2
β3
β4
(151)
(152)
(153)
(154)
(155)
(159)
(160)
(161)
(162)
(163)
A weakly connected network consisting of three sub-
Fig. 11.
networks. In this example, receiving agents 6 and 7 are influenced
by both sending networks, while agent 8 is only influenced by the
first sending network.
where
want to design TSR and TRR to attain the desired limiting beliefs
(cid:20) 0.2
0.8
Q =
(cid:21)
0.3
0.7
0.5
0.5
The matrix B is therefore of the following form:
B =
1
0
1
0
0
1
0
1
0.2
0.8
0.3
0.7
0.5
0.5
(cid:20) 1
0
(cid:21)
(140)
(141)
We start with agent 6. After eliminating entries to satisfy the sparsity
in the connections, we are reduced to finding t(cid:48)
RR,6 that
SR,6 and t(cid:48)
satisfy(cid:20) 1
(cid:124)
0
Let
t(cid:48)
SR,6
(cid:21)
0.5
0.5
(cid:21)
(cid:125)
(cid:20) t(cid:48)
(cid:20) 0.2
(cid:21)
(cid:124) (cid:123)(cid:122) (cid:125)
(cid:3)T
∆= (cid:2) α1 α2 α3 α4 α5
∆
= q6
0.8
RR,6
=
1
0
0
1
0.3
0.7
(cid:123)(cid:122)
(cid:21)
∆
= B6
(cid:20) t(cid:48)
t(cid:48)
SR,6
RR,6
(142)
(143)
Agent 6 is connected to the two sending sub-networks (case 1).
Therefore, the problem has a solution, where t(cid:48)
SR,6 (α1, α2 and α3)
can be expressed in terms of t(cid:48)
RR,6 (α4 and α5). More precisely,
from (142) and (110), we have:
where
α1 + α2 = 0.2 − 0.3α4 − 0.5α5
α3 = 0.8 − 0.7α4 − 0.5α5
(144)
(145)
0.2β3 + 0.5β4 ≤ 0.3
0.8β3 + 0.5β4 ≤ 0.7
β3 > 0, β4 > 0
According to (111), to ensure that α1, α2 and α3 can be chosen
as nonnegative numbers, the scalars α4 and α5 should be chosen to
satisfy
0.3α4 + 0.5α5 ≤ 0.2
0.7α4 + 0.5α5 ≤ 0.8
(146)
(147)
Note that what matters for scalars α1 and α2 (the weights with which
the data received from sending sub-network 1 is scaled) is that their
sum should be equal to 0.2 − 0.3α4 − 0.5α5 according to (144). In
other words, when a receiving agent is connected to many agents
from the same sending sub-network, it does not matter how much
weight is given to each of these agents as long as the sum of these
weights takes the required value. This is because the beliefs of agents
of the same sending sub-networks will converge to the same final
distribution. An alternative way to express (144) is to set α1 and α2
to the following:
α1 =
1
2
(0.2 − 0.3α4 − 0.5α5) + β
(148)
For this agent, we can choose for instance as a solution β3 = 0.2
and β4 = 0.1 (as they both satisfy (161) and (162)). In this case,
β1 = 0.3 − 0.2β3 − 0.5β4 = 0.21
β2 = 0.7 − 0.8β3 − 0.5β4 = 0.49
tSR,7 =(cid:2) 0
Therefore, a possible solution for tSR,7 is:
0.21
0
0.49
0
0.2
0
(164)
(165)
(166)
0.1 (cid:3)T
Agent 8 is connected to the first sending sub-network only (case
2). For this agent, we have:
(cid:20) 1
(cid:124)
0
(cid:20) t(cid:48)
(cid:21)
(cid:125)
(cid:21)
(cid:20) 0.5
(cid:21)
(cid:124) (cid:123)(cid:122) (cid:125)
0.5
∆
= q8
t(cid:48)
SR,8
RR,8
=
0.5
0.5
0.2
0.8
0.3
0.7
(cid:123)(cid:122)
∆
= B8
(167)
Let
(cid:20) t(cid:48)
t(cid:48)
SR,8
RR,8
(cid:21)
∆=
γ1
γ2
γ3
γ4
Therefore, from (167), (117) and (118), we have:
γ1 = 0.5 − 0.2γ2 − 0.3γ3 − 0.5γ4
0.8γ2 + 0.7γ3 + 0.5γ4 = 0.5
and any vector that satisfies (167) has the following form:
(cid:20) t(cid:48)
(cid:21)
t(cid:48)
SR,8
RR,8
=
γ1
γ2
γ3
γ4
=
0.5 − 0.2γ2 − 0.3γ3 − 0.5γ4
γ2
γ3
γ4
where
0.8γ2 + 0.7γ3 + 0.5γ4 = 0.5
(169)
(170)
(171)
(172)
Now to ensure that γ1 is non-negative, γ2, γ3 and γ4 should be
chosen as follows (as in (119)):
A =
0.2γ2 + 0.3γ3 + 0.5γ4 ≤ 0.5
(173)
Therefore, a solution in this case should satisfy (171) subject to
15
k = 6
k = 7
k = 8
(185)
calculation gives
(168)
lim
i→∞ µk,i(θ
◦
1 ) =
and
0.1070 + 0.0310 + 0.0620 = 0.2,
0.0258 + 0.2247 + 0.0494 = 0.3,
0.0443 + 0.0852 + 0.3705 = 0.5,
0.8,
0.7,
0.5,
k = 6
k = 7
k = 8
lim
i→∞ µk,i(θ
◦
2 ) =
Example II: Case 3 (agent k not
networks)
influenced by sending
Consider the network shown in Fig. 12, with the following
combination matrix:
0.2
0.5
0.3
0
0
0
0
0
0.2
0.4
0.4
0
0
0
0
0
0.8
0.1
0.1
0
0
0
0
0
0
0
0
0.4
0.6
0
0
0
0 × 0
0
0 × 0
0
0 × 0
0
0.3 × × 0
0
0.7
0
0 × ×
0
0 × 0 ×
0 × × 0
0
0.8γ2 + 0.7γ3 + 0.5γ4 = 0.5
0.2γ2 + 0.3γ3 + 0.5γ4 ≤ 0.5
γ2 > 0, γ3 > 0, γ4 > 0
(174)
(175)
(176)
For this example, finding γ2, γ3 and γ4 that satisfy (174)-(175) is
always possible. To see this, for any choice of γ2, γ3 and γ4 that
satisfy (174), condition (175) is automatically satisfied. Indeed, if
(174) is satisfied then
0.5γ2 + 0.5γ3 + 0.5γ4 ≤ 0.8γ2 + 0.7γ3 + 0.5γ4 = 0.5
=⇒ 0.5(γ2 + γ3 + γ4) ≤ 0.5 =⇒ γ2 + γ3 + γ4 ≤ 1
(177)
(178)
Therefore,
γ2 + γ3 + γ4 − 0.8γ2 − 0.7γ3 − 0.5γ4 ≤ 1 − 0.5
=⇒ 0.2γ2 + 0.3γ3 + 0.5γ4 ≤ 0.5
(179)
(180)
For instance, one possible choice for γ2, γ3 and γ4 that satisfies (174)
is
γ2 = γ3 = γ4 =
0.5
0.8 + 0.7 + 0.5
= 0.25
(181)
Then,
γ1 = 0.5 − 0.2γ2 − 0.3γ3 − 0.5γ4 = 0.25
Therefore, a possible solution for tSR,8 is:
tSR,8 =(cid:2) 0
0
0.25
0
Thus, the overall solution is:
0.25 (cid:3)T
(cid:21)
(cid:20) TSR
TRR
=
0
0.25
0.25
0.1
0
0.02
0.68
0
0
0.1
0.1
0
0.21
0
0.49
0
0.2
0
0.1
0
0
0.25
0
0
0.25
0.25
0.25
(182)
(183)
(184)
To verify that the beliefs of the receiving agents converge to the
desired beliefs, we compute W T from (13) and use (18) to determine
the limiting beliefs at θo
the receiving agents. This
1 and θo
2 at
A weakly connected network consisting of three sub-
Fig. 12.
networks. In this case, agent 8 is not influenced by any sending
network.
What is different now is that agent 8 does not have is not connected
to agent 3 (that is, agent 8 is not connected to any sending network).
We are still assuming in this example that we have the same desired
limiting beliefs:
(cid:20) 0.2
0.8
Q =
(cid:21)
0.3
0.7
0.5
0.5
(186)
For agents 6 and 7, the solutions found for their corresponding
columns are still valid here. However, in this example, tSR,8 should
have all its elements equal to zero and tRR,8 should have its third
element equal to zero. Therefore, for agent 8, the problem reduces
to finding t(cid:48)
RR,8 that satisfies the following relationship:
(cid:20) 0.2
(cid:124)
0.8
0.3
0.7
(cid:123)(cid:122)
∆
= B8
(cid:21)
(cid:125)
(cid:21)
(cid:20) 0.5
(cid:124) (cid:123)(cid:122) (cid:125)
0.5
∆
= q8
(cid:48)
t
RR,8 =
(187)
where the elements of t(cid:48)
RR,8 should be positive and add up to 1.
Any convex combination of 0.2 and 0.3 can only produce a number
between 0.2 and 0.3, but not 0.5. This is why in this case, the problem
does not have a solution. However, we can seek instead a least-squares
solution for agent 8:
subject to
(cid:13)(cid:13)B8t
min
t(cid:48)
RR,8
(cid:13)(cid:13)2
RR,8 − q8
(cid:48)
By choosing 8 = 0.01 and solving it numerically, we obtain:
RR,8 (cid:23) 81
(cid:48)
t
(cid:48)
1Tt
RR,8 = 1
(cid:21)
(cid:20) 0.01
0.99
(cid:48)
t
RR,8 =
This solution can be also deduced directly because [0.3; 0.7] is the
closer distribution to [0.5; 0.5] than any other distribution formed by
a convex combination of [0.2; 0.8] and [0.3; 0.7]. Because the entries
should be strictly greater than 0, the lowest possible value is given
to the first entry of t(cid:48)
RR,8. Therefore, with this choice:
(cid:21)
(cid:20) TSR
TRR
=
0.1
0
0.02
0.68
0
0
0.1
0.1
0
0.21
0
0.49
0
0.2
0
0.1
0
0
0
0
0
0.01
0.99
0
(188)
(189)
(190)
(191)
(192)
we verify the limiting beliefs of the agents as follows. We compute
W T from (13) and use (18) to determine the limiting beliefs at θo
1
and θo
2 at the receiving agents . This calculation gives
and
i→∞ µk,i(θ
lim
◦
1 ) =
lim
i→∞ µk,i(θ
◦
2 ) =
0.272,
0.271,
0.174,
0.826,
0.728,
0.729,
k = 6
k = 7
k = 8
k = 6
k = 7
k = 8
It is expected that the beliefs of agents 6 and 7 would not converge
to the desired beliefs, because the belief of agent 8 cannot converge
to its desired belief, which will definitely affect the limiting beliefs
of agents 6 and 7. We know that agent 8 will not converge to its
desired limiting belief because [0.5;0.5] cannot be obtained by any
convex combination of [0.2;0.8] and [0.3;0.7] (its neighbors' limiting
beliefs, (103)).
REFERENCES
[1] H. Salami and A. H. Sayed, "Belief control strategies for interactions
over weak graphs," in Proc. IEEE ICASSP, New Orleans, LA, March
2017, pp. 4232 -- 4236.
[2] X. Zhao and A. H. Sayed, "Learning over social networks via diffusion
in Proc. Asilomar Conference on Signals, Systems and
adaptation,"
Computers, Pacific Grove, CA, Nov 2012, pp. 709 -- 713.
[3] A. Jadbabaie, P. Molavi, A. Sandroni, and A. Tahbaz-Salehi, "Non-
Bayesian social learning," Games and Economic Behavior, vol. 76, no.
1, pp. 210 -- 225, 2012.
[4] D. Acemoglu, M. Dahleh, I. Lobel, and A. Ozdaglar, "Bayesian learning
in social networks," The Review of Economic Studies, vol. 78, no. 4,
pp. 1201 -- 1236, 2011.
[5] E. Yildiz, A. Ozdaglar, D. Acemoglu, A. Saberi, and A. Scaglione,
"Binary opinion dynamics with stubborn agents," ACM Trans. Econ.
Comput., vol. 1, no. 4, pp. 19:1 -- 19:30, Dec. 2013.
[6] D. Acemoglu and A. Ozdaglar, "Opinion dynamics and learning in
social networks," Dynamic Games and Applications, vol. 1, no. 1, pp.
3 -- 49, 2011.
[7] D. Acemoglu, A. Ozdaglar, and A. ParandehGheibi,
"Spread of
(mis)information in social networks," Games and Economic Behavior,
vol. 70, no. 2, pp. 194 -- 227, 2010.
16
[8] B. Golub and M. Jackson, "Naive learning in social networks and the
wisdom of crowds," American Economic Journal: Microeconomics, vol.
2, no. 1, pp. 112 -- 49, 2010.
[9] P. Molavi, A. Jadbabaie, K. R. Rad, and A. Tahbaz-Salehi, "Reaching
consensus with increasing information," IEEE Journal of Selected Topics
in Signal Processing, vol. 7, no. 2, pp. 358 -- 369, 2013.
[10] C. Chamley, A. Scaglione, and L. Li, "Models for the diffusion of beliefs
in social networks: An overview," IEEE Signal Processing Magazine,
vol. 30, no. 3, pp. 16 -- 29, May 2013.
[11] H. T. Wai, A. Scaglione, and A. Leshem, "Active sensing of social
networks," IEEE Transactions on Signal and Information Processing
over Networks, vol. 2, no. 3, pp. 406 -- 419, Sept. 2016.
[12] V. Krishnamurthy and H. V. Poor, "Social learning and Bayesian games
in multiagent signal processing," IEEE Signal Processing Magazine,
vol. 33, no. 3, pp. 43 -- 757, 2013.
[13] V. Krishnamurthy and H.V. Poor, "A tutorial on interactive sensing in
social networks," IEEE Transactions on Computational Social Systems,
vol. 1, no. 1, pp. 3 -- 21, March 2014.
[14] V. Krishnamurthy and M. Hamdi, "Mis-information removal in social
networks: Constrained estimation on dynamic directed acyclic graphs,"
IEEE Journal of Selected Topics in Signal Processing, vol. 7, no. 2, pp.
333 -- 346, April 2013.
[15] V. Krishnamurthy and W. Hoiles,
"Information diffusion in social
sensing," Numerical Algebra, Control and Optimization, vol. 6, no.
3, pp. 365 -- 411, Sept. 2016.
[16] P. Molavi, C. Eksin, A. Ribeiro, and A. Jadbabaie,
"Learning to
coordinate in social networks," Operations Research, vol. 64, no. 3,
pp. 605 -- 621, 2016.
[17] S. Pequito, S. Kar, and A. P. Aguiar, "Minimum number of information
gatherers to ensure full observability of a dynamic social network: A
in Proc. IEEE Global Conference on
structural systems approach,"
Signal and Information Processing, Atlanta, GA, Dec 2014, pp. 750 --
753.
[18] A. Lalitha, A. Sarwate, and T. Javidi, "Social learning and distributed
hypothesis testing," in Proc. IEEE International Symposium on Infor-
mation Theory, Honolulu, HI, June 2014, pp. 551 -- 555.
[19] L. G. Epstein, J. Noor, and A. Sandroni, "Non-Bayesian learning," The
B.E. Journal of Theoretical Economics, vol. 10, no. 1, pp. 1 -- 20, 2010.
[20] Q. Liu, A. Fang, L. Wang, and X. Wang, "Social learning with time-
varying weights," Journal of Systems Science and Complexity, vol. 27,
no. 3, pp. 581 -- 593, June 2014.
[21] L. Su and N. H. Vaidya, "Non-Bayesian learning in the presence of
Byzantine agents," in Proc. International Symposium on Distributed
Computing, Paris, France, September 2016, pp. 414 -- 427, Springer.
[22] S. Shahrampour and A. Jadbabaie, "Distributed online optimization in
dynamic environments using mirror descent," IEEE Transactions on
Automatic Control, vol. 63, no. 3, pp. 714 -- 725, March 2018.
[23] A. Nedi´c, A. Olshevsky, and C. A. Uribe, "Fast convergence rates for
IEEE Transactions on Automatic
distributed non-bayesian learning,"
Control, vol. 62, no. 11, pp. 5538 -- 5553, Nov 2017.
[24] A. Jadbabaie, P. Molavi, and A. Tahbaz-Salehi, "Information heterogene-
ity and the speed of learning in social networks," Columbia Business
School Research Paper, pp. 13 -- 28, May 2013.
[25] P. Molavi, K. R. Rad, A. Tahbaz-Salehi, and A. Jadbabaie,
"On
consensus and exponentially fast social learning," in Proc. American
Control Conference (ACC), Montr´eal, Canada, June 2012, pp. 2165 --
2170.
[26] A. Nedi´c and A. Olshevsky, "Distributed optimization over time-varying
directed graphs," IEEE Transactions on Automatic Control, vol. 60, no.
3, pp. 601 -- 615, March 2015.
[27] B. Ying and A. H. Sayed, "Information exchange and learning dynamics
IEEE Transactions on
over weakly connected adaptive networks,"
Information Theory, vol. 62, no. 3, pp. 1396 -- 1414, March 2016.
[28] H. Salami, B. Ying, and A. H. Sayed, "Diffusion social learning over
weakly-connected graphs," in Proc. IEEE ICASSP, Shanghai, China,
March 2016, pp. 4119 -- 4123.
[29] H. Salami, B. Ying, and A. H. Sayed, "Social learning over weakly-
IEEE Transactions on Signal and Information
connected graphs,"
Processing over Networks, vol. 3, no. 2, pp. 222 -- 238, June 2017.
[30] U. A. Khan, S. Kar, and J. M. F. Moura,
"Higher dimensional
consensus: Learning in large-scale networks," IEEE Transactions on
Signal Processing, vol. 58, no. 5, pp. 2836 -- 2849, May 2010.
[31] A. H. Sayed, "Adaptation, learning, and optimization over networks,"
Foundations and Trends in Machine Learning, vol. 7, no. 4-5, pp. 311 --
801, 2014.
[32] A. H. Sayed, "Adaptive networks," Proceedings of the IEEE, vol. 102,
no. 4, pp. 460 -- 497, 2014.
[33] S. Shahrampour and A. Jadbabaie,
"Exponentially fast parameter
estimation in networks using distributed dual averaging," in Proc. IEEE
Conference on Decision and Control, Firenze, Italy, Dec 2013, pp. 6196 --
6201.
[34] S. Shahrampour, A. Rakhlin, and A. Jadbabaie, "Distributed detection:
Finite-time analysis and impact of network topology," IEEE Transac-
tions on Automatic Control, vol. 61, no. 11, pp. 3256 -- 3268, Nov 2016.
[35] A. J. Laub, Matrix Analysis For Scientists And Engineers, SIAM,
Philadelphia, PA, USA, 2004.
17
|
1907.04396 | 1 | 1907 | 2019-07-09T20:29:48 | Informative Path Planning with Local Penalization for Decentralized and Asynchronous Swarm Robotic Search | [
"cs.MA",
"cs.AI",
"cs.RO"
] | Decentralized swarm robotic solutions to searching for targets that emit a spatially varying signal promise task parallelism, time efficiency, and fault tolerance. It is, however, challenging for swarm algorithms to offer scalability and efficiency, while preserving mathematical insights into the exhibited behavior. A new decentralized search method (called Bayes-Swarm), founded on batch Bayesian Optimization (BO) principles, is presented here to address these challenges. Unlike swarm heuristics approaches, Bayes-Swarm decouples the knowledge generation and task planning process, thus preserving insights into the emergent behavior. Key contributions lie in: 1) modeling knowledge extraction over trajectories, unlike in BO; 2) time-adaptively balancing exploration/exploitation and using an efficient local penalization approach to account for potential interactions among different robots' planned samples; and 3) presenting an asynchronous implementation of the algorithm. This algorithm is tested on case studies with bimodal and highly multimodal signal distributions. Up to 76 times better efficiency is demonstrated compared to an exhaustive search baseline. The benefits of exploitation/exploration balancing, asynchronous planning, and local penalization, and scalability with swarm size, are also demonstrated. | cs.MA | cs |
Informative Path Planning with Local Penalization for Decentralized
and Asynchronous Swarm Robotic Search
Payam Ghassemi1 and Souma Chowdhury2
Abstract -- Decentralized swarm robotic solutions to search-
ing for targets that emit a spatially varying signal promise
task parallelism, time efficiency, and fault tolerance. It is,
however, challenging for swarm algorithms to offer scalability
and efficiency, while preserving mathematical insights into the
exhibited behavior. A new decentralized search method (called
Bayes-Swarm), founded on batch Bayesian Optimization (BO)
principles, is presented here to address these challenges. Un-
like swarm heuristics approaches, Bayes-Swarm decouples the
knowledge generation and task planning process, thus preserv-
ing insights into the emergent behavior. Key contributions lie
in: 1) modeling knowledge extraction over trajectories, unlike in
BO; 2) time-adaptively balancing exploration/exploitation and
using an efficient local penalization approach to account for
potential interactions among different robots' planned samples;
and 3) presenting an asynchronous implementation of the algo-
rithm. This algorithm is tested on case studies with bimodal and
highly multimodal signal distributions. Up to 76 times better
efficiency is demonstrated compared to an exhaustive search
baseline. The benefits of exploitation/exploration balancing,
asynchronous planning, and local penalization, and scalability
with swarm size, are also demonstrated.
Index Terms -- Swarm Robotic Search, Informative Path
Planning, Bayesian Search, Gaussian Process, Asynchronous.
I. INTRODUCTION
Swarm robotic search is concerned with searching for
or localizing targets in unknown environments with a large
number of collaborative robots. There exists a class of search
problems in which the goal is to find the source or target
with maximum strength (often in the presence of weaker
sources), and where each source emits a spatially varying
signal. Potential applications include source localization of
gas leakage [1], nuclear meltdown tracking [2], chemical
plume tracing [3], and magnetic field and radio source lo-
calization [4], [5]. In such applications, decentralized swarm
robotic systems have been touted to provide mission effi-
ciency, fault tolerance, and scalable coverage advantages [6] --
[8], compared to sophisticated standalone systems. Decen-
tralized search subject to a signal with unknown spatial dis-
tribution usually requires both task inference and planning,
which must be undertaken in a manner that maximizes search
efficiency and mitigates inter-robot conflicts. This in turn
demands decision algorithms that are computationally light-
weight (i.e., amenable to onboard execution) [9], preferably
1Ph.D. Student, Dept. of Mechanical and Aerospace Engineering, Univer-
sity at Buffalo, Buffalo, NY 14260, USA. [email protected]
2Assistant Professor, Dept. of Mechanical and Aerospace Engineering,
University at Buffalo, Buffalo, NY 14260, USA. Corresponding Author.
[email protected]
explainable [10], and scalable [11] -- it is particularly chal-
lenging to meet these characteristics simultaneously.
In this paper, we perceive the swarm robotic1 search
process to consist of creating/updating a model of the signal
environment and deciding future waypoints thereof. Specif-
ically, we design, implement and test a novel decentralized
algorithm founded on a batch Bayesian search formalism.
This algorithm tackles the balance between exploration and
exploitation over trajectories (as opposed to over points,
which is typical in non-embodied search), while allowing
asynchronous decision-making. The remainder of this section
briefly surveys the literature on swarm search algorithms, and
converges on the contributions of this paper.
A. Swarm Robotic Search
In time-sensitive search applications under complex sig-
nal distributions, a team of robots can broaden the scope
of operational capabilities through distributed remote sens-
ing, scalability and parallelism (in terms of task execution
and information gathering) [12]. The multi-robot search
paradigm [11] uses concepts such as cooperative control [13],
model-driven strategies, [14], Bayesian filter by incorpo-
rating mutual information [15], strategies based on local
cues [16], and uncertainty reduction methods [17]. Scaling
these methods from the multi-robotic (<10 agents [11]) to
the swarm-robotic level (10 to 100 agents) often becomes
challenging in terms of online computational tractability.
A different class of approaches that is dedicated to guiding
the search behavior for larger teams is that based on nature-
inspired swarm intelligence (SI) principles
[18] -- [20]. SI-
based heuristics have been used to design algorithms both for
search in non-embodied n-dimensional space (e.g., particle
swarm optimization) and for swarm robotic search [21], [22].
Majority of the latter methods are targeted at
localizing
a single source [9], [23], with their effectiveness relying
on the usage of adaptive parameters (e.g., changing inertia
weight) [23]. The localization of the maximum strength
source in the presence of other weaker sources (i.e., under
a multi-modal spatial distribution), without making limiting
assumptions (e.g., distributed starting points [24]), has re-
ceived much less attention among SI-based approaches.
Translating optimization processes: Similar in principle
to some SI approaches, here we aim to translate an opti-
mization strategy [25] to perform search in the physical 2D
environment. In doing so, it is important to appreciate two
1Note that, we take a broader perspective of self-organizing swarm sys-
tems, one that is not limited to superficially-observable pattern-formations.
critical differences between these processes: 1) Movement
cost: unlike optimization, in swarm robotic search, moving
from one point to another may require a different energy/time
cost depending upon the environment (distance, barriers,
etc.) separating the current and next waypoints. 2) Sampling
over paths: robots usually gather multiple samples (signal
measurements) over the path from one waypoint to the next
(as sampling frequency >> waypoint frequency), unlike in
optimization where we sample only at their next planned
point. This "sampling over paths" characteristic has received
minimal attention in existing SI-based approaches.
Moreover, with SI-based methods, the resulting emergent
behavior, although often competitive, raises questions of
dependability (due to the use of heuristics) and mathematical
explainability [26]). The search problem can be thought
of as comprising two main steps: task inference (identi-
fying/updating the signal spatial model) and task selection
(waypoint planning). In SI methods, the two steps are not
separable, and a spatial model is not explicit. In our proposed
approach, the processes are inherently decoupled -- robots
exploit Gaussian Processes to model the signal distribution
knowledge (task inference) and solves a 2D optimization
over a special acquisition function to decide waypoints
(task selection). Such an approach is expected to provide
explainability, while preserving computational tractability.
B. Contributions of this Paper
The primary contributions of this paper, comprising what
we call the Bayes-Swarm algorithm, can be summarized as
the following. 1) We extend Gaussian process modeling to
update over trajectories and consider robot's motion con-
straints when using the GP to identify new samples. 2) We
formulate a novel batch-BO acquisition function, which not
only seeks to balance exploration and exploitation, but also
manages the interactions between samples in a batch, i.e.,
different robots' planned future paths, in a computationally
efficient manner. 3) We develop and test a simulated parallel
implementation of Bayes-Swarm for asynchronous search
planning over complex multi-modal signal distributions. The
performance of Bayes-Swarm, and its variations (synchro-
nized and purely explorative implementations), is analyzed
over signal environments of different complexity, and com-
pared with that of an exhaustive search baseline.
The remaining portion of the paper is organized as follows:
Our proposed decentralized algorithm (Bayes-Swarm) is de-
scribed next in Section II. Numerical experiments and results,
investigating the performance of Bayes-Swarm for different
case studies, are then presented in Sections III and IV. The
paper ends with concluding remarks. A summary background
of GP modeling is provided as Appendix I thereafter.
II. BAYES-SWARM ALGORITHM
A. Bayes-Swarm: Overview
Fig. 1. Bayes-Swarm: sequence of processes for each robot in the swarm
robot. The pseudocode of our proposed Bayes-Swarm algo-
rithm is given in Alg. 1. Each robot executes the Bayes-
Swarm algorithm after reaching a waypoint, to update its
knowledge and identify the next waypoint. This is accom-
plished by updating the GP model of the signal environment,
and using it
to maximize a special acquisition function.
Importantly, these planning instances need not be synchro-
nized across robots. The following assumptions are made
in implementing Bayes-Swarm: i) all robots are equipped
with precise localization; and ii) each robot can communicate
their knowledge and decisions to all peers after reaching a
waypoint (while full observability is assumed, the provision
of asynchronous planning reduces the communication depen-
dency, compared to synchronized algorithms [27], [28].
Assuming we have m robots, let's define the key param-
eters in Bayes-Swarm: Dkr
r ]: the observation
(cid:83)kr
r ) of robot-r
locations (Xkr
over its path connecting waypoints kr and kr−1; D1:kr =
i=1 Di
r: the cumulative information of robot-r up to
its arrival at the kr-th waypoint, including all self-recorded
and peer-reported observations; xkr
p : the next planned way-
point of robot-p, known to robot-r at the time when it's
p : reported next
(cid:83)m
at its kr-th waypoint; Xkr−r =(cid:83)
r ) and signal measurements (ykr
r = [Xkr
r , ykr
r=1
waypoints of robot-r's peers by that time.
p=1∧p(cid:54)=r xkr
B. Acquisition Function
r
Each robot-r takes an action (plans/travels-to next way-
point, xkr+1
) that maximizes an acquisition function. Since
the swarm's objective is to collectively explore a search area
to find the strongest signal source among multiple sources
in the least amount of time, the acquisition function must
balance exploration and exploitation. To this end, we propose
the following acquisition function formulation:
Figure 1 illustrates the sequence of processes (motion,
sensing, planning, communication, etc.), and associated flow
of information, encapsulating the behavior of each swarm
xkr+1
r
s.t.
(α · Ωr + (1 − α)βΣr) Γr
= arg max
x∈X kr
0 ≤ lkr
s = (cid:107)x − xkr
r (cid:107) ≤ V T
(1)
(2)
the first
term, Ωr = Ωr(x,D1:kr ),
Here,
leads robot-r
towards the location of the maximum signal strength expec-
tation (promotes exploitation); and the second term, Σr =
Σr(x,D1:kr , Xkr−r), minimizes the knowledge uncertainty
of robot-r w.r.t. the signal's spatial distribution (promotes
exploration). The multiplicative factor, Γ(x, ¯Xkr−r), penalizes
the interactions among the samples planned to be collected
by robot-r and its peers. How these terms are formulated
differently from standard acquisition functions used in BO
(to enable the unique characteristics of embodied search), are
described further in the following sub-sections.
The coefficient α ∈ [0, 1] in Eq. (1) is the exploitation
weight, i.e., α = 1 would be purely exploitative. Here, we
design α to be adaptive in a way (given by Eq. (3)) such that
the swarm behavior is strongly explorative at the start and
becomes increasingly exploitative over waypoint iterations,
e.g., it changes from ≈0.5 to ≈0.97 between 33% and 70%
of the expected mission time:
α = 1/(cid:0)1 + exp(−10(
Tmax
(3)
The term β in Eq. (1) is a prescribed scaling parameter that
is used to align the orders of magnitude of the exploitative
and explorative terms. For our case studies, we set β = 50.
Equation (2) constrains the length of robot-r's planned path,
s ), based on a set time-horizon (T ) for reaching the next
lkr
waypoint, and the maximum velocity (V ) of robots. In this
paper, the time-horizon is set such that 6×distance between
any consecutive waypoints does not exceed the arena length.
t
− 1
3
))(cid:1)
C. Source Seeking (Exploitative) Term
The robots model the signal's spatial distribution using a
GP with squared exponential kernel (further description of
this GP modeling is given in Appendix I). The GP model is
updated based on the robot's own measurements and those
r
- the next waypoint of robot-r.
Algorithm 1 Bayes-Swarm Algorithm
Input: GPr, xr, X kr−r -- current waypoint and recent observations
of robot-r (x), and planned waypoints of peers (X kr−r).
Output: xkr +1
1: procedure TAKEDECISION(r, kr, m, ∆θ)
2:
3:
4:
5:
6:
7:
8:
r ← TAKEFIRSTDECISION(r, kr, m, ∆θ)
xkr
if Size of Dkr
Down-sample Dkr
to Nmax n observations
x0 ← Location with the highest observed value
r ← Location maximizing the acquisition func.,
xkr
r > Nmax then
if kr = 0 then
(cid:46) Nmax = 1, 000
else
r
Eq.(1), by a gradient-based solver w/ initial x0
9:
10:
11: procedure TAKEFIRSTDECISION(r, m, ∆θ, Vr, T )
12:
13:
14:
15:
16:
17:
18:
kr ← kr + 1
return xkr
r , kr
d ← V T
if ∆θ = 360 then
θ ← r∆θ/m
else
θ ← r∆θ/(m + 1)
r ← [d cos θ, d sin θ]
x1
return x1
r,
(cid:46) ∆θ: Initial feasible dir. range
(cid:83)kr
communicated by its peers (each over their respective most
recent paths), thereby providing the following mean function:
r and y =(cid:83)m
where X =(cid:83)m
r=1
µr(x) = µ(x, X, y)
i=1 Xi
(4)
r. Due to
motion constraints (Eq. (2), a robot may not be able to reach
the location, ¯x∗, with the maximum expected signal strength
(estimated using their GP model), within the time horizon.
Therefore, the exploitative term is re-defined to get closest
to ¯x∗, as given by:
i=1 yi
r=1
(cid:83)kr
Ωr(x,D) =
1
1 + (x − ¯x∗)T (x − ¯x∗)
(5)
r = arg max¯x µr(¯x).
where ¯x∗
D. Knowledge-Uncertainty Reducing (Explorative) Term
Unlike in optimization,
in robotic search, sampling is
performed over the path of each agent. This concept is known
as informative path planning, where robots decide their path
such that the best possible information is extracted. The
(explorative) second term in Eq. (1) models the reduction in
uncertainty in the robots' belief (knowledge), thus facilitating
informative path planning. To this end, the path of the robot
is written in a parametric form as given below:
r ; u ∈ [0, 1]
s(u) = ux + (1 − u)xkr
is the current location of robot-r. In computing the
where xkr
r
self-reducible uncertainty in the belief of robot-r, we account
for the locations of both the past observations made by the
robot and its peers and the future observations of robot-r's
peers to be made over the paths to their planned waypoints
( ¯Xkr−r) -- both of these only consider what's currently known
to robot-r via communication from its peers. The knowledge-
gain can thus be expressed as:
Σr(x,D1:kr , ¯Xkr−r) =
r) and Xe
(cid:90)
(7)
r = X ∪ ¯Xkr−r. For further
where σr(x) = σ(x, Xe
details on computing the mean (Eq. (5)) and the variance
(Eq. (7)) of the GP, refer to Appendix I.
σr(s(u))du
(6)
s(x)
E. Local Penalizing Term
For a batch-BO implementation, it is necessary to account
for (and in our case mitigate) the interaction between the
batch of future samples. In swarm robotic search, this pro-
vides an added benefit of mitigating the overlap in planned
knowledge gain by robots in the swarm -- thereby promoting
a more efficient search process. Modeling it explicitly via
predictive distribution carries a significant computational
overhead of O(n3) [29]. Simultaneous optimization of future
candidate samples (in the batch) [30] is also not applicable
here since each robot must plan its future waypoint in a
decentralized manner. Recently, Gonzalez et al. [29] reported
a computationally tractable approximation to model the inter-
actions, using a local penalization term. We adopt and extend
this idea in our work through the local penalty factor.
This penalty factor, Γ(x, Xp−i) enables local exclusion
zones based on the Lipschitz properties of the signal spatial
function (f (x)), and thus tends to smoothly reduce the
acquisition function in the neighborhood of the existing batch
samples (the known planned waypoints of robot-r's peers
( Xkr−r), with the signal observations at those points being
not yet reported). To compute this penalty, we define a ball
Br with radius ρ around each peers' planned waypoints:
Br(x, xkp
p ∈ Xkr−r (8)
The local penalty associated with a point x is defined as the
probability that x does not belong to the ball Br:
p ) = {x ∈ X : (cid:107)xkp
p − x(cid:107) ≤ ρ}; xkp
γ(x, xkp
p ) = 1 − P (x ∈ Br)
(9)
r (xkp
p ))/L and variance σ2
We assume that the distribution of the ball radius ρ is Gaus-
sian with mean (M −µr(xkp
p )/L2.
Here, M = maxx f (x) is the maximum strength of the
source signal and L is a valid Lipschitz constant ((cid:107)f (x1) −
f (x2)(cid:107) ≤ L(cid:107)x1 − x2(cid:107)). Both M and L can be in general
readily set based on the knowledge of the application (we
know the expected maximum strength of the source signal
and the size of arena). By having these assumptions, we can
derive the following expressions for the local penalty:
γ(x, xkp
p ) = 1 − P ((cid:107)xkp
p − x(cid:107) ≤ ρ)
= P (N (0, 1) ≤ L(cid:107)x − xkp
− L(cid:107)x − xkp
(cid:113)
erfc
=
1
2
p (cid:107) − M + µr(xkp
p )
σr(xkp
p )
)
p (cid:107) − M + µr(xkp
p )
2σ2
r (xkp
p )
(10)
Here, erfc(.) is the complementary error function. The effec-
tive penalty factor is estimated based on the approximated
interactions with the waypoints of all peers of robot-r:
γ(x, xkp
p )
(11)
m(cid:89)
p=1∧p(cid:54)=r
Γr(x, Xkr−r) =
F. Information Sharing
Global inter-robot communication is assumed in this work.
However, given the bandwidth limitations of ad-hoc wireless
communication (likely in emergency response applications)
and its energy footprint of [31], the communication burden
needs to be kept low. Thus, along with asynchronous plan-
ning, robots share only a down-sampled set of observations.
Moreover, each robot broadcasts the following information
only after every planning instance: its next planned waypoint
r ) and observations made over its last path (Dkr
r ).
(xkr
III. CASE STUDIES
A. Distributed Implementation of Bayes-Swarm
In order to represent the decentralized manner in which
Bayes-Swarm operates, we develop a simulated environment
using MATLAB's (R2017b) parallel computing tools, and
deployed this environment in a dual 20 core workstation
(Intel R(cid:13) Xeon Gold 6148 27.5M Cache 2.40 GHz, 20 cores
processor, 196 GB RAM). Each robot executes its behavior,
as depicted in Fig. 1), in parallel with respect to the rest of
the swarm -- updating its own knowledge model after each
(a) Case study 1: small arena, non-
convex bimodal signal distribution
(b) Case study 2: large arena, multi-
modal signal distribution
Fig. 2. Two environment cases with different signal distributions.
waypoint and deciding its next waypoint based on its own
information and that received from its peers till that point.
The simulation time step is set at 1 ms. The observation
frequency over a path is set at 1 Hz. In order to have tractable
GP updating, each robot uses all observations dataset (D1:kr
)
if the size is less than 1,000, otherwise it is down-sampled
to 1,000 samples using a simple integer factor.
r
B. Case Studies
To evaluate the Bayes-Swarm algorithm,
two types of
experiments are conducted using two distinct signal environ-
ments. The two environments, shown in Fig. 2, respectively
provide a bimodal spatial distribution over a small arena, and
a complex multimodal spatial distribution over a larger arena.
In Experiment 1: Bayes-Swarm is run with 5 robots (small
swarm size chosen for ease of illustration), to analyse its
performance over the two environments, and compare with
that of two variations of Bayes-Swarm (Bayes-Swarm-Sync
and Bayes-Swarm-Explorative) and an exhaustive search
baseline. The synchronized planning (Bayes-Swarm-Sync)
version is implemented by changing the inequality con-
straint Eq. (2) to an equality constraint (fixed interval be-
tween waypoints) -- to investigate the hypothesized benefits
of asynchronous planning. The purely explorative version
(Bayes-Swarm-Explorative) is implemented by using α = 0
in Eq. (1) -- to highlight
the need for balancing explo-
ration/exploitation. In Experiment 2: a scalability analysis
is undertaken to explore the performance of Bayes-Swarm in
Case 2, across multiple swarm sizes.
The results of the experiments are evaluated and compared
in terms of relative completion time (τ) and mapping error
(δ). The relative completion time represents the search com-
pletion time (tachieved) relative to the idealized completion
time (tidealized), as given by:
τ = (tachieved − tidealized) /tidealized
(12)
Here, tidealized represents the time that a swarm robot would
hypothetically take to directly traverse the straight-line path
connecting the starting point and the signal source location.
Although the focus of this work is source localization (a
search problem), Bayes-Swarm can also be applied for
mapping purposes (a coverage problem) by setting α = 0.
Hence we report the mapping error δ (in terms of Root-
Mean Square Error or RMSE), which measures how the
response estimated using GP deviates from the actual signal
distribution over the arena. The RMSE is computed over a
set of 10, 000 test points uniformly distributed over the arena.
(a) σ1(x) at t = 5s (5 samples)
(b) σ5(x) at t = 5s (25 samples)
(c) σ5(x) at t = 20s (100 samples) (d) σ5(x) at t = 36s (180 samples)
(e) µ1(x) at t = 5s
(f) µ5(x) at t = 5s
(g) µ5(x) at t = 20s
(h) µ5(x) at t = 36s
For simulation termination purposes, two criteria are used.
The first criterion terminates the search if any robot arrives
within -vicinity of the signal source location. The second
criterion terminates the simulated mission, if a maximum
allowed search time (Tmax) is reached.
Fig. 3. Experiment 1, case 1: knowledge state and robot path snapshots. 3(a) -- 3(b): Top figures show the knowledge uncertainty map, in terms of σr(x)).
3(e) -- 3(h): Bottom figures show robot paths. Blue lines depict travelled paths, observations over which have been shared. Red solid lines depict travelled
paths, observations over which have not been shared. Red dashed lines depict planned paths, not yet travelled. Gray solid contours represent the actual
signal distribution and gray dashed contours represent the current signal distribution model of the stated robot. The green cross shows the source location.
correspondingly how the uncertainty in robot 5's knowledge
of the signal environment reduces (Figs. 3(b) to 3(d)). In this
Case, the five robots are able to build a relatively accurate
model of the environment and find the signal source in 36s
(using a total of 180 downsampled measurements); at that
point, the estimated and actual signal distributions mostly
coincide (Figs. 3(h)). The helpful role played by the adaptive
weight parameter (α) is also evident from Figs. 3(g) to
3(h), which show a more explorative behavior early on (e.g.,
t = 20s), and a more exploitive behavior later on (e.g., at
t = 36s) when two of the robots converge on the source.
A. Overall Performance of Bayes-Swarm
IV. RESULTS AND DISCUSSION
Figure 3 shows snapshots of the status of a 5-robot
swarm at different time points, based on the implementation
of Bayes-Swarm in the Case 1 environment (Fig. 2(a)).
They illustrate how the knowledge uncertainty (top plots),
represented by σr of any given robot-r, changes as the swarm
robots explore the arena while exchanging information with
each other and updating their GP models (bottom plots).
Looking from the perspective of robot 5 -- it can be
observed from Figs. 3(f) to 3(h) how the actions of the
swarm helps improve the model of the environment, and
TABLE I
ALGORITHM PERFORMANCE FOR 5-ROBOT SWARM ON THE TWO CASES
Case
1
2
Algorithm
Bayes-Swarm
Bayes-Swarm-Sync
Bayes-Explorative
Exhaustive Search
Bayes-Swarm
Bayes-Swarm-Sync
Bayes-Explorative
Exhaustive Search
Completion Time Mapping Error
0.009
0.008
0.007
-
0.22
0.16
0.50
6.55
0.41
0.71
6.11
†31.36
0.102
0.066
0.054
-
The maximum allowed search time, the idealized time, robot velocity, and
Lipschitz constant are set for both cases as follows: Case 1: Tmax = 100s,
Tidealized = 29.8s, T = 5s, V = 0.1m/s, M = 1, L = 20, = 0.05m;
Case 2: Tmax = 1000s, Tidealized = 140.6s, V = 0.2m/s, T = 20s,
M = 1, L = 200, = 0.2m. † Exhaustive search is done in parallel by 5
robots (2 in 1 qtr, other 3 separately in 3 qtrs of the arena).
B. Experiment 1: Comparative Analysis of Bayes-Swarm
Table I summarizes the performance of the complete,
synchronized, and explorative versions of Bayes-Swarm and
the baseline algorithm, in terms of completion time (τ) and
mapping error (δ). The results show that Bayes-Swarm and
its variations outperform the baseline exhaustive search (by
an order of magnitude better efficiency) in both case studies.
Although Bayes-Swarm requires greater completion time
than Bayes-Swarm-Sync in Case 1, the former is clearly
superior compared in Case 2. This performance benefit in a
complex environment is attributed to the planning flexibility
afforded by the absence of synchronization, which enables
planning paths of different length across the swarm.
Our asynchronous implementation, with communication
occurring only at waypoints and in a sequence among robots,
however introduces non-homogeneity in the models of the
environment across the swarm. An example of this is seen
from the discrepancies in the knowledge states of robot 1
(Fig. 3(a)) and robot 5 (Fig. 3(b)) around the 5s timepoint.
In the future, this issue could be addressed, while retaining
the asynchronous benefits, by designing the communication
schedule to be independent of the planning processes.
following: 1) knowledge gain over trajectories, as opposed
at points; 2) mitigating interactions among planned samples
of different robots; 3) time adaptive balance between explo-
ration /exploitation; and 4) accounting for motion constraints.
For evaluation, we used two cases with different arena
size and signal distribution complexity. Bayes-Swarm out-
performed an exhaustive search baseline by completing the
missions 17 times and 76 times faster than exhaustive search,
respectively in the simple and the complex cases. The
benefits of allowing asynchronous planning and exploita-
tion/exploration balance was also evident in the complex
case (noticeably lower mission completion time), studied by
setting control experiments (where synchronization and pure
exploration was enforced).
Scalability analysis of Bayes-Swarm demonstrated a some-
what superlinear reduction in completion time and mapping
error with increasing number of robots. The computing cost
per waypoint planning did increase sharply with increasing
swarm size, since swarm size exacerbates the cost of re-
fitting the GP (onboard swarmbots), which grows expo-
nentially along the mission as more samples are collected.
Efficient refitting approximations, e.g., particle learning, will
be explored in the future to address this concern. Another
direction of future research would be consideration of partial
observation (attributed to communication constraints), which
along with physical demonstration would allow more com-
prehensive appreciation of the Bayes-Swarm algorithm.
APPENDIX I
GAUSSIAN PROCESS MODEL
Gaussian process (GP) models provide non-parametric
surrogates [32] that can be used for Bayesian inference
over a function space
[33]. For a set of n observations,
D = xi, yii = 1 . . . n, GP expresses the observed values
yi as a summation of the approximating function f (xi)
and an additive noise i, i.e., yi = f (xi) + i. Assuming
the noise follows an independent,
identically distributed
n, we
Gaussian distribution with zero mean and variance, σ2
have ∼ N (0, σ2
n). The function f (x) can then be estimated
by a GP with mean µ(x) and a covariance kernel σ2(x):
f x, X, y ∼ N(cid:0)µ(x, X, y), σ2(x, X)(cid:1)
µ(x, X, y) = kn(x)T [K + σ2
(13)
n(x)I]−1(y − Φβ) (14)
σ2(x, X) = k(x, x) − kn(x)T [K + σ2
n(x)I]−1kn(x) (15)
Here Φ is the vector of explicit basis functions and K =
K(X, Xθ) is the covariance matrix, (K)ij = k(xi, xj),
with kn(x) = [k(x1, x), . . . , k(xn, x)]T . In this paper, the
squared exponential kernel is used to define the covariance
k(xi, x). The GP hyper-parameters are determined by max-
n, i.e.,
imizing the log-likelihood P as a function of β, θ, σ2
(16)
β, θ, σ2
n = arg max
β,θ,σ2
n
log P (yX, β, θ, σ2
n)
where
log P (yX, β, θ, σ2
n) = − 1
2
− Ns
2
(y − Φβ)T Λ(x)−1(y − Φβ)
log 2π − 1
2
log Λ(x)
(17)
Scalability analysis of Bayes-Swarm: Variation in performance
Fig. 4.
metrics (τ, ρ, computing time) with swarm sizes changing from 2 to 100.
While the purely explorative version (Bayes-Explorative)
expectedly provides lower mapping error by reducing the
knowledge uncertainty faster,
it falls significantly behind
both Bayes-Swarm-Sync and Bayes-Swarm in terms of search
completion time, for both environment cases (as evident from
Table I). This illustrates the importance of preserving the
exploitation/exploration balance.
To study the impact of the penalty factor, Γ, that promotes
waypoints away from those planned by peers, we ran Bayes-
Swarm without the penalty. Compared to Bayes-Swarm, the
"without penalty factor" version got stuck in the local signal
mode in case 1 and took 1.5 times the time to find the global
source in Case 2, with the latter's mapping error performance
being also poorer. These observations highlighted the value
of the penalty factor.
C. Experiment 2: Scalability Analysis of Bayes-Swarm
Here, we run Bayes-Swarm simulations on Case 2 with
swarm sizes varying from 2 to 100. Figure 4 illustrates the
results of this study in terms of the relative completion time,
mapping error, and computing time per planning instance.
The mapping error drops quickly with increasing swarm
size, given the resulting increased exploratory capability.
The mission completion time also reduces at a remarkable
rate between 2 to 50 robots, and then saturates (due to a
diminishing marginal utility). Some oscillations are observed
in this performance metric, since the penalty characteristics
become more aggressive as the swarm size (and thus robot
crowding) increases. The computation cost
increases, as
expected, due to the increasing size of sample sets over
which the GP model has to be updated by the robots at
each planning instance. This cost is however bounded, via
downsampling to a maximum of 1000 samples. Interestingly,
the downsampling does not noticeably affect the mission
performance improvements.
V. CONCLUSION
In this paper, we proposed an asynchronous and decentral-
ized swarm robotic algorithm to perform searching for the
maximum strength source of spatially distributed signals. To
this end, we exploit the batch Bayesian search concept, by
making important new modifications necessary to account for
the constraints and capabilities that differentiate embodied
search from a Bayesian Optimization process. Primarily,
a new acquisition function is designed to incorporate the
Swarm Size [22] E. Bonabeau, D. d. R. D. F. Marco, M. Dorigo, G. Th´eraulaz,
G. Theraulaz et al., Swarm intelligence: from natural to artificial
systems. Oxford university press, 1999, no. 1.
[23] W. Jatmiko, K. Sekiyama, and T. Fukuda, "A pso-based mobile sensor
network for odor source localization in dynamic environment: Theory,
simulation and measurement," in 2006 IEEE International Conference
on Evolutionary Computation.
IEEE, 2006, pp. 1036 -- 1043.
[24] K. Krishnanand, P. Amruth, M. Guruprasad, S. V. Bidargaddi, and
D. Ghose, "Glowworm-inspired robot swarm for simultaneous taxis
towards multiple radiation sources," in Proceedings 2006 IEEE Inter-
national Conference on Robotics and Automation, 2006. ICRA 2006.
IEEE, 2006, pp. 958 -- 963.
[25] P. Ghassemi, S. S. Lulekar, and S. Chowdhury, "Adaptive model
refinement with batch bayesian sampling for optimization of bio-
inspired flow tailoring," in AIAA Aviation 2019 Forum, 2019, p. 2983.
[26] A. Kolling, P. Walker, N. Chakraborty, K. Sycara, and M. Lewis,
"Human interaction with robot swarms: A survey," IEEE Transactions
on Human-Machine Systems, vol. 46, no. 1, pp. 9 -- 26, 2016.
[27] E. Klavins, "Communication complexity of multi-robot systems," in
Algorithmic Foundations of Robotics V. Springer, 2004, pp. 275 -- 291.
[28] M. C´ap, P. Nov´ak, M. Seleck`y, J. Faigl, and J. Vokffnek, "Asyn-
chronous decentralized prioritized planning for coordination in multi-
robot system," in 2013 IEEE/RSJ International Conference on Intelli-
gent Robots and Systems.
IEEE, 2013, pp. 3822 -- 3829.
[29] J. Gonz´alez, Z. Dai, P. Hennig, and N. Lawrence, "Batch bayesian
optimization via local penalization," in Artificial Intelligence and
Statistics, 2016, pp. 648 -- 657.
[30] J. Azimi, A. Fern, and X. Z. Fern, "Batch bayesian optimization via
simulation matching," in Advances in Neural Information Processing
Systems, 2010, pp. 109 -- 117.
[31] M. Li, K. Lu, H. Zhu, M. Chen, S. Mao, and B. Prabhakaran, "Robot
swarm communication networks: architectures, protocols, and appli-
cations," in 2008 Third International Conference on Communications
and Networking in China.
IEEE, 2008, pp. 162 -- 166.
[32] C. E. Rasmussen, "Gaussian processes in machine learning," in
Summer School on Machine Learning. Springer, 2003, pp. 63 -- 71.
[33] J. Snoek, H. Larochelle, and R. P. Adams, "Practical bayesian op-
timization of machine learning algorithms," in Advances in neural
information processing systems, 2012, pp. 2951 -- 2959.
REFERENCES
[1] W. Baetz, A. Kroll, and G. Bonow, "Mobile robots with active ir-
optical sensing for remote gas detection and source localization," in
2009 IEEE International Conference on Robotics and Automation.
IEEE, 2009, pp. 2773 -- 2778.
[2] K. Nagatani, S. Kiribayashi, Y. Okada, K. Otake, K. Yoshida, S. Ta-
dokoro, T. Nishimura, T. Yoshida, E. Koyanagi, M. Fukushima et al.,
"Emergency response to the nuclear accident at the fukushima daiichi
nuclear power plants using mobile rescue robots," Journal of Field
Robotics, vol. 30, no. 1, pp. 44 -- 63, 2013.
[3] W. Li, J. A. Farrell, S. Pang, and R. M. Arrieta, "Moth-inspired
chemical plume tracing on an autonomous underwater vehicle," IEEE
Transactions on Robotics, vol. 22, no. 2, pp. 292 -- 307, 2006.
[4] A. Viseras, T. Wiedemann, C. Manss, L. Magel, J. Mueller, D. Shutin,
and L. Merino, "Decentralized multi-agent exploration with online-
learning of gaussian processes," in 2016 IEEE International Confer-
ence on Robotics and Automation (ICRA).
IEEE, 2016, pp. 4222 --
4229.
[5] D. Song, C.-Y. Kim, and J. Yi, "Simultaneous localization of multiple
unknown and transient radio sources using a mobile robot," IEEE
Transactions on Robotics, vol. 28, no. 3, pp. 668 -- 680, 2012.
[6] O. De Silva, G. K. Mann, and R. G. Gosine, "Development of a
relative localization scheme for ground-aerial multi-robot systems," in
Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International
Conference on.
IEEE, 2012, pp. 870 -- 875.
[7] P. Ghassemi and S. Chowdhury, "Decentralized task allocation in
multi-robot systems via bipartite graph matching augmented with
fuzzy clustering," in ASME 2018 International Design Engineering
Technical Conferences and Computers and Information in Engineering
Conference. American Society of Mechanical Engineers, 2018, pp.
V02AT03A014 -- V02AT03A014.
[8] -- -- , "Decentralized informative path planning with exploration-
exploitation balance for swarm robotic search," arXiv preprint
arXiv:1905.09988, 2019.
[9] J. Pugh and A. Martinoli, "Inspiring and modeling multi-robot search
with particle swarm optimization," in Swarm Intelligence Symposium,
2007. SIS 2007. IEEE.
IEEE, 2007, pp. 332 -- 339.
[10] D. Gunning, "Explainable artificial intelligence (xai)," Defense Ad-
vanced Research Projects Agency (DARPA), nd Web, 2017.
[11] Y. Tan and Z.-y. Zheng, "Research advance in swarm robotics,"
Defence Technology, vol. 9, no. 1, pp. 18 -- 39, 2013.
[12] P. Odonkor, Z. Ball, and S. Chowdhury, "Distributed operation of
collaborating unmanned aerial vehicles for time-sensitive oil spill
mapping," Swarm and Evolutionary Computation, 2019.
[13] A. Sinha, R. Kaur, R. Kumar, and A. Bhondekar, "A cooperative con-
trol framework for odor source localization by multi-agent systems,"
2017.
[14] T. Wiedemann, D. Shutin, V. Hernandez, E. Schaffernicht, and A. J.
Lilienthal, "Bayesian gas source localization and exploration with a
multi-robot system using partial differential equation based model-
ing," in Olfaction and Electronic Nose (ISOEN), 2017 ISOCS/IEEE
International Symposium on.
IEEE, 2017, pp. 1 -- 3.
[15] B. Charrow, N. Michael, and V. Kumar, "Cooperative multi-robot
estimation and control for radio source localization," The International
Journal of Robotics Research, vol. 33, no. 4, pp. 569 -- 580, 2014.
[16] H. Hajieghrary, M. A. Hsieh, and I. B. Schwartz, "Multi-agent search
for source localization in a turbulent medium," Physics Letters A, vol.
380, no. 20, pp. 1698 -- 1705, 2016.
[17] P. Sujit and D. Ghose, "Negotiation schemes for multi-agent coopera-
tive search," Proceedings of the Institution of Mechanical Engineers,
Part G: Journal of Aerospace Engineering, vol. 223, no. 6, pp. 791 --
813, 2009.
[18] J. Kennedy, "Particle swarm optimization," Encyclopedia of machine
learning, pp. 760 -- 766, 2010.
[19] K. Krishnanand and D. Ghose, "Glowworm swarm optimisation: a new
method for optimising multi-modal functions," International Journal
of Computational Intelligence Studies, vol. 1, no. 1, pp. 93 -- 119, 2009.
[20] M. Senanayake, I. Senthooran, J. C. Barca, H. Chung, J. Kamruzza-
man, and M. Murshed, "Search and tracking algorithms for swarms
of robots: A survey," Robotics and Autonomous Systems, vol. 75, pp.
422 -- 434, 2016.
[21] J. Kennedy, "Swarm intelligence," in Handbook of nature-inspired and
innovative computing. Springer, 2006, pp. 187 -- 219.
|
1905.01361 | 1 | 1905 | 2019-04-23T14:20:44 | Fuzzy Q-Learning Based Multi-Agent System for Intelligent Traffic Control by a Game Theory Approach | [
"cs.MA",
"math.OC"
] | This paper introduces a multi-agent approach to adjust traffic lights based on traffic situation in order to reduce average delay time. In the traffic model, lights of each intersection are controlled by an autonomous agent. Since decision of each agent affects neighbor agents, this approach creates a classical non-stationary environment. Thus, each agent not only needs to learn from the past experience but also has to consider decision of neighbors to overcome dynamic changes of the traffic network. Fuzzy Q-learning and Game theory are employed to make policy based on previous experiences and decision of neighbor agents. Simulation results illustrate the advantage of the proposed method over fixed time, fuzzy, Q-learning and fuzzy Q-learning control methods. | cs.MA | cs |
Fuzzy Q-Learning Based Multi-Agent System for Intelligent Traffic Control by a
Game Theory Approach
Abolghasem Daeichian∗
Department of Electrical Engineering, Faculty of Engineering, Arak University, Arak, 38156-8-8349 Iran
Department of electrical engineering, Payam Institute of Higher Education,Golpayegan, Isfahan, Iran
Amir Haghani
This paper introduces a multi-agent approach to adjust traffic lights based on traffic situation in
order to reduce average delay time. In the traffic model, lights of each intersection are controlled by
an autonomous agent. Since decision of each agent affects neighbor agents, this approach creates
a classical non-stationary environment. Thus, each agent not only needs to learn from the past
experience but also has to consider decision of neighbors to overcome dynamic changes of the
traffic network. Fuzzy Q-learning and Game theory are employed to make policy based on previous
experiences and decision of neighbor agents. Simulation results illustrate the advantage of the
proposed method over fixed time, fuzzy, Q-learning and fuzzy Q-learning control methods.
Keywords: Traffic control, Multi-agent system, Game theory, Fuzzy Q-learning
I.
INTRODUCTION
Urbanization, increasing number of vehicles, and lack of transport infrastructures have increased travel time, fuel
consumption, and air pollution. Therefore, urban life equals with waste of time, less clean air, and acoustic pollution.
Conventional fixed traffic management systems are not able to fight complexity and dynamic of large traffic networks.
While, artificial intelligence (AI) are greatly employed to develop intelligent traffic systems (ITS)
[39, 40, 52, 57],
multi-agent system is an approach to model ITS [58, 63]. This framework consists of a population of intelligent and
autonomous agents work together in an environment [60]. Traffic lights [53], vehicles [36], and pedestrians [62] are
considered as agents in modeling of urban traffic networks. Each agent needs to learn from the past experiences which
is a key point to approximate a better decision-making policy. Multi-agent model-based [65] as well as model-free
[45] reinforcement learning (RL) techniques are widely used in researches on ITS [39, 56].
In a multitude of researches, any agent only considers its own traffic state in order to determine the control policy.
For example, single intersection with two phases is investigated in [35]. Length of vehicles queue waiting on the light
is considered as state which can be measured by the agent. It decides on extend green time or change it to the next
phase so that the number of vehicles waiting on the light is minimized. The results show superiority of Q-learning
agent over uniform traffic flows and constant-ratio traffic flows. In [65], traffic lights are considered as agents which
communicate with vehicles. The vehicles estimate their mean waiting time and transmit this time to traffic light
where a popular RL algorithm, namely Q-learning, is used to provide a control for traffic signal scheduling. Results
of this study show 22% reduction in waiting time compared to constant time lights. Multi-objective reinforcement
learning is utilized to control several traffic lights in [50]. Optimization goals include number of stops of a vehicle,
mean stopping time, and length of vehicles' queue on the next intersection. Its results indicate that multi-RL can
effectively prevent the queue spillovers under congested condition to avoid large-scale traffic jams. Bull et.al. used
learner classifiers to control light traffic including 4 intersections
In this research, traffic lights include two
phases at each intersection, where one phase is for moving north-south and one is for east-west. Controller at each
intersection, obtains optimum phase time through extracting if-then rules. Its results show that performance of the
traffic light using learner classifier system has improved significantly compared to constant time traffic light. In [61],
the learning purpose is modeled in such a way that states indications are based on the summation of the cars waiting
times. Obviously the more cars information is received, the model will be more complicated and state space will be
larger. This issue is one of the significant problems of large networks. Adaptive control, which is introduced in [56],
uses the approximate of a function as mapping of states to scheduling. Fuzzy inference engine is exploited to decrease
systematic faults of Q-algorithm in [55]. The results demonstrate that not only learning in fuzzy framework is done
faster than Q-learning but also delay in intersections is decreased considerably. A multi-agent fuzzy approach is
proposed in [51], where Q-learning updates the set of rule base in fussy inference engine. In [46] a new method which
[43].
∗ E-mail:[email protected], [email protected]; This is a pre-print of an article published in Arabian Journal for Science and
Engineering. The final authenticated version is available online at: https://doi.org/10.1007/s13369-017-3018-9
2
has the capability to estimate an incomplete model of environment is described for a given non-static environment.
This method is applied in a network composed of 9 intersections. The reported results show that this method has
better performance than the model-free methods and model-based methods, but could not be generalized and used
in larger networks.
In other researches, agents consider other agents in determination of their own control policy. For instance, coor-
dination among agents is desired in [54] where the agents not only consider number of waiting vehicles on its own
intersection but also they consider number of vehicles which have stopped in adjacent intersections. The RL is applied
on 5 intersections within three different scenario. The overall results show improvement in delay time. In [65], RL
is used to control the traffic in a grid where a type of cooperative learning simultaneously controls the traffic signals
and determines the optimal routes. One of the main drawbacks of this method is the high costs of communication
and information exchange, specifically when intersections of network are increased. Cooperative RL tries to extract
the knowledge from neighbor agents in a scheduling learning [59]. This method is implemented in an area of Dublin
including 64 intersections.
This paper introduces a hybrid fuzzy Q-learning and Game theory method for control of traffic lights in multi agent
framework. It exploits the benefits of fuzzification as well as interaction with other agents. The traffic network is
modeled by considering an autonomous agent controls in which each intersection decides on duration of green phase.
The number of vehicles in different inputs of the intersection are measured by the corresponding agent. Any agent
interacts with neighbour agents by getting a reward from each decision. This paper proposes that each agent fuzzify the
inputs and utilizes in a fuzzy inference system for fuzzy estimation of traffic model states. The agent uses a Q-learning
approach modified by Game theory to learn from the past experiences and consider the interaction with neighbor
agents. The agent gets a reward proportional to its own traffic state and a reward from each decision from neighbour
agents to update its Q-learning algorithm. The neighbour reward and its weighting in Q-value update is proposed
to be fuzzy in the proposed method. The proposed method is applied on a five-intersection traffic network. The
simulation results indicate that proposed method outperforms the fixed time, fuzzy, Q-learning and fuzzy Q-learning
control methods in the sense of average delay time.
This paper is unfolds as follows. After this introduction, Q-learning and its fuzzy version are described in the next
section. Section 3 is devoted to application of Game theory in ITS. Section 4 and 5 are about problem statement and
proposed solution, respectively. Simulation results are given in section 6. Finally, the paper is concluded in section 7.
II. Q-LEARNING AND FUZZY Q-LEARNING
The objective of agents which act in dynamic environments is making optimum decisions. If the agents are not
aware of rewards corresponding to various actions, selecting a proper action would be challenging. To achieve this
goal, learning adjusts agents' action selection based on collected data. Each agent tries to optimize its actions with
dynamic environment via trial and error in reinforcement learning (RL). The RL is actually how different situations
are mapped upon actions to receive the best results or the highest reward. In many cases, actions influence the reward
of next steps as well as affect the reward of its corresponding step. There are model-based [65] as well as model-free
[45] RL techniques.
In model-free RL, the agent does not need explicit modeling of the environment because its
actions could be directly selected based on rewards. Q-learning is a model-independent approach where the agent
does not access to transfer model
[34, 64]. Suppose that the agent is in a state s, performs an action a, from which it
gets the rewards r from the environment and the environment changes to state s(cid:48). This is given by a tuple in the form
of (s, a, r, s(cid:48)). State-action value which represents the expected total reward resulting from taking action a in state s
is denoted by Q-value Q(s, a). The agent starts with random value and after each action they receive a tuple in the
form of (s, a, r, s(cid:48)). For each tuple the value of state-action could be calculated according to the following equation:
Q(s, a) = (1 − α)Q(s, a) + α[r + γmaxQ(s(cid:48), a(cid:48)) − Q(s, a)]
(1)
where α ∈ [0, 1] is the learning rate of agent. α = 1 means that merely new information is considered and zero means
that the agent does not have any learning. γ ∈ [0, 1] is discount factor which determines future rewards. Zero value
for this factor makes the agent opportunist which means that the agent only considers current reward. On the other
hand, γ = 1 means that the agent will wait for a longer time to achieve a large reward. Q-learning will converge to
optimum value Q∗(s, a) with probability of one if all state-action pairs are experienced repetitively and learning rate
decrease during the time [55]. Generally, RL is useful for solving problems with small dimension discrete state and
action space. When the dimension of state and action space becomes larger, the size of search table will be so large
that it makes the algorithm very slow due to computational time. On the other hand, when the states or actions
are stated continuously, using search table will not be possible. To tackle this problem fuzzy theory is employed.
If the intelligent agent has a proper fuzzy set as expert knowledge about the desired area, the ambiguity could be
resolved. Thus, intelligent agent can understand vague objectives and unknown environment. In practice, the action
in large spaces is facilitated by eliminating Q-values table. In this method everything is based on quality values and
fuzzy inference. Fuzzy inference system (FIS) deals with input and Q-learning algorithm uses the follower section and
its active rules as states. Reward signal of Q-algorithm is built in accordance with fuzzy logic, environment reward
signal and performance estimation of current action. It is tried to select the action which maximizes the reward signal
[42, 47]. Learning system is able to select one action among j actions for each rule. j-th possible action in i-th rule
is denoted by a[i, j] and its value is shown by q[i, j] consider the following rules [42]:
3
If x is si then a[i, 1] with q[i, 1]
or a[i, 2] with q[i, 2]
...
or a[i, j] with q[i, j]
(2)
Learning should find the best result for each rule. If the agent selects an action which results in high value, it may
learn optimum policy. Thus, fuzzy inference system may obtain necessary action for each rule [42] .
III. GAME THEORY IN ITS
Relation between agent oriented environments and games theory originates from the fact that each state of agent-
oriented environments can be resembled to a game environment. Profit function of players would be current state
of the environment and goal of players is to move toward balanced or equilibrium point (reaching the best decision
making policy). Some scholars have studied the application of Game theory to control of traffic lights [48, 49]. They
integrate Game theory into the multi-agent interaction approach. Some of them suit the traffic problem into a rigorous
mathematical game model
[38, 41, 44] while others modify the learning method of agents based on Game theory
[66]. In [38], signalized intersections are modeled as finite controlled Markov chains and each intersection is seen
as non-cooperative game where each player try to minimize its queue. The solutions are given as Nash equilibrium
and Stackelberbg equilibrium and the simulation results indicate shorter queue length than adaptive control. In [41],
a two-player non-cooperative game is articulated between user seeking a path to minimize the expected trip cost
and choosing link performance scenarios to maximize the expected trip cost.
It shows that the Nash equilibrium
point measures network performance.
Intelligent traffic control is expressed as a Cournot game where the traffic
authority and the users choose their strategies simultaneously and as a bi-level Stackelberg game where the traffic
authority is the leader which determines the signal settings in anticipation of the user reactions.
In [66], Game
theory is used to address coordination between agents based on traffic signal control with Q-learning. It specifies
strategies (C(m) ={red light time plus 4sec, red light time plus 8sec, red light time minus 4s, red light time minus
8s,unchangeably}) and actions (S(n) ={east west straight and right turn, south north straight and right turn, east
west left turn, south north left turn}). Then, an interaction mathematical model via Game theory as a four parameter
group G = {B, A, I, U} is presented. B is a group of decision-makers as players. A is a group of any possible strategies
and actions, i.e. A = C(m) ∗ S(n). I represents the information which agents masters. U is the benefit function
which adopts Q-value. So, the Nash equilibrium is [66]:
i , a∗
−i) ≥ Ui(ai, a∗
−i)
Ui(a∗
(3)
where ai and a−i denote action of i-th agent and actions of other agents, respectively. a∗
−i represent the actions
at Nash equilibrium. The renewed Q-values in distributed reinforcement Q-learning is used to build the payoff values.
Q-value function is updated as:
i and a∗
n(cid:88)
j=1,j(cid:54)=i
Qi(si, ai) = (1 − αi)Qi(si, ai) + αi[ri(si, ai) +
f (i, j)rj(si, ai) + γmax (Qi(s(cid:48)
i, a(cid:48)
i) − Qi(si, ai))]
(4)
where α and γ are learning rate and discount factor, respectively. si and ai are current state of traffic environment
and current action, respectively. s(cid:48)
i is its next state, n is the number of traffic signal control agents surrounding i-th
agent, Qi(si, ai) is the Q-value function for i-th agent when selects action ai in state si. ri(si, ai) is reward function of
i-th agent and rj(si, ai) is reward function of j-th agent neighboring i-th agent. f (i, j) ∈ [0, 1] is a weighted function
which shows the effect of rj(si, ai) on i-th agent. Mathematical functions are suggested in [66] for r(s, a) and f (i, j).
Assumption of discrete action-state space and determination of reward and weighting functions are drawbacks of that
work.
IV. PROBLEM STATEMENTS
4
Consider a traffic network in which the lights of each intersection is controlled by an autonomous agents without
any centeralized management. Some sensors which are installed below the surface of surrounding streets or traffic
cameras of each intersection provide information about traffic situation for the corresponding agent. An agent has to
decide on duration of green light at North-South (NS) and West-East (WE) paths. Also, any agent interacts with
neighbour agents. Anyway, the agent is expected to schedule traffic lights optimally, in the sense of average delay,
based on the received information from its sensors and received information from neighbor agents.
The agents may have little knowledge about others' decision due to distribution of information. Even if an agent
has previous known information about others' decision, it is not valid as other agents are also learning. Thus, the
environment is dynamic and the behavior of other agents may change during time. Lack of prediction of other agents
causes uncertainty in problem solving procedure. This paper looks for a decision-making algorithm for lights control
agents which considers neighbour agents information in addition to its own information.
V. PROPOSED ALGORITHM
We consider a constant duration T for green plus red phases. So, if the agent determines the green phase duration
tg, then the red phase duration is tr = T − tg. Any typical agent i receives number of vehicles on the NS and WE
streets from its own sensors and the green phase duration of neighbour agent j in order to schedule its own green
phase duration. This paper proposes an autonomous agent with structure in Fig.1 to control each intersection.
FIG. 1. The proposed structure for a typical agent
The number of vehicles in WE and NS streets which are measured by sensors are fuzzified. Then, a fuzzy inference
engine with rules as Eq.2 are employed to fire the corresponding output membership functions. Finally, defuzzification
results to duration of green phase in NS path (tN S
). Thus, the duration of green phase in other path, WE, is
g = T − tN S
tW E
. We propose that, Q-value function which is updated by Eq.4 be the value of each action in Eq.2
which is denoted by q[i, j]. This update equation takes the neighbour agents' decision into account.
g
g
The i-th agent takes decision of neighbor agent j into account by reward rj(si, ai) and a weighting function f (i, j).
The reward is calculated based on average delay obtained from the decision made by the agent and current traffic
situation in a fuzzy manner. A fuzzy inference engine obtains these two inputs after fuzzification and gives the reward
after defuzzification; see Fig.1. weighting function f (i, j) ∈ [0, 1] shows the effect of rj(si, ai) on the decision of
i-th agent. This weight is also calculated by a fuzzy inference engine. This engine takes its own tg, the neighbour
agents' tg, and number of waited vehicles and gives f (i, j). Suitable choice for reward and weighting function plays a
significant role in agent learning. The agent with structure in Fig.1 runs the following algorithm:
1. Initial value of Qi-value for i-th traffic signal control agent is in the form of ∀(si, ai) : Qi(si, ai) = 0.
2. Observing si by WE and NS sensors which is the current state of i-th intersection.
3. Selecting a proper estimation for desired state by fuzzy inference system.
4. Calculating the reward related to i-th and j-th traffic signal control agent and the weighting function for
5
neighboring agents separately.
5. Observing new state s(cid:48)
i.
6. Updating Qi-value according to equation 4.
7. Returning to step 2 till the variation of Q-value becomes less than .
VI. SIMULATION RESULTS
Consider a traffic network with a center and four neighbor intersection. The delay in each intersection depends on
physical characteristics of the intersection, traffic light scheduling and number of cars in input streets. We utilized
traffic model which is given by the American Highway Capacity Manual (HCM) [37, Eq.20]:
d = 0.38
C(1 − λ)2
1 − λx
+ 173x2
(x − 1) +
(x − 1)2 +
(5)
(cid:34)
(cid:114)
(cid:35)
16x
C
c and x = v
where d, C, λ, and x are average delay (sec), cycle time (sec), green ratio, and degree of saturation, respectively.
λ = g
c , where c, g, and v are capacity (vehicle per hour), green time (sec), and input volume, respectively.
We use this model to calculate average delay based on the green phase duration and number of vehicles. For more
details of this equation we refer to [37].
Assume that C = T = 100sec and c = 3500veh/h. v is volume of vehicles entering each street which varies between
0 to 3500veh/h. g is duration of the green phase which each agent selects considering fuzzy Q-learning and interaction
with adjacent agents. The traffic network simulation algorithm is as follow:
1. The volume of vehicles entering each intersection (v) are randomly generated by a discrete uniform distribution
on the interval [0, 3500].
2. Average delay is calculated by Eq.5.
3. Each agent decides on the time of green phase g.
4. Go to step 1 until end of simulation time.
Assume structure of the agents as in Fig.1 with the Mamdani FIS with input membership function as in Fig.2
for number of input vehicles and Fig.3 for average delay to calculate the reward functions rj(si, ai). Centroid de-
fuzzification by the output membership function as in Fig.4 is considered to estimate a reward value in interval
[−3, 3].
FIG. 2. Membership function of number of vehicles enter the street for reward FIS
The weighting function FIS has number of vehicles, its own green phase duration and the neighbour agents' green
phase duration as inputs. Fig.2 shows the membership function for number of vehicles and Fig.5 depicts the member-
ship function for its own and neighbour green phase duration. Centroid defuzzification is applied to calculate weights
on output membership function as in Fig.6 which should be a value between 0 and 1.
Finally, the agent uses fuzzy Q-learning (Eq.2) with Q-value update rule (Eq.4) where learning and discount factor
are selected to be 0.5 and 0.7, respectively. The membership function for each measured number of vehicles is shown
in Fig.7. The output estimates green phase duration with membership functions as in Fig.8.
The proposed method is compared with Fuzzy Q-learning (using Eq.2 where q[i, j] is the Q-value which updates
with Eq.1), Q-learning (using Q-learning method with Q-value which updates with Eq.1), fuzzy(using traditional
Number of vehivles050010001500200025003000350000.20.40.60.81VLLMHVH 6
FIG. 3. Membership function of average delay for reward FIS
FIG. 4. Membership function of output for reward FIS
FIG. 5. Membership function of green phase duration for weighting function FIS
FIG. 6. Membership function of output for weighting function FIS
FIG. 7. Membership function of number of vehicles for fuzzy Q-learning
fuzzy inference method) and fixed time (tg = 60sec) in the sense of total average delay. Average delay in each time
interval is depicted in Fig.9 and the total average delay is illustrated in Fig.10. The results illustrate that total average
delay decrease from more than 50sec for fixed time scheduling to approximately 15sec for the proposed method.
Delay010203040506000.20.40.60.81VSSMHVHReward-3-2-1012300.20.40.60.81VSSMHVHGreen phase duration02040608010000.20.40.60.81LMHWeight00.20.40.60.8100.20.40.60.81VSSMHVHNumber of vehivles050010001500200025003000350000.20.40.60.81VLLHVH 7
FIG. 8. Membership function of green phase duration for fuzzy Q-learning
FIG. 9. Delay of the proposed method, fixed time, fuzzy Q-learning, Q-learning and fuzzy in each time step
FIG. 10. Average of delay for the proposed method, fixed time, fuzzy, Q-learning, fuzzy Q-learning
VII. CONCLUSION
In this study an intelligent control method of a controlling traffic network was performed to decrease average delay
time. Each traffic light is considered as a learning agent. This paper proposed a structure for the agents. Each agent
learn to decide on the duration of green phase through a fuzzy Q-learning algorithm which is modified by Game
theory. Each agent receives a reward from neighbour agents. The reward received from the neighbour and weighted
functions of neighboring agents are factors learning algorithm. These parameters are fuzzified through a FIS. Also,
the number of vehicles in each street is measured and fuzzified to be used in decision making process. The simulation
results were compared with fixed time method and other intelligent methods. The results revealed that our proposed
green phase duration02040608010000.20.40.60.81S1S2S3S4S5S6S7S8S9 method achieves considerable reduction of average delay in intersections.
8
[34] Abdoos M, Mozayani N, Bazzan AL (2011) Traffic light control in non-stationary environments based on multi agent
q-learning. In: 14th International IEEE Conference on Intelligent Transportation Systems (ITSC), IEEE, pp 1580 -- 1585
[35] Abdulhai B, Pringle R, Karakoulas GJ (2003) Reinforcement learning for true adaptive traffic signal control. Journal of
Transportation Engineering 129(3):278 -- 285
[36] Adler JL, Satapathy G, Manikonda V, Bowles B, Blue VJ (2005) A multi-agent approach to cooperative traffic management
and route guidance. Transportation Research Part B: Methodological 39(4):297 -- 318
[37] Akgungor AP, Bullen AGR (1999) Analytical delay models for signalized intersections. In: 69th ITE Annual Meeting,
Nevada, USA
[38] Alvarez I, Poznyak A, Malo A (2008) Urban traffic control problem a game theory approach. In: 47th IEEE Conference
on Decision and Control, IEEE, pp 2168 -- 2172
[39] Balaji P, German X, Srinivasan D (2010) Urban traffic signal control using reinforcement learning agents. IET Intelligent
Transport Systems 4(3):177 -- 188
[40] Bazzan AL, Klgl F (2014) A review on agent-based technology for traffic and transportation. The Knowledge Engineering
Review 29(03):375 -- 403
[41] Bell MG (2000) A game theory approach to measuring the performance reliability of transport networks. Transportation
Research Part B: Methodological 34(6):533 -- 545
[42] Bonarini A, Lazaric A, Montrone F, Restelli M (2009) Reinforcement distribution in fuzzy q-learning. Fuzzy sets and
systems 160(10):1420 -- 1443
[43] Bull L, ShaAban J, Tomlinson A, Addison JD, Heydecker BG (2004) Towards distributed adaptive control for road traffic
junction signals using learning classifier systems. In: Applications of Learning Classifier Systems, Springer, pp 276 -- 299
[44] Chen O, Ben-Akiva M (1998) Game-theoretic formulations of interaction between dynamic traffic control and dynamic
traffic assignment. Transportation Research Record: Journal of the Transportation Research Board (1617):179 -- 188
[45] Chin YK, Bolong N, Kiring A, Yang SS, Teo KTK (2011) Q-learning based traffic optimization in management of signal
timing plan. International Journal of Simulation, Systems, Science and Technology 12(3):29 -- 35
[46] Da Silva BC, Basso EW, Perotto FS, C Bazzan AL, Engel PM (2006) Improving reinforcement learning with context
detection. In: Proceedings of the fifth international joint conference on Autonomous agents and multiagent systems, ACM,
pp 810 -- 812
[47] Glowaty G (2005) Enhancements of fuzzy q-learning algorithm. Computer Science 7:77 -- 87
[48] Goyal T, Kaushal S (2017) An intelligent scheduling scheme for real-time traffic management using cooperative game
theory and ahp-topsis methods for next generation telecommunication networks. Expert Systems with Applications
[49] Groot N, Zaccour G, De Schutter B (2017) Hierarchical game theory for system-optimal control: Applications of reverse
stackelberg games in regulating marketing channels and traffic routing. IEEE Control Systems 37(2):129 -- 152
[50] Houli D, Zhiheng L, Yi Z (2010) Multiobjective reinforcement learning for traffic signal control using vehicular ad hoc
network. EURASIP journal on advances in signal processing 2010(1):724,035
[51] Iyer V, Jadhav R, Mavchi U, Abraham J (2016) Intelligent traffic signal synchronization using fuzzy logic and q-learning.
In: International Conference on Computing, Analytics and Security Trends (CAST), IEEE, pp 156 -- 161
[52] Kponyo J, Nwizege K, Opare K, Ahmed A, Hamdoun H, Akazua L, Alshehri S, Frank H (2016) A distributed intelligent
traffic system using ant colony optimization: A netlogo modeling approach. In: Systems Informatics, Modelling and
Simulation (SIMS), International Conference on, IEEE, pp 11 -- 17
[53] Liu Z (2007) A survey of intelligence methods in urban traffic signal control. IJCSNS International Journal of Computer
Science and Network Security 7(7):105 -- 112
[54] Medina JC, Hajbabaie A, Benekohal RF (2010) Arterial traffic control using reinforcement learning agents and information
from adjacent intersections in the state and reward structure. In: Intelligent Transportation Systems (ITSC), 2010 13th
International IEEE Conference on, IEEE, pp 525 -- 530
[55] Pacheco JC, Rossetti RJ (2010) Agent-based traffic control: a fuzzy q-learning approach. In: 13th International IEEE
Conference on Intelligent Transportation Systems (ITSC), IEEE, pp 1172 -- 1177
[56] Prashanth L, Bhatnagar S (2011) Reinforcement learning with function approximation for traffic signal control. IEEE
Transactions on Intelligent Transportation Systems 12(2):412 -- 421
[57] Rida M (2014) Modeling and optimization of decision-making process during loading and unloading operations at container
port. Arabian Journal for Science and Engineering 39(11):8395 -- 8408
[58] Roess RP, Prassas ES, McShane WR (2004) Traffic engineering. Prentice Hall
[59] Salkham A, Cunningham R, Garg A, Cahill V (2008) A collaborative reinforcement learning approach to urban traffic
control optimization. In: Proceedings of IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent
Agent Technology, IEEE Computer Society, pp 560 -- 566
[60] Schaefer M, Vokr´ınek J, Pinotti D, Tango F (2016) Multi-agent traffic simulation for development and validation of
autonomic car-to-car systems. In: Autonomic Road Transport Support Systems, Springer, pp 165 -- 180
[61] Steingrover M, Schouten R, Peelen S, Nijhuis E, Bakker B (2005) Reinforcement learning of traffic light controllers adapting
to traffic congestion. In: BNAIC, Citeseer, pp 216 -- 223
9
[62] Teknomo K (2006) Application of microscopic pedestrian simulation model. Transportation Research Part F: Traffic Psy-
chology and Behaviour 9(1):15 -- 27
[63] Vilarinho C, Tavares JP, Rossetti RJ (2017) Intelligent traffic lights: Green time period negotiation. Transportation
Research Procedia 22:325 -- 334
[64] Watkins CJ, Dayan P (1992) Q-learning. Machine learning 8(3-4):279 -- 292
[65] Wiering M (2000) Multi-agent reinforcement learning for traffic light control. In: ICML, pp 1151 -- 1158
[66] Xinhai X, Lunhui X (2009) Traffic signal control agent interaction model based on game theory and reinforcement learning.
In: International Forum on Computer Science-Technology and Applications, IEEE, vol 1, pp 164 -- 168
[34] Abdoos M, Mozayani N, Bazzan AL (2011) Traffic light control in non-stationary environments based on multi agent
q-learning. In: 14th International IEEE Conference on Intelligent Transportation Systems (ITSC), IEEE, pp 1580 -- 1585
[35] Abdulhai B, Pringle R, Karakoulas GJ (2003) Reinforcement learning for true adaptive traffic signal control. Journal of
Transportation Engineering 129(3):278 -- 285
[36] Adler JL, Satapathy G, Manikonda V, Bowles B, Blue VJ (2005) A multi-agent approach to cooperative traffic management
and route guidance. Transportation Research Part B: Methodological 39(4):297 -- 318
[37] Akgungor AP, Bullen AGR (1999) Analytical delay models for signalized intersections. In: 69th ITE Annual Meeting,
Nevada, USA
[38] Alvarez I, Poznyak A, Malo A (2008) Urban traffic control problem a game theory approach. In: 47th IEEE Conference
on Decision and Control, IEEE, pp 2168 -- 2172
[39] Balaji P, German X, Srinivasan D (2010) Urban traffic signal control using reinforcement learning agents. IET Intelligent
Transport Systems 4(3):177 -- 188
[40] Bazzan AL, Klgl F (2014) A review on agent-based technology for traffic and transportation. The Knowledge Engineering
Review 29(03):375 -- 403
[41] Bell MG (2000) A game theory approach to measuring the performance reliability of transport networks. Transportation
Research Part B: Methodological 34(6):533 -- 545
[42] Bonarini A, Lazaric A, Montrone F, Restelli M (2009) Reinforcement distribution in fuzzy q-learning. Fuzzy sets and
systems 160(10):1420 -- 1443
[43] Bull L, ShaAban J, Tomlinson A, Addison JD, Heydecker BG (2004) Towards distributed adaptive control for road traffic
junction signals using learning classifier systems. In: Applications of Learning Classifier Systems, Springer, pp 276 -- 299
[44] Chen O, Ben-Akiva M (1998) Game-theoretic formulations of interaction between dynamic traffic control and dynamic
traffic assignment. Transportation Research Record: Journal of the Transportation Research Board (1617):179 -- 188
[45] Chin YK, Bolong N, Kiring A, Yang SS, Teo KTK (2011) Q-learning based traffic optimization in management of signal
timing plan. International Journal of Simulation, Systems, Science and Technology 12(3):29 -- 35
[46] Da Silva BC, Basso EW, Perotto FS, C Bazzan AL, Engel PM (2006) Improving reinforcement learning with context
detection. In: Proceedings of the fifth international joint conference on Autonomous agents and multiagent systems, ACM,
pp 810 -- 812
[47] Glowaty G (2005) Enhancements of fuzzy q-learning algorithm. Computer Science 7:77 -- 87
[48] Goyal T, Kaushal S (2017) An intelligent scheduling scheme for real-time traffic management using cooperative game
theory and ahp-topsis methods for next generation telecommunication networks. Expert Systems with Applications
[49] Groot N, Zaccour G, De Schutter B (2017) Hierarchical game theory for system-optimal control: Applications of reverse
stackelberg games in regulating marketing channels and traffic routing. IEEE Control Systems 37(2):129 -- 152
[50] Houli D, Zhiheng L, Yi Z (2010) Multiobjective reinforcement learning for traffic signal control using vehicular ad hoc
network. EURASIP journal on advances in signal processing 2010(1):724,035
[51] Iyer V, Jadhav R, Mavchi U, Abraham J (2016) Intelligent traffic signal synchronization using fuzzy logic and q-learning.
In: International Conference on Computing, Analytics and Security Trends (CAST), IEEE, pp 156 -- 161
[52] Kponyo J, Nwizege K, Opare K, Ahmed A, Hamdoun H, Akazua L, Alshehri S, Frank H (2016) A distributed intelligent
traffic system using ant colony optimization: A netlogo modeling approach. In: Systems Informatics, Modelling and
Simulation (SIMS), International Conference on, IEEE, pp 11 -- 17
[53] Liu Z (2007) A survey of intelligence methods in urban traffic signal control. IJCSNS International Journal of Computer
Science and Network Security 7(7):105 -- 112
[54] Medina JC, Hajbabaie A, Benekohal RF (2010) Arterial traffic control using reinforcement learning agents and information
from adjacent intersections in the state and reward structure. In: Intelligent Transportation Systems (ITSC), 2010 13th
International IEEE Conference on, IEEE, pp 525 -- 530
[55] Pacheco JC, Rossetti RJ (2010) Agent-based traffic control: a fuzzy q-learning approach. In: 13th International IEEE
Conference on Intelligent Transportation Systems (ITSC), IEEE, pp 1172 -- 1177
[56] Prashanth L, Bhatnagar S (2011) Reinforcement learning with function approximation for traffic signal control. IEEE
Transactions on Intelligent Transportation Systems 12(2):412 -- 421
[57] Rida M (2014) Modeling and optimization of decision-making process during loading and unloading operations at container
port. Arabian Journal for Science and Engineering 39(11):8395 -- 8408
[58] Roess RP, Prassas ES, McShane WR (2004) Traffic engineering. Prentice Hall
[59] Salkham A, Cunningham R, Garg A, Cahill V (2008) A collaborative reinforcement learning approach to urban traffic
control optimization. In: Proceedings of IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent
Agent Technology, IEEE Computer Society, pp 560 -- 566
[60] Schaefer M, Vokr´ınek J, Pinotti D, Tango F (2016) Multi-agent traffic simulation for development and validation of
autonomic car-to-car systems. In: Autonomic Road Transport Support Systems, Springer, pp 165 -- 180
[61] Steingrover M, Schouten R, Peelen S, Nijhuis E, Bakker B (2005) Reinforcement learning of traffic light controllers adapting
to traffic congestion. In: BNAIC, Citeseer, pp 216 -- 223
[62] Teknomo K (2006) Application of microscopic pedestrian simulation model. Transportation Research Part F: Traffic Psy-
chology and Behaviour 9(1):15 -- 27
[63] Vilarinho C, Tavares JP, Rossetti RJ (2017) Intelligent traffic lights: Green time period negotiation. Transportation
Research Procedia 22:325 -- 334
[64] Watkins CJ, Dayan P (1992) Q-learning. Machine learning 8(3-4):279 -- 292
[65] Wiering M (2000) Multi-agent reinforcement learning for traffic light control. In: ICML, pp 1151 -- 1158
[66] Xinhai X, Lunhui X (2009) Traffic signal control agent interaction model based on game theory and reinforcement learning.
In: International Forum on Computer Science-Technology and Applications, IEEE, vol 1, pp 164 -- 168
10
|
1909.09461 | 1 | 1909 | 2019-09-12T11:32:48 | MCTS-based Automated Negotiation Agent | [
"cs.MA",
"cs.AI"
] | This paper introduces a new negotiating agent model for automated negotiation. We focus on applications without time pressure with multidi-mensional negotiation on both continuous and discrete domains. The agent bidding strategy relies on Monte Carlo Tree Search, which is a trendy method since it has been used with success on games with high branching factor such as Go. It also exploits opponent modeling techniques thanks to Gaussian process regression and Bayesian learning. Evaluation is done by confronting the existing agents that are able to negotiate in such context: Random Walker, Tit-for-tat and Nice Tit-for-Tat. None of those agents succeeds in beating our agent. Also, the modular and adaptive nature of our approach is a huge advantage when it comes to optimize it in specific applicative contexts. | cs.MA | cs | MCTS-based Automated Negotiation Agent
C´edric L.R. Buron∗, Zahia Guessoum†‡, Sylvain Ductor§
September 23, 2019
9
1
0
2
p
e
S
2
1
]
A
M
.
s
c
[
1
v
1
6
4
9
0
.
9
0
9
1
:
v
i
X
r
a
Abstract
This paper introduces a new negotiating agent
model for automated negotiation. We focus on
applications without time pressure with multidi-
mensional negotiation on both continuous and dis-
crete domains. The agent bidding strategy relies
on Monte Carlo Tree Search, which is a trendy
method since it has been used with success on
games with high branching factor such as Go. It
also exploits opponent modeling techniques thanks
to Gaussian process regression and Bayesian learn-
ing. Evaluation is done by confronting the exist-
ing agents that are able to negotiate in such con-
text: Random Walker, Tit-for-tat and Nice Tit-
for-Tat. None of those agents succeeds in beating
our agent. Also, the modular and adaptive nature
of our approach is a huge advantage when it comes
to optimize it in specific applicative contexts.
keywords: Automated negotiation, MCTS,
Supply Chain
1 Introduction
Negotiation is a form of interaction in which a
group of agents with conflicting interests and a
desire to cooperate try to reach a mutually ac-
ceptable agreement on an object of negotiation
[2]. The agents explore solutions according to a
predetermined protocol in order to find an accept-
able agreement. Being widely used in economic
France
[email protected]
∗Thales Research & Technology, Palaiseau France
†LIP6, Sorbonne Universit´e, Paris, France
‡CReSTIC, Universit´e de Reims Champagne Ardennes,
§Universidade Estadual do Cear´a
Cite as C´edric L. R. Buron, Zahia Guessoum, and
Sylvain Ductor.
"MCTS-based Automated Negotiation
Agent".
In The 22nd International Conference on Prin-
ciples and Practice of Multi-Agent Systems (PRIMA2019),
Turin, 2019
domains and with the rise of e-commerce applica-
tions, the question of automating negotiation has
gained a lot of interest in the field of artificial in-
telligence and multi-agent systems.
Many negotiation frameworks have been pro-
posed [13]. They may be characterized along dif-
ferent aspects, whether concerning the set of par-
bilateral or multilateral), agent
ticipants (e.g.
preferences (e.g.
linear or not), issues of nego-
tiated objects (e.g. discrete or continuous), or
even the characteristics of the interaction proto-
col (e.g. globally bounded in time or number of
rounds). They run negotiating agents that use
strategies to evaluate the received information and
make proposals. Several strategies have been pro-
posed. Either fixed or adaptive, most of them rely
on a known deadline (either in time or in rounds).
However in several applications, the deadline of an
agent may change over the negotiation. The ne-
gotiation horizon may vary depending on external
elements such as other opportunities. To the best
of our knowledge, these elements have not been
taken into account so far.
In this paper, we propose to handle this issue by
designing a loosely constrained adaptive strategy
for automated negotiation. This strategy consid-
ers that: 1) the agent preferences are nonlinear, 2)
the issues of negotiated objects can be both con-
tinuous and discrete and 3) the time pressure is
undefined, and therefore the deadline of the nego-
tiation. To cope with this objective, our agent is
based on General Game Playing [11] and Machine
Learning [2].
Its strategy relies on both Monte
Carlo Tree Search (MCTS), a heuristic technique
that has been used with success for many kinds
of games (see for instance [5, 23]), and opponent
modeling techniques in order to be more efficient.
The paper is organized as follows. Section 2
describes our targeted industrial application. Sec-
tion 3 provides some background on automated
negotiation and AI strategies for games. Section 4
gives the theoretical and formal setting for bar-
1
gaining in order to motivate the use of AI for
games. Section 5 introduces our strategy. Sec-
tion 6 gives some details on the agent implementa-
tion and shows its performance against a Random
Walker agent, Tit-for-Tat agent and Nice Tit-for-
Tat agent. The last section gives concluding re-
marks and perspectives.
2 Target application
This work is part of an industrial project that
addresses an economic application, the factoring.
This application requires a solution that complies
with the specific scope we consider here and which
is neglected by the literature.
When a company sells goods or services to an-
other company, it produces an invoice. The sell-
ing company is called supplier and the customer
is called debtor. Each country may define a legal
payment term of generally several weeks. More-
over, the principal may not pay within this pay-
ment term. In the supply chain, the debtor is often
much larger than the supplier. It can therefore im-
pose its own conditions at the expense of the sup-
plier. The consequences are quite harmful for the
latter: during those payment delays, its working
capital is reduced and hence its capacity to pro-
duce, fulfill future orders or pay its own suppliers.
Factoring is an interesting answer to this issue.
A funding company (called a factor) accepts to
fund the invoices of the supplier, by paying them
immediately less than their nominal amount and
assuming the delay of payment of the principal.
From the factor perspective -- generally a bank or
an investment fund -- the principal can be seen
as a short-term investment, where the risk of de-
faulting on payment depends on the reputation
of the principal. Since we consider the case where
the principal is much larger than the supplier, this
risk is lower and the rate is more affordable for the
supplier.
As this kind of funding may be recurrent, there
is a strong interest in automating the negotiation
between the supplier and the factor. Moreover,
some recent works have shown an increase in the
number of factoring marketplaces all around the
world [8]. However, there are several specificities
to this setting. The first specificity is the negoti-
ation domain. Several elements are negotiated at
the same time: the nominal amount to be funded
and the discount rate are the primary elements
of the negotiation. Also, for identical nominal
amounts, a factor may ask the supplier to sell in-
voices of certain principals it trusts. Finally, when
several invoices are available, the expected number
of financing days may also be negotiated. There-
fore, we consider complex issues that combine at
the same time elements of various kinds: contin-
uous (the discount rate), numeric (the numerical
amount and the financing days), and categorical
(the principal).
The second specificity is related to uncertainty
and resource availability. Automated negotiation
often considers a deadline, which defines the time
allocated to the negotiation. Most of the negotia-
tion strategies rely on this time pressure to com-
pute a concession rate. In our application, time
pressure is not constant over the negotiation. For
the factor, the time pressure depends both on the
money it has to invest and on the investment op-
portunities. If the factor has a lot of money and
few opportunities, the time pressure increases: the
factor sees not invested money as a loss. On the
contrary, when resources are limited and opportu-
nities are common, the factor tries to get a bet-
ter discount rate and the time pressure decreases.
For the supplier, the situation is even more unpre-
dictable. Time pressure depends on its opportuni-
ties to get new credit lines (including bank loans)
and even on the time the principal takes to pay
it: the negotiation may be brutally interrupted at
some point if the payment of an invoice makes it
useless for the supplier.
Automated negotiation components relying ei-
ther on a deadline commonly known by the agents,
or on a deadline private to each of them is there-
fore not applicable to our target applications. In
our target application the negotiation domain con-
sists of numerical, continuous and categorical is-
sues, the agent preferences are nonlinear, and the
time pressure is dynamic.
3 Related work
Our agent is at the meeting point of automated
negotiation and Monte Carlo methods applied to
games.
In this section, we introduce both do-
mains.
2
3.1 Automated Negotiation
Various authors have explored the negotiation
strategies with different perspectives. Most of
those works have identified three components that
make up the "BOA" architecture [1]: a Bidding
strategy defines the offers the agent sends to
its opponent, an Acceptance strategy defines
whether the agent accepts the offer it just received
or if it makes a counterproposal, and an Oppo-
nent modeling models some features of the op-
ponents, as its bidding strategy, its preference do-
main and its acceptance strategy. The latter aims
to improve the efficiency of the bidding strategy
and/or the acceptance strategy of the agent. The
following subsections present the related work for
each component.
3.1.1 Bidding Strategy
Bidding strategies may depend on several ele-
ments: the history, including the concessions made
by the opponent and/or a negotiation deadline,
the utility function of the agent, and the oppo-
nent model. Faratin et al. proposed the so-called
tactics that mainly rely on the criticality of the
resources, the remaining time before the deadline
is reached or the concessions made by the oppo-
nent. All of them except the last rely on a known
deadline. The latter has been extended to create
more complex strategies, as the Nice tit-For-Tat
agent [4], which uses learning techniques in order
to improve it. Genetic algorithms [15] use gener-
ated proposals as individuals. They rely on the
time pressure for the variation of their proposals
and make some that are acceptable for their op-
ponent.
This paper advances the state of the art in the
bidding strategies by introducing and evaluating a
strategy that considers the negotiation as a game
and uses the very efficient Monte Carlo Methods.
3.1.2 Acceptance Strategy
Acceptance strategies can be divided into two
main categories [3]. The first category is called
"myopic strategies" as they only consider the last
bid of the opponent. An agent may accept an offer
when (1) it is better than the new one produced
by its own bidding strategy, (2) it is better than
the last one made by the agent, (3) it is above
a predetermined threshold, or (4) it embodies any
combination of the previous ones. The second cat-
egory consists of "optimal strategies" [3] that rely
on an opponent bidding strategy in order to opti-
mize the expected utility. They are based on the
concessions made by the opponent and a predic-
tion that the expected utility of the agent should
increase while the deadline is getting closer. So,
the first category is not suitable for our context.
We therefore propose to use the second category
in our agent model.
3.1.3 Opponent modeling
[2] presents an exhaustive review of the opponent
modeling techniques related to automated negoti-
ation. They are generally used to model (1) op-
ponent bidding strategy, (2) its utility and (3) its
acceptance strategy as well as private deadline and
a reserve price, depending on which of these ele-
ments are relevant in each context.
There are two main methods to model adap-
tive bidding strategy which does not rely on the
deadline: neural networks and time series-based
techniques. Neural networks use a fixed number of
previous offers as input, and the expected value for
the next proposal as output. Time series methods
can be generalized very easily. Among them, the
Gaussian process regression is a stochastic tech-
nique which has been used with success by [24].
Due to its nature, it can generate various propos-
als at each negotiation turn. Those proposals are
proportional to a likelihood provided by the re-
gression. We select this technique to model the
opponent bidding strategy since it is particularly
adapted to the Monte Carlo Tree Search.
The opponent utility is generally considered as
the weighted average of partial utility functions
for each issue. Two families of methods are used
to model them. The first one is based on the fre-
quency of each value among the previous opponent
bids. The methods of this family make the hypoth-
esis that the most frequent values are the ones the
opponent prefers, and that the most stable issues
are the most important for it. They are relevant in
the cases where the negotiation domain only con-
sists of discrete issues; their extension to the con-
tinuous case is not suitable to complex domains,
as it requires the definition of a distance function
which depends on the negotiation domain. The
second family of methods is based on Bayesian
Learning [14].
It is well suited for the continu-
ous case and can be easily extended to categorical
3
domains. We use it as is for the numerical issues
and make an extension for the categorical ones.
The latter is presented in section 5.1.2.
ber of possible moves at each step to be finite.
Also,
[16] does not require opponent modeling,
since the opponent utility profile is known.
Their is two ways to learn the opponent accep-
tance strategy: either by assuming that the op-
ponent has a myopic strategy or by using neural
networks [9]. The latter is quite expensive in terms
of computation time. The weights of the network
must be updated each time the opponent makes a
new proposal.
3.2 Monte Carlo Tree Search
Monte Carlo methods are regularly used as heuris-
tics for games. R´emi Coulom [6] proposes a
method to combine the construction of a game tree
-- a traditional method for games that has proved
to be very to be effective -- with Monte Carlo tech-
niques. This method is called Monte Carlo Tree
Search (MCTS) and it has been improved using
various extensions [5], including pruning the less
promising branches of the tree. It has met great
success in games, particularly games with high
branching factor [23].
MCTS consists of 4 steps. A selection is dedi-
cated to the exploration of the already built part
of the tree, based on a predefined strategy. While
exploring a node, the algorithm chooses whether
to explore a lower-level branch or expand a new
branch. In the latter case, there is an expansion
of a new node; it is created just below the last ex-
plored one. Once a new node has been expanded,
a simulation of the game is performed until a fi-
nal state is reached. Finally, outcomes of the final
state are computed and a backpropagation of
them is made over all the nodes that have been
explored.
[16] presents a recent attempt to exploit MCTS
for General Game Playing, with Automated Ne-
gotiation as a potential application. The negotia-
tion domain considered is limited to a single-issue,
discrete domain with complete information (each
agent knows what is the optimal deal for its op-
ponent). Our work is specifically made for recent
evolution of Automated Negotiation, focusing thus
on multi-issues, combining continuous, numerical
and categorical domains with incomplete informa-
tion (the agent has no information on the oppo-
nent utility profile). These differences have con-
sequences on the technical aspects of these works.
The referenced works use Upper Confidence Trees,
inapplicable in our case since it imposes the num-
4
4 Negotiation as a game
Monte Carlo Tree Search has been applied with
success to extensive games.
In this section, we
show how it is possible to represent negotiation
using this model. We first associate each aspect of
negotiation to a game element. We then describe
specificities of negotiation that prevent us from
using most common MCTS selection, expansion
and simulation strategies.
An extensive game [20] consists of a set of play-
ers, the set of all possible game histories, a func-
tion mapping each non-terminal history to the
player who must play then and a preference pro-
file. By using this definition, it is possible to define
a bargaining B as an extensive game:
Definition 1 (Bargaining). A bargaining can
be represented as triple B =
where:
H, A, (ui)i∈(cid:74)1,2(cid:75)
(cid:16)
(cid:17)
1. A is the set of two players: the buyer (player
1) and the seller (player 2),
2. H is the set of possible histories of the
negotiation. Each history consists of the se-
quence of messages the agents sent to each
other:
proposals, acceptance and rejection
messages. Terminal histories are the histories
ending by acceptance or rejection, and infi-
nite histories. Each message is a pair (α, c)
where α is the speech act (performative) of
the message and c is the content of the mes-
sage, i.e. a list of couples (k, v) where k is
the key of an issue of the negotiation domain
and v is the corresponding value. The history
can be divided in two parts, each part corre-
sponding to the messages sent by one of the
agents: hi = (α, c)i,
3. the player function is based on the parity of
the size of the history (we suppose that the
buyer (player 1) always plays first). Therefore
∀h ∈ H, player(h) = 2 − (h mod 2) where
h is the size of h,
4. the preference profile, ui, i ∈ {1, 2}, is an
evaluation of terminal histories with regards
to each player.
If the history ends with an
strategy. Each of the agent submodules and their
interactions are described in this section.
Bidding strategy
Selection
Expansion
Simulation
Backpropagation
Strategy (Gaussian
Process Regression)
Utility (Bayesian
Learning)
Opponent modeling
Figure 1: Interaction between the modules of our
agent
A
c
c
e
p
t
a
n
c
e
s
t
r
a
t
e
g
y
5.1 Opponent modeling
In order to improve the efficiency of MCTS, we
model both the bidding strategy and the utility of
the opponent.
5.1.1 Bidding Strategy Modeling
The goal of this model is to predict what pro-
posal the opponent will make at turn x∗. To do
so, we use Gaussian Process Regression [21]. This
method produces a Gaussian with the mean corre-
sponding to the value predicted by the algorithm
and a standard deviation corresponding to the un-
certainty induced by the model.
The first step is to compute the covariance ma-
trix K which represents the proximity between the
turns (xi)i∈(cid:74)1,n(cid:75) of the sequence, based on a covari-
ance function, also called kernel k. Let
k(x1, x1)
...
(1)
. . .
k(x1, xn)
...
acceptance of the agent, ui returns the util-
ity associated by the agent, if not it returns
a specific value that may depend on the en-
gaged resources.
The representation of bargaining as a game has
already been investigated in [19, 22] for single is-
sue bargaining. More complex domains has been
initially dealt by making the assumption that the
agents proposals are independent of the oppo-
nent's ones [10]. However in recent advances in
automated negotiation, the agent bidding strate-
gies are generally adaptive i.e. the proposals made
by an agent depend on its opponent's ones.
However, negotiation is not a classical combi-
natorial game as Chess or Go.
Its resolution is
a challenge to the MCTS approach for three rea-
sons. First, it is a non-zero sum game: agents try
to find a mutually beneficial agreement which is
often much better for both agents than their re-
serve utility i.e.
the situation where the agents
do not find an agreement. Second, it is an in-
complete information game: the agent preference
profiles are unknown to their opponents and gen-
erally modeled by them. These two specificities
make it impossible to use the most common imple-
mentation of MCTS, the Upper Confidence Tree
[17]. Last, we consider a large and complex do-
main that encompasses numeric, continuous and
categorical issues, with nonlinear utility functions
and possibly infinite game trees. This has several
consequences on the way the tree is explored, in
particular on the criterion followed to expand a
new node.
5 MCTS-based agent
K =
As we explained in the previous section, nego-
tiation is a particular game.
It is therefore re-
quired to adapt the heuristics traditionally used
for games to its specificities. In this section, we
present our automated negotiation agent relying
on MCTS. The agent architecture is composed of
three modules presented in Figure 1. The bidding
strategy module implements MCTS and uses the
opponent modeling module. The latter consists of
two submodules: one models the opponent utility,
the other models its bidding strategy. The last
module is the acceptance strategy, which makes
a comparison between the last proposal from the
opponent and the bid generated by the bidding
k(xn, x1)
. . .
k(xn, xn)
we then compute the distance between the turn
of the predicted proposal x∗ and the previous
turns in the vector K∗:
K∗ = (k(x∗, x1), . . . , k(x∗, xn))
(2)
The Gaussian process regression relies on the
supposition that all these values are the dimen-
sions of a multivariate Gaussian. Using results on
multivariate Gaussian, we can compute
y∗ = K∗K−1y
σ2∗ = Var(y∗) = K∗∗ − K∗K−1K(cid:62)
∗
(3)
(4)
5
where K∗∗ = k(x∗, x∗). The result corresponds
to a Gaussian random variable with mean y∗ and
standard deviation σ∗.
One of the capital aspects of Gaussian process
regression is the choice of the kernel. The most
common ones are radial basis functions (RBF), ra-
tional quadratic functions (RQF), Mat´ern kernel
and exponential sine squared (ESS). These ker-
nels are used to define distance between the turns
of the bargaining. We tested these four kernels
on various negotiations among finalists of ANAC
2014 (bilateral general-purpose negotiation with
nonlinear utilities). The table 1 shows the re-
sults of GPR for each of the aforementioned ker-
nels. We generated randomly 25 negotiation ses-
sions and modeled two agents using each kernel.
We got a total of 50 models by kernel. Each bid
of each sequence is predicted using previous pro-
posals and is used to predict following ones. The
table shows average Euclidean distance between
the actual proposals and the predicted sequences.
The lower the value, the closer is the prediction to
the actual proposal. The kernel that got the best
result is the Rational Quadratic Function. Our
agent therefore uses this one.
Kernel
avg. distance
RBF
43.288
RQF
17.766
Mat´ern
43.228
• there is some c in [a, b] such that t is linear on
[a, c] and [c, b], t(a) = 0, t(b) = 0 and t(c) = 1.
The method can be divided into two steps. First,
the agent generates a predetermined number of hy-
potheses on the utility function. These hypothe-
ses are composed of weighted sums of triangular
functions (one per issue). Each issue is therefore
associated with a weight and a triangular function.
The estimated utility of the opponent is the
weighted sum of these hypotheses where each
weight is the probability computed using Bayesian
learning. This method does not make any suppo-
sition on the opponent strategy, which makes it
more general than the frequency based techniques.
This method can be naturally extended to the
categorical issues. Given a categorical issue C =
{C1, . . . , Cn}, the partial utility function is chosen
among the set of functions of [0, 1]C.
5.1.3 Acceptance Strategy Model
Acceptance strategy is modeled in a very simple
way: a simulated agent accepts the proposal from
its opponent if and only if its utility is better
than the utility generated by the bidding strat-
egy model. This method presented by Baarslag et
al. in [2] is not computationally expensive.
ESS
22.292
Table 1: Average Euclidean distance between ac-
tual proposals of a bargaining and values predicted
by GPR, depending on the used kernel.
This method also allows to make predictions
on the categorical issues. The method presented
in chapter 3 of [21] is also possible and relies on
Monte Carlo method for some integration estima-
tion.
5.1.2 Preference Profile Modeling
Bayesian learning described in [14] makes the only
supposition that an agent makes concessions at
roughly constant rate. Though this constraint
may seem tough, it is relatively low in compari-
son with other methods.
The opponent utility is approximated by a
weighted sum of triangular functions. A function
t of [a, b] ⊂ R in [0, 1] is called triangular if and
only if:
• t is linear and either t(a) = 0 and t(b) = 1 or
t(a) = 1 et t(b) = 0, or:
5.2 MCTS-Based Bidding Strategy
Monte Carlo methods are very adaptive, and
achieve promising results in various games, includ-
ing games with high branching factor. In this sec-
tion, we describe the way we have adapted these
methods to the negotiation context.
5.2.1 Raw MCTS
As we explained before, MCTS is a general algo-
rithm and relies on several strategies. Each time
an agent needs to take a decision, it generates a
new tree and explores it using MCTS. The most
common implementation of MCTS is Upper Con-
fidence Tree (UCT). This method has proved to
be very efficient, particularly for Go. Neverthe-
less, it expands a new node whenever it explores
a node whose children have not all been explored,
which is not applicable when issues are continuous.
Beyond that, in the case where the branching fac-
tor is not small in comparison with the number
of simulation, UCT keeps expanding a new node
without exploring deeper nodes, which loses the
6
interest of MCTS compared to flat Monte Carlo.
In our context, it is therefore necessary to define
a different implementation of MCTS:
Selection For
this
step, we use progressive
widening, as described by [7]. The expansion
criterion of the progressive widening states
that a new node is expanded if and only if:
p ≥ nc
nα
(5)
where np is the number of times the parent
has been simulated, nc is its number of chil-
dren and α is a parameter of the model. If the
result is that a new node is not expanded, the
selected node is the node i maximizing:
(cid:115)
Wi =
si
ni + 1
+ C × nα
ln(n)
ni + 1
(6)
where n is the total number of simulations of
the tree, si is the score of the node i and C is
also a parameter of the model.
Expansion The content of the expanded node is
chosen randomly among all the possible bids
of the domain with an even distribution
Simulation During the simulation, the model of
the opponent bidding strategy is used in order
to make the simulation more representative.
The model of the opponent utility and its ac-
ceptance strategy are used to decide when it
accepts a proposal.
Backpropagation The backpropagation also
uses the opponent utility model. The utility
of both agents is computed and the scores of
both agents is updated for each visited node.
The utility of the agent itself is computed us-
ing its real preference profile.
5.2.2 Pruning
In order to explore only the interesting nodes for
the agent, it is possible to use the knowledge of
the agent on the game to prune the less promising
branches of the tree. Though we do not have much
information in our context, we decide that the op-
ponent should find any offer it made acceptable.
We therefore decide to prune all the branches of
the tree where our agent makes a proposal less
interesting than the best proposal received from
the opponent: the goal is to use the best proposal
of the opponent as a lower bound and to try to
improve on this basic value.
7
6 Experiments
In order to evaluate our agent, we use the Genius
[18] framework. We confront our agent with the
only agents that do not require a deadline for their
strategy to the best of our knowledge: two vari-
ants of the Tit-for-Tat agent and a RandomWalker
agent. In this section, we describe the implemen-
tations of our agent and those three. We then
describe in detail the experimental protocol and
finally we present the achieved results.
6.1 Implementation
Our agent is developed in Java and consists of a
set of independent modules connected to Genius
platform thanks to interfaces. Figure 2 presents
this architecture (the R represents the messages
sent and received by the agent).
It consists of
the components of Section 5: a module for the
MCTS-based bidding strategy (with classes for the
Monte Carlo Tree, its nodes, etc.), a module for
opponent utility modeling and another one for the
opponent bidding strategy modeling. We use the
same acceptance strategy for both our agent and
the model it uses for its opponent. As they are
very simple, there is no module dedicated to them.
As MCTS are computationally expensive and
take quite a long time, we parallelized them. The
stochastic aspect of the opponent bidding strat-
egy model ensures that two simulations going from
the same branch of the tree are not similar, while
still prioritizing the most probable values. The
Gaussian process regression, which relies on ma-
trix computation, has been developed using the
Jama library1. The parameters of the kernel are
optimized using the Apache Commons Math li-
brary utilities2.
The time taken by our agent for each round has
been set empirically. When our agent makes a
choice among 200 bids, there is generally one that
results into a high utility both for it and its oppo-
nent. In order to explore the tree in depth, we let it
enough time to generate about 50'000 simulations.
Then, to make a proposal, our agent takes about
3 minutes. This duration meets the expectation of
the real-world applications of our industrial part-
ners. From these values, we get α = 0.489.
1http://math.nist.gov/javanumerics/
2http://commons.apache.org/proper/commons-math/
Strategy modeling
ǡ
Previous offers
Lj
Genius connector
R
Utility modeling
(cid:255)
MCTS-based
3
bidding
strategy
Agent
R
R
Figure 2: Software architecture of the modules of
our agent
6.2 Opponent's Description
In this section, we provide a description of the
confronted agents. It is impossible to compare our
agents with state-of-the-art ones, as they all rely
on the supposition that there is a publicly known
deadline. The only three agents able to negotiate
without a given deadline are the RandomWalker,
the Tit-for-Tat agent and its more evolved variant,
the Nice Tit-for-Tat.
6.2.1 RandomWalker
The RandomWalker is described by Baarslag in [3]
and makes random proposals.
6.2.2 Tit-for-Tat
Tit-for-tat agent was first described by Faratin et
al.
in [10]. This agent makes a concession when-
ever its opponent makes concessions itself. Several
possible implementations are given. Here, when
it has received fewer than 2 proposals the agent
makes the most interesting proposal from its own
perspective (it generates 10'000 random proposals
and chooses the best one for its utility function).
For the other proposals, the agent looks at the last
two proposals of its opponent and computes the
made concession. This concession may be positive
or negative. It adds this concession to the utility
of its own last proposal. It then searches for a pro-
posal with the closest utility to this target value.
In our implementation, 10'000 proposals are gen-
erated to this end.
6.2.3 Nice Tit-for-Tat
Nice Tit-for-Tat is somehow, a more evolved ver-
sion of the Tit-for-Tat. It has been described by
8
Baarslag et al.
in [4]. The goal of this work is
to comply with domains where mutual agreement
is possible. The agent then uses the same oppo-
nent utility modeling as our agent and uses it to
estimate the Nash point of the setting, i.e.
the
agreement maximizing the product of the utilities
of the agents. The concession rate is computed
between the first and the last bids of the oppo-
nent as the percentile of the distance between its
first bid and the Nash point. The corresponding
concession is made from the agent point of view.
Among the equivalent offers, the utility model is
also used in order to choose the best bid for the op-
ponent among equivalent ones for the agent. The
only difference with the version used in our exper-
iments and the agent proposed in [4] is the accep-
tance strategy. Indeed, the strategy proposed in
the original version which is presented in [3] de-
pends on the deadline of the negotiation in order
to take the time pressure into account. Here, we
provide the Nice Tit-for-Tat with a simplified ver-
sion of its acceptance strategy, which corresponds
to the same acceptance strategy as our agent.
6.3 Experimental Protocol
Genius makes it possible to negotiate on numeri-
cal or categorical issues, but not yet on continuous
ones. In order to evaluate our agent, we want to
target a negotiation domain that is at the same
time neutral enough to show the generic aspects
of our work, but complex enough to motivate its
use. ANAC is an international competition used
to determine the effective negotiation strategies.
The negotiation domain used in ANAC 2014 [12]
fits well to this objective. Subsequent competi-
tions focused on multilateral negotiation and spe-
cific application domains. In ANAC 2014 domain,
issues are numerical, varying from 1 to 10. Several
domains have been proposed, varying from 10 to
50 issues. In order to reduce computational com-
plexity which is not the concern of this work, we
use the 10-issue version. The utility functions are
non-linear, and the reserve utility is set to 0, which
is the minimal outcome value for the agents. As
the time pressure is supposed to vary over time,
we do not use a discount rate. While it does not
exactly correspond to the targeted application, its
complexity (1010 possible proposals) makes a suit-
able test bench for our negotiation strategy. In or-
der to simulate the fact that the time pressure is
unknown to the agent, we put a very large dead-
1
0.5
0
1
0.5
0
1
0.5
0
profile 1 profile 2
profile 1 profile 2
profile 1 profile 2
(a) Our agent vs Random
Walker
(b) Our agent vs Tit-for-tat
(c) Nice Tit-for-tat vs Random
Walker
Our agent Random Walker Tit-for-tat Nice Tit-for-tat
Figure 3: Average utility of negotiating agents
line, so that it is never reached.
6.4 Results
Figure 3 displays the utility of the agents when ne-
gotiating with each other using a histogram. The
results are averaged over 20 negotiation sessions
with each profile, with error bars representing the
standard deviation from the average.
Note that the two preference profiles are very
different from each other, and not symmetrical at
all. This specificity explains the fact that for all
the agents in all configurations, utility is always
higher with Profile 2 than it is with Profile 1.
As shown on Figure 3a, our agent is able to
beat the Random Walker in every situation, even
when its preference profile is Profile 1 and Random
Walker's is Profile 2. It is interesting to note that
the negotiations with Random Walker are very
short with only 3.1 proposals in average: 2.5 when
it gets Profile 1 and 3.7 proposals when our agent
gets Profile 2. This difference can be explained
by the fact that it is easy to find agreements with
very high results for Profile 1 (0.9 or more) and
high results for Profile 2 (0.6). In most of the ne-
gotiations, the first proposal of our agent is of this
kind. In that case, Random Walker is more likely
to generate a proposal with utility lower than the
one proposed by our agent and accepts it, gener-
ating a utility of 0.6 for Random Walker and a
negotiation session consisting of a single proposal.
Negotiating with Tit-for-Tat is harder, as we
can see on Figure 3b. Our agent gets a lower utility
than the Random Walker but is able to beat Tit-
for-Tat. The expectation level of Tit-for-Tat also
generates much longer negotiations: 34.2 propos-
als on average with 31.55 proposals when it gets
Profile 1 and 36.85 proposals when our agent gets
Profile 2. This result can be explained the same
way as the results of the negotiation with Random
Walker.
The negotiations with Nice Tit-for-Tat never
ends: the agents keep negotiating forever. Our
MCTS-based method refuses to make a concession
significant enough to have a chance to be accepted
by Nice Tit-for-Tat, considering the high expecta-
tion it has by using the Nash point. Reciprocally,
Nice Tit-for-Tat, without time pressure, and dy-
namic adaptation of its acceptance strategy, does
not accept the proposals of our agent. By looking
at the internal state of the Nice Tit-for-Tat, we
also see that its estimation of the Nash point is
incorrect: it expects a utility above the real one.
We propose instead an indirect evaluation
by confronting Nice Tit-for-Tat with Random
Walker, in the same setting. The results are rep-
resented on Figure 3c. The performances of both
agents are comparable, considering the standard
deviation of the series.
7 Conclusion
In this paper, we presented an automated negoti-
ation agent able to negotiate in a context where
agents do not have predetermined deadline, nei-
ther in time nor in rounds, and where the negoti-
ation domain can be composed of numerical, con-
tinuous and categorical issues. We described this
9
setting as an extensive game and described a ne-
gotiation strategy based on a specific implemen-
tation of MCTS relying on two opponent mod-
els and a pruning strategy. One of them is the
Gaussian process regression, which relies on a co-
variance function. We tested several covariance
functions and chose the one that provides better
results in context similar to ours.
Experiments were run in the context of a large
negotiation domain, with nonlinear utility func-
tions using different preference profiles. The ex-
perimental results are promising: against all the
agents that can negotiate in its negotiation do-
main, our agent outperformed Random Walker
and Tit-for-Tat and draws with Nice Tit-for-Tat.
This work therefore indicates that techniques from
games such as MCTS can be used with success in
automated negotiation. However, the modularity
of the architecture and the variety of strategies
proposed on General Game Playing and Machine
Learning areas are a huge advantage when it comes
to optimizing the agent for a specific application
domain.
Among the perspectives of this work, we would
like to create a customized version of our agent
and adapt it to the context where the deadline is
known, in order to make it available for these ap-
plications. Our agent can already be used in this
context, but the fact that it does not exploit this
information may make it less efficient than its op-
ponents. Another possible direction would be to
adapt it to the multilateral context. In fact, there
would be little modification to make it available
for a context of stacked alternate protocol or a
many-to-many bargaining scenario, since MCTS
has already been used in n-player games. The use
of our agent in its industrial context, in particu-
lar with corresponding negotiation domains would
yield very interesting results. Last, we would like
to improve our agent by using MCTS variations.
It would be interesting for instance to test other
kinds of pruning. The use of traditional MCTS
techniques such as Rapid Action Value Estima-
tion or All Moves As First to reduce the number
of simulations while keeping their intrinsic quali-
ties would be interesting.
References
[1] Baarslag, T.: Exploring the Strategy Space
of Negotiating Agents: A Framework for Bid-
10
ding, Learning and Accepting in Automated
Negotiation. Ph.D. thesis, Delft University of
Technology (2016)
[2] Baarslag, T., Hendrikx, M.J.C., Hindriks,
K.V., Jonker, C.M.: Learning about the op-
ponent in automated bilateral negotiation:
a comprehensive survey of opponent mod-
eling techniques. Autonomous Agents and
Multi-Agent Systems 20(1), 1 -- 50 (2015).
https://doi.org/10.1007/s10458-015-9309-1
[3] Baarslag, T., Hindriks, K.V.: Accepting op-
timally in automated negotiation with in-
complete information. In: AAMAS '13. pp.
715 -- 722. International Foundation for Au-
tonomous Agents and Multiagent Systems,
Richland, SC (2013), http://dl.acm.org/
citation.cfm?id=2484920.2485033
[4] Baarslag, T., Hindriks, K.V.,
for
A tit
Jonker,
C.:
tat negotiation strat-
egy for real-time bilateral negotiation. In:
Complex Automated Negotiations: Theo-
ries, Models, and Software Competitions,
vol. 435, pp. 229 -- 233. Springer Berlin Hei-
delberg (2013). https://doi.org/10.1007/978-
3-642-30737-9 18
[5] Browne, C.C., Powley, E., Whitehouse,
D., Lucas, S.M., Cowling, P.I., Rohlf-
shagen, P., Taverner,
S., Perez, D.,
Samothrakis, S., Colton, S.: A survey
of Monte Carlo tree search methods. IEEE
Transactions
Intelli-
gence and AI in games 4(1), 1 -- 43 (2012).
https://doi.org/10.1109/TCIAIG.2012.2186810
on Computational
[6] Coulom, R.: Efficient selectivity and backup
operators in monte-carlo tree search.
In:
International conference on computers and
games. pp. 72 -- 83. Springer (2006)
[7] Couetoux, A.: Monte Carlo Tree Search for
Continuous and Stochastic Sequential Deci-
sion Making Problems. Ph.D. thesis, Univer-
sit´e Paris XI (2013)
[8] Dziuba, D.T.:
Crowdfunding platforms
in invoice trading as alternative financial
markets. Roczniki Kolegium Analiz Eko-
nomicznych/Szko(cid:32)la G(cid:32)l´owna Handlowa 49,
455 -- 464 (2018)
[9] Fang, F., Xin, Y., Xia, Y., Haitao,
An opponent's negotiation behav-
X.:
to facilitate buyer-seller ne-
ior model
in supply chain management.
gotiations
In:
2008 International Symposium on
Electronic Commerce and Security (2008).
https://doi.org/10.1109/ISECS.2008.93
decision
[10] Faratin, P., Jennings, N.R., Sierra, C.:
au-
and Au-
agents. Robotics
Systems
159 -- 182
https://doi.org/10.1016/S0921-
Negotiation
tonomous
tonomous
(1998).
8890(98)00029-3
functions
24(3-4),
for
[11] Finnsson, H.: Generalized Monte Carlo
for general game
tree search extensions
playing. In: Proceedings of the Twenty-
Sixth AAAI Conference on Artificial Intelli-
gence. pp. 1550 -- 1556. AAAI'12, AAAI Press
(2012), http://dl.acm.org/citation.cfm?
id=2900929.2900948
[12] Fukuta, N.,
Ito, T., Zhang, M., Fujita,
K., Robu, V. (eds.): Recent Advances in
Agent-based Complex Automated Negotia-
tion, Studies in Computational Intelligence,
vol. 638. Springer International Publishing
(2016)
[13] Guttman, R.H., Moukas, A.G., Maes,
P.: Agent-mediated electronic commerce:
survey. The Knowledge Engineering
a
Review 13(02),
1998).
https://doi.org/10.1007/3-540-58266-5 1,
http://journals.cambridge.org/action/
displayAbstract?fromPage=online&aid=
71015&fileId=S0269888998002082
147 -- 159
(Jul
[14] Hindriks, K., Tykhonov, D.: Opponent mod-
elling in automated multi-issue negotiation
using bayesian learning. In: Proceedings of
the 7th International Joint Conference on Au-
tonomous Agents and Multiagent Systems.
vol. 1, pp. 331 -- 338 (2008), http://dl.acm.
org/citation.cfm?id=1402383.1402433
[15] de Jonge, D., Sierra, C.: Recent Advances in
Agent-based Complex Automated Negotia-
tion, Studies in Computational Intelligence,
vol. 638, chap. GANGSTER: An Automated
Negotiator Applying Genetic Algorithms, pp.
225 -- 234. Springer International Publishing
(2016).
30307-9 14,
978-3-319-30307-9_14
https://doi.org/10.1007/978-3-319-
https://doi.org/10.1007/
for general game playing.
[16] de Jonge, D., Zhang, D.: Automated ne-
gotiations
In:
AAMAS '17. pp. 371 -- 379. Richland, SC
(2017), http://dl.acm.org/citation.cfm?
id=3091125.3091183
[17] Kocsis, L., Szepesv´ari, C.: Bandit based
Monte Carlo planning. In: Furnkranz, J.,
Scheffer, T., Spiliopoulou, M. (eds.) Ma-
chine Learning: ECML 2006: 17th Euro-
pean Conference on Machine Learning Berlin,
Germany, September 18-22, 2006 Proceed-
ings. pp. 282 -- 293. Springer Berlin Heidelberg
(2006). https://doi.org/10.1007/11871842 29
[18] Lin, R., Kraus, S., Baarslag, T., Tykhonov,
D., Hindriks, K., Jonker, C.M.: Genius:
an integrated environment for supporting
the design of generic automated negotia-
tors. Computational Intelligence 30(1), 48 --
70 (2014). https://doi.org/10.1111/j.1467-
8640.2012.00463.x
[19] Nash Jr, J.F.: The bargaining problem.
Econometrica: Journal of the Econometric
Society pp. 155 -- 162 (1950)
[20] Osborne, M.J., Rubinstein, A.: A course in
game theory. MIT press, 12 edn. (1994)
[21] Rasmussen, C.E., Williams, C.K.I.: Gaussian
processes for machine learning. MIT Press
(2006)
[22] Rubinstein, A.: Perfect equilibrium in a bar-
gaining model. Econometrica: Journal of the
Econometric Society pp. 97 -- 109 (1982)
[23] Silver, D., Huang, A., Maddison, C.J., Guez,
A., Sifre, L., Van Den Driessche, G., Schrit-
twieser, J., al.: Mastering the game of go with
deep neural networks and tree search. Nature
529(7587), 484 -- 489 (2016)
to
Using
[24] Williams, C.R., Robu, V., Gerding, E.H.,
gaussian pro-
Jennings, N.R.:
cesses
concession in com-
plex negotiations against unknown oppo-
nents. In:
IJCAI'11. pp. 432 -- 438 (2011).
https://doi.org/10.5591/978-1-57735-516-
8/IJCAI11-080
optimise
11
|
1407.8269 | 4 | 1407 | 2016-09-12T01:02:43 | Justified Representation in Approval-Based Committee Voting | [
"cs.MA",
"cs.GT"
] | We consider approval-based committee voting, i.e. the setting where each voter approves a subset of candidates, and these votes are then used to select a fixed-size set of winners (committee). We propose a natural axiom for this setting, which we call justified representation (JR). This axiom requires that if a large enough group of voters exhibits agreement by supporting the same candidate, then at least one voter in this group has an approved candidate in the winning committee. We show that for every list of ballots it is possible to select a committee that provides JR. However, it turns out that several prominent approval-based voting rules may fail to output such a committee. In particular, while Proportional Approval Voting (PAV) always outputs a committee that provides JR, Reweighted Approval Voting (RAV), a tractable approximation to PAV, does not have this property. We then introduce a stronger version of the JR axiom, which we call extended justified representation (EJR), and show that PAV satisfies EJR, while other rules we consider do not; indeed, EJR can be used to characterize PAV within the class of weighted PAV rules. We also consider several other questions related to JR and EJR, including the relationship between JR/EJR and core stability, and the complexity of the associated algorithmic problems. | cs.MA | cs | Justified Representation in
Approval-Based Committee Voting
Haris Aziz · Markus Brill · Vincent Conitzer ·
Edith Elkind · Rupert Freeman · Toby Walsh
6
1
0
2
p
e
S
2
1
]
A
M
.
s
c
[
4
v
9
6
2
8
.
7
0
4
1
:
v
i
X
r
a
Received: date / Accepted: date
Abstract We consider approval-based committee voting, i.e. the setting where each
voter approves a subset of candidates, and these votes are then used to select a fixed-
size set of winners (committee). We propose a natural axiom for this setting, which
we call justified representation (JR). This axiom requires that if a large enough group
of voters exhibits agreement by supporting the same candidate, then at least one voter
in this group has an approved candidate in the winning committee. We show that for
every list of ballots it is possible to select a committee that provides JR. However,
it turns out that several prominent approval-based voting rules may fail to output
such a committee. In particular, while Proportional Approval Voting (PAV) always
outputs a committee that provides JR, Reweighted Approval Voting (RAV), a tractable
approximation to PAV, does not have this property. We then introduce a stronger
version of the JR axiom, which we call extended justified representation (EJR), and
show that PAV satisfies EJR, while other rules we consider do not; indeed, EJR can
be used to characterize PAV within the class of weighted PAV rules. We also consider
several other questions related to JR and EJR, including the relationship between
JR/EJR and core stability, and the complexity of the associated algorithmic problems.
Keywords Approval voting · committee selection · representation.
H. Aziz · T. Walsh
Data61, CSIRO and UNSW Australia, Sydney 2052 , Australia
Tel.: +61-2-8306 0490
Fax: +61-2-8306 0405
E-mail: {haris.aziz,toby.walsh}@data61.csiro.au
M. Brill · E. Elkind
University of Oxford
Oxford OX1 3QD, UK
E-mail: {mbrill,elkind}@cs.ox.ac.uk
V. Conitzer · R. Freeman
Duke University
Durham, NC 27708, USA
E-mail: {conitzer, rupert}@cs.duke.edu
2
Haris Aziz et al.
JEL Classification: C70 · D61 · D71
1 Introduction
The aggregation of preferences is a central problem in the field of social choice.
While the most-studied scenario is that of selecting a single candidate out of many, it
is often the case that one needs to select a fixed-size set of winners (committee): this
includes domains such as parliamentary elections, the hiring of faculty members, or
(automated) agents deciding on a set of plans (LeGrand et al. 2007; Davis et al. 2014;
Elkind et al. 2014, 2015; Skowron et al. 2015a). The study of algorithmic complexity
of voting rules that output committees is an active research direction (Procaccia et al.
2008; Meir et al. 2008; Caragiannis et al. 2010; Lu and Boutilier 2011; Cornaz et al.
2012; Betzler et al. 2013; Skowron et al. 2015b,c).
In this paper we consider approval-based rules, where each voter lists the subset
of candidates that she approves of. There is a growing literature on voting rules that
are based on approval ballots: the Handbook on Approval Voting (Laslier and Sanver
2010) provides a very useful survey of pre-2010 research on this topic, and after this
seminal book was published, various aspects of approval voting continued to attract
a considerable amount of attention (see, e.g., the papers of Caragiannis et al. 2010;
Endriss 2013; Duddy 2014). One of the advantages of approval ballots is their sim-
plicity: such ballots reduce the cognitive burden on voters (rather than providing a full
ranking of the candidates, a voter only needs to decide which candidates to approve)
and are also easier to communicate to the election authority. The most straightforward
way to aggregate approvals is to have every approval for a candidate contribute one
point to that candidate's score and select the candidates with the highest score. This
rule is called Approval Voting (AV). AV has many desirable properties in the single-
winner case (Brams et al. 2006; Endriss 2013), including its "simplicity, propensity
to elect Condorcet winners (when they exist), its robustness to manipulation and its
monotonicity" (Brams 2010, p. viii). However, for the case of multiple winners, the
merits of AV are "less clear" (Brams 2010, p. viii). For example, AV may fail pro-
portional representation: if the goal is to select k winners, k > 1, 51% of the voters
approve the same k candidates, and the remaining voters approve a disjoint set of k
candidates, then the voters in minority do not get any of their approved candidates
selected.
As a consequence, over the years, several multi-winner rules based on approval
ballots have been proposed (see, e.g., the survey by Kilgour 2010); we will now
briefly describe the rules that will be considered in this paper (see Section 2 for for-
mal definitions). Under Proportional Approval Voting (PAV), each voter's contribu-
tion to the committee's total score depends on how many candidates from the voter's
approval set have been elected. In the canonical variant of this rule the marginal util-
ity of the ℓ-th approved candidate is 1
ℓ , i.e. this rule is associated with the weight
vector (1, 1
3 , . . . ); other weight vectors can be used as well, resulting in the family
of weighted PAV rules. A sequential variant of PAV is known as Reweighted Ap-
proval Voting (RAV); again, by varying the weight vector, we obtain the family of
weighted RAV rules. Another way to modulate the approvals is through computing a
2 , 1
Justified Representation in Approval-Based Committee Voting
3
satisfaction score for each voter based on the ratio of the number of their approved
candidates appearing in the committee and their total number of approved candi-
dates; this idea leads to Satisfaction Approval Voting (SAV). One could also use a
distance-based approach: Minimax Approval Voting (MAV) selects a set of k candi-
dates that minimizes the maximum Hamming distance from the submitted ballots.
Finally, one could adapt classic rules that provide fully proportional representation,
such as the Chamberlin–Courant rule (Chamberlin and Courant 1983) or the Monroe
rule (Monroe 1995), to work with approval ballots, by using each voter's ballot as
a scoring vector. All the rules informally described above have a more egalitarian
objective than AV. For example, Steven Brams, a proponent of AV in single-winner
elections, has argued that SAV is more suitable for equitable representation in multi-
winner elections (Brams and Kilgour 2014).
The relative merits of approval-based multi-winner rules and the complexity
of winner determination under these rules have been examined in great detail in
both economics and computer science in recent years (Brams and Fishburn 2007;
LeGrand et al. 2007; Meir et al. 2008; Caragiannis et al. 2010; Aziz et al. 2015;
Byrka and Sornat 2014; Misra et al. 2015). On the other hand, there has been lim-
ited axiomatic analysis of these rules from the perspective of representation (see,
however, Section 7).
In this paper, we introduce the notion of justified representation (JR) in approval-
based voting. Briefly, a committee is said to provide justified representation for a
given set of ballots if every large enough group of voters with shared preferences
is allocated at least one representative. A rule is said to satisfy justified representa-
tion if it always outputs a committee that provides justified representation. This con-
cept is related to the Droop proportionality criterion (Droop 1881) and Dummett's
solid coalition property (Dummett 1984; Tideman and Richardson 2000; Elkind et al.
2014), but is specific to approval-based elections.
We show that every set of ballots admits a committee that provides justified rep-
resentation; moreover, such a committee can be computed efficiently, and checking
whether a given committee provides JR can be done in polynomial time as well. This
shows that justified representation is a reasonable requirement. However, it turns out
that many popular multi-winner approval-based rules fail JR; in particular, this is the
case for AV, SAV, MAV and the canonical variant of RAV. On the positive side, JR is
satisfied by some of the weighted PAV rules, including the canonical PAV rule, as well
as by the weighted RAV rule associated with the weight vector (1, 0, . . . ) and by the
Monroe rule. Also, MAV satisfies JR for a restricted domain of voters' preferences.
We then consider a strengthening of the JR axiom, which we call extended justified
representation (EJR). This axiom captures the intuition that a very large group of
voters with similar preferences may deserve not just one, but several representatives.
EJR turns out to be a more demanding property than JR: of all voting rules considered
in this paper, only the canonical PAV rule satisfies EJR. Thus, in particular, EJR char-
acterizes the canonical PAV rule within the class of weighted PAV rules. However,
we show that it is computationally hard to check whether a given committee provides
EJR.
We also consider other strengthenings of JR, which we call semi-strong justified
representation and strong justified representation; however, it turns out that for some
4
Haris Aziz et al.
inputs the requirements imposed by these axioms are impossible to satisfy. Finally,
we explore the relationship between JR/EJR and core stability in a non-transferable
utility game that can be associated with a multiwinner approval voting scenario. We
show that, even though EJR may appear to be similar to core stability, it is, in fact,
a strictly weaker condition. Indeed, the core stability condition appears to be too de-
manding, as none of the voting rules considered in our work is guaranteed to produce
a core stable outcome, even when the core is known to be non-empty. We conclude
the paper by showing how JR can be used to formulate other attractive approval-
based multi-winner rules, discussing related work, and identifying several directions
for future work.
2 Preliminaries
We consider a social choice setting with a set N = {1, . . . , n} of voters and a set C
of candidates. Each voter i ∈ N submits an approval ballot Ai ⊆ C, which represents
the subset of candidates that she approves of. We refer to the list A = (A1, . . . , An) of
approval ballots as the ballot profile. We will consider approval-based multi-winner
voting rules that take as input a tuple (N, C, A, k), where k is a positive integer that
satisfies k ≤ C, and return a subset W ⊆ C of size k, which we call the winning set,
or committee (Kilgour and Marshall 2012). We omit N and C from the notation when
they are clear from the context. Several approval-based multi-winner rules are defined
below. Whenever the description of the rule does not uniquely specify a winning set,
we assume that ties are broken according to some deterministic procedure; however,
most of our results do not depend on the tie-breaking rule.
2.1 Approval-Based Multi-Winner Rules
Approval Voting (AV) Under AV, the winners are the k candidates that receive the
largest number of approvals. Formally, the approval score of a candidate c ∈ C is
defined as {i c ∈ Ai}, and AV outputs a set W of size k that maximizes Pc∈W {i
tion scores. Formally, SAV outputs a set W ⊆ C of size k that maximizesPi∈N
c ∈ Ai}. AV has been adopted by several academic and professional societies such
as the Institute of Electrical and Electronics Engineers (IEEE) and the International
Joint Conference on Artificial Intelligence (IJCAI).
Satisfaction Approval Voting (SAV) A voter's satisfaction score is the fraction of
her approved candidates that are elected. SAV maximizes the sum of voters' satisfac-
W∩Ai
.
This rule was proposed with the aim of "representing more diverse interests" than AV
(Brams and Kilgour 2014).
Proportional Approval Voting (PAV) Under PAV, a voter is assumed to derive
a utility of 1 + 1
j from a committee that contains exactly j of her
approved candidates, and the goal is to maximize the sum of the voters' utilities.
Formally, the PAV-score of a set W ⊆ C is defined as Pi∈N r(W ∩ Ai), where r(p) =
Pp
1
j , and PAV outputs a set W ⊆ C of size k with the highest PAV-score. Though
sometimes attributed to Forest Simmons, PAV was already proposed by the Danish
Ai
2 + 1
3 + · · · + 1
j=1
Justified Representation in Approval-Based Committee Voting
5
polymath Thorvald N. Thiele in the 19th century (Thiele 1895).1 PAV captures the
idea of diminishing returns: an individual voter's preferences should count less the
more she is satisfied.
2 , 1
We can generalize the definition of PAV by using an arbitrary score vector in place
of (1, 1
3 , . . . ). Specifically, for every vector2 w = (w1, w2, . . . ), where w1, w2, . . . are
non-negative reals, we define a voting rule w-PAV that operates as follows. Given a
ballot profile (A1, . . . , An) and a target number of winners k, w-PAV returns a set W
of size k with the highest w-PAV score, defined by Pi∈N rw(W ∩ Ai), where rw(p) =
Pp
j=1 w j. Usually, it is required that w1 = 1 and w1 ≥ w2 ≥ . . . . The latter constraint is
appropriate in the context of representative democracy: it is motivated by the intuition
that once an agent already has one or more representatives in the committee, that
agent should have less priority for further representation. It what follows, we will
always impose the constraint w1 = 1 (as we can always rescale the weight vector, this
is equivalent to requiring that w1 > 0; while the case w1 = 0 may be of interest in
some applications, we omit it in order to keep the length of the paper manageable3)
and explicitly indicate which of our results require that w1 ≥ w2 ≥ . . . ; in particular,
for our characterization of PAV in Theorem 11 this is not the case.
Reweighted Approval Voting (RAV) RAV converts PAV into a multi-round rule, by
selecting a candidate in each round and then reweighing the approvals for the subse-
quent rounds. Specifically, RAV starts by setting W = ∅. Then in round j, j = 1, . . . , k,
1+W∩Ai , selects a
candidate with the highest approval weight, and adds him to W. After k rounds, it
outputs the set W. RAV has also been referred to as "sequential proportional AV"
(Brams and Kilgour 2014), and was used briefly in Sweden during the early 1900s.
it computes the approval weight of each candidate c as Pi:c∈Ai
1
Thiele (1895) proposed RAV as a tractable approximation to PAV (see Section 2.2
for a discussion of the computational complexity of these rules and the relation-
ship between them). We note that there are several other examples of voting rules
that were conceived as approximate versions of other rules, yet became viewed as
legitimate voting rules in and of themselves; two representative examples are the
Simplified Dodgson rule of Tideman (2006), which was designed as an approxi-
mate version of the Dodgson rule (see the discussion by Caragiannis et al. 2014), and
the Greedy Monroe rule of Skowron et al. (2015b), which approximates the Monroe
rule (Monroe 1995).
Just as for PAV, we can extend the definition of RAV to score vectors other than
(1, 1
2 , 1
3 , . . . ): every vector w = (w1, w2, . . . ) defines a sequential voting rule w-RAV,
which proceeds as RAV, except that it computes the approval weight of a candidate c
in round j asPi:c∈Ai wW∩Ai +1, where W is the winning set after the first j − 1 rounds.
Again, we impose the constraint w1 = 1 (note that if w1 = 0, then w-RAV can pick an
arbitrary candidate at the first step, which is obviously undesirable).
1 We are grateful to Xavier Mora and Svante Janson for pointing this out to us.
2
It is convenient to think of w as an infinite vector; note that for an election with m candidates only the
first m entries of w matter. To analyze the complexity of w-PAV rules, one would have to place additional
requirements on w; however, we do not consider algorithmic properties of such rules in this paper.
3 Generalizations of PAV with w1 = 0 have been considered by Fishburn and Pekec (2004) and
Skowron et al. (2015d). We note that such rules do not satisfy justified representation (as defined in Sec-
tion 3).
6
Haris Aziz et al.
A particularly interesting rule in this class is (1, 0, . . . )-RAV: this rule, which
we will refer to as Greedy Approval Voting (GAV), can be seen as a variant of the
SweetSpotGreedy (SSG) algorithm of Lu and Boutilier (2011), and admits a very
simple description: we pick candidates one by one, trying to 'cover' as many cur-
rently 'uncovered' voters as possible. In more detail, a winning committee under this
rule can be computed by the following algorithm. We start by setting C′ = C, A′ = A,
and W = ∅. As long as W < k and A′ is non-empty, we pick a candidate c ∈ C′ that
has the highest approval score with respect to A′, and set W := W ∪ {c}, C′ := C′ \ {c}.
Also, we remove from A′ all ballots Ai such that c ∈ Ai. If at some point we have
W < k and A′ is empty, we add an arbitrary set of k − W candidates from C′ to W
and return W; if this does not happen, we terminate after having picked k candidates.
We will also consider a variant of GAV, where, at each step, after selecting a
candidate c, instead of removing all voters in Ac = {i c ∈ Ai} from A′, we remove
a subset of Ac of size minn⌈ n
k ⌉, Aco. This rule can be seen as an adaptation of the
classic STV rule to approval ballots, and we will refer to it as GAVT (where T stands
for 'threshold').4
Minimax Approval Voting (MAV) MAV returns a committee W that minimizes
the maximum Hamming distance between W and the voters' ballots; this rule was
proposed by Brams et al. (2007). Formally, let d(Q, T ) = Q \ T + T \ Q and define
the MAV-score of a set W ⊆ C as max (d(W, A1), . . . , d(W, An)). MAV outputs a size-k
set with the lowest MAV-score.
Chamberlin–Courant and Monroe Approval Voting (CCAV and MonAV) The
Chamberlin–Courant rule (Chamberlin and Courant 1983) is usually defined for the
setting where each voter provides a full ranking of the candidates. Each voter i ∈ N
is associated with a scoring vector ui = (ui
m) whose entries are non-negative
reals; we think of ui
j as voter i's satisfaction from being represented by candidate c j. A
voter's satisfaction from a committee W is defined as maxc j∈W ui
j, and the rule returns
a committee of size k that maximizes the sum of voters' satisfactions. For the case of
approval ballots, it is natural to define the scoring vectors by setting ui
j = 1 if c j ∈ Ai
and ui
j = 0 otherwise; that is, a voter is satisfied by a committee if this committee
contains one of her approved candidates. Thus, the resulting rule is equivalent to
(1, 0, . . . )-PAV (and therefore we will not discuss it separately).
1, . . . , ui
The Monroe rule (Monroe 1995) is a modification of the Chamberlin–Courant
rule where each committee member represents roughly the same number of vot-
ers. Just as under the Chamberlin–Courant rule, we have a scoring vector ui =
(ui
m) for each voter i ∈ N. Given a committee W of size k, we say that a
k ⌋, ⌈ n
k ⌉o for each c ∈ W.
π(i), and the Monroe score
of a committee W is the maximum Monroe score of a valid mapping from N to
W. The Monroe rule returns a size-k committee with the maximum Monroe score.
For approval ballots, we define the scoring vectors in the same manner as for the
1, . . . , ui
mapping π : N → W is valid if it satisfies π−1(c) ∈ n⌊ n
The Monroe score of a valid mapping π is given by Pi∈N ui
4 For readability, we use the Hare quota ⌈ n
k ⌉; however, all our proofs go through if we use the Droop
k+1 ⌉+1 instead. For a discussion of differences between these two quotas, see the article of Tideman
quota ⌈ n
(1995).
Justified Representation in Approval-Based Committee Voting
7
Chamberlin–Courant Approval Voting rule; we call the resulting rule the Monroe
Approval Voting rule (MonAV).
We note that for k = 1, AV, PAV, RAV, GAV, GAVT and MonAV produce the same
output if there is a unique candidate with the highest approval score. However, such
a candidate need not be a winner under SAV or MAV.
2.2 Computational Complexity
The rules listed above differ from an algorithmic perspective. For some of these
rules, namely, AV, SAV, RAV, GAV and GAVT , a winning committee can be com-
puted in polynomial time; this is also true for w-RAV as long the entries of the
weight vector are rational numbers that can be efficiently computed given the num-
ber of candidates. In contrast, PAV, MAV, and MonAV are computationally hard
(Aziz et al. 2015; Skowron et al. 2015a; LeGrand et al. 2007; Procaccia et al. 2008);
for w-PAV, the hardness result holds for most weight vectors, including (1, 0, . . . ),
(i.e., it holds for CCAV). However, both PAV and MAV admit efficient approxima-
tion algorithms (i.e., algorithms that output committees which are approximately
optimal with respect to the optimization criteria of these rules) and have been an-
alyzed from the perspective of parameterized complexity. Specifically, w-PAV admits
an efficient (cid:16)1 − 1
e(cid:17)-approximation algorithm as long as the weight vector w is ef-
ficiently computable and non-increasing; in fact, such an algorithm is provided by
w-RAV (Skowron et al. 2015a). For MAV, LeGrand et al. (2007) propose a simple 3-
approximation algorithm; Caragiannis et al. (2010) improve the approximation ratio
to 2 and Byrka and Sornat (2014) develop a polynomial-time approximation scheme.
Misra et al. (2015) show that MAV is fixed-parameter tractable for a number of natu-
ral parameters; Elkind and Lackner (2015) obtain fixed parameter tractability results
for PAV when voters' preferences are, in some sense, single-dimensional. There is
also a number of tractability results for CCAV, and, to a lesser extent, for MonAV; we
refer the reader to the work of Skowron et al. (2015a) and references therein.
3 Justified Representation
We will now define one of the main concepts of this paper.
Definition 1 (Justified representation (JR)) Given a ballot profile A = (A1, . . . , An)
over a candidate set C and a target committee size k, we say that a set of candidates
W of size W = k provides justified representation for (A, k) if there does not exist
a set of voters N∗ ⊆ N with N∗ ≥ n
all i ∈ N∗. We say that an approval-based voting rule satisfies justified representation
(JR) if for every profile A = (A1, . . . , An) and every target committee size k it outputs
a winning set that provides justified representation for (A, k).
k such that Ti∈N∗ Ai , ∅ and Ai ∩ W = ∅ for
The logic behind this definition is that if k candidates are to be selected, then, intu-
itively, each group of n
k voters
that have at least one candidate in common should not be completely unrepresented.
We refer the reader to Section 6 for a discussion of alternative definitions.
k voters "deserves" a representative. Therefore, a set of n
8
Haris Aziz et al.
3.1 Existence and Computational Properties
We start our analysis of justified representation by observing that, for every ballot
profile A and every value of k, there is a committee that provides justified represen-
tation for (A, k), and, moreover, such a committee can be computed efficiently given
the voters' ballots. In fact, both GAV and GAVT output a committee that provides JR.
Theorem 1 GAV and GAVT satisfy JR.
Proof We present a proof that applies to both GAV and GAVT . Suppose for the sake
of contradiction that for some ballot profile A = (A1, . . . , An) and some k > 0, GAV
(respectively, GAVT ) outputs a committee that does not provide justified representa-
tion for (A, k). Then there exists a set N∗ ⊆ N with N∗ ≥ n
k such that Ti∈N∗ Ai , ∅
and, when GAV (respectively, GAVT ) terminates, every ballot Ai such that i ∈ N∗ is
still in A′. Consider some candidate c ∈ Ti∈N∗ Ai. At every point in the execution of
our algorithm, c's approval score is at least N∗ ≥ n
k . As c was not elected, at every
stage the algorithm selected a candidate whose approval score was at least as high as
that of c. Thus, at the end of each stage the algorithm removed from A′ at least ⌈ n
k ⌉
ballots containing the candidate added to W at that stage, so altogether the algorithm
has removed at least k · n
k ballots from A′. This contradicts the assumption that A′
contains at least n
⊓⊔
k ballots when the algorithm terminates.
Theorem 1 shows that it is easy to find a committee that provides justified repre-
sentation for a given ballot profile. It is also not too hard to check whether a given
committee W provides JR. Indeed, while it may seem that we need to consider every
subset of voters of size n
k , in fact it is sufficient to consider the candidates one by one,
and, for each candidate c, compute s(c) = {i ∈ N c ∈ Ai, Ai ∩ W = ∅}; the set W
fails to provide justified representation for (A, k) if and only if there exists a candidate
c with s(c) ≥ n
k . We obtain the following theorem.
Theorem 2 There exists a polynomial-time algorithm that, given a ballot profile A
over a candidate set C, and a committee W, W = k, decides whether W provides
justified representation for (A, k).
3.2 Justified Representation and Unanimity
A desirable property of single-winner approval-based voting rules is unanimity: a
voting rule is unanimous if, given a ballot profile (A1, . . . , An) with ∩i∈N Ai , ∅, it
outputs a candidate in ∩i∈N Ai. This property is somewhat similar in spirit to JR, so
the reader may expect that for k = 1 it is equivalent to JR. However, it turns out that
the JR axiom is strictly weaker than unanimity for k = 1: while unanimity implies
JR, the converse is not true, as illustrated by the following example.
Example 1 Let N = {1, . . . , n}, C = {a, b1, . . . , bn}, Ai = {a, bi} for i ∈ N. Consider
a voting rule that for k = 1 outputs b1 on this profile and coincides with GAV in
all other cases. Clearly, this rule is not unanimous; however, it satisfies JR, as it is
impossible to find a group of n
k = n unrepresented voters for (A1, . . . , An).
Justified Representation in Approval-Based Committee Voting
9
It is not immediately clear how to define unanimity for multi-winner voting rules;
however, any reasonable definition would be equivalent to the standard definition of
unanimity when k = 1, and therefore would be different from justified representation.
We remark that a rule can be unanimous for k = 1 and provide JR for all values
of k: this is the case, for instance, for GAV.
4 Justified Representation under Approval-Based Rules
We have argued that justified representation is a reasonable condition: there always
exists a committee that provides it, and, moreover, such a committee can be computed
efficiently. It is therefore natural to ask whether prominent voting rules satisfy JR. In
this section, we will answer this question for AV, SAV, MAV, PAV, RAV, and MonAV.
We will also identify conditions on w that are sufficient/necessary for w-PAV and
w-RAV to satisfy JR.
In what follows, for each rule we will try to identify the range of values of k for
which this rule satisfies JR. Trivially, all rules that we consider satisfy JR for k = 1.
It turns out that AV fails JR for k > 2, and for k = 2 the answer depends on the
tie-breaking rule.
Theorem 3 For k = 2, AV satisfies JR if ties are broken in favor of sets that provide
JR. For k ≥ 3, AV fails JR.
Proof Suppose first that k = 2. Fix a ballot profile A. If every candidate is approved
by fewer than n
2 voters in A, JR is trivially satisfied. If some candidate is approved
by more than n
2 voters in A, then AV selects some such candidate, in which case no
group of ⌈ n
2 ⌉ voters is unrepresented, so JR is satisfied in this case as well. It remains
to consider the case where n = 2n′, some candidates are approved by n′ voters, and
no candidate is approved by more than n′ voters. Then AV necessarily picks at least
one candidate approved by n′ voters; denote this candidate by c. In this situation
JR can only be violated if the n′ voters who do not approve c all approve the same
candidate (say, c′), and this candidate is not elected. But the approval score of c′ is
n′, and, by our assumption, the approval score of every candidate is at most n′, so
this is a contradiction with our tie-breaking rule. This argument also illustrates why
the assumption on the tie-breaking rule is necessary: it can be the case that n′ voters
approve c and c′′, and the remaining n′ voters approve c′, in which case the approval
score of {c, c′′} is the same as that of {c, c′}.
For k ≥ 3, we let C = {c0, c1, . . . , ck}, n = k, and consider the profile where the
first voter approves c0, whereas each of the remaining voters approves all of c1, . . . , ck.
JR requires c0 to be selected, but AV selects {c1, . . . , ck}.
⊓⊔
On the other hand, SAV and MAV fail JR even for k = 2.
Theorem 4 SAV and MAV do not satisfy JR for k ≥ 2.
Proof We first consider SAV. Fix k ≥ 2, let X = {x1, . . . , xk, xk+1}, Y = {y1, . . . , yk},
C = X ∪ Y, and consider the profile (A1, . . . , Ak), where A1 = X, A2 = {y1, y2},
Ai = {yi} for i = 3, . . . , k. JR requires each voter to be represented, but SAV will
10
Haris Aziz et al.
choose Y: the SAV-score of Y is k − 1, whereas the SAV-score of every committee W
with W ∩ X , ∅ is at most k − 2 + 1
k+1 < k − 1. Therefore, the first voter will
remain unrepresented.
2 + 1
For MAV, we use the following construction. Fix k ≥ 2, let X = {x1, . . . , xk},
Y = {y1, . . . , yk}, C = X ∪ Y ∪ {z}, and consider the profile (A1, . . . , A2k), where
Ai = {xi, yi} for i = 1, . . . , k, Ai = {z} for i = k + 1, . . . , 2k. Every committee of size
k that provides JR for this profile contains z. However, MAV fails to select z. Indeed,
the MAV-score of X is k + 1: we have d(X, Ai) = k for i ≤ k and d(X, Ai) = k + 1 for
i > k. Now, consider some committee W with W = k, z ∈ W. We have Ai ∩W = ∅ for
some i ≤ k, so d(W, Ai) = k + 2. Thus, MAV prefers X to any committee that includes
z.
⊓⊔
The constructions used in the proof of Theorem 4 show that MAV and SAV may be-
have very differently: SAV appears to favor voters who approve very few candidates,
whereas MAV appears to favor voters who approve many candidates.
Interestingly, we can show that MAV satisfies JR if we assume that each voter
approves exactly k candidates and ties are broken in favor of sets that provide JR.
Theorem 5 If the target committee size is k, Ai = k for all i ∈ N, and ties are broken
in favor of sets that provide JR, then MAV satisfies JR.
Proof Consider a profile A = (A1, . . . , An) with Ai = k for all i ∈ N.
Observe that if there exists a set of candidates W with W = k such that W ∩Ai , ∅
for all i ∈ N, then MAV will necessarily select some such set. Indeed, for any such set
W we have d(W, Ai) ≤ 2k − 1 for each i ∈ N, whereas if W ′ ∩ Ai = ∅ for some set W ′
with W ′ = k and some i ∈ N, then d(W ′, Ai) = 2k. Further, by definition, every set
W such that W = k and W ∩ Ai , ∅ for all i ∈ N provides justified representation for
(A, k).
On the other hand, if there is no k-element set of candidates that intersects each
Ai, i ∈ N, then the MAV-score of every set of size k is 2k, and therefore MAV can pick
an arbitrary size-k subset. Since we assumed that the tie-breaking rule favors sets that
provide JR, our claim follows.
⊓⊔
While Theorem 5 provides an example of a setting where MAV satisfies JR, this
result is not entirely satisfactory: first, we had to place a strong restriction on voters'
preferences, and, second, we used a tie-breaking rule that was tailored to JR.
We will now show that PAV satisfies JR, for all ballot profiles and irrespective of
the tie-breaking rule.
Theorem 6 PAV satisfies JR.
Proof Fix a ballot profile A = (A1, . . . , An) and a k > 0 and let s = ⌈ n
k ⌉. Let W be
the output of PAV on (A, k). Suppose for the sake of contradiction that there exists a
set N∗ ⊂ N, N∗ ≥ s, such that Ti∈N∗ Ai , ∅, but W ∩ Si∈N∗ Ai = ∅. Let c be some
candidate approved by all voters in N∗.
For each candidate w ∈ W, define its marginal contribution as the difference
between the PAV-score of W and that of W \{w}. Let m(W) denote the sum of marginal
contributions of all candidates in W. Observe that if c were to be added to the winning
Justified Representation in Approval-Based Committee Voting
11
set, this would increase the PAV-score by at least s. Therefore, it suffices to argue that
the marginal contribution of some candidate in W is less than s: this would mean that
swapping this candidate with c increases the PAV-score, a contradiction. To this end,
we will prove that m(W) ≤ s(k − 1); as W = k, our claim would then follow by the
pigeonhole principle.
Consider the set N \ N∗; we have n ≤ sk, so N \ N∗ ≤ n − s ≤ s(k − 1). Pick
a voter i ∈ N \ N∗, and let j = Ai ∩ W. If j > 0, this voter contributes exactly 1
j to
the marginal contribution of each candidate in Ai ∩ W, and hence her contribution to
m(W) is exactly 1. If j = 0, this voter does not contribute to m(W) at all. Therefore,
we have m(W) ≤ N \ N∗ ≤ s(k − 1), which is what we wanted to prove.
⊓⊔
The reader may observe that the proof of Theorem 6 applies to all voting rules of
the form w-PAV where the weight vector satisfies w1 = 1, w j ≤ 1
j for all j ≥ 1. In
Section 5 we will see that this condition on w is also necessary for w-PAV to satisfy
JR.
Next, we consider RAV. As this voting rule can be viewed as a tractable approx-
imation of PAV (recall that PAV is NP-hard to compute), one could expect that RAV
satisfies JR as well. However, this turns out not to be the case, at least if k is suffi-
ciently large.
Theorem 7 RAV satisfies JR for k = 2, but fails it for k ≥ 10.
Proof For k = 2, we can use essentially the same argument as for AV; however, we
do not need to assume anything about the tie-breaking rule. This is because if there
are three candidates, c, c′, and c′′, such that c and c′′ are approved by the same n
2
voters, whereas c′ is approved by the remaining n
2 voters, and RAV selects c in the
first round, then in the second round RAV favors c′ over c′′.
Now, suppose that k = 10. Consider a profile over a candidate set C = {c1, . . . , c11}
with 1199 voters who submit the following ballots:
81×{c1, c2},
81×{c4, c5},
49×{c7, c8},
96×{c8},
81×{c1, c3},
81×{c4, c6},
49×{c7, c9},
96×{c9},
80×{c2},
80×{c5},
49×{c7, c10},
96×{c10},
80×{c3},
80×{c6},
120×{c11}.
Candidates c1 and c4 are each approved by 162 voters, the most of any candidate,
and these blocks of 162 voters do not overlap, so RAV selects c1 and c4 first. This
reduces the RAV scores of c2, c3, c5 and c6 from 80 + 81 = 161 to 80 + 40.5 = 120.5,
so c7, whose RAV score is 147, is selected next. Now, the RAV scores of c8, c9 and c10
become 96 + 24.5 = 120.5. The selection of any of c2, c3, c5, c6, c8, c9 or c10 does not
affect the RAV score of the others, so all seven of these candidates will be selected
before c11, who has 120 approvals. Thus, after the selection of 10 candidates, there
are 120 > 1199
k unrepresented voters who jointly approve c11.
10 = n
To extend this construction to k > 10, we create k − 10 additional candidates and
120(k − 10) additional voters such that for each new candidate, there are 120 new
voters who approve that candidate only. Note that we still have 120 > n
k . RAV will
proceed to select c1, . . . , c10, followed by k − 10 additional candidates, and c11 or one
of the new candidates will remain unselected.
⊓⊔
12
Haris Aziz et al.
While RAV itself does not satisfy JR, one could hope that this can be fixed by tweak-
ing the weights, i.e. that w-RAV satisfies JR for a suitable weight vector w. However,
it turns out that (1, 0, . . . ) is essentially the only weight vector for which this is the
case: Theorem 7 extends to w-RAV for every weight vector w with w1 = 1, w2 > 0.
Theorem 8 For every vector w = (w1, w2, . . . ) with w1 = 1, w2 > 0, there exists a
value of k0 > 0 such that w-RAV does not satisfy JR for k > k0.
Proof Pick a positive integer s ≥ 8 such that w2 ≥ 1
s . Let C = C1 ∪ C2 ∪ {x, y}, where
C1 = {ci, j i = 1, . . . , 2s + 3, j = 1, . . . , 2s + 1},
C2 = {ci i = 1, . . . , 2s + 3}.
For each i = 1, . . . , 2s + 3 and each j = 1, . . . , 2s + 1 we construct 2s3 − s voters who
approve ci, j only and s2 voters who approve ci, j and ci only. Finally, we construct
2s3 − 1 voters who approve x only and s2 − 7s − 5 voters who approve y only (note
that the number of voters who approve y is positive by our choice of s).
Set k0 = (2s + 2)(2s + 3) = C1 ∪ C2. Note that the number of voters n is given by
(2s + 3)(2s + 1)(2s3 + s2 − s) + (2s3 − 1) + (s2 − 7s − 5)
= (2s + 2)(2s + 3)(2s3 − 1) = (2s3 − 1)k0,
= 2s3 − 1.
and hence n
k0
Under w-RAV initially the score of each candidate in C2 is s2(2s + 1) = 2s3 + s2,
the score of each candidate in C1 is 2s3 + s2 − s, the score of x is 2s3 − 1, and the score
of y is s2 −7s−5, so in the first 2s+3 rounds the candidates from C2 get elected. After
that, the score of every candidate in C1 becomes 2s3 − s + w2s2 ≥ 2s3 − s + s = 2s3,
while the scores of x and y remains unchanged. Therefore, in the next (2s + 3)(2s + 1)
rounds the candidates from C1 get elected. At this point, k candidates are elected, and
x is not elected, even though the 2s3 − 1 = n
voters who approve him do not approve
k0
of any of the candidates in the winning set.
To extend this argument to larger values of k, we proceed as in the proof of The-
orem 7: for k > k0, we add k − k0 new candidates, and for each new candidate we
construct 2s3 − 1 new voters who approve that candidate only. Let the resulting num-
ber of voters be n′; we have n′
k = 2s3 − 1, so w-RAV will first select the candidates in
C2, followed by the candidates in C1, and then it will choose k −k0 winners among the
new candidates and x. As a result, either x or one of the new candidates will remain
unselected.
⊓⊔
Remark 1 Theorem 8 partially subsumes Theorem 7: it implies that RAV fails JR, but
the proof only shows that this is the case for k ≥ 18·19 = 342, while Theorem 7 states
that RAV fails JR for k ≥ 10 already. We chose to include the proof of Theorem 7
because we feel that it is useful to know what happens for relatively small values of
k. Note, however, that Theorem 7 leaves open the question of whether RAV satisfies
JR for k = 3, . . . , 9. Very recently, Sanch´ez-Fernand´ez et al. (2016) have answered
this question by showing that RAV satisfies JR for k ≤ 5 and fails it for k ≥ 6.
Justified Representation in Approval-Based Committee Voting
13
If we allow the entries of the weight vector to depend on the number of voters n, we
can obtain another class of rules that provide justified representation: the argument
used to show that GAV satisfies JR extends to w-RAV where the weight vector w
satisfies w1 = 1, w j ≤ 1
n2 , . . . , )-RAV is
somewhat more appealing than GAV: for instance, if Ti∈N Ai = {c} and k > 1, GAV
will pick c, and then behave arbitrarily, whereas (1, 1
n2 , . . . , )-RAV will also pick
c, but then it will continue to look for candidates approved by as many voters as
possible.
n for j > 1. In particular, the rule (1, 1
n , 1
n , 1
We conclude this section by showing that MonAV satisfies JR.
Theorem 9 MonAV satisfies JR.
Proof Fix a ballot profile A = (A1, . . . , An) and a k > 0. Let W be an output of
MonAV on (A, k). If Ai ∩ W , ∅ for all i ∈ N, then W provides justified representation
for (A, k). Thus, assume that this is not the case, i.e. there exists some voter i with
Ai ∩ W = ∅. Consider a valid mapping π : N → W whose Monroe score equals the
Monroe score of W, let c = π(i), and set s = π−1(c); note that s ∈ {⌊ n
k ⌋, ⌈ n
k ⌉}.
Suppose for the sake of contradiction that W does not provide justified represen-
tation for (A, k). Then by our choice of s there exists a set N∗ ⊂ N, N∗ = s, such
that Ti∈N∗ Ai , ∅, but W ∩ (Si∈N∗ Ai) = ∅. Let c′ be some candidate approved by all
voters in N∗, and set W ′ = (W \ {c}) ∪ {c′}. To obtain a contradiction, we will argue
that W ′ has a higher Monroe score than W.
To this end, we will modify π by first swapping the voters in N∗ with voters
in π−1(c) and then assigning the voters in N∗ to c′. Formally, let σ : π−1(c) \ N∗ →
N∗ \ π−1(c) be a bijection between π−1(c) \ N∗ and N∗ \ π−1(c). We construct a mapping
π : N → W ′ by setting
π(σ(i))
c′
π(i)
for i ∈ π−1(c) \ N∗,
for i ∈ N∗,
for i < π−1(c) ∪ N∗.
π(i) =
Note that π is a valid mapping: we have π−1(c′) = s and π−1(c′′) = π−1(c′′) for
each c′′ ∈ W ′ \ {c′}. Now, let us consider the impact of this modification on the
Monroe score. The s voters in N∗ contributed nothing to the Monroe score of π,
and they contribute s to the Monroe score of π. By our choice of c, the voters in
π−1(c) contributed at most s − 1 to the Monroe score of π, and their contribution to
the Monroe score of π is non-negative. For all other voters their contribution to the
Monroe score of π is equal to their contribution to the Monroe score of π′. Thus, the
total Monroe score of π is higher than that of π. Since the Monroe score of W is equal
to the Monroe score of π, and, by definition, the Monroe score of W ′ is at least the
Monroe score of π, we obtain a contradiction.
⊓⊔
5 Extended Justified Representation
We have identified four (families of) voting rules that satisfy JR for arbitrary ballot
profiles: w-PAV with w1 = 1, w j ≤ 1
j for j > 1 (this class includes PAV), w-RAV
14
Haris Aziz et al.
n for j > 1 (this class includes GAV), GAVT and MonAV. The
with w1 = 1, w j ≤ 1
obvious advantage of GAV and GAVT is that their output can be computed efficiently,
whereas computing the outputs of PAV or MonAV is NP-hard. However, GAV puts
considerable emphasis on representing every voter, at the expense of ensuring that
large sets of voters with shared preferences are allocated an adequate number of rep-
resentatives. This approach may be problematic in a variety of applications, such as
selecting a representative assembly, or choosing movies to be shown on an airplane,
or foods to be provided at a banquet (see the discussion by Skowron et al. 2015a).
In particular, it may be desirable to have several assembly members that represent a
widely held political position, both to reflect the popularity of this position, and to
highlight specific aspects of it, as articulated by different candidates. Consider, for
instance, the following example.
Example 2 Let k = 3, C = {a, b, c, d}, and n = 100. One voter approves c, one voter
approves d, and 98 voters approve a and b. GAV would include both c and d in the
winning set, whereas in many settings it would be more reasonable to choose both a
and b (and one of c and d); indeed, this is exactly what GAVT would do.
This issue is not addressed by the JR axiom, as this axiom does not care if a given
voter is represented by one or more candidates. Thus, if we want to capture the intu-
ition that large cohesive groups of voters should be allocated several representatives,
we need a stronger condition. Recall that JR says that each group of n
k voters that all
approve the same candidate "deserves" at least one representative. It seems reason-
able to scale this idea and say that, for every ℓ > 0, each group of ℓ · n
k voters that
all approve the same ℓ candidates "deserves" at least ℓ representatives. This approach
can be formalized as follows.
Definition 2 (Extended justified representation (EJR)) Given a ballot profile
(A1, . . . , An) over a candidate set C, a target committee size k, k ≤ C, and a posi-
tive integer ℓ, ℓ ≤ k, we say that a set of candidates W, W = k, provides ℓ-justified
representation for (A, k) if there does not exist a set of voters N∗ ⊆ N with N∗ ≥ ℓ · n
k
such that Ti∈N∗ Ai ≥ ℓ, but Ai ∩ W < ℓ for each i ∈ N∗; we say that W provides
extended justified representation (EJR) for (A, k) if it provides ℓ-JR for (A, k) for all
ℓ, 1 ≤ ℓ ≤ k. We say that an approval-based voting rule satisfies ℓ-justified represen-
tation (ℓ-JR) if for every profile A = (A1, . . . , An) and every target committee size k it
outputs a committee that provides ℓ-JR for (A, k). Finally, we say that a rule satisfies
extended justified representation (EJR) if it satisfies ℓ-JR for all ℓ, 1 ≤ ℓ ≤ k.
Observe that EJR implies JR, because the latter coincides with 1-JR.
The definition of EJR interprets "a group N∗ deserves at least ℓ representatives"
as "at least one voter in N∗ gets ℓ representatives". Of course, other interpretations
are also possible: for instance, we can require that each voter in N∗ is represented by
ℓ candidates in the winning committee or, alternatively, that the winning committee
contains at least ℓ candidates each of which is approved by some member of N∗.
However, the former requirement is too strong: it differs from JR for ℓ = 1 and
in Section 6 we show that there are ballot profiles for which commitees with this
property do not exist even for ℓ = 1. The latter approach, which was very recently
Justified Representation in Approval-Based Committee Voting
15
proposed by Sanch´ez-Fernand´ez et al. (2016) (see the discussion in Section 7), is not
unreasonable; in particular, it coincides with JR for ℓ = 1. However, it is strictly
less demanding than the approach we take: clearly, every rule that satisfies EJR also
satisfies this condition. As it turns out (Theorem 10) that every ballot profile admits a
committee that provides EJR, the EJR axiom offers more guidance in choosing a good
winning committee than its weaker cousin, while still leaving us with a non-empty set
of candidate committees to choose from. Finally, the EJR axiom in its present form is
very similar to a core stability condition for a natural NTU game associated with the
input profile (see Section 5.2); it is not clear if the axiom of Sanch´ez-Fernand´ez et al.
(2016) admits a similar interpretation.
5.1 Extended Justified Representation under Approval-Based Rules
It is natural to ask which of the voting rules that satisfy JR also satisfy EJR. Example 2
immediately shows that for GAV the answer is negative. Consequently, no w-RAV rule
such that the entries of w do not depend on n satisfies EJR: if w2 = 0, this rule is GAV
and if w2 > 0, our claim follows from Theorem 8. Moreover, Example 2 also implies
that w-RAV rules with w j ≤ 1
n for j > 1 also fail EJR.
The next example shows that MAV fails EJR even if each voter approves exactly
k candidates (recall that under this assumption MAV satisfies JR).
Example 3 Let k = 4, C = C1 ∪ C2 ∪ C3 ∪ C4, where C1 = C2 = C3 = C4 = 4
and the sets C1, C2, C3, C4 are pairwise disjoint. Let A = (A1, . . . , A8), where Ai = Ci
for i = 1, 2, 3, and Ai = C4 for i = 4, 5, 6, 7, 8. MAV will select exactly one candidate
from each of the sets C1, C2, C3 and C4, but EJR dictates that at least two candidates
from C4 are chosen.
Further, MonAV fails EJR as well.
Example 4 Let k = 4, C = {c1, c2, c3, c4, a, b}, N = {1, . . . , 8}, Ai = {ci} for
i = 1, . . . , 4, Ai = {ci−4, a, b} for i = 5, . . . , 8. MonAV outputs {c1, c2, c3, c4} on this
profile, as this is the unique set of candidates with the maximum Monroe score. Thus,
every voter is represented by a single candidate, though the voters in N∗ = {5, 6, 7, 8}
"deserve" two candidates.
Example 4 illustrates the conflict between the EJR axiom and the requirement to rep-
resent all voters whenever possible. We discuss this issue in more detail in Section 7.
For GAV T , it is not hard to construct an example where this rule fails EJR for
some way of breaking intermediate ties.
Example 5 Let N = {1, . . . , 8}, C = {a, b, c, d, e, f }, A1 = A2 = {a}, A3 = A4 =
{a, b, c}, A5 = A6 = {d, b, c}, A7 = {d, e}, A8 = {d, f }. Suppose that k = 4. Note that
all voters in N∗ = {3, 4, 5, 6} approve b and c, and N∗ = 2 · n
k . Under GAVT , at the
first step candidates a, b, c and d are tied, so we can select a and remove voters 3 and
4. Next, we have to select d; we can then remove voters 5 and 6. In the remaining
two steps, we add e and f to the committee. The resulting committee violates EJR,
as each voter in N∗ = {3, 4, 5, 6} is only represented by a single candidate.
16
Haris Aziz et al.
We note that in Example 5 we can remove voters 1 and 2 after selecting a, which
enables us to select b or c in the second step and thereby obtain a committee that
provides EJR. In fact, we were unable to construct an example where GAVT fails
EJR for all ways of breaking intermediate ties; we now conjecture that it is always
possible to break intermediate ties in GAVT so as to satisfy EJR. However, it is not
clear if a tie-breaking rule with this property can be formulated in a succinct manner.
Thus, GAVT does not seem particularly useful if we want to find a committee that
provides EJR: even if our conjecture is true, we may have to explore all ways of
breaking intermediate ties.
In contrast, we will now show that PAV satisfies EJR irrespective of the tie-
breaking rule.
Theorem 10 PAV satisfies EJR.
Proof Suppose that PAV violates EJR for some value of k, and consider a ballot
profile A1, . . . , An, a value of ℓ > 0 and a set of voters N∗, N∗ = s ≥ ℓ · n
k , that
witness this. Let W, W = k, be the winning set. We know that at least one of the ℓ
candidates approved by all voters in N∗ is not elected; let c be some such candidate.
Each voter in N∗ has at most ℓ − 1 representatives in W, so the marginal contribution
of c (if it were to be added to W) would be at least s · 1
k . On the other hand, the
ℓ
argument in the proof of Theorem 6 can be modified to show that the sum of marginal
contributions of candidates in W is at most n.
≥ n
k . If it is strictly less than n
Now, consider some candidate w ∈ W with the smallest marginal contribution;
clearly, his marginal contribution is at most n
k , we are
done, as we can improve the total PAV-score by swapping w and c, a contradiction.
Therefore suppose it is exactly n
k , and therefore the marginal contribution of each
candidate in W is exactly n
k . Since PAV satisfies JR, we know that Ai ∩ W , ∅ for
some i ∈ N∗. Pick some candidate w′ ∈ W ∩ Ai, and set W ′ = (W \ {w′}) ∪ {c}.
Observe that after w′ is removed, adding c increases the total PAV-score by at least
(s − 1) · 1
k . Indeed, i approves at most ℓ − 2 candidates in W \ {w′} and
1
ℓ−1 to her satisfaction. Thus, the
therefore adding c to W \ {w′} contributes at least
PAV-score of W ′ is higher than that of W, a contradiction again.
⊓⊔
ℓ
+ 1
ℓ−1 > n
Interestingly, Theorem 10 does not extend to weight vectors other than
2 , 1
3 , . . . ): our next theorem shows that PAV is essentially the unique w-PAV rule
(1, 1
that satisfies EJR.
Theorem 11 For every weight vector w with w1 = 1, w , (1, 1
w-PAV does not satisfy EJR.
2 , 1
3 , . . . ), the rule
Theorem 11 follows immediately from Lemmas 1 and 2, which are stated below.
Lemma 1 Consider a weight vector w with w1 = 1. If w j > 1
w-PAV fails JR.
j for some j > 1, then
Proof Suppose that w j = 1
ε j ⌉ + 1 so that j
divides k; let t = k
j . Let C = C0 ∪ C1 ∪ · · · ∪ Ct, where C0 = {c}, C1 = · · · = Ct = j,
and the sets C0, C1, . . . , Ct are pairwise disjoint. Note that C = t j + 1 = k + 1. Also,
j + ε for some j > 1 and ε > 0. Pick k > ⌈ 1
Justified Representation in Approval-Based Committee Voting
17
construct t +1 pairwise disjoint groups of voters N0, N1, . . . , Nt so that N0 = k, N1 =
· · · = Nt = j(k−1), and for each i = 0, 1, . . . , t the voters in Ni approve the candidates
in Ci only. Observe that the total number of voters is given by n = k + t j(k − 1) = k2.
k , so every committee that provides justified representation
for this profile must elect c. However, we claim that w-PAV elects all candidates in
C \ {c} instead. Indeed, if we replace an arbitrary candidate in C \ {c} with c, then
under w-PAV the total score of our committee changes by
We have N0 = k = n
k − j(k − 1) · 1
j
+ ε! = 1 − j(k − 1)ε < 1 − jε& 1
ε j' ≤ 0,
i.e. C \ {c} has a strictly higher score than any committee that includes c.
⊓⊔
Lemma 2 Consider a weight vector w with w1 = 1. If w j < 1
w-PAV fails j-JR.
j for some j > 1, then
ε
j − ε for some j > 1 and ε > 0. Pick k > j + ⌈ 1
Proof Suppose that w j = 1
⌉. Let C =
C0 ∪ C1, where C0 = j, C1 = {c1, . . . , ck− j+1} and C0 ∩ C1 = ∅. Note that C = k + 1.
Also, construct k − j + 2 pairwise disjoint groups of voters N0, N1, . . . , Nk− j+1 so that
N0 = j(k − j+1), N1 = · · · = Nk− j+1 = k − j, the voters in N0 approve the candidates
in C0 only, and for each i = 1, . . . , k − j + 1 the voters in Ni approve ci only. Note that
the number of voters is given by n = j(k − j + 1) + (k − j + 1)(k − j) = k(k − j + 1).
k , so every committee that provides EJR must
select all candidates in C0. However, we claim that w-PAV elects all candidates from
C1 and j − 1 candidates from C0 instead. Indeed, let c be some candidate in C0, let c′
be some candidate in C1, and let W = C \ {c}, W ′ = C \ {c′}. The difference between
the total score of W and that of W ′ is
k = k − j + 1 and N0 = j · n
We have n
j(k − j + 1) 1
j
− ε! − (k − j) < 1 − j ·
1
ε
· ε < 1 − j < 0,
i.e. w-PAV assigns a higher score to W. As this argument does not depend on the
choice of c in C0 and c′ in C1, the proof is complete.
⊓⊔
5.2 JR, EJR and Core Stability
One can view (extended) justified representation as a stability condition, by associat-
ing committees that provide JR/EJR with outcomes of a certain NTU game that are
resistant to certain types of deviations.
Specifically, given a pair (A, k), where A = (A1, . . . , An), we define an NTU game
G(A, k) with the set of players N as follows. We assume that each coalition of size x,
k ≤ x < (ℓ +1) n
ℓ n
k , where ℓ ∈ {1, . . . , k}, can "purchase" ℓ alternatives. Moreover, each
player evaluates a committee of size ℓ, ℓ ∈ {1, . . . , k}, using the PAV utility function,
i.e. i derives a utility of 1 + 1
j from a committee that contains exactly j of
her approved alternatives (the argument goes through for w-PAV utilities, as long as
w1 ≥ · · · ≥ wk > 0). Thus, for each coalition S with ℓ n
k a payoff
vector x ∈ Rn is considered to be feasible for S if and only if there exists a committee
k ≤ S < (ℓ + 1) n
2 + · · · + 1
18
Haris Aziz et al.
W ⊆ A with W ≤ ℓ such that xi = ui(W) for each i ∈ S , where ui(W) = 1+· · ·+ 1
Ai∩W .
We denote the set of all payoff vectors that are feasible for a coalition S ⊆ N by V(S ).
We say that a coalition S ⊆ N has a profitable deviation from a payoff vector
x ∈ V(N) if there exists a payoff vector y ∈ V(S ) such that yi > xi for all i ∈ S . A
payoff vector x is stable if it is feasible for N and no coalition S ⊆ N has a profitable
deviation from it; the set of all stable payoff vectors is the core of G(A, k).
The following theorem describes the relationship between JR, EJR, and outcomes
of G(A, k).
Theorem 12 A committee W, W = k, provides justified representation for (A, k) if
and only if no coalition of size ⌈ n
k ⌉ or less has a profitable deviation from the payoff
vector x associated with W. Moreover, W provides extended justified representation
k ≤ N∗ < (ℓ + 1) · n
for (A, k) if and only if for every ℓ ≥ 0 no coalition N∗ with ℓ · n
k ,
∩i∈N∗ Ai ≥ ℓ has a profitable deviation from x.
Proof Suppose that W fails to provide justified representation for (A, k), i.e. there
exists a set of voters N∗, N∗ = ⌈ n
k ⌉, who all approve some candidate c < W, but none
of them approves any of the candidates in W. Then we have xi = 0 for each i ∈ N∗,
and players in N∗ can successfully deviate: the payoff vector y that is associated with
the committee {c} is feasible for N∗ and satisfies yi = 1 for each i ∈ N∗.
Conversely, suppose that W provides justified representation for (A, k), and con-
sider a coalition N∗. If N∗ < ⌈ n
k ⌉, then for every y ∈ V(N∗) we have yi = 0 for all
i ∈ N∗, so N∗ cannot profitably deviate. On the other hand, if N∗ = ⌈ n
k ⌉, then every
payoff vector y ∈ V(N∗) is associated with a committee of size 1. Hence, for every
y ∈ V(N∗) we have yi ≤ 1 for all i ∈ N∗, and if yi = 1 for all i ∈ N∗, then ∩i∈N∗ Ai , ∅,
and therefore, since W provides JR, we have xi ≥ 1 for some i ∈ N∗.
tation for (A, k), there exists an ℓ > 0 and a set of voters N∗, N∗ ≥ ℓ · n
For EJR the argument is similar. If W fails to provide extended justified represen-
k , such that
Ti∈N∗ Ai ≥ ℓ, but Ai ∩W < ℓ for each i ∈ N∗. Then we have xi < 1 + · · · + 1
ℓ for each
i ∈ N∗, and players in N∗ can successfully deviate: if S is a committee that consists
of some ℓ candidates in Ti∈N∗ Ai, then the payoff vector y that is associated with S is
feasible for N∗ and satisfies yi = 1 + · · · + 1
Conversely, suppose that W provides extended justified representation for (A, k),
and consider some ℓ ≥ 0 and some coalition N∗ with ℓ · n
k . We have
argued above that if ℓ = 0, then N∗ cannot profitably deviate. Thus, assume ℓ > 0.
Every payoff vector y ∈ V(N∗) is associated with a committee of size ℓ. Hence, for
every y ∈ V(N∗) we have yi ≤ 1 + · · · + 1
ℓ for all
i ∈ N∗, then ∩i∈N∗ Ai ≥ ℓ. Since W provides EJR, we have xi ≥ 1 + · · · + 1
ℓ for some
i ∈ N∗.
⊓⊔
ℓ for all i ∈ S , and if yi = 1 + · · · + 1
ℓ for each i ∈ N∗.
k ≤ N∗ < (ℓ + 1) n
The second part of Theorem 12 considers deviations by cohesive coalitions.
The reader may wonder if it can be strengthened to arbitrary coalitional deviations,
i.e. whether a committee provides EJR if and only if the associated payoff vector is in
the core of G(A, k). The following example shows that this is not the case.
Justified Representation in Approval-Based Committee Voting
19
Example 6 Let k = 10, C = {x1, x2, . . . , x10, y, z}, N = {1, 2, . . . , 20}, and
A1 = A2 = A3 = {x1, y},
A4 = A5 = A6 = {x1, z},
A7 = . . . = A20 = {x2, . . . , x10}.
Then PAV outputs the committee W = {x1, x2, . . . , x10} for (A, k); in particular, W
provides EJR for (A, k). However, the associated payoff vector x is not in the core, as
the players in {1, 2, 3, 4, 5, 6}, a coalition of size 3 n
k , can successfully deviate: the pay-
off vector associated with {x1, y, z} is feasible for {1, 2, 3, 4, 5, 6} and provides a higher
payoff than x to each of the first six players. We remark that the core of G(A, k) is not
empty: in particular, it contains the payoff vector associated with {x1, . . . , x8, y, z}.
It remains an open question whether the core of G(A, k) is non-empty for every
pair (A, k). Further, while it would be desirable to have a voting rule that outputs a
committee whose associated payoff vector is in the core whenever the core is not
empty, we are not aware of any such rule: every voting rule that fails EJR also fails
this more demanding criterion, and Example 6 illustrates that PAV fails this criterion
as well.
5.3 Computational Issues
In Section 3 we have argued that it is easy to find a committee that provides JR for a
given ballot profile, and to check whether a specific committee provides JR. In con-
trast, for EJR these questions appear to be computationally difficult. Specifically, we
were unable to design an efficient algorithm for computing a committee that provides
EJR; while PAV is guaranteed to find such a committee, computing its output is NP-
hard. We remark, however, that when ℓ is bounded by a constant, we can efficiently
compute a committee that provides ℓ-JR, i.e. the challenge is in handling large values
of ℓ.
Theorem 13 A committee satisfying ℓ-JR can be computed in time polynomial in n
and Cℓ.
Proof Consider the following greedy algorithm, which we will refer to as ℓ-GAV. We
start by setting C′ = C, A′ = A, and W = ∅. As long as W ≤ k − ℓ, we check if
there exists a set of candidates {c1, . . . , cℓ} ⊂ C′ that is unanimously approved by at
least ℓ n
k voters in A′ (this can be done in time n · Cℓ+1). If such a set exists, we set
W := W ∪ {c1, . . . , cℓ} and we remove from A′ all ballots Ai such that Ai ∩ W ≥ ℓ
(note that this includes all ballots Ai with {c1, . . . cℓ} ⊆ Ai). If at some point we have
W ≤ k − ℓ and there is no {c1, . . . , cℓ} that satisfy our criterion or W > k − ℓ, we add
an arbitrary k − W candidates from C′ to W and return W; if this does not happen,
we terminate after having picked k candidates.
Suppose for the sake of contradiction that for some profile A = (A1, . . . , An)
and some k > 0, ℓ-GAV outputs a committee that does not provide ℓ-JR for (A, k).
Then there exists a set N∗ ⊆ N with N∗ ≥ ℓ n
k such that Ti∈N∗ Ai ≥ ℓ and, when
20
Haris Aziz et al.
ℓ-GAV terminates, every ballot Ai such that i ∈ N∗ is still in A′. Consider some
subset of candidates {c1, . . . , cℓ} ⊆ Ti∈N∗ Ai. At every point in the execution of ℓ-GAV
this subset is unanimously approved by at least N∗ ≥ ℓ n
k ballots in A′. As at least
one of {c1, . . . , cℓ} was not elected, at every stage the algorithm selected a set of ℓ
candidates that was approved by at least ℓ n
k ballots (until more than k − ℓ candidates
were selected). Since at the end of each stage the algorithm removed from A′ all
ballots containing the candidates that had been added to W at that stage, it follows
k = n − ℓ n
that altogether the algorithm has removed at least ⌊ k
k
ℓ
ballots from A′. This is a contradiction, since we assumed that, when the algorithm
terminates, the ℓ n
k ballots (Ai)i∈N∗ are still in A′.
− 1) · ℓ n
⌋ · ℓ n
k > ( k
ℓ
For the problem of checking whether a given committee provides EJR for a given
input, we can establish a formal hardness result.
Theorem 14 Given a ballot profile A, a target committee size k, and a committee W,
W = k, it is coNP-complete to check whether W provides EJR for (A, k).
Proof It is easy to see that this problem is in coNP: to show that W does not provide
EJR for (A, k), it suffices to guess an integer ℓ and a set of voters N∗ of size at least
ℓ · n
k such that Ti∈N∗ Ai ≥ ℓ, but Ai ∩ W < ℓ for all i ∈ N∗.
To prove coNP-completeness, we reduce the classic Balanced Biclique problem
(Garey and Johnson 1979, [GT24]) to the complement of our problem. An instance
of Balanced Biclique is given by a bipartite graph (L, R, E) with parts L and R and
edge set E, and an integer ℓ; it is a "yes"-instance if we can pick subsets of vertices
L′ ⊆ L and R′ ⊆ R so that L′ = R′ = ℓ and (u, v) ∈ E for each u ∈ L′, v ∈ R′;
otherwise, it is a "no"-instance.
Given an instance h(L, R, E), ℓi of Balanced Biclique with R = {v1, . . . , vs}, we
create an instance of our problem as follows. Assume without loss of generality that
1, C2, so
s ≥ 3, ℓ ≥ 3. We construct 4 pairwise disjoint sets of candidates C0, C1, C′
1 ∪ C2.
1 = ℓ − 1, C2 = sℓ + ℓ − 3s, and set C = C0 ∪ C1 ∪ C′
that C0 = L, C1 = C′
We then construct 3 sets of voters N0, N1, N2, so that N0 = {1, . . . , s}, N1 = ℓ(s − 1),
N2 = sℓ + ℓ − 3s (note that N2 > 0 as we assume that ℓ ≥ 3). For each i ∈ N0 we set
1. The candidates
Ai = {u j (u j, vi) ∈ E} ∪ C1, and for each i ∈ N1 we set Ai = C0 ∪ C′
in C2 are matched to voters in N2: each voter in N2 approves exactly one candidate
in C2, and each candidate in C2 is approved by exactly one voter in N2. Denote the
resulting list of ballots by A. Finally, we set k = 2ℓ − 2, and let W = C1 ∪ C′
1. Note
that the number of voters n is given by s + ℓ(s − 1) + sℓ + ℓ − 3s = 2s(ℓ − 1), so n
k = s.
Suppose first that we started with a "yes"-instance of Balanced Biclique, and
let (L′, R′) be the respective ℓ-by-ℓ biclique. Let C∗ = L′, N∗ = N0 ∪ N1. Then
N∗ = ℓs, all voters in N∗ approve all candidates in C∗, C∗ = ℓ, but each voter in
N∗ is only represented by ℓ − 1 candidates in W. Hence, W fails to provide ℓ-justified
representation for (A, k).
Conversely, suppose that W fails to provide EJR for (A, k). That is, there exists a
value j > 0, a set N∗ of js voters and a set C∗ of j candidates so that all voters in N∗
approve of all candidates in C∗, but for each voter in N∗ at most j of her approved
candidates are in W. Note that, since s > 1, we have N∗ ∩ N2 = ∅. Further, each voter
in N \ N2 is represented by ℓ − 1 candidates in W, so j ≥ ℓ. As N∗ = js ≥ ℓs ≥ s, it
Justified Representation in Approval-Based Committee Voting
21
follows that N∗ ∩ N0 ≥ ℓ, N∗ ∩ N1 > 0. Since N∗ contains voters from both N0 and
N1, it follows that C∗ ⊆ C0. Thus, there are at least ℓ voters in N∗ ∩ N0 who approve
the same j ≥ ℓ candidates in C0; any set of ℓ such voters and ℓ such candidates
corresponds to an ℓ-by-ℓ biclique in the input graph.
⊓⊔
6 Variants of Justified Representation
The definition of JR requires that if there is a group of ⌈ n
k ⌉ voters who jointly approve
some candidate, then the elected committee has to contain at least one candidate ap-
proved by some member of this group. This condition may appear to be too weak;
it may seem more natural to require that every group member approves some can-
didate in the committee, or-stronger yet-that the committee contains at least one
candidate approved by all group members. This intuition is captured by the following
definitions.
Definition 3 Given a ballot profile (A1, . . . , An) and a target committee size k, we say
that a committee W of size k provides
– semi-strong justified representation for (A, k) if for each group N∗ ⊆ N with
By definition, a committee providing strong justified representation also provides
semi-strong justified representation, and a committee providing semi-strong justified
representation also provides (standard) justified representation.
However, it turns out that satisfying these stronger requirements is not always fea-
sible: there are ballot profiles for which no committee provides semi-strong justified
representation.
Example 7 Let k = 3 and consider the following profile with n = 9 and C =
{a, b, c, d}.
A1 = A2 = {a}
A6 = {c}
A3 = {a, b}
A7 = {c, d}
A4 = {b}
A8 = A9 = {d}
A5 = {b, c}
For each candidate x ∈ C, there are n
k = 3 voters such that ∩iAi = {x}, and at least
one of those voters has Ai = {x}. Thus, a committee that satisfies semi-strong justified
representation would have to contain all four candidates, which is impossible.
While Example 7 shows that no approval-based voting rule can always find a commit-
tee that provides strong or semi-strong justified representation, it may be interesting
to identify voting rules that output such committees whenever they exist.
Finally, we remark that strong justified representation does not imply EJR.
N∗ ≥ n
k and Ti∈N∗ Ai , ∅ it holds that W ∩ Ai , ∅ for all i ∈ N∗.
– strong justified representation for (A, k) if for each group N∗ ⊆ N with N∗ ≥ n
k
andTi∈N∗ Ai , ∅ it holds that W ∩ (∩i∈N∗ Ai) , ∅.
22
Haris Aziz et al.
Example 8 Let C = {a, b, c, d, e}, n = 4, k = 4, and consider the following ballot
profile.
A1 : {a, b}
A2 : {a, b}
A3 : {c}
A4 : {d, e}
EJR requires that we choose both a and b, but {a, c, d, e} provides strong justified
representation.
7 Related Work
It is instructive to compare JR and EJR to alternative approaches towards fair repre-
sentation, such as representativeness (Duddy 2014) and proportional justified repre-
sentation (Sanch´ez-Fernand´ez et al. 2016).
Duddy (2014) proposes the notion of representativeness, which applies to prob-
abilistic voting rules. The property Duddy proposes is incomparable with JR: in sit-
uations he considers (k = 2, n voters approve x, n + 1 voters approve y and z), JR
requires that one of y and z should be selected, whereas Duddy requires x to be se-
lected with positive probability. Both are reasonable requirements, but they address
different concerns. Duddy's axiom say nothing about situations where voters are split
equally (say, n voters approve {x, y}, n voters approve {z, t}), whereas JR requires that
each voter is represented. Another obvious difference is that he allows for randomized
rules.
k such that Ti∈N∗ Ai ≥ ℓ, but (Si∈N∗ Ai) ∩ W < ℓ. In
Very recently (after the conference version of our paper was published),
Sanch´ez-Fernand´ez et al. (2016) came up with the notion of proportional justified
representation (PJR), which can be seen as an alternative to EJR. A committee is
said to provide PJR for a ballot profile (A1, . . . , An) over a candidate set C and a tar-
get committee size k if, for every positive integer ℓ, ℓ ≤ k, there does not exist a set of
voters N∗ ⊆ N with N∗ ≥ ℓ · n
contrast to EJR, the PJR condition does not require one of the voters in N∗ to have ℓ
representatives. Rather, a committee provides PJR as long as it contains ℓ candidates
that are approved by (possibly different) voters in N∗, for every group N∗ satisfying
the size and cohesiveness constraints. An attractive feature of PJR is that it is compat-
ible with the idea of perfect representation: a committee W provides perfect represen-
tation for a group of n voters and a target committee size k if n = ks for some positive
integer s and the voters can be split into k pairwise disjoint groups N1, . . . , Nk of size
s each in such a way that there is a one-to-one mapping µ : W → {N1, . . . , Nk} such
that for each candidate a ∈ W all voters in µ(a) approve a. Sanch´ez-Fernand´ez et al.
(2016) prove that every committee that provides perfect representation also provides
PJR; in contrast, EJR may rule out all committees that provide perfect representation,
as illustrated by Example 4. It is easily seen that PJR is a weaker requirement than
EJR, and a stronger one than JR. Interestingly, Sanch´ez-Fernand´ez et al. (2016) show
that many results that we have established for EJR also hold for PJR: in particular,
w-RAV violates PJR for every weight vector w, and w-PAV satisfies PJR if and only
if w = (1, 1
2 , 1
3 , . . . ).
Justified Representation in Approval-Based Committee Voting
23
8 Conclusions
Rule
AV
SAV
MAV
RAV
GAV
GAVT
MonAV
PAV
JR
EJR
Complexity
–
–
–
–
X
X
X
X
–
–
–
–
–
–∗
–
X
in P
in P
NP-hard
in P
in P
in P
NP-hard
NP-hard
Table 1 Satisfaction of JR and EJR and complexity of approval-based voting rules; the superscript '∗'
indicates that the rule fails the respective axiom for some way of breaking intermediate ties.
We have formulated a desirable property of approval-based committee selection rules,
which we called justified representation (JR). While JR is fairly easy to satisfy, it
turns out that many well-known approval-based rules fail it. A prominent exception
is the PAV rule, which also satisfies a stronger version of this property, namely ex-
tended justified representation (EJR). Indeed, EJR characterizes PAV within the class
of w-PAV rules, and we are not aware of any other natural voting rule that satisfies
EJR irrespective of the tie-breaking rule (of course, we can construct voting rules
that differ from PAV, yet satisfy EJR, by modifying the output of PAV on profiles
on which EJR places no constraints on the output). Perhaps the most pressing open
question suggested by our work is whether there is an efficient algorithm for finding
a committee that provides EJR for a given profile. In particular, we would like to un-
derstand whether we can break ties in the execution of GAVT to always produce such
a committee, and whether some tie-breaking rule with this property is polynomial-
time computable. Also, it would be interesting to see if EJR, in combination with
other natural axioms, can be used to axiomatize PAV. Concerning (semi-)strong jus-
tified representation, an interesting algorithmic problem is whether there are efficient
algorithm for checking the existence of committees satisfying these requirements.
Justified representation can also be used to formulate new approval-based rules.
We mention two rules that seem particularly attractive:
The utilitarian (E)JR rule returns a committee that, among all committees
that satisfy (E)JR, has the highest AV score.
The egalitarian (E)JR rule returns a committee that, among all committees
that satisfy (E)JR, maximizes the number of representatives of the voter who
has the least number of representatives in the winning committee.
The computational complexity of winner determination for these rules is an interest-
ing problem.
Since PAV is NP-hard to compute, our study also provides additional motiva-
tion for the use of approximation and parameterized algorithms to compute PAV out-
comes. Finally, analyzing the compatibility of JR with other important properties,
24
Haris Aziz et al.
such as, e.g., strategyproofness for dichotomous preferences, is another avenue of
future research.
Acknowledgements The authors thank the anonymous reviewers of Multidisciplinary Workshop on Ad-
vances in Preference Handling (MPREF 2014), and the Twenty-Ninth AAAI Conference (AAAI 2015) for
their helpful feedback on earlier versions of the paper. We further thank Martin Lackner and Piotr Skowron
for valuable discussions.
Brill, Conitzer, and Freeman were supported by NSF and ARO under grants CCF-1101659, IIS-
0953756, CCF-1337215, W911NF-12-1-0550, and W911NF-11-1-0332, by a Feodor Lynen research fel-
lowship of the Alexander von Humboldt Foundation, and by COST Action IC1205 on Computational
Social Choice. Brill and Elkind were partially supported by ERC-StG 639945. Walsh also receives sup-
port from the Asian Office of Aerospace Research and Development (AOARD 124056) and the German
Federal Ministry for Education and Research through the Alexander von Humboldt Foundation.
References
H. Aziz, S. Gaspers, J. Gudmundsson, S. Mackenzie, N. Mattei, and T. Walsh. Computational aspects of
multi-winner approval voting. In Proceedings of the 14th International Conference on Autonomous
Agents and Multiagent Systems (AAMAS), pages 107–115, 2015.
N. Betzler, A. Slinko, and J. Uhlmann. On the computation of fully proportional representation. Journal
of Artificial Intelligence Research, 47:475–519, 2013.
S. J. Brams. Preface. In J.-F. Laslier and M. R. Sanver, editors, Handbook on Approval Voting, Studies in
Choice and Welfare, pages vii–ix. Springer, 2010.
S. J. Brams and P. C. Fishburn. Approval Voting. Springer, 2nd edition, 2007.
S. J. Brams and D. M. Kilgour. Satisfaction approval voting. In R. Fara, D. Leech, and M. Salles, editors,
Voting Power and Procedures, Studies in Choice and Welfare, pages 323–346. Springer, 2014.
S. J. Brams, D. M. Kilgour, and M. R. Sanver. How to elect a representative committee using approval
balloting. In B. Simeone and F. Pukelsheim, editors, Mathematics and democracy: Recent advances
in voting systems and collective choice, pages 83–96. Springer, 2006.
S. J. Brams, D. M. Kilgour, and R. M. Sanver. A minimax procedure for electing committees. Public
Choice, 132(3-4):401–420, 2007.
J. Byrka and K. Sornat. PTAS for minimax approval voting.
In Proceedings of the 10th International
Conference on Web and Internet Economics (WINE), pages 203–217, 2014.
I. Caragiannis, D. Kalaitzis, and E. Markakis. Approximation algorithms and mechanism design for min-
imax approval voting. In Proceedings of the 24th AAAI Conference on Artificial Intelligence (AAAI),
pages 737–742, 2010.
I. Caragiannis, C. Kaklamanis, N. Karanikolas, and A. D. Procaccia. Socially desirable approximations
for Dodgson's voting rule. ACM Transactions on Algorithms, 10(2), 2014.
B. Chamberlin and P. Courant. Representative deliberations and representative decisions: Proportional
representation and the Borda rule. American Political Science Review, 77(3):718–733, 1983.
D. Cornaz, L. Galand, and O. Spanjaard. Bounded single-peaked width and proportional representation.
In Proceedings of the 20th European Conference on Artificial Intelligence (ECAI), pages 270–275,
2012.
M. Davis, M. E. Orrison, and F. E. Su. Voting for committees in agreeable societies. Contemporary
Mathematics, 624:147–157, 2014.
H. R. Droop. On methods of electing representatives. Journal of the Statistical Society of London, 44(2):
141–196, 1881.
C. Duddy. Electing a representative committee by approval ballot: An impossibility result. Economic
Letters, 124:14–16, 2014.
M. Dummett. Voting Procedures. Oxford University Press, 1984.
E. Elkind and M. Lackner. Structure in dichotomous preferences. In Proceedings of the 24th International
Joint Conference on Artificial Intelligence (IJCAI), pages 2019–2025, 2015.
E. Elkind, P. Faliszewski, P. Skowron, and A. Slinko. Properties of multiwinner voting rules. In Proceed-
ings of the 13th International Conference on Autonomous Agents and Multiagent Systems (AAMAS),
pages 53–60, 2014.
Justified Representation in Approval-Based Committee Voting
25
E. Elkind, J. Lang, and A. Saffidine. Condorcet winning sets. Social Choice and Welfare, 44(3):493–517,
2015.
U. Endriss. Sincerity and manipulation under approval voting. Theory and Decision, 74(4):335–355, 2013.
P. C. Fishburn and A. S. Pekec. Approval voting for committees: Threshold approaches. Technical report,
2004.
M. Garey and D. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. W.
H. Freeman and Company, 1979.
D. M. Kilgour. Approval balloting for multi-winner elections. In Handbook on Approval Voting, chapter 6.
Springer, 2010.
D. M. Kilgour and E. Marshall. Approval balloting for fixed-size committees.
In D. S. Felsenthal and
M. Machover, editors, Electoral Systems, Studies in Choice and Welfare, pages 305–326. Springer,
2012.
J.-F. Laslier and M. R. Sanver, editors. Handbook on Approval Voting. Studies in Choice and Welfare.
Springer, 2010.
R. LeGrand, E. Markakis, and A. Mehta. Some results on approximating the minimax solution in approval
voting. In Proceedings of the 6th International Conference on Autonomous Agents and Multi-Agent
Systems (AAMAS), pages 1185–1187, 2007.
T. Lu and C. Boutilier. Budgeted social choice: From consensus to personalized decision making.
In
Proceedings of the 22nd International Joint Conference on Artificial Intelligence (IJCAI), pages 280–
286. AAAI Press, 2011.
R. Meir, A. D. Procaccia, J. S. Rosenschein, and A. Zohar. Complexity of strategic behavior in multi-
winner elections. Journal of Artificial Intelligence Research, 33:149–178, 2008.
N. Misra, A. Nabeel, and H. Singh. On the parameterized complexity of minimax approval voting. In
Proceedings of the 14th International Conference on Autonomous Agents and Multiagent Systemsn
(AAMAS), pages 97–105, 2015.
B. Monroe. Fully proportional representation. American Political Science Review, 89(4):925–940, 1995.
A. Procaccia, J. Rosenschein, and A. Zohar. On the complexity of achieving proportional representation.
Social Choice and Welfare, 30(3):353–362, 2008.
L. Sanch´ez-Fernand´ez, N. Fern´andez, J. A. Fisteus, and P. Basanta-Val. Some notes on justified represen-
tation. In Proceedings of the 10th Multidisciplinary Workshop on Advances in Preference Handling
(MPREF), 2016.
P. Skowron, P. Faliszewski, and J. Lang. Finding a collective set of items: From proportional multirep-
In Proceedings of the 29th AAAI Conference on Artificial
resentation to group recommendation.
Intelligence (AAAI), pages 2131–2137, 2015a.
P. Skowron, P. Faliszewski, and A. Slinko. Achieving fully proportional representation: Approximability
results. Artificial Intelligence, 222:67–103, 2015b.
P. Skowron, L. Yu, P. Faliszewski, and E. Elkind. The complexity of fully proportional representation for
single-crossing electorates. Theoretical Computer Science, 569:43–57, 2015c.
P. K. Skowron, P. Faliszewski, and J. Lang. Finding a collective set of items: From proportional mul-
tirepresentation to group recommendation. In Proceedings of the 29th International Conference on
Autonomous Agents and Multi-Agent Systems (AAMAS), pages 2131–2137. AAAI Press, 2015d.
T. N. Thiele. Om flerfoldsvalg. In Oversigt over det Kongelige Danske Videnskabernes Selskabs Forhan-
dlinger, pages 415–441. 1895.
N. Tideman. The single transferable vote. Journal of Economic Perspectives, 9(1):27–38, 1995.
N. Tideman. Collective Decisions and Voting. Ashgate, 2006.
N. Tideman and D. Richardson. Better voting methods through technology: The refinement-manageability
trade-off in the single transferable vote. Public Choice, 103(1-2):13–34, 2000.
|
1210.4849 | 1 | 1210 | 2012-10-16T17:35:25 | Toward Large-Scale Agent Guidance in an Urban Taxi Service | [
"cs.MA",
"cs.AI",
"cs.GT"
] | Empty taxi cruising represents a wastage of resources in the context of urban taxi services. In this work, we seek to minimize such wastage. An analysis of a large trace of taxi operations reveals that the services' inefficiency is caused by drivers' greedy cruising behavior. We model the existing system as a continuous time Markov chain. To address the problem, we propose that each taxi be equipped with an intelligent agent that will guide the driver when cruising for passengers. Then, drawing from AI literature on multiagent planning, we explore two possible ways to compute such guidance. The first formulation assumes fully cooperative drivers. This allows us, in principle, to compute systemwide optimal cruising policy. This is modeled as a Markov decision process. The second formulation assumes rational drivers, seeking to maximize their own profit. This is modeled as a stochastic congestion game, a specialization of stochastic games. Nash equilibrium policy is proposed as the solution to the game, where no driver has the incentive to singly deviate from it. Empirical result shows that both formulations improve the efficiency of the service significantly. | cs.MA | cs | Toward Large-Scale Agent Guidance in an Urban Taxi Service
Lucas AGUSSURJA
School of Information Systems
Singapore Management University
80 Stamford Road, S178902
Hoong Chuin LAU
School of Information Systems
Singapore Management University
80 Stamford Road, S178902
Abstract
Empty taxi cruising represents a wastage of
resources in the context of urban taxi ser-
vices.
In this work, we seek to minimize
such wastage. An analysis of a large trace of
taxi operations reveals that the services’ inef-
ficiency is caused by drivers’ greedy cruising
behavior. We model the existing system as
a continuous time Markov chain. To address
the problem, we propose that each taxi be
equipped with an intelligent agent that will
guide the driver when cruising for passengers.
Then, drawing from AI literature on multi-
agent planning, we explore two possible ways
to compute such guidance. The first formula-
tion assumes fully cooperative drivers. This
allows us, in principle, to compute system-
wide optimal cruising policy. This is modeled
as a Markov decision process. The second
formulation assumes rational drivers, seeking
to maximize their own profit. This is mod-
eled as a stochastic congestion game, a spe-
cialization of stochastic games. Nash equi-
librium policy is proposed as the solution to
the game, where no driver has the incentive
to singly deviate from it. Empirical result
shows that both formulations improve the ef-
ficiency of the service significantly.
1 INTRODUCTION
Taxis are a ma jor mode of transport in every urban
city in the world. In Singapore, as of April 2009, there
were about 24,000 taxis and 87,000 licensed drivers,
providing around 850,000 trips daily. These are op-
erated by a small number of companies. The largest
company, ComfortDelgro, operates over 15,000 taxis,
and captures the ma jority of the market share in terms
of ridership. Like many congested cities in the world,
commuters in Singapore often view taxis as a more
efficient mode of transport compared to private cars.
Data from the Singapore Land Transport Authority
(the regulatory agency for land transportation) show
that taxis on average chalk up higher mileage than pri-
vate cars, with single-shift taxis traveling some 120,000
km a year. More than a third of this travel is empty
cruising, as shown in the next section, see (Tan et al.,
2009) as well, which represents a significant wastage
of resources.
Taxi services have been studied extensively in the lit-
erature. Research has been conducted to investigate
the services’ demand-supply interaction and the re-
sulting market equilibrium (Cairns and Liston-Heyes,
1996; Yang et al., 2002). These studies aim to pre-
dict the economic consequences of regulatory policies,
such as entry restriction and fare control. At the op-
erational level, quantitative models have been built to
capture, for example, passengers’ and drivers’ bilateral
searching behavior (Wong et al., 2005). While these
models tend to be descriptive, our work is operative
by nature. We seek to provide solutions to a specific
problem, namely, we improve the inefficiency of exist-
ing cruising policy by providing alternative policies de-
rived from multi-agent planning and decision making
models. We believe the public transport arena offers
a rich domain for application of multi-agent concepts
and methodologies.
We start by analyzing a dataset obtained from a ma-
jor taxi operator, which traces the movement of a large
number of taxis in Singapore. Using the data for the
month of July 2009, our analysis shows that the cur-
rent system’s inefficiency, measured in terms of cruis-
ing hours, is due to the inherently greedy behavior of
drivers. This, in turn, is caused by the lack of visibil-
ity in regards to the distribution of cruising taxis on
the network at different time periods. We then model
the existing taxi service by a continuous-time Markov
chain, whose parameters are derived from the dataset.
The result is used as the baseline for empirical com-
parison with our approach.
games remains an important open problem.
We proceed to propose two models: a cooperative and
a non-cooperative model.
In our formulations, each
driver is endowed with a guidance agent that provides
suggestions on how to cruise in search of passengers.
In the first formulation, we assume that each driver is
fully cooperative, and therefore willing to follow a cen-
tralized policy (that could be suboptimal for him/her
individually). This allows us, in principle, to compute
the system-wide optimal policy, where the ob jective is
to maximize overall occupied time. For this purpose,
we model the problem as a Markov decision process. In
the second formulation, we assume that drivers are ra-
tional who seek to maximize their respective occupied
time. Here, there is an implicit competition among
drivers, simply because an increase in the number of
cruising taxis in a zone decreases each driver’s chance
of finding passengers within a given time. This leads
directly to a game theoretic formulation. We model
the problem as a stochastic congestion game (a special-
ization of stochastic games) and seek to find a Nash
equilibrium policy. Given an equilibrium policy, no
driver has the incentive to singly deviate from it. Both
problems are solved for finite horizon using value iter-
ation coupled with a sampling technique.
There has been a surge of interest, in recent years,
among AI and complexity theory communities in com-
putational problems related to game theory. The cen-
tral problem in the area is the computation of Nash
equilibrium in different game settings. Classical algo-
rithms to solve the standard simultaneous games have
relied on homotopy methods, which solve fixpoint (one
instance of which is the Nash equilibrium) problems.
The most well-known of such methods is the Lemke-
Howson algorithm. See (Herings and Peeters, 2009)
for a recent survey, and (Goldberg et al., 2011) for a
discussion on the complexity of such methods. Non-
homotopy attempts in the literature include enumer-
ation of strategy supports (Porter et al., 2004) and
mixed-integer programming (Sandholm et al., 2005).
specifically in
In the setting of dynamic games,
stochastic games (Shapley, 1953), the problem of com-
puting Nash equilibrium has been cast in the context
of multiagent reinforcement learning, first introduced
by (Littman, 1994), who shows the convergence of
value iteration in 2-player zero-sum stochastic games.
(Hu and Wellman, 2003) attempts to generalize the
result to n-player general-sum games by extending Q-
learning algorithm. Their algorithm converges to Nash
equilibrium in a restrictive case. Recent attempt by
(Kearns et al., 2000) proposes an algorithm that con-
verges to an equilibrium-like joint policy but not Nash
equilibrium. Finding an algorithm that converges to
Nash equilibrium in n-player general-sum stochastic
Despite the tremendous interest, many interesting
real-world problems (such as the one presented here)
are so large that even the best algorithms have no hope
of computing an equilibrium directly. Furthermore, it
has been shown that the problem is likely to be in-
tractable even for the case of two-player simultaneous
games (Chen and Deng, 2006), it being complete for
the complexity class PPAD. The standard approach
to overcome this problem is to construct a smaller
game that is similar to the original game, and solve
the smaller game. Then, the solution is mapped to
a strategy profile in the original game (Ganzfried and
Sandholm, 2011).
In this work, we tackle a moderate-size real-world
problem, by solving for finite horizon approximate
equilibrium. Our aim is to extend this solution ap-
proach to cover the full scale problem in the future.
Our experimental results show that the optimal pol-
icy (derived from the cooperative model) manages to
reduce the cruising time of the existing policy by ap-
proximately 30%. This result may be viewed as what
can be achieved theoretically. In real-life implementa-
tion, however, the improvement will be closer to that
of equilibrium policy, which is approximately 20%.
2 ANALYSIS OF CURRENT
SYSTEM
Figure 1 and 2 summarize the taxi operation on week-
days (Mon-Thu) for the month of July 2009. In Figure
1a, we see the comparison of average daily time spent
between delivering passengers, cruising, and picking
up passengers (for booked trips) for different periods
of the day. The average number of operating taxis
is also shown in the figure. To book a taxi, passen-
gers may send their request to the central operator
and specify their current location. The central oper-
ator will then broadcast the request to cruising taxis
around the specified location. Drivers who receive the
request may bid for the job by specifying the time re-
quired to reach the pickup point. The job is given to
the driver with the shortest pickup time. The pro-
portion of street-hail vs. booked trips, for different
periods of the day, is shown in Figure 1b. From Fig-
ure 1a, we observe that taxi cruising time is almost
constant throughout the day, accounting for roughly
half of the total operating hours. On the other hand,
pickup time accounts for only a small percentage of
operating hours. For this reason, in this work, we are
focusing on reducing the cruising time for street-hail
jobs.
Figure 2 describes the cruising behavior of drivers. In
Figure 2a, we see that cruising time is high for zones
with a high number of available trips. This indicates
that drivers are spending time cruising in high trip-
frequency zones. The same can be observed from Fig-
ure 2b and 2c. Figure 2b shows that taxis are entering
high trip-frequency zones in anticipation for the surge
of passengers. The number of cruising taxis will drop
with the surge, but starts to build up again for the next
surge, and the cycle continues. On the other hand,
in low trip-frequency zones, except for early morning
hours, the number of cruising taxis stays proportional
to trip-frequency. We can conclude that drivers are
employing a greedy cruising policy, spending a large
amount of time cruising in high trip-frequency zone.
This is one of the causes for the system’s inefficiency,
which is verified in the empirical study, when this pol-
icy is substituted for better cruising policies.
2.1 A MODEL FOR THE EXISTING
SYSTEM
In this section, we model the aggregate behavior of a
taxi service as a continuous time Markov chain.
In
our model, the taxi service operates on a road net-
work which can be divided into logical cruising zones.
We denote the network of zones by a directed graph
G = (N , E ). At any point of time, a taxi is in one of
the following states, which corresponds to the states in
the Markov chain S = {Okl , Ck , Wk k , l ∈ N }: (1) the
taxi is occupied and is delivering its passenger from
zone k to l, denoted by the state Okl , or (2) the taxi is
empty and cruising in a zone k , denoted by Ck , or (3)
the taxi is in zone k but not in operation, denoted by
Wk . Here, we assume that drivers have uniform cruis-
ing behavior that are independent of each other. As an
example, Figure 3 shows a subset of Singapore’s road
network and the corresponding Markov chain, model-
ing the taxi service operating on the network.
In a continuous time Markov chain, the time spent in
a state before moving to another state is a continu-
ous random variable that is exponentially distributed.
The rates of transition between the states constitute
the generator matrix (or Q-matrix) of the Markov
chain.
In our model, the generator matrix consists
of the following four components. The first compo-
nent, {λkl (k , l) ∈ E }, describes a driver’s cruising
behavior. Since a driver does not have visibility in
regards to the state and location of other drivers, we
can assume that they are approximately independent
of each other. The second component, {πkl k , l ∈ N },
describes the likelihood of finding passengers in a zone,
∑
where πkl is the rate of finding passengers in zone k
with a destination point in zone l. Assuming indepen-
dence for πkl , the total rate of finding passengers in
l πkl (a combination of Poisson
zone k is given by
(a) The histogram shows the average total time spent
daily, in hours, on: delivering passengers, cruising, and
reaching pickup points (for booked trips) respectively, in
different periods of the day. The curve shows the number
of operating taxis.
(b) The histogram shows the average daily number of
street-hail and booked trips respectively in different pe-
riods.
Figure 1: Summary of weekday data for July 2009
(a) The correlation between trip fre-
quency and cruising time in a zone
(b) Number of
trips vs.
cruising taxis in a high trip-
frequency zone with peaks
(c) Number of
trips vs.
cruising taxis in a low trip-
frequency zone
Figure 2: Inefficiency of existing cruising policy
0 5000 10000 15000 20000 250000-33-66-99-1212-1515-1818-2121-24Period of the DayOccupied TimeCruising TimePickup TimeOperating Taxis 0 10000 20000 30000 40000 50000 600000-33-66-99-1212-1515-1818-2121-24Period of the DayStreet-Hail TripsBooked Trips 0 500 1000 1500 2000 2500 3000 3500 4000 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045Cruising Time (hours)Number of Trips (normalized)ZoneLeast Squares Fit 0 100 200 300 400 500 600 0 10 20 30 40 50 60 70 80 90 100Period of the DayNumber of TripsNumber of Cruising Taxis 0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90 100Period of the DayNumber of TripsNumber of Cruising Taxis Figure 3: An example of a road network, divided
into its cruising zones, and the corresponding Markov
chain, modeling a taxi operating on the network. The
shaded nodes indicate the states where the taxi is oc-
cupied. Here, we assume that taxis are uniform and
independent of each other.
Figure 4: The frequency histograms (for one day) for
cruising time in a randomly chosen zone, and delivery
time between two randomly chosen zones, respectively.
can be estimated accurately with exponential distribu-
tions. The other variables, however, are closer to Er-
lang distributions than exponential distributions. Fig-
ure 4 shows, for example, two frequency histograms,
one for each of these cases. In this work, we approx-
imate both cases by exponential distributions using
maximum likelihood estimate. These approximations
are better when the zones are adequately close to each
other, which is the case for Singapore.
processes). The third component, {ρkl k , l ∈ N }, de-
scribes the time needed to deliver passengers to their
destination point. The nondeterminism of these vari-
ables is due to the variability of pickup and drop-off
points within zones and the congestion on the road
network. The forth component, {σk , σ
k ∈ N }, de-
′
k
scribes the likelihood of a driver taking a break within
′
a zone (σk ) and the length of the break (σ
k ). Most
taxis are operated by two or more drivers taking turns
continuously. This accounts for the nondeterminism
of the forth component.
One can derive the system efficiency once the param-
eters of the Markov chain are obtained. Since we also
like to compute efficiency under different customer ar-
rival rates and different numbers of operating taxis
(for empirical study), we model the passenger-driver
dynamics in a zone as an M/M/1 queue. Consider
the subset of the Markov chain consisting only the
states {Ck k ∈ N } and the transitions rate between
them {λkl (k , l) ∈ E }. Each zone is modeled as a
FIFO queue, where its arrival rate is the arrival rate
of cruising taxis in the zone, and its service rate is the
passengers arrival rate in the zone. Let µk denote the
We define the system efficiency as the steady state
passengers arrival rate in zone k , then the waiting time
(stationary) probability of a taxi being in the occupied
in the queue (queueing time + service time) is expo-
state. Let θ(s), for s ∈ S , denote the steady state
,
∑
nentially distributed with the following rate, which we
probability of being in the state s. This probability
associate with πk :
∑
∑
distribution can be obtained by solving the following
system of equations:
θ(Ck ),
∑
k θ(Wk )
ρlk θ(Olk ) + σ
λlk θ(Cl ) +
∀k ∈ N ,
′
l:(l;k)∈E
∑
l:l∈N
l:(l;k)∈E
where n is the number of taxis in operation and ϕ(Cl )
the steady state probability of being in the state Cl .
∑
∑
=
λkl +
πkl + σk
The steady state probabilities ϕ can be obtained, sim-
l:l∈N
l:(k;l)∈E
ilar to θ, by solving the following system of equations:
πkl θ(Ck ) = ρkl θ(Okl ),
∑
λl;k ϕ(Cl ) = ϕ(Ck )
∀k ∈ N ,
k θ(Wk ),
σk θ(Ck ) = σ
∑
′
l:(l;k)∈E
l:(k;l)∈E
∑
k∈N ϕ(Ck ) = 1. Now, given the drivers’
s∈S θ(s) = 1.
and normalizing the solution such that
sub ject to
The steady state probability of a taxi being in the
k;l∈N θ(Okl ).
cruising behaviors (in the form of smaller Markov
chain parameterized by {λkl (k , l) ∈ E }), travel time
occupied state is thus given by
information {ρkl k , l ∈ N }, and nonoperating profile
We estimate the transition rates of the Markov chain
{σk , σ
k ∈ N }, we can derive the system efficiency as
′
using the dataset. The random variables cruising time
the function of passenger arrival rates {µk k ∈ N } and
k
(with rate λkl ) and passenger find time (with rate πk )
∀k , l ∈ N ,
∀k ∈ N ,
λkl
,
πk = µk − n
ϕ(Cl )λlk
0 50 100 150 200 250 300 350 400 450 0 5 10 15 20 25FrequencyCruising Time (Minutes)Cruising Time in a chosen zone 0 500 1000 1500 2000 2500 0 5 10 15 20 25FrequencyDelivery Time (Minutes)Delivery time between two chosen zones the total number of taxis n, by first constructing the
full Markov chain, and then deriving its steady state
probabilities.
3 A COOPERATIVE MODEL
The cooperative model is based on Markov decision
process, which is a special case of stochastic games,
where there is only one player. The player, in this
case, is a central operator whose actions are all pos-
sible joint actions of the drivers, and whose ob jective
is the overall occupied time of the service. The model
is similar to the noncooperative case. The difference
lies in the formulation of the utility function. In the
noncooperative case, we have a set of utility functions
to be optimized simultaneously, while in the coopera-
tive case, we have a single aggregated utility function.
We choose to present our models under the more gen-
eral setting of stochastic games (next section). We will
highlight the differences for the cooperative case.
4 A NONCOOPERATIVE MODEL
Stochastic games (Shapley, 1953; Littman, 1994) are a
generalization of Markov decision processes to a multi-
agent setting by allowing the state transitions to be
influenced by their joint action. They are also a gen-
eralization of sequential games with perfect informa-
tion, by introducing different states. The state (and
thus the payoff matrix) changes as a result of both na-
ture and the joint action of the agents. In congestion
games, agents share a set of facilities (facilities can be
viewed as resources, such as machines), and the utility
an agent derives from using a facility reduces as the
number of agents using the same facility increases. In
the taxi context, the facilities correspond to the zones
where taxi agents can cruise. Each zone has a fixed
arrival rate of passengers, and therefore, an increase
in the number of cruising taxis in the zone, decreases
the likelihood that each would find passengers within
a certain time period. Stochastic congestion games
therefore are a generalization of congestion games, al-
lowing the games to be played indefinitely. They are
also a specialization of stochastic games, to the case
where the underlying games are congestion games. A
state in a stochastic congestion game defines an assign-
ment of agents to facilities, and transitions to a new
state are dependent on the old state and the agents’
joint action. In this work, we consider only games with
finite horizon.
4.1 STOCHASTIC CONGESTION GAMES
Formally, a finite stochastic congestion game (SCG) Γ
is a tuple (I , J, P , R), where
• I = {1, . . . , n} is the set of agents.
• J = {1, . . . , m} is the set of facilities. An action
of an agent i, denoted by ai where ai ∈ J , is
to choose a facility to which it would like to be
assigned. We denote an agents’ joint action by
a = (a1 , . . . , an ) and the set of possible joint ac-
tions by A, that is, A = J n . Next, we denote by
S , the set of possible states, where a state s ∈ S
is an assignment of agents to facilities, and we de-
fine sj as the number of agents assigned to facility
j in the state s.
• P : S × A → △(S ) is the state transition function.
′
For convenience, we will also write P (s, a, s
) for
′
the probability that the next state is s
given that
the current state is s and the players’ joint action
is a = (a1 , . . . , an ) ∈ A,
• R = {rj }j∈J is the reward function, where each
rj : {0, . . . , n} → R is a nonincreasing function
that maps the number of agents assigned to fa-
cility j to the reward that each of the agents in
j receives. We define s(i, j ) such that s(i, j ) = 1
∑
if agent i is assigned to facility j in the state s,
and s(i, j ) = 0 otherwise. The reward received
by agent i in a state s, denoted by Ri (s), is thus
j∈J s(i, j )rj (sj ).
given by
The game proceeds in steps, starting from some ini-
tial state sT (the subscript indicates the number of re-
maining steps). In each step, the agents first observe
the current state st , the number of remaining steps t,
and simultaneously choose actions according to their
respective policy. An agent i’s policy is the function
π i : S × {1, . . . , T } → △(J ), where π i (st , t) computes
agent i’s mixed strategy, i.e., a probability distribution
over the set of facilities, given the current state st and
the number of remaining steps t. Each agent i’s ac-
tion in this step, denoted by ai
t , is then drawn from the
distribution given by π i (st , t), forming the joint action
at = (ai
t , . . . , an
t ). Nature then selects the next state
st−1 according to the probabilities given by P (st , at ).
In the new state st−1 , each agent i receives Ri (st−1 )
as its reward, and the game proceeds to the next step.
Given a possible outcome of the game, agent i’s utility,
T −1∑
denoted by U i , is defined such that
t=0
T (sT , aT , sT −1 , . . . , s1 , a1 , s0 ) =
U i
Ri (st ).
A Nash equilibrium of the game is a joint policy π ,
such that, for each agent i, π i maximizes the expected
utility of agent i given that the other players follow
their respective policy specified by π . The expectation
is taken over all possible outcomes of the game. For
the cooperative case, there is only one utility formed
by summing up individual agent’s utility. Here, we
seek to find a joint policy that maximizes the expected
value of this utility.
4.2 MODELING TAXI SYSTEM AS AN
SCG
The drivers correspond to the set of agents I . The set
of facilities J corresponds to the states of the Markov
chain, S = {Ck , Okl , Wk k , l ∈ N }, defined previously.
We will refer to them as facilities here. A state of the
game is an assignment of agents to facilities. When an
agent is assigned to facility Ck , it means that the cor-
responding taxi is cruising in zone k . Similarly, when
it is assigned to facility Okl and Wk , it means that
the corresponding taxi is occupied and not in opera-
tion, respectively. One step in the SCG represents the
period of one minute.
Next, we define available actions and the state transi-
tion function. The set of available actions for an agent
depends on the state of the agent. When in facility Ck ,
the agent may take one of the following actions: (a1)
continue cruising in the current zone, or (a2) make an
attempt to move to an adjacent zone. In this state,
the agent has the chance of getting a passenger and
be moved to one of the facilities Okl in the next step.
Given that an agent is in facility Ck and takes the ac-
tion a1, the following may happen in a step: (1) The
agent manages to find a passenger with l as the desti-
nation zone. In the next step, the agent will move to
facility Okl . The probability of this event happening
is given by:
.
γkl ,
µk
n(Ck )
where γkl is the probability that a passenger’s desti-
nation zone is l given that its starting zone is k , and
n(Ck ) is the number of agents in facility Ck in the cur-
πkl∑
rent state. γkl can be estimated by the following:
γkl ≈
l∈N πkl
The expression µk /n(Ck ) also defines the reward of the
agent in this state. (2) The agent doesn’t find any pas-
senger and stays in the same facility. The probability
of this event is given by: (1 − µk /n(Ck ))e
−(cid:27)k , or (3)
the agent doesn’t find any passenger and moves to fa-
cility Wk with probability (1 − µk /n(Ck ))(1 − e
−(cid:27)k ).
On the other hand, if the agent chooses a2, one of the
following may occur: (1) a passenger, with destina-
tion l, is found before the agent manages to move to
a new zone. The agent moves to facility Okl in the
next step. The probability of this happening is given
by µk γkl /n(Ck ). (2) Otherwise, no passenger is found,
and the agent moves to a new zone.
Algorithm FiniteNashPolicies(Γ, T ):
Initialization phase:
For all s ∈ S , joint action a ∈ A:
π(s, 0)[a] ← 1A ;
For all i ∈ I :
Vs;0 (a, i) ← 0;
Vs;t (a, i) ← ∑
Iteration phase:
For t = 1 to T , all s ∈ S :
{
∑
For all pure strategy profile a ∈ A, i ∈ I :
)×
′
s′ P (s, a, s
, t − 1)[a
′
′
′
′
Ri (s
) +
a′ π(s
π(s, t) ← FindNash(Vs;t );
Return π ;
]Vs′ ;t−1 (a
, i)
}
;
Figure 5: Value Iteration for Finite-horizon SCGs
When in facility Okl , the agent has only one available
action. It will try to deliver its passengers to their des-
tination point and move to facility Cl . The probability
of this happening in the next step is given by 1 − e
−(cid:26)kl ,
while the probability of staying in the same facility is
−(cid:26)kl . The reward for the agent in this state is one.
e
Similarly, when in facility Wk , the agent may move
to Ck with probability 1 − e
−(cid:27)
′
k , or otherwise stays in
the same zone. The reward for being in this facility
is zero. Note that in our model, the move to facility
Wk is involuntary on the part of the agents. An agent
will never choose to stop its operation because of the
way the reward function is structured. Rather than
being an available action, the move is treated as a re-
quirement instead, for example, when the driver has
to change shift or rest.
4.3 COMPUTING EQUILIBRIUM POLICY
In this section, we describe an algorithm to computing
equilibrium policy in SCGs based on value iteration al-
gorithm, shown in Figure 5. This algorithm takes an
SCG Γ, and a time horizon T as input. It outputs a
vector of policies (π i (s, t))i∈I , where each π i maps any
state s, and the number of remaining steps t ≤ T to
a mixed strategy for player i, which is a probability
distribution over the set of facilities. The algorithm
outputs a vector of policies that is a Nash equilibrium
for the T -step SCG Γ from any start state. This algo-
rithm is a generalization of the classical finite-horizon
value iteration for Markov decision processes (Kael-
bling et al., 1998). Instead of backup values, the al-
gorithm maintains backup matrices, denoted by Vs;t .
Each Vs;t (a, i) is a function that takes a joint action a,
a player i, and returns the expected utility of player
i in the t-step SCG Γ, if the game starts at the state
s and the first joint action of the agents is a. We can
view, therefore, each Vs;t as defining a game in the
usual sense, and associated with each Vs;t is a Nash
equilibrium π(s, t), where π i (s, t) denotes the mixed
strategy of player i. We denote by π(s, t)[a] the prob-
ability assigned by π(s, t) to the joint strategy a. Both
Vs;t and π(s, t) are constructed iteratively. The algo-
rithm uses the function FindNash that returns a Nash
equilibrium given the game Vs;t . For the cooperative
case, the function FindNash is replaced by FindOpt
which computes social optimum policy instead of Nash
equilibrium.
Computing Nash equilibrium in stochastic games
is a very difficult problem.
See,
for example,
(Bowling, 2000; Hu and Wellman, 2003; Ganzfried and
Sandholm, 2009) for recent works in this area. And the
value iteration algorithm (for finite horizon stochas-
tic games) is computationally very expensive for large
scale problem.
In this work, we employ the sparse
sampling technique proposed by (Kearns et al., 2000).
5 EMPIRICAL RESULT
We run a simulation (on a macro level) to evaluate the
performance of our proposed solutions. We choose, as
the case study, 15 connected zones that represent the
congested central business district of Singapore and
its surrounding areas. We consider only passengers
with origins and destinations within these zones, and
restrict the cruising area of drivers to these zones as
well. The number of operating taxis is 500. From
the dataset we estimate the the passenger arrival rates
{µk k ∈ N }, delivery time rate {ρkl k , l ∈ N }, and
taxis nonoperating profile {σk , σ
k ∈ N } for these
′
k
zones in different periods. Using this setting we com-
pare existing cruising policy (also derived from the
dataset), the optimal policy (computed from the co-
operative model), and the equilibrium policy (com-
puted from the noncooperative model). The measure
of comparison is cruising time. For both cooperative
and noncooperative model, joint actions are computed
every minute with 2-step look ahead (T = 2).
We simulate one day of operation on a weekday. The
driver-passenger dynamics in each zone is implemented
as a FIFO queue. As a passenger appears in a zone,
it joins the queue associated with the zone. If there
are some taxis cruising in the zone, the passenger at
the head of the queue is removed and assigned to a
randomly chosen taxi. The average time spent in the
queue models the average passengers waiting time.
The experimental results are shown in Figures 6, 7 and
8. Figure 6 compares the average cruising time (over
multiple runs) of the three policies. On average, the
optimal policy and the equilibrium policy reduce the
cruising time of existing policy by approximately 30%
and 20% respectively. Most of the savings are obtained
during the peak periods. This matches our intuition,
Figure 6: Comparison of cruising time between three
policies. The optimal and equilibrium policies reduce
the cruising time by approximately 30% and 20% re-
spectively.
Figure 7: Number of trips vs. number of cruising taxi
in a high trip-frequency zone.
In general, the new
policies send lower number of cruising taxis to high
trip-frequency zones compared to existing policy.
Figure 8: Number of trips vs. number of cruising taxi
in a low trip-frequency zone. In general, the new poli-
cies send higher number of cruising taxis to low trip-
frequency zones compared to existing policy.
200 300 400 500 600 7000-33-66-99-1212-1515-1818-2121-24Cruising Time (hours)Period of the DayExisting PolicyOptimal PolicyEquilibrium Policy 0 50 100 150 200 250 300 20 30 40 50 60 70 80 90 100Period of the DayNumber of TripsNumber of Cruising Taxis (Existing Policy)Number of Cruising Taxis (Optimal Policy) 0 2 4 6 8 10 12 14 16 20 30 40 50 60 70 80 90 100Period of the DayNumber of TripsNumber of Cruising Taxis (Existing Policy)Number of Cruising Taxis (Optimal Policy) that the new policies are able to better distribute cruis-
ing taxis among the zones, while the greedy policy
sends too many cruising taxis to high trip-frequency
zones. This is confirmed when we look at the number
of trips vs. cruising taxis in both high and low trip-
frequency zone (see Figure 7 and 8). The new policies
are able to maintain balance between fulfilling trips
requirement and managing cruising time.
6 Conclusion
We presented in this paper an interesting and useful
application of Markov chains to manage an urban taxi
service. The basic premise is the need to minimize
cruising time (and therefore maximize utilization) of
the taxis. There are several assumptions we made for
this to work. We assume that each taxi has a device
that guides the taxi driver. Using MDPs the system
then generates a policy that optimizes cruising among
all taxis. We showed, with the use of real data from a
taxi operator in Singapore for the study, a cooperative
model where taxis follow instructions and a noncoop-
erative one where drivers compete with one another.
Our study also assumes rational drivers that try to
maximize the time their cars are occupied. For future
works, we would need to incorporate real-world behav-
ior of the passengers as well as relax the independence
assumption of passengers and their destinations.
References
[1] Michael Bowling.
Convergence problems of
general-sum multiagent reinforcement learning.
In Proceedings of the 17st International Confer-
ence on Machine Learning, pages 89–94, 2000.
[2] Robert D. Cairns and Catherine Liston-Heyes.
Competition and regulation in the taxi industry.
Journal of Public Economics, 59(1):1–15, 1996.
[3] Xi Chen and Xiaotie Deng. Settling the complex-
ity of two-player nash equilibrium. In Proceedings
of the 47th Symposium on Foundations of Com-
puter Science, pages 261–272, 2006.
[4] Sam Ganzfried and Tuomas Sandholm. Com-
puting Nash equilibria in multiplayer stochastic
games of imperfect information.
In Proceedings
of the 21st International Joint Conference on Ar-
ti(cid:12)cial Intel ligence, pages 140–146, 2009.
[5] Sam Ganzfried and Tuomas Sandholm. Game
opponent modeling
theory-based
large
in
imperfect-information games.
In Proceedings
of
the 10th International Conference on Au-
tonomous Agents and Multiagent Systems, pages
533–540, 2011.
[6] Paul W. Goldberg, Christos H. Papadimitriou,
and Rahul Savani. The complexity of the homo-
topy method, equilibrium selection, and Lemke-
Howson solutions.
In Proceedings of the 52nd
Symposium on Foundations of Computer Science,
pages 67–76, 2011.
[7] P. Jean-Jacques Herings and Ronald Peeters. Ho-
motopy methods to compute equilibria in game
theory. Economic Theory, 42(1):119–156, 2009.
[8] Junling Hu and Michael P. Wellman. Nash Q-
learning for general-sum stochastic games. Jour-
nal of Machine Learning Research, 4:1039–1069,
2003.
[9] Leslie P. Kaelbling, Michael L. Littman, and An-
thony R. Cassandra. Planning and acting in par-
tially observable stochastic domains. Arti(cid:12)cial In-
tel ligence, 101:99–134, 1998.
[10] Michael Kearns, Yishay Mansour, and Satinder
Singh. Fast planning in stochastic games. In Pro-
ceedings of the 16th Conference on Uncertainty in
Arti(cid:12)cial Intel ligence, pages 309–316, 2000.
[11] Michael L. Littman. Markov games as a frame-
work for multi-agent reinforcement learning.
In
Proceedings of the 11th International Conference
on Machine Learning, pages 157–163, 1994.
[12] Ryan Porter, Eugene Nudelman, and Yoav
Shoham. Simple search methods for finding a
Nash equilibrium.
In Proceedings of the 19th
AAAI Conference on Arti(cid:12)cial Intel ligence, pages
664–669, 2004.
[13] Tuomas Sandholm, Andrew Gilpin, and Vincent
Conitzer. Mixed-integer programming methods
for finding nash equilibria. In Proceedings of the
20th AAAI Conference on Arti(cid:12)cial Intel ligence,
pages 495–501, 2005.
[14] Lloyd S. Shapley. Stochastic games. Proceedings
of National Academy of Science, 39:1095–1100,
1953.
[15] S. T. Tan, L. F. Lin, A. L. Chan, C. M. Tan,
and D. H. Tan. Chapter 3 : Land transportation.
In Economics in Public Policy - The Singapore
Story. Marshall Cavendish Educ., 2009.
[16] K. I. Wong, S. C. Wong, M. G. H. Bell, and Hai
Yang. Modeling the bilateral micro-searching be-
havior for urban taxi services using the absorbing
Markov chain approach. Advanced Transporta-
tion, 39:81–104, 2005.
[17] Hai Yang, S. C. Wong, and K. I. Wong. Demand-
supply equilibrium of taxi services in a network
under competition and regulation. Transporta-
tion Research Part B: Methodological, 36:799–819,
2002.
|
1901.11454 | 1 | 1901 | 2019-01-31T16:26:48 | Efficient Ridesharing Order Dispatching with Mean Field Multi-Agent Reinforcement Learning | [
"cs.MA"
] | A fundamental question in any peer-to-peer ridesharing system is how to, both effectively and efficiently, dispatch user's ride requests to the right driver in real time. Traditional rule-based solutions usually work on a simplified problem setting, which requires a sophisticated hand-crafted weight design for either centralized authority control or decentralized multi-agent scheduling systems. Although recent approaches have used reinforcement learning to provide centralized combinatorial optimization algorithms with informative weight values, their single-agent setting can hardly model the complex interactions between drivers and orders. In this paper, we address the order dispatching problem using multi-agent reinforcement learning (MARL), which follows the distributed nature of the peer-to-peer ridesharing problem and possesses the ability to capture the stochastic demand-supply dynamics in large-scale ridesharing scenarios. Being more reliable than centralized approaches, our proposed MARL solutions could also support fully distributed execution through recent advances in the Internet of Vehicles (IoV) and the Vehicle-to-Network (V2N). Furthermore, we adopt the mean field approximation to simplify the local interactions by taking an average action among neighborhoods. The mean field approximation is capable of globally capturing dynamic demand-supply variations by propagating many local interactions between agents and the environment. Our extensive experiments have shown the significant improvements of MARL order dispatching algorithms over several strong baselines on the gross merchandise volume (GMV), and order response rate measures. Besides, the simulated experiments with real data have also justified that our solution can alleviate the supply-demand gap during the rush hours, thus possessing the capability of reducing traffic congestion. | cs.MA | cs | Efficient Ridesharing Order Dispatching with
Mean Field Multi-Agent Reinforcement Learning
Minne Li1, Zhiwei (Tony) Qin2, Yan Jiao2, Yaodong Yang1, Zhichen Gong1, Jun Wang1,
Chenxi Wang3, Guobin Wu3, Jieping Ye3
1University College London, 2DiDi Research America, 3DiDi Research
9
1
0
2
n
a
J
1
3
]
A
M
.
s
c
[
1
v
4
5
4
1
1
.
1
0
9
1
:
v
i
X
r
a
ABSTRACT
A fundamental question in any peer-to-peer ridesharing system is
how to, both effectively and efficiently, dispatch user's ride requests
to the right driver in real time. Traditional rule-based solutions
usually work on a simplified problem setting, which requires a
sophisticated hand-crafted weight design for either centralized au-
thority control or decentralized multi-agent scheduling systems.
Although recent approaches have used reinforcement learning to
provide centralized combinatorial optimization algorithms with
informative weight values, their single-agent setting can hardly
model the complex interactions between drivers and orders. In this
paper, we address the order dispatching problem using multi-agent
reinforcement learning (MARL), which follows the distributed na-
ture of the peer-to-peer ridesharing problem and possesses the
ability to capture the stochastic demand-supply dynamics in large-
scale ridesharing scenarios. Being more reliable than centralized
approaches, our proposed MARL solutions could also support fully
distributed execution through recent advances in the Internet of
Vehicles (IoV) and the Vehicle-to-Network (V2N). Furthermore, we
adopt the mean field approximation to simplify the local inter-
actions by taking an average action among neighborhoods. The
mean field approximation is capable of globally capturing dynamic
demand-supply variations by propagating many local interactions
between agents and the environment. Our extensive experiments
have shown the significant improvements of MARL order dispatch-
ing algorithms over several strong baselines on the gross merchan-
dise volume (GMV), and order response rate measures. Besides,
the simulated experiments with real data have also justified that
our solution can alleviate the supply-demand gap during the rush
hours, thus possessing the capability of reducing traffic congestion.
CCS CONCEPTS
• Computing methodologies → Multi-agent reinforcement
learning; • Applied computing → Transportation;
KEYWORDS
Multi-Agent Reinforcement Learning, Mean Field Reinforcement
Learning, Order Dispatching
ACM Reference Format:
Minne Li1, Zhiwei (Tony) Qin2, Yan Jiao2, Yaodong Yang1, Zhichen Gong1,
Jun Wang1, Chenxi Wang3, Guobin Wu3, Jieping Ye3. 2019. Efficient Rideshar-
ing Order Dispatching with Mean Field Multi-Agent Reinforcement Learn-
ing. In Proceedings of ACM WWW conference (The Web Conference 2019).
ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn
The Web Conference 2019, May 2019, San Francisco, USA
2019. ACM ISBN 978-x-xxxx-xxxx-x/YY/MM...$15.00
https://doi.org/10.1145/nnnnnnn.nnnnnnn
1 INTRODUCTION
Real-time ridesharing refers to the task of helping to arrange one-
time shared rides on very short notice [4, 10]. Such technique,
embedded in popular platforms including Uber, Lyft, and DiDi
Chuxing, has greatly transformed the way people travel nowadays.
By exploiting the data of individual trajectories in both space and
time dimensions, it offers more efficiency on traffic management,
and the traffic congestion can be further alleviated as well [18].
One of the critical problems in large-scale real-time ridesharing
systems is how to dispatch orders, i.e., to assign orders to a set
of active drivers on a real-time basis. Since the quality of order
dispatching will directly affect the utility of transportation capacity,
the amount of service income, and the level of customer satisfaction,
therefore, solving the problem of order dispatching is the key to any
successful ride-sharing platform. In this paper, our goal is to develop
an intelligent decision system to maximize the gross merchandise
volume (GMV), i.e., the value of all the orders served in a single
day, with the ability to scale up to a large number of drivers and
robust to potential hardware or connectivity failures.
The challenge of order dispatching is to find an optimal trade-off
between the short-term and long-term rewards. When the number
of available orders is larger than that of active drivers within the or-
der broadcasting area (shown as the grey shadow area in the center
of Fig. 1b), the problem turns into finding an optimal order choice
for each driver. Taking an order with a higher price will contribute
to the immediate income; however, it might also harm the GMV in
the long run if this order takes the driver to a sparsely populated
area. As illustrated in Fig. 1a, considering the two orders starting
from the same area but to two different destinations, a neighbor-
ing central business district (CBD) and a distant suburb. A driver
taking the latter one may have a higher one-off order price due to
the longer travel distance, but the subsequent suburban area with
little demand could also prevent the driver from further sustaining
income. Dispatching too many such orders will, therefore, reduce
the number of orders taken and harm the long-term GMV. The
problem becomes more serious particularly during the peak hours
when the situation in places with the imbalance between vehicle
supply and order demand gets worse. As such, an intelligent order
dispatching system should be designed to not only assign orders
with high prices to the drivers, but also to anticipate the future
demand-supply gap and distribute the imbalance among different
destinations. In the meantime, the pick-up distance should also be
minimized, as the drivers will not get paid during the pick-up pro-
cess; on the other hand, long waiting time will affect the customer
experience.
One direction to tackle the order dispatching challenge has been
to apply hand-crafted features to either centralized dispatching
authorities (e.g., the combinatorial optimization algorithm [28, 44])
The Web Conference 2019, May 2019, San Francisco, USA
M. Li et al.
(a)
(b)
Figure 1: The order dispatching problem. (a) Two order choices (black triangle, departing from orii,1 and orii,2 respectively)
within a driver i's (black dot, located at loci) order receiving area (grey shadow), where one ends at a neighboring CBD (located
at desi,1) with high demand, and the other ends at a distant suburb (located at desi,2) with low demand. Both orders have the
same pick-up distance for driver i. (b) Details within a neighborhood, where the radius of the order receiving area and the
neighborhood (black circle) are r and 2r respectively.
or distributed multi-agent scheduling systems [3, 39], in which a
group of autonomous agents that share a common environment
interact with each other. However, the system performance relies
highly on the specially designed weighting scheme. For centralized
approaches, another critical issue is the potential "single point of
failure" [23], i.e., the failure of the centralized authority control
will fail the whole system. Although the multi-agent formulation
provides a distributed perspective by allowing each driver to choose
their order preference independently, existing solutions require
rounds of direct communications between agents during execution
[30], thus being limited to a local area with a small number of
agents.
Recent attempts have been made to formulate this problem with
centralized authority control and model-free reinforcement learn-
ing (RL) [33], which learns a policy by interacting with a complex
environment. However, existing approaches [40] formulate the or-
der dispatching problem with the single-agent setting, which is un-
able to model the complex interactions between drivers and orders,
thus being oversimplifications of the stochastic demand-supply
dynamics in large-scale ridesharing scenarios. Also, executing an
order dispatching system in a centralized manner still suffers from
the reliability issue mentioned above that is generally inherent to
the centralized architecture.
Staying different from these approaches, in this work we model
the order dispatching problem with multi-agent reinforcement
learning (MARL) [7], where agents share a centralized judge (the
critic) to rate their decisions (actions) and update their strategies
(policies). The centralized critic is no longer needed during the
execution period as agents can follow their learned policies inde-
pendently, making the order dispatching system more robust to
potential hardware or connectivity failures. With the recent devel-
opment of the Internet of Things (IoT) [5, 11, 43] and the Internet
of Vehicles (IoV) [22, 41], the fully distributed execution could be
practically deployed by distributing the centralized trained policy to
each vehicle through the Vehicle-to-Network (V2N) [1]. By allowing
each driver to learn by maximizing its cumulative reward through
time, the reinforcement learning based approach relieves us from
designing a sophisticated weighting scheme for the matching algo-
rithms. Also, the multi-agent setting follows the distributed nature
of the peer-to-peer ridesharing problem, providing the dispatching
system with the ability to capture the stochastic demand-supply dy-
namics in large-scale ridesharing scenarios. Meanwhile, such fully
distributed executions also enable us to scale up to much larger
scenarios with many more agents, i.e., a scalable real-time order
dispatching system for ridesharing services with millions of drivers.
Nonetheless, the major challenge in applying MARL to order
dispatching lies in the changing dynamics of two components: the
size of the action set and the size of the population. As illustrated in
Fig. 1b, the action set for each agent is defined as a set of neighboring
active orders within a given radius r from each active driver (shown
as the shadow area around each black dot). As the active orders will
be taken and new orders keep arriving, the size and content of this
set will constantly change over time. The action set will also change
when the agent moves to another location and arrives in a new
neighborhood. On the other hand, drivers can also switch between
online and offline in the real-world scenario, the population size
for the order dispatching task is therefore also changing over time.
In this paper, we address the two problems of variable action sets
and population size by extending the actor-critic policy gradient
methods. Our methods tackle the order dispatching problem within
the framework of centralized training with decentralized execution.
The critic is provided with the information from other agents to
incorporate the peer information, while each actor behaves inde-
pendently with local information only. To resolve the variable popu-
lation size, we adopt the mean field approximation to transform the
interactions between agents to the pairwise interaction between
an agent and the average response from a sub-population in the
neighborhood. We provide the convergence proof of mean field
reinforcement learning algorithms with function approximations
to justify our algorithm in theory. To solve the issue of changing ac-
tion set, we use the vectorized features of each order as the network
input to generate a set of ranking values, which are fed into a Boltz-
mann softmax selector to choose an action. Experiments on the
large-scale simulator show that compared with variant multi-agent
learning benchmarks, the mean field multi-agent reinforcement
learning algorithm gives the best performance towards the GMV,
the order response rate, and the average pick-up distance. Besides
the state of the art performance, our solution also enjoys the ad-
vantage of distributed execution, which has lower latency and is
easily adaptable to be deployed in the real world application.
Motivation•Goal: maximize GMV
•# Order > # Driver: find an optimal order choice for each driver
•Scenario:order fee & potentialDESi,1 = [DXi,1, DYi,1]DESi,2 = [DXi,2, DYi,2]o1locidesi,1desi,2orii,1orii,2r2r Efficient Ridesharing Order Dispatching with MFRL
The Web Conference 2019, May 2019, San Francisco, USA
2 METHOD
In this section, we first illustrate our definition of order dispatch-
ing as a Markov game, and discuss two challenges when applying
MARL to this game. We then propose a MARL approach with the
independent Q-learning, namely the independent order dispatch-
ing algorithm (IOD), to solve this game. By extending IOD with
mean field approximations, which capture dynamic demand-supply
variations by propagating many local interactions between agents
and the environment, we finally propose the cooperative order
dispatching algorithm (COD).
2.1 Order Dispatching as a Markov Game
2.1.1 Game Settings. We model the order dispatching task by a
Partially Observable Markov Decision Process (POMDP) [20] in a
fully cooperative setting, defined by a tuple Γ = ⟨S, P, A, R, O, N , γ⟩,
where S, P, A, R, O, N , γ are the sets of states, transition probabil-
ity functions, sets of joint actions, reward functions, sets of private
observations, number of agents, and a discount factor respectively.
Given two sets X and Y, we use X × Y to denote the Cartesian
product of X and Y, i.e., X × Y = {(x, y)x ∈ X, y ∈ Y}. The
definitions are given as follows:
• N : N homogeneous agents identified by i ∈ I ≡ {1, ..., N} are
defined as the active drivers in the environment. As the drivers
can switch between online and offline via a random process, the
number of agents N could change over time.
• S, O: At each time step t, agent i draws private observations ot
i ∈
O correlated with the true environment state st ∈ S according
to the observation function S × I → O. The initial state of
the environment is determined by a distribution p1(s
1) : S →
[0, 1]. A typical environmental state s includes the order and
driver distribution, the global timestamp, and other environment
dynamics (traffic congestion, weather conditions, etc.). In this
work, we define the observation for each agent i with three
components: agent i's location loci, the timestamp t, and an
on-trip flag to show if agent i is available to take new orders.
• P, A: In the order dispatching task, each driver can only take
active orders within its neighborhood (inside a given radius, as
illustrated in Fig. 1b). Hence, agent i's action set Ai is defined
as its own active order pool based on the observation oi. Each
action candidate ai,m ∈ Ai is parameterized by the normalized
vector representation of corresponding order's origin orii,m and
destination desi,m, i.e., ai,m ≡ [orii,m, desi,m]. At time step t,
i ∈ Ai, forming a set of joint driver-
each agent takes an action at
order pair at = A1 × ... × AN , which induces a transition in
the environment according to the state transition function
P(st +1st , at) : S × A1 × ... × AN → S.
(1)
For simplicity, we assume no order cancellations and changes
during the trip, i.e., at
keeps unchanged if agent i is on its way
i
to the destination.
• R: Each agent i obtains rewards r t
i
by a reward function
Ri(st , at) : S × A1 × ... × AN → R.
(2)
As described in Section 1, we want to maximize the total income
by considering both the charge of each order and the potential
opportunity of the destination. The reward is then defined as the
combination of driver i's own income 0
r from its order choice
and the order destination potential 1
r , which is determined
by all agents' behaviors in the environment. Considering the
credit-assignment problem [2] arises in MARL with many agents,
i.e., the contribution of an agent's behavior is drowned by the
noise of all the other agents' impact on the reward function, we
set each driver's own income instead of the total income of all
drivers as the reward.
To encourage cooperation between agents and avoid agents'
being selfish and greedy, we use the order destination's demand-
supply gap as a constraint 1
r on the behavior of agents. Precisely,
we compare the demand-supply status between the order ori-
gin and destination, and encourage the driver to choose the
order destination with a larger demand-supply gap. The order
destination potential (DP) is defined as
DP = #DD − #DS,
(3)
where #DD and #DS is the demand and the supply of the desti-
nation respectively. We consider the DP only if the number of
orders are larger than that of drivers at the origin. If the order
destination has more drivers than orders, we penalize this order
with the demand-supply gap at the destination, and vice versa.
To provide better customer experience, we also add the pick-up
distance 2
r as a regularizer. The ratio of the DP and the pick-up
distance to order price are defined as 1
α respectively, i.e.,
+1
. We typically choose the regularization
r t
i
ratio to scale different reward terms into approximately the same
range, although in practice a grid-search could be used to get
better performance. The effectiveness of our reward function
setting is empirically verified in Section 3.1.3.
• γ: Each agent i aims to maximize its total discounted reward
(4)
from time step t onwards, where γ ∈ [0, 1] is the discount factor.
We denote joint quantities over agents in bold, and joint quanti-
ties over agents other than a given agent i with the subscript −i,
e.g., a ≡ (at−i , at
i ). To stabilize the training process, we maintain an
experience replay buffer D containing tuples (ot , at , rt , ot +1) as
described in [26].
i =∞
γ k−t r k
i
α and 2
+2
α
= 0
r t
i
Gt
1
r t
i
α
2
r t
i
k =t
2.1.2 Dynamic of the Action Set Elements. In the order dispatch-
ing problem, an action is defined as an active order within a given
radius from the agent. Hence, the content and size of the active
order pool for each driver are changing with both the location and
time, i.e., the action set for an agent in the order dispatching MDP
is changing throughout the training and execution process. This
aspect of order dispatching refrains us from using the Q-table to log
the Q-value because of the potentially infinite size of the action set.
On the other hand, a typical policy network for stochastic actions
makes use of a softmax output layer to produce a set of probabilities
of choosing each action among a fixed set of action candidates, thus
is unable to fit into the order dispatching problem.
2.1.3 Dynamic of the Population Size. To overcome the non-
stationarity of the multi-agent environment, Lowe et al. [21] uses Q-
Learning to approximate the discounted reward Gt
, and rewrites the
i
gradient of the expected return for agent i following a deterministic
The Web Conference 2019, May 2019, San Francisco, USA
M. Li et al.
(a)
(b)
(c)
Figure 2: Overview of our COD approach. (a) The information flow between RL agents and the environment (shown with two
agents). The centralized authority is only needed during the training stage (dashed arrows) to gather the average response ¯a;
agents could behave independently during the execution stage (solid arrows), thus being robust to the "single point of failure".
(b) and (c) Architectures of the critic and the actor.
policy µi (parameterized by θi) as
∇θi J(θi) = Eo,a∼D[∇θi µi(aioi)∇ai Qi(oi , a)].
(5)
Here Qi(oi , a) is a centralized action-value function that takes as
input the observation oi of agent i and the joint actions a of all
agents, and outputs the Q-value for agent i. However, as off-line
drivers cannot participate in the order dispatching procedure, the
number of agents N in the environment is changing over time. Also,
in a typical order dispatching task, which involves thousands of
agents, the high dynamics of interactions between a large number
of agents is intractable. Thus, a naive concatenation of all other
agents' actions cannot form a valid input for the value network and
is not applicable to the order dispatching task.
i
J(θi) = Eo,a∼D(cid:2)Gt
(cid:3) by taking steps in the direction of ∇θi J(θi).
2.2 Independent Order Dispatching
To solve the order dispatching MDP, we first propose the indepen-
dent order dispatching algorithm (IOD), a straightforward MARL
approach with the independent Q-learning. We provide each learner
with the actor-critic model, which is a popular form of policy gra-
dient (PG) method. For each agent i, PG works by directly adjust-
ing the parameters θi of the policy µi to maximize the objective
In MARL, independent actor-critic uses the action-value function
Qi(oi , ai) to approximate the discounted reward Gt
by Q-Learning
i
[38]. Here we use temporal-difference learning [32] to approximate
the true Qi(oi , ai), leading to a variety of actor-critic algorithms
with Qi(oi , ai) called as the critic and µi called as the actor.
To solve the problem of variable action sets, we use a policy
network with both observation and action embeddings as the input,
derived from the in-action approximation methods (shown in Fig.
2b). As illustrated in Fig. 2c, we use a deterministic policy µθi
: O × A → R∗, abbreviated as µi) to generate
(denoted by µθi
ranking values of each observation-action pair (oi , ai,m) for each
of the Mi candidates within agent i's action set Ai. To choose
an action ai, these values are then fed into a Boltzmann softmax
selector
πi(ai, joi) =
Mi
exp(βµi(oi , ai, j))
m=0 exp(βµi(oi , ai,m)) ,
where β is the temperature to control the exploration rate. Note
that a typical policy network with out-action approximation is
equivalent to this approach, where we can use an Mi-dimension
one-hot vector as the embedding to feed into the policy network.
The main difference between these approaches is the execution effi-
ciency, as we need exactly Mi forward passes in a single execution
step. Meanwhile, using order features naturally provide us with an
informative form of embeddings. As each order is parameterized
by the concatenation of the normalized vector representation of
its origin orii,m and destination desi,m, i.e., ai,m ≡ [orii,m, desi,m].
Similar orders will be close to each other in the vector space and
produce similar outputs from the policy network, which improves
the generalization ability of the algorithm.
In IOD, each critic takes input the observation embedding oi
by combining agent i's location and the timestamp t. The action
embedding is built with the vector representation of the order des-
tination and the distance between the driver location and the order
origin. The critic is a DQN [26] using neural network function ap-
proximations to learn the action-value function Qϕ (parameterized
by ϕ, abbreviated as Qi for each agent i) by minimizing the loss
′
′
L(ϕi) = Eo,a,r,o′[(y − Qi(oi , ai))2], y = ri + γ Q
i)),
i , a
(7)
where Q−
is the target network for the action-value function, and
µ−
i
is the target network for the deterministic policy. These earlier
i
snapshots of parameters are periodically updated with the most
recent network weights and help increase learning stability by
decorrelating predicted and target Q-values and deterministic pol-
icy values.
−
i (o
Following Silver et al. [31], the gradient of the expected return
′
i , µ
−
i (o
for agent i following a deterministic policy µi is
∇θi J(θi) = Eo,a∼D[∇θi µi(oi , ai)∇ai Qi(oi , ai)].
(8)
Here Qi(oi , ai) is an action-value function that takes as input the
observation oi and action ai, and outputs the Q-value for agent i.
2.3 Cooperative Order Dispatching with Mean
Field Approximation
for j = 1, . . . , Mi
(6)
To fully condition on other agents' policy in the environment with
variable population size, we propose to integrate our IOD algorithm
Actor 2o2a2o2Critic 2r2a2a1EnvironmentActor 1o1r1Critic 1o1a1a1a2q1q2oiaiqioi,ai,ai()φioiai,1πioi,ai,1()oiai,Mπioi,ai,M()……πioi,ai()Selectorθiθiq1a1EnvironmentActor 1Actor No1aNoNr1Critic 1o1rNCritic NoNqN……a1aNaioiaiqioi,ai()ϕioiai,1µioi,ai,1()oiai,Mµioi,ai,M()……aiSelectorθiθiq1a1EnvironmentActor 1Actor No1aNoNr1Critic 1o1rNCritic NoNqN……a1aN Efficient Ridesharing Order Dispatching with MFRL
The Web Conference 2019, May 2019, San Francisco, USA
Algorithm 1 Cooperative Order Dispatching (COD)
, µθi
, Q−
ϕi
for all i ∈ {1, . . . , N}
, and µ−
Initialize Qϕi
θi
while training not finished do
For each agent i, sample action ai using the Boltzmann soft-
max selector πi(aioi) from Eq. (6)
Take the joint action a = [a1, . . . , aN ] and observe the re-
ward r = [r1, . . . , rN ] and the next observations o′
Compute the new mean action ¯a = [¯a1, . . . , ¯aN ]
Store ⟨o, a, r, o′, ¯a⟩ in replay buffer D
for i = 1 to N do
Sample K experiences ⟨o, a, r, o′, ¯a⟩ from D
Update the critic by minimizing the loss from Eq. (9)
Update the actor using the policy gradient as Eq. (11)
end for
Update the parameters of the target networks for each agent
i with updating rates τϕ and τθ :
ϕ
−
−
i ← τϕϕi + (1 − τϕ)ϕ
i
−
−
i ← τθ θi + (1 − τθ)θ
i
θ
end while
with mean field approximations, following the Mean Field Rein-
forcement Learning (MFRL) [42]. MFRL addresses the scalability
issue in the multi-agent reinforcement learning with a large num-
ber of agents, where the interactions are approximated pairwise by
the interaction between an agent and the average response ¯a from
a sub-population in the neighborhood. As this pairwise approxima-
tion shadows the exact size of interacting counterparts, the use of
mean field approximation can help us model other agents' policies
directly in the environment with variable population sizes.
In the order dispatching task, agents are interacting with each
other by choosing order destinations with a high demand to opti-
mize the demand-supply gap. As illustrated in Fig. 1b, the range of
the neighborhood is then defined as twice the length of the order
receiving radius, because agents within this area have intersections
between their action sets and interact with each other. The average
response ¯ai is therefore defined as the number of drivers arriving
at the same neighborhood as agent i, divided by the number of
available orders for agent i. For example, when agent i finishes the
order and arrives at the neighborhood in Fig. 1b (the central agent),
the average response ¯ai is 2/3, as there are two agents within the
neighborhood area and three available orders.
The introduction of mean field approximations enables agents to
learn with the awareness of interacting counterparts, thus helping
to improve the training stability and robustness of agents after train-
ing for the order dispatching task. Note that the average response ¯a
only serves for the model update; thus the centralized authority is
only needed during the training stage. During the execution stage,
agents could behave in a fully distributed manner, thus being robust
to the "single point of failure".
We propose the cooperative order dispatching algorithm (COD)
as illustrated in Fig. 2a, and present the pseudo code for COD in
Algorithm 1. Each critic is trained by minimizing the loss
′
i) − Qi(oi ,(¯ai , ai)))2],
L(ϕi) = Eo,a,r,o′[(ri + γ vMF
i− (o
(9)
π
a′
i
[Q
i(a′
′
io
−
i (a
′
i)E ¯a′
−i)∼π−
−i
i) =
i− (o′
where vMF
and µ−
Q−
i
i
′
i− (o
vMF
i) is the mean field value function for the target network
(shown as the Boltzmann selector π−
from Eq. 6)
i
′
−
′
′
i (o
i ,(¯a
i))],
i , a
∇θi J(θi) = Eo,a∼D[∇θi µi(oi , ai)∇ai Qi(oi ,(¯ai , ai))].
(10)
and ¯ai is the average response within agent i's neighborhood. The
actor of COD learns the optimal policy by using the policy gradient:
(11)
In the current decision process, active agents sharing duplicated
order receiving areas (e.g., the central and upper agent in Fig. 1b)
might select the same order following their own strategy. Such
collisions could lead to invalid order assignment and force both dri-
vers and customers to wait for a certain period, which equals to the
time interval between each dispatching iteration. Observe that the
time interval of decision-making also influences the performance
of dispatching; a too long interval will affect the passenger experi-
ence, while a too short interval without enough order candidates
is not conducive to the decision making. To solve this problem,
our approach works in a fully distributed manner with asynchro-
nous dispatching strategy, allowing agents have different decision
time interval for individual states, i.e., agents assigned with invalid
orders could immediately re-choose a new order from updated
candidates pool.
To theoretically support the efficacy of our proposed COD algo-
rithm, we provide the convergence proof of MFRL with function
approximations as shown below.
2.4 Convergence of Mean Field Reinforcement
Learning with Function Approximations
Inspired by the previous proof of MFRL convergence in a tabular
Q-function setting [42], we further develop the proof towards the
converge when the Q-function is represented by other function
approximators. In addition to the Markov Game setting in Section
2.1.1, let Q = {Qϕ} be a family of real-valued functions defined on
S × A × ¯A, where ¯A is the action space for the mean actions com-
puted from the neighbors. For simplicity, we assume the environ-
ment is a fully observable MDP Γ, i.e., each agent i ∈ I ≡ {1, ..., N}
can observe the global state s instead of the local observation oi.
Assuming that the function class is linearly parameterized, for
each agent i, the Q-function can be expressed as the linear span of a
: S × A × ¯A → R.
fixed set of P linearly independent functions ω
Given the parameter vector ϕi ∈ RP , the function Qϕi
(abbreviated
as Qi) is thus defined as
Qi(s,(¯ai , ai)) =
i (s,(¯ai , ai))ϕi(p) = ωi(s,(¯ai , ai))⊤
P
(12)
ϕi .
p
i
ω
p
p=1
In the function approximation setting, we apply the update rules:
(13)
′
i = ϕi + α∆∇ϕi Qi(s,(¯ai , ai)) = ϕi + α∆ωi(s,(¯ai , ai)),
ϕ
where ∆ is the temporal difference:
i (s
∆ = ri + γ vMF
a∼πMF[Qi(s
= ri + γ E
′
′) − Qi(s,(¯ai , ai))
(14)
Our goal is to derive the parameter vector φ = {ϕi} such that
ω⊤φ approximates the (local) Nash Q-values. Under the main as-
sumptions and the lemma as introduced below, Yang et al. [42]
,(¯ai , ai))] − Qi(s,(¯ai , ai)).
The Web Conference 2019, May 2019, San Francisco, USA
M. Li et al.
proved that the policy π is K2T
Lipschitz continuous with respect to
ϕ, where T = 1/β and K ≥ 0 is the upper bound of the observed
reward.
Assumption 1. Each state-action pair is visited infinitely often,
and the reward is bounded by some constant K.
Assumption 2. Agent's policy is Greedy in the Limit with Infinite
Exploration (GLIE). In the case with the Boltzmann policy, the policy
becomes greedy w.r.t. the Q-function in the limit as the temperature
decays asymptotically to zero.
t (s), ..., Q N
1
Assumption 3. For each stage game [Q
t (s)] at time
t and in state s in training, for all t, s, j ∈ {1, . . . , N}, the Nash
equilibrium π∗ = [π
1∗ , . . . , π N∗ ] is recognized either as 1) the global
optimum or 2) a saddle point expressed as:
t(s)] ≥ Eπ[Q
1. Eπ∗[Q
2. Eπ∗[Q
t(s)] ≥ Eπ j E
t(s)] ≤ E
Eπ∗[Q
π j∗
Lemma 2.1. The random process {∆t } defined in R as
∆t +1(x) = (1 − αt(x))∆t(x) + αt(x)Ft(x)
t(s)], ∀π ∈ Ω(cid:0)
k Ak(cid:1);
t(s)], ∀π j ∈ Ω(cid:0)Aj(cid:1) and
t(s)], ∀π−j ∈ Ω(cid:0)
k(cid:44)j Ak(cid:1).
[Q
−j∗
π−j[Q
π
E
j
j
j
j
j
j
t α
converges to zero with probability 1 (w.p.1) when
t (x) < ∞;
2
2. x ∈ X, the set of possible states, and X < ∞;
3. ∥E[Ft(x)Ft]∥W ≤ γ ∥∆t ∥W + ct , where γ ∈ [0, 1) and ct con-
4. var[Ft(x)Ft] ≤ K(1 + ∥∆t ∥2
Here Ft denotes the filtration of an increasing sequence of σ-fields
including the history of processes; αt , ∆t , Ft ∈ Ft and ∥ · ∥W is a
weighted maximum norm [6].
W ) with constant K > 0.
verges to zero w.p.1;
1. 0 ≤ αt(x) ≤ 1,
t αt(x) = ∞,
Proof. See Theorem 1 in Jaakkola et al. [13] and Corollary 5 in
Szepesvári and Littman [35] for detailed derivation. We include it
here to stay self-contained.
□
In contrast to the previous work Yang et al. [42], we establish
convergence of Eq. (13) by adopting an ordinary differentiable equa-
tion (ODE) with a globally asymptotically stable equilibrium point
where the trajectories closely follow, following the framework of
the convergence proof of single-agent Q-learning with function
approximation [25].
Theorem 2.2. Given the MDP Γ, {ωp , p = 1, ..., P}, and the learn-
ing policy π that is K2T Lipschitz continuous with respect to ϕ, if the
Assumptions 1, 2 & 3, and Lemma 2.1's first and second conditions are
met, then there exists C0 such that the algorithm in Eq. (13) converges
w.p.1 if K2T < C0.
Proof. We first re-write the Eq. (13) as on ODE:
dφ
dt
(cid:0)r(s, a, s
(cid:2)ω⊤
(cid:2)ω⊤
s
= Eφ
= Eφ
= Aϕ φ + bϕ
s (γ ω⊤
s′ − ω⊤
s φ(cid:1)(cid:3)
(cid:2)ω⊤
s′φ − ω⊤
′) + γ ω⊤
s )(cid:3) φ + Eφ
s (r(s, a, s
′))(cid:3)
(15)
Notice that we use a vector for considering the updating rule for
the Q-function of each agent. We can easily know that necessity
condition of the equilibrium is φ∗ = A−1
ϕ∗ bϕ∗. The existence of the
such equilibrium has been restricted in the scenario that meets
Assumption 3. Yang et al. [42] proved that under the Assumption 3,
the existing equilibrium, either in the form of a global equilibrium
or in the form of a saddle-point equilibrium, is unique.
Let φ = φt − φ∗, we have:
− 2φ∗ · dφ
φ2 = 2φ · dφ
dt
dt
d
dt
φ
(cid:33)
(cid:32)
Aϕ∗ + sup
= 2 φ⊤(Aϕ φ + bϕ − Aϕ∗ φ∗ − bϕ∗)
= 2 φ⊤(Aϕ∗ φ − Aϕ∗ φ + Aϕ φ + bϕ − Aϕ∗ φ∗ − bϕ∗)
= 2 φ⊤ Aϕ∗ φ + 2 φ⊤(Aϕ − Aϕ∗)φ + 2 φ⊤(bϕ − bϕ∗)
bϕ − b∗
ϕ2
≤ 2 φ⊤
φ − φ∗2
Aϕ − Aϕ∗2 + sup
φ.
(16)
As we know that the policy πt is Lipschitz continuous w.r.t ϕt , this
implies that Aϕ and bϕ are also Lipschitz continuous w.r.t to ϕ.
In other words, if K2T ≤ C0 is sufficiently small and close to zero,
bϕ − b∗
ϕ2
then the norm term of
φ − φ∗2
goes to zero. Considering near the equilibrium point ϕ∗, Aϕ∗ is a
negative definite matrix, the Eq. (16) tends to be negative definite
as well, so the ODE in Eq.(15) is globally asymptotically stable and
the conclusion of the theorem follows.
□
φ Aϕ − Aϕ∗2 + sup
sup
(cid:32)
(cid:33)
φ
φ
While in practice, we might break the linear condition by the
use of nonlinear activation functions, the Lipschitz continuity will
still hold as long as the nonlinear add-on is limited to a small scale.
3 EXPERIMENT
To support the training and evaluation of our MARL algorithm,
we adopt two simulators with a grid-based map and a coordinate-
based map respectively. The main difference between these two
simulators is the design of the driver pick-up module, i.e., the pro-
cess of a driver reaching the origin of the assigned order. In the
grid-based simulator (introduced by Lin et al. [19]), the location
state for each driver and order is represented by a grid ID. Hence,
the exact coordinate for each instance inside a grid is shadowed by
the state representation. This simplified setting ensures there will
be no pick-up distance (or arriving time) difference inside a grid; it
also brings an assumption of no cancellations before driver pick-up.
Whereas in the coordinate-based simulator, the location of each
driver and order instance is represented by a two-value vector from
the Geographic Coordinate System, and the cancellation before
pick-up is also taken into account. In this setting, taking an order
within an appropriate pick-up distance is crucial to each driver,
as the order may be canceled if the driver takes too long time to
arrive. We present experiment details of the grid-based simulator
in Section 3.1, and the coordinate-based simulator in Section 3.2.
3.1 Grid-based Experiment
3.1.1 Environment Setting. In the grid-based simulator, the city
is covered by a hexagonal grid-world as illustrated in Fig. 3. At each
simulation time step t, the simulator provides an observation ot
with a set of active drivers and a set of available orders. Each order
Efficient Ridesharing Order Dispatching with MFRL
The Web Conference 2019, May 2019, San Francisco, USA
Table 1: Performance comparison regarding the normalized
Gross Merchandise Volume (GMV) on the test set with re-
spect to the performance of RAN.
Figure 3: Illustration of the grid-based simulator. All loca-
tion features are represented by the corresponding grid ID,
where the order origin orii,m is the same as the driver loca-
tion loci.
feature includes the origin grid ID and the destination grid ID, while
each driver has the grid ID as the location feature loci. Drivers are
regarded as homogeneous and can switch between online (active)
and offline via a random process learned from the history data. As
the travel distance between neighboring grids is approximately 1.2
kilometers and the time step interval t△ is 10 minutes, we assume
that drivers will not move to other grids before taking a new order,
and define the order receiving area and the neighborhood as the
grid where the agent stays. The order dispatching algorithm then
generates an optimal list of driver-order pairs at for the current
policy, where ai is an available order ai,m selected from the order
candidate pool Ai. In the grid-based setting, the origin of each order
is already embedded as the location feature loci in oi, thus ai,m is
parameterized by the destination grid ID desi,m. After receiving
the driver-order pairs from the algorithm, the simulator will then
return a new observation ot +1 and a list of order fees. Stepping on
this new observation, the order dispatching algorithm will calculate
a set of rewards r t for each agent, store the record (ot , at , rt , ot +1)
to replay buffer, and update the network parameters with respect
to a batch of samples from replay buffer.
The data source of this simulator (provided by DiDi Chuxing) in-
cludes order information and trajectories of vehicles in three weeks.
Available orders are generated by bootstrapping from real orders
occurred in the same period during the day given a bootstrapping
ratio psample. More concretely, suppose the simulation time step
interval is t△, at each simulation time step t, we randomly sample
psample orders with replacement from real orders happened be-
tween t△ ∗ t to t△ ∗ (t + 1). Also, drivers are set between online
and offline following a distribution learned from real data using a
maximum likelihood estimation. On average, the simulator has 7K
drivers and 136K dispatching events per time step.
The effectiveness of the grid-based simulator is evaluated by Lin
et al. [19] using the calibration against the real data regarding the
most important performance measurement: the gross merchandise
2 between simu-
volume (GMV). The coefficient of determination r
lated GMV and real GMV is 0.9331 and the Pearson correlation is
0.9853 with p-value p < 0.00001.
psample
100%
50%
10%
−4.60 ± 0.01 −4.65 ± 0.04 −4.71 ± 0.03
RES
+4.78 ± 0.00 +4.87 ± 0.02 +5.24 ± 0.07
REV
+6.27 ± 0.01 +6.21 ± 0.03 +5.73 ± 0.07
IOD
COD +7.78 ± 0.01 +7.76 ± 0.03 +7.52 ± 0.14
• Response-based (RES): This response-based method aims to
achieve higher order response rate by assigning drivers to short
duration orders. During each time step, all available orders start-
ing from the same grid will be sorted by the estimated trip time.
Multiple orders with the same expected duration will be further
sorted by the order price to balance the performance.
• Revenue-based (REV): The revenue-based algorithm focuses
on a higher GMV. Orders with higher prices will be given pri-
ority to get dispatched first. Following the similar principle as
described above, orders with shorter estimated trip time will be
assigned first if multiple orders have the same price.
• IOD: The independent order dispatching algorithm as described
in Section 2.2. The action-value function approximation (i.e., the
Q-network) is parameterized by an MLP with four hidden layers
(512, 256, 128, 64) and the policy network is parameterized by an
MLP with three hidden layers (256, 128, 64). We use the ReLU [27]
activation between hidden layers, and transform the final linear
output of Q-network and policy network with ReLU and sigmoid
function respectively. To find an optimal parameter setting, we
use the Adam Optimizer [14] with a learning rate of 0.0001 for
the critic and 0.001 for the actor. The discounted factor γ is 0.95,
and the batch size is 2048. We update the network parameters
after every 3K samples are added to the replay buffer (capacity
5 ∗ 105). We use a Boltzmann softmax selector for all MARL
methods and set the initial temperature as 1.0, then gradually
reduce the temperature until 0.01 to limit exploration.
• COD: Our proposed cooperative order dispatching algorithm
with mean field approximation as described in Section 2.3. The
network architecture is identical to the one described in IOD,
except a mean action is fed as another input to the critic network
as illustrated in Fig. 2b.
As described in the reward setting in Section 2.1.1, we set the
regularization ratio 1
α = 0 for the order
waiting time penalty as we don't consider the pick-up distance
in the grid-based experiment. Because of our homogeneous agent
setting, all agents share the same Q-network and policy network
for efficient training. During the execution in the real-world envi-
ronment, each agent can keep its copy of policy parameters and
receive updates periodically from a parameter server.
α = 0.01 for DP and 2
3.1.2 Model Setting. We use the grid-based simulator to com-
pare the performance of following methods.
• Random (RAN): The random dispatching algorithm considers
no additional information. It only assigns all active drivers with
an available order at each time step.
3.1.3 Result Analysis. For all learning methods, we run 20 episodes
for training, store the trained model periodically, and conduct the
evaluation on the stored model with the best training performance.
The training set is generated by bootstrapping 50% of the original
real orders unless specified otherwise. We use five random seeds
locidesi,2desi,1 The Web Conference 2019, May 2019, San Francisco, USA
M. Li et al.
Table 2: Performance comparison regarding the order re-
sponse rate (OOR) on the test set. The percentage difference
shown for all methods is with respect to RAN.
Table 3: Performance comparison in terms of the average
destination potential (ADP) on the test set. The percentage
difference shown for all methods is with respect to RAN.
psample
100%
50%
10%
+3.15 ± 0.01 +3.11 ± 0.02 +2.59 ± 0.04
RES
−6.21 ± 0.00 −6.14 ± 0.03 −5.22 ± 0.06
REV
+4.67 ± 0.01 +4.68 ± 0.03 +4.53 ± 0.04
IOD
COD +5.35 ± 0.00 +5.38 ± 0.03 +5.42 ± 0.09
psample
100%
50%
10%
+6.36 ± 0.15 +7.40 ± 0.54 +9.97 ± 0.43
RES
−20.55 ± 0.16 −20.23 ± 0.52 −20.55 ± 1.58
REV
+54.33 ± 0.09 +54.71 ± 0.20 +53.33 ± 0.54
IOD
COD +66.74 ± 0.05 +66.90 ± 0.29 +64.69 ± 0.61
(a) COD
(b) REV
Figure 4: An example of the demand-supply gap in the
city center during peak hours. Grids with more drivers are
shown in green (in red if opposite) and the gap is propor-
tional to the shade of colors.
for testing and present the averaged result. We compare the per-
formance of different methods by three metrics, including the total
income in a day (GMV), the order response rate (ORR), and the
average order destination potential (ADP). ORR is the number of
orders taken divided by the number of orders generated, and ADP
is the sum of destination potential (as described in Section 2.1.1) of
all orders divided by the number of orders taken.
Gross Merchandise Volume. As shown in Table 1, the performance
of COD largely surpasses all rule-based methods and IOD in GMV
metric. RES suffers from lowest GMV among all methods due to its
preference of short distance trips with lower average order value.
On the other hand, REV aims to pick higher value orders with
longer trip time, thus enjoying a higher GMV. However, both RES
and REV cannot find a balance between getting higher income per
order and taking more orders, while RAN falls into a sub-optimal
trade-off without favoring either side. Instead, our proposed MARL
methods (IOD and COD) achieve higher growths in terms of GMV
by considering each order's price and the destination potential
concurrently. Orders with relatively low destination potential will
be less possible to get picked, thus avoiding harming GMV by
preventing the driver from trapping in areas with very few future
orders. By direct modeling other agents' policies and capturing the
interaction between agents in the environment, the COD algorithm
with mean field approximation gives the best performance among
all comparing methods.
Order Response Rate. In Table 2 we compare the performance of
different models in terms of the order response rate (OOR), which
is the number of orders taken divided by the number of orders
generated. RES has a higher OOR than the random strategy as it
Figure 5: Normalized hourly income of REV, IOD and COD
with respect to the average hourly income of RAN.
focuses on reducing the trip distance to take more orders. On the
other hand, REV aims to pick higher value orders with longer trip
time, leading to sacrifice on OOR. Although REV has a relatively
higher GMV than other two rule-based methods, its lower OOR
indicates a lower customer satisfaction rate, thus failing to meet
the requirement of an optimal order dispatching algorithm. By con-
sidering both the average order value and the destination potential,
IOD and COD achieve higher OOR as well. High-priced orders with
low destination potential, i.e., a long trip to a suburban area, will
be less possible to get picked, thus avoiding harming OOR when
trying to take a high-priced order.
Average Order Destination Potential. To better present our method
on optimizing the demand-supply gap, we list the average order
destination potential (ADP) in Table 3. Note that all ADP values
are negative, which indicates the supply still cannot fully satisfy
the demand on average. However, IOD and COD largely alleviate
the problem by dispatching orders to places with higher demand.
As shown in Fig. 4, COD largely fills the demand-supply gap in the
city center during peak hours, while REV fails to assign drivers to
take orders with high-demand destination, thus leaving many grids
with unserved orders.
Sequential Performance Analysis. To investigate the performance
change of our proposed algorithms regarding the time of the day,
Fig. 5 shows the normalized income of all algorithms with respect
to the average hourly revenue of RAN. We eliminate RES from
this comparison since it does not focus on getting a higher GMV
and has relatively low performance. A positive value of the bar
graph illustrates an increase in income compared to RAN, and
vice versa. As shown in Fig. 5, MARL methods (IOD and COD)
0246810121416182022TimH (HRurV)−5051015201RrmDOizHd IncRmH (%)5(VI2DC2D Efficient Ridesharing Order Dispatching with MFRL
The Web Conference 2019, May 2019, San Francisco, USA
Table 4: Performance comparison of different reward set-
tings applied to IOD with respect to RAN.
Table 5: Performance comparison in terms of the GMV and
the average arriving time (AAT) with respect to HOD.
Method
AVE
IND
GMV (%)
OOR (%)
−0.36 ± 0.09 −0.14 ± 0.07
+0.36 ± 0.10 +0.30 ± 0.09
AVE + DP −0.26 ± 0.09 −0.10 ± 0.06
+2.97 ± 0.02
+2.54 ± 0.03
IND + DP
ADP (%)
+1.11 ± 1.06
−1.26 ± 0.87
+1.17 ± 1.08
+59.44 ± 1.33
GMV2
AAT
Method GMV1
Q-IOD −0.02% −0.77% +6.76%
+0.02% −0.48% +5.90%
IOD
COD
+0.32% +0.06% +5.49%
outperform RAN in most of the hours in a day (except for the late
night period between 12 a.m. to 4 a.m., when very few orders are
generated). During the peak hours in the morning and at night,
MARL methods achieve a much higher hourly income than RAN.
REV achieves a higher income than MARL methods between late
night and early morning (12 a.m. to 8 a.m.) because of its aggressive
strategy to take high-priced orders. However, this strategy ignores
the destination quality of orders and assigns many drivers with
orders to places with low demand, resulting in a significant income
drop during the rest of the day (except for the evening peak hours
when the possibility of encountering high-priced orders is large
enough for REV to counteract the influence of the low response
rate). On the other hand, IOD and COD earn significantly higher
than RAN and REV for the rest of the day, possibly because of
MARL methods' ability to recognize a better order in terms of both
the order fee and the destination potential. As the order choice of
MARL methods will prevent the agent from choosing a destination
with lower potential, agents following MARL methods are thus
enjoying more sustainable incomes. Also, COD outperforms IOD
constantly, showing the effectiveness of explicitly conditioning on
other agents' policies to capture the interaction between agents.
Effectiveness of Reward Settings. As described in Section 2.1.1, we
use the destination potential as a reward function regularizer to
encourage cooperation between agents. To show the effectiveness
of this reward setting, we compare the GMV, OOR, and ADP of
setting average income (AVE) and independent income (IND) as
each agent's reward for IOD respectively. We also measure the
performance of adding DP as a regularizer for both settings. For
this experiment, we bootstrap 10% of the original real orders for
the training and test set separately. As shown in Table 4, the perfor-
mance of AVE in terms of all metrics is relatively lower than those
of IND methods and RAN (even with DP added). This is possibly
because of the credit assignment problem, where the agent's behav-
ior is drowned by the noise of other agents' impact on the reward
function. On the other hand, setting the individual income as the
reward helps to distinguish each agent's contribution to the global
objective from others, while adding DP as a regularizer further
encourages the coordination between agents by arranging them to
places with higher demand.
3.2 Coordinate-based Experiment
3.2.1 Environment Setting. As the real-world environment is
coordinate-based rather than grid-based, we also conduct experi-
ments on a more complex coordinate-based simulator provided by
DiDi Chuxing. At each time step t, the coordinate-based simulator
provides an observation ot including a set of active drivers and a set
of available orders. Each order feature includes the coordinate for
the origin and the destination, while each driver has the coordinate
as the location feature. The order dispatching algorithm works the
same as described in Section 3.1.1. To better approximate the real-
world scenario, this simulator also considers order cancelations, i.e.,
an order might be canceled during the pick-up process. This dy-
namic is controlled by a random variable which is positively related
to the arriving time. The data resource of this simulator is based
on historical dispatching events, including order generation events,
driver logging on/off events and order fee estimation. During the
training stage, the simulator will load five weekdays data with 350K
dispatching events and generate 9K drivers. For evaluation, our
model is applied on future days which are not used in the training
phase.
3.2.2 Model Setting. We evaluate the performance of following
MARL based methods including IOD, COD, and a DQN variation
of IOD (Q-IOD), i.e., without the policy network. We also compare
these MARL methods with a centralized combinatorial optimiza-
tion method based on the Hungarian algorithm (HOD). The HOD
method focuses on minimizing the average arriving time (AAT)
by setting the weight of each driver-order pair with the pick-up
distance. For all MARL based methods, the same network architec-
ture setting as described in Section 3.1.2 is applied. Except that we
use a mini-batch size of 200 because of the shorter simulation gap.
α = −0.1 in this
The regularization ratio for pick-up distance is 2
experiment.
3.2.3 Result Analysis. We train all MARL methods for 400K
iterations and apply the trained model in a test set (consists of three
weekdays) for comparison. We compare different algorithms in
terms of the total income in a day (GMV) and the average arriving
time (AAT). GMV2 considers the cancellation while GMV1 doesn't.
All the above metrics are normalized with respect to the result of
HOD.
As shown in Table 5, the result of COD largely outperforms Q-
IOD and IOD in both GMV1 and GMV2, showing the effectiveness of
direct modeling of other agents' policies in MARL. In addition, COD
outperforms HOD in both GMV settings as well; this justifies the
advantage of MARL algorithms that exploit the interaction between
agents and the environment to maximize the cumulative reward.
The performance improvement of GMV2 is smaller than that of
GMV1 for MARL methods. This is possibly because that HOD works
by minimizing the global pick-up distance and has a shorter waiting
time. On the other hand, MARL methods only consider the pick-up
distance as a regularization term, thus performing comparatively
worse than HOD regarding AAT. As shown in Table 5, the AAT of all
The Web Conference 2019, May 2019, San Francisco, USA
M. Li et al.
MARL methods are relatively longer than that of the combinatorial
optimization method. However, as the absolute values of GMV
are orders of magnitude higher than ATT, the increase in ATT
is relatively minor and is thus tolerable in the order dispatching
task. Also, MARL methods require no centralized control during
execution, thus making the order dispatching system more robust
to potential hardware or connectivity failures.
4 RELATED WORK
Order Dispatching. Several previous works addressed the order
dispatching problem by either centralized or decentralized ruled-
based approaches. Lee et al. [16] and Lee et al. [17] chose the pick-up
distance (or time) as the basic criterion, and focused on finding the
nearest option from a set of homogeneous drivers for each order on
a first-come, first-served basis. These approaches only focus on the
individual order pick-up distance; however, they do not account
for the possibility of other orders in the waiting queue being more
suitable for this driver. To improve global performance, Zhang et al.
[44] proposed a novel model based on centralized combinatorial
optimization by concurrently matching multiple driver-order pairs
within a short time window. They considered each driver as het-
erogenous by taking the long-term behavior history and short-term
interests into account. The above methods work with centralized
control, which is prone to the potential "single point of failure" [23].
With the decentralized setting, Seow et al. [30] addressed the
problem by grouping neighboring drivers and orders in a small
multi-agent environment, and then simultaneously assigning or-
ders to drivers within the group. Drivers in a group are considered
as agents who conduct negotiations by several rounds of collab-
orative reasoning to decide whether to exchange current order
assignments or not. This approach requires rounds of direct com-
munications between agents, thus being limited to a local area with
a small number of agents. Alshamsi and Abdallah [3] proposed an
adaptive approach for the multi-agent scheduling system to en-
able negotiations between agents (drivers) to re-schedule allocated
orders. They used a cycling transfer algorithm to evaluate each
driver-order pair with multiple criteria, requiring a sophisticated
design of feature selection and weighting scheme.
Different from rule-based approaches, which require additionally
hand-crafted heuristics, we use a model-free RL agent to learn an
optimal policy given the rewards and observations provided by the
environment. A very recent work by Xu et al. [40] proposed an
RL-based dispatching algorithm to optimize resource utilization
and user experience in a global and more farsighted view. However,
they formulated the problem with the single-agent setting, which
is unable to model the complex interactions between drivers and
orders. On the contrary, our multi-agent setting follows the dis-
tributed nature of the peer-to-peer ridesharing problem, providing
the dispatching system with the ability to capture the stochastic
demand-supply dynamics in large-scale ridesharing scenarios. Dur-
ing the execution stage, agents will behave under the learned policy
independently, thus being more robust to potential hardware or
connectivity failures.
Multi-Agent Reinforcement Learning. One of the most straightfor-
ward approaches to adapt reinforcement learning in the multi-agent
environment is to make each agent learn independently regardless
of the other agents, such as independent Q-learning [36]. They,
however, tend to fail in practice [24] because of the non-stationary
nature of the multi-agent environment. Several approaches have
been attempted to address this problem, including sharing the pol-
icy parameters [12], training the Q-function with other agent's
policy parameters [37], or using importance sampling to learn from
data gathered in a different environment [9]. The idea of central-
ized training with decentralized execution has been investigated
by several works [8, 21, 29] recently for MARL using policy gradi-
ents [34], and deep neural networks function approximators, based
on the actor-critic framework [15]. Agents within this paradigm
learn a centralized Q-function augmented with actions of other
agents as the critic during training stage, and use the learned policy
(the actor) with local observations to guide their behaviors during
execution. Most of these approaches limit their work to a small
number of agents usually less than ten. To address the problem
of the increasing input space and accumulated exploratory noises
of other agents in large-scale MARL, Yang et al. [42] proposed a
novel method by integrating MARL with mean field approxima-
tions and proved its convergence in a tabular Q-function setting.
In this work, we further develop MFRL and prove its convergence
when the Q-function is represented by function approximators.
5 CONCLUSION
In this paper, we proposed the multi-agent reinforcement learning
solution to the order dispatching problem. Results on two large-
scale simulation environments have shown that our proposed al-
gorithms (COD and IOD) achieved (1) a higher GMV and OOR
than three rule-based methods (RAN, RES, REV); (2) a higher GMV
than the combinatorial optimization method (HOD), with desirable
properties of fully distributed execution; (3) lower supply-demand
gap during the rush hours, which indicates the ability to reduce
traffic congestion. We also provide the convergence proof of ap-
plying mean field theory to MARL with function approximations
as the theoretical justification of our proposed algorithms. Fur-
thermore, our MARL approaches could achieve fully decentralized
execution by distributing the centralized trained policy to each
vehicle through Vehicle-to-Network (V2N). For future work, we are
working towards controlling ATT while maximizing the GMV with
the proposed MARL framework. Another interesting and practical
direction to develop is to use a heterogeneous agent setting with
individual specific features, such as the personal preference and
the distance from its own destination.
REFERENCES
[1] Khadige Abboud, Hassan Aboubakr Omar, and Weihua Zhuang. 2016. Inter-
working of DSRC and cellular network technologies for V2X communications: A
survey. IEEE transactions on vehicular technology 65, 12 (2016), 9457 -- 9470.
[2] Adrian K. Agogino and Kagan Tumer. 2008. Analyzing and Visualizing Multiagent
Rewards in Dynamic and Stochastic Environments. Journal of Autonomous Agents
and Multiagent Systems (2008), 320 -- 338.
[3] Aamena Alshamsi and Sherief Abdallah. 2009. Multiagent self-organization for a
taxi dispatch system. In Proceedings of 8th International Conference of Autonomous
Agents and Multiagent Systems, 2009. 89 -- 96.
[4] Andrew Amey, John Attanucci, and Rabi Mishalani. 2011. Real-time ridesharing:
opportunities and challenges in using mobile phone technology to improve
rideshare services. Transportation Research Record: Journal of the Transportation
Research Board 2217 (2011), 103 -- 110.
[5] Luigi Atzori, Antonio Iera, and Giacomo Morabito. 2010. The Internet of Things:
A survey. Computer Networks 54, 15 (2010), 2787 -- 2805. https://doi.org/10.1016/
j.comnet.2010.05.010
Efficient Ridesharing Order Dispatching with MFRL
The Web Conference 2019, May 2019, San Francisco, USA
[29] Peng Peng, Quan Yuan, Ying Wen, Yaodong Yang, Zhenkun Tang, Haitao Long,
and Jun Wang. 2017. Multiagent Bidirectionally-Coordinated Nets for Learning
to Play StarCraft Combat Games. arXiv preprint arXiv:1703.10069 (2017).
[30] Kiam Tian Seow, Nam Hai Dang, and Der-Horng Lee. 2010. A collaborative
multiagent taxi-dispatch system. IEEE Transactions on Automation Science and
Engineering 7, 3 (2010), 607 -- 616.
[31] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and
Martin Riedmiller. 2014. Deterministic policy gradient algorithms. In ICML.
387 -- 395.
[32] Richard S Sutton. 1988. Learning to predict by the methods of temporal differences.
Machine learning 3, 1 (1988), 9 -- 44.
[33] Richard S Sutton and Andrew G Barto. 1998. Reinforcement learning: An intro-
duction. Vol. 1. MIT press Cambridge.
[34] Richard S Sutton, David A McAllester, Satinder P Singh, Yishay Mansour, et al.
1999. Policy gradient methods for reinforcement learning with function approxi-
mation.. In NIPS, Vol. 99. 1057 -- 1063.
[35] Csaba Szepesvári and Michael L Littman. 1999. A unified analysis of value-
function-based reinforcement-learning algorithms. Neural computation 11, 8
(1999), 2017 -- 2060.
[36] Ming Tan. 1993. Multi-agent reinforcement learning: Independent vs. cooperative
agents. In Proceedings of the tenth international conference on machine learning.
330 -- 337.
[37] Gerald Tesauro. 2004. Extending Q-learning to general adaptive multi-agent
systems. In NIPS. 871 -- 878.
8, 3-4 (1992), 279 -- 292.
Publishing.
[38] Christopher JCH Watkins and Peter Dayan. 1992. Q-learning. Machine learning
[39] Michael Wooldridge. 2009. An Introduction to MultiAgent Systems (2nd ed.). Wiley
[40] Zhe Xu, Zhixin Li, Qingwen Guan, Dingshui Zhang, Qiang Li, Junxiao Nan,
Chunyang Liu, Wei Bian, and Jieping Ye. 2018. Large-Scale Order Dispatch in
On-Demand Ride-Hailing Platforms: A Learning and Planning Approach. In
Proceedings of the 24th ACM SIGKDD International Conference on Knowledge
Discovery & Data Mining (KDD '18). ACM, New York, NY, USA, 905 -- 913.
https://doi.org/10.1145/3219819.3219824
[41] F. Yang, S. Wang, J. Li, Z. Liu, and Q. Sun. 2014. An overview of Internet of
Vehicles. China Communications 11, 10 (Oct 2014), 1 -- 15. https://doi.org/10.1109/
CC.2014.6969789
[42] Yaodong Yang, Rui Luo, Minne Li, Ming Zhou, Weinan Zhang, and Jun Wang.
2018. Mean Field Multi-Agent Reinforcement Learning. In Proceedings of the
35th International Conference on Machine Learning (Proceedings of Machine Learn-
ing Research), Jennifer Dy and Andreas Krause (Eds.), Vol. 80. PMLR, Stock-
holmsmÃďssan, Stockholm Sweden, 5567 -- 5576.
[43] A. Zanella, N. Bui, A. Castellani, L. Vangelista, and M. Zorzi. 2014. Internet of
Things for Smart Cities. IEEE Internet of Things Journal 1, 1 (Feb 2014), 22 -- 32.
https://doi.org/10.1109/JIOT.2014.2306328
[44] Lingyu Zhang, Tao Hu, Yue Min, Guobin Wu, Junying Zhang, Pengcheng Feng,
Pinghua Gong, and Jieping Ye. 2017. A Taxi Order Dispatch Model Based On
Combinatorial Optimization. In Proceedings of the 23rd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining (KDD '17). ACM, New York,
NY, USA, 2151 -- 2159. https://doi.org/10.1145/3097983.3098138
[6] Dimitri P Bertsekas. 2012. Weighted sup-norm contractions in dynamic pro-
gramming: A review and some new applications. Dept. Elect. Eng. Comput. Sci.,
Massachusetts Inst. Technol., Cambridge, MA, USA, Tech. Rep. LIDS-P-2884 (2012).
[7] Lucian Buşoniu, Robert Babuška, and Bart De Schutter. 2010. Multi-agent Rein-
forcement Learning: An Overview. Springer Berlin Heidelberg, Berlin, Heidelberg,
183 -- 221. https://doi.org/10.1007/978-3-642-14435-6_7
[8] Jakob Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and
Shimon Whiteson. 2017. Counterfactual Multi-Agent Policy Gradients. arXiv
preprint arXiv:1705.08926 (2017).
[9] Jakob Foerster, Nantas Nardelli, Gregory Farquhar, Triantafyllos Afouras,
Philip HS Torr, Pushmeet Kohli, and Shimon Whiteson. 2017. Stabilising Expe-
rience Replay for Deep Multi-Agent Reinforcement Learning. In International
Conference on Machine Learning. 1146 -- 1155.
[10] Masabumi Furuhata, Maged Dessouky, Fernando Ordóñez, Marc-Etienne Brunet,
Xiaoqing Wang, and Sven Koenig. 2013. Ridesharing: The state-of-the-art and
future directions. Transportation Research Part B: Methodological 57 (2013), 28 -- 46.
[11] Jayavardhana Gubbi, Rajkumar Buyya, Slaven Marusic, and Marimuthu
Palaniswami. 2013. Internet of Things (IoT): A vision, architectural elements, and
future directions. Future Generation Computer Systems 29, 7 (2013), 1645 -- 1660.
https://doi.org/10.1016/j.future.2013.01.010 Including Special sections: Cyber-
enabled Distributed Computing for Ubiquitous Cloud and Network Services &
Cloud Computing and Scientific Applications -- Big Data, Scalable Analytics,
and Beyond.
[12] Jayesh K Gupta, Maxim Egorov, and Mykel Kochenderfer. 2017. Cooperative
multi-agent control using deep reinforcement learning. In AAMAS. Springer,
66 -- 83.
[13] Tommi Jaakkola, Michael I Jordan, and Satinder P Singh. 1994. Convergence of
[15] Vijay R. Konda and John N. Tsitsiklis. 2000. Actor-Critic Algorithms.
[14] Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti-
stochastic iterative dynamic programming algorithms. In NIPS. 703 -- 710.
mization. arXiv preprint arXiv:1412.6980 (2014).
In
Advances in Neural Information Processing Systems 12, S. A. Solla, T. K. Leen,
and K. Müller (Eds.). MIT Press, 1008 -- 1014.
http://papers.nips.cc/paper/
1786-actor-critic-algorithms.pdf
[16] Der-Horng Lee, Hao Wang, Ruey Cheu, and Siew Teo. 2004. Taxi dispatch
system based on current demands and real-time traffic conditions. Transportation
Research Record: Journal of the Transportation Research Board 1882 (2004), 193 --
200.
[17] Junghoon Lee, Gyung-Leen Park, Hanil Kim, Young-Kyu Yang, Pankoo Kim,
and Sang-Wook Kim. 2007. A Telematics Service System Based on the Linux
Cluster. In Computational Science -- ICCS 2007, Yong Shi, Geert Dick van Albada,
Jack Dongarra, and Peter M. A. Sloot (Eds.). Springer Berlin Heidelberg, Berlin,
Heidelberg, 660 -- 667.
[18] Ziru Li, Yili Hong, and Zhongju Zhang. 2016. An empirical analysis of on-demand
ride sharing and traffic congestion. In 2016 International Conference on Information
Systems, ICIS 2016. Association for Information Systems.
[19] Kaixiang Lin, Renyu Zhao, Zhe Xu, and Jiayu Zhou. 2018. Efficient Large-Scale
Fleet Management via Multi-Agent Deep Reinforcement Learning. arXiv preprint
arXiv:1802.06444 (2018).
[20] Michael L. Littman. 1994. Markov Games As a Framework for Multi-agent
Reinforcement Learning. In Proceedings of the Eleventh International Conference
on International Conference on Machine Learning (ICML'94). Morgan Kaufmann
Publishers Inc., San Francisco, CA, USA, 157 -- 163. http://dl.acm.org/citation.
cfm?id=3091574.3091594
[21] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor
Mordatch. 2017. Multi-agent actor-critic for mixed cooperative-competitive
environments. In NIPS. 6382 -- 6393.
[22] N. Lu, N. Cheng, N. Zhang, X. Shen, and J. W. Mark. 2014. Connected Vehicles:
Solutions and Challenges. IEEE Internet of Things Journal 1, 4 (Aug 2014), 289 -- 299.
https://doi.org/10.1109/JIOT.2014.2327587
[23] Gary S Lynch. 2009. Single point of failure: The 10 essential laws of supply chain
risk management. John Wiley & Sons.
[24] Laetitia Matignon, Guillaume J Laurent, and Nadine Le Fort-Piat. 2012. Indepen-
dent reinforcement learners in cooperative Markov games: a survey regarding
coordination problems. The Knowledge Engineering Review 27, 1 (2012), 1 -- 31.
[25] Francisco S Melo, Sean P Meyn, and M Isabel Ribeiro. 2008. An analysis of
reinforcement learning with function approximation. In Proceedings of the 25th
international conference on Machine learning. ACM, 664 -- 671.
[26] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness,
Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg
Ostrovski, et al. 2015. Human-level control through deep reinforcement learning.
Nature 518, 7540 (2015), 529 -- 533.
[27] Vinod Nair and Geoffrey E Hinton. 2010. Rectified linear units improve re-
stricted boltzmann machines. In Proceedings of the 27th international conference
on machine learning (ICML-10). 807 -- 814.
[28] Christos H. Papadimitriou and Kenneth Steiglitz. 1982. Combinatorial Optimiza-
tion: Algorithms and Complexity. Prentice-Hall, Inc., Upper Saddle River, NJ,
USA.
|
1003.1792 | 1 | 1003 | 2010-03-09T06:35:17 | A Hybrid System based on Multi-Agent System in the Data Preprocessing Stage | [
"cs.MA"
] | We describe the usage of the Multi-agent system in the data preprocessing stage of an on-going project, called e-Wedding. The aim of this project is to utilize MAS and various approaches, like Web services, Ontology, and Data mining techniques, in e-Business that want to improve responsiveness and efficiency of systems so as to extract customer behavior model on Wedding Businesses. However, in this paper, we propose and implement the multi-agent-system, based on JADE, to only cope data preprocessing stage specified on handle with missing value techniques. JADE is quite easy to learn and use. Moreover, it supports many agent approaches such as agent communication, protocol, behavior and ontology. This framework has been experimented and evaluated in the realization of a simple, but realistic. The results, though still preliminary, are quite. | cs.MA | cs | (IJCSIS) International Journal of Computer Science and Information Security,
Vol. 7, No. 2, 2010
A Hybrid System based on Multi-Agent System in the
Data Preprocessing Stage
Kobkul Kularbphettong, Gareth Clayton and Phayung Meesad
The Information Technology Faculty,
King Mongkut’s University of Technology
North Bangkok, Thailand
Abstract—We describe the usage of the Multi-agent system in the
data preprocessing stage of an on-going project, called e-
Wedding. The aim of this project is to utilize MAS and various
approaches, like Web services, Ontology, and Data mining
techniques, in e-Business that want to improve responsiveness
and efficiency of systems so as to extract customer behavior
model on Wedding Businesses. However, in this paper, we
propose and implement the multi-agent-system, based on JADE,
to only cope data preprocessing stage specified on handle with
missing value techniques. JADE is quite easy to learn and use.
Moreover, it supports many agent approaches such as agent
communication, protocol, behavior and ontology. This
framework has been experimented and evaluated
in the
realization of a simple, but realistic. The results, though still
preliminary, are quite.
Keywords-component; multi-agent system, data preprocessing
stage, Web services, ontology, data mining techniques, e-Wedding,
JADE.
I.
INTRODUCTION
Data Preprocessing is one of the significant factors that
affects on the success of Data Mining and Machine Learning
approaches. Generally, data preprocessing stage represents the
quality of data. The more incorrect and incomplete data
presents, the more result is unreliable. Moreover, data
preprocessing task is time consuming because it includes
many phases like data cleaning, data integrating, data
transforming and data
the best
reducing. However,
performance of the data pre-processing algorithms is relied on
the nature of each data set. Hence, it would be nice if it have
the interested methodology to adapt for choosing the best
performance of the data preprocessing algorithms for each
data set.
Recently, although there are much of researches applied
MAS (Multi-agent system) in a wide range of problem in Data
Mining and Machine Learning techniques, very few researches
are focused on using MAS in the data preprocessing step. A
multi agent system is a computational system, or a loosely
coupled network in which two or more agents interact or work
together to perform a set of tasks or to satisfy a set of goals.
Each agent is considered as a locus of a problem-solving
activity which operates asynchronously with respect to the
other agents [1]. Therefore, in this paper we propose the MAS
framework of an on-going project, called e-Wedding, that
merely focuses on using MAS to handle in the problems of
data preprocessing stage.
The remainder of this paper is organized as follows. Section
2 reviews about related literatures and research works in the
use of multi-agent system for data mining. Section 3 presents
the related methodologies used in this work. Section 4 presents
the experimental results based on the purposed model based
on multi-agent framework. This prototype demonstrates how
to success
in data
for adapting multi-agent system
preprocessing stage. Finally, we conclude the paper with
future research issues in section 5.
II. RELATED WORKS
A literature search shows that most of the related
researches have been deployed multi-agent to cope with the
data mining techniques specified on the data preprocessing
algorithms by following this:
According to [5], they showed a prototype of the system
using the JADE platform in the context of travel industry.
Furthermore, other
research works
show
that agent
technologies are deployed as a significant tool for developing
e-Commerce
[2]-[7]. Hence, multi-agent
applications
technology, a promising approach, trends to handle internet
transaction for customers.
Moreover, other researchers propose an agent-based
framework representing in various ways that are related with
data mining techniques [8]-[10]. For instances, Chien-Ho Wu,
Yuehjen E. Shao, Jeng-Fu Liu, and Tsair-Yuan Chang[[11]
applied agent technology to collect and integrate data
distributed over various computing platforms to facilitate
statistical data analysis in replacing the missing values by
using either Approximate Bayesian Bootstrap (ABB) or Ratio
Imputation and using MAS improves the execution time at
different by focusing on spatial knowledge in order to extract
knowledge
in Predictive Modeling Markup Language
(PMML) format [9].
From previous literature works, it appears that there are
many research studies exploiting various techniques blended
with multi-agent technology and data mining techniques.
Consequently, in order to success on e-Commerce, agent
should have abilities to perform as a behalf of user to handle
199http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security,
Vol. 7, No. 2, 2010
phrase "Garbage In, Garbage Out" is particularly applicable
[15].
to data mining and machine learning projects
with business tasks such as planning, reasoning and learning.
Also, data mining techniques is the important way to make a
reason for agent under uncertainty and with incomplete
information situations. Notwithstanding, data preprocessing
step acts as an crucial task to filter and select suitable
information before processing any mining algorithms.
III. THE METHODOLOGIES
In this section, we illustrate the specified methodologies
used in this project but it is only focused on the approaches
using in the data preprocessing stage specified in dealing with
missing values.
A. Multi-Agent System
Agent is the software program that enables to autonomous
action in some environment so as to meet its design objectives.
According to N. R. Jennings and M. Wooldridge [12], the
essential characters of each agent are following: reactive, pro-
active, autonomous, object-oriented and social ability. Each
agent can play as a behalf of the user and execute the
particular task. Also, Padghan and Winikopff [13] described
that the concept of Agent refers to an entity acting on behalf of
other entities or organizations and having the ability to
perceive relevant information, and following and handling the
objectives to be accomplished. However, in open and dynamic
environment like internet, a multi-agent system is one of the
promising means to help reduce cost, increase efficiency,
reduce errors and achieve optimal deal.
There are two issues related to the design of MAS: Agent
Communication Language and agent development platform.
The former concerns with the message interchange between
different agent such as KQML, and FIPA ACL. The latter is
related with the platform development to provide an effective
framework, such as IBM Aglets, ObjectSpace Voyager and
etc, for the dispatching, communications, and management of
multiple agents in the open and dynamic environment.
(Java Agent
JADE
this proposed project,
For
Development Framework) will be deployed as the prototype
development
tool.
JADE
(Java Agent Development
Framework) is a software environment fully implemented in
JAVA language aiming at the development of multi-agent
systems that comply with FIPA specifications [14]. The goal
of JADE is to simplify development while ensuring standard
compliance through a comprehensive set of system services
and agents. Each running instance of the JADE runtime
environment is called a container as it can contain several
agents. The set of active containers is called a platform. A
single special container must always be active in a platform
and all other containers register with it as soon as they start.
Hence, the development framework based on JADE is
considered very suitable for implementing applications that
require distributing computation tasks over the network.
B. Data Preprocessing Techniuques
Data pre-processing is an often neglected but important step
in the data mining process, as depicted in figure. 1. The
Figure 1. The data mining steps [16].
There are many tasks in data preprocessing like data
cleaning, data transformation, data reduction, data integration
and etc. Nevertheless, in this paper, we emphasizes on the data
cleaning stage so as to handle missing data.
Missing data is a common problem for data quality in real
datasets. However, there are several methods for handling
missing data and according Little and Rubin[17], missing data
treatment can be divided to three categories a) Ignoring and
discarding data, which known discarding method can be
categorized to the complete case analysis method and the pair
wise deletion method. The former discards all instances with
missing data and the latter discards only the instances with
high level of missing data, determined the extent of missing
data before b) Parameter estimation, which Maximum
likelihood procedures that use variants of the Expectation-
Maximization algorithm can cope estimation parameter in the
presence of missing and c) Imputation techniques, which
missing values are filled with estimated ones based on
information available in the data set.
Also, there are some popular missing data treatment
methods that researcher have to choose by following this:
• Mean or mode substitution: replacing all missing data
with the mean (numeric attribute) or mode (nominal
attribute) of all observed cases. However, the
drawbacks of using these methods are to changing
the characteristic of the original dataset and ignoring
the relationship among attributes that affect on the
performance of the data mining algorithms.
• Regression substitution: replacing all missing data
with a statistics relied on the assumption of linear
relationship between attributes.
• Hot deck imputation: replacing all missing data with
an estimated distribution from the current data. In
random hot deck, a missing value is replaced by a
observed value (the donor) of the attribute chosen
randomly, similarly to hot deck, but in cold deck
methods, the imputed value must be different from
the current data source.
• KNN Imputation: replacing all missing data with k-
nearest neighbor algorithm
that determines
the
200http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal of Computer Science and Information Security,
Vol. 7, No. 2, 2010
Figure 4. The operation of JADE in e-Wedding System.
Recently, there are a lot of algorithms dealing with missing
data. The basic approaches about these popular algorithms
have been introduced above section. In this paper, we
implement a composite imputation method between hot deck
and nearest neighbor methods based on mean substitution,
shown figure 4.
Hot deck imputation technique is commonly used in
statistic for item non response. The main concept of the hot
deck method is to use the current data, called donors, to
provide imputed values for records with missing values. The
procedure through which we find the donor that matches the
record with missing values is different according to the
particular techniques used [20].
similarity of two instances by using a distance
function.
• Classification methods: replacing all missing data
with classification models, like decision tree, C4.5
and etc, and using all relevant features as predictors.
IV. THE PURPOSED FRAMEWORK AND EXPERIMENTAL
RESULTS
This section displayed the purposed framework of this
project and compares the result of the chosen missing value
algorithms. For illustration of framework as figure 2-4 [18,
19], we select the wedding businesses and its environment.
There are several issues in this system such as multi-agent
system, web services, ontology, and data mining techniques,
as shown in figure 3, but in this paper we present merely a
multi-agent system dealing with data preprocessing steps and
focusing on the missing value techniques.
Figure 2. The purposed architecture of the e-Wedding system.
Figure 3. The web page in e-Wedding System.
Figure 5. The process of the purposed Imputation technique.
A hot deck imputation method can be described by two
factors: the way in which donors are selected for each missing
instance and the way in which the weight of the donor is
defined for each missing instance [21].
According to Jae Kwang Kim, the first way can determine by
the distribution of d equals
d
d
;
=
ij
jAi
∈
R
,
(1)
∈
A
m
201http://sites.google.com/site/ijcsis/ ISSN 1947-5500 ij
i
=
Y
(2)
where
ywd
*
ij
ij
AR = the set of indices of the sample respondents
Am = the set of indices of the sample nonrespondents
And dij = the number of times that Yi is used as donor for Yj
The second way can determine the weight of the donor
specified for each missing item. For missing item j:
∑
RAi
∈
ijw
*
= the fraction of the original weight assigned to
Let
donor i as a donor for element j.
Also, to tailor the hot deck imputation process, predictive
mean matching is applied to this process. Predicted means are
then calculated for both records where the item is missing and
records where it is non-missing. Donors for those records
requiring imputation are selected by matching on the predicted
means, according to some specified distance metric. The
imputed value is then the value of the item on the donor
record.
Figure 6. The operation of JADE in the purposed Imputation technique.
V. CONCLUSION AND FUTURE WORKS
In this paper we presented our preliminary ideas of building
multi-agent system with data preprocessing steps by specified
in a missing value step, based on e-Wedding system. In the
part of MAS, we have implemented this prototype by using
JADE platform. JADE is quite easy to learn and use.
Moreover, it supports many agent approaches such as agent
communication, protocol, behavior and ontology. As for the
future work, we need to explore more reasonable and data
mining technologies
REFERENCES
[1] Sandholm, T. and Lesser, “Advantages of a Leveled Commitment
Contracting Protocol.” Thirteenth National Conference on Artificial
Intelligence (AAAI-96), pp. 126--133, Portland, OR.
[2] Bala M. Balachandran and Majigsuren Enkhsaikhan, “Developing Multi-
agent E-Commerce Applications with JADE”, Lecture Notes
in
Computer Science, Springer Berlin / Heidelberg 2009.
[3] Mu-Kun Cao, Yu-Qiang Feng, Chun-Yan Wang, “Designing Intelligent
Agent for e-Business Oriented Multi-agent Automated Negotiation.”
Proceedings of the Fourth International Conference on Machine
Learning and Cybernetics, Guangzhou, 18-21 August 2005.
[4] Patrick C. K. Hung, Ji-Ye Mao. “Modeling of E-negotiation Activities
with Petri Nets.”, HICSS 2002.
Y. Yuan, J. B. Rose, N. Archer, and H. Suarga, “A Web-Based
Negotiation Support System.” International Journal of Electronic
Markets,1998.
[5]
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 7, No. 2, 2010
[6] Giacomo
Preist.
and Chris
Piccinelli, Claudio Bartolini
"E-service composition: supporting dynamic definition of process-
oriented negotiation" In Proc. 12th International Workshop on Database
and Expert Systems Applications (DEXA 2001), Munich, Germany,
September 2001. IEEE Computer Society 2001.
[7] Akkermans, H. “Intelligent E-Business - From Technology to Value.”,
IEEE Intelligent Systems, 16(4):8-10, 2001.
[8] Huang Xin Li; Chosler, R. “Application of Multilayered Multi-Agent
Data Mining Architecture to Bank Domain”, Wireless Communications,
NetworkingandMobileComputing,2007.InternationalConferenceonVolu
me, Issue , 21-25 Sept. 2007 Page(s):6721 – 6724.
[9] H. Baazaoui Zghal, S. Faiz, and H. Ben Ghezala, “A Framework for
Data Mining Based Multi-Agent: An Application to Spatial Data”,
World Academy of Science, Engineering and Technology 5 2005.
[10] Zili Zhang, Chengqi Zhang and Shichaozhang, “An Agent-Based Hybrid
Framework for Database Mining”, Applied Artificial Intelligence,
17:383–398, 2003.
[11] Chien-Ho Wu, Yuehjen E. Shao, Jeng-Fu Liu, and Tsair-Yuan Chang ,
“On Supporting Cross-Platform Statistical Data Analysis Using JADE”,
Book Series Studies in Computational Intelligence, Springer Berlin
Heidelberg,
issn1860-949X
(Print) 1860-9503
(Online), Volume
214/2009.
[12] N. R. Jennings and M. Wooldridge. “Software Agents”, IEE Review
42(1), pages 17-21. January 1996.
[13] Padghan, L.
and Winikopff, M.,
AgentSystems.”, Wiley.2004.
[14] JADE,
Agent
Java
Environment,2006, http://jade.tilab.com
[15] http://en.wikipedia.org/wiki/Data_Pre-processing
[16] http://www.infovis-wiki.net/index.php?title=Image:Fayyad96kdd-
process.png
[17] Little, R. J. A. and Rubin, D. B.,"Statistical Analysis with Missing Data"
2nd Edition, John Wiley & Sons, New York, 2002
[18] Kobkul Kilarbphettong,
“e-Negotiation based on Multi-agent
ssyetm”,JCSSE 2007 –The international Joint Conference on Computer
Science and software Engineer,Thailand.
[19] Kobkul Kilarbphettong, Gareth Clayton, and Phayung Meesad, “e-
Wedding System based on Multi-System”, Advances in Intelligent and
Soft-Computing ,series of Springer, 2010
[20] Andrea Piesse, David Judkins, and Zizhong Fan., “Item Imputation
Made Easy”, Proceedings of
the Survey Research Methods
Section,2005.
[21] Jae Kwang Kim, “Variance Estimation for Nearest Neighbor Imputation
with Application to Census Long Form Data”, Proceedings of the
Survey Research Methods Section,2002.
AUTHORS PROFILE
Development
“Developing
Intelligent
Kobkul Kularbphettong received the B.S. degree in
Computer Business, M.S. degree in Computer Science.
She
is Currently Ph.D. Student
in
Information
Technology. Her current research interests Multi-agent
System, Web Services, Semantic Web Services,
Ontology and Data mining techniques.
Dr.Gareth
Clayton
Statistician,
a
is
so for IT students any project involving statistics,
including the following, but not excluding other areas:
Data Mining, Simulation
studies, Design
of
Experiments, Model Fitting Parameter Estimation.
Phayung Meesad received the B.S.,M.S., and Ph.D.
degree in Electrical Engineering. His current research
interests Fuzzy Systems and Neural Networks,
Evolutionary Computation and Discrete Control
System.
202http://sites.google.com/site/ijcsis/ ISSN 1947-5500 |
1611.05638 | 1 | 1611 | 2016-11-17T11:28:27 | Fictitious play for cooperative action selection in robot teams | [
"cs.MA"
] | A game theoretic distributed decision making approach is presented for the problem of control effort allocation in a robotic team based on a novel variant of fictitious play. The proposed learning process allows the robots to accomplish their objectives by coordinating their actions in order to efficiently complete their tasks. In particular, each robot of the team predicts the other robots' planned actions while making decisions to maximise their own expected reward that depends on the reward for joint successful completion of the task. Action selection is interpreted as an $n$-player cooperative game. The approach presented can be seen as part of the \emph{Belief Desire Intention} (BDI) framework, also can address the problem of cooperative, legal, safe, considerate and emphatic decisions by robots if their individual and group rewards are suitably defined. After theoretical analysis the performance of the proposed algorithm is tested on four simulation scenarios. The first one is a coordination game between two material handling robots, the second one is a warehouse patrolling task by a team of robots, the third one presents a coordination mechanism between two robots that carry a heavy object on a corridor and the fourth one is an example of coordination on a sensors network. | cs.MA | cs | Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
Journal
Logo
Fictitious play for cooperative action selection in robot teams
aDepartment of Automatic Control and Systems Engineering The University of Sheffield Mappin Street Sheffield, S1 3JD United Kingdom
M. Smyrnakisa,, S. Veresa
Abstract
A game theoretic distributed decision making approach is presented for the problem of control effort allocation in a robotic team
based on a novel variant of fictitious play. The proposed learning process allows the robots to accomplish their objectives by
coordinating their actions in order to efficiently complete their tasks. In particular, each robot of the team predicts the other robots'
planned actions while making decisions to maximise their own expected reward that depends on the reward for joint successful
completion of the task. Action selection is interpreted as an n-player cooperative game. The approach presented can be seen as part
of the Belief Desire Intention (BDI) framework, also can address the problem of cooperative, legal, safe, considerate and emphatic
decisions by robots if their individual and group rewards are suitably defined. After theoretical analysis the performance of the
proposed algorithm is tested on four simulation scenarios. The first one is a coordination game between two material handling
robots, the second one is a warehouse patrolling task by a team of robots, the third one presents a coordination mechanism between
two robots that carry a heavy object on a corridor and the fourth one is an example of coordination on a sensors network.
Keywords: Robot team coordination; Fictitious play; Extended Kalman filters; Game theory; Distributed optimisation
1. Introduction
Recent advances in industrial automation technology often require distributed optimisation in a multi-agent system
where each agent controls a machine. An application of particular interest, addressed in this paper through a game-
theoretic approach, is the coordination of robot teams. Teams of robots can be used in many domains such as mine
detection [42], medication delivery in medical facilities [11], formation control [27] and exploration of unknown
environments [31, 32, 20]. In these cases teams of intelligent robots should coordinate in order to accomplish a desired
task. When autonomy is a desired property of a multi-robot system then self-coordination is necessary between the
robots of the team. Applications of these methodologies also include wireless sensor networks [21, 17, 41, 18], smart
grids [37, 2], water distribution system optimisation [40] and scheduling problems [34].
Game theory has also been used to design optimal controllers when the objective is coordination, see e.g. [30].
Using this approach the agents/robots eventually reach the Nash equilibrium of a coordination game. In [30, 29] local
and global components in the agents' cost function are used and the Nash equilibrium of the game is reached. Another
approach is presented in [3], based on agents' cost functions, which use local components and the assumption that the
states of the other agents are constant.
Fictitious play is an iterative learning process where players choose an action that maximises their expected re-
wards based on their beliefs about their opponents' strategies. The players update these beliefs after observing their
opponents' actions. Even though fictitious play converges to the Nash equilibrium for certain categories of games
[28, 22, 24, 23, 13], this convergence can be very slow because of the assumption that players use a fixed strategy
Email addresses: [email protected] (M. Smyrnakis), [email protected] (S. Veres)
1
6
1
0
2
v
o
N
7
1
]
A
M
.
s
c
[
1
v
8
3
6
5
0
.
1
1
6
1
:
v
i
X
r
a
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
2
in the whole game [13]. Speed up of the convergence can be facilitated by an alternative approach, which was pre-
sented in [33], where opponents' strategies vary through time and players use particle filters to predict them. Though
providing faster convergence, this approach has the drawback of high computational costs of the particle filters. In
applications where the computational cost is important, as the coordination of many UAVs, the particle filters ap-
proach is intractable. An alternative, we propose here, is to use extended Kalman filters (EKF) instead of predicting
opponents' strategies using particle filters. EKFs have much smaller computational costs than the particle filter variant
of fictitious play algorithm that has been proposed in [33]. Moreover in contrast to [33] we provide a proof of conver-
gence to Nash equilibrium of the proposed learning algorithm for potential games. Potential games are of particular
interest as many distributed optimisation tasks can be cast as potential games. Therefore convergence of an algorithm
to the the Nash equilibrium of a potential game is equivalent to convergence to the global or local optimum of the
distributed optimisation problem.
Thus the proposed learning process can be used as a design methodology for cooperative control based on game
theory, which overlaps with solutions in the area of distributed optimisation [8]. Each agent i strives to maximise a
global control reward, as negative of the control cost, through minimising its private control cost, which is associated
with the global one. The private cost function of an agent i incorporates terms that not only depend on agent i, but also
on costs associated with the actions of other agents. As the agents strive to minimise a common cost function through
their individual ones, the problem addressed here can also be seen as a distributed optimisation problem. In this work
we enable the agents to learn how they minimise their cost function through communication and interaction with other
agents, instead of finding the Nash equilibrium of the game, which is not possible in polynomial time for some games
[10]. In the proposed scheme robots learn and change their behaviour according to the other robots' actions. The
learning algorithm, which is based on fictitious play [7], serves as the coordination mechanism of the controllers of
team members. Thus, in the proposed cooperative control methodology there is an implicit coordination phase where
agents learn other agents' policies and then they use this knowledge to decide on the action that minimises their cost
functions. Additionally the proposed control module can be seen as a part of the BDI framework since each agent
updates his beliefs about his opponents' strategies given the state of the environment and based on a decision rule that
can represent his desires perform the selected actions
The remainder of this paper is organised as follows. We start with a brief description of relevant game-theoretic
definitions. Section 3 present some background material about Rational agents. Section 4 introduces the learning
algorithm that we use in our controller, Section 5 contains the main theoretical results and Section 6 contains the
simulation results in order to define the parameters of the proposed algorithm. Section 7 presents simulation results
before conclusions are drawn.
2. Game theoretical definitions
In this section we will briefly present some basic definitions from game theory, since the learning block of our
controller is based on these. A game Γ is defined by a set of players I, i ∈ {1, 2, . . . ,I}, who can choose an action,
si, from a finite discrete set S i. We then can define the joint action s, s = (s1, . . . , sI), that is played in a game as an
element of the product set S = ×i=I
i=1 S i. Each player i receives a reward, ri, after choosing an action si. The reward, also
called the utility, is a map from the joint action space to the real numbers, ri : S → R. We will often write s = (si, s−i),
where si is the action of player i and s−i is the joint action of player i's opponents. When players select their actions
using a probability distribution they use mixed strategies. The mixed strategy of a player i, σi, is an element of the
set ∆i, where ∆i is the set of all the probability distributions over the action space S i. The joint mixed strategy, σ, is
then an element of ∆ = ×i=I
∆i. A strategy where a specific action is chosen with probability 1 is referred to as pure
strategy. Analogously to the joint actions we will write σ = (σi, σ−i) for mixed strategies. The expected utility a
player i will gain if it chooses a strategy σi (resp. si), when its opponents choose the joint strategy σ−i, is denoted by
ri(σi, σ−i) (resp. ri(si, σ−i)).
i=1
A game, depending on the structure of its reward functions, can be characterised either as competitive or as a
coordination game. In competitive games players have conflicted interest and there is not a single joint action where
all players maximise their utilities. Zero sum games are a representative example of competitive games where the
reward of a player i is the loss of other players . An example of a zero-sum game is presented in Table 1. On the other
hand in coordination games players either share a common reward function or their rewards are maximised in the
same joint action. A very simple example of a coordination game where players share the same rewards, is depicted in
2
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
3
Head Tails
1,-1
-1,1
1,-1
-1,1
Head
Tails
Table 1. Rewards of two players in a zero sum game as function of the outcome of throwing a coin: matching pennies game.
L
U 1,1
D 0,0
R
0,0
1,1
Table 2. Rewards of two players in a simple coordination game as function of joint moves to the left & up (L,U) or right & down (R).
Table 2. Even though competitive games are the most studied games, we will focus our work on coordination games
because they naturally formulate a solution to distributed optimisation and coordination.
2.1. Best response and Nash Equilibrium
A common decision rule in game theory is best response. Best response is defined as the action that maximises
players' expected utility given their opponents' strategies. Thus for a specific mixed strategy σ−i we evaluate the best
response as:
σi
−i) = argmax
ri(si, σ
−i)
(1)
Nash in [25] showed that every game has at least one equilibrium. A joint mixed strategy σ = ( σi, σ−i) is called a
Nash equilibrium when
pure(σ
si∈S
(2)
Equation (2) implies that if a strategy σ is a Nash equilibrium then it is not possible for a player to increase its utility
by unilaterally changing its strategy. When all the robotic players in a game select their actions using pure strategies
then the equilibrium is referred as pure Nash equilibrium.
−i) ≥ ri(σi, σ
−i)
for any σi ∈ ∆i
ri( σi, σ
2.2. Optimisation tasks as potential games
It is possible to cast distributed optimisation tasks as potential games [1, 9], thus the task of finding an optimal
solution of the distributed optimisation task can be seen as the search of a Nash equilibrium in a game. An optimi-
sation problem can be solved distributively if it can be divided into D coupled or independent sub-problems with the
following property [4]:
r(s) − r( s) > 0 ⇔ ri(si) − ri( si) > 0, i ∈ I,∀s, s
(3)
where s and s are any sets of actions by the agents, r represents the global reward function or global utility, and
ri, i ∈ I, represent players' reward or utility. Equation (3) implies that a joint action s, should have the similarly
positive or negative impact in the global and the local task. Thus a solution s should increase or decrease both the
local and global utility when it is compared with another joint action s.
There is a direct analogy between (3) and ordinal potential games. Ordinal potential games are games where their
reward function have a potential function with the following property [23]
ri(si, s−i) − ri( si, s−i) > 0 ⇔ φ(si, s−i) − φ( si, s−i) > 0,∀s = (si, s−i), ∀ s = ( si, s−i)
(4)
where φ is a potential function and the above equality stands for every player i. Exact potential games (or potential
games thereafter), is a subclass of ordinal potential games which can be used to solve distributed optimisation prob-
lems where the difference in the global reward between two joint actions is the same as the difference in the potential
function [23]:
ri(si, s−i) − ri( si, s−i) = φ(si, s−i) − φ( si, s−i), ∀s = (si, s−i), ∀ s = ( si, s−i)
(5)
where similarly to ordinal potential games φ is a potential function and the above equality stands for every player i.
An advantage of potential games is that they have at least one pure Nash equilibrium, hence there is at least one joint
action s where no player can increase their reward, i.e. their potential function, through a unilateral change of action.
3
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
4
It is often feasible to choose appropriate forms of agents' utility functions and also the global utility in order to
enable the existence and use of a potential function of the system. In this paper we assume that all players/robots
share the same global reward. There are cases where, because of communication or other physical constraints, only
reward functions that are shared among groups of robots can be used. But even in these cases it is possible to create
reward functions that can act as potentials. Wonderful life utility is such a utility function. It was introduced in [38]
and applied in [1] to formulate distributed optimisation tasks as potential games. Player i's utility, when wonderful
life utility is used, can be defined as the difference between the global utility rg and the utility of the system when a
reference action is used as Player i's action. More formally when Player i chooses an action si one can define
ri(si) = rg(si, s−i) − rg(si
0, s−i),∀si, ∀s−i
(6)
where si
no action to represents the case when the player chooses to take no action.
0 denotes a reference action for player i. A reference action is introduced because in strategic games there is
3. Rational agent cooperation
3.1. Agent Definitions
By analogy to previous definitions [19? ? ? ] of AgentSpeak-like architectures, we define our agents as a tuple:
R = {F , B, L, Π, A}
(7)
where:
• F = {p1, p2, . . . , pnp} is the set of all predicates.
• B ⊂ F is the total set of belief predicates. The current belief base at time t is defined as Bt ⊂ B. Beliefs that are
added, deleted or modified can be either called internal or external depending on whether they are generated
from an internal action, in which case are referred to as "mental notes", or from an external input, in which case
they are called "percepts".
• L = {l1, l2, . . . lnl} is a set of logic-based implication rules.
• Π = {π1, π2, . . . , πnπ} is the set of executable plans or plans library.
Current applicable plans at time t are part of the subset applicable plan Πt ⊂ Π or "desire set".
• A = {a1, a2, . . . , ana} ⊂ F \ B is a set of all available actions. Actions can be either internal, when they modify
the belief base or fata in memory objects, or external, when they are linked to external functions that operate in
the environment.
AgentSpeak like languages, including the limited instruction set agent [? ], can be fully defined and implemented
by listing the following items:
• Initial Beliefs.
The initial beliefs and goals B0 ⊂ F are a set of literals that are automatically copied into the belief base Bt (that
is the set of current beliefs) when the agent mind is first run.
• Initial Actions.
The initial actions A0 ⊂ A are a set of actions that are executed when the agent mind is first run. The actions are
generally goals that activate specific plans.
The following three operations are repeated for each reasoning cycle.
• Maintenance of Percepts. This means generation of perception predicates for Bt and data objects such as the
world model used here W.
4
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
5
• Logic rules.
A set of logic based implication rules L describes theoretical reasoning to improve the agent current knowledge
about the world.
• Executable plans.
A set of executable plans or plan library Π. Each plan π j is described in the form:
p j : c j ← a1, a2, . . . , an j
(8)
where p j ∈ B is a triggering predicate, which allows the plan to be retrieved from the plan library whenever
it comes true, c j ∈ B is a logic formula of a context, which allows the agent to check the state of the world,
described by the current belief set Bt, before applying a particular plan sequence a1, a2, . . . , an j ∈ A with a list
of actions. Each a j can be one of (1) predicate of an external action with arguments of names of data objects,
(2) internal (mental note) with a preceding + or - sign to indicate whether the predicate needs to be added or
taken away from the belief set Bt (3) conditional set of items from (1)-(2). The set of all triggers p j in a program
is denoted by Etr
The reasoning cycle of LISA used in this paper consists of the following steps (Fig. 1):
1. Belief base update.
The agent updates the belief base by retrieving information about the world through perception and commu-
nication. The job is done by two functions, called Belief Update Function (BUF) and Belief Review Function
(BRF). The BUF takes care of adding and removing beliefs from the belief base; the BRF updates the set of
current events Et by looking at the changes in the belief base.
2. Application of logic rules.
The logic rules in L are applied in a round-robin fashion (restarting at the beginning of the list) until there are
new predicates generated for Bt. This means that rules need to be verified not to lead to infinite loops.
3. Trigger Event Selection.
For every reasoning cycle a function called Intention Set Function selects the current event set Et:
(9)
where ℘(·) is the so called power operator and represents the set of all possible subset of a particular set. We
will call the current selected trigger event S t(B) = Tt and the associated plans the Intention Set.
S t : ℘(Bt) → ℘(Et)
4. Plan Selection.
All the trigger plans in Tt are checked for their context to form the Applicable Plans set Πt,
S O : Et → ℘(S t)
(10)
We will call the current selected plan S O(Πt) = πt.
5. Plan Executions.
All plans in S O : Et are started to be executed concurrently by going through the plan items a1, a2, . . . , an j
one-by-one sequentially.
3.2. Cooperation cycle
The robots review their actions at a much slower rate than their reasoning cycle operates. During a cooperation
cycle each of the agents observes the other's actions, learns from it, defines a reward function relevant to the environ-
mental circumstances, performs optimisation to select its own action takes or continues the action undergoing. In most
practical situations the currently executed actions is one of the options for selection. The agent theory described in the
previous subsection can be used to make the process well coordinated among the agents using logic rules, perception
cycles and executable plans allocated as follows for each of the cyclical steps the agents need to perform collectively.
Slower than the reasoning cycle, is the asynchronous operational cycle. Its length in time can vary depending on
the agent and circumstances, it is asynchronous across the agent set as availability of communication is not guaranteed
5
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
6
Figure 1. Reasoning cycle of LISA: the plans run in a multi-threaded way, avoiding the need of S E and S I
between all the agents at all times. Each agent executes the following steps during its operational cycle, which is an
integer multiple No ≥ 3 of its reasoning cycle. Note that update of a simultaneous localisation and mapping-based
world model W. The following steps are carried out during the last of the reasoning cycles in the operational cycle.
This will allow enoug useful data to be
1. Agent i performs estimation of other agent's actions set s−i by analysing changes in W in time. The plan
associated with this is
estim agents : cooperate ← estim movements(W, Am), in f er activities(Am, S ni), Opt(R, S ni, Act), trigger(Act).
where Am is available movement description data for the rest of the agents in the team and S ni is a notation for
s−i. Each item in s ∈ s−i can be a pair describing the situation and the action the agent is taking. The reward
function ri(si, s−i) is represented by a data object R.
2. trigger(Act) adds a predicate pAct to Bt+1 to trigger the action plan associated with Act:
pAct : active(pAct) & goal(reached) ← per f orm(pAct).
Here the predicate goal(reached) can be established by logic inference in L during reasoning cycles involving
predicates generated by perception.
3. Finally there is a mission completion plan to be automatically executed when the mission is completed.
goal(reached) : mission started ← return to base(B),−misson started.
where return to base(B) is a repeated action of returning with collision avoidance to a base described by data
object B.
The above schema is generally applicable, can be particularised for its actions and can be added to make and agent
cooperative by fictitious play.
3.3. Example
Consider the following case of a UAV team as an example. Assume that one of the UAVs in the team suddenly
find itself with limited battery power and it can choose between two actions. One is that it quickly lands as this has
6
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
7
Figure 2. Vehicles should patrol five areas of a warehouse. Each area has two objects of different attribute and therefore different value. The
vehicles, according to their capabilities, should choose a joint action that will minimise their common cost function.
small control cost and can therefore can be performed safely. The second action is to pick up parcel first, which has
high control cost and can result in unsafe landing later. In addition consider the case where the action of landing will
affect negatively the performance of other UAVs and the team's mission will fail while picking up the parcel would
result in a successful performance of the desired joint task. The costs and rewards of this decision dilemma can be
formulated in terms of the above mentioned energy and environmental costs.
In order to ensure that the robot team will accomplish its mission the impact of the control cost in robots' decisions
should be smaller than the one of the environmental cost. This can be achieved either by setting wm << we or by setting
the cost of a mission failure to be significantly higher than the maximum control cost. This example is to point out
the variety of possible decisions that can be accommodated when reward functions are used to solve the distributed
optimisation tasks.
The following case study, which has also studied in simulation, shows how the proposed game theoretic decision
making mechanism. We consider a team of robots who are to coordinate their team in order to identify possible threats
in a warehouse, where we assume that there are N rooms with some hazardous items. The materials have different
attributes of the following categories: flammable, chemical, radioactive and security sensitive. In each room there can
be items that belong up to two different categories but there are constraints. For instance, are not allowed in the same
room flammable, radioactive and chemical materials together . Each of the I robots of the team is equipped with
sensors which can sense the different attributes and capabilities. Figure 2 depicts an example of this scenario. The
set of rooms correspond to the set of possible placement actions of each robot and is denoted by C. In each of these
delivery states there are control actions of the robots described by dynamical models. As an example we can consider
the task of moving towards the region of interest. Each robot i should choose one of the available rooms n, based on
the cost of the total control costs of moving to room n, but also the estimated decisions of the other robots and the
suitability of robot i to examine room n for its hazardous materials.
In order to create the performance model of the proposed decision making process we need to define a utility
function. Utility functions have been used as metrics of the robots coordination efficacy in applications, such as
[26, 43, 35, 6, 36]. Since the players of the game will maximise their expected reward the utility function can be
seen as the negative of a cost function. Therefore we can include in the utility any control costs of the robots and
environmental elements of the robots' tasks and the constraints that might arise from specific tasks. The function of
the control cost should take into account the spatial characteristics of the problem like the cost to the robot to move
towards to a specific position. The environmental part of the cost function takes into account costs that arose from the
nature of the coordination problem and can also include the quality of the sensors of a robot, the aptness of the robots
to perform specific tasks, etc.
We will use cooperative robot teams, robots who have different sensors and capabilities, but it is possible to use a
similar controller in swarms of robots that have identical specifications. The differences between the robots can also
be expressed in terms of their endurance in a specific environment, their efficiency to accomplish a specific task and
the presence of the correct sensor to identify a specific threat. Fuzzy variables can be used to quantify robots' efficacy.
Two fuzzy variables that can be used are sensors' quality and battery life. The values of battery life are short, fair,
long and values of sensors' quality are low, medium and high. If a robot is not equipped with a specific sensor then its
7
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
8
efficiency to detect an event is zero. Each robot, using a controller like the one that is depicted in Figure ??, should
then coordinate with the other robots in order to efficiently choose a region to patrol.
The robots should coordinate and choose a region which they will patrol based on the possible threats that should
be detected in each area, their sensors' specifications and the actions of the other robots. Each robot can choose only
one region to patrol, but many robots can choose the same area.
Each material in the warehouse has different significance and therefore each room has a different value depending
on the objects that are stored there which can be incorporated in Ei(·). In a game theoretic terminology we can define
a potential game Γ with I players which have N available actions. The utility function that is associated with room n
can be defined as:
(cid:88)
∀i,si=n
(cid:88)
∀si=n,n∈N
rn(s) =
rm
i (si) +
i (si, s−i).
re
where rm
robot's cost to move toward a specific area and re
i (si), is a function that depends on the initial position of the robot and the room n, n ∈ N and represent the
Logic constraints such as the fact that a robot should not visit an area which it has not the appropriate sensors to
patrol or that robots with the same attributes (sensors) should not choose the same area can be included in (25) using
i (si, s−i)
penalty terms in re
should have a negative or zero value. Similarly if two robots i and j with the same sensors choose the same area
j(s j, s− j) should also have a negative or zero value. The global utility that the robots will share is
i (si, s−i) and re
then re
defined as:
i (si, s−i), i.e if a robot i chooses to select an area which is not suitable of its sensors or then re
i (si, s−i) is the environmental part of the cost.
n=N(cid:88)
rg(s) =
rn(s)
(11)
4. The learning process
n=1
In this section we present a combination of fictitious play and extended Kalman filters as the algorithm that we
will use in the learning block of the proposed decision making module. We briefly present the classic fictitious play
algorithm and how it can be combined with extended Kalman filters in a decision making algorithm.
4.1. Fictitious play
Fictitious play [7], is a widely used learning technique in game theory. In fictitious play each player chooses his
action according to the best response of his beliefs about his opponent's strategy.
Initially each player has some prior beliefs about the strategy that his opponent uses to choose an action. These
(s j) denotes the weight function of Player i
beliefs are expressed by a weighting function κ
that Player j will play action s j at the tth iteration. The players, after each iteration, update the weight func-
tions, and therefore their beliefs, about their opponents' strategy and play again the best response to their beliefs.
More formally, in the beginning of a game players maintain some arbitrary non-negative initial weight functions
i→ j
κ
0
(s j), i = 1, . . . ,I , j ∈ {1, . . . ,I}\{i} that are updated using the formula [13]:
(s j). Where κ
i→ j
t
i→ j
t
(cid:40)
for each j, where Is j
t =s j =
1
0
t = s j
ifs j
otherwise.
i→ j
κ
t
(s j) = κ
i→ j
t−1 (s j) + Is j
t =s j
(12)
(cid:41)
.
Equation (12) suggests that all the observed actions have the same impact. Therefore players assume that their
opponents choose their actions using a fixed mixed strategy. It is natural then to use a multinomial distribution to
approximate an opponent's mixed strategy. The parameters of the multinomial distribution can be estimated using the
maximum likelihood method. The mixed strategy of opponent j is then estimated from the following formula:
i→ j
t
σ
(s j) =
i→ j
κ
t
s(cid:48)∈S j κ
(s j)
i→ j
t
(s(cid:48))
.
(13)
(cid:80)
8
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
9
4.2. Fictitious play as a state space model
A more realistic assumption than using (12) is to presume that players are intelligent and change their strategies
according to the other players' actions. We follow [33] and will represent the fictitious play process as a state-space
model. According to the state space model each player has a propensity Qi
t(si) to play each of their available actions
si ∈ S i, and then to form a strategy based on these propensities. Finally players can choose an action based on
their strategy and the best response decision rule. Because players have no information about the evolution of their
opponents' propensities, and under the assumption that the changes in propensities are small from one iteration of
the game to another, we model propensities using a Gaussian autoregressive prior on all propensities [33]. We set
Q0 ∼ N(0, I), where I is the identity matrix, and recursively update the value of Qt according to the value of Qt−1 as
follows:
Q(st) = Q(st−1) + ηt
(14)
where ηt ∼ N(0, χ2I).
Similarly to the sigmoid function that is widely used in the neural network literature [5], to relate the weights
and the observation layer of a neural network, we assume that propensities are connected with players' actions by the
following Boltzman formula for every s j ∈ S j
(cid:80)
e(Q j(s j)/τ)
s∈S j e(Qt( s)/τ) .
(15)
4.3. Kalman filters and Extended Kalman filters
Is j
t =s j =
Our objective is to estimate Player i's opponent propensity and thus to estimate the marginal probability p(Qt, s1:t).
This objective can be represented as a Hidden Markov Model (HMM). HMMs are used to predict the value of an un-
observed variable ct, the hidden state, using the observations of another variable z1:t. There are two main assumptions
in the HMM representation. The former one is that the probability of being at any state ct at time t depends only at
the state of time t − 1, ct−1. The latter one is that an observation at time t depends only on the current state ct. One
of the most common methods to estimate p(c1:t, z1:t) is Kalman filters and its variations. Kalman filter [16] is based
on two assumptions, the first is that the state variable is Gaussian. The second is that the observations are the result
of a linear combination of the state variable. Hence Kalman filters can be used in cases which are represented as the
following state space model:
ct = Act−1 + ξt−1 hidden layer
yt = Bct + ζt observations
(16)
where ξt and ζt follow a zero mean normal distribution with covariance matrices Ξ = qtI and Z = rtI respectively, and
A, B are linear transformation matrices. When the distribution of the state variable ct is Gaussian then p(cty1:t) is also
a Gaussian distribution, since yt is a linear combination of ct. Therefore it is enough to estimate its mean and variance
to fully characterise p(cty1:t).
Nevertheless in the state space model we want to implement, the relation between Player i's opponent propensity
and his actions is not linear (15). Thus a more general form of state space model should be used, such as:
ct = f (ct−1) + ξt
yt = h(xt) + ζt
(17)
where f (·) and h(·), are non-linear functions, ξt and ζt are the hidden and observation state noise respectively, with
zero mean and covariance matrices Ξ = qtI and Z = rtI respectively. The distribution of p(cty1:t) is not a Gaussian
distribution because f (·) and h(·) are non-linear functions. Extended Kalman filter (EKF) provides a simple method to
overcome this shortcoming by using a first order Taylor expansion to approximate the distributions of the sate space
model in (17). In particular if we let ct = mt−1 + , where mt denotes the mean of ct and ∼ N(0, P), we can rewrite
(17) as:
ct = f (mt−1 + ) + wt−1 = f (mt−1) + Fc(mt−1) + ξt−1
yt = h(mt + ) + ζt = h(mt) + Hc(mt) + ζt
(18)
where Fc(mt−1) and Hc(mt) is the Jacobian matrix of f and h evaluated at mt−1 and mt, respectively. If we use the
transformations of (18) then p(cty1:t) is a Gaussian distribution.
9
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
10
Since p(cty1:t) is a Gaussian distribution to fully characterise it we need to evaluate its mean and its variance.
The EKF process [15, 14] estimates this mean and variance in two steps the prediction and the update step. In the
prediction step at any iteration t the distribution of the state variable is estimated based on all the observations until
time t − 1, p(cty1:t−1). The distribution of p(cty1:t−1) is Gaussian and we will denote its mean and variance as m−
t
and P−
t respectively. During the update step the estimation of the prediction step is corrected in the light of the new
observation at time t, so we estimate p(cty1:t). This is also a Gaussian distribution and we will denote its mean and
variance as mt and Pt respectively.
The estimate of the state c for and its variance prediction for p(cty1:t−1) and p(cty1:t) are evaluated based on the
prediction and update steps of the EKF process [15, 14] respectively as follows:
Prediction Step
¯ct = f (ct−1)
P−
t =F(mt−1)Pt−1F(mt−1) + Wt−1
where ¯ct and ct−1 are the estimates of c for p(cty1:t−1) and p(ct−y1:t−1) respectively, and the j, j(cid:48) element of F(mt) is
defined as
[F(m−
t )] j, j(cid:48) =
∂ f (c j, r)
∂c j(cid:48)
c=m−
t ,q=0
Update Step
t HT (¯ct) + Z
vt = zt − h(¯ct)
S t = H(¯ct)P−
Kt = P−
ct = ¯ct + Ktvt
Pt = P−
t HT (¯ct)S −1
t − KtS tKT
t
t
where zt is the observation vector and the ( j, j)(cid:48) element of H(ct) is defined as:
[H(m−
t )] j, j(cid:48) =
∂h(c j, r)
∂c j(cid:48)
c=m−
t ,r=0
4.4. Fictitious play and EKF
For the rest of this paper we will only consider inference over a single opponent mixed strategy in fictitious play.
Separate estimates will be formed identically and independently for each opponent. We therefore consider only one
opponent, and we will drop all dependence on player i, and write st, σt and Qt for Player i's opponent's action,
strategy and propensity respectively. Moreover for any vector x, x[ j] will denote the jth element of the vector and for
any matrix y, y[i, j] will denote the (i, j)th element of the matrix.
We can use the following state space model to describe the fictitious play process:
Qt
Is j
t =s j
= Qt−1 + ξt−1
= h(Qt) + ζt
(19)
where ξt−1 ∼ N(0, Ξ), is the noise of the state process and ζt ∼ N(0, Z) is is the error of the observation state with
zero mean and covariance matrix Z, which occurs because we approximate a discrete process like best response (1),
using a continuous function h(·), where hs(k)(Q) = exp(Qt[s(k)]/τ)
s∈S exp(Qt[ s]/τ), where τ is a temperature parameter. Hence we can
combine the EKF with fictitious play as follows.
At time t − 1 Player i has observed action st−1 and based on the update step of the EKF process has an estimate
of his opponent's propensity, Qt−1 with the variance Pt−1. Then at time t he uses EKF prediction step to estimate his
opponent's propensity. The estimate and its variance are:
(cid:80)
Qt−1
¯Qt =
P−
t = Pt−1 + Ξ
10
(20)
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
11
t − 1, Q j
t−1, with covariance P j
2. Agent i predicts his opponents' propensities ¯Q j
1. At time t agent i maintains estimates of his opponent's propensities up to time
tk, j ∈ {1, ...,I}\{i}, k ∈ S j using
3. Based on the propensities in 2 each agent updates his beliefs about his oppo-
t−1 of his distribution.
(20).
nents' strategies using (21).
4. Agent i chooses an action based on the beliefs in 3 and applies best response
decision rule.
5. The agent i observes his opponents' action s j
tk,
6. The agent update its estimates of all of its opponents' propensities using ex-
j ∈ {1, ...,I}\{i}.
tended Kalman Filtering to obtain Q j
t ,
j ∈ {1, ...,I}\{i}.
Table 3. EKF based fictitious play algorithm.
Player i then evaluates his opponents strategies using his estimations as:
exp( ¯Q(st = k)/τ)
s∈S exp( ¯Q( s)/τ)
σt(st = k) =
(cid:80)
.
(21)
Player i then uses the estimate of his opponent strategy (21) and best responses (1), to choose an action. After
observing his opponent's action st, Player i correct his estimate about his opponent's propensity using the update
equations of EKF process. The update equations are:
Table 3 summarises the fictitious play algorithm when EKF is used to predict opponents strategies.
4.5. EKF fictitious play as part of the reasoning cycle of Jason
As described in Section 3.2 robots actions and reasoning cycles are not synchronised. After each action of the
robot team the utility function can be updated and a new reasoning cycle that will include the EKF fictitious play
learning algorithm can be started. In particular steps 1 and 2 of the the reasoning cycle of Jason, (Section 3.1), can
be updated using the proposed learning process. The update of the propensities can be seen as Belief base update,
since when agents update their propensities, they update their beliefs about other agents actions for a given state of the
environment. Then the action selection of EKF fictitious play is the Trigger Event Selection. Note here that instead of
Best Response other alternatives can be used with EKF fictitious play algorithm as part of the Trigger Event Selection
such as Smooth Best Response [13].
5. The Main Results
In this section we present our convergence results for games with at least one pure Nash equilibrium. The results
are for the EKF fictitious play algorithm of Table 3, when the covariance matrices Ξ and Z are defined as Ξ = (ξ + )I
11
t HT ( ¯Qt) + Z
vt = Is j
S t = H( ¯Qt)P−
Kt = P−
Qt =
Pt = P−
t =s j − h( ¯Qt)
t HT ( ¯Qt)S −1
¯Qt + Ktvt
t − KtS tKT
t
t
The Jacobian matrix H( ¯Qt) is defined as
[H( ¯Qt)] j, j(cid:48) =
(cid:80)
((cid:80)
j(cid:44) j(cid:48) exp( ¯Qt[ j]) exp( ¯Qt)
((cid:80)
− exp( ¯Qt[ j]) exp( ¯Qt[ j(cid:48)])
j exp( ¯Qt[ j]))2
j exp( ¯Qt[ j]))2
if j = j(cid:48)
if j (cid:44) j'
.
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
12
and Z = (1/t)I respectively, where ξ is a constant, is an arbitrarily small Gaussian random variable, ∼ N(0, Ψ), t is
the tth iteration of fictitious play, and I is the identity matrix.
The EKF fictitious play algorithm has the following properties:
Proposition 1. If at iteration t of the EKF fictitious play algorithm, action k is played from Player i's opponent, then
the estimate of his opponent propensity to play action k increases, Qt−1[k] < Qt[k]. Moreover if we denote ∆[i] as
∆[i] = Qt[i]− Qt−1[i], then ∆[k] > ∆[ j]∀ j ∈ S j, where S j is the action space of the jth opponent of Player i. Therefore
since(cid:80)
k∈S j σ(s j = k) = 1, σ
j
t will be also increased.
Proof. The proof of Proposition 1 is on Appendix Appendix A.
Proposition 1 implies that player i, when he uses EKF fictitious play, learn his opponent's strategy and eventually
they will choose the action that will maximise their reward base on their estimation. Nevertheless there are cases
where players may change their action simultaneously and become trapped in a cycle instead of converging in a pure
Nash equilibrium. As an example we consider the symmetric game that is depicted in Table 2. This is a simple
coordination game with two pure Nash equilibria the joint actions (U, L) and (D, R). In the case were the two players
start from joint action (U, R) or (D, L) and they always change their action simultaneously then they will never reach
one of the two pure Nash equilibria of the game.
Proposition 2. When the players of a game Γ use EKF fictitious play process to choose their actions, then with high
probability they will not change their action simultaneously infinitely often.
Proof. The proof of Proposition 2 is on Appendix Appendix B.
Based on Proposition 1 and 2 we can infer the following propositions and theorem.
Proposition 3. (a) If s is a pure Nash equilibrium of a game Γ, and s is played at date t in the process of EKF fictitious
play, s will be played at all subsequent dates. That is, strict Nash equilibria are absorbing for the process of EKF
fictitious play.
(b) Any pure strategy steady state of EKF fictitious play must be a Nash equilibrium.
Proof. Consider the case where players beliefs σt, are such that their optimal choices correspond to a strict Nash equi-
librium s. In EKF fictitious play process players' beliefs are formed identically and independently for each opponent
based on equation (21). By Proposition 1 we know that players' estimations about their opponents' propensities and
therefore their strategies will increase for the actions that are included in s. Thus the best response to their beliefs σt+1
will be again s and since s is a Nash equilibrium they will not deviate from it. Conversely, if a player remains at a
pure strategy profile, then eventually the assessments will become concentrated at that profile, because of Proposition
1 and so if the profile is not a Nash equilibrium, one of the players would eventually want to deviate.
Proposition 4. Under EKF fictitious play, if the beliefs over each player's choices converge, the strategy profile
corresponding to the product of these distributions is a Nash equilibrium.
Proof. Suppose that the beliefs of the players at time t, σt, converges to some profile σ. If σ were not a Nash equi-
librium, some player would eventually want to deviate and the beliefs would also deviate since based on Proposition
1 players eventually learn their opponents actions.
Based on the propositions (1-4) we can show that EKF fictitious play converges to the the Nash equilibrium of
games with a better reply path. A game with a better reply path can be represented as a graph were its edges are the
join actions of the game s and there is a vertex that connects s with s(cid:48) iff only one player i can increasing his payoff
by changing his action [39]. Potential games have a better reply path [39].
Theorem 1. The EKF fictitious play process converges to the Nash equilibrium in games with a better reply path.
12
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
13
Proof. If the initial beliefs of the players are such that their initial joint action s0 is a Nash equilibrium, from Propo-
sition 3 and equation (21), we know that they will play the joint action which is a Nash equilibrium for the rest of the
game.
Moreover, the initial beliefs of the players are such that their initial joint action s0 is not a Nash equilibrium based
on Proposition 1 and Proposition 2 after a finite number of iterations because the game has a better reply path the only
player that can improve his pay-off by changing his actions will choose a new action which will result in a new joint
action s. If this action is not the a Nash equilibrium then again after finite number of iterations the player who can
improve his pay-off will change action and a new joint action s(cid:48) will be played. Thus after the search of the vertices
of a finite graph, and therefore after a finite number of iterations, players will choose a joint action which is a Nash
equilibrium. Eventually after a finite number of time steps, T, the process will end up in a pure Nash equilibrium. The
maximum number of iterations that is needed is the cardinality of the joint action set multiplied with the total number
of iterations that is needed in order not to have simultaneous changes,
(t1 + t2 + t3 + . . . + T)
(cid:16)n
(cid:17)
2
6. Parameters selection
The covariance matrix of the state space error Ξ = ξI and the measurement error Z = ζI are two parameters that
we should define in the beginning of the EKF fictitious play algorithm and they affect its performance. Our aim is to
find values, or range of values, of ξ and ζ that can efficiently track opponents' strategy when it smoothly or abruptly
change, instead of choosing them heuristically for each opponent when we use the EKF algorithm. Nevertheless it
is possible that for some games the results of the EKF algorithm will be improved for other combinations of ξ and ζ
than the ones that we propose in this section.
+1
We examine the impact of EKF fictitious play algorithm parameters in its performance, by measuring the mean
square error (MSE) between EKf estimation and the real opponent strategy, in the following two tracking scenarios.
In the former scenario a single opponent chooses his actions using a mixed strategy which changes smoothly and
has a sinusoidal form over the iterations of the tracking scenario. In particular for t = 1, 2, . . . , 5000 iterations of the
= 1− σt(2), where n = 100. In the second toy example Player i's opponent change his strategy
game: σt(1) = cos 2πt
n
2
abruptly and chooses action 1 with probability σ2
t (1) = 1 during the first and the last 1250 iterations' of the game and
for the rest iterations of the game σ2
t (1).
Figure 3 depicts the mean square error for both tracking scenarios for 0.005 ≤ ξ ≤ 0.5 and 0.005 ≤ ζ ≤ 0.5.
Additionally we examine the performance of EKF fictitious play for ζ = 1/t which is depicted as the last element of
the x-axis of Figure 3. There is a wide range of combinations of ξ and ζ that minimise the MSE where the mean for
both examples. The MSE is minimised for a wider range of ξ when ζ = 1/t than the all the cases where a single value
of ζ was used. Moreover the minimum value of the MSE was observed when ζ = 1
t and ξ = 0.1. Figures 4 and 5
depict the opponent's strategy tracking of EKF fictitious play in both examples for these parameters.
t (1) = 0. The probability of the second action is calculated as: σ2
t (2) = 1 − σ2
7. Simulations
This section contains the results of two simulation scenarios, on which the performance of the proposed learning
algorithm was tested. Based on the results of the previous section, we used the following values for Ξ and Z in our
simulations, Ξt = (ξ + )I and Zt = ζI, where I is the identity matrix, ξ = 0.1, ζ = 1/t and is an arbitrarily small
Gaussian random number. For both simulation scenarios we report the average results for 100 replications of each
learning instance. In each learning instance the agents were negotiating for 50 iterations.
7.1. Results in symmetric games
We initially tested the performance of the proposed algorithm in symmetric games as they are presented in Tables
2 and 4. Even though these are games that involve only two players, they can be ensued from realistic applications.
The game of Table 2 can be seen as a collision avoidance game. Consider the case where two UAVs, the players of the
game, should fly in opposite directions. They will gain some reward if they accomplish their task and fly in different
altitudes. Therefore players should coordinate and choose one of the two joint actions that maximise their reward, UL
and DR, who represent the choices of different altitudes from the UAVs.
13
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
14
Figure 3. Contour plot of the combined mean square error for the pure and mixed strategy case. The range of ξ and ζ is [0.005, 0.5] . The last value
of x-axis corresponds to ζ = 1/t the
Figure 4. Tracking of opponent strategy at a pure strategy environment, where the pre-specified strategy of the opponent and its prediction are
depicted as the red and blue line respectively.
Weak
Fair
Strong
Weak
1,1
0,0
0,0
Fair
0,0
1,1
0,0
Strong
0,0
0,0
1,1
Table 4. Players' rewards of a symmetric game.
14
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
15
Figure 5. Tracking of opponent strategy at a mixed strategy environment, where the pre-specified strategy of the opponent and its prediction are
depicted as the red and blue line respectively.
Another example of how a symmetric game can be ensued from a realistic application comes from the area of
material handling robots. Consider the case where two robots should coordinate in order to move some objects to a
desired destination. The robots can either push or pull the objects depending on the direction of the destination of the
object. Moreover, each robot can apply different forces to the objects. The amount of the force that a robot applies to
an object can be represented as a fuzzy variable that can take the values: Weak, Fair and Strong. This scenario can
lead to the game that is described in Table 4.
Since in this paper we focus on the decision making module of a robot's controller, we will compare the proposed
learning algorithm of EKF fictitious play with the particle filter alternative [33]. We will use it as a baseline criterion
for the greedy choice of the players, i.e the action that maximises the reward of a player independently of what the
others are doing. In particular for the games that are presented in Tables 2 and 4 this is the random action since the
rewards for all the actions are the same.
Figures 6 and 7 depict the percentage of the replications where the algorithms converged to one of the optimal
solutions for the games depicted in Tables 2 and 4 respectively. Because of the symmetry in the utility function,
the particle filter fictitious play behaves similarly to the the random action. Thus when the initial beliefs were not
supporting one of the pure Nash equilibria of the games, it failed to converge to joint action with reward. EKF
fictitious play on the other hand, always converged to a joint action that maximises players rewards. Figure 8 shows
the average utility that players obtain in each iteration of the game when the initial joint actions have zero reward. As
we can observe, after the maximum of 35 iterations, the utility that players receive is 1 and therefore it converges to
one of the pure Nash equilibria.
7.2. Results in a robots team warehouse patrolling task
We also examined the performance of EKF fictitious play in the task allocation scenario we described in Section
2. A robot team, which consists of I robots, should examine an area A that is divided in N regions with hazardous
objects. We present simulation results for various combinations of number of regions and sizes of the robot's team.
In particular, we examined the performance of EKF learning algorithm when the robot's team consists of I robots,
I ∈ {5, 10, 15, 20, 25, 30, 35, 40}, who should patrol one of the N regions, N ∈ {5, 10, 20}.
Each region contains two objects with different attributes. Each robot i then can be equipped with up to three of
the following sensors: fire detector, chemical detector, Geiger counter and vision system. Therefore robots with a fire
detector should patrol areas with flammable objects, robots with Geiger should patrol areas with radioactive material
etc. Moreover each robot can move towards a region n with a velocity Vin. In our case study we assume that the
15
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
16
Figure 6. Percentage of the times that each algorithm converged to one of the two optimum joint actions for the game of Table 2.
Figure 7. Percentage of the times that each algorithm converged to one of the three optimum joint actions for the game of Table4.
16
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
17
Figure 8. Average reward that players obtain when their initial action is either UR orDL for the game of Table 2 .
Sensors' quality
Medium
Low
High
Battery capacity
Short
0.2
0.3
0.5
Fair Long
0.3
0.5
07
0.5
0.7
0.9
Table 5. Efficiency that a robot efficiently patrol the area for a specific threat.
rg(s) = −(cid:80)N
+(cid:80)N
(cid:80)
(cid:80)
i:si=n c1Tin + c2Vin
k∈Kn Vnk(1 −(cid:81)
velocity of a robot i, towards a region n can be either slow, medium or fast. Therefore the time that a robot needs to
reach a region n, Tin, depends both in the distance between the robot and the region and the velocity that the robot will
choose to move towards n as well. The global reward that is shared among the robots is defined as:
(22)
where c1 = 1, c2 = 1/6 and Eink is a metric of robot i's efficiency to patrol a region n based on its battery capacity and
its sensor quality. It can also be seen as the probability that robot i has to detect a threat k in region n. We define Eink
as a function of robot i's battery capacity and its sensors' quality as it is presented in Table 5. If a robot has no suitable
sensor to patrol a specific area Eink = 0. Note that the first term in he summation represents the control cost the robots
have when they choose a specific region to allocate and the second term corresponds to the environmental cost.
i:si=n(1 − Eink))
n=1
n=1
Even though this article considers distributed optimisation tasks, a centralised alternative, genetic algorithm (GA),
will be also used. Since the centralised optimisation algorithms are expected to perform equally good or better than
distributed optimisation algorithms, the performance of GA can be seen as a benchmark solution. Thus a central
decision maker will choose in advance the area that each robot will patrol and the robot completes the pre-optimised
allocation task. In our comparison we used the genetic algorithm function of Matlab's optimisation toolbox. Since
genetic algorithms are stochastic processes, in each repetition of the same replication of the task allocation problem it
will converge to a different solution. Thus we used two instances of the genetic algorithm, in the former one we used
a single repetition of the genetic algorithm per replication of our problem, in the latter one for each replication of the
task allocation problem we chose the allocation of the maximum utility among 50 runs of the genetic algorithm.
players obtained for each learning algorithm they used: F = 100 × rg(s)
have obtained if all all areas had been patrolled with 100% efficiency.
In order to be able to average across the 100 replications, we used the following scores for the global rewards
, where rmax is the reward that players would
rmax
17
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
18
Figure 9. Robot team task allocation outcome for the case where N = 5. The x-axis represent the number of robots that was used in the simulation
scenario and the y-axis the corresponding average score of the 100 replications of the allocation task.
Figures 9, 10 and 11 depict the score players obtain in the final iteration of the algorithm as a function of the
number of robots, I, for the tested algorithms when there are 5,10 and 20 areas of interest respectively.
In all cases when players chose the greedy actions, the outcome was very poor, which also indicates the need of
the coordination controller that we propose when distributed solutions are considered. When the case with 5 areas
is considered, Figure 9, all algorithms performed similarly, when I ≤ 30. When I ≥ 35 the best performance was
observed when the proposed algorithm was used. When the searching areas were increased to 10, Figure 10, the EKF
algorithm and the variants of the genetic algorithm have similar performance independently of the number of robots.
But its performance is better than the particle filters algorithm when I ≥ 20. Finally when we consider the case with
20 search areas EKF fictitious play perform significantly better than both variants of the genetic algorithms we used.
But its score is 3% smaller than the one of particle filter alternative when we consider the case where 40 robots had to
coordinate.
We also test the computational cost, in seconds of the 50 iterations per agent of the algorithms we tested. For
the genetic algorithms, since they are centralised algorithms, the cost per robot is the cost the algorithm needed to
terminate. In order to have comparable results with the centralised cases, for the two distributed algorithms it is
assumed that robots are able to communicate. Therefore, before they take an action there is a "negotiation" period
among the robots. During this period the robots exchange messages indicating the areas, which they intent to patrol.
Then, when the final iteration of the negotiation period is reached, they execute the last joint action. Under this
process the difference in the computational time of each robot is the total computational time of each algorithm. The
experiments were performed on a PC with Intel i7 processor at 2.5 GHz processors and 8 GB memory.
Even though the performance of the proposed algorithm is not always improving the results of the existing algo-
rithms we tested, it has the smallest computational cost among them, as it is depicted in Tables 6,7 and 8. This is
important in robotics applications where restrictions, such as battery life, should be taken into account.
7.3. Transportation of a heavy object through corridors.
In this section we describe how the proposed decision making module can be implemented as a coordination
mechanism between two robots which carry a heavy item through a set of possible unmapped corridors. Based on
a game-theoretic formulation, after the robots sense that an object or wall block their way, they have to coordinate
in order to choose the action that will allow them to continue their mission even when they are not aware of the
18
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
19
Figure 10. Robot team task allocation outcome for the case where N = 10. The x-axis represent the number of robots that was used in the
simulation scenario and the y-axis the corresponding average score of the 100 replications of the allocation task.
Figure 11. Robot team task allocation outcome for the case where N = 20. The x-axis represent the number of robots that was used in the
simulation scenario and the y-axis the corresponding average score of the 100 replications of the allocation task.
19
5
10
15
20
25
30
35
40
5
10
15
20
25
30
35
40
5
10
15
20
25
30
35
40
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
20
One-shot Genetic algorithm Genetic algorithm Particle filter fictitious play EKF fictitious play
1.097
2.081
2.255
2.358
2.366
2.698
2.728
2.839
56.9
105.9
110.6
117.5
121.5
159.9
132.1
136.7
0.185
0.399
0.616
0.886
1.089
1.374
1.558
1.917
0.159
0.165
0.246
0.326
0.428
0.503
0.589
0.660
Table 6. Average computational time that each robot was needed per replication, 5 areas
One-shot Genetic algorithm Genetic algorithm Particle filter fictitious play EKF fictitious play
2.056
4.353
4.318
3.296
4.313
5.092
5.275
6.023
106.9
190.5
199.85
240.25
246.75
248.5
255.5
296.0
0.628
1.165
1.881
2.579
3.208
3.762
4.445
5.271
0.674
0.705
1.105
1.704
2.119
3.143
3.899
4.487
Table 7. Average computational time that each robot was needed per replication, 10 areas
One-shot Genetic algorithm Genetic algorithm Particle filter fictitious play EKF fictitious play
3.455
12.023
20.141
20.681
22.014
22.211
26.146
27.276
136.9
556.9
749.2
870.8
875.1
909.8
911.6
944.5
0.826
1.906
2.932
3.276
4.955
6.138
6.754
8.399
0.706
1.105
1.704
2.119
2.229
3.143
3.899
4.487
Table 8. Average computational time that each robot was needed per replication, 20 areas
20
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
21
Figure 12. The configuration the corridor used in our simulation in a MATLABT M figure and illustration of robot moves.
environment's structure. We assume that the two robots can move in corridors using the moves that are depicted in
Figure 13. In order to take into account obstacles and varying widths of corridors we allowed two possible ways
to turn left or right. In particular Action 1 represents the move forward, Action 2 represents the move forward for a
distance z meters and right z meters, Action 3 is a z meter diagonal right move of 45 degrees slope, Action 4 represents
the move forward for z meters and left z meters and Action 5 is a z meter diagonal left move of 45 degrees slope. This
also introduces the need of a coordination mechanism because there can be cases, such as the set of corridors which
are depicted in Figure 12, were both moves can be used by the robots to change direction.
Figure 13. Available moves to each robot
Robots in this scenario have to make decisions in various areas of Figure 12. Each of these areas is assumed to
be a checkpoint, at which when the robots arrive, they use the EKF algorithm in order to choose the joint action that
will lead them in the next checkpoint. The checkpoints can either be predefined before the task starts, or it can be
dynamically generated based on set time intervals or the distance that the robots have covered. In this game when the
robots fail to coordinate, they receive zero utility, and when they coordinate the reward their gain is a constant c if the
joint action is feasible and zero otherwise. Thus in each checkpoint the robots play a game with rewards defined as:
(cid:40)
r(s) =
c
0
if the joint action s is feasible and s1 = s2
otherwise
(23)
A joint action s is considered feasible when it is executed and the task can proceed, i.e. during the execution of the
action neither the robots nor the hit a wall.
We present the results over 100 replications of the above coordination task. In this task, because of the symmetry
of the game, when the robots use the particle filter variant of fictitious play, fail to coordinate the majority of the times.
21
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
22
Figure 14. Trajectory of the robots through the checkpoints.
In particular only in 34% of the replications of the coordination tasks were successfully completed when the particle
filters algorithm was used. On the other hand, when the proposed algorithm was used, the task was completed in all
replications.
Figure 14 displays a sample trajectory of the robots for the dynamically generated checkpoints. In this simulation
scenario we assume that 0.1 distance units corresponds to 10 meters. The checkpoints are generated when robots
move 5 meters from the previous checkpoint, i.e. z = 5. This can be seen also as a new checkpoint every 2 seconds
when the robots are moving with speed of 2.5m/s.
Table 9 shows the joint actions who can produce some reward (successful joint action) for each check point and
the average number of iterations the EKF fictitious play algorithm needed to converge to a solution.
7.4. Ad-hoc sensor network
The simulation scenario of this section is based on a coordination task of a power constrained sensor network,
where sensors can be either in a sensinf or sleeping mode [12]. When the sensors are in a sensing mode, they can
observe the events that occur in their range. During their sleeping mode the sensors harvest the energy they need in
order to be able function when they are in the sensing mode. The sensors then should coordinate and choose their
sensing/sleeping schedule in order to maximise the coverage of the events. This optimisation task can be cast as a
potential game. In particular we consider the case where I sensors are deployed in an area where E events occur. If
an event e, e ∈ E, is observed from the sensors, then it will produce some utility Ve. Each of the sensors i = 1, . . . ,I
should choose an action si = j, from one of the j = 1, . . . , J time intervals which they can be in sensing mode. Each
sensor i, when it is in sensing mode, can observe an event e, if it is in its sensing range, with probability pie = 1
,
die
where die is the distance between the sensor i and the event e. We assume that the probability that each sensor observes
an event is independent from the other sensors' probabilities. If we denote by iin the sensors that are in sense mode
when the event e occurs and e is in their sensing range, then we can write the probability an event e to be observed
from the sensors, iin as:
(24)
1 −(cid:89)
i∈iin
(1 − pie)
The expected utility that is produced from the event e is the product of its utility Ve and the probability it has to be
observed by the sensors, iin that are in sensing mode when the event e occurs and e is in their sensing range. More
formally, we can express the utility that is produced from an event e as:
(1 − pie))
re(s) = Ve(1 −(cid:89)
(25)
i∈iin
22
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
23
Checkpoint 1
Checkpoint 2
Checkpoint 3
Checkpoint 4
Checkpoint 5
Checkpoint 6
Checkpoint 7
Checkpoint 8
Checkpoint 9
Checkpoint 10
Checkpoint 11
Checkpoint 12
Checkpoint 13
Checkpoint 14
Checkpoint 15
Checkpoint 16
Checkpoint 17
Checkpoint 18
Checkpoint 19
Checkpoint 20
Checkpoint 21
Checkpoint 22
Checkpoint 23
Checkpoint 24
Checkpoint 25
Checkpoint 26
Checkpoint 27
Successful joint actions
Iterations
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
1
1
1
(Action 2,Action 2),(Action 3,Action 3)
25.6
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
1
1
1
1
1
1
1
(Action 4,Action 4),(Action 5,Action 5)
29.8
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
1
1
1
1
(Action 2,Action 2),(Action 3,Action 3)
22.6
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
1
1
1
(Action 2,Action 2),(Action 3,Action 3)
27.9
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
(Action 1,Action 1)
1
1
1
1
1
1
Table 9. Average number of iterations that was needed by EKF fictitious play algorithm in order for the robots to reach consensus.
23
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
24
Figure 15. Average reward as a function of time of the three learning algorithms. The x-axis represents the Iteration and the y-axis the average
normalised reward.
The global utility is then the sum of the utilities that all events, e ∈ E, produce
rglobal(s) =
re(s).
(26)
(cid:88)
e
(cid:88)
Each sensor, after each iteration of the game, receives some utility, which is based on the sensors and the events that
are inside its communication and sensing range respectively. For a sensor i we denote e the events that are in its
sensing range and s−i the joint action of the sensors that are inside his communication range. The utility that sensor i
will receive if his sense mode is j will be
ri(si = j, s−i) =
re(si = j, s−i).
(27)
e
In this simulation scenario 40 sensors are placed on a unite square, each have to choose one time interval of the
day when they will be in sensing mode and use the rest time intervals to harvest energy. We consider cases where
sensors had to choose their sensing mode among 4 available time intervals. Sensors are able to communicate with
other sensors that are at most 0.6 distance units away, and can only observe events that are at most 0.3 distance units
away. The case where 20 events are taking place is considered. Those events were uniformly distributed in space
and time, so an event could evenly appear in any point of the unit square area and it could occur at any time with the
same probability. The duration of each event was uniformly chosen between (0 − 6] hours and each event had a value
Ve ∈ (0 − 1].
The results of EKF fictitious play are compared with the ones of the particle filter based algorithm. The number of
particles in particle filter fictitious play algorithm, play an important role in the computational time of the algorithm.
For that reason comparisons were made with three variants of the particle filter algorithm with 500-1000 particles.
The results presented in Figure 15 and Table 10 are the averaged reward of 100 instances of the above described
scenario. Figure 15 depicts the the average global reward. The global reward in each instance was normalised using
the maximum reward which is achieved when all the sensors are in sense mode. As it is depicted in Figure 15 the
average performance of the particle filter algorithm is not affected by the number of particles in this particular example.
But the EKF fictitious play performed better than the particle filters alternatives. In addition, as it is shown in Table
10, the computational cost of the two particle filter's variants is greater than the EKF fictitious play algorithm.
8. Conclusions
A novel variant of fictitious play has been presented. This variant is based on extended Kalman filters and can
be used as a decision making module of a robot's controller. The relation between the proposed learning process and
24
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
25
EKF fictitious play
Particle filter fictitious play 500 particles
Particle filter 1000 particles
Average time
0.33
5.37
8.43
Table 10. Average computational time for each agent in seconds.
the BDI framework was also considered. The main theoretical result has been, that the learning algorithm associated
with the cooperative game converges to the Nash equilibrium of potential games. The application of this method-
ology has been illustrated on four examples. The filtering methods used for learning have also been compared for
their performance and computational costs. The methods presented can potentially find widespread applications in
the robotics industry. Possible improvements could be obtained by the adoption of more sophisticated reward/cost
functions without altering the essential features of the methodology.
Appendix A. Proof of Proposition 1
Appendix A.1. Case where players have two available actions
In the case where players have only 2 available actions Player i's predictions about his opponent's propensity are:
without loss of generality we can assume that his opponent in iteration t chooses action 2. Then the update step will
be :
since Players i's opponent played action 2 and hs(k)(Q) = exp(Qt[s(k)]/τ)
(cid:33)
+ qI
(A.1)
(A.2)
(cid:33)
s∈S exp(Qt[ s]/τ) we can write vt and Ht( ¯Qt) as:
(cid:33)
σt−1(1)
1 − σt−1(1)
(cid:33)
(A.3)
(A.4)
(A.5)
(A.6)
(A.7)
(A.8)
(cid:32)
(cid:33)
(cid:32) P−
Qt−1[1]
¯Qt =
Qt−1[2]
t−1[1, 1] P−
P−
t−1[2, 1] P−
t−1[1, 2]
t−1[2, 2]
P−
t =
vt =
=
(cid:32) 0
1
−
vt = zt − h( ¯Qt)
(cid:80)
(cid:33)
(cid:32)
(cid:32) −σt−1(1)
(cid:32) at
(cid:32) b
σt−1(1)
−at
−b
−b
b
S t = a2
Ht( ¯Qt) =
−at
at
(cid:33)
+ Z
(cid:33)(cid:32) at
t [1, 2]. The Kalaman gain, Kt = P−
−at
(cid:33)
−c
d
(cid:32) P−
(cid:32) c
P−
t [2, 2]
t [1, 1]
−d
k
k
where at is defined at = σt−1(1)σt−1(2). The estimation of S t = H( ¯Qt)P−
t HT ( ¯Qt) + Z will be:
where b = P−
t [1, 1] + P−
t [2, 2] − 2P−
1
Kt =
2rb + r2
up to a multiplicative constant we can write
can be written as
t
(cid:33)(cid:32) b + r
t HT ( ¯Qt)S −1
−at
at
b
(cid:33)
b
b + r
where c = P−
t [1, 1] − P−
(cid:32)
t [1, 2] and d = P−
Qt =
K1 ∼
(cid:33)
t [2, 2] − P−
t [1, 2]. The EKF estimate of the opponent's propensity is:
Qt[1]
Qt[2]
¯Qt[1] − 2σ(1)
¯Qt[2] + 2σ(1)
4a2(b−k)+(r+)
4a2(b−k)+(r+)
a(b−k)
a(b−k)
=
Based on the above we observe that Qt(1) < Qt−1(1) and Qt(2) > Qt−1(2).
25
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
26
Appendix A.2. Case where players have more than two available actions
H( ¯Qt)P−
diagonal elements and negative off diagonal elements. In particular we can write :
In the case where there are more than 2 available actions, when 1/t << 1 then the covariance matrix S t (cid:39)
t HT ( ¯Qt) and then Kt (cid:39) H( ¯Qt). From the definition of H( ¯Qt we know that it is invertible, with positive
(cid:88)
j∈S /i
H( ¯Qt)[i, i] =
σt−1(si)σt−1(s j)
H( ¯Qt)[i, j] = −σt−1(si)σt−1(s j).
Suppose that action s j is played from Player i's opponent then the update of Qt[ j] is the following:
Qt[ j] = ¯Qt + H( ¯Qt)[ j,·]y.
The only positive element of y is y[ j]. The multiplication of H( ¯Qt)[ j,·] and y is the sum of I positive coefficients and
therefore the value of Qt[ j] will be increased. In order to complete the proof we should show that ∆[ j] > ∆[i]∀i ∈ S /i,
where ∆[i] = Qt[i] − Qt−1[i], i ∈ S . For simplicity of notation for the rest of the proof we will write H[i, j] instead of
H( ¯Qt)[i, j] and σ(i) instead of σt−1(si).
∆[i] = H[i,·]y
= H[i, 1]σ(1) + H[i, 2]σ(2) + . . . +
H[i, i − 1]σ(i − 1) − H[i, i]σ(i)+
H[i, i + 1]σ[i + 1] + . . . +
H[i, j − 1]σ( j − 1) − H[i, j](1 − σ( j))+
H[i, j + 1]σ( j + 1) + . . . + H[i,I]σ(I)
∆[ j] = H[ j,·]y
= H[ j, 1]σ(1) + H[ j, 2]σ(2) + . . . +
H[ j, j](1 − σ( j)) + . . . + H[i,I]σ(I)
∆[i] − ∆[ j] = σ(1)(H[i, 1] − H[ j, 1]) + . . . +
σ(i − 1)(H[i, i − 1] − H[ j, i − 1])−
σ(i)(H[i, i] + H[ j, i])+
σ(i + 1)(H[i, i + 1] − H[i, i + 1]) + . . . +
σ( j − 1)(H[i, j − 1] − H[ j, j − 1])−
(1 − σ( j))(H[[i, j] + H[ j, j]])+
σ( j + 1)(H[i, j + 1] − H[ j, j + 1]) + . . . +
σ(I)(H[i,I] − H[ j,I])
(σ(i))2(((cid:80)
(1 − σ( j))σ( j)(((cid:80)
(σ(i − 1))2(σ(i) − σ( j))−
j∈S /i σ( j)) + σ( j))+
(σ(i + 1))2(σ(i) − σ( j)) + . . . +
(σ( j − 1))2(σ(i) − σ( j))−
(σ( j + 1))2(σ(i) − σ( j)) + . . . +
(σ(I))2(σ(i) − σ( j))
j∈S / j σ( j)) + σ(i))+
If we substitute H[·,·] with its equivalent and we can write ∆[i] − ∆[ j] as:
∆[i] − ∆[ j] = (σ(1))2(σ(i) − σ( j)) + . . . +
solving the inequality ∆[i] − ∆[ j] < 0 we obtain:
∆[i] − ∆[ j] < 0 ⇔
(σ(i) − σ( j))(((cid:80)
((σ(i))2(((cid:80)
(1 − σ( j))(σ( j))(((cid:80)
26
j∈S /{i, j}(σ( j)2))) <
j∈S /i σ( j)) + σ( j))+
j∈S / j σ( j)) + σ(i)))
(A.9)
(A.10)
(A.11)
(A.12)
(A.13)
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
27
In the case where σ(i) < σ( j) the inequality is satisfied always because the left hand side of the inequality is always
negative and the right hand side is always positive. In the case where σ(i) > σ( j) inequality (A.13) we will show by
contradiction that (A.13) is satisfied ∀i, i (cid:44) j. Therefore we assume that:
∆[i] − ∆[ j] ≥ 0 ⇔
(σ(i) − σ( j))(((cid:80)
((σ(i))2 + (1 − σ( j))(σ( j)))(((cid:80)
j∈S /{i, j}(σ( j)2))) ≥
j∈S /i σ( j)) + σ( j))
Since (((cid:80)
j∈S /{i, j}(σ( j)2))) < (((cid:80)
(A.14)
(A.15)
(A.16)
(A.17)
j∈S /i σ( j)) + σ( j)) in order to complete the proof we only need to show that:
σ(i) − σ( j) ≥ (σ(i))2 + (1 − σ( j))σ( j)
σ(i) − (σ(i))2 ≥ 2σ( j) − (σ( j))2
σ(i)(1 − σ(i)) ≥ σ( j)(2 − σ( j))
Inequality (A.15) will be satisfied if the following inequality is satisfied:
σ(i) (2 − σ( j))
(1 − σ(i)) ≥ (2 − σ( j))
Inequality (A.16) will be satisfied if
(1 − σ(i)) ≥ σ( j)
σ( j) − σ(i) ≥ 1
since σ(i) > σ( j), (A.17), is false ∀i and thus by contradiction ∆[i] − ∆[ j] < 0, which completes the proof.
Appendix B. Proof of Proposition 2
Similarly to the proof of Proposition 1 we consider only one opponent and a game with more than one pure Nash
equilibria. We assume that a joint action s = (s1( j(cid:48)), s2( j)) is played which is not a Nash equilibrium. Since players
use EKF fictitious play because of Proposition 1 they will eventually change their action to s = (s1( j(cid:48)]), s2( j)) which
will be the best response to action s2( j) and s1( j(cid:48)) for players 1 and 2 respectively. But if this change is simultaneous
their is no guarantee that the resulted joint action s will increase the expected reward of the players. We will show that
with high probability the two players will not change actions simultaneously infinitely often.
Without loss of generality we assume that at time t Player 2 change his action from s2( j) to s2( j). We want to show
that the probability that Player 1 will change his action to s1( j(cid:48)) with probability less than 1. Player 1 will change
his action if he thinks that Player 2 will play action s2( j) with high probability such that his utility will be maximised
when he plays action s1( j(cid:48)). Therefore Player 1 will change his action to s1( j(cid:48)) if σ2(s2( j)) > λ, where σ2(·) is the
estimation of his Player 1 about Player 2's strategy. Thus we want to show
exp( Qt−1[ j])
p(σ2 > λ) < 1 ⇔
(cid:80)
j(cid:48)(cid:48)∈S 2 exp( Qt−1[ j(cid:48)(cid:48)]) > λ) < 1 ⇔
p(
+(cid:80)
p( Qt−1[ j] > ln λ
1−λ
p( Qt−2[ j] + (Kt−1yt−1)[ j] > . . .
j(cid:48)(cid:48)∈S 2/ j exp( Qt−1[ j(cid:48)(cid:48)])) < 1
+(cid:80)
ln λ
1−λ
j(cid:48)(cid:48)∈S 2/ j exp( Qt−1[ j(cid:48)(cid:48)])) < 1 ⇔
we can expand Kt−1yt−1[ j] as follows
(Kt−1yt−1)[ j] = ((Pt−2 + (ξ + )I)HS −1yt−1)[ j]
= (Pt−2HS −1yt−1)[ j] + ((ξ + )HS −1yt−1)[ j]
If we substitute Kt−1yt−1[ j] in (B.1) with its equivalent we obtain the following inequality:
p( Qt−2[ j] + (Pt−2HS −1yt−1)[ j] + ((ξ + )HS −1yt−1)[ j] > . . .
ln λ
1−λ
p(((ξ + )HS −1yt−1)[ j] > − Qt−2[ j] − (Pt−2HS −1yt−1)[ j] . . .
ln λ
1−λ
j(cid:48)(cid:48)∈S 2/ j exp( Qt−1[ j(cid:48)(cid:48)])) < 1 ⇔
j(cid:48)(cid:48)∈S 2/ j exp( Qt−1[ j(cid:48)(cid:48)])) < 1 ⇔
+(cid:80)
+(cid:80)
p( > C) < 1
27
(B.1)
(B.2)
(B.3)
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
28
where C is defined as:
C =
+
(cid:80)
ln λ
1−λ
(HS −1yt−1)[ j]
j(cid:48)(cid:48)∈S 2 / j exp( Qt−1[ j(cid:48)(cid:48)])
−
(HS −1yt−1)[ j]
− Qt−2[ j]−(Pt−2HS −1yt−1)[ j]
q(HS −1yt−1)[ j]
(HS −1yt−1)[ j]
(HS −1yt−1)[ j]
−
(B.4)
Inequality (B.3) is always satisfied since is a Gaussian random variable. To conclude the proof we define χt as
the event that both players change their action at time t simultaneously, and assume that the two players have change
their actions simultaneously at the following iterations t1, t2, . . . , tT , then the probability that they will also change
their action simultaneously at time tT +1, P(χt1 , χt2 , . . . , χtT , χtT +1) is almost zero for large but finite T.
References
[1] Arslan, G., Marden, J. R., Shamma, J. S., 2007. Autonomous vehicle-target assignment: A game-theoretical formulation. Journal of Dynamic
Systems, Measurement, and Control 129 (5), 584 -- 596.
[2] Ayken, T., Imura, J.-i., 2012. Asynchronous distributed optimization of smart grid. In: SICE Annual Conference (SICE), 2012 Proceedings
[3] Bauso, D., Giarre, L., Pesenti, R., 2006. Mechanism design for optimal consensus problems. In: Decision and Control, 2006 45th IEEE
of. IEEE, pp. 2098 -- 2102.
Conference on. pp. 3381 -- 3386.
[4] Bertsekas, D. P., 1982. Distributed dynamic programming. Automatic Control, IEEE Transactions on 27 (3), 610 -- 616.
[5] Bishop, C. M., 1995. Neural Networks for Pattern Recognition. Oxford University Press.
[6] Botelho, S., Alami, R., 1999. M+: a scheme for multi-robot cooperation through negotiated task allocation and achievement. In: Robotics
and Automation, 1999. Proceedings. 1999 IEEE International Conference on. Vol. 2. pp. 1234 -- 1239 vol.2.
[7] Brown, G. W., 1951. Iterative solutions of games by fictitious play. In: Koopmans, T. C. (Ed.), Activity Analysis of Production and Allocation.
Wiley, pp. 374 -- 376.
[8] Chapman, A. C., Rogers, A., Jennings, N. R., Leslie, D. S., 2011. A unifying framework for iterative approximate best-response algorithms
for distributed constraint optimization problems. Knowledge Engineering Review 26 (4), 411 -- 444.
[9] Chapman, A. C., Rogers, A., Jennings, N. R., Leslie, D. S., 2011. A unifying framework for iterative approximate best-response algorithms
for distributed constraint optimization problems. Knowledge Engineering Review 26 (4), 411 -- 444.
[10] Daskalakis, C., Goldberg, P. W., Papadimitriou, C. H., 2006. The complexity of computing a nash equilibrium. In: Proceedings of the
thirty-eighth annual ACM symposium on Theory of computing. pp. 71 -- 78.
[11] Evans, J., Krishnamurthy, B., 1998. Helpmate, the trackless robotic courier: A perspective on the development of a commercial autonomous
mobile robot. In: Autonomous Robotic Systems. Vol. 236. Springer London, pp. 182 -- 210.
[12] Farinelli, A., Rogers, A., Jennings, N. R., June 2008. Maximising sensor network efficiency through agent-based coordination of sense/sleep
schedules. In: Workshop on Energy in Wireless Sensor Networks in conjuction with DCOSS 2008. pp. 43 -- 56.
[13] Fudenberg, D., Levine, D., 1998. The theory of Learning in Games. The MIT Press.
[14] Grewal, M., Andrews, A., 2011. Kalman filtering: theory and practice using MATLAB. Wiley-IEEE press.
[15] Jazwinski, A., 1970. Stochastic processes and filtering theory. Vol. 63. Academic press.
[16] Kalman, R., et al., 1960. A new approach to linear filtering and prediction problems. Journal of basic Engineering 82 (1), 35 -- 45.
[17] Kho, J., Rogers, A., Jennings, N. R., 2009. Decentralized control of adaptive sampling in wireless sensor networks. ACM Trans. Sen. Netw.
[18] Kho, J., Rogers, A., Jennings, N. R., 2009. Decentralized control of adaptive sampling in wireless sensor networks. ACM Transactions on
[19] Lincoln, N. K., Veres, S. M., Dennis, L., Fisher, M., 2013. Autonomous asteroid exploration by rational agents. IEEE Computational Intelli-
[20] Madhavan, R., Fregene, K., Parker, L., 2004. Distributed cooperative outdoor multirobot localization and mapping. Autonomous Robots
5 (3), 1 -- 35.
Sensor Networks (TOSN) 5 (3), 19.
gence Magazine 8, 25 -- 38.
17 (1), 23 -- 39.
[21] Makarenko, A., Durrant-Whyte, H., 2004. Decentralized data fusion and control in active sensor networks. In: Pro. 7th International Confer-
ence on Information Fusion. Vol. 1. pp. 479 -- 486.
[22] Miyasawa, K., 1961. On the convergence of learning process in a 2x2 non-zero-person game.
[23] Monderer, D., Shapley, L., 1996. Potential games. Games and Economic Behavior 14, 124 -- 143.
[24] Nachbar, J., 1990. Evolutionary' selection dynamics in games: Convergence and limit properties. International Journal of Game Theory 19,
[25] Nash, J., 1950. Equilibrium points in n-person games. In: Proceedings of the National Academy of Science, USA. Vol. 36. pp. 48 -- 49.
[26] Parker, L., 1998. Alliance: an architecture for fault tolerant multirobot cooperation. Robotics and Automation, IEEE Transactions on 14 (2),
59 -- 89.
220 -- 240.
[27] Raffard, R. L., Tomlin, C. J., Boyd, S. P., 2004. Distributed optimization for cooperative agents: Application to formation flight. In: Decision
and Control, 2004. CDC. 43rd IEEE Conference on. Vol. 3. IEEE, pp. 2453 -- 2459.
[28] Robinson, J., 1951. An iterative method of solving a game. Annals of Mathematics 54, 296 -- 301.
28
M. Smyrnakis and S. Veres / Engineering Applications of Artificial Intelligence 00 (2018) 1 -- 29
29
[29] Semsar-Kazerooni, E., Khorasani, K., 2008. Optimal consensus algorithms for cooperative team of agents subject to partial information.
[30] Semsar-Kazerooni, E., Khorasani, K., 2009. Multi-agent team cooperation: A game theory approach. Automatica 45 (10), 2205 -- 2213.
[31] Simmons, R., Apfelbaum, D., Burgard, W., Fox, D., Moors, M., Thrun, S., Younes, H., 2000. Coordination for multi-robot exploration and
[32] Simmons, R., Apfelbaum, D., Burgard, W., Fox, D., Moors, M., Thrun, S., Younes, H., 2000. Coordination for multi-robot exploration and
Automatica 44 (11), 2766 -- 2777.
mapping. In: AAAI/IAAI. pp. 852 -- 858.
mapping. In: AAAI/IAAI. pp. 852 -- 858.
[33] Smyrnakis, M., Leslie, D. S., 2010. Dynamic Opponent Modelling in Fictitious Play. The Computer Journal 53, 1344 -- 1359.
[34] Stranjak, A., Dutta, P. S., Ebden, M., Rogers, A., Vytelingum, P., May 2008. A multi-agent simulation system for prediction and scheduling of
aero engine overhaul. In: AAMAS '08: Proceedings of the 7th international joint conference on Autonomous agents and multiagent systems.
pp. 81 -- 88.
[35] Timofeev, A., Kolushev, F., Bogdanov, A., 1999. Hybrid algorithms of multi-agent control of mobile robots. In: Neural Networks, 1999.
IJCNN '99. International Joint Conference on. Vol. 6. pp. 4115 -- 4118.
[36] Tsalatsanis, A., Yalcin, A., Valavanis, K., 2009. Optimized task allocation in cooperative robot teams. In: Control and Automation, 2009.
MED '09. 17th Mediterranean Conference on. pp. 270 -- 275.
[37] Voice, T., Vytelingum, P., Ramchurn, S. D., Rogers, A., Jennings, N. R., 2011. Decentralised control of micro-storage in the smart grid. In:
AAAI. pp. 1421 -- 1427.
[38] Wolpert, D., Tumer, K., 2004. .A survey of collectives. Springer, Ch. Collectives and the Design of Complex Systems., pp. 1 -- 42.
[39] Young, H. P., 2005. Strategic Learning and Its Limits. Oxford University Press.
[40] Zecchin, A. C., Simpson, A. R., Maier, H. R., Leonard, M., Roberts, A. J., Berrisford, M. J., 2006. Application of two ant colony optimisation
algorithms to water distribution system optimisation. Math. Comput. Model. 44 (5), 451 -- 468.
[41] Zhang, P., Sadler, C. M., Lyon, S. A., Martonosi, M., 2004. Hardware design experiences in zebranet. In: Proc. SenSys'04. ACM, pp.
227 -- 238.
[42] Zhang, Y., Schervish, M., Acar, E., Choset, H., 2001. Probabilistic methods for robotic landmine search. In: Proceedings of the 2001
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS '01). pp. 1525 -- 1532.
[43] Zlot, R., Stentz, A., Dias, M. B., Thayer, S., 2002. Multi-robot exploration controlled by a market economy. In: Robotics and Automation,
2002. Proceedings. ICRA'02. IEEE International Conference on. Vol. 3.
29
|
cs/0602045 | 1 | 0602 | 2006-02-12T22:11:14 | Emergence Explained | [
"cs.MA",
"cs.DC",
"cs.GL"
] | Emergence (macro-level effects from micro-level causes) is at the heart of the conflict between reductionism and functionalism. How can there be autonomous higher level laws of nature (the functionalist claim) if everything can be reduced to the fundamental forces of physics (the reductionist position)? We cut through this debate by applying a computer science lens to the way we view nature. We conclude (a) that what functionalism calls the special sciences (sciences other than physics) do indeed study autonomous laws and furthermore that those laws pertain to real higher level entities but (b) that interactions among such higher-level entities is epiphenomenal in that they can always be reduced to primitive physical forces. In other words, epiphenomena, which we will identify with emergent phenomena, do real higher-level work. The proposed perspective provides a framework for understanding many thorny issues including the nature of entities, stigmergy, the evolution of complexity, phase transitions, supervenience, and downward entailment. We also discuss some practical considerations pertaining to systems of systems and the limitations of modeling. | cs.MA | cs | DRAFT.
DRAFT
Comments
are welcome.
2/12/2006
Emergence Explained:
Getting epiphenomena to do real work
Russ Abbott
Department of Computer Science
California State University, Los Angeles
Los Angeles, California
[email protected]
Abstract. Emergence— macro-level ef-
fects from micro-level causes —
is at the
heart of the conflict between reduction-
ism and functionalism. How can there be
autonomous higher level laws of nature
(the functionalist claim) if everything
can be reduced to the fundamental forces
of physics (the reductionist position)?
We cut through this debate by applying a
computer science lens to the way we
view nature. We conclude (a) that what
functionalism calls the special sciences
(sciences other than physics) do indeed
study autonomous laws and furthermore
that those laws pertain to real higher
level entities but (b) that interactions
Emergence Explained: ___________________1
Getting epiphenomena to do real work___1
Russ Abbott _________________________1
Introduction_______________________3
Background and foundations _________5
2.1
Functionalism ________________5
Reductionism ________________6
2.2
2.3
Epiphenomena _______________7
Supervenience ________________8
2.4
2.5
Fundamental forces and strong
emergence _________________________10
2.6
Historical accidents and the
environment _______________________12
Emergence in the Game of Life ______13
1
2
among such higher-level entities is epi-
phenomenal in that they can always be
reduced to primitive physical forces. In
other words, epiphenomena, which we
will identify with emergent phenomena,
do real higher-level work. The proposed
perspective provides a framework for
understanding many thorny issues in-
cluding the nature of entities, stigmergy,
the evolution of complexity, phase tran-
sitions, supervenience, and downward
entailment. We also discuss some practi-
cal considerations pertaining to systems
of systems and the limitations of model-
ing.
Epiphenomenal gliders _______ 14
3.1
Gliders in our physics world ___ 14
3.2
3.3
The Game of Life as a
programming platform ______________ 15
Game of Life anthropologists __ 16
3.4
3.5
Keeping score _______________ 17
3.6
Defining emergence __________ 17
Implications of emergence __________ 19
Non-reductive regularities ____ 20
4.1
4.2
Downward entailment ________ 20
Reduction proofs ____________ 21
4.3
4.4
Downward entailment as science21
4.5
The reality of higher level
abstractions _______________________ 22
4
3
Emergence Explained
1/67
DRAFT
2/12/2006
Abstractions and abstract designs
4.6
24
Phase transitions _____________24
4.7
The constructionist hypothesis
4.8
revisited ___________________________25
Entities __________________________26
5.1
Entities, entropy, designs, and
functionality _______________________27
Entities at an energy equilibrium
5.2
28
5.3
Entities and emergence are
fundamental _______________________29
Dissipative structures _________30
5.4
5.5
Integrating dissipative structures
and at-equilibrium entities. ___________31
Autonomous entities __________31
5.6
5.7
A naturally occurring autonomous
entity that is neither biological nor social 33
5.8
Natural and artificial autonomous
entities 33
5.9
Theseus’s ship _______________34
5.10
Autonomous entities may act in
the world __________________________35
5.11
Autonomous entities tend not to
supervene over their static components _35
Entities, objects, and agents____36
5.12
5.13
Thermodynamic computing: nihil
ex nihilo ___________________________37
5.14 Minimal autonomous entities __38
The evolution of complexity _________38
Stigmergy __________________38
6.1
Design and evolution _________40
6.2
6.3
The accretion of complexity ____40
6.4
Increasing complexity increasing
efficiency, and historical contingency ___42
Entities, emergence, and science _____43
Entities and the sciences_______43
7.1
5
6
7
8
9
Science and emergence _______ 43
7.2
Varieties of Emergence ____________ 45
Static emergence ____________ 45
8.1
8.2
Dynamic emergence _________ 46
8.3
Non-stigmergic dynamic
emergence _________________________ 46
8.4
Dynamic emergence and grand
reductionism _______________________ 46
8.5
Stigmergic emergence ________ 47
Some practical considerations _______ 48
Emergence and software ______ 48
9.1
Bricolage as design __________ 48
9.2
9.3
Infrastructure-centric
development _______________________ 49
9.4
Service refactoring and the age of
services 50
9.5
A possible undesirable unintended
consequence _______________________ 50
9.6
Modeling: the difficulty of looking
downward _________________________ 51
9.7
Modeling: the difficulty of looking
upward 52
10
Concluding remarks ____________ 53
Computer Science and Philosophy
10.1
53
Constructive science _________ 54
10.2
11
Acknowledgement ______________ 55
Appendix. Game of Life Patterns __ 60
12
12.1
Live cell groups _____________ 60
12.2
Basic patterns: temporal
sequences of live cell groups __________ 60
12.3
BP is recursively enumerable __ 62
12.4
Game of Life patterns:
combinations of basic patterns ________ 63
Emergence Explained
2/67
DRAFT
2/12/2006
1 Introduction
Although the field of complex systems is
relatively young, the sense of the term
emergence that is commonly associated
that micro phenomena often
with it —
give rise to macro phenomena1—
has
been in use for well over a century. The
article on Emergent Properties in the
valuable online Stanford Encyclopedia
of Philosophy [O'Connor] begins as fol-
lows.
(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:3)(cid:6)(cid:7)(cid:3)(cid:8)(cid:9)(cid:10)(cid:11)(cid:12)(cid:8)(cid:13)(cid:3)(cid:3)(cid:6)(cid:14)(cid:8)(cid:11)(cid:8)(cid:6)(cid:15)(cid:16)(cid:15)(cid:4)(cid:17)(cid:15)(cid:18) (cid:12)(cid:8)(cid:19) (cid:10)(cid:17)(cid:20)(cid:15)(cid:12)(cid:15)(cid:21)
(cid:19) (cid:10)(cid:17)(cid:7)(cid:11)(cid:20)(cid:8)(cid:16)(cid:3)(cid:4)(cid:2)(cid:8)(cid:15)(cid:22)(cid:8)(cid:11)(cid:4)(cid:16)(cid:8)(cid:9)(cid:12)(cid:17)(cid:6)(cid:7)(cid:3)(cid:8)(cid:23) (cid:24) (cid:25) (cid:26) (cid:14)(cid:27)(cid:8)(cid:28)
(cid:8)(cid:29) (cid:3)(cid:8)(cid:2)(cid:17)(cid:5)(cid:10)(cid:16)(cid:8)
(cid:4)(cid:15)(cid:18) (cid:5)(cid:10)(cid:20)(cid:30)(cid:8) (cid:7)(cid:10)(cid:11)(cid:4)(cid:11)(cid:7)(cid:16)(cid:3)(cid:4)(cid:17)(cid:31) (cid:3)(cid:8) (cid:9)(cid:17)(cid:16)(cid:12)(cid:14)(cid:8) (cid:2)(cid:3)(cid:11)(cid:6)(cid:17)(cid:6)(cid:5)(cid:8) (cid:16)(cid:10)(cid:18) (cid:12) (cid:8)
(cid:3)(cid:2)(cid:3)(cid:4)(cid:5)(cid:3)(cid:6)(cid:16)(cid:8) (cid:3)(cid:6)(cid:16)(cid:17)(cid:16)(cid:17)(cid:3)(cid:12)(cid:8)
! (cid:19) (cid:4)(cid:15)(cid:19) (cid:3)(cid:4)(cid:16)(cid:17)(cid:3)(cid:12)(cid:8) (cid:15)(cid:4)(cid:8) (cid:12)(cid:18) (cid:13)(cid:21)
(cid:12)(cid:16)(cid:11)(cid:6)(cid:7)(cid:3)(cid:12)" (cid:8) #(cid:11)(cid:4)(cid:17)(cid:12)(cid:3)$(cid:8) (cid:15)(cid:18) (cid:16)(cid:8) (cid:15)(cid:22)(cid:8) (cid:2)(cid:15)(cid:4)(cid:3)(cid:8) (cid:22)(cid:18) (cid:6)% (cid:11)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:20)(cid:8)
(cid:3)(cid:6)(cid:16)(cid:17)(cid:16)(cid:17)(cid:3)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:30)(cid:3)(cid:16)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) #(cid:6)(cid:15)& (cid:3)(cid:20)$(cid:8) (cid:15)(cid:4)(cid:8) #(cid:17)(cid:4)(cid:4)(cid:3)% (cid:18) (cid:7)(cid:17)(cid:13)(cid:20)(cid:3)$(cid:8)
’ (cid:17)(cid:16)(cid:10)(cid:8)(cid:4)(cid:3)(cid:12)(cid:19) (cid:3)(cid:7)(cid:16)(cid:8)(cid:16)(cid:15)(cid:8)(cid:16)(cid:10)(cid:3)(cid:2)(cid:27)(cid:8)(cid:28)
(cid:8)(cid:1)(cid:11)(cid:7)(cid:10)(cid:8)(cid:15)(cid:22)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)( (cid:18) (cid:15)(cid:16)(cid:3)% (cid:8)
(cid:16)(cid:3)(cid:4)(cid:2)(cid:12)(cid:8) (cid:17)(cid:12)(cid:8) (cid:12)(cid:20)(cid:17)(cid:19) (cid:19) (cid:3)(cid:4)(cid:30)(cid:8) (cid:17)(cid:6)(cid:8) (cid:17)(cid:16)(cid:12)(cid:8) (cid:15)’ (cid:6)(cid:8) (cid:4)(cid:17)(cid:5)(cid:10)(cid:16)(cid:8) (cid:28)
(cid:8) (cid:27)(cid:8) ) (cid:10)(cid:3)(cid:4)(cid:3)(cid:8)
(cid:10)(cid:11)(cid:12)(cid:8) (cid:13)(cid:3)(cid:3)(cid:6)(cid:8) (cid:4)(cid:3)(cid:6)(cid:3)’ (cid:3)% (cid:8) (cid:17)(cid:6)(cid:16)(cid:3)(cid:4)(cid:3)(cid:12)(cid:16)(cid:8) (cid:17)(cid:6)(cid:8) (cid:3)(cid:2)(cid:3)(cid:4)(cid:5)(cid:3)(cid:6)(cid:7)(cid:3)(cid:8)
’ (cid:17)(cid:16)(cid:10)(cid:17)(cid:6)(cid:8) % (cid:17)(cid:12)(cid:7)(cid:18) (cid:12)(cid:12)(cid:17)(cid:15)(cid:6)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:13)(cid:3)(cid:10)(cid:11)& (cid:17)(cid:15)(cid:4)(cid:8) (cid:15)(cid:22)(cid:8) (cid:7)(cid:15)(cid:2)(cid:21)
(cid:19) (cid:20)(cid:3)*(cid:8)(cid:12)(cid:30)(cid:12)(cid:16)(cid:3)(cid:2)(cid:12)(cid:27)(cid:8)
In a 1998 book-length perspective on his
life’ s work [Holland], John Holland, the
inventor of genetic algorithms and one
of the founders of the field of complex
systems, offered an admirably honest ac-
count of the state of our understanding
of emergence.
+(cid:16)(cid:8) (cid:17)(cid:12)(cid:8) (cid:18) (cid:6)(cid:20)(cid:17), (cid:3)(cid:20)(cid:30)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:11)(cid:8) (cid:16)(cid:15)(cid:19) (cid:17)(cid:7)(cid:8) (cid:11)(cid:12)(cid:8) (cid:7)(cid:15)(cid:2)(cid:19) (cid:20)(cid:17)(cid:7)(cid:11)(cid:16)(cid:3)% (cid:8) (cid:11)(cid:12)(cid:8)
(cid:3)(cid:2)(cid:3)(cid:4)(cid:5)(cid:3)(cid:6)(cid:7)(cid:3)(cid:8) ’ (cid:17)(cid:20)(cid:20)(cid:8)(cid:12)(cid:18) (cid:13)(cid:2)(cid:17)(cid:16)(cid:8)(cid:2)(cid:3)(cid:3), (cid:20)(cid:30)(cid:8) (cid:16)(cid:15)(cid:8)(cid:11)(cid:8)(cid:7)(cid:15)(cid:6)(cid:7)(cid:17)(cid:12)(cid:3)(cid:8)
% (cid:3)(cid:22)(cid:17)(cid:6)(cid:17)(cid:16)(cid:17)(cid:15)(cid:6)-(cid:8) (cid:11)(cid:6)% (cid:8) +(cid:8) (cid:10)(cid:11)& (cid:3)(cid:8) (cid:6)(cid:15)(cid:8) (cid:12)(cid:18) (cid:7)(cid:10)(cid:8) % (cid:3)(cid:22)(cid:17)(cid:6)(cid:17)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:16)(cid:15)(cid:8)
(cid:15)(cid:22)(cid:22)(cid:3)(cid:4)(cid:27)(cid:8)
In a review of Holland’ s book, Cosma
Shalizi wrote the following.
. (cid:15)(cid:2)(cid:3)(cid:19) (cid:20)(cid:11)(cid:7)(cid:3)(cid:8) (cid:28)
(cid:8) ’ (cid:10)(cid:3)(cid:4)(cid:3)(cid:8) ( (cid:18) (cid:11)(cid:6)(cid:16)(cid:18) (cid:2)(cid:8) (cid:22)(cid:17)(cid:3)(cid:20)% (cid:8) (cid:16)(cid:10)(cid:3)(cid:15)(cid:4)(cid:30)(cid:8)
(cid:2)(cid:3)(cid:3)(cid:16)(cid:12)(cid:8) (cid:5)(cid:3)(cid:6)(cid:3)(cid:4)(cid:11)(cid:20)(cid:8) (cid:4)(cid:3)(cid:20)(cid:11)(cid:16)(cid:17)& (cid:17)(cid:16)(cid:30)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:11)(cid:16)(cid:15)(cid:2)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8)
1 Recently the term multiscale has gained favor as a
less mysterious-sounding way to refer to this
macro-micro interplay. The fact that multiscale
sounds more scientific, however, does not reflect a
correspondingly clearer understanding of the phe-
nomenon itself.
& (cid:15)(cid:17)% (cid:8) (cid:2)(cid:3)(cid:4)(cid:5)(cid:3)(cid:8) (cid:17)(cid:6)(cid:16)(cid:15)(cid:8) (cid:15)(cid:6)(cid:3)(cid:8) (cid:11)(cid:6)(cid:15)(cid:16)(cid:10)(cid:3)(cid:4)-(cid:8) ’ (cid:3)(cid:8) (cid:2)(cid:11)(cid:30)(cid:8) (cid:16)(cid:11), (cid:3)(cid:8)
/ (cid:16)(cid:10)(cid:3)(cid:8)(cid:4)(cid:18) (cid:20)(cid:3)(cid:12)(cid:8)(cid:15)(cid:22)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:5)(cid:11)(cid:2)(cid:3)0 (cid:8)(cid:16)(cid:15)(cid:8)(cid:13)(cid:3)(cid:8)(cid:5)(cid:17)& (cid:3)(cid:6)(cid:27)(cid:8)1 (cid:18) (cid:16)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)
(cid:4)(cid:3)(cid:12)(cid:16)(cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:15)(cid:13)(cid:12)(cid:3)(cid:4)& (cid:11)(cid:13)(cid:20)(cid:3)-(cid:8) (cid:3)*(cid:19) (cid:20)(cid:15)(cid:17)(cid:16)(cid:11)(cid:13)(cid:20)(cid:3)(cid:8) (cid:15)(cid:4)% (cid:3)(cid:4)(cid:8) (cid:17)(cid:6)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:8)
(cid:18) (cid:6)(cid:17)& (cid:3)(cid:4)(cid:12)(cid:3)(cid:8) 2(cid:8) (cid:13)(cid:3)(cid:6)(cid:31) (cid:3)(cid:6)(cid:3)(cid:8) (cid:2)(cid:15)(cid:20)(cid:3)(cid:7)(cid:18) (cid:20)(cid:3)(cid:12)-(cid:8)
(cid:1)(cid:2)(cid:3)(cid:4)(cid:3)(cid:5) (cid:6) (cid:7) (cid:8)(cid:8)(cid:12)(cid:6)(cid:15)’ (cid:22)(cid:20)(cid:11), (cid:3)(cid:12)-(cid:8)(cid:7)(cid:30)(cid:7)(cid:20)(cid:15)(cid:6)(cid:17)(cid:7)(cid:8)(cid:12)(cid:16)(cid:15)(cid:4)(cid:2)(cid:12)-(cid:8), (cid:17)(cid:16)(cid:21)
(cid:16)(cid:3)(cid:6)(cid:12)-(cid:8) (cid:7)(cid:11)(cid:16)(cid:12)-(cid:8) (cid:30)(cid:15)(cid:18) (cid:6)(cid:5)(cid:8) (cid:20)(cid:15)& (cid:3)-(cid:8) (cid:2)(cid:17)% % (cid:20)(cid:3)(cid:21)(cid:11)(cid:5)(cid:3)% (cid:8) (cid:4)(cid:3)(cid:21)
(cid:2)(cid:15)(cid:4)(cid:12)(cid:3)-(cid:8)
(cid:22)(cid:17)(cid:6)(cid:11)(cid:6)(cid:7)(cid:17)(cid:11)(cid:20)(cid:8) (cid:3)(cid:18) (cid:19) (cid:10)(cid:15)(cid:4)(cid:17)(cid:11)(cid:8) (cid:11)(cid:7)(cid:7)(cid:15)(cid:2)(cid:19) (cid:11)(cid:6)(cid:17)(cid:3)% (cid:8)
’ (cid:17)(cid:16)(cid:10)(cid:8) (cid:11)(cid:7)(cid:18) (cid:16)(cid:3)(cid:8) (cid:5)(cid:18) (cid:20)(cid:20)(cid:17)(cid:13)(cid:17)(cid:20)(cid:17)(cid:16)(cid:30)-(cid:8) (cid:19) (cid:4)(cid:3)& (cid:11)(cid:4)(cid:17)(cid:7)(cid:11)(cid:16)(cid:17)(cid:6)(cid:5)(cid:8) (cid:7)(cid:11)(cid:6)% (cid:17)(cid:21)
% (cid:11)(cid:16)(cid:3)(cid:12)(cid:8) (cid:22)(cid:15)(cid:4)(cid:8) (cid:19) (cid:18) (cid:13)(cid:20)(cid:17)(cid:7)(cid:8) (cid:15)(cid:22)(cid:22)(cid:17)(cid:7)(cid:3)-(cid:8) (cid:16)(cid:11)(cid:19) (cid:3)’ (cid:15)(cid:4)(cid:2)(cid:12)-(cid:8) 3(cid:3)(cid:16)(cid:21)(cid:20)(cid:11)(cid:5)-(cid:8)
(cid:11)(cid:6)% (cid:8) (cid:18) (cid:6)(cid:22)(cid:15)(cid:20)% (cid:17)(cid:6)(cid:5)(cid:8) (cid:7)(cid:10)(cid:3)(cid:4)(cid:4)(cid:30)(cid:8) (cid:13)(cid:20)(cid:15)(cid:12)(cid:12)(cid:15)(cid:2)(cid:12)(cid:8) 2(cid:8) ’ (cid:10)(cid:3)(cid:4)(cid:3)(cid:8)
% (cid:15)(cid:8) (cid:11)(cid:20)(cid:20)(cid:8) (cid:16)(cid:10)(cid:3)(cid:12)(cid:3)(cid:8) (cid:4)(cid:3)(cid:5)(cid:18) (cid:20)(cid:11)(cid:4)(cid:17)(cid:16)(cid:17)(cid:3)(cid:12)(cid:8) (cid:7)(cid:15)(cid:2)(cid:3)(cid:8) (cid:22)(cid:4)(cid:15)(cid:2)4 (cid:8) 5 (cid:11)(cid:20)(cid:20)(cid:8)
(cid:16)(cid:10)(cid:17)(cid:12)(cid:8) (cid:9) (cid:10)(cid:9) (cid:11) (cid:12) (cid:9) (cid:5) (cid:13) (cid:9) (cid:8) (cid:17)(cid:22)(cid:8) (cid:30)(cid:15)(cid:18) (cid:8) (cid:20)(cid:17), (cid:3)(cid:27)(cid:8) +(cid:16)$(cid:12)(cid:8) (cid:11)(cid:8) (cid:22)(cid:17)(cid:6)(cid:3)(cid:21)
(cid:12)(cid:15)(cid:18) (cid:6)% (cid:17)(cid:6)(cid:5)(cid:8) ’ (cid:15)(cid:4)% -(cid:8) (cid:11)(cid:6)% (cid:8) (cid:13)(cid:4)(cid:17)(cid:6)(cid:5)(cid:12)(cid:8) (cid:16)(cid:15)(cid:8) (cid:2)(cid:17)(cid:6)% (cid:8) (cid:12)(cid:15)(cid:18) (cid:16)(cid:10)(cid:21)
’ (cid:3)(cid:12)(cid:16)(cid:3)(cid:4)(cid:6)(cid:8) (cid:7)(cid:4)(cid:3)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:2)(cid:30)(cid:16)(cid:10)(cid:12)(cid:8) (cid:17)(cid:6)(cid:8) (cid:11)(cid:6)(cid:8) (cid:15)% % (cid:20)(cid:30)(cid:8) (cid:11)(cid:19) (cid:16)(cid:8)
’ (cid:11)(cid:30)(cid:27)2(cid:8)
The preceding is a poetic echo of the po-
sition expressed in a landmark paper
[Anderson] by Philip Anderson when he
distinguished reductionism from what he
called
the constructionist hypothesis,
with which he disagrees, which holds
that the
/(cid:11)(cid:13)(cid:17)(cid:20)(cid:17)(cid:16)(cid:30)(cid:8) (cid:16)(cid:15)(cid:8) (cid:4)(cid:3)% (cid:18) (cid:7)(cid:3)(cid:8) (cid:3)& (cid:3)(cid:4)(cid:30)(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:8) (cid:16)(cid:15)(cid:8) (cid:12)(cid:17)(cid:2)(cid:19) (cid:20)(cid:3)(cid:8) (cid:22)(cid:18) (cid:6)(cid:21)
% (cid:11)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:20)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8) (cid:28)
(cid:8) (cid:17)(cid:2)(cid:19) (cid:20)(cid:17)(cid:3)(cid:12)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)(cid:11)(cid:13)(cid:17)(cid:20)(cid:17)(cid:16)(cid:30)(cid:8) (cid:16)(cid:15)(cid:8)(cid:12)(cid:16)(cid:11)(cid:4)(cid:16)(cid:8)
(cid:22)(cid:4)(cid:15)(cid:2)(cid:8) (cid:16)(cid:10)(cid:15)(cid:12)(cid:3)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:4)(cid:3)(cid:7)(cid:15)(cid:6)(cid:12)(cid:16)(cid:4)(cid:18) (cid:7)(cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:18) (cid:6)(cid:17)(cid:21)
& (cid:3)(cid:4)(cid:12)(cid:3)0 (cid:8)
In a statement that is strikingly consis-
tent with the Stanford Encyclopedia of
Philosophy “ common understanding” of
emergence offered above, Anderson ex-
plained his anti-constructionist position.
6(cid:16)(cid:8) (cid:3)(cid:11)(cid:7)(cid:10)(cid:8) (cid:20)(cid:3)& (cid:3)(cid:20)(cid:8) (cid:15)(cid:22)(cid:8) (cid:7)(cid:15)(cid:2)(cid:19) (cid:20)(cid:3)*(cid:17)(cid:16)(cid:30)(cid:8) (cid:3)(cid:6)(cid:16)(cid:17)(cid:4)(cid:3)(cid:20)(cid:30)(cid:8) (cid:6)(cid:3)’ (cid:8)
(cid:19) (cid:4)(cid:15)(cid:19) (cid:3)(cid:4)(cid:16)(cid:17)(cid:3)(cid:12)(cid:8) (cid:11)(cid:19) (cid:19) (cid:3)(cid:11)(cid:4)(cid:27)(cid:8) (cid:28)
(cid:8) (cid:9)7 (cid:14)(cid:6)(cid:3)(cid:8) (cid:2)(cid:11)(cid:30)(cid:8) (cid:11)(cid:4)(cid:4)(cid:11)(cid:30)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)
(cid:12)(cid:7)(cid:17)(cid:3)(cid:6)(cid:7)(cid:3)(cid:12)(cid:8) (cid:4)(cid:15)(cid:18) (cid:5)(cid:10)(cid:20)(cid:30)(cid:8) (cid:20)(cid:17)(cid:6)(cid:3)(cid:11)(cid:4)(cid:20)(cid:30)(cid:8) (cid:17)(cid:6)(cid:8) (cid:9)(cid:16)(cid:10)(cid:3)(cid:8) (cid:22)(cid:15)(cid:20)(cid:20)(cid:15)’ (cid:17)(cid:6)(cid:5)(cid:14)(cid:8)
(cid:10)(cid:17)(cid:3)(cid:4)(cid:11)(cid:4)(cid:7)(cid:10)(cid:30)(cid:8) (cid:9)(cid:17)(cid:6)(cid:8) ’ (cid:10)(cid:17)(cid:7)(cid:10)(cid:14)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)(cid:3)(cid:20)(cid:3)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:4)(cid:30)(cid:8) (cid:3)(cid:6)(cid:16)(cid:17)(cid:16)(cid:17)(cid:3)(cid:12)(cid:8)
(cid:15)(cid:22)(cid:8) (cid:9)(cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:7)(cid:17)(cid:3)(cid:6)(cid:7)(cid:3)(cid:8) (cid:11)(cid:16)(cid:8) (cid:20)(cid:3)& (cid:3)(cid:20)(cid:8) (cid:6)8 (cid:23) (cid:14)(cid:8) (cid:15)(cid:13)(cid:3)(cid:30)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8)
(cid:15)(cid:22)(cid:8) (cid:9)(cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:7)(cid:17)(cid:3)(cid:6)(cid:7)(cid:3)(cid:8) (cid:11)(cid:16)(cid:8) (cid:20)(cid:3)& (cid:3)(cid:20)(cid:8) (cid:6)(cid:14) (cid:8) (cid:3)(cid:20)(cid:3)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:4)(cid:30)(cid:8) (cid:19) (cid:11)(cid:4)(cid:21)
2 Shalizi also offers his own definition of emergence
on his website [Shalizi] as follows.
One set of variables, A, emerges from another, B if
(1) A is a function of B, i.e., at a higher level of
abstraction, and (2) the higher-level variables can
be predicted more efficiently than the lower-level
ones, where "efficiency of prediction" is defined
using information theory.
Emergence Explained
3/67
DRAFT
2/12/2006
(cid:16)(cid:17)(cid:7)(cid:20)(cid:3)(cid:8) (cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:12)-(cid:8) (cid:12)(cid:15)(cid:20)(cid:17)% (cid:8) (cid:12)(cid:16)(cid:11)(cid:16)(cid:3)(cid:8) !(cid:15)(cid:4)(cid:8) (cid:2)(cid:11)(cid:6)(cid:30)(cid:8) (cid:13)(cid:15)% (cid:30)" (cid:8)
(cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:12)-(cid:8) (cid:7)(cid:10)(cid:3)(cid:2)(cid:17)(cid:12)(cid:16)(cid:4)(cid:30)-(cid:8) (cid:2)(cid:15)(cid:20)(cid:3)(cid:7)(cid:18) (cid:20)(cid:11)(cid:4)(cid:8) (cid:13)(cid:17)(cid:15)(cid:20)(cid:15)(cid:5)(cid:30)-(cid:8) (cid:7)(cid:3)(cid:20)(cid:20)(cid:8)
(cid:13)(cid:17)(cid:15)(cid:20)(cid:15)(cid:5)(cid:30)-(cid:8) (cid:28)
-(cid:8) (cid:19) (cid:12)(cid:30)(cid:7)(cid:10)(cid:15)(cid:20)(cid:15)(cid:5)(cid:30)-(cid:8) (cid:12)(cid:15)(cid:7)(cid:17)(cid:11)(cid:20)(cid:8) (cid:12)(cid:7)(cid:17)(cid:3)(cid:6)(cid:7)(cid:3)(cid:12)(cid:27)(cid:8)
1 (cid:18) (cid:16)(cid:8) (cid:16)(cid:10)(cid:17)(cid:12)(cid:8) (cid:10)(cid:17)(cid:3)(cid:4)(cid:11)(cid:4)(cid:7)(cid:10)(cid:30)(cid:8) % (cid:15)(cid:3)(cid:12)(cid:8) (cid:6)(cid:15)(cid:16)(cid:8) (cid:17)(cid:2)(cid:19) (cid:20)(cid:30)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:12)(cid:7)(cid:17)(cid:21)
(cid:3)(cid:6)(cid:7)(cid:3)(cid:8) (cid:9)(cid:6)8 (cid:23) (cid:14)(cid:8) (cid:17)(cid:12)(cid:8) #3(cid:18) (cid:12)(cid:16)(cid:8) (cid:11)(cid:19) (cid:19) (cid:20)(cid:17)(cid:3)% (cid:8) (cid:9)(cid:12)(cid:7)(cid:17)(cid:3)(cid:6)(cid:7)(cid:3)(cid:8) (cid:6)(cid:14)(cid:27)$(cid:8) 6(cid:16)(cid:8)
(cid:3)(cid:11)(cid:7)(cid:10)(cid:8) (cid:9)(cid:20)(cid:3)& (cid:3)(cid:20)(cid:14)(cid:8) (cid:3)(cid:6)(cid:16)(cid:17)(cid:4)(cid:3)(cid:20)(cid:30)(cid:8) (cid:6)(cid:3)’ (cid:8) (cid:20)(cid:11)’ (cid:12)-(cid:8) (cid:7)(cid:15)(cid:6)(cid:7)(cid:3)(cid:19) (cid:16)(cid:12)-(cid:8)
(cid:11)(cid:6)% (cid:8) (cid:5)(cid:3)(cid:6)(cid:3)(cid:4)(cid:11)(cid:20)(cid:17)(cid:31) (cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:6)(cid:3)(cid:7)(cid:3)(cid:12)(cid:12)(cid:11)(cid:4)(cid:30)(cid:27)(cid:8) (cid:28)
(cid:8) 9 (cid:12)(cid:30)(cid:21)
(cid:7)(cid:10)(cid:15)(cid:20)(cid:15)(cid:5)(cid:30)(cid:8) (cid:17)(cid:12)(cid:8) (cid:6)(cid:15)(cid:16)(cid:8) (cid:11)(cid:19) (cid:19) (cid:20)(cid:17)(cid:3)% (cid:8) (cid:13)(cid:17)(cid:15)(cid:20)(cid:15)(cid:5)(cid:30)-(cid:8) (cid:6)(cid:15)(cid:4)(cid:8) (cid:17)(cid:12)(cid:8) (cid:13)(cid:17)(cid:15)(cid:20)(cid:21)
(cid:15)(cid:5)(cid:30)(cid:8) (cid:11)(cid:19) (cid:19) (cid:20)(cid:17)(cid:3)% (cid:8) (cid:7)(cid:10)(cid:3)(cid:2)(cid:17)(cid:12)(cid:16)(cid:4)(cid:30)(cid:27)(cid:8) (cid:28)
(cid:8) ) (cid:10)(cid:3)(cid:8) ’ (cid:10)(cid:15)(cid:20)(cid:3)(cid:8) (cid:13)(cid:3)(cid:21)
(cid:7)(cid:15)(cid:2)(cid:3)(cid:12)(cid:8) (cid:6)(cid:15)(cid:16)(cid:8) (cid:15)(cid:6)(cid:20)(cid:30)(cid:8) (cid:2)(cid:15)(cid:4)(cid:3)(cid:8) (cid:16)(cid:10)(cid:11)(cid:6)(cid:8) (cid:13)(cid:18) (cid:16)(cid:8) & (cid:3)(cid:4)(cid:30)(cid:8) % (cid:17)(cid:22)(cid:22)(cid:3)(cid:4)(cid:21)
(cid:3)(cid:6)(cid:16)(cid:8)(cid:22)(cid:4)(cid:15)(cid:2)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:12)(cid:18) (cid:2)(cid:8)(cid:15)(cid:22)(cid:8)(cid:17)(cid:16)(cid:12)(cid:8)(cid:19) (cid:11)(cid:4)(cid:16)(cid:12)(cid:27)(cid:8)
Although not so labeled, the preceding
provides a good summary of the position
known as functionalism (or in other con-
texts as non-reductive physicalism),
which argues that autonomous laws of
nature appear at many levels.
Anderson thought that the position he
was taking was radical enough—
how
can one be a reductionist, which he
claimed to be, and at the same time ar-
gue that there are autonomous sci-
that it was important to reaffirm
ences —
his adherence
to
reductionism. He
summed this up as follows.
/ (cid:9)) (cid:10)(cid:3)(cid:14)(cid:8) ’ (cid:15)(cid:4), (cid:17)(cid:6)(cid:5)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:11)(cid:20)(cid:20)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:11)(cid:6)(cid:17)(cid:2)(cid:11)(cid:16)(cid:3)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:17)(cid:6)(cid:21)
(cid:11)(cid:6)(cid:17)(cid:2)(cid:11)(cid:16)(cid:3)(cid:8) (cid:2)(cid:11)(cid:16)(cid:16)(cid:3)(cid:4)(cid:8) (cid:15)(cid:22)(cid:8) ’ (cid:10)(cid:17)(cid:7)(cid:10)(cid:8) ’ (cid:3)(cid:8) (cid:10)(cid:11)& (cid:3)(cid:8) (cid:11)(cid:6)(cid:30)(cid:8) % (cid:3)(cid:21)
(cid:16)(cid:11)(cid:17)(cid:20)(cid:3)% (cid:8) , (cid:6)(cid:15)’ (cid:20)(cid:3)% (cid:5)(cid:3)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:11)(cid:20)(cid:20)(cid:8) (cid:28)
(cid:8) (cid:7)(cid:15)(cid:6)(cid:16)(cid:4)(cid:15)(cid:20)(cid:20)(cid:3)% (cid:8) (cid:13)(cid:30)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:8)(cid:12)(cid:11)(cid:2)(cid:3)(cid:8)(cid:12)(cid:3)(cid:16)(cid:8)(cid:15)(cid:22)(cid:8)(cid:22)(cid:18) (cid:6)% (cid:11)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:20)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8)(cid:9)(cid:15)(cid:22)(cid:8)(cid:19) (cid:10)(cid:30)(cid:12)(cid:21)
(cid:17)(cid:7)(cid:12)(cid:14)(cid:27)(cid:8) (cid:28)
(cid:14)(cid:3)(cid:8) (cid:2)(cid:18) (cid:12)(cid:16)(cid:8) (cid:11)(cid:20)(cid:20)(cid:8) (cid:12)(cid:16)(cid:11)(cid:4)(cid:16)(cid:8) ’ (cid:17)(cid:16)(cid:10)(cid:8) (cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:21)
(cid:8) (cid:9)(cid:29)
(cid:17)(cid:12)(cid:2)-(cid:8)’ (cid:10)(cid:17)(cid:7)(cid:10)(cid:8)+(cid:8)(cid:22)(cid:18) (cid:20)(cid:20)(cid:30)(cid:8)(cid:11)(cid:7)(cid:7)(cid:3)(cid:19) (cid:16)(cid:27)0 (cid:8)
In the rest of this paper, we elaborate
and extend the position that Anderson
set forth. We hope to offer a coherent
explanation for how nature can be both
reductive and non-reductive simultane-
ously.
Much of our approach is derived from
concepts borrowed from Computer Sci-
ence— which more than any other hu-
man endeavor has had to deal, on the
most concrete terms, with the question
of how one can operate on many levels
simultaneously. How is it possible for
such the amazing panoply of functional-
ity created by software to emerge from
(nothing but) electrons in motion?
The rest of this paper is organized as fol-
lows.
• Section 2 develops basic concepts. It
explores the notions of reductionism
and functionalism, and it characterizes
their differences and points of agree-
ment. It defines the term epiphenome-
non. It explicates the notion of super-
venience and points out an important
limitation. It argues that one must
chose between force reductionism and
the position that new forces of nature
come into being mysteriously.
• Section 3 uses the Game of Life to il-
lustrate and then to define emergence.
• Section 4 explores some of the impli-
cations of our definition. It defines the
notion of downward entailment. It dis-
cusses the reality of higher level ab-
stractions, and it offers a novel view of
phase transitions.
• Section 5 defines the notion of an en-
tity as a persistent region of reduced
entropy. It relates the concepts of enti-
ties, dissipative structures, and auton-
omy. It shows why emergence is a
fundamental feature of nature. It dis-
tinguishes natural
from
artificial
autonomous entities. It shows why su-
pervenience is not as powerful a con-
cept as one might have hoped. It dis-
cusses
the conceptual
limitations
Computer Science suffers as a result of
its self-imposed exile to a world of free
energy.
• Section 6 discusses stigmergy, histori-
cal contingency, and the evolution of
complexity.
Emergence Explained
4/67
DRAFT
2/12/2006
• Section 7 presents additional implica-
tions for science of entities, emer-
gence, and complexity.
• Section 8 presents a framework for the
varieties of emergence that we discuss.
• Section 9 offers some practical advice.
about service-oriented architectures,
stove-piped systems, and the limita-
tions of modeling.
• Section 10 provides brief a summary
and includes a remark about an area
for future investigation.
• The Appendix offers a formal defini-
tion of Games of Life patterns such as
the glider. It shows how such patterns
can be used to create an API of pat-
terns. It presents some basic unsolva-
bility results.
2 Background and founda-
tions
We begin by contrasting reductionism
and functionalism. We use papers writ-
ten by Steven Weinberg, a reductionist
physicist, and Jerrold (Jerry) Fodor, a
functionalist philosopher, as our points
of departure.
• Weinberg, a professor of physics at
the University of Texas, Austin, was
awarded the Nobel prize in physics
for his work on unifying the weak
and electromagnetic forces. He is an
articulate spokesperson for a con-
structivist-like form of reductionism,
the position that all of science can be
reduced to and are mathematical
consequences of the laws of physics.
He presented his views in “ Reduc-
tionism Redux ” [Weinberg], an es-
say in which he responded to posi-
tion papers of participants in a debate
on reductionism held in 1992 at Je-
sus College, Cambridge University.
• Fodor, a professor of philosophy at
Rutgers, is one of the founders of
functionalism, the position that regu-
larities appear at all levels of science
and that these regularities are not re-
ducible to physics. He reviewed [Fo-
dor 98] his position in “ Special Sci-
ences; Still Autonomous after All
These Years,” which he wrote in re-
ply to a more reductionist position
taken [Kim ’ 92, ‘ 93] by Jaegwon
Kim, a professor of philosophy at
Brown University.
In the end we will claim that it was nei-
ther the reductionist Weinberg nor the
functionalist Fodor but Anderson — who
proposed a marriage between reduction-
ism and functionalism — who was right.3
Our job will be to show how the improb-
able couple can live happily ever after.
2.1 Functionalism
Functionalism [Fodor 74] holds that
there are so-called ‘ special sciences ’ (in
fact, all sciences other than physics and
perhaps chemistry) that study regulari-
ties in nature that are in some sense
autonomous of physics. In [Fodor 98]
Fodor wrote the following reaffirmation
of functionalism.
) (cid:10)(cid:3)(cid:8) (cid:14) (cid:9) (cid:11) (cid:15) (cid:3) (cid:9) (cid:16) (cid:17)(cid:18) (cid:19)(cid:9) (cid:5) (cid:13) (cid:9) (cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:19) (cid:3)(cid:7)(cid:17)(cid:11)(cid:20)(cid:8) (cid:12)(cid:7)(cid:17)(cid:3)(cid:6)(cid:7)(cid:3)(cid:12)(cid:8)
(cid:16)(cid:3)(cid:12)(cid:16)(cid:17)(cid:22)(cid:17)(cid:3)(cid:12)(cid:8) (cid:16)(cid:15)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:4)(cid:3)(cid:20)(cid:17)(cid:11)(cid:13)(cid:20)(cid:3)(cid:8) (cid:2)(cid:11)(cid:7)(cid:4)(cid:15)(cid:20)(cid:3)& (cid:3)(cid:20)(cid:8) (cid:4)(cid:3)(cid:5)(cid:18) (cid:20)(cid:11)(cid:4)(cid:17)(cid:21)
(cid:16)(cid:17)(cid:3)(cid:12)(cid:8)
(cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:4)(cid:3)(cid:11)(cid:20)(cid:17)(cid:31) (cid:3)% (cid:8) (cid:13)(cid:30)(cid:8) (cid:2)(cid:3)(cid:7)(cid:10)(cid:11)(cid:6)(cid:17)(cid:12)(cid:2)(cid:12)(cid:8)
’ (cid:10)(cid:15)(cid:12)(cid:3)(cid:8) (cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:11)(cid:20)(cid:8) (cid:12)(cid:18) (cid:13)(cid:12)(cid:16)(cid:11)(cid:6)(cid:7)(cid:3)(cid:8) (cid:17)(cid:12)(cid:8) ( (cid:18) (cid:17)(cid:16)(cid:3)(cid:8) (cid:16)(cid:30)(cid:19) (cid:17)(cid:7)(cid:11)(cid:20)(cid:20)(cid:30)(cid:8)
(cid:10)(cid:3)(cid:16)(cid:3)(cid:4)(cid:15)(cid:5)(cid:3)(cid:6)(cid:3)(cid:15)(cid:18) (cid:12)(cid:27)(cid:8) : (cid:15)(cid:3)(cid:12)(cid:8)(cid:11)(cid:6)(cid:30)(cid:13)(cid:15)% (cid:30)(cid:8) (cid:4)(cid:3)(cid:11)(cid:20)(cid:20)(cid:30)(cid:8) % (cid:15)(cid:18) (cid:13)(cid:16)(cid:8)
(cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:2)(cid:15)(cid:18) (cid:6)(cid:16)(cid:11)(cid:17)(cid:6)(cid:12)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:2)(cid:11)% (cid:3)(cid:8) (cid:15)(cid:22)(cid:8) (cid:11)(cid:20)(cid:20)(cid:8) (cid:12)(cid:15)(cid:4)(cid:16)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8)
(cid:12)(cid:16)(cid:18) (cid:22)(cid:22)4 (cid:8): (cid:15)(cid:3)(cid:12)(cid:8)(cid:11)(cid:6)(cid:30)(cid:13)(cid:15)% (cid:30)(cid:8)(cid:4)(cid:3)(cid:11)(cid:20)(cid:20)(cid:30)(cid:8)(cid:16)(cid:10)(cid:17)(cid:6), (cid:8)(cid:16)(cid:10)(cid:11)(cid:16)-(cid:8)(cid:12)(cid:17)(cid:6)(cid:7)(cid:3)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:30)(cid:8) (cid:11)(cid:4)(cid:3)-(cid:8) (cid:5)(cid:3)(cid:6)(cid:3)(cid:4)(cid:11)(cid:20)(cid:17)(cid:31) (cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8) (cid:2)(cid:15)(cid:18) (cid:6)(cid:16)(cid:11)(cid:17)(cid:6)(cid:12)(cid:21)
(cid:11)(cid:12)(cid:21)(cid:12)(cid:18) (cid:7)(cid:10)(cid:8) ’ (cid:15)(cid:6)$(cid:16)(cid:8) (cid:7)(cid:15)(cid:6)(cid:16)(cid:17)(cid:6)(cid:18) (cid:3)(cid:8) (cid:16)(cid:15)(cid:8) (cid:12)(cid:3)(cid:4)& (cid:3)(cid:8) (cid:5)(cid:3)(cid:15)(cid:20)(cid:15)(cid:5)(cid:30)(cid:8) (cid:17)(cid:6)(cid:8)
(cid:5)(cid:15)(cid:15)% (cid:8) (cid:12)(cid:16)(cid:3)(cid:11)% 4 (cid:8) : (cid:11)(cid:2)(cid:6)(cid:8) (cid:6)(cid:3)(cid:11)(cid:4)(cid:8) (cid:3)& (cid:3)(cid:4)(cid:30)(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:8) ’ (cid:3)(cid:8)
, (cid:6)(cid:15)’ (cid:8) (cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) ’ (cid:15)(cid:4)(cid:20)% (cid:8) (cid:12)(cid:18) (cid:5)(cid:5)(cid:3)(cid:12)(cid:16)(cid:12)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:18) (cid:6)(cid:17)(cid:21)
3 All of the extracts from Anderson, Fodor, and
Weinberg are from the three papers cited above.
The emphases in the extracts are all in the origi-
nals.
Emergence Explained
5/67
DRAFT
2/12/2006
(cid:2)(cid:11)(cid:5)(cid:17)(cid:6)(cid:11)(cid:13)(cid:20)(cid:30)(cid:8) (cid:7)(cid:15)(cid:2)(cid:19) (cid:20)(cid:17)(cid:7)(cid:11)(cid:16)(cid:3)% (cid:8) (cid:16)(cid:15)(cid:21)(cid:17)(cid:6)(cid:5)(cid:12)(cid:8)(cid:11)(cid:6)% (cid:8) (cid:22)(cid:4)(cid:15)(cid:21)(cid:17)(cid:6)(cid:5)(cid:12)(cid:8)
(cid:15)(cid:22)(cid:8) (cid:13)(cid:17)(cid:16)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:19) (cid:17)(cid:3)(cid:7)(cid:3)(cid:12)(cid:8) (cid:11)(cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:3)*(cid:16)(cid:4)(cid:3)(cid:2)(cid:3)(cid:8) (cid:10)(cid:17)(cid:20)
(cid:13) (cid:11) (cid:21) (cid:20)(cid:3)& (cid:3)(cid:20)(cid:8) (cid:2)(cid:11)(cid:6)(cid:11)(cid:5)(cid:3)(cid:8) (cid:12)(cid:15)(cid:2)(cid:3)(cid:10)(cid:15)’ (cid:8) (cid:16)(cid:15)(cid:8) (cid:7)(cid:15)(cid:6)& (cid:3)(cid:4)(cid:5)(cid:3)(cid:8) (cid:15)(cid:6)(cid:8)
(cid:12)(cid:16)(cid:11)(cid:13)(cid:20)(cid:3)(cid:8)(cid:10)(cid:22) (cid:13) (cid:11) (cid:21) (cid:23)(cid:9) (cid:14) (cid:9) (cid:23)(cid:8)(cid:19) (cid:4)(cid:15)(cid:19) (cid:3)(cid:4)(cid:16)(cid:17)(cid:3)(cid:12)(cid:27)(cid:8)
Although Fodor does not use the term,
the phenomena studied by the special
sciences are the same sort of phenomena
that we now call multiscale, i.e., emer-
gent.
Why is there emergence? Fodor contin-
ues as follows.
(cid:9)) (cid:14)(cid:10)(cid:3)(cid:8) #(cid:12)(cid:15)(cid:2)(cid:3)(cid:10)(cid:15)’ $(cid:8) (cid:9)(cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:19) (cid:4)(cid:3)(cid:7)(cid:3)% (cid:17)(cid:6)(cid:5)(cid:8) (cid:3)*(cid:16)(cid:4)(cid:11)(cid:7)(cid:16)(cid:14)(cid:8)
(cid:4)(cid:3)(cid:11)(cid:20)(cid:20)(cid:30)(cid:8)(cid:17)(cid:12)(cid:8)(cid:3)(cid:6)(cid:16)(cid:17)(cid:4)(cid:3)(cid:20)(cid:30)(cid:8)(cid:2)(cid:30)(cid:12)(cid:16)(cid:3)(cid:4)(cid:17)(cid:15)(cid:18) (cid:12)(cid:8)(cid:28)
(cid:8)(cid:27)(cid:8)(cid:8)
. (cid:15)-(cid:8) (cid:16)(cid:10)(cid:3)(cid:6)-(cid:8) (cid:24) (cid:25) (cid:15) (cid:3) (cid:17)(cid:18) (cid:3) (cid:19)(cid:25) (cid:9) (cid:11) (cid:9) (cid:3) (cid:22) (cid:5) (cid:15) (cid:19)(cid:25) (cid:17)(cid:5) (cid:12) (cid:3) (cid:9) (cid:16) (cid:13) (cid:9) (cid:26) (cid:19)(cid:3)
(cid:26) (cid:25) (cid:15) (cid:18) (cid:17)(cid:13) (cid:18) (cid:27) (cid:8) (cid:28)
(cid:8) (cid:29) (cid:3)(cid:20)(cid:20)-(cid:8) +(cid:8) (cid:11)% (cid:2)(cid:17)(cid:16)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) +(cid:8) % (cid:15)(cid:6)$(cid:16)(cid:8) , (cid:6)(cid:15)’ (cid:8)
’ (cid:10)(cid:30)(cid:27)(cid:8) +(cid:8) % (cid:15)(cid:6)$(cid:16)(cid:8) (cid:3)& (cid:3)(cid:6)(cid:8) , (cid:6)(cid:15)’ (cid:8) (cid:10)(cid:15)’ (cid:8) (cid:16)(cid:15)(cid:8) (cid:16)(cid:10)(cid:17)(cid:6), (cid:8) (cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8)
’ (cid:10)(cid:30)(cid:27)(cid:8) +(cid:8) (cid:3)*(cid:19) (cid:3)(cid:7)(cid:16)(cid:8) (cid:16)(cid:15)(cid:8) (cid:22)(cid:17)(cid:5)(cid:18) (cid:4)(cid:3)(cid:8) (cid:15)(cid:18) (cid:16)(cid:8) ’ (cid:10)(cid:30)(cid:8) (cid:16)(cid:10)(cid:3)(cid:4)(cid:3)(cid:8) (cid:17)(cid:12)(cid:8)
(cid:11)(cid:6)(cid:30)(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:8) (cid:3)*(cid:7)(cid:3)(cid:19) (cid:16)(cid:8) (cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:12)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) % (cid:11)(cid:30)(cid:8) (cid:13)(cid:3)(cid:22)(cid:15)(cid:4)(cid:3)(cid:8) +(cid:8)
(cid:22)(cid:17)(cid:5)(cid:18) (cid:4)(cid:3)(cid:8)(cid:15)(cid:18) (cid:16)(cid:8)’ (cid:10)(cid:30)(cid:8)(cid:16)(cid:10)(cid:3)(cid:4)(cid:3)(cid:8)(cid:17)(cid:12)(cid:8)(cid:11)(cid:6)(cid:30)(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:8)(cid:11)(cid:16)(cid:8)(cid:11)(cid:20)(cid:20)(cid:8)(cid:28)
(cid:8)(cid:27)(cid:8)
So, like Holland, Fodor throws up his
hands with respect to explaining emer-
gence.
One of the tenets of functionalism is that
within any domain, it is the regularities
that appear at the level of phenomena
with which the domain is concerned that
are important. It is not significant how
those regularities are realized in terms of
lower level phenomena —
both because
that doesn’ t matter and because they can
often be implemented in any of a num-
ber of ways. A term commonly used in
Functionalism is multiple realizability,
which refers to the notion that many
regularities (many functions) can be re-
alized in multiple ways. As Fodor re-
marks,
(cid:16)(cid:10)(cid:11)(cid:16)$(cid:12)(cid:8) ’ (cid:10)(cid:30)(cid:8) (cid:4)(cid:3)(cid:22)(cid:3)(cid:4)(cid:3)(cid:6)(cid:7)(cid:3)(cid:12)(cid:8) (cid:16)(cid:15)(cid:8) (cid:7)(cid:11)(cid:6)(cid:8) (cid:15)(cid:19) (cid:3)(cid:6)(cid:3)(cid:4)(cid:12)-(cid:8)
(cid:2)(cid:15)(cid:18) (cid:12)(cid:3)(cid:16)(cid:4)(cid:11)(cid:19) (cid:12)-(cid:8) (cid:7)(cid:11)(cid:2)(cid:12)(cid:10)(cid:11)(cid:22)(cid:16)(cid:12)-(cid:8) (cid:7)(cid:11)(cid:20)(cid:7)(cid:18) (cid:20)(cid:11)(cid:16)(cid:15)(cid:4)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8)
(cid:16)(cid:10)(cid:3)(cid:8) (cid:20)(cid:17), (cid:3)(cid:8) (cid:13)(cid:3)(cid:12)(cid:16)(cid:4)(cid:3)’ (cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:19) (cid:11)(cid:5)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:22)(cid:18) (cid:6)(cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:11)(cid:20)(cid:17)(cid:12)(cid:16)(cid:8)
(cid:19) (cid:10)(cid:17)(cid:20)(cid:15)(cid:12)(cid:15)(cid:19) (cid:10)(cid:30)(cid:27)(cid:8) ) (cid:15)(cid:8)(cid:2)(cid:11), (cid:3)(cid:8)(cid:11)(cid:8)(cid:13)(cid:3)(cid:16)(cid:16)(cid:3)(cid:4)(cid:8)(cid:2)(cid:15)(cid:18) (cid:12)(cid:3)(cid:16)(cid:4)(cid:11)(cid:19) (cid:8) (cid:17)(cid:18) (cid:8)
(cid:16)(cid:15)(cid:8) % (cid:3)& (cid:17)(cid:12)(cid:3)(cid:8) (cid:11)(cid:8) (cid:6)(cid:3)’ (cid:8) , (cid:17)(cid:6)% (cid:8) (cid:15)(cid:22)(cid:8)(cid:2)(cid:3)(cid:7)(cid:10)(cid:11)(cid:6)(cid:17)(cid:12)(cid:2)(cid:8) ’ (cid:10)(cid:15)(cid:12)(cid:3)(cid:8)
(cid:13)(cid:3)(cid:10)(cid:11)& (cid:17)(cid:15)(cid:4)(cid:8) (cid:17)(cid:12)(cid:8) (cid:4)(cid:3)(cid:20)(cid:17)(cid:11)(cid:13)(cid:20)(cid:3)(cid:8) ’ (cid:17)(cid:16)(cid:10)(cid:8) (cid:4)(cid:3)(cid:12)(cid:19) (cid:3)(cid:7)(cid:16)(cid:8) (cid:16)(cid:15)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)
(cid:10)(cid:17)(cid:5)(cid:10)(cid:21)(cid:20)(cid:3)& (cid:3)(cid:20)(cid:8) (cid:4)(cid:3)(cid:5)(cid:18) (cid:20)(cid:11)(cid:4)(cid:17)(cid:16)(cid:30)(cid:8) / (cid:20)(cid:17)& (cid:3)(cid:8) (cid:2)(cid:15)(cid:18) (cid:12)(cid:3)(cid:8) (cid:17)(cid:6)-(cid:8) % (cid:3)(cid:11)% (cid:8)
(cid:2)(cid:15)(cid:18) (cid:12)(cid:3)(cid:8)(cid:15)(cid:18) (cid:16)(cid:27)0 (cid:8)(cid:8)
2.2 Reductionism
Taking the other side of the debate is
Steven Weinberg, one of the most articu-
late defenders of reductionism. Wein-
berg distinguishes two kinds of reduc-
tionism.
(cid:29) (cid:3)(cid:8) (cid:15)(cid:18) (cid:5)(cid:10)(cid:16)(cid:8) (cid:22)(cid:17)(cid:4)(cid:12)(cid:16)(cid:8) (cid:15)(cid:22)(cid:8) (cid:11)(cid:20)(cid:20)(cid:8) (cid:16)(cid:15)(cid:8) % (cid:17)(cid:12)(cid:16)(cid:17)(cid:6)(cid:5)(cid:18) (cid:17)(cid:12)(cid:10)(cid:8) (cid:13)(cid:3)(cid:16)’ (cid:3)(cid:3)(cid:6)(cid:8)
’ (cid:10)(cid:11)(cid:16)(cid:8) ! (cid:16)(cid:15)(cid:8) (cid:13)(cid:15)(cid:4)(cid:4)(cid:15)’ (cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:20)(cid:11)(cid:6)(cid:5)(cid:18) (cid:11)(cid:5)(cid:3)(cid:8) (cid:15)(cid:22)(cid:8) (cid:7)(cid:4)(cid:17)(cid:2)(cid:17)(cid:6)(cid:11)(cid:20)(cid:8)
(cid:20)(cid:11)’ " (cid:8)+(cid:8) (cid:20)(cid:17), (cid:3)(cid:8)(cid:16)(cid:15)(cid:8)(cid:7)(cid:11)(cid:20)(cid:20)(cid:8)(cid:5)(cid:4)(cid:11)(cid:6)% (cid:8)(cid:11)(cid:6)% (cid:8)(cid:19) (cid:3)(cid:16)(cid:16)(cid:30)(cid:8)(cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:21)
(cid:17)(cid:12)(cid:2)(cid:27)(cid:8); (cid:4)(cid:11)(cid:6)% (cid:8)(cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:17)(cid:12)(cid:2)(cid:8) (cid:17)(cid:12)(cid:8)(cid:28)
(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)& (cid:17)(cid:3)’ (cid:8)(cid:16)(cid:10)(cid:11)(cid:16)(cid:8)
(cid:11)(cid:20)(cid:20)(cid:8) (cid:15)(cid:22)(cid:8) (cid:6)(cid:11)(cid:16)(cid:18) (cid:4)(cid:3)(cid:8) (cid:17)(cid:12)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) ’ (cid:11)(cid:30)(cid:8) (cid:17)(cid:16)(cid:8) (cid:17)(cid:12)(cid:8) ! ’ (cid:17)(cid:16)(cid:10)(cid:8) (cid:7)(cid:3)(cid:4)(cid:16)(cid:11)(cid:17)(cid:6)(cid:8)
( (cid:18) (cid:11)(cid:20)(cid:17)(cid:22)(cid:17)(cid:7)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)(cid:8) (cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8) (cid:17)(cid:6)(cid:17)(cid:16)(cid:17)(cid:11)(cid:20)(cid:8) (cid:7)(cid:15)(cid:6)% (cid:17)(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8)
(cid:10)(cid:17)(cid:12)(cid:16)(cid:15)(cid:4)(cid:17)(cid:7)(cid:11)(cid:20)(cid:8) (cid:11)(cid:7)(cid:7)(cid:17)% (cid:3)(cid:6)(cid:16)(cid:12)" (cid:8) (cid:13)(cid:3)(cid:7)(cid:11)(cid:18) (cid:12)(cid:3)(cid:8) (cid:15)(cid:22)(cid:8) (cid:12)(cid:17)(cid:2)(cid:19) (cid:20)(cid:3)(cid:8)
(cid:18) (cid:6)(cid:17)& (cid:3)(cid:4)(cid:12)(cid:11)(cid:20)(cid:8) (cid:20)(cid:11)’ (cid:12)-(cid:8) (cid:16)(cid:15)(cid:8) ’ (cid:10)(cid:17)(cid:7)(cid:10)(cid:8) (cid:11)(cid:20)(cid:20)(cid:8) (cid:15)(cid:16)(cid:10)(cid:3)(cid:4)(cid:8) (cid:12)(cid:7)(cid:17)(cid:3)(cid:6)(cid:16)(cid:17)(cid:22)(cid:17)(cid:7)(cid:8)
(cid:20)(cid:11)’ (cid:12)(cid:8)(cid:2)(cid:11)(cid:30)(cid:8) (cid:17)(cid:6)(cid:8)(cid:12)(cid:15)(cid:2)(cid:3)(cid:8)(cid:12)(cid:3)(cid:6)(cid:12)(cid:3)(cid:8)(cid:13)(cid:3)(cid:8)(cid:4)(cid:3)% (cid:18) (cid:7)(cid:3)% (cid:27)(cid:8)9 (cid:3)(cid:16)(cid:16)(cid:30)(cid:8)
(cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:17)(cid:12)(cid:2)(cid:8) (cid:17)(cid:12)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:2)(cid:18) (cid:7)(cid:10)(cid:8) (cid:20)(cid:3)(cid:12)(cid:12)(cid:8) (cid:17)(cid:6)(cid:16)(cid:3)(cid:4)(cid:3)(cid:12)(cid:16)(cid:17)(cid:6)(cid:5)(cid:8)
% (cid:15)(cid:7)(cid:16)(cid:4)(cid:17)(cid:6)(cid:3)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:12)(cid:8) (cid:13)(cid:3)(cid:10)(cid:11)& (cid:3)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) ’ (cid:11)(cid:30)(cid:8) (cid:16)(cid:10)(cid:3)(cid:30)(cid:8)
% (cid:15)(cid:8) (cid:13)(cid:3)(cid:7)(cid:11)(cid:18) (cid:12)(cid:3)(cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:19) (cid:4)(cid:15)(cid:19) (cid:3)(cid:4)(cid:16)(cid:17)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:17)(cid:4)(cid:8) (cid:7)(cid:15)(cid:6)(cid:21)
(cid:12)(cid:16)(cid:17)(cid:16)(cid:18) (cid:3)(cid:6)(cid:16)(cid:12) (cid:8) (cid:22)(cid:15)(cid:4)(cid:8) (cid:17)(cid:6)(cid:12)(cid:16)(cid:11)(cid:6)(cid:7)(cid:3)-(cid:8) (cid:11)(cid:8) % (cid:17)(cid:11)(cid:2)(cid:15)(cid:6)% (cid:8) (cid:17)(cid:12)(cid:8) (cid:10)(cid:11)(cid:4)% (cid:8)
(cid:13)(cid:3)(cid:7)(cid:11)(cid:18) (cid:12)(cid:3)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:7)(cid:11)(cid:4)(cid:13)(cid:15)(cid:6)(cid:8) (cid:11)(cid:16)(cid:15)(cid:2)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) ’ (cid:10)(cid:17)(cid:7)(cid:10)(cid:8) (cid:17)(cid:16)(cid:8) (cid:17)(cid:12)(cid:8)
(cid:7)(cid:15)(cid:2)(cid:19) (cid:15)(cid:12)(cid:3)% (cid:8)(cid:7)(cid:11)(cid:6)(cid:8)(cid:22)(cid:17)(cid:16)(cid:8)(cid:16)(cid:15)(cid:5)(cid:3)(cid:16)(cid:10)(cid:3)(cid:4)(cid:8)(cid:6)(cid:3)(cid:11)(cid:16)(cid:20)(cid:30)(cid:27)(cid:8)(cid:28)
(cid:8)
9 (cid:3)(cid:16)(cid:16)(cid:30)(cid:8) (cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:17)(cid:12)(cid:2)(cid:8) (cid:17)(cid:12)(cid:8)(cid:6)(cid:15)(cid:16)(cid:8) ’ (cid:15)(cid:4)(cid:16)(cid:10)(cid:8)(cid:11)(cid:8) (cid:22)(cid:17)(cid:3)(cid:4)(cid:7)(cid:3)(cid:8) % (cid:3)(cid:21)
(cid:22)(cid:3)(cid:6)(cid:12)(cid:3)(cid:27)(cid:8) (cid:28)
(cid:8) +(cid:6)(cid:8) (cid:22)(cid:11)(cid:7)(cid:16)-(cid:8) (cid:19) (cid:3)(cid:16)(cid:16)(cid:30)(cid:8) (cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:17)(cid:12)(cid:2)(cid:8) (cid:17)(cid:6)(cid:8)
(cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:12)(cid:8) (cid:10)(cid:11)(cid:12)(cid:8) (cid:19) (cid:4)(cid:15)(cid:13)(cid:11)(cid:13)(cid:20)(cid:30)(cid:8) (cid:4)(cid:18) (cid:6)(cid:8) (cid:17)(cid:16)(cid:12)(cid:8) (cid:7)(cid:15)(cid:18) (cid:4)(cid:12)(cid:3)(cid:27)(cid:8) < (cid:18) (cid:12)(cid:16)(cid:8)
(cid:11)(cid:12)(cid:8) (cid:17)(cid:16)(cid:8) % (cid:15)(cid:3)(cid:12)(cid:6)=(cid:16)(cid:8) (cid:2)(cid:11), (cid:3)(cid:8) (cid:12)(cid:3)(cid:6)(cid:12)(cid:3)(cid:8) (cid:16)(cid:15)(cid:8) (cid:16)(cid:11)(cid:20), (cid:8) (cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)
(cid:10)(cid:11)(cid:4)% (cid:6)(cid:3)(cid:12)(cid:12)(cid:8) (cid:15)(cid:4)(cid:8) (cid:16)(cid:3)(cid:2)(cid:19) (cid:3)(cid:4)(cid:11)(cid:16)(cid:18) (cid:4)(cid:3)(cid:8) (cid:15)(cid:4)(cid:8) (cid:17)(cid:6)(cid:16)(cid:3)(cid:20)(cid:20)(cid:17)(cid:5)(cid:3)(cid:6)(cid:7)(cid:3)(cid:8) (cid:15)(cid:22)(cid:8)
(cid:17)(cid:6)% (cid:17)& (cid:17)% (cid:18) (cid:11)(cid:20)(cid:8) > (cid:3)(cid:20)(cid:3)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:4)(cid:30)> (cid:8) (cid:19) (cid:11)(cid:4)(cid:16)(cid:17)(cid:7)(cid:20)(cid:3)(cid:12)-(cid:8) (cid:17)(cid:16)(cid:8) (cid:17)(cid:12)(cid:8) (cid:11)(cid:20)(cid:12)(cid:15)(cid:8)
(cid:6)(cid:15)(cid:16)(cid:8) (cid:19) (cid:15)(cid:12)(cid:12)(cid:17)(cid:13)(cid:20)(cid:3)(cid:8) (cid:16)(cid:15)(cid:8) (cid:5)(cid:17)& (cid:3)(cid:8) (cid:11)(cid:8) (cid:19) (cid:4)(cid:3)(cid:7)(cid:17)(cid:12)(cid:3)(cid:8) (cid:2)(cid:3)(cid:11)(cid:6)(cid:17)(cid:6)(cid:5)(cid:8) (cid:16)(cid:15)(cid:8)
(cid:12)(cid:16)(cid:11)(cid:16)(cid:3)(cid:2)(cid:3)(cid:6)(cid:16)(cid:12)(cid:8) (cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8)
(cid:19) (cid:11)(cid:4)(cid:16)(cid:17)(cid:7)(cid:20)(cid:3)(cid:12)(cid:8) (cid:13)(cid:3)(cid:17)(cid:6)(cid:5)(cid:8) (cid:7)(cid:15)(cid:2)(cid:21)
(cid:19) (cid:15)(cid:12)(cid:3)% (cid:8) (cid:15)(cid:22)(cid:8) (cid:15)(cid:16)(cid:10)(cid:3)(cid:4)(cid:8) (cid:19) (cid:11)(cid:4)(cid:16)(cid:17)(cid:7)(cid:20)(cid:3)(cid:12)(cid:27)(cid:8) (cid:29) (cid:3)(cid:8) % (cid:15)(cid:8) (cid:12)(cid:19) (cid:3)(cid:11), (cid:8)
(cid:20)(cid:15)(cid:15)(cid:12)(cid:3)(cid:20)(cid:30)(cid:8) (cid:15)(cid:22)(cid:8) (cid:11)(cid:8) (cid:19) (cid:4)(cid:15)(cid:16)(cid:15)(cid:6)(cid:8) (cid:11)(cid:12)(cid:8) (cid:13)(cid:3)(cid:17)(cid:6)(cid:5)(cid:8) (cid:7)(cid:15)(cid:2)(cid:19) (cid:15)(cid:12)(cid:3)% (cid:8) (cid:15)(cid:22)(cid:8)
(cid:16)(cid:10)(cid:4)(cid:3)(cid:3)(cid:8)( (cid:18) (cid:11)(cid:4), (cid:12)-(cid:8)(cid:13)(cid:18) (cid:16)(cid:8) (cid:17)(cid:22)(cid:8)(cid:30)(cid:15)(cid:18) (cid:8) (cid:20)(cid:15)(cid:15), (cid:8) & (cid:3)(cid:4)(cid:30)(cid:8)(cid:7)(cid:20)(cid:15)(cid:12)(cid:3)(cid:20)(cid:30)(cid:8)(cid:11)(cid:16)(cid:8)
(cid:11)(cid:8) ( (cid:18) (cid:11)(cid:4), (cid:8) (cid:30)(cid:15)(cid:18) (cid:8) ’ (cid:17)(cid:20)(cid:20)(cid:8) (cid:22)(cid:17)(cid:6)% (cid:8) (cid:17)(cid:16)(cid:8) (cid:12)(cid:18) (cid:4)(cid:4)(cid:15)(cid:18) (cid:6)% (cid:3)% (cid:8) ’ (cid:17)(cid:16)(cid:10)(cid:8) (cid:11)(cid:8)
(cid:7)(cid:20)(cid:15)(cid:18) % (cid:8) (cid:15)(cid:22)(cid:8) ( (cid:18) (cid:11)(cid:4), (cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:11)(cid:6)(cid:16)(cid:17)(cid:21)( (cid:18) (cid:11)(cid:4), (cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:15)(cid:16)(cid:10)(cid:3)(cid:4)(cid:8)
(cid:19) (cid:11)(cid:4)(cid:16)(cid:17)(cid:7)(cid:20)(cid:3)(cid:12)-(cid:8) (cid:15)(cid:7)(cid:7)(cid:11)(cid:12)(cid:17)(cid:15)(cid:6)(cid:11)(cid:20)(cid:20)(cid:30)(cid:8) (cid:13)(cid:15)(cid:18) (cid:6)% (cid:8) (cid:17)(cid:6)(cid:16)(cid:15)(cid:8) (cid:19) (cid:4)(cid:15)(cid:16)(cid:15)(cid:6)(cid:12)?(cid:8)
(cid:12)(cid:15)(cid:8)(cid:11)(cid:16)(cid:8) (cid:20)(cid:3)(cid:11)(cid:12)(cid:16)(cid:8)(cid:22)(cid:15)(cid:4)(cid:8)(cid:11)(cid:8)(cid:13)(cid:4)(cid:17)(cid:3)(cid:22)(cid:8)(cid:2)(cid:15)(cid:2)(cid:3)(cid:6)(cid:16)(cid:8)’ (cid:3)(cid:8)(cid:7)(cid:15)(cid:18) (cid:20)% (cid:8)(cid:12)(cid:11)(cid:30)(cid:8)
(cid:16)(cid:10)(cid:11)(cid:16)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)( (cid:18) (cid:11)(cid:4), (cid:8)(cid:17)(cid:12)(cid:8)(cid:2)(cid:11)% (cid:3)(cid:8)(cid:15)(cid:22)(cid:8)(cid:19) (cid:4)(cid:15)(cid:16)(cid:15)(cid:6)(cid:12)(cid:27)(cid:8)(cid:8)
Weinberg continues his explication of
grand reductionism by using the weather
as an example.
(cid:9)) (cid:14)(cid:10)(cid:3)(cid:8)(cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:17)(cid:12)(cid:16)(cid:8)(cid:4)(cid:3)(cid:5)(cid:11)(cid:4)% (cid:12)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:5)(cid:3)(cid:6)(cid:3)(cid:4)(cid:11)(cid:20)(cid:8)(cid:16)(cid:10)(cid:3)(cid:15)(cid:21)
(cid:4)(cid:17)(cid:3)(cid:12)(cid:8) (cid:5)(cid:15)& (cid:3)(cid:4)(cid:6)(cid:17)(cid:6)(cid:5)(cid:8) (cid:11)(cid:17)(cid:4)(cid:8) (cid:11)(cid:6)% (cid:8) ’ (cid:11)(cid:16)(cid:3)(cid:4)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:4)(cid:11)% (cid:17)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8)
Emergence Explained
6/67
DRAFT
2/12/2006
(cid:11)(cid:12)(cid:8) (cid:13)(cid:3)(cid:17)(cid:6)(cid:5)(cid:8) (cid:11)(cid:16)(cid:8) (cid:11)(cid:8) % (cid:3)(cid:3)(cid:19) (cid:3)(cid:4)(cid:8) (cid:20)(cid:3)& (cid:3)(cid:20)(cid:8) (cid:16)(cid:10)(cid:11)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:15)(cid:4)(cid:17)(cid:3)(cid:12)(cid:8)
(cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8) (cid:7)(cid:15)(cid:20)% (cid:8) (cid:22)(cid:4)(cid:15)(cid:6)(cid:16)(cid:12)(cid:8) (cid:15)(cid:4)(cid:8) (cid:16)(cid:10)(cid:18) (cid:6)% (cid:3)(cid:4)(cid:12)(cid:16)(cid:15)(cid:4)(cid:2)(cid:12)-(cid:8) (cid:6)(cid:15)(cid:16)(cid:8) (cid:17)(cid:6)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:3)(cid:6)(cid:12)(cid:3)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:30)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:2)(cid:15)(cid:4)(cid:3)(cid:8) (cid:18) (cid:12)(cid:3)(cid:22)(cid:18) (cid:20)-(cid:8) (cid:13)(cid:18) (cid:16)(cid:8)
(cid:15)(cid:6)(cid:20)(cid:30)(cid:8) (cid:17)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)(cid:12)(cid:3)(cid:6)(cid:12)(cid:3)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:20)(cid:11)(cid:16)(cid:16)(cid:3)(cid:4)(cid:8)(cid:7)(cid:11)(cid:6)(cid:8) (cid:17)(cid:6)(cid:8)(cid:19) (cid:4)(cid:17)(cid:6)(cid:21)
(cid:7)(cid:17)(cid:19) (cid:20)(cid:3)(cid:8) (cid:13)(cid:3)(cid:8) (cid:18) (cid:6)% (cid:3)(cid:4)(cid:12)(cid:16)(cid:15)(cid:15)% (cid:8) (cid:11)(cid:12)(cid:8) (cid:2)(cid:11)(cid:16)(cid:10)(cid:3)(cid:2)(cid:11)(cid:16)(cid:17)(cid:7)(cid:11)(cid:20)(cid:8) (cid:7)(cid:15)(cid:6)(cid:21)
(cid:12)(cid:3)( (cid:18) (cid:3)(cid:6)(cid:7)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:22)(cid:15)(cid:4)(cid:2)(cid:3)(cid:4)(cid:27)(cid:8) ) (cid:10)(cid:3)(cid:8) (cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:17)(cid:12)(cid:16)(cid:8)
(cid:19) (cid:4)(cid:15)(cid:5)(cid:4)(cid:11)(cid:2)(cid:8) (cid:15)(cid:22)(cid:8) (cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:12)(cid:8) (cid:17)(cid:12)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:3)(cid:11)(cid:4)(cid:7)(cid:10)(cid:8) (cid:22)(cid:15)(cid:4)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)
(cid:7)(cid:15)(cid:2)(cid:2)(cid:15)(cid:6)(cid:8)(cid:12)(cid:15)(cid:18) (cid:4)(cid:7)(cid:3)(cid:8)(cid:15)(cid:22)(cid:8)(cid:11)(cid:20)(cid:20)(cid:8)(cid:3)*(cid:19) (cid:20)(cid:11)(cid:6)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)(cid:27)(cid:8)(cid:28)
(cid:8)(cid:8)
@ (cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:17)(cid:12)(cid:2)(cid:8) (cid:28)
(cid:8) (cid:19) (cid:4)(cid:15)& (cid:17)% (cid:3)(cid:12)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:6)(cid:3)(cid:7)(cid:3)(cid:12)(cid:12)(cid:11)(cid:4)(cid:30)(cid:8)
(cid:17)(cid:6)(cid:12)(cid:17)(cid:5)(cid:10)(cid:16)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:4)(cid:3)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:6)(cid:15)(cid:8) (cid:11)(cid:18) (cid:16)(cid:15)(cid:6)(cid:15)(cid:2)(cid:15)(cid:18) (cid:12)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8)
(cid:15)(cid:22)(cid:8)’ (cid:3)(cid:11)(cid:16)(cid:10)(cid:3)(cid:4)(cid:8)(cid:16)(cid:10)(cid:11)(cid:16)(cid:8)(cid:11)(cid:4)(cid:3)(cid:8) (cid:20)(cid:15)(cid:5)(cid:17)(cid:7)(cid:11)(cid:20)(cid:20)(cid:30)(cid:8) (cid:17)(cid:6)% (cid:3)(cid:19) (cid:3)(cid:6)% (cid:3)(cid:6)(cid:16)(cid:8)(cid:15)(cid:22)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:8)(cid:19) (cid:4)(cid:17)(cid:6)(cid:7)(cid:17)(cid:19) (cid:20)(cid:3)(cid:12)(cid:8)(cid:15)(cid:22)(cid:8)(cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:12)(cid:27)(cid:8)(cid:28)
(cid:8)(cid:29) (cid:3)(cid:8)% (cid:15)(cid:6)=(cid:16)(cid:8), (cid:6)(cid:15)’ (cid:8)
(cid:16)(cid:10)(cid:3)(cid:8) (cid:22)(cid:17)(cid:6)(cid:11)(cid:20)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:6)(cid:11)(cid:16)(cid:18) (cid:4)(cid:3)-(cid:8) (cid:13)(cid:18) (cid:16)(cid:8) ’ (cid:3)(cid:8) , (cid:6)(cid:15)’ (cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:30)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:6)(cid:15)(cid:16)(cid:8) (cid:3)*(cid:19) (cid:4)(cid:3)(cid:12)(cid:12)(cid:3)% (cid:8) (cid:17)(cid:6)(cid:8) (cid:16)(cid:3)(cid:4)(cid:2)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:7)(cid:15)(cid:20)% (cid:8)
(cid:22)(cid:4)(cid:15)(cid:6)(cid:16)(cid:12)(cid:8)(cid:15)(cid:4)(cid:8)(cid:16)(cid:10)(cid:18) (cid:6)% (cid:3)(cid:4)(cid:12)(cid:16)(cid:15)(cid:4)(cid:2)(cid:12)(cid:27)(cid:8)(cid:28)
(cid:8)(cid:8)
(cid:1)& (cid:3)(cid:4)(cid:30)(cid:8)(cid:22)(cid:17)(cid:3)(cid:20)% (cid:8)(cid:15)(cid:22)(cid:8)(cid:12)(cid:7)(cid:17)(cid:3)(cid:6)(cid:7)(cid:3)(cid:8)(cid:15)(cid:19) (cid:3)(cid:4)(cid:11)(cid:16)(cid:3)(cid:12)(cid:8)(cid:13)(cid:30)(cid:8)(cid:22)(cid:15)(cid:4)(cid:2)(cid:18) (cid:20)(cid:11)(cid:16)(cid:21)
(cid:17)(cid:6)(cid:5)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:16)(cid:3)(cid:12)(cid:16)(cid:17)(cid:6)(cid:5)(cid:8) (cid:5)(cid:3)(cid:6)(cid:3)(cid:4)(cid:11)(cid:20)(cid:17)(cid:31) (cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8)
(cid:12)(cid:15)(cid:2)(cid:3)(cid:16)(cid:17)(cid:2)(cid:3)(cid:12)(cid:8) % (cid:17)(cid:5)(cid:6)(cid:17)(cid:22)(cid:17)(cid:3)% (cid:8) (cid:13)(cid:30)(cid:8) (cid:13)(cid:3)(cid:17)(cid:6)(cid:5)(cid:8) (cid:7)(cid:11)(cid:20)(cid:20)(cid:3)% (cid:8) (cid:19) (cid:4)(cid:17)(cid:6)(cid:7)(cid:17)(cid:21)
(cid:19) (cid:20)(cid:3)(cid:12)(cid:8)(cid:15)(cid:4)(cid:8)(cid:20)(cid:11)’ (cid:12)(cid:27)(cid:8)) (cid:10)(cid:3)(cid:8)(cid:20)(cid:17)(cid:13)(cid:4)(cid:11)(cid:4)(cid:30)(cid:8)(cid:15)(cid:22)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)A (cid:6)(cid:17)& (cid:3)(cid:4)(cid:12)(cid:17)(cid:16)(cid:30)(cid:8)(cid:15)(cid:22)(cid:8)
) (cid:3)*(cid:11)(cid:12)(cid:8) (cid:10)(cid:11)(cid:12)(cid:8) (cid:16)(cid:10)(cid:17)(cid:4)(cid:16)(cid:30)(cid:21)(cid:22)(cid:17)& (cid:3)(cid:8) (cid:13)(cid:15)(cid:15), (cid:12)(cid:8) ’ (cid:17)(cid:16)(cid:10)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:16)(cid:17)(cid:16)(cid:20)(cid:3)(cid:8)
> 9 (cid:4)(cid:17)(cid:6)(cid:7)(cid:17)(cid:19) (cid:20)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) 5 (cid:10)(cid:3)(cid:2)(cid:17)(cid:12)(cid:16)(cid:4)(cid:30)> (cid:8) (cid:11)(cid:6)% (cid:8) (cid:3)(cid:17)(cid:5)(cid:10)(cid:16)(cid:3)(cid:3)(cid:6)(cid:8)
(cid:13)(cid:15)(cid:15), (cid:12)(cid:8) ’ (cid:17)(cid:16)(cid:10)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:16)(cid:17)(cid:16)(cid:20)(cid:3)(cid:8) > 9 (cid:4)(cid:17)(cid:6)(cid:7)(cid:17)(cid:19) (cid:20)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) 9 (cid:12)(cid:30)(cid:7)(cid:10)(cid:15)(cid:20)(cid:21)
(cid:15)(cid:5)(cid:30)> (cid:27)(cid:8)1 (cid:18) (cid:16)(cid:8)(cid:16)(cid:10)(cid:3)(cid:4)(cid:3)(cid:8)(cid:11)(cid:4)(cid:3)(cid:8)(cid:6)(cid:15)(cid:8)(cid:19) (cid:4)(cid:17)(cid:6)(cid:7)(cid:17)(cid:19) (cid:20)(cid:3)(cid:12)(cid:8)(cid:15)(cid:22)(cid:8)(cid:7)(cid:10)(cid:3)(cid:2)(cid:17)(cid:12)(cid:21)
(cid:16)(cid:4)(cid:30)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:12)(cid:17)(cid:2)(cid:19) (cid:20)(cid:30)(cid:8) (cid:12)(cid:16)(cid:11)(cid:6)% (cid:8) (cid:15)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:17)(cid:4)(cid:8) (cid:15)’ (cid:6)-(cid:8) ’ (cid:17)(cid:16)(cid:10)(cid:15)(cid:18) (cid:16)(cid:8)
(cid:6)(cid:3)(cid:3)% (cid:17)(cid:6)(cid:5)(cid:8) (cid:16)(cid:15)(cid:8) (cid:13)(cid:3)(cid:8) (cid:3)*(cid:19) (cid:20)(cid:11)(cid:17)(cid:6)(cid:3)% (cid:8) (cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)& (cid:3)(cid:20)(cid:30)(cid:8) (cid:22)(cid:4)(cid:15)(cid:2)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:8) (cid:19) (cid:4)(cid:15)(cid:19) (cid:3)(cid:4)(cid:16)(cid:17)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:3)(cid:20)(cid:3)(cid:7)(cid:16)(cid:4)(cid:15)(cid:6)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:11)(cid:16)(cid:15)(cid:2)(cid:17)(cid:7)(cid:8) (cid:6)(cid:18) (cid:21)
(cid:7)(cid:20)(cid:3)(cid:17)-(cid:8) (cid:11)(cid:6)% (cid:8) (cid:17)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:11)(cid:2)(cid:3)(cid:8) ’ (cid:11)(cid:30)(cid:8) (cid:16)(cid:10)(cid:3)(cid:4)(cid:3)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:6)(cid:15)(cid:8)
(cid:19) (cid:4)(cid:17)(cid:6)(cid:7)(cid:17)(cid:19) (cid:20)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:19) (cid:12)(cid:30)(cid:7)(cid:10)(cid:15)(cid:20)(cid:15)(cid:5)(cid:30)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:22)(cid:4)(cid:3)(cid:3)(cid:21)
(cid:12)(cid:16)(cid:11)(cid:6)% (cid:17)(cid:6)(cid:5)-(cid:8) (cid:17)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:3)(cid:6)(cid:12)(cid:3)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:30)(cid:8) % (cid:15)(cid:8) (cid:6)(cid:15)(cid:16)(cid:8)
(cid:6)(cid:3)(cid:3)% (cid:8) (cid:18) (cid:20)(cid:16)(cid:17)(cid:2)(cid:11)(cid:16)(cid:3)(cid:20)(cid:30)(cid:8) (cid:16)(cid:15)(cid:8) (cid:13)(cid:3)(cid:8) (cid:18) (cid:6)% (cid:3)(cid:4)(cid:12)(cid:16)(cid:15)(cid:15)% (cid:8) (cid:16)(cid:10)(cid:4)(cid:15)(cid:18) (cid:5)(cid:10)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:8)(cid:12)(cid:16)(cid:18) % (cid:30)(cid:8)(cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)(cid:10)(cid:18) (cid:2)(cid:11)(cid:6)(cid:8)(cid:13)(cid:4)(cid:11)(cid:17)(cid:6)-(cid:8) ’ (cid:10)(cid:17)(cid:7)(cid:10)(cid:8) (cid:17)(cid:6)(cid:8) (cid:16)(cid:18) (cid:4)(cid:6)(cid:8)
(cid:2)(cid:18) (cid:12)(cid:16)(cid:8)(cid:18) (cid:20)(cid:16)(cid:17)(cid:2)(cid:11)(cid:16)(cid:3)(cid:20)(cid:30)(cid:8)(cid:13)(cid:3)(cid:8)(cid:18) (cid:6)% (cid:3)(cid:4)(cid:12)(cid:16)(cid:15)(cid:15)% (cid:8)(cid:15)(cid:6)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:13)(cid:11)(cid:12)(cid:17)(cid:12)(cid:8)
(cid:15)(cid:22)(cid:8)(cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:12)(cid:8)(cid:11)(cid:6)% (cid:8)(cid:7)(cid:10)(cid:3)(cid:2)(cid:17)(cid:12)(cid:16)(cid:4)(cid:30)(cid:27)(cid:8)(cid:8)
Thus the battle is joined: are the higher
level sciences derived from physics?
Before approaching this question, it is
reasonable to ask whether any common
ground exists between reductionism and
functionalism? To do that, we first ex-
amine the concepts of epiphenomena
and supervenience. Then we explore the
status of the fundamental forces of phys-
ics and the possibility of the emergence
of higher level forces. Finally we look at
the role the environment and historical
accidents play in both reductionism and
functionalism.
2.3 Epiphenomena
If one doesn ’ t already have a sense of
what it means, the term epiphenomenon
is quite difficult to understand. Here is
the WordNet
definition [WordNet],
which is representative.
6(cid:8) (cid:12)(cid:3)(cid:7)(cid:15)(cid:6)% (cid:11)(cid:4)(cid:30)(cid:8) (cid:19) (cid:10)(cid:3)(cid:6)(cid:15)(cid:2)(cid:3)(cid:6)(cid:15)(cid:6)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:17)(cid:12)(cid:8) (cid:11)(cid:8) (cid:13)(cid:30)(cid:21)
(cid:19) (cid:4)(cid:15)% (cid:18) (cid:7)(cid:16)(cid:8)(cid:15)(cid:22)(cid:8)(cid:11)(cid:6)(cid:15)(cid:16)(cid:10)(cid:3)(cid:4)(cid:8)(cid:19) (cid:10)(cid:3)(cid:6)(cid:15)(cid:2)(cid:3)(cid:6)(cid:15)(cid:6)(cid:27)(cid:8)
It is not clear that this definition pins
much down. This definition is especially
troublesome because the terms secon-
dary and by-product should not be inter-
preted to mean that an epiphenomenon is
separate from and a consequence of the
state of affairs characterized by the
“ other” phenomenon.
We suggest that a better way to think of
an epiphenomenon is as an alternative
way of apprehending or perceiving a
given state of affairs. Consider Brownian
motion, which appears to be motion that
very small particles of non-organic ma-
terials are able to engage in on their
own. Before Einstein, Brownian motion
was a mystery. How could inanimate
matter move on its own? We now know
that Brownian motion is an epiphe-
nomenon of collisions of particles with
atoms or molecules.
With this usage of epiphenomenon as a
guide we define an epiphenomenon as a
phenomenon
that can be described
(sometimes formally but sometimes only
informally) in terms that do not depend
on
the underlying phenomena from
which it emerges.4
4 We use the term emerges advisedly. We will soon
define emergent and epiphenomenal to be syno-
nyms.
Emergence Explained
7/67
DRAFT
2/12/2006
This is familiar territory for Engineering
and Computer Science. Requirements
and specifications are by intention writ-
ten in terms that do not depend on the
design or implementation of the systems
that realize them. Requirements are writ-
ten before systems are designed, and
specifications are intended specifically
to be implementation-independent.
Although use of the term epiphenomena
in this context sounds strange, a re-
quirements document or a system speci-
fication is intended to describe epiphe-
nomena of systems that satisfy those re-
quirements or that meet that specifica-
tion.5
Even though we just claimed that a
specification describes epiphenomena,
our notion of epiphenomena is not the
same as the notion of functionality.
From our perspective epiphenomena ex-
ist only when there is an implementa-
tion. Epiphenomena are, first of all, phe-
nomena. A specification for a system
that did not exist would not describe
epiphenomena. It would describe the
epiphenomena of any system that satis-
fies that specification, but that ’ s as far as
one could go. From our perspective,
there are no epiphenomena unless they
are epiphenomena of something. That is
not the case with functionality, which is
understood in the abstract.
It is reasonable to say that the function-
ality of most executing software is
epiphenomenal. The computation con-
sidered as an abstraction may be defined
independently of the implementation.
But the only real action is at the very
lowest level. No matter how abstract
one’ s software, one can always stop a
5 We are not holding our breath waiting for a con-
tracting officer to refer to a system’ s inadequate
epiphenomena.
computation by pulling the plug on the
computer. The computation as an ab-
stract epiphenomenon exists only be-
cause electrons are actually flowing.
2.4 Supervenience
A closely related term from the philoso-
phical literature is supervenience. The
intended use of this term is to relate a
presumably higher level set of predicates
or properties6 (call this set H for higher)
to a presumably lower level set of predi-
cates or properties (call this set L for
lower). The properties or predicates in H
and L are all presumed to be applicable
to some common domain of discourse.
H and L are each ways of characterizing
the state of affairs of the underlying do-
main. For any particular state of affairs
in the domain of discourse, the predi-
cates in H and L will each be either true
or false (or perhaps not applicable).
One says that H supervenes on (or over)
L if it is never7 the case that two states of
affairs will assign the same configura-
tion of values to the elements of L but
different configuration of values to the
elements of H. In other words when a
state of affairs assigns values to predi-
cates in L, that fixes the assignments of
values to predicates of H.
Consider the following simple example.
Let the domain be a sequence of n bits.
Let L be the statements: bit 1 is on; bit 2
is on; etc. Let H be statements of the
6 Since properties (e.g., the color of an object) may
be expressed in terms of predicates about that
property, from here on we will speak only of
predicates.
7 Some definitions require that not only is it never
the case, it never can be the case. It does make a
formal difference whether we base supervenience
on a logical impossibility or on empirical facts.
We finesse that distinction by adopting the rule of
thumb of fundamental particle physicists: if some-
thing can happen it will.
Emergence Explained
8/67
DRAFT
2/12/2006
sort: exactly 5 bits are on; an even num-
ber of bits are on; no two successive bits
are on; the bits that are on form the ini-
tial values in the Fibonacci sequence;
etc.
H supervenes on L since any configura-
tion of values of the statements in L de-
termines the values of the statements in
H.
However, if we remove one of the
statements from L, e.g., we don’ t include
in L a statement about bit 3, but we leave
the statements in H alone, then H does
not supervene on L.
To see why, consider the H statement
h1: an even number of bits is on.
For concreteness, let ’ s assume that there
are exactly 5 bits. Let ’ s assume first, as
in the first line of Figure 1, that all the
bits except bit 3, the one for which there
is no L statement, are on. Thus since
there is no L statement about bit 3, all
the L statements are true even though bit
3 is off. Since 4 of the 5 bits are on, h1
is also true.
Now, assume that bit 3 is on as in the
second line of Figure 1. All the L state-
ments are still true. But since 5 bits are
now on, h1 is now false.
Since we have found an H statement that
has two different values for a single con-
figuration of values of the L statements,
H does not supervene over L.
Although we have not attempted to trace
its history, the notion of supervenience
may have originated as an attempt to
capture the relationship between epiphe-
nomena and their underlying phenom-
ena. Presumably epiphenomena super-
vene on underlying phenomena: distinct
epiphenomena must be associated with
distinct underlying phenomena, which is
what one wants.8 You can’ t get two dif-
ferent sets of epiphenomena from the
same underlying phenomena.9
Note that the reverse is not true. Two
different states of the underlying phe-
nomena may result in the same epiphe-
nomena. In our bit example, there are
many different ways in which an even
number of bits may be on.
The position known as supervenience
physicalism might be understood as
claiming that any higher-level descrip-
tion of nature supervenes over some set
of primitive descriptions.
It would appear that the relationship de-
fined by supervenience will be useful in
analyzing multi-scale phenomena. To
some extent this is the case. When we
consider autonomous entities, however,
we shall see that supervenience is not as
useful as one might have hoped. This is
one reason that emergence has been so
difficult to pin down.
One reason that supervenience is less
useful than one might have hoped may
be related to the difficulty one encoun-
ters when using supervenience for infi-
nite domains. Consider our bit example
again, but imagine that we have a count-
ably infinite number of bits. Consider
the H statement
8 On the other hand, searches for the phrase
“ epiphenomena supervene” on Google, Yahoo
msn.com, and AskJeeves conducted on 9/17/2005
found no references—
other than to an online draft
of this paper. This suggests that none of these ser-
vices had ever scanned a document that contained
the sentence fragment “
epiphenomena super-
vene over …
. ”
It ’ s not really the epiphenomena that supervene
over their underlying phenomena. It ’ s statements
about the epiphenomena or properties of the
epiphenomena that supervene over statements
about the underlying phenomena or their proper-
ties. As a short-hand, however, we will talk about
epiphenomena as supervening (or not) over under-
lying phenomena.
9
Emergence Explained
9/67
…
DRAFT
2/12/2006
h2: the bits that are on are prime.
Clearly the H set consisting solely of h2
supervenes over the entire set of L
statements. Just as clearly, that H set
does not supervene over any proper sub-
set of the L statements, and certainly not
over any finite subset of the L state-
ments —
one needs to look at all of the
bits to determine whether it is exactly
the prime bits that are on.
So even though we can conclude that H,
which contains a single relatively simple
statement, supervenes over the infinite
set of statements in L, that information
doesn’ t buy us much. On the contrary,
supervenience of this sort is like describ-
ing a tapestry by enumerating the threads
that make it up. The epiphenomena are
lost. We will see another example of this
later.
and
2.5 Fundamental
forces
strong emergence
Returning to Weinberg and Fodor, pre-
sumably both would agree that phenom-
ena of the special sciences supervene on
phenomena in physics. A given set of
phenomena at the level of fundamental
physics is associated with no more than
one set of phenomena at the level of any
of the special sciences. Or looking top-
down, two different states of affairs in
some special science must be associated
with two different states of affairs at the
level of fundamental physics. Superven-
ience of this sort seems to correspond
more or less to Weinberg’ s petty reduc-
tionism, a doctrine that he finds of only
minor importance. So perhaps agreement
at this level is not very significant.
Where Weinberg and Fodor disagree is
not about supervenience but about
whether the principles of the special sci-
ences can be derived from the principles
of physics.
That disagreement aside, we do have a
fundamental area of agreement — which
has some quite significant implications.
Weinberg makes his case sarcastically.
B (cid:3)(cid:6)(cid:4)(cid:30)(cid:8) 1 (cid:3)(cid:4)(cid:5)(cid:12)(cid:15)(cid:6)(cid:8) (cid:11)(cid:6)% (cid:8) : (cid:11)(cid:4)(cid:16)(cid:10)(cid:8) C (cid:11)% (cid:3)(cid:4)(cid:8) (cid:6)(cid:15)(cid:16)’ (cid:17)(cid:16)(cid:10)(cid:21)
(cid:12)(cid:16)(cid:11)(cid:6)% (cid:17)(cid:6)(cid:5)-(cid:8) (cid:16)(cid:10)(cid:3)(cid:4)(cid:3)(cid:8) (cid:17)(cid:12)(cid:8) (cid:6)(cid:15)(cid:8) (cid:20)(cid:17)(cid:22)(cid:3)(cid:8) (cid:22)(cid:15)(cid:4)(cid:7)(cid:3)(cid:27)(cid:8) ) (cid:10)(cid:17)(cid:12)(cid:8) (cid:17)(cid:12)(cid:8) (cid:9)(cid:16)(cid:10)(cid:3)(cid:14)(cid:8)
(cid:17)(cid:6)& (cid:11)(cid:20)(cid:18) (cid:11)(cid:13)(cid:20)(cid:3)(cid:8)(cid:6)(cid:3)(cid:5)(cid:11)(cid:16)(cid:17)& (cid:3)(cid:8) (cid:19) (cid:3)(cid:4)(cid:12)(cid:19) (cid:3)(cid:7)(cid:16)(cid:17)& (cid:3)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:17)(cid:12)(cid:8) (cid:19) (cid:4)(cid:15)(cid:21)
& (cid:17)% (cid:3)% (cid:8)(cid:13)(cid:30)(cid:8)(cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:17)(cid:12)(cid:2)(cid:27)(cid:8)
What I believe Weinberg is getting at is
that the current standard model of phys-
ics postulates four elementary forces: the
strong force, the weak force, the elec-
tromagnetic force, and gravity. Since
Weinberg’ s Nobel prize was for his
work on unifying the weak and electro-
magnetic force, perhaps one should say
there are only three fundamental forces.
Either way, what ’ s important is that ac-
cording to physics there is a small fixed
number of fundamental forces. I doubt
that Fodor would disagree.
Weinberg ’ s sarcastic reference to a life
force is an implicit criticism of an obso-
lete strain of thinking about emergence.
The notion of vitalism—
the emergence
of
chemicals —
lifeless
from
life
postulates a new force of nature that ap-
pears at the level of biology and that is
not reducible to lower level phenomena.
Emergence of this sort is what Bedau
[Bedau] has labeled “ strong emergence. ”
As Bedau also points out, no one takes
this kind of emergence seriously.
It is worth noting, however, that even
were evidence of strong emergence to be
found, science would not shrivel up and
die. Dark energy, the apparently extra
force that seems to be pushing the Uni-
verse to expand may be a new force of
nature.
Furthermore, even if other (spooky)
forces of nature like vitalism were (mys-
teriously) to appear at various levels of
complexity, science would carry on. We
Emergence Explained
10/67
DRAFT
2/12/2006
would do our best to understand such
forces by measuring and characterizing
them in any way that we could.
What one doesn’ t want is to have new
forces of nature popping up indiscrimi-
nately. Strong emergence would not be
so terrible as long as it didn’ t happen too
often—
after all, the existing primitive
forces just seemed to pop up out of no-
where and we have taken them in
stride— or if it did happen often, if we
could find a theory that let us predict
when they would appear.
What would be especially upsetting
would be a new force of nature that vio-
lated the conservation laws. Suppose one
could put certain components together
and produce energy that didn’ t depend
on the energy or mass of the compo-
nents —
the mass and energy of the com-
ponents remained intact, but the combi-
nation of those components created new
energy. That would be scientifically
quite disturbing—
although it would cer-
tainly help with the energy problem.
Since the appearance of new forces of
nature that do not upset the conservation
laws could be taken simply as an accept-
able extension of physics (but one that
we have not seen and do not know how
to produce), and since the appearance of
new forces that violate the conservation
laws is currently unimaginable, from
now on we ignore the possibility of
strong emergence.
If one dismisses the possibility of strong
emergence and agrees that the only
forces of nature are the fundamental
forces as determined by physics, then
Fodor must also agree (no doubt he
would) that any force-like construct pos-
tulated by any of the special sciences
must be strictly reducible to the funda-
mental forces of physics. As Weinberg
says, there is no life force.
Note that this is a truly stark choice:
strict reductionism with respect to forces
or strong emergence. There is no third
way.
This leads to an important conclusion.
Any cause-like effect that results from a
force-like phenomenon in the domain of
any of the special (i.e., higher level) sci-
ences must be epiphenomenal.10 Since
epiphenomenal forces supervene on fun-
damental forces, distinct epiphenomenal
interactions must be manifestations of
distinct fundamental force actions. In
other words, anything that happens at the
most fundamental level (if there is one)
has no more than one manifestation at
any higher level.
It is important to note that this perspec-
tive establishes one of the basic claims
of reductionism: forces at all levels must
be explicable in terms of—
i.e., they are
epiphenomenal of and reducible to —
the
fundamental forces of physics.11 There
are no magical mystery forces; there is
no strong emergence.
The fact that all special (i.e., higher
level) science interactions are epiphe-
nomenal is not a problem for functional-
ism. Nor is it news to Fodor, who speaks
freely of “ the to-ings and fro-ings of bits
and pieces at the extreme microlevel.”
Rather, the functionalist claim is that the
regularities (be they epiphenomenal or
10 Kim [Kim ‘ 93] (denigratingly) used the term
epiphenomenal causation to refer to interactions of
this sort. We also consider such interactions to be
epiphenomenal, but we don ’ t find them less wor-
thy as a result.
11 Compare this with the conclusion Hume reached
[Hume] in his considerations of causality—
that
when one looks carefully at any allegedly direct
causal connection, one will find intermediary
links. Since Hume did not presume what we now
consider to be a bottom level of fundamental
physical forces, he dismissed the notion of causal-
ity entirely.
Emergence Explained
11/67
DRAFT
2/12/2006
not) that appear at the level of any spe-
cial science are of significance on their
own. The fact that the interactions to
which one refers in describing those
regularities are in fact implemented by
lower level interactions does not dimin-
ish either the importance or the lawful-
ness of the higher level regularities —
even though as we have seen Fodor has
no idea how those regularities come
about.
2.6 Historical accidents and the
environment
There is a second area of at least implicit
agreement between Weinberg and Fo-
dor. Consider
from
following
the
Weinberg.
(cid:9)6(cid:14)(cid:19) (cid:11)(cid:4)(cid:16)(cid:8) (cid:22)(cid:4)(cid:15)(cid:2)(cid:8) (cid:10)(cid:17)(cid:12)(cid:16)(cid:15)(cid:4)(cid:17)(cid:7)(cid:11)(cid:20)(cid:8) (cid:11)(cid:7)(cid:7)(cid:17)% (cid:3)(cid:6)(cid:16)(cid:12)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:13)(cid:30)(cid:8)
% (cid:3)(cid:22)(cid:17)(cid:6)(cid:17)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8)(cid:7)(cid:11)(cid:6)(cid:6)(cid:15)(cid:16)(cid:8)(cid:13)(cid:3)(cid:8)(cid:3)*(cid:19) (cid:20)(cid:11)(cid:17)(cid:6)(cid:3)% -(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:9)(cid:10)(cid:18) (cid:2)(cid:11)(cid:6)(cid:14)(cid:8)
(cid:6)(cid:3)(cid:4)& (cid:15)(cid:18) (cid:12)(cid:8)(cid:12)(cid:30)(cid:12)(cid:16)(cid:3)(cid:2)(cid:8)(cid:9)(cid:10)(cid:11)(cid:12)(cid:14)(cid:8)(cid:3)& (cid:15)(cid:20)& (cid:3)% (cid:8)(cid:16)(cid:15)(cid:8)’ (cid:10)(cid:11)(cid:16)(cid:8)(cid:9)(cid:17)(cid:16)(cid:8) (cid:17)(cid:12)(cid:14)(cid:8)
(cid:3)(cid:6)(cid:16)(cid:17)(cid:4)(cid:3)(cid:20)(cid:30)(cid:8) (cid:13)(cid:3)(cid:7)(cid:11)(cid:18) (cid:12)(cid:3)(cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:19) (cid:4)(cid:17)(cid:6)(cid:7)(cid:17)(cid:19) (cid:20)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:2)(cid:11)(cid:7)(cid:21)
(cid:4)(cid:15)(cid:12)(cid:7)(cid:15)(cid:19) (cid:17)(cid:7)(cid:8) (cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:7)(cid:10)(cid:3)(cid:2)(cid:17)(cid:12)(cid:16)(cid:4)(cid:30)-(cid:8) ’ (cid:10)(cid:17)(cid:7)(cid:10)(cid:8) (cid:17)(cid:6)(cid:8)
(cid:16)(cid:18) (cid:4)(cid:6)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) ’ (cid:10)(cid:11)(cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:30)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:3)(cid:6)(cid:16)(cid:17)(cid:4)(cid:3)(cid:20)(cid:30)(cid:8) (cid:13)(cid:3)(cid:7)(cid:11)(cid:18) (cid:12)(cid:3)(cid:8) (cid:15)(cid:22)(cid:8)
(cid:19) (cid:4)(cid:17)(cid:6)(cid:7)(cid:17)(cid:19) (cid:20)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:12)(cid:16)(cid:11)(cid:6)% (cid:11)(cid:4)% (cid:8)(cid:2)(cid:15)% (cid:3)(cid:20)(cid:8) (cid:15)(cid:22)(cid:8) (cid:3)(cid:20)(cid:3)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:4)(cid:30)(cid:8)
(cid:19) (cid:11)(cid:4)(cid:16)(cid:17)(cid:7)(cid:20)(cid:3)(cid:12)(cid:27)(cid:8)
Note Weinberg’ s reference to historical
accidents — which we also saw earlier, in
both his definition of grand reductionism
and his discussion of
the weather.
Weinberg gives historical accidents as
important a role in shaping the world as
he gives to the principles of physics. But
he gives them a significantly lesser bill-
ing—
one might say below the title in-
stead of above.
The importance of historical accidents is
especially clear when thinking about
evolution. We suggest that contrary to
Weinberg’ s claim, the human nervous
system (and human anatomy in general)
evolved to what they are not primarily
because of the principles of macroscopic
physics and chemistry but primarily be-
cause of the environment in which that
evolution took place—
and in which the
nervous system and anatomy must func-
tion.
Because of the crucial role the environ-
ment plays in evolution, one cannot start
with the principles of physics and chem-
istry and, considering them in isolation,
derive human anatomy. Certainly human
anatomy must be consistent with the
principles of physics and chemistry. But
human anatomy cannot be derived ex-
clusively from the principles of physics
and chemistry. One must also consider
the environment in which that anatomy
is intended to function.
We don’ t think we are putting words
into Weinberg’ s mouth when we take his
references to “ historical accidents ” to
mean the environment and the context in
which something occurs. If one grants us
this, then as Weinberg acknowledges,
the environment is often as significant a
consideration in how things turn out as
are the principles of physics.
The same perspective is nearly (but not
quite) explicit in the extracts from Fo-
dor. Recall Fodor’ s list of artifacts, in-
cluding a mouse trap, a can opener, etc.
All those artifacts are defined function-
ally, i.e., in terms of the functions they
perform — which
necessarily means
functions performed in the environment
in which they exist.
A can opener in an environment in
which cans are of an entirely different
size or shape from our cans would not be
a can opener. The same is true for a
mousetrap in an environment in which
mice were either the size of fleas or the
size of elephants.
And of course, the very name of the
school of thought, functionalism, under-
lines its concern with how things func-
tion. Functionality is by definition a rela-
Emergence Explained
12/67
DRAFT
2/12/2006
tionship between something and its envi-
ronment —
even if we limit the notion of
an environment to the “ inputs ” and “ out-
puts ” of the element under consideration.
It is also important to note that when one
speaks of an environment, one must un-
derstand the environment at the level at
which the object under consideration in-
teracts with it —
even though as we saw
earlier those interactions are epiphe-
nomenal. A can opener opens cans, and
a mouse trap traps mice. Neither deals
with swarms of quarks and other funda-
mental particles. As Fodor says, “ live
mouse in, dead mouse out. ”
Although neither side of this debate fo-
cuses on this issue, they are both appar-
ently in agreement that the environment
within which something exists is impor-
tant.
Interaction with an environment will
turn out to be quite significant. We ex-
plore it further when we discuss autono-
mous entities and stigmergy.
3 Emergence in the Game of
Life
In this section we use the Game of Life12
[Gardner] to illustrate emergence: the
implementation of a new level of ab-
straction on top of an existing substrate.
• The Game of Life is a totalistic13
two-dimensional cellular automaton.
12 The Game of Life is a popular example in discus-
sions of emergence. Bedau uses it as the primary
example in “ Downward causation and the auton-
omy of weak emergence” [Bedau]. We return to
Bedau later. Dennett refers to it in “ Real Patterns”
[Dennett] when discussing how his intentional
stance perspective compares to the perspectives of
other philosophers with respect to the reality of be-
liefs. (His position is that beliefs should be consid-
ered to be “ mildly real, ” which is intermediate
within the spectrum of positions he examines.)
13 Totalistic means that the action taken by an agent
depends on the number of its neighbors in certain
The Game of Life grid is assumed to
be unbounded in each direction, like
the tape of a Turing Machine.
• An agent occupies each grid cell.
Agents are fixed and cannot move
around on the grid. This is not un-
usual in agent-based modeling.
• Each agent is in one of two states:
“ alive” or “ dead” or more simply on
or off.
• The 8 surrounding agents are an
agent ’ s neighbors.
• At each time step an agent deter-
mines whether it will be alive or
dead at the next time step according
to the following rules.
• A live agent with two or three
live neighbors stays alive; other-
wise it dies.
• A dead agent with exactly three
live neighbors is (miraculously)
(re)born and becomes alive.
• All agents update themselves simul-
taneously based on the values of
their neighbors at that time step.
It is useful to think of the Game of Life
in the following three ways.
1. Treat the Game of Life is a serious
agent-based model —
of something,
perhaps life and death phenomena.
For our purposes it doesn’ t matter
that the Game of Life isn ’ t a realistic
model —
of anything. Many agent-
based models are at the same time
quite simple and quite revealing.
2. Treat the Game of Life as a trivial
physical universe. Recall Shalizi:
states and not on which neighbors are in which
state.
Emergence Explained
13/67
DRAFT
2/12/2006
(cid:8) ’ (cid:10)(cid:3)(cid:4)(cid:3)(cid:8) ( (cid:18) (cid:11)(cid:6)(cid:16)(cid:18) (cid:2)(cid:8) (cid:22)(cid:17)(cid:3)(cid:20)% (cid:8)
. (cid:15)(cid:2)(cid:3)(cid:19) (cid:20)(cid:11)(cid:7)(cid:3)(cid:8) (cid:28)
(cid:16)(cid:10)(cid:3)(cid:15)(cid:4)(cid:30)(cid:8)(cid:2)(cid:3)(cid:3)(cid:16)(cid:12)(cid:8) (cid:5)(cid:3)(cid:6)(cid:3)(cid:4)(cid:11)(cid:20)(cid:8) (cid:4)(cid:3)(cid:20)(cid:11)(cid:16)(cid:17)& (cid:17)(cid:16)(cid:30)(cid:8) (cid:28)
(cid:8) ’ (cid:3)(cid:8)
(cid:2)(cid:11)(cid:30)(cid:8) (cid:16)(cid:11), (cid:3)(cid:8) / (cid:16)(cid:10)(cid:3)(cid:8) (cid:4)(cid:18) (cid:20)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:5)(cid:11)(cid:2)(cid:3)0 (cid:8) (cid:16)(cid:15)(cid:8)
(cid:13)(cid:3)(cid:8)(cid:5)(cid:17)& (cid:3)(cid:6)(cid:27) ”
The Game of Life rules will be those
“ rules of the game.” The rules that
determine how cells turn on and off
will be taken as the most primitive
operations of the physics of the
Game of Life universe.14 When
thinking in terms of this perspective,
rather than thinking of a grid cell as
occupied by (immobile) agents, we
think of the cells themselves as
something like primitive particles.
The reductionist agenda within such
a Game of Life universe would be to
reduce every higher level phenome-
non to the underlying Game of Life
rules.
3. Treat the Game of Life as a pro-
gramming platform.
Although these three perspectives will
yield three different approaches to the
phenomena generated, the phenomena
themselves will be identical. It will al-
ways be the same Game of Life rules
which determine what happens.
3.1 Epiphenomenal gliders
Figure 2 shows a sequence of 5 time
steps in a Game of Life run. The dark
cells (agents) are “ alive; ” the light cells
(agents) are “ dead.” One can apply the
rules manually and satisfy oneself that
they produce the sequence as shown.
Notice that the fifth configuration shows
the same pattern of live and dead cells as
the first except that the pattern is offset
14 This is the basis of what is sometimes called “ digi-
tal physics ” (see [Zuse], [Fredkin], and [Wolf-
ram]), which attempts to understand nature in
terms of cellular automata.
by one cell to the right and one cell
down.
If there are no other live cells on the
grid, this process could be repeated in-
definitely, producing a glider-like effect.
Such a glider is an epiphenomenon of
the Game of Life rules. If one thinks
about it —
and forgets that one already
knows that the Game of Life can pro-
duce gliders —
gliders are quite amazing.
A pattern that traverses the grid arises
from very simple (and local) rules for
turning cells on and off.
We should be clear that gliders are
epiphenomenal. The rules of the Game
of Life do nothing but turn individual
cells on and off. There is nothing in the
rules about waves of cells turning on and
off sweeping across the grid. Such
epiphenomenal gliders exemplify emer-
gence.
• Gliders are not generated explicitly:
there is no glider algorithm. There is
no “ code” that explicitly decides
which cells should be turned on and
off to produce a glider.
• Gliders are not visible in the rules.
None of the rules are formulated, ei-
ther explicitly or implicitly, in terms
of gliders.
When looked at from our agent-based
modeling perspective, gliders may repre-
sent epidemics or waves of births and
deaths. If one were attempting to dem-
onstrate that such waves could be gener-
ated by simple agent-agent interactions,
one might be quite pleased by this result.
It might merit a conference paper.
3.2 Gliders in our physics world
From our physics perspective, we note
that the rules are the only forces in our
Game of Life universe. Being epiphe-
nomenal, gliders are causally power-
Emergence Explained
14/67
DRAFT
2/12/2006
less.15 The existence of a glider does not
change either how the rules operate or
which cells will be switched on and off.
Gliders may be emergent, but they do
not represent a new force of nature in the
Game of Life universe. It may appear to
us as observers that a glider looks like it
is moving across the grid and that when
it reaches a certain cell it will turn that
cell on.
But that ’ s not true. It is only the rules
that turn cells on and off. A glider
doesn’ t “ go to an cell and turn it on.” A
Game of Life run will proceed in exactly
the same way whether one notices the
gliders or not. This is a very reductionist
position. Things happen only as a result
of the lowest level forces of nature,
which in this case are the rules.
3.3 The Game of Life as a pro-
gramming platform
Amazing as they are, gliders are also
trivial. Once one knows how to produce
a glider, it ’ s a simple matter to make as
many as one wants. If we look at the
Game of Life as a programming plat-
form —
imagine that we are kids fooling
around with a new toy— we might ex-
periment with it to see whether we can
make other sorts of patterns. If we find
some, which we will, we might want to
see what happens when patterns crash
into each other—
boys will be boys.
After some time and effort, we might
compile a library of Game of Life pat-
terns, including the API16 of each pattern,
15 All epiphenomena are causally powerless. Since
epiphenomena are simply another way of perceiv-
ing underlying phenomena, an epiphenomenon it-
self cannot have an effect on anything. It is the un-
derlying phenomena that act.
16 Application Programming Interface
which describes what happens when that
pattern collides with other patterns.17
Since its introduction three decades ago,
a community of such Game of Life pro-
grammers has developed. That commu-
nity has created such libraries —
at least
on an informal basis.18
It has even been shown [Rendell] that by
suitably arranging Game of Life pat-
terns, one can simulate a Turing Ma-
chine.
Moreover, and this is a crucial point, the
emulation of a Turing Machine with
Game of Life patterns is also an example
of emergence. There is no algorithm.
The Turing Machine appears as a conse-
quence of epiphenomenal interactions
among epiphenomenal patterns!
What did we just say? What does it
mean to say that epiphenomenal gliders
and other epiphenomenal patterns simu-
late a Turing Machine? How can it mean
anything? The patterns aren’ t real; the
Turing Machine isn’ t real; they are all
(nothing but) epiphenomena.
Furthermore, even the interactions be-
tween and among patterns aren’ t real ei-
ther. They’ re also epiphenomenal —
and
epiphenomenal in the sense described
above: the only real action is at the most
fundamental level, the Game of Life
rules. Pattern APIs notwithstanding the
only thing that happens on a Game of
Life grid is that the Game of Life rules
determine which cells are to be on and
17 Note, however, that interactions among patterns
are quite fragile. If two patterns meet in slightly
different ways, the results will generally be quite
different.
18 Many of these libraries are available on the web.
To explore this world, a good place to start and
one that seems to be kept up to date is Jason ’ s Life
Page [Summers]. The patterns available on that
page and on the pages to which that page links are
quite amazing.
Emergence Explained
15/67
DRAFT
2/12/2006
which cells are to be off. No matter how
real the patterns look to us, interaction
among them is always epiphenomenal.
So what are we talking about?
What does one do to show that a Game
of Life emulation of a Turing machine is
correct? What one must do is to adopt a
operational perspective and treat the pat-
terns and their interactions, i.e., the de-
sign itself, independently of the Game of
Life.
It is the design, i.e., the way in which the
patterns —
be they real or epiphenome-
nal —
interact that we want to claim sim-
ulates a Turing Machine. To show that
we must do two things.
1. Show that the abstract design con-
sisting of patterns and their interac-
tions actually does simulate a Turing
Machine. That is, we reify the de-
sign, i.e., treat it as real, and argue
about its properties.
2. Show that the design can be imple-
mented on a Game-of-Life platform.
Note what this perspective does. It un-
shackles the design from its moorings as
a Game of Life epiphenomenon, reifies
it as an independent abstraction, and lets
it float free. (The protestors in the streets
chanting “ Free the design” can now
lower their picket signs and go home.)
The design becomes an abstraction, an
abstract construct on its own. Once we
have such a design as an abstraction we
can then reason about its properties, i.e.,
(a) that it accomplishes what we want,
namely that it simulates a Turing Ma-
chine and (b) that it can be reattached to
its moorings and be implemented on a
Game of Life platform. In other words,
emergence is getting epiphenomena to
do real (functional) work.
3.4 Game of Life anthropologists
Let ’ s forget everything we just said
about the Game of Life, and let ’ s pretend
we are anthropologists. Let ’ s imagine
that a lost tribe of what turn out to be
Game of Life creatures has been discov-
ered in a remote wilderness. Of course,
when they are first discovered, we don’ t
know that they are Game of Life crea-
tures. All we know is that their strange
grid-like faces are made up of cells that
blink on and off.
We get a grant to study these creatures.
We travel to their far-off village, and we
learn their language. They can’ t seem to
explain what makes their cells blink on
and off; we have to figure that out for
ourselves.
After months of study, we come up with
the Game of Life rules as an explanation
for how the grid cells are controlled.
That seems to work. Every single mem-
ber of the tribe operates in a way that is
consistent with those rules. The rules
even explain the unusual patterns that we
observe—
them, glider-like,
some of
traversing the entire grid. Thrilled with
our analysis, we return home and publish
our results.
But one thing continues to nag. One of
the
teenage girls —
she calls herself
Hacka for reasons that we do not under-
stand— has a pattern of activities on her
grid that seems more complex than most
of the others. The Game of Life rules
fully explain every light that goes on and
every light that goes off on Hacka ’ s
pretty face. But somehow that explana-
tion doesn’ t seem to capture everything
that ’ s going on. It just seems more com-
plex than that. Did we miss something?
To make a long story short, it turns out
that the tribe was not as isolated as we
had thought. In fact they have an Internet
connection. Hacka had learned not only
Emergence Explained
16/67
DRAFT
2/12/2006
that she was a Game of Life system but
that the Game of Life can emulate a Tur-
ing Machine. She had decided to pro-
gram herself to do just that. Her parents
disapproved. But girls just want to have
fun—
especially teenage girls.
No wonder we felt uncertain about our
results. Even though the Game of Life
rules explained every light that went on
and off on Hacka’ s face, it said nothing
about the functionality implemented by
Hacka ’ s Turing Machine emulation.
The rules explained everything about
how the system worked; but they said
nothing about what the system did. The
rules didn’ t have a way even to begin to
talk about the functionality of the sys-
tem — which was logically independent
of the rules. The rules simply don’ t talk
about Turing Machines.
A Turing machine is an autonomous
functional abstraction
that we (and
Hacka) built on top of the rules of the
Game of Life. Our reductive explana-
tion, that a certain set of rules make the
cells go on and off, had no way to cap-
ture this sort of additional functionality.19
3.5 Keeping score
In the debate between reductionism and
functionalism, the fact that one can build
a Game of Life Turing Machine may be
scored as follows.
19 In “ Real Patterns” Dennett [Dennett ‘ 91] uses the
fact that a Turing Machine may be implemented in
terms of Game of Life patterns to argue that the
position he takes in The Intentional Stance [Den-
nett ‘ 87] falls midway along a spectrum of posi-
tions ranging from what he calls
“ Industrial
strength Realism” to eliminative materialism (that
beliefs are nothing but convenient fictions).
Our focus in this paper differs from Dennett ’ s in
that it is not on psychological states or mental
events but on
regularities—
the nature of
independently of whether those regularities are the
subject matter of anyone’ s beliefs.
• Reductionism scores a point — with
which it was credited earlier—
in that
the only forces operating are (and
must be) reducible to the fundamen-
tal operational rules of the Game of
Life.
In philosophical terms, this would be
considered an supervenience physi-
calism: any apparent higher level
forces or interactions are nothing
more than conceptually convenient
ways of packaging lower level forces
and interactions.
• Functionalism scores a point in that
the functionality of a Turing Ma-
chine is beyond the realm of and can
neither be reduced to nor derived
from the Game of Life rules.
Functionalism gets only partial credit
because it is quite clear (if im-
mensely complicated)—
but not at all
mysterious and not on a par with the
question of why anything exists at
all — how a Turing Machine may be
implemented in terms of Game of
Life rules. All that is required is that
a design for a Turing Machine emu-
lator be implemented by using the
Game of Life as a software devel-
opment platform.
3.6 Defining emergence
With this background, we can define
emergence as a relationship between a
phenomenon and a model.
By a model we will mean a collection of
elements with certain interrelationships,
e.g., the grid cells and rules of the Game
of Life.
To begin, given some model we will say
that a phenomenon is emergent over that
model if it is epiphenomenal with re-
spect to that model. In other words, all
epiphenomena are emergent, and all
Emergence Explained
17/67
DRAFT
2/12/2006
emergent phenomena are epiphenome-
nal. In short, we define epiphenomenal
and emergent to be synonyms.
Recall that we defined an epiphenome-
non to be a phenomenon that is appre-
hended, conceptualized, or perceived in-
dependently of the forces that bring it
about. Thus a phenomenon is emergent
if it is conceptualized independently of
the phenomena that implement it. As we
pointed out in our discussion of epiphe-
nomena earlier, this is familiar territory
for Engineering and Computer Science.
Requirements and specifications are by
intention written in terms that do not de-
pend on the design or implementation of
the systems that realize them.
More concretely we define a phenome-
non as emergent over a model if it satis-
fies two conditions.
1. The phenomenon as a phenomenon
may be understood on its own. Its
conceptualization does not depend
on the model from which it emer-
ges.20
2. There is an implementation of the
phenomenon in terms of elements of
the model.21
Our prototypical examples are gliders
and Turing Machines, which we will say
are emergent over the Game of Life.
20 We are not requiring that there be a person to do
the conceptualizing, only that such an independent
conceptualization be possible.
21 Our concern with implementations distinguishes
our position from both functionalism and Den-
nett ’ s intentional stance, neither of which cares
much about how higher level phenomena are im-
plemented.
It may be that emergent phenomena may have
multiple realizations— Turing Machines may be
epiphenomena of many computing substrates—
but
this will not be the most commonly occurring
situation in nature. Non-computational epiphe-
nomena will usually occur as the consequence of a
particular implementation.
Notice that we do not require an emer-
gent phenomenon to have a formal de-
scription or
specification. Certainly
some epiphenomena may be formalized
and understood as formalized abstrac-
tions —
the Turning Machine for exam-
ple. Epiphenomena about which such in-
dependent formal theories may be de-
veloped often tend to be particularly use-
ful cases of emergence from a scientific
or mathematical perspective.
On the other hand, we shall see later that
there are epiphenomena about which it is
either not feasible or not useful to define
abstract theories but which are important
to us nevertheless. As an immediate and
familiar (if informal) example, consider
the so-called edge-of-chaos behavior of
certain cellular automata. (See, for ex-
ample, [Langton].) Edge-of-chaos auto-
mata are those whose runs become nei-
ther trivially predictable nor apparently
chaotic. We tend to find this sort of
(emergent) behavior interesting but dif-
ficult to formalize. We are not aware of
any formal characterization of properties
that distinguish edge-of-chaos phenom-
ena from others.
To be clear, our definition of emergence
allows us to label as emergent anything
that is computable from a model state or
sequence of states. Thus any conceptual
construct that one can impose on a
model counts as an emergent phenome-
non.
This may seem overly simple or overly
broad, and
it certainly seems anti-
climactic. But this definition does seem
to include everything we think of as
emergent. Whether it also includes phe-
nomena that we would not want to con-
sider emergent is open to debate.22
22 Let ’ s contrast our definition of emergence with
Bedau ’ s. For Bedau, a property or phenomenon is
Emergence Explained
18/67
DRAFT
2/12/2006
More to the point, though, we are pro-
posing that the essence of emergence is
the independent conceptualization of the
emergent phenomenon.
• Any phenomenon that may be under-
stood in its own terms and whose
understanding does not depend on
knowing how it is implemented is
emergent under our definition.
• No phenomenon that cannot be so
understood will be emergent for us.
Emergent phenomena are often associ-
ated with a sense of surprise: where did
that come from? how did that happen?
Clearly the surprise of an observer has
little to do with how one should under-
stand nature. Surprise is not uncommon,
however, because phenomena are typi-
cally thought of as emergent when we
perceive them before we understand how
they are brought about. Sometimes it
may appear that there is no way for a
emergent [Bedau] if it is applicable at some macro
level but not at any micro sublevel. If one identi-
fies macro with epiphenomenal and micro with
underlying, our definition and Bedau ’ s are quite
similar except that Bedau explicitly excludes what
he calls trivially emergent properties. A property is
trivially emergent for Bedau if (to use our terms) it
is essentially the same property at the epiphe-
nomenal and underlying levels. As an example,
Bedau ’ s definition excludes the mass of an aggre-
gate from being an emergent property of the ag-
gregate because mass is also a property of the
components. Our definition does not make that ex-
clusion. For us the mass and (perhaps more inter-
estingly) the center of gravity of an aggregate both
qualify as emergent.
Our definition is similar to but less restrictive than
Shalizi ’ s [Shalizi 2001] and [Shalizi 2005], which
requires that an emergent phenomenon be compu-
tationally simpler than the phenomena from which
it emerges. If an epiphenomenon has its own for-
malizable abstraction, it will often (but even then
not always) be the case that computations per-
formed in terms of that abstraction will be simpler
than the equivalent computations performed in
terms of the underlying model. But we don ’ t re-
quire this. In addition, many epiphenomena do not
have easily formalizable abstractions.
surprising result to be brought about. Of
course there always is, but this reaction
is reminiscent of the remark by Arthur
C. Clarke that
(cid:11)(cid:6)(cid:30)(cid:8) (cid:12)(cid:18) (cid:22)(cid:22)(cid:17)(cid:7)(cid:17)(cid:3)(cid:6)(cid:16)(cid:20)(cid:30)(cid:8) (cid:11)% & (cid:11)(cid:6)(cid:7)(cid:3)% (cid:8) (cid:16)(cid:3)(cid:7)(cid:10)(cid:6)(cid:15)(cid:20)(cid:15)(cid:5)(cid:30)(cid:8) (cid:17)(cid:12)(cid:8) (cid:17)(cid:6)(cid:21)
% (cid:17)(cid:12)(cid:16)(cid:17)(cid:6)(cid:5)(cid:18) (cid:17)(cid:12)(cid:10)(cid:11)(cid:13)(cid:20)(cid:3)(cid:8)(cid:22)(cid:4)(cid:15)(cid:2)(cid:8)(cid:2)(cid:11)(cid:5)(cid:17)(cid:7)(cid:27)(cid:8)
4 Implications of emergence
The technique of implementing an ab-
stract design by using lower level con-
structs is not new; it is the bread and but-
ter of Computer Science. In a recent re-
view [CFCS] of the status of Computer
Science, the Committee on the Funda-
mentals of Computer Science: Chal-
lenges and Opportunities, National Re-
search Council wrote (p. 65) the follow-
ing.
(cid:9)6(cid:14)(cid:13)(cid:12)(cid:16)(cid:4)(cid:11)(cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:17)(cid:12)(cid:8) (cid:11)(cid:8) ( (cid:18) (cid:17)(cid:6)(cid:16)(cid:3)(cid:12)(cid:12)(cid:3)(cid:6)(cid:16)(cid:17)(cid:11)(cid:20)(cid:8) (cid:11)(cid:7)(cid:16)(cid:17)& (cid:17)(cid:16)(cid:30)(cid:8) (cid:15)(cid:22)(cid:8)
(cid:7)(cid:15)(cid:2)(cid:19) (cid:18) (cid:16)(cid:3)(cid:4)(cid:8) (cid:12)(cid:7)(cid:17)(cid:3)(cid:6)(cid:7)(cid:3)(cid:8) (cid:28)
(cid:8) (cid:27)(cid:8) 5 (cid:15)(cid:2)(cid:19) (cid:18) (cid:16)(cid:3)(cid:4)(cid:8) (cid:12)(cid:7)(cid:17)(cid:3)(cid:6)(cid:16)(cid:17)(cid:12)(cid:16)(cid:12)(cid:8)
(cid:7)(cid:4)(cid:3)(cid:11)(cid:16)(cid:3)(cid:8) (cid:11)(cid:6)% (cid:8) % (cid:17)(cid:12)(cid:7)(cid:11)(cid:4)% (cid:8) (cid:11)(cid:13)(cid:12)(cid:16)(cid:4)(cid:11)(cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)(cid:8) (cid:11)(cid:12)(cid:8) (cid:22)(cid:4)(cid:3)(cid:3)(cid:20)(cid:30)(cid:8)
(cid:11)(cid:12)(cid:8) (cid:3)(cid:6)(cid:5)(cid:17)(cid:6)(cid:3)(cid:3)(cid:4)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:11)(cid:4)(cid:7)(cid:10)(cid:17)(cid:16)(cid:3)(cid:7)(cid:16)(cid:12)(cid:8) (cid:7)(cid:4)(cid:3)(cid:11)(cid:16)(cid:3)(cid:8) (cid:11)(cid:6)% (cid:8)
(cid:8)
% (cid:17)(cid:12)(cid:7)(cid:11)(cid:4)% (cid:8)% (cid:3)(cid:12)(cid:17)(cid:5)(cid:6)(cid:8)(cid:12), (cid:3)(cid:16)(cid:7)(cid:10)(cid:3)(cid:12)(cid:27)(cid:8)(cid:28)
(cid:9). (cid:14)(cid:15)(cid:22)(cid:16)’ (cid:11)(cid:4)(cid:3)(cid:21)% (cid:3)(cid:12)(cid:17)(cid:5)(cid:6)(cid:8) (cid:16)(cid:3)(cid:7)(cid:10)(cid:6)(cid:17)( (cid:18) (cid:3)(cid:12)(cid:8) (cid:9)’ (cid:10)(cid:17)(cid:7)(cid:10)(cid:8) (cid:11)(cid:20)(cid:20)(cid:15)’ (cid:8)
(cid:15)(cid:6)(cid:3)(cid:8) (cid:16)(cid:15)(cid:8) (cid:4)(cid:3)(cid:16)(cid:11)(cid:17)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:17)(cid:6)% (cid:3)(cid:19) (cid:3)(cid:6)% (cid:3)(cid:6)(cid:7)(cid:3)(cid:8) (cid:15)(cid:22)(cid:8) (cid:11)(cid:6)(cid:8) (cid:11)(cid:13)(cid:21)
(cid:12)(cid:16)(cid:4)(cid:11)(cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:13)(cid:30)(cid:8) (cid:2)(cid:11)(cid:17)(cid:6)(cid:16)(cid:11)(cid:17)(cid:6)(cid:17)(cid:6)(cid:5)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) % (cid:17)(cid:12)(cid:16)(cid:17)(cid:6)(cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:13)(cid:3)(cid:21)
(cid:16)’ (cid:3)(cid:3)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:11)(cid:13)(cid:12)(cid:16)(cid:4)(cid:11)(cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:17)(cid:16)(cid:12)(cid:8) (cid:17)(cid:2)(cid:19) (cid:20)(cid:3)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:21)
(cid:16)(cid:17)(cid:15)(cid:6)(cid:14)23(cid:8) (cid:19) (cid:4)(cid:15)& (cid:17)% (cid:3)(cid:8) ’ (cid:11)(cid:30)(cid:12)(cid:8) (cid:16)(cid:15)(cid:8) (cid:15)(cid:4)(cid:5)(cid:11)(cid:6)(cid:17)(cid:31) (cid:3)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:11)(cid:13)(cid:21)
(cid:12)(cid:16)(cid:4)(cid:11)(cid:7)(cid:16)(cid:8) % (cid:3)(cid:22)(cid:17)(cid:6)(cid:17)(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:17)(cid:6)(cid:22)(cid:15)(cid:4)(cid:2)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:30)(cid:8)
(cid:7)(cid:15)(cid:6)(cid:16)(cid:4)(cid:15)(cid:20)(cid:27)(cid:8) (cid:28)
(cid:8) (cid:9)9 (cid:14)(cid:4)(cid:15)(cid:5)(cid:4)(cid:11)(cid:2)(cid:2)(cid:17)(cid:6)(cid:5)(cid:8) (cid:20)(cid:11)(cid:6)(cid:5)(cid:18) (cid:11)(cid:5)(cid:3)(cid:12)(cid:8) (cid:19) (cid:4)(cid:15)(cid:21)
& (cid:17)% (cid:3)(cid:8) (cid:11)(cid:8) (cid:6)(cid:15)(cid:16)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:16)(cid:15)(cid:8) (cid:3)(cid:6)(cid:7)(cid:15)% (cid:3)(cid:8) (cid:11)(cid:13)(cid:12)(cid:16)(cid:4)(cid:11)(cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)(cid:8) (cid:12)(cid:15)(cid:8)
(cid:11)(cid:12)(cid:8) (cid:16)(cid:15)(cid:8) (cid:11)(cid:20)(cid:20)(cid:15)’ (cid:8) (cid:16)(cid:10)(cid:3)(cid:17)(cid:4)(cid:8) % (cid:17)(cid:4)(cid:3)(cid:7)(cid:16)(cid:8) (cid:3)*(cid:3)(cid:7)(cid:18) (cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:13)(cid:30)(cid:8) (cid:7)(cid:15)(cid:2)(cid:21)
(cid:19) (cid:18) (cid:16)(cid:3)(cid:4)(cid:27)(cid:8)
This section explores the implications of
emergence that our Turing Machine ex-
ample illustrates.
23 Although the bracketed texts says a bit more than
was literally in the original, we believe it expresses
the authors’ intentions.
Emergence Explained
19/67
DRAFT
2/12/2006
4.1 Non-reductive regularities
Recall Weinberg’ s statement.
(cid:9)) (cid:14)(cid:10)(cid:3)(cid:4)(cid:3)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:6)(cid:15)(cid:8) (cid:11)(cid:18) (cid:16)(cid:15)(cid:6)(cid:15)(cid:2)(cid:15)(cid:18) (cid:12)(cid:8)
(cid:20)(cid:11)’ (cid:12)(cid:8) (cid:15)(cid:22)(cid:8)
’ (cid:3)(cid:11)(cid:16)(cid:10)(cid:3)(cid:4)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:20)(cid:15)(cid:5)(cid:17)(cid:7)(cid:11)(cid:20)(cid:20)(cid:30)(cid:8) (cid:17)(cid:6)% (cid:3)(cid:19) (cid:3)(cid:6)% (cid:3)(cid:6)(cid:16)(cid:8) (cid:15)(cid:22)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:8)(cid:19) (cid:4)(cid:17)(cid:6)(cid:7)(cid:17)(cid:19) (cid:20)(cid:3)(cid:12)(cid:8)(cid:15)(cid:22)(cid:8)(cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:12)(cid:27)(cid:8)(cid:8)
Clearly there are lots of autonomous
“ laws ” of Turing Machines (namely
computability theory), and they are all
logically independent of the rules of the
Game of Life.
The fact that one can emulate a Turing
Machine on a Game of Life platform
tells us nothing about Turing Ma-
chines — other than that they can be emu-
lated by using the Game of Life.
An emulation of a Turing Machine on
the Game of Life is an example of what
might be called a non-reductive regular-
ity. The Turing Machine and its emula-
tion is certainly a kind of regularity, but
the regularity that it embodies (i.e., that
it is a model for computability) is not a
logical consequence of (i.e., is not re-
ducible to and cannot be deduced from)
the Game of Life rules.
Facts about Turing Machines, i.e., the
theorems of computability theory, are
derived de novo. They are made up out
of whole cloth; they are not based on the
Game of Life rules. The fact that such
abstract designs can be realized using
Game of Life rules as an implementation
platform tells us nothing about comput-
ability theory that we don’ t already
know.
4.2 Downward entailment
On the other hand, the fact that a Turing
Machine can be implemented using the
Game of Life rules as primitives does
tell us something about the Game of
Life—
namely that the results of comput-
ability theory can be applied to the
Game of Life. The property of being
Turing complete applies to the Game of
Life precisely because a Turing Machine
can be shown to be one of its possible
epiphenomena.
In other words, epiphenomena are
downward
entailing. Properties of
epiphenomena are also properties of the
phenomena from which they spring. This
is not quite as striking as downward cau-
sation24 would be, but it is a powerful in-
tellectual tool.
Earlier, we dismissed the notion that a
glider may be said to “ go to a cell and
turns it on.” The only things that turn on
Game of Life grid cells are the Game of
Life rules. But because of downward en-
tailment, there is hope for talk of this
sort.
To prove that a Turing Machine emula-
tion on a Game of Life platform does
what we claim it does we disengaged the
design of the emulation from its imple-
mentation, and we reasoned about the
emulation as an abstract design. We can
do the same thing with gliders. We can
establish a domain of discourse about
gliders as abstract entities. Within that
domain of discourse we can reason about
gliders, and in particular we can reason
about which cells a glider will turn on
and when it will turn them on. Our the-
ory will tell us how fast an in which di-
rection a glider moves
Having developed facts and rules about
gliders as independent abstract elements,
we can then use the fact that gliders are
epiphenomena of the Game of Life and
by appeal to downward entailment apply
those facts and rules to the Game of Life
cells that implement gliders.
24 See, for example [Emmeche] for a number of so-
phisticated discussions of downward causation.
Emergence Explained
20/67
DRAFT
2/12/2006
This sounds more complex than it really
is. What it really amounts to is that
downward entailment justifies what we
as human beings tend to do anyway: no-
tice regularities in the world and then
engage the world in terms of those regu-
larities.
4.3 Reduction proofs
Consider how the unsolvability of the
halting problem applies to the Game of
Life. The fact that a Turing Machine can
be implemented on a Game of Life plat-
form means, among other things that the
halting problem for the Game of Life—
which we can define as determining
whether a game of Life run ever reaches
a stable (unchanging or repeating) con-
figuration—
is unsolvable.
When reasoning about insolvability one
often talks of reducing one problem to
another. In this case, because we can
implement Turing Machines using the
Game of Life, we know that we can re-
duce the halting problem for Turing Ma-
chines to the halting problem for the
Game of Life: if we could solve the
Game of Life halting problem, we could
solve the Turing Machine halting prob-
lem. But we know that the Turing Ma-
chine halting problem is unsolvable.
Therefore the Game of Life halting prob-
lem is also unsolvable. This sort of
downward entailment reduction gives us
a lot of intellectual leverage since it ’ s
not at all clear how difficult it would be
to prove “ directly” that the halting prob-
lem for the Game of Life is unsolvable.
Thus another consequence of downward
entailment is that reducibility cuts both
ways. One can conclude that if some-
thing is impossible at a higher level it
must be impossible at the lower (imple-
mentation) level as well. But the only
way to reach that conclusion is to reason
about the higher level as an independent
abstraction and then to reconnect that
abstraction to the lower level. Logically
independent higher level abstractions,
i.e., functionality and design, matter on
their own.
4.4 Downward entailment as sci-
ence
A striking example of downward entail-
ment is the kind of computation we do
when computing the effect of one bil-
liard ball on another in a Newtonian uni-
verse. It ’ s a simple calculation involving
vectors and the transfer of kinetic en-
ergy.
In truth there is no fundamental force of
physics corresponding to kinetic energy.
If one had to compute the consequences
of a billiard ball collision in terms of
quantum states and the electromagnetic
force, which is the one that applies, the
task would be impossibly complex. But
the computation is easy to do at the
epiphenomenal level of billiard balls.
We know that the computation we do at
the billiard ball level applies to the real
world because of downward entailment:
billiard balls are epiphenomena of the
underlying reality. So even though it
may be literally incorrect to say that a
glider “ turns on a Game of Life cell ” or
that “ one billiard ball pushes another one
in a particular direction,” because of
downward entailment this sort of con-
ceptual shorthand is not only reasonable
but essential for how we think about the
world. But see the discussion of Newto-
nian mechanics in the next section for a
somewhat stronger version of this per-
spective.
Downward entailment in the form of the
process just sketched is, in fact, a rea-
sonable description of how we do sci-
ence: we build models, which we then
apply to the world around us.
Emergence Explained
21/67
DRAFT
2/12/2006
We are not saying that there are forces in
the world that operate according to bil-
liard ball rules or that there are forces in
the Game of Life that operate according
to glider rules. That would be downward
causation, a form of strong emergence,
which we have already ruled out. What
we are saying is that billiard balls, glid-
ers, Turing Machines, and their interac-
tions can be defined in the abstract. We
can reason about them as abstractions,
and then through downward entailment
we can apply the results of that reason-
ing to any implementation of those ab-
stractions whenever the implementation
preserves the assumptions required by
the abstraction.
4.5 The reality of higher level ab-
stractions
In “ Real Patterns ” [Dennett ‘ 91], Den-
nett argues that when compared with the
work required to compute the equivalent
results in terms of primitive forces, one
gets a “ stupendous ” “ scale of compres-
sion” when one adopts his notion of an
intentional stance [Dennett ‘ 87]. Al-
though “ Real Patterns ” doesn’ t spell out
the link explicitly, Dennett ’ s position
appears to be that because of that intel-
lectual advantage, one should treat the
ontologies offered by the intentional
stance as what he calls “ mildly real ”
although he doesn’ t spell out in any de-
tail what
something as
regarding
“ mildly real ” involves.
Our position contrasts with Dennett ’ s in
that we claim that nature is often best
the theories of science are
understood—
best expressed—
in terms of two-level
(or perhaps multi-level) theories. One
level is an abstract design; the other
level is the implementation of that de-
sign. Whether or not one wants to say
that the abstract design is “ mildly real ”
(or real with some other adjective ap-
plied) is not our focus.25
As we shall see below, our claim will
also be that the entities (such as billiard
balls) about which higher level abstrac-
tions are formulated are real in an objec-
tive sense (they have reduced entropy)
but that interactions among those entities
are epiphenomenal —
the only
since
forces in nature are the fundamental
forces.
In a recent book [Laughlin], Laughlin
argues for what he calls collective prin-
ciples of organization, which he finds to
be at least as important as reductionist
principles. For example in discussing
Newton’ s laws he concludes from the
fact that (p. 31)
(cid:16)(cid:10)(cid:3)(cid:12)(cid:3)(cid:8) (cid:9)(cid:15)(cid:16)(cid:10)(cid:3)(cid:4)’ (cid:17)(cid:12)(cid:3)(cid:14)(cid:8) (cid:15)& (cid:3)(cid:4)’ (cid:10)(cid:3)(cid:20)(cid:2)(cid:17)(cid:6)(cid:5)(cid:20)(cid:30)(cid:8) (cid:12)(cid:18) (cid:7)(cid:7)(cid:3)(cid:12)(cid:12)(cid:21)
(cid:22)(cid:18) (cid:20)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8) (cid:28)
(cid:8) (cid:2)(cid:11), (cid:3)(cid:8) (cid:19) (cid:4)(cid:15)(cid:22)(cid:15)(cid:18) (cid:6)% (cid:20)(cid:30)(cid:8) ’ (cid:4)(cid:15)(cid:6)(cid:5)(cid:8) (cid:19) (cid:4)(cid:3)% (cid:17)(cid:7)(cid:21)
(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)(cid:8)(cid:11)(cid:16)(cid:8)(cid:9)(cid:16)(cid:10)(cid:3)(cid:8)( (cid:18) (cid:11)(cid:6)(cid:16)(cid:18) (cid:2)(cid:14)(cid:8)(cid:12)(cid:7)(cid:11)(cid:20)(cid:3)(cid:8)
that
D (cid:3)’ (cid:16)(cid:15)(cid:6)$(cid:12)(cid:8) (cid:20)(cid:3)(cid:5)(cid:3)(cid:6)% (cid:11)(cid:4)(cid:30)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8) (cid:10)(cid:11)& (cid:3)(cid:8) (cid:16)(cid:18) (cid:4)(cid:6)(cid:3)% (cid:8) (cid:15)(cid:18) (cid:16)(cid:8)
(cid:16)(cid:15)(cid:8) (cid:13)(cid:3)(cid:8) (cid:3)(cid:2)(cid:3)(cid:4)(cid:5)(cid:3)(cid:6)(cid:16)(cid:27)(cid:8) ) (cid:10)(cid:3)(cid:30)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:6)(cid:15)(cid:16)(cid:8) (cid:22)(cid:18) (cid:6)% (cid:11)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:20)(cid:8)
(cid:11)(cid:16)(cid:8)(cid:11)(cid:20)(cid:20)(cid:8)(cid:13)(cid:18) (cid:16)(cid:8)(cid:11)(cid:8)(cid:7)(cid:15)(cid:6)(cid:12)(cid:3)( (cid:18) (cid:3)(cid:6)(cid:7)(cid:3)(cid:8)(cid:15)(cid:22)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:11)(cid:5)(cid:5)(cid:4)(cid:3)(cid:5)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8)
(cid:15)(cid:22)(cid:8) ( (cid:18) (cid:11)(cid:6)(cid:16)(cid:18) (cid:2)(cid:8) (cid:2)(cid:11)(cid:16)(cid:16)(cid:3)(cid:4)(cid:8) (cid:17)(cid:6)(cid:16)(cid:15)(cid:8) (cid:2)(cid:11)(cid:7)(cid:4)(cid:15)(cid:12)(cid:7)(cid:15)(cid:19) (cid:17)(cid:7)(cid:8) (cid:22)(cid:20)(cid:18) (cid:17)% (cid:12)(cid:8)
(cid:11)(cid:6)% (cid:8) (cid:12)(cid:15)(cid:20)(cid:17)% (cid:12)2(cid:11)(cid:8) (cid:7)(cid:15)(cid:20)(cid:20)(cid:3)(cid:7)(cid:16)(cid:17)& (cid:3)(cid:8) (cid:15)(cid:4)(cid:5)(cid:11)(cid:6)(cid:17)(cid:31) (cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:11)(cid:20)(cid:8)
(cid:19) (cid:10)(cid:3)(cid:6)(cid:15)(cid:2)(cid:3)(cid:6)(cid:15)(cid:6)(cid:27)(cid:8) (cid:28)
(cid:8) (cid:9)D (cid:3)’ (cid:16)(cid:15)(cid:6)$(cid:12)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:14)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:11)(cid:12)(cid:8)
(cid:3)*(cid:11)(cid:7)(cid:16)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:16)(cid:4)(cid:18) (cid:3)(cid:8) (cid:11)(cid:12)(cid:8) (cid:11)(cid:6)(cid:30)(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:8) ’ (cid:3)(cid:8) , (cid:6)(cid:15)’ (cid:8) (cid:17)(cid:6)(cid:8)
(cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:12)2(cid:30)(cid:3)(cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:30)(cid:8) & (cid:11)(cid:6)(cid:17)(cid:12)(cid:10)(cid:8) (cid:17)(cid:6)(cid:16)(cid:15)(cid:8) (cid:6)(cid:15)(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:6)(cid:3)(cid:12)(cid:12)(cid:8)
’ (cid:10)(cid:3)(cid:6)(cid:8) (cid:3)*(cid:11)(cid:2)(cid:17)(cid:6)(cid:3)% (cid:8) (cid:16)(cid:15)(cid:15)(cid:8) (cid:7)(cid:20)(cid:15)(cid:12)(cid:3)(cid:20)(cid:30)(cid:27)(cid:8) (cid:28)
(cid:8) (cid:9)E (cid:14)(cid:11)(cid:6)(cid:30)(cid:8)
(cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:17)(cid:12)(cid:16)(cid:12)(cid:8) (cid:4)(cid:3)(cid:2)(cid:11)(cid:17)(cid:6)(cid:8) (cid:17)(cid:6)(cid:8) % (cid:3)(cid:6)(cid:17)(cid:11)(cid:20)(cid:27)(cid:8) ) (cid:15)(cid:8) (cid:16)(cid:10)(cid:17)(cid:12)(cid:8) % (cid:11)(cid:30)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:30)(cid:8) (cid:15)(cid:4)(cid:5)(cid:11)(cid:6)(cid:17)(cid:31) (cid:3)(cid:8) (cid:7)(cid:15)(cid:6)(cid:22)(cid:3)(cid:4)(cid:3)(cid:6)(cid:7)(cid:3)(cid:12)(cid:8) (cid:15)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:18) (cid:13)3(cid:3)(cid:7)(cid:16)(cid:8)
(cid:11)(cid:6)% (cid:8) (cid:4)(cid:15)(cid:18) (cid:16)(cid:17)(cid:6)(cid:3)(cid:20)(cid:30)(cid:8) (cid:12)(cid:19) (cid:3)(cid:11), (cid:8) (cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8) D (cid:3)’ (cid:16)(cid:15)(cid:6)$(cid:12)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8)
(cid:13)(cid:3)(cid:17)(cid:6)(cid:5)(cid:8) (cid:11)(cid:6)(cid:8) /(cid:11)(cid:19) (cid:19) (cid:4)(cid:15)*(cid:17)(cid:2)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)0 (cid:8) (cid:22)(cid:15)(cid:4)(cid:8) ( (cid:18) (cid:11)(cid:6)(cid:16)(cid:18) (cid:2)(cid:8) (cid:2)(cid:3)(cid:21)
(cid:7)(cid:10)(cid:11)(cid:6)(cid:17)(cid:7)(cid:12)-(cid:8) & (cid:11)(cid:20)(cid:17)% (cid:8) ’ (cid:10)(cid:3)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:30)(cid:12)(cid:16)(cid:3)(cid:2)(cid:8) (cid:17)(cid:12)(cid:8) (cid:20)(cid:11)(cid:4)(cid:5)(cid:3)2
(cid:3)& (cid:3)(cid:6)(cid:8) (cid:16)(cid:10)(cid:15)(cid:18) (cid:5)(cid:10)(cid:8) (cid:6)(cid:15)(cid:8) (cid:20)(cid:3)(cid:5)(cid:17)(cid:16)(cid:17)(cid:2)(cid:11)(cid:16)(cid:3)(cid:8) (cid:11)(cid:19) (cid:19) (cid:4)(cid:15)*(cid:17)(cid:2)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8)
(cid:12)(cid:7)(cid:10)(cid:3)(cid:2)(cid:3)(cid:8)(cid:10)(cid:11)(cid:12)(cid:8)(cid:3)& (cid:3)(cid:4)(cid:8)(cid:13)(cid:3)(cid:3)(cid:6)(cid:8)(cid:22)(cid:15)(cid:18) (cid:6)% (cid:27)(cid:8)(cid:8)
25 Furthermore, our position is that the question of
whether or not anyone would find it advantageous
to have beliefs about epiphenomena (which is
Dennett ’ s primary concern) has nothing to do with
the issues we are considering.
Emergence Explained
22/67
—
DRAFT
2/12/2006
A second example to which Laughlin
frequently returns is the solid state of
matter, which, as he points out, exhibits
properties of rigidity and elasticity. The
solid state of matter may be character-
ized as material that may be understood
as a three dimensional lattice of compo-
nents held together by forces acting
among those components.
Once one has defined an abstract struc-
ture of this sort, one can derive proper-
ties of matter having this structure. One
can do so without knowing anything
more about either (a) the particular ele-
ments at the lattice nodes or (b) how the
binding forces are implemented. All one
needs to know are the strengths of the
forces and the shape of the lattice.
From our perspective, both Newton ’ s
laws and the solid state of matter are ab-
stract organizational designs, i.e., epi-
phenomena. They are abstractions that
apply to nature in much the same way as
a Turing Machine as an abstraction ap-
plies to certain cell configurations in the
Game of Life. Laughlin calls the imple-
mentation of such an abstraction a pro-
tectorate.
Laughlin points out that protectorates
tend to have feasibility ranges, which are
often characterized by size, speed, and
temperature. A few molecules of H2O
won’ t have the usual properties of ice.
And ice, like most solids, melts when
heated to a point at which the attractive
forces are no longer able to preserve the
lattice configuration of the elements.
Similarly Newton’ s laws fail at the quan-
tum level.
The existence of such feasibility ranges
does not reduce the importance of either
the solid matter abstraction or the New-
tonian physics abstraction. They just
limit the conditions under which those
abstractions apply, i.e., under which na-
ture is able to implement them.
The more general point is that nature
implements a great many such abstract
designs. As is the case with computabil-
ity theory, which includes many sophis-
ticated results about the Turing machine
abstraction, there are often sophisticated
theories that characterize the properties
of such naturally occurring abstractions.
These theories may have nothing to do
with how the abstract designs are im-
plemented. They are functional theories
that apply to the abstract designs them-
selves. To apply such theories to a real
physical example (through downward
entailment), all one needs is for the
physical example to implement the ab-
stract designs.
Furthermore and perhaps more impor-
tantly, these abstract designs are neither
derivable from nor logical consequences
of their implementations. Abstract de-
signs and the theories built on them are
new and creative constructs and are not
consequences of the platform on which
they are implemented. The Game of Life
doesn’ t include the concept of a Turing
machine, and quantum physics doesn’ t
include the concept of a solid.
The point of all this is to support Laugh-
lin position: when nature implements an
abstraction, the epiphenomena described
by that abstraction become just as real
any other phenomena, and the abstrac-
tion that describes them is just as valid a
description of that aspect of nature as
any other description of any other aspect
of nature.
Our notion that much of nature is best
understood in terms of implementations
of abstractions suggests that many scien-
tific theories are best expressed at two
levels: (1) the level of an abstraction it-
self, i.e., how it is specified, how it
Emergence Explained
23/67
DRAFT
2/12/2006
works on the abstract level, and what its
implications are, and (2) the level that
explains (a) under what circumstances
(when and where) that abstraction may
be found implemented and (b) how that
implementation works.
The use of agent-based modeling in the
social sciences illustrates this methodol-
ogy. Agent-based models are typically
used to show both (a) how a higher level
abstraction functions in its context, i.e.,
what its consequences as a theory are,
and (b) how that abstraction may come
to be realized in terms of some particular
set of lower level interactions.
4.6 Abstractions and abstract
designs
We have intentionally used the terms ab-
straction and abstract design somewhat
interchangeably. Abstraction
is, of
course, a much more general term. Yet
from the perspective of scientific expla-
nation, what one is looking for is, as
Weinberg says, a description of why na-
ture is the way it is.
To some extent, this includes the ques-
tion of why nature is made up of what-
ever makes it up. But for the most part,
this question is generally taken to be
asking why nature works the way it
does.
As summarized by Woodward [Wood-
ward], scientific explanations are in-
tended to explain
(cid:24) (cid:25) (cid:15) (cid:8) (cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:12)(cid:8) (cid:10)(cid:11)(cid:19) (cid:19) (cid:3)(cid:6)-(cid:8) ’ (cid:10)(cid:3)(cid:4)(cid:3)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) / (cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:12)0 (cid:8) (cid:17)(cid:6)(cid:8)
( (cid:18) (cid:3)(cid:12)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:7)(cid:11)(cid:6)(cid:8) (cid:13)(cid:3)(cid:8) (cid:3)(cid:17)(cid:16)(cid:10)(cid:3)(cid:4)(cid:8) (cid:19) (cid:11)(cid:4)(cid:16)(cid:17)(cid:7)(cid:18) (cid:20)(cid:11)(cid:4)(cid:8) (cid:3)& (cid:3)(cid:6)(cid:16)(cid:12)(cid:8) (cid:15)(cid:4)(cid:8)
(cid:12)(cid:15)(cid:2)(cid:3)(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:8) (cid:2)(cid:15)(cid:4)(cid:3)(cid:8) (cid:5)(cid:3)(cid:6)(cid:3)(cid:4)(cid:11)(cid:20)2(cid:3)(cid:27)(cid:5)(cid:27)-(cid:8) (cid:4)(cid:3)(cid:5)(cid:18) (cid:20)(cid:11)(cid:4)(cid:17)(cid:16)(cid:17)(cid:3)(cid:12)(cid:8)
(cid:15)(cid:4)(cid:8)(cid:4)(cid:3)(cid:19) (cid:3)(cid:11)(cid:16)(cid:11)(cid:13)(cid:20)(cid:3)(cid:8)(cid:19) (cid:11)(cid:16)(cid:16)(cid:3)(cid:4)(cid:6)(cid:12)(cid:8)(cid:17)(cid:6)(cid:8)(cid:6)(cid:11)(cid:16)(cid:18) (cid:4)(cid:3)(cid:27)(cid:8)
Typically we are looking for an opera-
tional explanation: how does nature
work?26 Thus when we speak of an ab-
stract design, we are referring to the de-
sign of an operational mechanism that
can be understood as bringing about
some consequence. From here on, when
we talk about abstractions or abstract
designs, this is what we have in mind:
the abstract design of a mechanism that
produces certain kinds of results.
Note that it is not just the functional re-
sults that matter; the abstract operational
design matters also. A Turing Machine
is not just a device whose input-output
functionality satisfies the theorems of
computability theory; it is an abstract
device that operates in a particular way.
The theorems of computability theory
follow from the way a Turing Machine
operates, not the other way around. It is
the abstract operational design
that
comes first. The theory follows.
This operational perspective seems to be
somewhat different from that of tradi-
tional functionalism. Functionalism is
concerned with the functions and regu-
larities that characterize the special sci-
ences. It tends not to be concerned with
how those functions or regularities are
realized. In fact the opposite seems to be
the case. By emphasizing the possibility
of multiple realizability, functionalism
minimizes the importance of any par-
ticular realization and dismisses the im-
portance of understanding operational
design issues.
4.7 Phase transitions
Those of us in Computer Science know
that implementations of abstract designs
are often not perfect. Most include some
compromises, and some have bugs.
26 As Woodward explains, though, the question of
what should be taken as a scientific explanation is
the subject of continuing investigation.
Emergence Explained
24/67
DRAFT
2/12/2006
What about nature’ s implementation of
abstract designs?
Nature’ s implementation of abstract de-
signs are not always perfect either—
especially, as Laughlin points out, either
(a) when the conditions under which an
implementation is feasible are not in ef-
fect or (b) when one approaches the
boundaries of such feasibility regions.
There will almost always be borderline
situations in which the implementation
of an abstract design is on the verge of
breaking down. These borderline situa-
tions frequently manifest as what we call
phase transitions —
regions or points (re-
lated to a parameter such as size, speed,
temperature, and pressure) where multi-
ple distinct and incompatible abstrac-
tions may to be implemented.
Newton’ s laws fail at both the quantum
level and at relativistic speeds. If as
Laughlin suggests, the Newtonian ab-
straction cannot be shown to be an ap-
proximation of quantum theory, phase
transitions should appear as one ap-
proaches the quantum realm.
As explained by Sachdev [Sachdev], the
transition from a Newtonian gas to a
Boise-Einstein condensate (such as su-
per-fluid liquid helium) illustrates such a
phase transition.
6(cid:16)(cid:8)(cid:4)(cid:15)(cid:15)(cid:2)(cid:8)(cid:16)(cid:3)(cid:2)(cid:19) (cid:3)(cid:4)(cid:11)(cid:16)(cid:18) (cid:4)(cid:3)-(cid:8)(cid:11)(cid:8)(cid:5)(cid:11)(cid:12)(cid:8)(cid:12)(cid:18) (cid:7)(cid:10)(cid:8)(cid:11)(cid:12)(cid:8)(cid:10)(cid:3)(cid:20)(cid:17)(cid:18) (cid:2)(cid:8)
(cid:7)(cid:15)(cid:6)(cid:12)(cid:17)(cid:12)(cid:16)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:4)(cid:11)(cid:19) (cid:17)% (cid:20)(cid:30)(cid:8) (cid:2)(cid:15)& (cid:17)(cid:6)(cid:5)(cid:8) (cid:11)(cid:16)(cid:15)(cid:2)(cid:12)-(cid:8) (cid:11)(cid:6)% (cid:8) (cid:7)(cid:11)(cid:6)(cid:8)
(cid:13)(cid:3)(cid:8) & (cid:17)(cid:12)(cid:18) (cid:11)(cid:20)(cid:17)(cid:31) (cid:3)% (cid:8) (cid:11)(cid:12)(cid:8) (cid:7)(cid:20)(cid:11)(cid:12)(cid:12)(cid:17)(cid:7)(cid:11)(cid:20)(cid:8) (cid:13)(cid:17)(cid:20)(cid:20)(cid:17)(cid:11)(cid:4)% (cid:8) (cid:13)(cid:11)(cid:20)(cid:20)(cid:12)(cid:8)
’ (cid:10)(cid:17)(cid:7)(cid:10)(cid:8)(cid:7)(cid:15)(cid:20)(cid:20)(cid:17)% (cid:3)(cid:8)’ (cid:17)(cid:16)(cid:10)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)’ (cid:11)(cid:20)(cid:20)(cid:12)(cid:8)(cid:15)(cid:22)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:7)(cid:15)(cid:6)(cid:16)(cid:11)(cid:17)(cid:6)(cid:3)(cid:4)(cid:8)
(cid:11)(cid:6)% (cid:8)(cid:15)(cid:7)(cid:7)(cid:11)(cid:12)(cid:17)(cid:15)(cid:6)(cid:11)(cid:20)(cid:20)(cid:30)(cid:8)’ (cid:17)(cid:16)(cid:10)(cid:8)(cid:3)(cid:11)(cid:7)(cid:10)(cid:8)(cid:15)(cid:16)(cid:10)(cid:3)(cid:4)(cid:27)(cid:8)
6(cid:12)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:16)(cid:3)(cid:2)(cid:19) (cid:3)(cid:4)(cid:11)(cid:16)(cid:18) (cid:4)(cid:3)(cid:8) (cid:17)(cid:12)(cid:8) (cid:20)(cid:15)’ (cid:3)(cid:4)(cid:3)% -(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:11)(cid:16)(cid:15)(cid:2)(cid:12)(cid:8)
(cid:12)(cid:20)(cid:15)’ (cid:8) % (cid:15)’ (cid:6)(cid:8) (cid:9)(cid:11)(cid:6)% (cid:14)(cid:8) (cid:16)(cid:10)(cid:3)(cid:17)(cid:4)(cid:8) ( (cid:18) (cid:11)(cid:6)(cid:16)(cid:18) (cid:2)(cid:21)(cid:2)(cid:3)(cid:7)(cid:10)(cid:11)(cid:6)(cid:17)(cid:7)(cid:11)(cid:20)(cid:8)
(cid:7)(cid:10)(cid:11)(cid:4)(cid:11)(cid:7)(cid:16)(cid:3)(cid:4)(cid:17)(cid:12)(cid:16)(cid:17)(cid:7)(cid:12)(cid:8) (cid:13)(cid:3)(cid:7)(cid:15)(cid:2)(cid:3)(cid:8) (cid:17)(cid:2)(cid:19) (cid:15)(cid:4)(cid:16)(cid:11)(cid:6)(cid:16)(cid:27)(cid:8) D (cid:15)’ (cid:8) ’ (cid:3)(cid:8)
(cid:10)(cid:11)& (cid:3)(cid:8) (cid:16)(cid:15)(cid:8) (cid:16)(cid:10)(cid:17)(cid:6), (cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:11)(cid:16)(cid:15)(cid:2)(cid:12)(cid:8) (cid:11)(cid:12)(cid:8) (cid:15)(cid:7)(cid:7)(cid:18) (cid:19) (cid:30)(cid:17)(cid:6)(cid:5)(cid:8)
(cid:12)(cid:19) (cid:3)(cid:7)(cid:17)(cid:22)(cid:17)(cid:7)(cid:8)
( (cid:18) (cid:11)(cid:6)(cid:16)(cid:18) (cid:2)(cid:8) (cid:12)(cid:16)(cid:11)(cid:16)(cid:3)(cid:12)(cid:8)
’ (cid:10)(cid:17)(cid:7)(cid:10)(cid:8) (cid:3)*(cid:16)(cid:3)(cid:6)% (cid:8)
(cid:11)(cid:7)(cid:4)(cid:15)(cid:12)(cid:12)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:3)(cid:6)(cid:16)(cid:17)(cid:4)(cid:3)(cid:8) & (cid:15)(cid:20)(cid:18) (cid:2)(cid:3)(cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:7)(cid:15)(cid:6)(cid:16)(cid:11)(cid:17)(cid:6)(cid:3)(cid:4)(cid:27)(cid:8)
(cid:28)
(cid:8) (cid:11)(cid:12)(cid:8) (cid:17)(cid:12)(cid:8) (cid:10)(cid:3)(cid:21)
(cid:8) (cid:9)+(cid:14)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:11)(cid:16)(cid:15)(cid:2)(cid:12)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) #(cid:13)(cid:15)(cid:12)(cid:15)(cid:6)(cid:12)$(cid:8) ! (cid:28)
(cid:20)(cid:17)(cid:18) (cid:2)" (cid:8) (cid:11)(cid:6)(cid:8) (cid:11)(cid:4)(cid:13)(cid:17)(cid:16)(cid:4)(cid:11)(cid:4)(cid:30)(cid:8) (cid:6)(cid:18) (cid:2)(cid:13)(cid:3)(cid:4)(cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:2)(cid:8) (cid:7)(cid:11)(cid:6)(cid:8) (cid:15)(cid:7)(cid:21)
(cid:8) +(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)
(cid:7)(cid:18) (cid:19) (cid:30)(cid:8) (cid:11)(cid:6)(cid:30)(cid:8) (cid:12)(cid:17)(cid:6)(cid:5)(cid:20)(cid:3)(cid:8) ( (cid:18) (cid:11)(cid:6)(cid:16)(cid:18) (cid:2)(cid:8) (cid:12)(cid:16)(cid:11)(cid:16)(cid:3)(cid:8) (cid:28)
(cid:8) (cid:9) (cid:14) (cid:9) (cid:11) (cid:15) (cid:11)(cid:16)(cid:15)(cid:2)(cid:8)
(cid:16)(cid:3)(cid:2)(cid:19) (cid:3)(cid:4)(cid:11)(cid:16)(cid:18) (cid:4)(cid:3)(cid:8) (cid:17)(cid:12)(cid:8) (cid:20)(cid:15)’ (cid:8) (cid:3)(cid:6)(cid:15)(cid:18) (cid:5)(cid:10)(cid:8) (cid:28)
’ (cid:17)(cid:20)(cid:20)(cid:8) (cid:15)(cid:7)(cid:7)(cid:18) (cid:19) (cid:30)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:18) (cid:22) (cid:10)(cid:9) (cid:20)(cid:15)’ (cid:3)(cid:12)(cid:16)(cid:8) (cid:3)(cid:6)(cid:3)(cid:4)(cid:5)(cid:30)(cid:8) (cid:28)
(cid:8)(cid:8)
( (cid:18) (cid:11)(cid:6)(cid:16)(cid:18) (cid:2)(cid:8)(cid:12)(cid:16)(cid:11)(cid:16)(cid:3)(cid:27)(cid:8)
On the other hand, since Newton’ s laws
are indeed an approximation of relativis-
tic physics, there are no Newtonian-
related phase transitions as one ap-
proaches relativistic speeds.
These considerations suggest that when-
ever data that suggests a phase transition
appears, one should look for two or more
abstractions with implementations hav-
ing overlapping or adjacent feasibility
regions.
Speculatively, one might even view
quantum probability amplitude waves as
phase transition phenomena. Quantum
states are discrete. Matter may not oc-
cupy states that are intermediate between
them. Consequently, matter cannot tran-
sition smoothly from one quantum state
to another. As Hardy suggests [Hardy],
by making such transitions probabilisti-
cally continuous,
( (cid:18) (cid:11)(cid:6)(cid:16)(cid:18) (cid:2)(cid:8)(cid:16)(cid:10)(cid:3)(cid:15)(cid:4)(cid:30)(cid:8)(cid:15)(cid:22)(cid:22)(cid:3)(cid:4)(cid:12)(cid:8)(cid:18) (cid:12)(cid:8)(cid:11)(cid:8)’ (cid:11)(cid:30)(cid:8)(cid:16)(cid:15)(cid:8)(cid:10)(cid:11)& (cid:3)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)
(cid:11)% & (cid:11)(cid:6)(cid:16)(cid:11)(cid:5)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) % (cid:17)(cid:12)(cid:7)(cid:4)(cid:3)(cid:16)(cid:3)(cid:6)(cid:3)(cid:12)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:7)(cid:15)(cid:6)(cid:16)(cid:17)(cid:6)(cid:18) (cid:17)(cid:16)(cid:30)(cid:8)
(cid:11)(cid:16)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:12)(cid:11)(cid:2)(cid:3)(cid:8)(cid:16)(cid:17)(cid:2)(cid:3)(cid:27)(cid:8)
Quantum waves exhibit interference pat-
terns. It might be worthwhile to attempt
to characterize non-quantum phase tran-
sitions as wave phenomena and to look
for interference patterns there.
4.8 The constructionist hypothe-
sis revisited
(cid:1)(cid:2)(cid:3)(cid:4)(cid:5) (cid:3)(cid:2)(cid:5)(cid:4)(cid:6)(cid:7)(cid:3)(cid:2)(cid:8)(cid:9)(cid:5)(cid:10)(cid:11)(cid:7)(cid:12)(cid:5)(cid:12)(cid:13)(cid:14)(cid:12)(cid:4)(cid:15)(cid:3)(cid:7)(cid:2)(cid:16)(cid:5)(cid:17)(cid:18)(cid:19)(cid:5)(cid:14)(cid:4)(cid:5)
(cid:20) (cid:8)(cid:4)(cid:12)(cid:4)(cid:16)(cid:3)(cid:4)(cid:6)(cid:5)(cid:18)(cid:12)(cid:5)(cid:18)(cid:5)(cid:12)(cid:7)(cid:6)(cid:4)(cid:14)(cid:18)(cid:8)(cid:9)(cid:5)
Earlier we noted Anderson’ s rejection of
what he called the constructionist hy-
pothesis,
/ (cid:9)(cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:14)(cid:8) (cid:11)(cid:13)(cid:17)(cid:20)(cid:17)(cid:16)(cid:30)(cid:8) (cid:16)(cid:15)(cid:8) (cid:4)(cid:3)% (cid:18) (cid:7)(cid:3)(cid:8) (cid:3)& (cid:3)(cid:4)(cid:30)(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:8) (cid:16)(cid:15)(cid:8)
(cid:12)(cid:17)(cid:2)(cid:19) (cid:20)(cid:3)(cid:8) (cid:22)(cid:18) (cid:6)% (cid:11)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:20)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8) (cid:28)
(cid:8) (cid:17)(cid:2)(cid:19) (cid:20)(cid:17)(cid:3)(cid:12)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)
Emergence Explained
25/67
DRAFT
2/12/2006
(cid:11)(cid:13)(cid:17)(cid:20)(cid:17)(cid:16)(cid:30)(cid:8) (cid:16)(cid:15)(cid:8) (cid:12)(cid:16)(cid:11)(cid:4)(cid:16)(cid:8) (cid:22)(cid:4)(cid:15)(cid:2)(cid:8) (cid:16)(cid:10)(cid:15)(cid:12)(cid:3)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:4)(cid:3)(cid:7)(cid:15)(cid:6)(cid:21)
(cid:12)(cid:16)(cid:4)(cid:18) (cid:7)(cid:16)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:18) (cid:6)(cid:17)& (cid:3)(cid:4)(cid:12)(cid:3)0 (cid:8)
Our considerations in this section clarify
this rejection.
The constructionist agenda is simply not
within the realm of science. Even were it
theoretically possible, it simply is not
one of the tasks that science sets for it-
self to reconstruct the universe from the
principles of physics and some initial
state of the universe.27 Science simply
does not attempt to imagine all the pos-
sible configurations of matter and energy
that are consistent with the constraints
imposed by the fundamental laws of
physics, i.e., all the possible abstract de-
signs that nature may possibly imple-
ment, and predict which will come to
pass.
It is worth considering what the role of
experiments would be were it possible to
start with the fundamental laws of phys-
ics and reconstruct the universe. We can
think of two.
1. To help refine our understanding of
the fundamental laws of physics.
2. If the laws of physics require that
one of a number of possible situa-
tions be the case, to decide which of
them actually is the case.
Most experiments are done for neither of
these reasons. Most experiments are
done to help establish whether some
higher level regularity actually holds.
Recently in the news were the results of
the impact of a projectile with the comet
Tempel 1. As reported by Cowan
[Cowan], the experiment
27 Although see our comments about theories of the
origin of the universe and biological evolution in
the final section.
(cid:8)(cid:11)(cid:4)(cid:3)(cid:8)
(cid:4)(cid:3)& (cid:3)(cid:11)(cid:20)(cid:3)% (cid:8)(cid:12)(cid:3)& (cid:3)(cid:4)(cid:11)(cid:20)(cid:8)(cid:12)(cid:18) (cid:4)(cid:19) (cid:4)(cid:17)(cid:12)(cid:3)(cid:12)(cid:27)(cid:8)) (cid:10)(cid:3)(cid:8)% (cid:11)(cid:16)(cid:11)(cid:8)(cid:28)
(cid:11)(cid:16)(cid:8) (cid:15)% % (cid:12)(cid:8) ’ (cid:17)(cid:16)(cid:10)(cid:8) (cid:11)(cid:8) (cid:20)(cid:3)(cid:11)% (cid:17)(cid:6)(cid:5)(cid:8)(cid:2)(cid:15)% (cid:3)(cid:20)(cid:8) (cid:22)(cid:15)(cid:4)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:16)(cid:4)(cid:18) (cid:7)(cid:21)
(cid:16)(cid:18) (cid:4)(cid:3)(cid:8) (cid:15)(cid:22)(cid:8) (cid:7)(cid:15)(cid:2)(cid:3)(cid:16)(cid:12)(cid:8) (cid:7)(cid:11)(cid:20)(cid:20)(cid:3)% (cid:8) (cid:16)(cid:10)(cid:3)(cid:8) % (cid:17)(cid:4)(cid:16)(cid:30)(cid:21)(cid:12)(cid:6)(cid:15)’ (cid:13)(cid:11)(cid:20)(cid:20)(cid:8)
(cid:2)(cid:15)% (cid:3)(cid:20)(cid:27)(cid:8)(cid:28)
(cid:8)
(cid:9)) (cid:14)(cid:10)(cid:3)(cid:8) % (cid:11)(cid:16)(cid:11)(cid:8) (cid:22)(cid:4)(cid:15)(cid:2)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) : (cid:3)(cid:3)(cid:19) (cid:8) +(cid:2)(cid:19) (cid:11)(cid:7)(cid:16)(cid:8) (cid:2)(cid:17)(cid:12)(cid:12)(cid:17)(cid:15)(cid:6)(cid:8)
(cid:17)(cid:6)% (cid:17)(cid:7)(cid:11)(cid:16)(cid:3)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:11)(cid:20)(cid:16)(cid:10)(cid:15)(cid:18) (cid:5)(cid:10)(cid:8) ) (cid:3)(cid:2)(cid:19) (cid:3)(cid:20)(cid:8) (cid:23) (cid:8) (cid:7)(cid:15)(cid:6)(cid:16)(cid:11)(cid:17)(cid:6)(cid:12)(cid:8)
(cid:12)(cid:15)(cid:2)(cid:3)(cid:8) (cid:17)(cid:7)(cid:3)(cid:12)-(cid:8) (cid:17)(cid:16)(cid:12)(cid:8) (cid:19) (cid:4)(cid:17)(cid:2)(cid:11)(cid:4)(cid:30)(cid:8) (cid:7)(cid:15)(cid:6)(cid:12)(cid:16)(cid:17)(cid:16)(cid:18) (cid:3)(cid:6)(cid:16)(cid:8) (cid:2)(cid:11)(cid:30)(cid:8) (cid:13)(cid:3)(cid:8)
% (cid:18) (cid:12)(cid:16)(cid:8)(cid:19) (cid:11)(cid:4)(cid:16)(cid:17)(cid:7)(cid:20)(cid:3)(cid:12)(cid:8)(cid:22)(cid:17)(cid:6)(cid:3)(cid:4)(cid:8)(cid:16)(cid:10)(cid:11)(cid:6)(cid:8)(cid:16)(cid:11)(cid:20)(cid:7)(cid:18) (cid:2)(cid:8)(cid:19) (cid:15)’ (cid:3)(cid:4)(cid:27)(cid:8)(cid:28)
(cid:8)
(cid:29) (cid:10)(cid:11)(cid:16)=(cid:12)(cid:8) (cid:2)(cid:15)(cid:4)(cid:3)-(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:7)(cid:15)(cid:2)(cid:3)(cid:16)(cid:8) (cid:17)(cid:12)(cid:6)=(cid:16)(cid:8) (cid:11)(cid:8) (cid:2)(cid:3)(cid:4)(cid:3)(cid:8)
(cid:10)(cid:15)% (cid:5)(cid:3)(cid:19) (cid:15)% (cid:5)(cid:3)(cid:8) (cid:15)(cid:22)(cid:8)
% (cid:17)(cid:22)(cid:22)(cid:3)(cid:4)(cid:3)(cid:6)(cid:16)(cid:8) (cid:2)(cid:11)(cid:16)(cid:3)(cid:4)(cid:17)(cid:11)(cid:20)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8)
(cid:12)(cid:16)(cid:4)(cid:18) (cid:7)(cid:16)(cid:18) (cid:4)(cid:3)(cid:12)(cid:27)(cid:8) > ) (cid:10)(cid:3)(cid:8) % (cid:11)(cid:2)(cid:6)(cid:8) (cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:8) (cid:17)(cid:12)(cid:8) (cid:20)(cid:11)(cid:30)(cid:3)(cid:4)(cid:3)% (cid:8) (cid:20)(cid:17), (cid:3)(cid:8)
(cid:11)(cid:8)(cid:22)(cid:4)(cid:15)(cid:31) (cid:3)(cid:6)(cid:8)(cid:15)(cid:6)(cid:17)(cid:15)(cid:6)-> (cid:8)(cid:12)(cid:11)(cid:30)(cid:12)(cid:8): (cid:3)(cid:3)(cid:19) (cid:8)+(cid:2)(cid:19) (cid:11)(cid:7)(cid:16)(cid:8)(cid:12)(cid:7)(cid:17)(cid:3)(cid:6)(cid:16)(cid:17)(cid:12)(cid:16)(cid:8)
< (cid:15)(cid:12)(cid:3)(cid:19) (cid:10)(cid:8)C (cid:3)& (cid:3)(cid:4), (cid:11)(cid:8)(cid:15)(cid:22)(cid:8)5 (cid:15)(cid:4)(cid:6)(cid:3)(cid:20)(cid:20)(cid:8)A (cid:6)(cid:17)& (cid:3)(cid:4)(cid:12)(cid:17)(cid:16)(cid:30)(cid:27)(cid:8)(cid:28)
(cid:8)
: (cid:3)(cid:3)(cid:19) (cid:8) +(cid:2)(cid:19) (cid:11)(cid:7)(cid:16)=(cid:12)(cid:8) (cid:4)(cid:3)& (cid:3)(cid:20)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)(cid:8) > (cid:11)(cid:4)(cid:3)(cid:8) (cid:5)(cid:15)(cid:17)(cid:6)(cid:5)(cid:8) (cid:16)(cid:15)(cid:8)
(cid:7)(cid:10)(cid:11)(cid:6)(cid:5)(cid:3)(cid:8) (cid:20)(cid:15)(cid:16)(cid:8) (cid:15)(cid:22)(cid:8) (cid:15)(cid:18) (cid:4)(cid:8) (cid:17)% (cid:3)(cid:11)(cid:12)(cid:8) (cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8) (cid:7)(cid:15)(cid:2)(cid:3)(cid:16)(cid:12)-> (cid:8)
(cid:19) (cid:4)(cid:3)% (cid:17)(cid:7)(cid:16)(cid:12)(cid:8) (cid:9)< (cid:11)(cid:30)(cid:14)(cid:8) E (cid:3)(cid:20)(cid:15)(cid:12)(cid:10)(cid:8) (cid:9)(cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) A (cid:6)(cid:17)& (cid:3)(cid:4)(cid:12)(cid:17)(cid:16)(cid:30)(cid:8) (cid:15)(cid:22)(cid:8)
6(cid:4)(cid:17)(cid:31) (cid:15)(cid:6)(cid:11)(cid:8)(cid:17)(cid:6)(cid:8)) (cid:18) (cid:7)(cid:12)(cid:15)(cid:6)(cid:14)(cid:27)(cid:8)(cid:8)
These are not the words of scientists in
the process of deriving results about
comets from the fundamental laws of
physics. These are scientists building
higher level models.
It is just as wrong for Weinberg to deni-
grate the notion that there can be inde-
pendent principles of chemistry or psy-
chology as it would be for him to claim
that the principles of Mathematics and
Computer Science are all reducible to
and derivable from physics.
Most sciences are like Mathematics and
Computer Science. They really do stand
on their own— which is why (as Ander-
son says) the constructionist hypothesis
is wrong and why (as Fodor says) the
special sciences are autonomous..
5 Entities
So far, we have discussed what one
might characterize as emergence in the
large. There is also emergence on a
smaller and more local scale. That sort
of emergence is related to what we intui-
tively think of as entities. This section
Emergence Explained
26/67
DRAFT
2/12/2006
discusses entities and how they relate to
emergence.
We think in terms of entities, i.e., things
or objects. It seems like the most natural
thing in the world. Yet the question of
how one might characterize what should
and should not be considered an entity
has long been a subject of philosophical
study. A brief review of the recent litera-
ture (for example, [Boyd], [Laylock],
[Miller], [Rosen], [Varzi Fall ‘ 04]) sug-
gests that no consensus about how to
understand the notion of “ an entity” has
yet been reached.
One might adopt a very general position.
For example, Laylock quotes Lowe
[Lowe] as follows.
#) (cid:10)(cid:17)(cid:6)(cid:5)$-(cid:8)
(cid:17)(cid:6)(cid:8)
(cid:17)(cid:16)(cid:12)(cid:8) (cid:2)(cid:15)(cid:12)(cid:16)(cid:8) (cid:5)(cid:3)(cid:6)(cid:3)(cid:4)(cid:11)(cid:20)(cid:8) (cid:12)(cid:3)(cid:6)(cid:12)(cid:3)-(cid:8)
(cid:17)(cid:12)(cid:8)
(cid:17)(cid:6)(cid:16)(cid:3)(cid:4)(cid:7)(cid:10)(cid:11)(cid:6)(cid:5)(cid:3)(cid:11)(cid:13)(cid:20)(cid:3)(cid:8) ’ (cid:17)(cid:16)(cid:10)(cid:8) #(cid:3)(cid:6)(cid:16)(cid:17)(cid:16)(cid:30)$(cid:8) (cid:15)(cid:4)(cid:8) #(cid:13)(cid:3)(cid:17)(cid:6)(cid:5)$(cid:8) (cid:11)(cid:6)% (cid:8)
(cid:17)(cid:12)(cid:8) (cid:11)(cid:19) (cid:19) (cid:20)(cid:17)(cid:7)(cid:11)(cid:13)(cid:20)(cid:3)(cid:8) (cid:16)(cid:15)(cid:8) (cid:11)(cid:6)(cid:30)(cid:8) (cid:17)(cid:16)(cid:3)(cid:2)(cid:8) ’ (cid:10)(cid:15)(cid:12)(cid:3)(cid:8) (cid:3)*(cid:17)(cid:12)(cid:16)(cid:3)(cid:6)(cid:7)(cid:3)(cid:8)
(cid:17)(cid:12)(cid:8) (cid:11)(cid:7), (cid:6)(cid:15)’ (cid:20)(cid:3)% (cid:5)(cid:3)% (cid:8) (cid:13)(cid:30)(cid:8) (cid:11)(cid:8) (cid:12)(cid:30)(cid:12)(cid:16)(cid:3)(cid:2)(cid:8) (cid:15)(cid:22)(cid:8) (cid:15)(cid:6)(cid:16)(cid:15)(cid:20)(cid:15)(cid:5)(cid:30)-(cid:8)
’ (cid:10)(cid:3)(cid:16)(cid:10)(cid:3)(cid:4)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:17)(cid:16)(cid:3)(cid:2)(cid:8) (cid:13)(cid:3)(cid:8) (cid:19) (cid:11)(cid:4)(cid:16)(cid:17)(cid:7)(cid:18) (cid:20)(cid:11)(cid:4)-(cid:8) (cid:18) (cid:6)(cid:17)& (cid:3)(cid:4)(cid:12)(cid:11)(cid:20)-(cid:8)
(cid:11)(cid:13)(cid:12)(cid:16)(cid:4)(cid:11)(cid:7)(cid:16)-(cid:8) (cid:15)(cid:4)(cid:8) (cid:7)(cid:15)(cid:6)(cid:7)(cid:4)(cid:3)(cid:16)(cid:3)(cid:27)(cid:8) +(cid:6)(cid:8) (cid:16)(cid:10)(cid:17)(cid:12)(cid:8) (cid:12)(cid:3)(cid:6)(cid:12)(cid:3)-(cid:8) (cid:6)(cid:15)(cid:16)(cid:8)
(cid:15)(cid:6)(cid:20)(cid:30)(cid:8) (cid:2)(cid:11)(cid:16)(cid:3)(cid:4)(cid:17)(cid:11)(cid:20)(cid:8) (cid:13)(cid:15)% (cid:17)(cid:3)(cid:12)(cid:8) (cid:13)(cid:18) (cid:16)(cid:8) (cid:11)(cid:20)(cid:12)(cid:15)(cid:8) (cid:19) (cid:4)(cid:15)(cid:19) (cid:3)(cid:4)(cid:16)(cid:17)(cid:3)(cid:12)-(cid:8)
(cid:4)(cid:3)(cid:20)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)-(cid:8) (cid:3)& (cid:3)(cid:6)(cid:16)(cid:12)-(cid:8) (cid:6)(cid:18) (cid:2)(cid:13)(cid:3)(cid:4)(cid:12)-(cid:8) (cid:12)(cid:3)(cid:16)(cid:12)-(cid:8) (cid:11)(cid:6)% (cid:8)
(cid:19) (cid:4)(cid:15)(cid:19) (cid:15)(cid:12)(cid:17)(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)(cid:8)(cid:11)(cid:4)(cid:3)2(cid:17)(cid:22)(cid:8)(cid:16)(cid:10)(cid:3)(cid:30)(cid:8)(cid:11)(cid:4)(cid:3)(cid:8)(cid:11)(cid:7), (cid:6)(cid:15)’ (cid:20)(cid:3)% (cid:5)(cid:3)% (cid:8)
(cid:11)(cid:12)(cid:8)(cid:3)*(cid:17)(cid:12)(cid:16)(cid:17)(cid:6)(cid:5)2(cid:16)(cid:15)(cid:8)(cid:13)(cid:3)(cid:8)(cid:11)(cid:7)(cid:7)(cid:15)(cid:18) (cid:6)(cid:16)(cid:3)% (cid:8)#(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:12)$(cid:27)(cid:8)
For our purposes, this is too broad. In
this paper we want to exclude properties,
relations, events, numbers, sets and
propositions from our notion of entity.
We don’ t want to think of, say, the
American Civil War or happiness as an
entity in the same way that we think of
an atom is an entity.
On the other hand, we don’ t want to
limit ourselves to strictly material ob-
jects. We want to include countries,
teams, corporations, and families, for
example, as well as what may seem like
quasi-physical entities such as people
and hurricanes, whose physical makeup
undergoes continual change.
For our purposes, entities, by fiat, will
always have some material aspect. That
is, an entity will at any time consist of
physical elements arranged in a particu-
lar way. With this decision we are ex-
cluding from our notion of entity strictly
mental constructs such as sets, numbers,
concepts, propositions, relationships, de-
signs, abstractions, etc.
If the preceding does not formally ex-
clude instants, events, and durations, we
will explicitly exclude them too. Entities
for us will be required to persist in time,
but they will not be aspects of time, i.e.,
instants or durations, or events, whatever
an event is.
An entity for us will be either atomic
(not in the sense of being a chemical
element but in the more generic sense of
having no constituents —
if indeed there
are atomic physical elements in nature),
or, if an entity has constituents, it will be
an epiphenomenon of its constituents.
Thus for us non-atomic entities will rep-
resent one of the most common forms of
emergence.
Our purpose in this section is not to set-
tle the grand philosophical question of
what one should mean by the terms
thing, object or entity but to sketch out
what it means to be an entity in our
sense. Of course we hope that the
framework we develop will offer a use-
ful way of thinking about some of the
uses to which we commonly put the
terms thing, object, and entity.
5.1 Entities, entropy, designs,
and functionality
The standard model of physics includes
fundamental particles such as electrons,
photons, quarks, etc. These are entities
which have no constituents. Beyond
these, one has atomic nuclei, atoms, and
Emergence Explained
27/67
DRAFT
2/12/2006
molecules, all of which we want to in-
clude in our notion of entity.
For functionalism, entities are every-
where: mice and cans are good exam-
ples. The higher level sciences speak of
all sorts of entities, including biological
entities (e.g., you and me) and social,
political, and economic entities, such as
families, states, and corporations.
We propose to characterize an entity as
either atomic or as any demarcatable re-
gion that exhibits a persistent and self-
perpetuating reduced level of entropy.
Since we are not prepared to define the
term demarcatable region, perhaps de-
fining entity in terms of a notion as
loosely defined as demarcatable region
doesn’ t get one very far.28 But the notion
of an entity always seems to imply a
boundary that distinguishes the entity
from its surroundings. Entities in our
sense always have an “ inside.”
We discuss two kinds of entities: entities
at an energy equilibrium and entities that
are far from equilibrium.29
It is important to note that since non-
atomic entities have a reduced level of
entropy, they always have an internal
structure, i.e., a design. Furthermore, the
design of an entity often allows it to as-
sume one or more states. A good exam-
ple is the design of an atom: a nucleus
along with associated electrons in vari-
ous orbitals. Among the states of an
atom are those differentiated by the dif-
fering energy levels of its electrons.
It seems pretty clear that we (and other
animals) have evolved the ability to per-
28 As Varzi [Varzi Spring ‘ 04] points out, the notion
of a boundary is itself quite difficult to pin down.
Some boundaries, Mt. Everest ’ s, for example, are
quite vague.
29 We first proposed this in [Abbott].
ceive entities in this sense. Our intuitive
sense of entity seems to map fairly well
onto the notion of a persistent demarcat-
able region that displays some special
order that distinguishes it from its envi-
ronment, i.e., an area that has an internal
design.
We use the term design deliberately. En-
tities
in
implement abstract designs
much the same way as abstract designs
such as Newtonian mechanics are im-
plemented on a larger scale. Because an
entity implements a particular design, it
exhibits the functionality that its design
produces. One of the tasks of science,
then, is to decide for any entity (or cate-
gory of entities), what design it embod-
ies and what the implications of that de-
sign are for the behavior of that entity
(or those entities).
When nature implements an abstract de-
sign such as solid matter or Newtonian
mechanics it is the functionalities that
come along with that design— what the
design implies about how matter that
implements it behaves —
that make us in-
terested in it. These larger scale abstract
designs are typically embodied by sub-
stances or by arbitrary collections of
things. In contrast, the designs that enti-
ties implement produce a particular kind
of functionality in a constrained and
bounded region.
5.2 Entities at an energy equilib-
rium
The entities of physics and chemistry are
at an energy equilibrium. A distinguish-
ing feature of these entities is that the
mass of any one of them is strictly
smaller than the sum of the masses of its
components. This may be seen most
clearly in nuclear fission and fusion, in
which one starts and ends with the same
number of atomic components, i.e., elec-
trons, protons, and neutrons — which
Emergence Explained
28/67
DRAFT
2/12/2006
raises the obvious question: which mass
was converted to energy?
The answer has to do with the strong nu-
clear force, which implements what is
called the “ binding energy” of nucleons
within a nucleus. Without going into de-
tails, the bottom line is that the mass of,
say, a helium nucleus (also known as an
alpha particle, two protons and two neu-
trons), which is one of the products of
hydrogen fusion, is less than the sum of
the masses of the protons and neutrons
that make up an alpha particle when not
bound together as an alpha particle.30
The same entity-mass relationship holds
for all physical and chemical entities.
The mass of an atom or molecule is
(negligibly) less than the sum of the
masses of its components taken sepa-
rately. The mass of the solar system is
(negligibly) less than the mass of the sun
and the planets when taken separately.
This fact implies that the entropy of
these entities is lower than the entropy of
the components taken separately. In
other words, an entity at an energy equi-
librium is distinguishable by the fact that
it has lower mass and lower entropy than
its components taken separately.
These entities are trivially self-perpetuat-
ing in that they are in what is often
called an energy well and require energy
to pull their components apart. This
gives us a nice metric of entityness for
at-equilibrium entities: the amount of
energy required to pull it apart.
30 It turns out that the atomic nucleus with the least
mass per nucleon is iron. Energy from fusion is
possible for elements lighter than iron; energy
from fission is possible for elements heavier than
iron. (See [Nave] for a discussion of these mat-
ters.)
5.3 Entities and emergence are
fundamental
The mechanisms (gravity, the strong nu-
clear force, and the electromagnetic
force) that expel entropy from at-equilib-
rium entities and that hold these entities
together are the fundamental forces of
nature.
One can say that these mechanisms in
some sense run for free. To the extent
that we understand how they work at all,
we attribute their operation to virtual
particles that pop into and out of exis-
tence and that do the work of the force—
with no extra effort expended anywhere
else.
There really is a free lunch. Atomic nu-
clei form, atoms form, solar systems and
galaxies form —
all without depleting any
energy reservoirs. We are so used to this
fact that we hardly notice it. But if one
stands back and observes
that at-
equilibrium entities exemplify emer-
gence at its most basic—
an atom is
emergent from, it is an epiphenomenon
of, and it supervenes over its compo-
nents — we may conclude that spontane-
ous emergence is fundamental to how
nature works.
Even so, one might suppose that beyond
combining
in
these basic ways (as
atomic nuclei, atoms, and astronomical
aggregations held together by gravity),
at-equilibrium entities are not very inter-
esting. Standing back again makes it
clear that this is not the case. Given what
we have learned during the past half cen-
tury (and what we still don ’ t know)—
especially about condensed matter phys-
ics and including, as we said earlier, the
startling fact that the same matter is ca-
pable of implementing multiple abstrac-
tions with radically different proper-
ties —
at-equilibrium entities are far from
boring.
Emergence Explained
29/67
DRAFT
2/12/2006
5.4 Dissipative structures
In [Prigogine] (and elsewhere) Prigogine
discussed what he called a dissipative
structure. We see dissipative structures
as the essential stepping stone from at-
equilibrium entities to autonomous enti-
ties.
Intuitively, a dissipative structure typi-
cally manifests when energy is pumped
into a bounded region. Dissipative struc-
tures typically involve structured activi-
ties internal to the region. A standard ex-
ample consists of the Bénard convection
cycles that form in a liquid when one
surface is heated and the opposite sur-
face is kept cool. (See Figure 4.)
A number of interesting phenomena may
be understood as dissipative structures.
Consider the distribution of water over
the earth. Water is transported from
place to place via processes that include
evaporation, atmospheric weather sys-
tem movements, precipitation, ground-
water flows, ocean current flows, etc.
Taken as a global system, these cycles
may be understood as a dissipative struc-
ture that is driven primarily by solar en-
ergy, which is pumped into the earth ’ s
atmosphere and surface structures. All of
this is played out against a static frame-
work defined and held in place by the
earth’ s surface and its gravitational field.
We note that our definition of a dissipa-
tive structure is quite broad. It includes
virtually any energy-consuming device
that operates according to some design.
Consider a digital watch. It converts an
inflow of energy into an ongoing series
of structured internal activities. Does a
digital watch define a dissipative struc-
ture? One may argue that the design of a
digital watch limits the ways in which it
can respond to an energy inflow. There-
fore the structured activity that arises as
energy is pumped into it should not be
characterized as a dissipative structure.
But any bounded region has only a lim-
ited number of ways in which it can re-
spond to an inflow of energy. We sug-
gest that it would be difficult if not im-
possible to formalize a principled dis-
tinction between the Bénard convection
cycles that arise in a liquid when energy
is pumped into it and the structured ac-
tivities within a digital watch.31 The pri-
mary difference seems to be that a digi-
tal watch has a much more constrained
static structure and can respond in far
fewer ways.
Recall that we previously characterized
Newtonian mechanics and
the solid
phase of matter as abstractions that mat-
ter implements under various conditions.
We can do the same thing for dissipative
structures and say that a dissipative
structure appears within a bounded re-
gion when the materials within that re-
gion implement an energy-driven ab-
stract design.
An apparent difference between the ab-
stract designs that dissipative structures
implement and the abstract designs dis-
cussed earlier is that the abstract designs
of dissipative structures seem to appear
unbidden— we don ’ t expect
them —
whereas the abstract designs discussed
earlier are commonplace. The issue for
the more commonplace abstract designs
is how to conceptualize them, not why
they appeared at all, whereas the abstract
design that appear as dissipative struc-
tures seem to demand an answer to the
question: why did they appear at all? In
fact, both kinds of abstract design are
31 One of the other common examples of a dissipa-
tive structure is the Belousov-Zhabotinsky (BZ)
reaction, which in some ways is a chemical watch.
Emergence Explained
30/67
DRAFT
2/12/2006
part of nature. The difference is that
some are familiar; others aren’ t.32
If we understand a dissipative structure
to be the implementation of an energy-
driven abstract design, the question for
any dissipative structure becomes: what
abstract design does it implement? In
other words, how does it work— which
is the same question one must ask about
any abstract design.
Like most abstract designs, those associ-
ated with dissipative structures generally
exist only within limited energy ranges.
Thus phase transitions may be expected
as materials transform themselves be-
tween configurations in which they are
and are not implanting the abstract de-
sign of a particular dissipative structure.
In this section we have referred, some-
what awkwardly, to bounded regions
within which dissipative structures form.
We have refrained from calling these
bounded regions entities. This may be
pickiness on our part, but our notion is
that an entity perpetuates itself. As de-
fined, bounded regions of materials that
are capable of implementing dissipative
structure abstract designs need not have
the capacity to perpetuate themselves.
Their boundaries may be imposed artifi-
cially. We shall have more to say about
this in the section on natural vs. artificial
autonomous entities.
32 Reliance on a surprise factor as characteristic of
emergence is, in our opinion, a common error. One
often hears that emergent phenomena appear un-
bidden, that one’ s surprise is fundamental to
whether something should be considered emer-
gent. On the contrary, we claim that an observer ’ s
surprise of lack of surprise should have nothing to
do with how we should understand a phenomenon
of nature.
5.5
Integrating dissipative struc-
tures and at-equilibrium enti-
ties.
A dissipative structure is a physical
manifestation of a region of energy sta-
bility in an environment in which energy
is flowing at a relatively constant rate.
An at-equilibrium entity is similarly a
physical manifestation of a region of
stability—
but
in an environment
in
which there is no energy flow. We
would welcome a formal integration of
the two in which at-equilibrium entities
are understood as dissipative structures
in an environment in which the rate of
energy flow is zero. Perhaps another
way of putting this would be to charac-
terize the energy wells that exist in envi-
ronments that include energy flows.
5.6 Autonomous entities
The notion of an autonomous entity
seems central to how we look at the
world.
• For millennia we have found it con-
venient to partition the world into two
realms: the animate and the inanimate.
The inanimate world is ruled by exter-
nal forces; the animate world is capa-
ble of autonomous action. Recall that
this is why Brownian motion posed
such a problem: how can inanimate
particles look so much like they are
moving autonomously?
• For the past half-millennium western
civilization (and more recently civili-
zation world-wide) has pursued, with
significant success, the dream of creat-
ing autonomous sources of action. We
have built machines about which it can
be said that in varying degrees they act
on their own. We do not yet confuse
our machines with biological life, and
we have not yet managed to construct
biological life “ from scratch.” But the
differences between human artifacts
Emergence Explained
31/67
DRAFT
2/12/2006
and natural biological life are becom-
ing more and more subtle—
and they
are likely to disappear within the life-
times of many of us.
• Most people will acknowledge that the
kinds of entities that the biological and
social sciences deal with seem some-
how different from those of physics
and chemistry. A major part of that
difference is the apparent ability of the
entities in those sciences to act on their
own, i.e., their autonomy.
So, what do we mean by autonomy?
Certainly, we no longer believe in any-
thing like vitalism, i.e., that there is such
a thing as a “ life force” the possession of
which differentiates the animate from
the inanimate. But when we speak of
autonomous entities, have we done much
more than substitute the word autono-
mous for other words? Do we have a
serviceable definition of what it means
to be autonomous?
In non-political contexts, the term auton-
omous is generally taken to mean some-
thing like self-directed or not controlled
by outside forces.33 But definitions of
this sort don’ t help much. Perhaps self-
directed is what we mean by autono-
mous. But what do we mean by self-
directed?
Furthermore any entity (in our sense of
an entity as having some material aspect)
is subject
to outside,
i.e., physical,
forces. Nothing is free from the laws of
physics. So it may not make any sense to
demand that to be autonomous an entity
must not be controlled by outside forces.
The intuition behind self-directed and
the connection to outside forces may
33 See, for example, the American Heritage® Dic-
tionary definition. URL as of 9/15/2005:
http://www.bartleby.com/61/86/A0538600.html.
give us a clue, however. Perhaps one can
require that an autonomous entity con-
trol —
at least to some extent and in what
may be considered a self-directed way,
although without implying willfulness —
how it is affected by outside forces.
Putting these ideas together, we suggest
that a useful way to think about auton-
omy may be that an entity is autonomous
to the extent that it shapes the way it is
affected by outside forces.
But this is pretty much how we have de-
fined a dissipative structure. A dissipa-
tive structure results from the operation
of an energy-driven abstract design. In
other words, a dissipative structure re-
sults when an energy-driven abstract de-
sign shapes the way outside forces oper-
ate within a bounded region.
Because this seems to be such a nice fit
with our intuition of what it means for an
entity to be autonomous, we will define
an autonomous entity as an entity that is
implementing the abstract design of a
dissipative structure.34
In other words, we define an autono-
mous entity as a self-perpetuating region
of reduced entropy that is implementing
a dissipative structure’ s abstract design.
By definition, autonomous entities con-
sume energy and are far from equilib-
rium. We suggest that most if not all of
the entities of the higher level sciences
satisfy our definition of an autonomous
entity.
Note that most biological, social, and
economic autonomous entities are even
more autonomous than our definition
suggests. Most of these entities acquire
energy in some “ frozen ” form such as
34 This intuitive fit may be one reason that the notion
of a dissipative structure generated as much enthu-
siasm as it has.
Emergence Explained
32/67
DRAFT
2/12/2006
food or money35 and convert it to energy
according to their internal designs. Thus
they do more than simply shape how
“ raw ” energy that they encounter affects
them. They are often able to save energy
and to chose in some sense when to use
it.
5.7 A naturally occurring autono-
mous entity that is neither
biological nor social
We suggest that a hurricane qualifies as
an autonomous entity. (See Figure 3.) In
simple
terms
(paraphrased
from
[NASA]), the internal design of a hurri-
cane involves a greater than normal
pressure differential between the ocean
surface and the upper atmosphere. That
pressure differential causes moist surface
air to rise. When the moisture-laden air
reaches the upper atmosphere, which is
cooler, it condenses, releasing heat. The
heat warms the air and reduces the pres-
sure, thereby maintaining the pressure
differential —
a marvelous design for a
self-perpetuating process.
In effect, a hurricane is a heat engine in
which condensation, which
replaces
combustion as the source of heat, occurs
in the upper atmosphere.36 Thus, al-
though physically very large, a hurricane
has a relatively simple design, which
causes it to consume energy and which
allows it to perpetuate itself as an area of
reduced entropy.
5.8 Natural and artificial autono-
mous entities
Most of our energy consuming machines
also qualify as autonomous entities. The
primary difference between human pro-
35 The maxim follow the money is really advising to
follow the energy.
36 A characterization of hurricanes as “ vertical heat
engines” may be found in Wikipedia. URL as of
9/1/2005: http://en.wikipedia.org/wiki/Hurricane
duced autonomous entities and naturally
occurring ones is that the naturally oc-
curring autonomous entities use at least
some of the energy they consume to per-
petuate themselves as entities. In con-
trast, human-produced autonomous enti-
ties are almost always at-equilibrium en-
tities through which energy flows. In
other words,
the nature of human-
produced autonomous entities is that
their persistence as entities tends to be
independent of their use of the energy
that flows through them. This tends not
to be the case with naturally occurring
autonomous entities.
One of the senses of the word natural is
to have properties characteristic of ele-
ments found in nature. We suggest that
the distinction between entities that rely
on an at-equilibrium frame and those
that more actively construct
their
framework is one of the central intuitive
differences between what we call artifi-
cial and what we call natural. A hurri-
cane would thus be considered a natu-
rally occurring autonomous entity which
is neither biological nor social.
As an example of a naturally occurring
at-equilibrium
entity
that becomes
autonomous, consider an atom that is be-
ing excited by a photon stream. Because
of its design it captures the energy of the
photons, which it releases at some later
time in what may be a slightly different
form. This is the basis of the laser.
and
5.9 Autonomous
entities
phase transitions
Many autonomous entities exhibit the
equivalent of phases —
and phase transi-
tions. Such phases differ from phases in
at-equilibrium entities in that they reflect
different ways in which the autonomous
entity makes use of the energy that is
flowing through it. Examples include
gaits (walking, running, etc.), heart beats
Emergence Explained
33/67
DRAFT
2/12/2006
(regular and fibrillation), and possibly
psychological conditions such as mania,
depression and psychosis.
The primary concern about global warm-
ing is not that the temperature will rise
by a degree or two—
although the melt-
ing of the ice caps resulting from that is
potentially destructive—
but the possibil-
ity that if the temperature warms suffi-
ciently, a phase transition will occur, and
the global climate structure, including
atmospheric and oceanic currents, will
change abruptly—
and possibly disas-
trously.
As we suggest later, the fact that paral-
lels exist between autonomous and at-
equilibrium entities leads to the sugges-
tion that one might be able to integrate
the two and see at-equilibrium entities as
one end of a continuum that includes
both at-equilibrium and autonomous en-
tities.
5.10 Autonomous entities and en-
ergy flows
Autonomous entities
require energy
flows for survival. But the kinds of en-
ergy flows available are limited. The
most familiar (at least here on earth) is
the flow of energy from the sun. Plants
exploit it. We are also familiar with arti-
ficial energy flows, as in the flow of
electricity to a device when the switch is
turned on. Other than these, what other
flows of energy support autonomous en-
tities?
Thermal vents in the ocean are one pos-
sibility. Yet the primary food producers
in thermal vents are bacteria that convert
chemicals from the vents to more use-
able forms of energy.37 It is not clear
what role, if any, is played by the flow
37 See, for example, Comm Tech Lab and University
of Delaware.
of thermal energy itself. It would be sig-
nificant if a life-form were found that
used thermal energy directly to power an
internal process in a way that paralleled
the way plants use energy from the sun.
It may be that some of the chemical re-
actions that occur in inhabitants of vent
ecologies depend on a high ambient
temperature. But that seems to be a dif-
ferent sort of dependency than using a
direct energy flow.
Most biological autonomous entities ac-
quire their energy in a packaged form,
e.g., as “ food ” of some sort rather than
as a direct energy flow. Once the energy
resource has been ingested, energy is ex-
tracted from it. This is even the case
with our hurricane example. The energy
of condensation is produced within the
hurricane after warm moist air is “ in-
gested. ”
This seems to be another distinction be-
tween naturally occurring and artificial
autonomous entities. No artificial enti-
ties procure their own energy resources.
Other than plants, all naturally occurring
autonomous entities do.
5.11 Theseus’s ship
The distinction between natural and arti-
ficial entities sheds some light on the
paradox of Theseus ’ s ship, a ship that
was maintained (repaired, repainted,
etc.) in a harbor for so long that all of its
original material had been replaced.
Does one say that it is “ the same ship”
from year to year?
We would like to distinguish between
two ways of looking at Theseus ’ s ship.
One way is to consider the material ship
as it exists at any one moment. By our
although it
definition, this is an entity—
is not an autonomous entity—
since it is
at an energy equilibrium. It is held to-
gether by a large number of relatively
Emergence Explained
34/67
DRAFT
2/12/2006
shallow energy wells. Entities of this
sort are particularly vulnerable to every-
day weathering and wear and tear. It
doesn’ t take much to push some of the
energy wells beyond their limits.
A second way to look at Theseus ’ s ship
is to include the maintenance process as
part of a larger autonomous ship entity.
The ship along with its maintenance
process is an entity because it is a self-
perpetuating region of reduced entropy.
It is a relatively simple example of a so-
cial autonomous entity. Both materials
and people cycle through it, but the
process perpetuates itself by using en-
ergy from the society in which it is em-
bedded.
So our answer to the question of whether
“ the same ship” is in the harbor from
year to year is “ No” if we are thinking
about the material ship and “ Yes ” if we
are thinking about the larger ship-plus-
maintenance entity.
By our definition, the larger ship-plus-
maintenance entity would be considered
natural rather than artificial because it as
a social process and is not at-equilib-
rium; it uses some of the energy it con-
sumes to perpetuate itself. We would
consider most social entities to be natu-
ral in this sense even though they are
constructed and maintained by people.
5.12 Autonomous entities may act
in the world
As we know, hurricanes can cause sig-
nificant damage. So far we haven ’ t
talked about how that might happen.
Since energy flows through autonomous
entities, part of that flowing through in-
volves flowing out. In other words,
autonomous entities may include as part
of their designs means for projecting
force into the world by directing outward
flows of energy.38
Furthermore, the internal design of most
autonomous entities enable them (a) to
store energy, (b) to move it about inter-
nally, and (c) to tap it as needed.
5.13 Autonomous entities tend not
to supervene over their static
components
As we said earlier, an at-equilibrium en-
tity consists of a fixed collection of
component elements over which it su-
pervenes. In contrast, autonomous enti-
ties for the most part tend not to consist
of a fixed collection of matter. Our hur-
ricane is a good example. A hurricane
may be relatively stable as a reduced en-
tropy region—
even though its bounda-
ries may be somewhat vague. But how-
ever its boundaries are defined, the ma-
terial within its boundaries tends to vary
from moment to moment as the hurri-
cane’ s winds move air and water about.
Similarly, most biological entities recy-
cle their physical components, and most
social entities (e.g., families) and eco-
nomic entities (e.g., corporations) re-
main intact as the people who fill vari-
ous roles cycle through them. Theseus ’ s
ship — when understood as including its
maintenance process as discussed above
is another example of an autonomous
entity that recycles its physical compo-
nents.
Because of this recycling property, most
autonomous entities don’ t supervene
38 This solves a problem that concerned Leibniz with
respect to monads: how do they interact. Leibniz’ s
answer was that they don ’ t. Our autonomous enti-
ties interact with each other and with the rest of the
world though energy flows over which they have
the ability to exert some control. Of course our
autonomous entities can exert that control because
they have internal designs; Leibniz’ s monads
didn ’ t.
Emergence Explained
35/67
—
DRAFT
2/12/2006
over any collection of matter that gives
us any intellectual leverage.
It is easiest to see this when we consider
gliders in the Game of Life, about which
this is true as well. In the Appendix we
show how to formalize the notion of a
Game of Life pattern. In simplest terms
we define what we call a live cell group
to be a connected group of live (i.e.,
“ on” ) cells. We define a pattern as a
connected sequence of live cell groups.
In general, such sequences may branch
or terminate, but the glider pattern is a
linear sequence of live cell groups. (See
the Appendix for the details, which
pretty much match one ’ s intuition.) A
glider is such a pattern.
One may define the state of a glider pat-
tern to be the particular configuration it
is in (See Figure 2 earlier for the four
possible configurations.) Alternatively,
one may also define the state of a glider
pattern in either of two ways: the con-
figuration (of the four) in which the pat-
tern exists or the configuration along
with the pattern’ s location on the grid.
To satisfy supervenience, for a glider
pattern to supervene over a set of Game
of Life cells requires that if the glider is
in different states then the grid cells
must also be in a different state.
Given either of our two definitions of
state, gliders (if undisturbed) do not su-
pervene over any finite set of grid cells.
Given any such finite set of cells, a
glider may assume multiple states when
beyond that set, thereby violating super-
venience.
The only sets of cells over which a glider
supervenes is a superset of (an infinite
subset of cells within) what one might
call the glider ’ s “ glide path,” the strip of
cells that a glider will traverse if undis-
turbed. The parenthetical qualification
allows for the possibility that one can
differentiate states without looking at the
entire glider pattern.
In other words, any set of cells over
which a glider supervenes must include a
potentially infinite subset of the cells
with which the glider comes in contact
over its lifetime. This may be super-
venience, but it is supervenience in a not
very useful way.
To connect this to autonomous entities,
imagine a glider pattern as fixed with the
grid moving underneath it, i.e., as if the
glider cycles grid cells through itself.
This
is quite similar
to how most
autonomous entities operate. These enti-
ties typically cycle matter through them-
selves. The same reasoning shows that
such autonomous entities don’ t super-
vene over any useful subset of matter
other than the collection of all mater
with which they may come in contact
during their lifetimes.
It appears that the concept of superven-
ience may not be as useful as one might
have hoped for thinking about epiphe-
nomena and emergence —
at least in the
case of autonomous entities.
5.14 Entities, objects, and agents
Computer Science has also developed a
distinction between entities that do and
do not act autonomously. Recall that our
definition of entity depended on distin-
guishing an entity from its environment,
i.e., it was a region of reduced entropy.
We may therefore refer to the “ inside” of
an entity and to whatever internal struc-
ture and state it may have. This also al-
lows us also to speak of the interface
(boundary) between an entity and its en-
vironment.
If an entity has an internal state, what, if
anything, may cause
to
that state
change? Are there outside influences
Emergence Explained
36/67
DRAFT
2/12/2006
that may cause an entity to change state?
If so, what mechanism enables those in-
fluences to act on the entity? Alterna-
tively, may an entity change state as a
result of purely internal activity?
In Computer Science two concepts have
emerged as fundamental to these issues:
objects and agents. There is a reasonable
consensus in Computer Science about
what we mean by an object, namely an
encapsulation of a mechanisms for as-
suming and changing states along with
means for acting on that encapsulated
mechanism.
There is far less agreement about the no-
tion of an agent. For our purposes, we
will construe an agent as simply as pos-
sible. An agent for us will be an object
(as defined above) that may act on its
own. In software terms, this means that
an agent is a software object that has an
internal thread.39
Given these definitions of object and
agent, we suggest that to a first very
rough approximation40 objects are the
software equivalent of at-equilibrium en-
tities and agents are the software equiva-
lent of autonomous entities.
5.15 Thermodynamic computing:
nihil ex nihilo
Note that when discussing software ob-
jects and agents, there is no concern with
entropy: the software system maintains
the integrity (and internal structure) of
objects and agents. Similarly, we did not
claim that gliders or Turing Machines
were entities in the Game of Life.
39 In adopting this definition, we are deliberately by-
passing issues of goals, beliefs, plans, etc., which
appear in some formulations of agent-based mod-
eling frameworks.
40 See the next section for a discussion of why this
approximation is indeed very rough.
The problem has to do with the way we
do Computer Science. In Computer Sci-
ence we assume that one can specify a
Turing Machine, a Finite State Automa-
ton, a Cellular Automaton, or a piece of
software, and it will do its thing—
for
free. Software runs for free. Turing ma-
chines run for free. Cellular Automata
run for free. Gliders run for free. Agents
in agent-based models run for free. Al-
though that may be a useful abstraction,
we should recognize that we are leaving
out something important. In the real
world one needs energy to drive proc-
esses. To run real software in the real
world requires a real computer, which
uses real energy. We suggest that a the-
ory of thermodynamic computation is
needed to integrate the notions of en-
ergy, entities, and computing.
How do we capture the notion of the
“ energy” that enables software to do its
“ symbolic work? ”
Is computational
complexity an equivalent concept?
Computational complexity is concerned
primarily with finding measures for how
intrinsically difficult particular kinds of
computations are. The focus seems dif-
ferent.
Performance analysis is somewhat closer
to what we are attempting to get at. But
performance analysis is typically satis-
fied with relatively gross results, not
with the fine details of how a computa-
tional energy budget is spent.
The problem seems to be that the com-
putational energy that software uses is
not visible to the software itself. Soft-
ware does not have to pay its energy bill;
the rest of nature does.
However this issue is resolved, for now a
thread seems to be a useful software ana-
log for the energy flow that powers a
dissipative structure. It also seems rea-
Emergence Explained
37/67
DRAFT
2/12/2006
sonable to use the term agent as syn-
onymous with autonomous entity.
With this in mind, though, we should
point out that the parallel between ob-
jects and agents on the one hand and at-
equilibrium and autonomous entities on
the other isn’ t perfect. An object in soft-
ware is not completely controlled by ex-
ternal forces. An object ’ s methods do
shape how energy (in the form of threads
that execute them) affects the object.
Objects differ from agents in that they
don’ t have what might be considered an
internal source of energy. Agents do. But
our analogy breaks down entirely if an
object is allowed to create a thread when
one of its methods is executed. (Most
multi-threaded programming languages
allow the arbitrary creation of threads.)
For an object to create a thread would be
equivalent to an entity in nature creating
an unlimited internal source of energy
for itself once it came in contact with
any external energy at all.
As we said, a real theory of thermody-
namic computing is needed.
5.16 Minimal autonomous entities
In [Kauffman] Kauffman asks what the
basic characteristics are of what he (also)
calls autonomous agents. He suggests
that the ability to perform a thermody-
namic (Carnot engine) work cycle is
fundamental.
In what may turn out to be the same an-
swer we suggest looking for the minimal
biological organism that perpetuates it-
self by consuming energy. Bacteria seem
to be too complex. Viruses41 and prions
41 Viruses are an interesting contrast to our lactose
example, however. In both cases, an at-equilibrium
element in the environment triggers a process in an
autonomous entity. In the case of lactose, the proc-
ess is advantageous to the entity; in the case of vi-
ruses, it is not advantageous to the entity.
don’ t consume energy.42 Is there any-
thing in between? We suggest that such
a minimal autonomous entity may help
us understand the yet-to-be-discovered
transition from the inanimate to the ani-
mate.
Since self-perpetuation does not imply
reproduction (as hurricanes illustrate),
simple self-perpetuating organisms may
not be able to reproduce. That means
that if they are to exist, it must be rela-
tively easy for them to come into being
directly from inorganic materials. Simi-
larly, simple self-perpetuating organisms
may not include any stable internal re-
like DNA— of their design (as
cord—
hurricanes
again
illustrate). One
wouldn ’ t expect to see evolution among
such organisms —
at least not evolution
that depends on modifications of such
design descriptions..
6 The evolution of complex-
ity
6.1 Stigmergy
Once one has autonomous entities (or
agents) that persist in their environment,
the ways in which complexity can de-
velop grows explosively.
to
Prior
agents, to get something new, one had to
build it as a layer on top of some exist-
ing substrate. As we have seen, nature
has found a number of amazing abstrac-
tions along with some often surprising
ways to implement them. Nonetheless,
this construction mechanism is relatively
ponderous. Layered hierarchies of ab-
stractions are powerful, but they are not
what one might characterize as light-
weight or responsive to change. Agents
change all that.
42 Hurricanes aren ’ t biological.
Emergence Explained
38/67
DRAFT
2/12/2006
Half a century ago, Pierre(cid:1)Paul Grasse
invented [Grasse] the term stigmergy to
help describe how social insect societies
function. The basic insight is that when
the behavior of an entity depends to at
least some extent on the state of its envi-
ronment, it is possible to modify that en-
tity’ s behavior by changing the state of
the environment. Grasse used the term
“ stigmergy” for this sort of indirect
communication and control. This sort of
interplay between agents and their envi-
ronment often produces epiphenomenal
effects that are useful to the agents. Of-
ten those effects may be understood in
terms of formal abstractions. Sometimes
it is easier to understand them less for-
mally.
Two of the most widely cited examples
of stigmergic interaction are ant foraging
and bird flocking. In ant foraging, ants
that have found a food source leave
pheromone markers that other ants use to
make their way to that food source. In
bird flocking, each bird determines how
it will move at least in part by noting the
positions and velocities of its neighbor-
ing birds.
The resulting epiphenomena are that
food is gathered and flocks form. Pre-
sumably these epiphenomena could be
formalized in terms of abstract effects
that obeyed a formal set of rules —
in the
same way that the rules for gliders and
Turing Machines can abstracted away
from their implementation by Game of
Life rules. But often the effort required
to generate such abstract theories doesn’ t
seem worth the effort —
as long as the re-
sults are what one wants.
Here are some additional examples of
stigmergy.
• When buyers and sellers interact in a
market, one gets market epiphe-
nomena. Economics attempts to for-
malize how those interactions may be
abstracted into theories.
• We often find that laws, rules, and
regulations have both intended and un-
intended consequences. In this case the
laws, rules, and regulations serve as
the environment within which agents
act. As the environment changes, so
does the behavior of the agents.
• Both sides of the evo-devo (evolution-
development) synthesis [Carroll] ex-
hibit stigmergic emergence. On the
“ evo ” side, species create environ-
mental effects for each other as do
sexes within species.
• The “ devo ” side is even more stigmer-
gic. Genes, the switches that control
gene expression, and the proteins that
genes produce when expressed all have
environmental effects on each other.
• Interestingly enough, the existence of
gene switches was discovered in the
investigation of another stigmergic
phenomenon. Certain bacteria generate
an enzyme to digest lactose, but they
do it only when lactose is present.
How do the bacteria “ know” when to
generate the enzyme?
It turns out to be simple. The gene for
the enzyme exists in the bacteria, but
its expression is normally blocked by a
protein that is attached to the DNA se-
quence just before the enzyme gene.
This
is called a gene expression
switch.
When lactose is in the environment, it
infuses into the body of the bacteria
and binds to the protein that blocks the
expression of the gene. This causes the
protein
to detach from
the DNA
thereby “ turning on” the gene and al-
lowing it to be expressed.
Emergence Explained
39/67
DRAFT
2/12/2006
The lactose enzyme switch is a lovely
illustration of stigmergic design. As
we described the mechanism above, it
seems that lactose itself turns on the
switch that causes the lactose-digesting
enzyme to be produced. If one were
thinking about the design of such a
system, one might imagine that the lac-
tose had been designed so that it would
bind to that switch. But of course, lac-
tose wasn’ t “ designed” to do that. It
existed prior to the switch. The bacte-
ria evolved a switch that lactose would
bind to. So the lactose must be under-
stood as being part of the environment
to which
the bacteria adapted by
evolving a switch to which lactose
would bind. How clever; how simple;
how stigmergic!
• Cellular automata operate stigmergi-
cally. Each cell serves as an environ-
ment for its neighbors. As we have
seen, epiphenomena may include glid-
ers and Turing Machines.
• Even the operation of the Turing Ma-
chine as an abstraction may be under-
stood stigmergically. The head of a
Turing Machine (the equivalent of an
autonomous agent) consults the tape,
which serves as its environment, to de-
termine how to act. By writing on the
tape, it leaves markers in its environ-
ment to which it may return — not
unlike the way foraging ants leave
pheromone markers in their environ-
ment. When the head returns to a
marker, that marker helps the head de-
termine how to act at that later time.
• In fact, one may understand all compu-
tations as being stigmergic with re-
spect to a computer’ s instruction exe-
cution cycle. Consider the following
familiar code fragment.
(cid:16)(cid:3)(cid:2)(cid:19) (cid:8) F (cid:8)*?(cid:8)
F (cid:8)(cid:30)?(cid:8)
*(cid:8)
(cid:30)(cid:8)
F (cid:8)(cid:16)(cid:3)(cid:2)(cid:19) ?(cid:8)
The epiphenomenal result is that * and
(cid:30) are exchanged. But this result is not a
consequence of any one statement. It is
an epiphenomenon of the three state-
ments being executed in sequence by a
computer’ s instruction execution cy-
cle.
Just as there in nothing in the rules of
the Game of Life about gliders, there is
nothing in a computer ’ s instruction
execution cycle about exchanging the
values of * and (cid:30)— or about any other
algorithm that software implements.
Those effects are all epiphenomenal.
• The instruction execution cycle itself is
epiphenomenal over the flow of elec-
trons through gates — which knows no
more about the instruction execution
cycle than the instruction execution
cycle knows about algorithms.
In all of the preceding examples it is
relatively easy to identify the agent(s),
the environment, and
the
resulting
epiphenomena.
6.2 Design and evolution
It is not surprising that designs appear in
nature. It is almost tautologous to say
that those things whose designs work in
the environments in which they find
themselves will persist in those envi-
ronments. This is a simpler (and more
accurate) way of saying that it is the
entities with designs that fit their
fit —
environment —
that survive.
6.3 The accretion of complexity
An entity that suits its environment per-
sists in that environment. But anything
that persists in an environment by that
very fact changes that environment for
Emergence Explained
40/67
DRAFT
2/12/2006
everything else. This phenomenon is
commonly referred to as an ever chang-
ing fitness landscape.
What has been less widely noted in the
complexity literature is that when some-
thing is added to an environment it may
enable something else to be added lat-
ter—
something that could not have ex-
isted in that environment prior to the ear-
lier addition.
This is an extension of notions from
ecology, biology, and the social sci-
ences. A term for this phenomenon from
the ecology literature, is succession.
(See, for example, [Trani].) Historically
succession has been taken to refer to a
fairly rigid sequence of communities of
species, generally leading to what is
called a climax or (less dramatically) a
steady state.
Our notion is closer to that of bricolage,
a notion that originated with the struc-
turalism movement of the early 20th cen-
tury [Wiener] and which is now used in
both biology and the social sciences.
Bricolage means the act or result of
tinkering,
improvising, or building
something out of what is at hand.
In genetics bricolage refers to the evolu-
tionary process as one that tinkers with
an existing genome to produce some-
thing new. [Church].
John Seely Brown, former chief scientist
for the Xerox Corporation and former di-
rector of the Xerox Palo Alto Research
Center captured its sense in a recent talk.
(cid:9)(cid:29)
(cid:14)(cid:17)(cid:16)(cid:10)(cid:8) (cid:28) (cid:11) (cid:17)(cid:13) (cid:21) (cid:23)(cid:22) (cid:12) (cid:9) (cid:8) (cid:30)(cid:15)(cid:18) (cid:8) (cid:11)(cid:19) (cid:19) (cid:4)(cid:15)(cid:19) (cid:4)(cid:17)(cid:11)(cid:16)(cid:3)(cid:8) (cid:12)(cid:15)(cid:2)(cid:3)(cid:21)
(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:27)(cid:8) ) (cid:10)(cid:11)(cid:16)(cid:8) (cid:2)(cid:3)(cid:11)(cid:6)(cid:12)(cid:8) (cid:30)(cid:15)(cid:18) (cid:8) (cid:13)(cid:4)(cid:17)(cid:6)(cid:5)(cid:8) (cid:17)(cid:16)(cid:8) (cid:17)(cid:6)(cid:16)(cid:15)(cid:8) (cid:30)(cid:15)(cid:18) (cid:4)(cid:8)
(cid:12)(cid:19) (cid:11)(cid:7)(cid:3)-(cid:8) (cid:30)(cid:15)(cid:18) (cid:8) (cid:16)(cid:17)(cid:6), (cid:3)(cid:4)(cid:8) ’ (cid:17)(cid:16)(cid:10)(cid:8) (cid:17)(cid:16)-(cid:8) (cid:11)(cid:6)% (cid:8) (cid:30)(cid:15)(cid:18) (cid:8) (cid:4)(cid:3)(cid:19) (cid:18) (cid:4)(cid:21)
(cid:19) (cid:15)(cid:12)(cid:3)(cid:8) (cid:17)(cid:16)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:4)(cid:3)(cid:19) (cid:15)(cid:12)(cid:17)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:17)(cid:16)(cid:27)(cid:8) (cid:29) (cid:10)(cid:3)(cid:6)(cid:8) (cid:30)(cid:15)(cid:18) (cid:8) (cid:4)(cid:3)(cid:19) (cid:18) (cid:4)(cid:21)
(cid:19) (cid:15)(cid:12)(cid:3)(cid:8)(cid:12)(cid:15)(cid:2)(cid:3)(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)-(cid:8)(cid:17)(cid:16)(cid:8)(cid:17)(cid:12)(cid:8)(cid:30)(cid:15)(cid:18) (cid:4)(cid:12)(cid:27)43(cid:8)(cid:8)
Ciborra [Ciborra] uses bricolage
to
characterize the way that organizations
tailor their information systems to their
changing needs through continual tinker-
ing.
This notion of building one thing upon
another applies to our framework in that
anything that persists in an environment
changes that environment for everything
else. The Internet provides many inter-
esting illustrations.
• Because the Internet exists at all, ac-
cess to a very large pool of people is
available. This enabled the develop-
ment of websites such as eBay.
• The establishment of eBay as a persis-
tent feature of the Internet environment
enabled the development of enterprises
whose only sales outlet was eBay.
These are enterprises with neither
brick and mortar nor web storefronts.
The only place they sell is on eBay.
This is a nice example of ecological
succession.
• At the same time—
and again because
the Internet provides access to a very
large number of people — other organi-
zations were able to establish what are
known as massively multi-player
online games. Each of these games is a
simulated world in which participants
interact with the game environment
and with each other. In most of these
games, participants seek to acquire vir-
tual game resources, such as magic
swords. Often it takes a fair amount of
43 In passing, Brown claims that this is how most
new technology develops.
(cid:9)) (cid:14)(cid:10)(cid:11)(cid:16)(cid:8) (cid:17)(cid:12)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) ’ (cid:11)(cid:30)(cid:8) ’ (cid:3)(cid:8) (cid:13)(cid:18) (cid:17)(cid:20)% (cid:8) (cid:11)(cid:20)(cid:2)(cid:15)(cid:12)(cid:16)(cid:8) (cid:11)(cid:20)(cid:20)(cid:8) (cid:16)(cid:3)(cid:7)(cid:10)(cid:6)(cid:15)(cid:20)(cid:15)(cid:5)(cid:30)(cid:8) (cid:16)(cid:15)% (cid:11)(cid:30)-(cid:8)
(cid:3)& (cid:3)(cid:6)(cid:8) (cid:16)(cid:10)(cid:15)(cid:18) (cid:5)(cid:10)(cid:8) (cid:2)(cid:30)(cid:8) (cid:20)(cid:11)’ (cid:30)(cid:3)(cid:4)(cid:12)(cid:8) % (cid:15)(cid:6)=(cid:16)(cid:8) ’ (cid:11)(cid:6)(cid:16)(cid:8) (cid:16)(cid:15)(cid:8) (cid:10)(cid:3)(cid:11)(cid:4)(cid:8) (cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8) (cid:17)(cid:16)(cid:27)(cid:8)
(cid:29) (cid:3)(cid:8) (cid:13)(cid:15)(cid:4)(cid:4)(cid:15)’ (cid:8) (cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:12)?(cid:8) ’ (cid:3)(cid:8) (cid:16)(cid:17)(cid:6), (cid:3)(cid:4)(cid:8) ’ (cid:17)(cid:16)(cid:10)(cid:8) (cid:16)(cid:10)(cid:3)(cid:2)?(cid:8) ’ (cid:3)(cid:8) (cid:2)(cid:15)% (cid:17)(cid:22)(cid:30)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:2)?(cid:8)’ (cid:3)(cid:8)3(cid:15)(cid:17)(cid:6)(cid:8)(cid:16)(cid:10)(cid:3)(cid:2)?(cid:8)’ (cid:3)(cid:8)(cid:13)(cid:18) (cid:17)(cid:20)% (cid:8)(cid:12)(cid:16)(cid:18) (cid:22)(cid:22)(cid:27)(cid:8)
Emergence Explained
41/67
DRAFT
2/12/2006
time, effort, and skill to acquire such
resources.
• The existence of all of these factors re-
sulted, though a creative leap, in an
eBay market in which players sold vir-
tual game assets for real money. This
market has become so large that there
are now websites dedicated exclu-
sively to trading in virtual game assets.
[Wallace]
• BBC News reported [BBC] that there
are companies
low-wage
that hire
Mexican and Chinese teenagers to earn
virtual assets, which are then sold in
these markets. How long will it be be-
fore a full-fledged economy develops
around these assets? There may be
brokers and retailers who buy and sell
these assets for their own accounts
even though they do not intend to play
the game. (Perhaps they already exist.)
Someone may develop a service that
tracks the prices of these assets. Per-
haps futures and options markets will
develop along with the inevitable in-
vestment advisors.
The point is that once something fits
well enough into its environment to per-
sist it adds itself to the environment for
everything else. This creates additional
possibilities and a world with ever in-
creasing complexity.
In each of the examples mentioned
above, one can identify what we have
been calling an autonomous entity. In
these entities are self-
most cases,
perpetuating
in
that
the amount of
money they extract from the environ-
ment (by selling either products, ser-
vices, or advertising)
than
is more
enough to pay for the resources needed
to keep it in existence.
In other cases, some Internet entities run
on time and effort contributed by volun-
teers. But the effect is the same. As long
as an entity is self-perpetuating, it be-
comes part of the environment and can
serve as the basis for the development of
additional entities.
6.4
Increasing
in-
complexity
creasing efficiency, and his-
torical contingency
The phenomenon whereby new entities
are built on top of existing entities is
now so widespread and commonplace
that it may seem gratuitous even to
comment on it. But it is an important
phenomenon, and one that has not re-
ceived the attention it deserves.
Easy though this phenomenon is to un-
derstand once one sees it, it is not trivial.
After all, the second law of thermody-
namics tells us that overall entropy in-
creases and complexity diminishes. Yet
we see complexity, both natural and man
made, continually increasing. For the
most part, this increasing complexity
consists of the development of new
autonomous entities, entities that imple-
ment the abstract designs of dissipative
structures.
This does not contradict the Second
Law. Each autonomous entity maintains
its own internally reduced entropy by us-
ing energy imported from the environ-
ment to export entropy to the environ-
ment. Overall entropy increases. Such a
process works only in an environment
that itself receives energy from outside
itself. Within such an environment,
complexity increases.
Progress in science and technology and
the bountifulness of the marketplace all
exemplify
increasing
this pattern of
complexity. One might refer to this kind
of pattern as a meta-epiphenomenon
since it is an epiphenomenon of the
process that creates epiphenomena.
Emergence Explained
42/67
DRAFT
2/12/2006
This creative process also tends to ex-
hibit a second meta-epiphenomenon.
Overall energy utilization becomes con-
tinually more efficient. As new autono-
mous entities find ways to use previ-
ously unused or under-used energy flows
(or forms of energy flows that had not
existed until
some newly created
autonomous entity generated them, per-
haps as a waste product), more of the
energy available to the system as a
whole is put to use.
The process whereby new autonomous
entities come into existence and perpetu-
ate themselves is non-reductive. It is
creative, contingent, and almost entirely
a sequence of historical accidents. As
they say, history is just one damn thing
after another—
to which we add, and na-
ture is a bricolage. We repeat the obser-
vation Anderson made more than three
decades ago.
) (cid:10)(cid:3)(cid:8) (cid:11)(cid:13)(cid:17)(cid:20)(cid:17)(cid:16)(cid:30)(cid:8) (cid:16)(cid:15)(cid:8) (cid:4)(cid:3)% (cid:18) (cid:7)(cid:3)(cid:8) (cid:3)& (cid:3)(cid:4)(cid:30)(cid:16)(cid:10)(cid:17)(cid:6)(cid:5)(cid:8) (cid:16)(cid:15)(cid:8) (cid:12)(cid:17)(cid:2)(cid:19) (cid:20)(cid:3)(cid:8)
(cid:22)(cid:18) (cid:6)% (cid:11)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:20)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8) (cid:9)% (cid:15)(cid:3)(cid:12)(cid:8)(cid:6)(cid:15)(cid:16)(cid:8) (cid:17)(cid:2)(cid:19) (cid:20)(cid:30)(cid:14)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)(cid:11)(cid:13)(cid:17)(cid:20)(cid:21)
(cid:17)(cid:16)(cid:30)(cid:8) (cid:16)(cid:15)(cid:8) (cid:12)(cid:16)(cid:11)(cid:4)(cid:16)(cid:8) (cid:22)(cid:4)(cid:15)(cid:2)(cid:8) (cid:16)(cid:10)(cid:15)(cid:12)(cid:3)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8)(cid:11)(cid:6)% (cid:8) (cid:4)(cid:3)(cid:7)(cid:15)(cid:6)(cid:12)(cid:16)(cid:4)(cid:18) (cid:7)(cid:16)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:8)(cid:18) (cid:6)(cid:17)& (cid:3)(cid:4)(cid:12)(cid:3)(cid:27)(cid:8)
7 Entities, emergence, and
science
7.1 Entities and the sciences
One reason that the sciences at levels
higher than physics and chemistry seem
somehow softer than physics and chem-
istry is that they work with autonomous
entities, entities that for the most part do
not supervene over any conveniently
compact collection of matter. Entities in
physics and chemistry are satisfyingly
solid—
or at least they seemed to be be-
fore quantum theory. In contrast, the en-
tities of the higher level sciences are not
defined in terms of material boundaries.
These entities don’ t exist as stable
clumps of matter; it ’ s hard to hold them
or in the grip
completely in one’ s hand —
of an instrument.
The entities of the special sciences are
objectively real —
there is some objective
measure (their reduced entropy relative
to their environment) by which they
qualify as entities. But as we saw earlier,
the processes through which these enti-
ties interact and by means of which they
perpetuate themselves are epiphenome-
nal. Even though the activities of higher
level entities may be described in terms
that are independent of the forces that
produce them (recall that this is our
definition of epiphenomenal), the fun-
damental forces of physics are the only
forces in nature. There is no strong
emergence. All other force-like effects
are epiphenomenal.
Consequently we find ourselves in the
position of claiming that the higher level
sciences study epiphenomenal interac-
tions among real if often somewhat ethe-
real entities.
7.2 Science and emergence
The idea that one can use heat and the
expansion of gases that it produces to
implement a particular function is not a
concept of fundamental physics. Of
course the Carnot engine is a conse-
quence of fundamental physics, but it is
not a concept of fundamental physics.
The idea of using a force to implement
new functionality is simply not within
the realm of fundamental physics.
Physics, like most science, does not con-
sider new functionality. It examines ex-
isting phenomena, and it asks how they
are brought about. It does not ask how
knowledge gained from such an analysis
can be used to implement something
new.
Here is a representative definition of the
term science.
Emergence Explained
43/67
DRAFT
2/12/2006
• ) (cid:10)(cid:3)(cid:8) (cid:15)(cid:13)(cid:12)(cid:3)(cid:4)& (cid:11)(cid:16)(cid:17)(cid:15)(cid:6)-(cid:8) (cid:17)% (cid:3)(cid:6)(cid:16)(cid:17)(cid:22)(cid:17)(cid:7)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)-(cid:8) % (cid:3)(cid:12)(cid:7)(cid:4)(cid:17)(cid:19) (cid:21)
(cid:16)(cid:17)(cid:15)(cid:6)-(cid:8) (cid:3)*(cid:19) (cid:3)(cid:4)(cid:17)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:20)(cid:8)
(cid:17)(cid:6)& (cid:3)(cid:12)(cid:16)(cid:17)(cid:5)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)-(cid:8) (cid:11)(cid:6)% (cid:8)
(cid:16)(cid:10)(cid:3)(cid:15)(cid:4)(cid:3)(cid:16)(cid:17)(cid:7)(cid:11)(cid:20)(cid:8) (cid:3)*(cid:19) (cid:20)(cid:11)(cid:6)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:15)(cid:22)(cid:8) (cid:19) (cid:10)(cid:3)(cid:6)(cid:15)(cid:2)(cid:3)(cid:6)(cid:11)(cid:27)(cid:8)
(cid:9)6(cid:2)(cid:3)(cid:4)(cid:17)(cid:7)(cid:11)(cid:6)(cid:8)B (cid:3)(cid:4)(cid:17)(cid:16)(cid:11)(cid:5)(cid:3)(cid:14)(cid:8)
Science is thus the study of nature, how
it is designed, i.e., organized, and how
its designs work. Science does not have
as part of its charter to take what is
known about nature and to create some-
thing new.
Recall our discussion of hurricanes. Ap-
parently they are the only kind of
weather system with an internal power
plant. Let ’ s imagine that no hurricane
ever existed—
at least not anywhere that
an earthbound scientist could observe it.
Under those circumstances no scientist
would hypothesize the possibility of
such a weather system. Doing so just
isn’ t part of the scientific agenda; it is
not the kind of task that scientists set for
themselves.
Why waste one’ s time thinking about
something so strange—
a weather system
that not only contains its own built-in
power plant but one in which the heat is
generated by condensation rather than
combustion and the “ furnace ” in which
the heat is generated is located in the up-
per atmosphere. Thinking through such a
possibility might make interesting sci-
ence fiction.
+(cid:6)(cid:8) (cid:11)(cid:8) (cid:5)(cid:11)(cid:20)(cid:11)*(cid:30)(cid:8) (cid:22)(cid:11)(cid:4)(cid:8) (cid:11)’ (cid:11)(cid:30)-(cid:8) (cid:15)(cid:6)(cid:8) (cid:11)(cid:8) (cid:19) (cid:20)(cid:11)(cid:6)(cid:3)(cid:16)(cid:8) (cid:15)(cid:22)(cid:8) (cid:11)(cid:8) (cid:2)(cid:3)(cid:21)
% (cid:17)(cid:18) (cid:2)(cid:8)(cid:12)(cid:17)(cid:31) (cid:3)(cid:8)(cid:12)(cid:16)(cid:11)(cid:4)(cid:8)(cid:6)(cid:3)(cid:11)(cid:4)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:3)% (cid:5)(cid:3)(cid:8)(cid:15)(cid:22)(cid:8)(cid:16)(cid:10)(cid:11)(cid:16)(cid:8)(cid:5)(cid:11)(cid:20)(cid:11)*(cid:30)-(cid:8)
(cid:11)(cid:8) (cid:19) (cid:20)(cid:11)(cid:6)(cid:3)(cid:16)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:10)(cid:11)% (cid:8) (cid:12)(cid:16)(cid:15)(cid:4)(cid:2)(cid:12)(cid:8) ’ (cid:17)(cid:16)(cid:10)(cid:8) (cid:16)(cid:10)(cid:3)(cid:17)(cid:4)(cid:8) (cid:15)’ (cid:6)(cid:8)
(cid:13)(cid:18) (cid:17)(cid:20)(cid:16)(cid:21)(cid:17)(cid:6)(cid:8)(cid:10)(cid:3)(cid:11)(cid:16)(cid:8)(cid:3)(cid:6)(cid:5)(cid:17)(cid:6)(cid:3)(cid:12)-(cid:8)(cid:28)
(cid:27)(cid:8)(cid:8)
Certainly nothing so bizarre could ever
occur naturally. It would not be consid-
ered science.
Imagine also how bizarre phase transi-
tions would seem if they weren’ t so
common— matter sometimes obeying
one set of rules and sometimes obeying
another set. It wouldn’ t make any sense.
What would happen at the boundaries?
How would transitions occur? If phase
transitions didn’ t happen naturally, sci-
ence almost certainly wouldn’ t invent
them.
If we conceive of science as the study of
existing phenomena, science is reduc-
tionism. To paraphrase Weinberg, the
goal of science is
(cid:16)(cid:15)(cid:8) (cid:22)(cid:17)(cid:6)% (cid:8) (cid:12)(cid:17)(cid:2)(cid:19) (cid:20)(cid:3)(cid:8) (cid:18) (cid:6)(cid:17)& (cid:3)(cid:4)(cid:12)(cid:11)(cid:20)(cid:8) (cid:20)(cid:11)’ (cid:12)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:3)*(cid:19) (cid:20)(cid:11)(cid:17)(cid:6)(cid:8)
’ (cid:10)(cid:30)(cid:8)(cid:6)(cid:11)(cid:16)(cid:18) (cid:4)(cid:3)(cid:8)(cid:17)(cid:12)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)’ (cid:11)(cid:30)(cid:8)(cid:17)(cid:16)(cid:8)(cid:17)(cid:12)(cid:27)(cid:8)
When science is understood in this way,
mathematics, computer science, and en-
gineering, all of which create and study
conceptual structures that need not exist,
are not science. Indeed scientists and
mathematicians are often surprised when
they find that a mathematical construct
that had been studied simply because it
seemed mathematically interesting has a
scientific application.
Fortunately for us, nature is not a scien-
tist. Like computer scientists and engi-
neers, she too creates things that need
not exist — people and hurricanes, for ex-
ample.
What about this paper? We would cate-
gorize this paper as science because one
of its goals is to help explain, i.e., to
provide some intellectual leverage for
understanding, why the nature is the way
it is.
This immediately raises another ques-
tion: if this is science, are we happy with
it? Let ’ s assume that the simplest and
most universal way to understand nature
is in terms of multilevel abstractions. Is
this satisfactory? Is this approach to sci-
entific explanation as real and as con-
crete as explaining nature in terms of
more absolute single-level laws? Isn’ t
there something unreal about explaining
nature at least in part as implementations
of abstractions?
Emergence Explained
44/67
DRAFT
2/12/2006
One way to argue for the reality of these
abstractions is to show that they build
upon each other. When a new abstrac-
tion is implemented in terms of the func-
tionalities embodied
in existing ab-
stractions, there seems little choice but
to acknowledge the reality of the imple-
menting abstractions.
The obvious place to look for sciences
building upon other sciences is the hier-
archy of the sciences. To take the most
concrete case, chemistry is built on the
abstraction of the atom as an entity with
an internal structure. Yet we certainly
don’ t wonder about whether chemistry is
a real science.
Molecules are the (emergent) entities of
chemistry. They form when combina-
tions of atoms in are in a lower energy
state than the atoms would be in isola-
tion. How does nature implement this?
It does it in terms of abstract structures
known as orbitals.
Molecules form when orbitals from pairs
of atoms merge. What is an orbital? It is
part of the abstract design—
the design
that determines how electrons and pro-
tons relate to each other—
that matter
implements by following the rules of
quantum mechanics. Thus molecular
bonds are
implemented by nature
through the quantum mechanical mecha-
nisms of orbitals. Like the formation of
atoms themselves, chemical bonding is
part of the free lunch that nature sets out
for us —
and another illustration that
emergence is a fundamental aspect of
nature.
Of course, this is just one example. As
we shall saw above, the complexity that
we see around us is a direct result of the
fact that new abstractions may be built
on top of the functionalities provided by
existing abstractions.
8 Varieties of Emergence
In this section we stand back and review
the kinds of emergence we have dis-
cussed. In particular we discuss two
categories of emergence: static emer-
gence and dynamic emergence. We also
suggest that these categories correspond
to Weinberg’ s notion of petty and grand
reductionism.
As in the case of Weinberg’ s petty and
grand reductionism, static emergence,
while of great importance, is of lesser in-
terest. It is dynamic emergence, and es-
pecially stigmergic dynamic emergence
that is central to complex systems.
Recall that we defined a phenomenon as
emergent over a underlying model if
(a) it has an independent conceptualiza-
tion and (b) it can be implemented in
terms of elements of that model.
8.1 Static emergence
An emergent phenomenon is statically
emergent if its implementation does not
depend on time.
As an interesting example of static
emergence, consider cloth as a collection
of threads woven together. Cloth has the
emergent property that it is able to cover
a surface. This property is implicitly two
dimensional. The components of cloth,
i.e., threads, do not have (or at least are
not understood in terms of) that prop-
erty. A thread is understood in terms of
the property length. Yet when threads
are woven together the resulting cloth
has this new property, which effectively
converts a collection of one dimensional
components to a two dimensional object.
Many human manufactured or con-
structed artifacts exhibit static emer-
gence. A house has the statically emer-
gent
property
number-of-bedrooms.
More generally, a house has the emer-
gent property that it can serve as a resi-
Emergence Explained
45/67
DRAFT
2/12/2006
dence. Static emergence also occurs in
nature. As Weinberg points out,
(cid:9)6(cid:14)(cid:8) % (cid:17)(cid:11)(cid:2)(cid:15)(cid:6)% (cid:8) (cid:17)(cid:12)(cid:8) (cid:10)(cid:11)(cid:4)% (cid:8) (cid:13)(cid:3)(cid:7)(cid:11)(cid:18) (cid:12)(cid:3)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:7)(cid:11)(cid:4)(cid:13)(cid:15)(cid:6)(cid:8)
(cid:11)(cid:16)(cid:15)(cid:2)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) ’ (cid:10)(cid:17)(cid:7)(cid:10)(cid:8) (cid:17)(cid:16)(cid:8) (cid:17)(cid:12)(cid:8) (cid:7)(cid:15)(cid:2)(cid:19) (cid:15)(cid:12)(cid:3)% (cid:8) (cid:7)(cid:11)(cid:6)(cid:8) (cid:22)(cid:17)(cid:16)(cid:8) (cid:16)(cid:15)(cid:21)
(cid:5)(cid:3)(cid:16)(cid:10)(cid:3)(cid:4)(cid:8)(cid:6)(cid:3)(cid:11)(cid:16)(cid:20)(cid:30)(cid:8)(cid:9)(cid:3)& (cid:3)(cid:6)(cid:8)(cid:16)(cid:10)(cid:15)(cid:18) (cid:5)(cid:10)(cid:14)(cid:8)(cid:17)(cid:16)(cid:8)% (cid:15)(cid:3)(cid:12)(cid:6)=(cid:16)(cid:8)(cid:2)(cid:11), (cid:3)(cid:8)
(cid:12)(cid:3)(cid:6)(cid:12)(cid:3)(cid:8)(cid:16)(cid:15)(cid:8)(cid:16)(cid:11)(cid:20), (cid:8)(cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:10)(cid:11)(cid:4)% (cid:6)(cid:3)(cid:12)(cid:12)(cid:8)(cid:28)
(cid:8)(cid:15)(cid:22)(cid:8) (cid:17)(cid:6)% (cid:17)(cid:21)
& (cid:17)% (cid:18) (cid:11)(cid:20)(cid:8)#(cid:3)(cid:20)(cid:3)(cid:2)(cid:3)(cid:6)(cid:16)(cid:11)(cid:4)(cid:30)$(cid:8)(cid:19) (cid:11)(cid:4)(cid:16)(cid:17)(cid:7)(cid:20)(cid:3)(cid:12)(cid:27)(cid:8)
Since statically emergent phenomena
must be implemented in terms of some
underlying model, and since time is by
definition excluded from that implemen-
tation, static emergence is equivalent to
Weinberg’ s petty reductionism.
8.2 Dynamic emergence
Properties or phenomena of a model are
dynamically emergent if they are defined
in terms of how the model changes (or
doesn’ t change) over some time. Dy-
namic emergence occurs either with or
without autonomous entities. We call the
former stigmergic emergence, but we
look at non-stigmergic dynamic emer-
gence first.
dynamic
8.3 Non-stigmergic
emergence
Interactions among at-equilibrium enti-
ties result in non-stigmergic dynamic
emergence.
are:
examples
Two
(a) objects moving in space and interact-
ing according to Newtonian mechanics
and (b) the quantum wave function.
What appears to be distinctive about
such systems is that they are not charac-
terized in terms of discrete states that
their elements assume. Elements do not
transition from one state to another.
Such systems may be defined in terms of
continuous equations.
The quantum wave function is an espe-
cially interesting example. As long as it
does not undergo decoherence, i.e., in-
teraction with an environment, the wave
function encompasses all possibilities,
but it realizes none of them.
Since quantum states are discrete (hence
the term quantum), objects cannot transi-
tion smoothly from one quantum state to
another. So how does that transition oc-
cur? Quantum theory turns these transi-
tions into probabilities. As Hardy points
out [Hardy], by making such transitions
probabilistically continuous,
( (cid:18) (cid:11)(cid:6)(cid:16)(cid:18) (cid:2)(cid:8)(cid:16)(cid:10)(cid:3)(cid:15)(cid:4)(cid:30)(cid:8)(cid:15)(cid:22)(cid:22)(cid:3)(cid:4)(cid:12)(cid:8)(cid:18) (cid:12)(cid:8)(cid:11)(cid:8)’ (cid:11)(cid:30)(cid:8)(cid:16)(cid:15)(cid:8)(cid:10)(cid:11)& (cid:3)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)
(cid:11)% & (cid:11)(cid:6)(cid:16)(cid:11)(cid:5)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) % (cid:17)(cid:12)(cid:7)(cid:4)(cid:3)(cid:16)(cid:3)(cid:6)(cid:3)(cid:12)(cid:12)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:7)(cid:15)(cid:6)(cid:16)(cid:17)(cid:6)(cid:18) (cid:17)(cid:16)(cid:30)(cid:8)
(cid:11)(cid:16)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:12)(cid:11)(cid:2)(cid:3)(cid:8)(cid:16)(cid:17)(cid:2)(cid:3)(cid:27)(cid:8)(cid:8)
For all practical purposes, actually to as-
sume a state requires something more,
the so-called collapse of the wave func-
tion. That happens stigmergically. At the
quantum level, stigmergy is equivalent
to decoherence.
and
emergence
8.4 Dynamic
grand reductionism
As static emergence corresponds
to
Weinberg’ s petty reductionism, dynamic
emergence seems to correspond nicely to
Weinberg’ s grand reductionism. Wein-
berg explains grand reductionism as fol-
lows
(cid:9)) (cid:14)(cid:10)(cid:3)(cid:8)(cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:17)(cid:12)(cid:16)(cid:8)(cid:4)(cid:3)(cid:5)(cid:11)(cid:4)% (cid:12)(cid:8)(cid:16)(cid:10)(cid:3)(cid:8)(cid:5)(cid:3)(cid:6)(cid:3)(cid:4)(cid:11)(cid:20)(cid:8)(cid:16)(cid:10)(cid:3)(cid:15)(cid:21)
(cid:4)(cid:17)(cid:3)(cid:12)(cid:8) (cid:5)(cid:15)& (cid:3)(cid:4)(cid:6)(cid:17)(cid:6)(cid:5)(cid:8) (cid:11)(cid:17)(cid:4)(cid:8) (cid:11)(cid:6)% (cid:8) ’ (cid:11)(cid:16)(cid:3)(cid:4)(cid:8) (cid:11)(cid:6)% (cid:8) (cid:4)(cid:11)% (cid:17)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8)
(cid:11)(cid:12)(cid:8) (cid:13)(cid:3)(cid:17)(cid:6)(cid:5)(cid:8) (cid:11)(cid:16)(cid:8) (cid:11)(cid:8) % (cid:3)(cid:3)(cid:19) (cid:3)(cid:4)(cid:8) (cid:20)(cid:3)& (cid:3)(cid:20)(cid:8) (cid:16)(cid:10)(cid:11)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:15)(cid:4)(cid:17)(cid:3)(cid:12)(cid:8)
(cid:11)(cid:13)(cid:15)(cid:18) (cid:16)(cid:8) (cid:7)(cid:15)(cid:20)% (cid:8) (cid:22)(cid:4)(cid:15)(cid:6)(cid:16)(cid:12)(cid:8) (cid:15)(cid:4)(cid:8) (cid:16)(cid:10)(cid:18) (cid:6)% (cid:3)(cid:4)(cid:12)(cid:16)(cid:15)(cid:4)(cid:2)(cid:12)-(cid:8) (cid:6)(cid:15)(cid:16)(cid:8) (cid:17)(cid:6)(cid:8)
(cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:3)(cid:6)(cid:12)(cid:3)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:30)(cid:8) (cid:11)(cid:4)(cid:3)(cid:8) (cid:2)(cid:15)(cid:4)(cid:3)(cid:8) (cid:18) (cid:12)(cid:3)(cid:22)(cid:18) (cid:20)-(cid:8) (cid:13)(cid:18) (cid:16)(cid:8)
(cid:15)(cid:6)(cid:20)(cid:30)(cid:8) (cid:17)(cid:6)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)(cid:12)(cid:3)(cid:6)(cid:12)(cid:3)(cid:8) (cid:16)(cid:10)(cid:11)(cid:16)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:20)(cid:11)(cid:16)(cid:16)(cid:3)(cid:4)(cid:8)(cid:7)(cid:11)(cid:6)(cid:8) (cid:17)(cid:6)(cid:8)(cid:19) (cid:4)(cid:17)(cid:6)(cid:21)
(cid:7)(cid:17)(cid:19) (cid:20)(cid:3)(cid:8) (cid:13)(cid:3)(cid:8) (cid:18) (cid:6)% (cid:3)(cid:4)(cid:12)(cid:16)(cid:15)(cid:15)% (cid:8) (cid:11)(cid:12)(cid:8) (cid:2)(cid:11)(cid:16)(cid:10)(cid:3)(cid:2)(cid:11)(cid:16)(cid:17)(cid:7)(cid:11)(cid:20)(cid:8) (cid:7)(cid:15)(cid:6)(cid:21)
(cid:12)(cid:3)( (cid:18) (cid:3)(cid:6)(cid:7)(cid:3)(cid:12)(cid:8) (cid:15)(cid:22)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:22)(cid:15)(cid:4)(cid:2)(cid:3)(cid:4)(cid:27)(cid:8) ) (cid:10)(cid:3)(cid:8) (cid:4)(cid:3)% (cid:18) (cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:17)(cid:12)(cid:16)(cid:8)
(cid:19) (cid:4)(cid:15)(cid:5)(cid:4)(cid:11)(cid:2)(cid:8) (cid:15)(cid:22)(cid:8) (cid:19) (cid:10)(cid:30)(cid:12)(cid:17)(cid:7)(cid:12)(cid:8) (cid:17)(cid:12)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8) (cid:12)(cid:3)(cid:11)(cid:4)(cid:7)(cid:10)(cid:8) (cid:22)(cid:15)(cid:4)(cid:8) (cid:16)(cid:10)(cid:3)(cid:8)
(cid:7)(cid:15)(cid:2)(cid:2)(cid:15)(cid:6)(cid:8)(cid:12)(cid:15)(cid:18) (cid:4)(cid:7)(cid:3)(cid:8)(cid:15)(cid:22)(cid:8)(cid:11)(cid:20)(cid:20)(cid:8)(cid:3)*(cid:19) (cid:20)(cid:11)(cid:6)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:12)(cid:27)(cid:8)(cid:28)
(cid:8)(cid:8)
We hope that Weinberg would not object
to the following paraphrase.
The reductionist goal (with respect to
reducing weather terminology to the
terminology of physics) is to build a
model (a) whose elements include air,
Emergence Explained
46/67
DRAFT
2/12/2006
water vapor,
and
radiation
and
(b) whose elements interact according
to the principles of physics. When that
model is run it will generate the emer-
gent phenomena that we would recog-
nize as cold fronts and thunderstorms.
Grand reductionism is thus the explana-
tion of phenomena at one level in terms
of phenomena at a more fundamental
level. One shows that when the laws at
the more fundamental level are applied,
the result will be the phenomena of in-
terest at the less fundamental level.
Since these sorts of models are inevita-
bly dynamic this is dynamic emergence
but expressed in other terms.
Nor should the equating of grand reduc-
tionism with dynamic emergence sur-
prise anyone in the field of complex sys-
tems. After all, the presumed reason to
build a model is to show that a set of
lower level rules will produce higher
level results — which is exactly the grand
reductionist agenda.
Of course the terminology that we would
use is not just that lower level rules pro-
duce higher level results, that lower level
rules may implement a higher level ab-
straction.
But no matter how it is expressed, with-
out the lower level substrate, the higher
level phenomena would not exist.
8.5 Stigmergic emergence
Stigmergic emergence is dynamic emer-
gence that involves autonomous entities.
What tends to be most interesting about
autonomous entities are (a) they may as-
sume discrete states and (b) they change
state as they interact with their environ-
ments.44
44 One might liken an isolated quantum wave system
to the inside of an autonomous entity. It assumes a
Furthermore, not only do autonomous
entities depend on their environments as
sources of energy and other resources,
the environment on which any autono-
mous entity depends
includes other
autonomous entities. Of course these
other autonomous entities also depend
on their environment, etc. These de-
pendencies form networks of enormous
and complexity in which the dependency
links are frequently not higher depend-
ing on lower.
Static and non-stigmergic dynamic
emergence is fairly well-behaved. One
can often write down equations that
characterize entire systems in which it
occurs —
even though it may not be prac-
tical to solve those equations for other
than trivial cases. Stigmergic emergence
is far worse. Because of the relative in-
terdependence of the components, it is
virtually impossible to provide a global
equation-like characterization of the sys-
tem as a whole. Stigmergic emergence is
the source of the complexity in nature. It
is because of stigmergic emergence that
complex systems are complex.
This would seem to put a final stake in
the heart of Laplace’ s demon, the hypo-
thetical computing device that if given
details about the initial state of the uni-
verse would be able to compute all fu-
ture states. Laplace ’ s demon may suc-
ceed in a Newtonian universe, for which
it was invented. Laplace’ s demon may
even succeed in a quantum mechanical
universe in that the quantum wave equa-
tion is deterministic—
even though it
characterizes probability amplitudes and
hence its collapse is not. But if nature
includes asynchronously acting autono-
mous entities, some of which may them-
state (i.e., collapses) when it interacts with its en-
vironment.
Emergence Explained
47/67
DRAFT
2/12/2006
selves embody quantum probability tran-
sitions, many of which are mutually in-
terdependent, and all of which depend
on their environment, which includes
other autonomous entities for their op-
eration and persistence, Laplace’ s de-
mon will be way beyond its depth.
One possible simple formal model for
such a computational system is a shared
tape Turing Machine community: a col-
lection of asynchronously operating Tur-
ing Machines that share a single tape.45
Some proponents of agent-based model-
ing argue for that approach on the
grounds that even though some domains
may have global characterizations, those
characterizations are much too complex
to compute. Our position is that agent-
based modeling is appropriate because
that ’ s how nature is.
9 Some practical considera-
tions
9.1 Emergence and software
As noted earlier, the computation that
results when software is executed is
emergent. It is an epiphenomenon of the
operation of the (actual or virtual) ma-
chine that executes the software.
Earlier we defined emergence as syn-
onymous with epiphenomenon. At that
time we suggested that formalizable
epiphenomena are often of significant
interest. We also said that formalization
may not always be in the cards. Soft-
ware, which one would imagine to be a
perfect candidate for formalization, now
seems to be a good example of an
45 Wegner ’ s work [Wegner] on non-traditional Tur-
ing Machine models begins to explore his own
models. Cockshott and Michaelson [Cockshott]
dispute whether Wegner ’ s models extend the
power of the Turing machine.
epiphenomenon that is unlikely to be
formalized.
It had once been hoped that software de-
velopment could evolve to a point at
which one need only write down a for-
mal specification of what one wanted the
software to do. Then some automatic
process would produce software that sat-
isfied that specification.
That dream now seems quite remote.
Besides the difficulty of developing (a) a
satisfactory specification language and
(b) a system that can translate specifica-
tions written in such a language into ex-
ecutable code, the real problem is that it
has turned out to be at least as difficult
and complex to write formal specifica-
tions as it is to write the code that pro-
duces the specified results.
Even if one could write software by
writing specifications, in many cases —
especially cases that involve large and
complex systems, the kinds of cases for
which
it
really matters —
doing
so
doesn’ t seem to result in much intellec-
tual leverage, if indeed it produces any at
all.
This illustrates quite nicely that we often
find ourselves in the position of wanting
to produce epiphenomena
(epiphe-
nomena, which may be very important to
us), whose formalization as an abstrac-
tion we find to be either infeasible or not
particularly useful.
9.2 Bricolage as design
The process of building one capability
on top of another not only drives the
overall increase in complexity, it also
provides guidance to designers about
how to do good design work. Any good
designer—
a developer, an architect, a
programmer, or an engineer—
knows
that it is often best if one can take advan-
Emergence Explained
48/67
DRAFT
2/12/2006
tage of forces and processes already in
existence as part of one’ s design.
But even before engineering, we as hu-
man beings made use of pre-existing ca-
pabilities. Agriculture and animal hus-
bandry use both plant reproduction and
such animal capabilities as locomotion
or material (i.e., skin) production for our
own purposes. The exploitation of exist-
ing capabilities for our own purposes is
not a new idea.
An interesting example of this approach
to engineering involves recent develop-
ments
in robotics. Collins reported
[Collins] that a good way to make a ro-
bot walk is by exploiting gravity through
what he called passive-dynamic mo-
tion—
raise the robot ’ s leg and let gravity
pull it back down—
rather than by direct-
ing the robot ’ s limbs to follow a pre-
defined trajectory.
This illustrates in a very concrete way
the use of an existing force in a design.
Instead of building a robot whose every
motion was explicitly programmed,
Collins built a robot whose motions were
controlled in part by gravity, a pre-
existing force.
9.3
Infrastructure-centric devel-
opment
Building new capabilities on top of ex-
isting ones is not only good design, it is
highly leveraged design. But now that
we are aware of this strategy a further
lesson can be drawn. New systems
should be explicitly designed to serve as
a possible basis for systems yet to come.
Another way of putting this is that every
time we build a new system, it should be
built so that it becomes part of our envi-
ronment, i.e., our infrastructure, and not
just a piece of closed and isolated func-
tionality.
By infrastructure we mean systems such
as the Internet, the telephone system, the
electric power distribution system, etc.
Each of these systems can be character-
ized in isolation in terms of the particu-
lar functions they perform. But more
important than the functional characteri-
zation of any of these individual systems
is the fact that they exist in the environ-
ment in such a way that other systems
can use them as services.
We should apply this perspective to all
new systems that we design: design them
as infrastructure services and not just as
bits of functionality. Clearly Microsoft
understands this. Not only does it posi-
tion the systems it sells as infrastructure
services, it also maintains tight owner-
ship and control over them. When such
systems become widely used elements of
the economy, the company makes a lot
of money. The tight control it maintains
and the selfishness with which it controls
these systems earns it lots of resentment
as well. Society can’ t prosper when any
important element of its infrastructure is
controlled primarily for selfish purposes.
The US Department of Defense (DoD) is
currently reinventing itself [Dick] to be
more infrastructure-centric. This requires
it to transform what is now a
(cid:10)(cid:18) (cid:5)(cid:3)(cid:8) (cid:7)(cid:15)(cid:20)(cid:20)(cid:3)(cid:7)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:15)(cid:22)(cid:8) (cid:17)(cid:6)% (cid:3)(cid:19) (cid:3)(cid:6)% (cid:3)(cid:6)(cid:16)(cid:8) /(cid:12)(cid:16)(cid:15)& (cid:3)(cid:19) (cid:17)(cid:19) (cid:3)0 (cid:8)
(cid:17)(cid:6)(cid:22)(cid:15)(cid:4)(cid:2)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)(cid:8) (cid:12)(cid:30)(cid:12)(cid:16)(cid:3)(cid:2)(cid:12)-(cid:8) (cid:3)(cid:11)(cid:7)(cid:10)(cid:8) (cid:12)(cid:18) (cid:19) (cid:19) (cid:15)(cid:4)(cid:16)(cid:17)(cid:6)(cid:5)(cid:8) (cid:15)(cid:6)(cid:20)(cid:30)(cid:8)
(cid:17)(cid:16)(cid:12)(cid:8) (cid:15)(cid:4)(cid:17)(cid:5)(cid:17)(cid:6)(cid:11)(cid:20)(cid:8) (cid:19) (cid:4)(cid:15)(cid:7)(cid:18) (cid:4)(cid:3)(cid:2)(cid:3)(cid:6)(cid:16)(cid:8) (cid:12)(cid:19) (cid:3)(cid:7)(cid:17)(cid:22)(cid:17)(cid:7)(cid:11)(cid:16)(cid:17)(cid:15)(cid:6)-(cid:8) (cid:16)(cid:15)(cid:8) (cid:11)(cid:8)
(cid:18) (cid:6)(cid:17)(cid:22)(cid:17)(cid:3)% (cid:8)(cid:11)(cid:12)(cid:12)(cid:3)(cid:2)(cid:13)(cid:20)(cid:30)(cid:8)(cid:15)(cid:22)(cid:8)(cid:17)(cid:6)(cid:16)(cid:3)(cid:4)(cid:15)(cid:19) (cid:3)(cid:4)(cid:11)(cid:16)(cid:17)(cid:6)(cid:5)(cid:8)(cid:12)(cid:30)(cid:12)(cid:16)(cid:3)(cid:2)(cid:12)(cid:27)(cid:8)(cid:8)
The evocative term stovepipe is intended
to distinguish the existing situation—
in
which the DoD finds that it has acquired
and deployed a large number of func-
tionally isolated systems (the “ stove-
pipes ” )—
from the more desirable situa-
tion in which all DoD systems are avail-
able to each other as an infrastructure of
services.
Emergence Explained
49/67
DRAFT
2/12/2006
9.4 Service refactoring and the
age of services
The process whereby infrastructure ser-
vices build on other infrastructure ser-
vices leads not only to new services, it
also leads to service refactoring. The
corporate trend toward outsourcing func-
tions that are not considered part of the
core competence of the corporation illus-
trates this. Payroll processing is a typical
example.
Because many organizations have em-
ployees who must be paid, these organi-
zations must provide a payroll service
for themselves. It has now become fea-
sible to factor out that service and offer
it as part of our economic infrastructure.
This outsourcing of internal processes
leads to economic efficiencies in that
many such processes can be done more
efficiently when performed by special-
ized organizations. Such specialized or-
ganizations can
take advantage of
economies of scale. They can also serve
as focal points where expertise in their
specialized service can be concentrated
and the means of providing those ser-
vices improved.
As this process establishes itself ever
more firmly, more and more organiza-
tions will focus more on offering ser-
vices rather than functions, and organi-
zations will become less stovepiped.
We frequently speak of the “ service in-
dustries. ” For the most part this term has
been used to refer to low level ser-
vices —
although even the fast food in-
dustry can be seen as the “ outsourcing”
of the personal food preparation func-
tion. With our more general notion of
service in mind, historians may look
back to this period as the beginning of
the age of services.
Recall that a successful service is an
autonomous entity. It persists as long as
it is able to extract from its environment
enough resources, typically money, to
perpetuate itself.
9.5 A possible undesirable unin-
tended consequence
The sort of service refactoring we just
discussed tends to make the overall eco-
nomic system more efficient. It also
tends to improve reliability: the payroll
service organizations are more reliable
than the average corporate payroll de-
partment.
On the other hand, by eliminating re-
dundancy, efficiency makes the overall
economic system more vulnerable to
large scale failure. If a payroll service
organization has a failure, it is likely to
have a larger impact than the failure of
any one corporate payroll department.
This phenomenon seems to be quite
common—
tending to transform failure
statistics from a Gaussian to a scale free
distribution: the tails are longer and fat-
ter. [Colbaugh] Failures may be less
frequent, but when they occur they may
be more global.
This may be yet another unintended and
unexpected emergent phenomenon—
a
modern example of the tragedy of the
commons.
Increased economic effi-
ciency leads to increased vulnerability to
major disasters at the societal-level.
On the other hand, perhaps our growing
realization that catastrophic failures may
occur along with our ability to factor out
commonly needed services will help us
solve this problem as well. We now see
increasing number of disaster planning
services being offered.
Emergence Explained
50/67
DRAFT
2/12/2006
9.6 Modeling:
the difficulty of
looking downward
The perspective we have described
yields two major implications for model-
ing. We refer to them as the difficulty of
looking downwards and the difficulty of
looking upwards. In both cases, the
problem is that it is very difficult to
model significant creativity—
notwith-
standing the fact that surprises do appear
in some of our models. In this section we
examine the difficulty of looking down-
ward. In the next we consider the diffi-
culty of looking upward.
Strict reductionism, our conclusion that
all forces and actions are epiphenomenal
over forces and actions at the fundamen-
tal level of physics, implies that it is im-
possible to find a non-arbitrary base
level for models. One never knows what
unexpected effects one may be leaving
out by defining a model in which inter-
actions occur at some non-fundamental
level.
Consider a model of computer security.
Suppose that by analyzing the model one
could guarantee that a communication
line uses essentially unbreakable encryp-
tion technology. Still it is possible for
someone inside to transmit information
to someone outside.
How? By sending messages in which the
content of the message is ignored but the
frequency of transmission carries the in-
formation, e.g., by using Morse code.
The problem is that the model didn’ t in-
clude that level of detail. This is the
problem of looking downward.
A further illustration of this difficulty is
that there are no good models of biologi-
cal arms races. (There don’ t seem to be
any good models of significant co-
evolution at all.) There certainly are
models of population size effects in
predator-prey simulations. But by bio-
logical arms races we are talking about
not just population sizes but actual evo-
lutionary changes.
Imagine a situation in which a plant spe-
cies comes under attack from an insect
species. In natural evolution the plant
may “ figure out ” how to grow bark. Can
we build a computer model in which this
solution would emerge? It
is very
unlikely. To do so would require that our
model have built into it enough informa-
tion about plant biochemistry to enable it
to find a way to modify that biochemis-
try to produce bark, which itself is de-
fined implicitly in terms of a surface that
the insect cannot penetrate. Evolving
bark would require an enormous amount
of information —
especially if we don’ t
want to prejudice the solution the plant
comes up with.
The next step, of course, is for the insect
to figure out how to bore through bark.
Can our model come up with something
like that? Unlikely. What about the
plant ’ s next step: “ figuring out ” how to
produce a compound that is toxic to the
insect? That requires that the model in-
clude information about both plant and
insect biochemistry—
and how the plant
can produce a compound that interferes
with the insect ’ s internal processes. This
would be followed by the development
by the insect of an anti-toxin defense.
To simulate this sort of evolutionary
process would require an enormous
amount of low level detail —
again espe-
cially if we don’ t want to prejudice the
solution in advance.
Other than Tierra (see [Ray]) and its
successors, which seem to lack the rich-
ness to get very far off the ground, as far
as we know, there are no good computer
models of biological arms races. A
seemingly promising approach would be
an agent-based system in which each
Emergence Explained
51/67
DRAFT
2/12/2006
agent ran its own internal genetic pro-
gramming model. But we are unaware of
any such work.46
Finally, consider the fact that geckos
climb walls by taking advantage of the
Van der Walls “ force. ” (We put force in
quotation marks because there is no Van
der Walls force. It is an epiphenomenon
of relatively rarely occurring quantum
phenomena.) To build a model of evolu-
tion in which creatures evolve to use the
Van der Walls force to climb walls
would require that we build quantum
physics into what is presumably in-
tended to be a relatively high-level bio-
logical model in which macro geckos
climb macro walls
It ’ s worth noting that the use of the Van
der Walls force was apparently not an
extension of some other gecko process.
Yet the gecko somehow found a way to
reach directly down to a quantum-level
effect to find a way to climb walls.
The moral is that any base level that we
select for our models will be arbitrary,
and by choosing that base level, we may
miss
important possibilities. Another
moral is that models used when doing
computer security or terrorism analy-
46 Genetic programming is relevant because we are
assuming that the agent has an arbitrarily detailed
description of how the it functions and how ele-
ments in its environment function.
Notice how difficult it would be implement such a
system. The agent ’ s internal model of the envi-
ronment would have to be updated continually as
the environment changed. That requires a means to
perceive the environment and to model changes in
it. Clearly that ’ s extraordinarily sophisticated. Al-
though one could describe such a system without
recourse to the word consciousness, the term does
come to mind.
Nature’ s approach is much simpler: change during
reproduction and see what happens. If the result is
unsuccessful, it dies out; if it is successful it per-
sists and reproduces. Of course that requires an en-
tire generation for each new idea.
sis — or virtually anything else that in-
cludes the possibility of creative adapta-
tion — will always be incomplete. We
will only be able to model effects on the
levels for which our models are defined.
The imaginations of any agents that we
model will be limited to the capabilities
built into the model.
9.7 Modeling:
the difficulty of
looking upward
We noted earlier that when a glider ap-
pears in the Game of Life, it has no ef-
fect on the how the system behaves. The
agents don ’ t see a glider coming and
duck. More significantly we don’ t know
how to build systems so that agents will
be able to notice gliders and duck.
It would be an extraordinary achieve-
ment in artificial intelligence to build a
modeling system that could notice emer-
gent phenomena and see how they could
be exploited. Yet we as human beings do
this all the time. The dynamism of a
free-market economy depends on our
ability to notice newly emergent patterns
and to find ways to exploit them.
Al Qaeda noticed that our commercial
airlines system can be seen as a network
of flying bombs. Yet no model of terror-
ism that doesn’ t have something like that
built into it will be able to make that sort
of creative leap. Our models are blind to
emergence even as it occurs within
them.
Notice that this is not the same as the
difficulty of looking downward. In the
Al Qaeda example one may assume that
one’ s model of the airline system in-
cludes the information that an airplane
when loaded with fuel will explode
when it crashes. The creative leap is to
notice that one can use that phenomenon
for new purposes. This is easier than the
Emergence Explained
52/67
DRAFT
2/12/2006
problem of looking downward. But it is
still a very difficult problem.
The moral is the same as before. Models
will always be incomplete. We will only
be able to model effects on the levels for
which our models are defined. The
imaginations of any agents that we
model will be limited to the capabilities
built into the model.
10 Concluding remarks
10.1 Computer Science and Phi-
losophy
• For centuries after Newton, nature was
seen as a perfect clockwork mecha-
nism.
• After the invention of the telephone,
the brain was likened to a switchboard.
Science and technology tends to shape
our view of the world. It ’ s not surprising
that the perspectives developed in this
article reflect those of Computer Sci-
ence, the predominant perspective of this
age. Is this parochialism? It ’ s difficult to
tell from so close. One thing is clear.
Because Computer Science has wres-
tled— with some success — with many
serious intellectual challenges, it is not
unreasonable to hope that the field may
contribute something to the broader in-
tellectual community.
It is useful to think of computers as reifi-
cation machines: they make the abstract
concrete. As such they are similar to the
drawing
tools of an animator— who
draws anything that crosses his or her
mind, no matter how wild or fanciful. In
the hands of a skilled and creative pro-
grammer,47 a computer and an associated
47 We use the term programmer deliberately. Fancier
terms like software developer (or software engi-
neer or worse information technology specialist)
lose the focus of what is really happening when
one creates software. A programmer (writes text
software development environment in-
vite virtually unlimited creativity.
But there is a major difference between
the product produced by an animator and
that produced by a programmer. An
animator produces images that have
meaning only when they enter the mind
of another human being. They do not
stand on their own. They are meant
strictly as communication from one hu-
man being to another. Nor are either the
physical medium on which the images
are stored or the mechanisms that causes
the images to be displayed of much sig-
nificance. The only thing that matters is
that the images arrive as intended in the
minds of the viewers.
A computer program is stuck in the real
world. It ’ s work is to shape the activity
of a computer, a physical device. Many
programs execute without even being
observed by people—
programs in cars
and other machinery, for example. And
even programs
that perform strictly
symbolic transformations, do work that
may be understood as constraining the
forces of nature (the motion of electrons)
to some formal shape.
To make the computer a useful reifica-
tion device—
to make it possible for pro-
grammers to write text that causes a
computer to convert programmers ’ fan-
tasies to some concrete form — Computer
Science has had to deal with some of
philosophy’ s thorniest issues.
• Computer Science has created lan-
guages that are both formally de-
that) programs, i.e., shapes, the way a computer
behaves. That ’ s really all that matters: what does
the text tell the computer to do?
In the preceding, note the stigmergy and down-
ward entailment. A program doesn ’ t tell a com-
puter anything. Successful programmers work
from that perspective all the time.
Emergence Explained
53/67
DRAFT
2/12/2006
fined— with formal syntax and seman-
tics —
and operational, i.e., they actu-
ally function in the real world.
• Computer Science has figured out how
to represent information in databases
in ways that allow that information to
hang together meaningfully.
• Computer Science has faced—
and to a
significant extent resolved—
the prob-
lem of working on many levels of ab-
straction and complexity simultane-
ously.
If insights gained from these and other
intellectual wrestling matches can be ap-
plied in a wider context, it is only Com-
puter Science paying back the debt that
it owes to the engineers, scientists,
mathematicians, and philosophers who
set the stage for and participated in its
development.
10.2 Constructive science
For most of its history, science has pur-
sued the goal of explaining existing phe-
nomena in terms of simpler phenomena.
That ’ s the reductionist agenda.
The approach we have taken is to ask
how new phenomena may be con-
structed from and implemented in terms
of existing phenomena. That ’ s the crea-
tive impulse of artists, computer scien-
tists, engineers —
and of nature. It is
these new phenomena that are often
thought of as emergent.
When thinking in the constructive direc-
tion, a question arises that is often under-
appreciated: what allows one to put ex-
isting things together to get something
new—
and something new that will per-
sist in the world? What binding forces
and binding strategies do we (and na-
ture) have at our disposal?
Our answer has been that there are two
sorts of binding strategies: energy wells
and energy-consuming processes. En-
ergy wells are reasonably well under-
stood —
although it is astonishing how
many different epiphenomena nature and
technology have produced through the
use of energy wells.
We have not even begun to catalog the
ways in which energy-consuming proc-
esses may be used to construct stable,
self-perpetuating, autonomous entities.
Earlier we wrote that science does not
consider it within its realm to ask con-
structivist questions. That is not com-
pletely true. Science asks about how we
got here from the big bang, and science
asks about biological evolution. These
are both constructivist questions. Since
science is an attempt to understand na-
ture, and since constructive processes
occur in nature, it is quite consistent with
the overall goals of science to ask how
these constructive processes work. As
far as we can determine, there is no sub-
discipline of science that asks, in gen-
eral, how the new arise from the exist-
ing.
Science has produced some specialized
answers to this question. The biological
evolutionary explanation involves ran-
dom mutation and crossover of design
records. The cosmological explanation
involves falling into energy wells of
various sorts. Is there any more to say
about how nature finds and then explores
new possibilities? If as Dennett argues in
[Dennett ‘ 96] this process may be fully
explicated as generalized Darwinian
evolution, questions still remain. Is there
any useful way to characterize the search
space that nature is exploring? What
search strategies does nature use to ex-
plore that space? Clearly one strategy is
human inventiveness.
Emergence Explained
54/67
DRAFT
2/12/2006
11 Acknowledgement
We are grateful for numerous enjoyable
and insightful discussions with Debora
Shuger during which many of the ideas
in this paper were developed and re-
fined.
We also wish to acknowledge the fol-
lowing websites and services, which we
used repeatedly.
• Google (www.google.com);
• The Stanford Encyclopedia of Phi-
losophy (plato.stanford.edu);
• OneLook
Dictionary
(onelook.com).
Search
References
Abbott, R., “ Emergence, Entities, En-
tropy, and Binding Forces,” The Agent
2004 Conference on: Social Dynamics:
Interaction, Reflexivity, and Emergence,
Argonne National Labs and University
of Chicago, October 2004. URL as of
4/2005:
http://abbott.calstatela.edu/PapersAndTa
lks/abbott_agent_2004.pdf.
American Heritage, The American Heri-
tage® Dictionary of the English Lan-
guage, 2000. URL as of 9/7/2005:
http://www.bartleby.com/61/67/S014670
0.html.
Anderson, P.W., “ More is Different,”
Science, 177 393-396, 1972.
BBC News, “ Gamer buys $26,500 vir-
tual land, ” BBC News, Dec. 17, 2004.
URL as of 2/2005:
http://news.bbc.co.uk/1/hi/technology/41
04731.stm.
Bedau, M.A., “ Downward causation and
the autonomy of weak emergence” .
Principia 6 (2002): 5-50. URL as of
11/2004:
http://www.reed.edu/~mab/papers/princi
pia.pdf.
Boyd, Richard, "Scientific Realism",
The Stanford Encyclopedia of Philoso-
phy (Summer 2002 Edition), Edward N.
Zalta (ed.), URL as of 9/01/2005:
http://plato.stanford.edu/archives/sum20
02/entries/scientific-realism/.
Brown, J.S., Talk at San Diego State
University, January 18, 2005. URL as of
6/2005:
http://ctl.sdsu.edu/pict/jsb_lecture18jan0
5.pdf
Carroll, S.B., Endless Forms Most Beau-
tiful: The New Science of Evo Devo and
the Making of the Animal Kingdom, W.
W. Norton, 2005.
Chaitin, G. Algorithmic Information
Theory, reprinted 2003. URL as of Sept.
6, 2005:
http://www.cs.auckland.ac.nz/CDMTCS/
chaitin/cup.pdf.
CFCS, Committee on the Fundamentals
of Computer Science: Challenges and
Opportunities, National Research Coun-
cil, Computer Science: Reflections on
the Field, Reflections from the Field,
2004. URL as of 9/9/2005:
http://www.nap.edu/books/0309093015/
html/65.html.
Church, G. M., “ From systems biology
to synthetic biology, ” Molecular
Systems Biology, March, 29, 2005. URL
as of 6/2005:
http://www.nature.com/msb/journal/v1/n
1/full/msb4100007.html.
Ciborra, C. "From Thinking to
Tinkering: The Grassroots of Strategic
Information Systems", The Information
Society 8, 297-309, 1992.
Clarke, A. C., "Hazards of Prophecy:
The Failure of Imagination,” Profiles of
The Future, Bantam Books, 1961.
Cockshott, P. and G. Michaelson, “ Are
There New Models of Computation: Re-
Emergence Explained
55/67
DRAFT
2/12/2006
ply to Wegner and Eberbach.” URL as
of Oct. 10, 2005:
http://www.dcs.gla.ac.uk/~wpc/reports/
wegner25aug.pdf.
Colbaugh, R. and Kristen Glass, “ Low
Cognition Agents: A Complex Networks
Perspective,” 3rd Lake Arrowhead Con-
ference on Human Complex Systems,
2005.
Collins, Steven, Martijn Wisse, and
Andy Ruina, “ A Three-Dimensional
Passive-Dynamic Walking Robot with
Two Legs and Knees,” The International
Journal of Robotics Research Vol. 20,
No. 7, July 2001, pp. 607-615, URL as
of 2/2005:
http://ruina.tam.cornell.edu/research/topi
cs/locomotion_and_robotics/papers/3d_p
as-
sive_dynamic/3d_passive_dynamic.pdf
Comm Tech Lab and the Center for Mi-
crobial Ecology, The Microbe Zoo, URL
as of Oct 10, 2005:
http://commtechlab.msu.edu/sites/dlc-
me/zoo/microbes/riftiasym.html
Comte, A. “ Positive Philosophy,” trans-
lated by Harriet Martineau, NY: Calvin
Blanchard, 1855. URL: as of 7/2005:
http://www.d.umn.edu/cla/faculty/jhamli
n/2111/ComteSimon/Comtefpintro.html.
Cowan, R., “ A spacecraft breaks open a
comet's secrets, ” Science News Online,
Vol. 168, No. 11 , p. 168, Sept. 10, 2005.
URL as of 9/9/2005:
http://www.sciencenews.org/articles/200
50910/bob9.asp.
Dennett, D. C., The Intentional Stance,
MIT Press/Bradford Books, 1987.
Dennett, D. C. “ Real Patterns,” The
Journal of Philosophy, (88, 1), 1991.
Dennett, D. C., Darwin's Dangerous
Idea: Evolution and the Meanings of
Life, V, 1996.
Dick, D., et. al., “ C2 Policy Evolution at
the U.S. Department of Defense,” 10th
International Command and Control Re-
search and Technology Symposium, Of-
fice of the Assistant Secretary of De-
fense, Networks and Information Inte-
gration (OASD-NII), June 2005. URL as
of 6/2005:
http://www.dodccrp.org/events/2005/10t
h/CD/papers/177.pdf.
Einstein, A., Sidelights on Relativity, An
address delivered at the University of
Leyden, May 5th, 1920. URL as of
6/2005:
http://www.gutenberg.org/catalog/world/
readfile?fk_files=27030.
Emmeche, C, S. Kø ppe and F. Stjernfelt,
“ Levels, Emergence, and Three Versions
of Downward Causation, ” in Andersen,
P.B., Emmeche, C., N. O. Finnemann
and P. V. Christiansen, eds. (2000):
Downward Causation. Minds, Bodies
and Matter. Å rhus: Aarhus University
Press. URL as of 11/2004:
http://www.nbi.dk/~emmeche/coPubl/20
00d.le3DC.v4b.html.
Fodor, J. A., “ Special Sciences (or the
disunity of science as a working hy-
pothesis),” Synthese 28: 97-115. 1974.
Fodor, J.A., “ Special Sciences; Still
Autonomous after All These Years,”
Philosophical Perspectives, 11, Mind,
Causation, and World, pp 149-163,
1998.
Fredkin, E., "Digital Mechanics",
Physica D, (1990) 254-270, North-
Holland. URL as of 6/2005: This and re-
lated papers are available as of 6/2005 at
the Digital Philosophy website, URL:
http://www.digitalphilosophy.org/.
Emergence Explained
56/67
DRAFT
2/12/2006
Gardner, M., Mathematical Games: “ The
fantastic combinations of John Conway's
new solitaire game ‘ life’ ," Scientific
American, October, November, Decem-
ber, 1970, February 1971. URL as of
11/2004:
http://www.ibiblio.org/lifepatterns/octob
er1970.html.
Grasse, P.P., “ La reconstruction du nid
et les coordinations inter-individuelles
chez Bellicosi-termes natalensis et Cubi-
termes sp. La theorie de la stigmergie:
Essai d'interpretation des termites cons-
tructeurs.” Ins. Soc., 6, 41-83, 1959.
Hardy, L., “ Why is nature described by
quantum theory? ” in Barrow, J.D.,
P.C.W. Davies, and C.L. Harper, Jr. Sci-
ence and Ultimate Reality, Cambridge
University Press, 2004.
Holland, J. Emergence: From Chaos to
Order, Addison-Wesley, 1997.
Hume, D. An Enquiry Concerning Hu-
man Understanding, Vol. XXXVII, Part
3. The Harvard Classics. New York: P.F.
Collier & Son, 1909– 14; Bartleby.com,
2001. URL a of 6/2005::
www.bartleby.com/37/3/.
Kauffman, S. “ Autonomous Agents,” in
Barrow, J.D., P.C.W. Davies, and C.L.
Harper, Jr. Science and Ultimate Reality,
Cambridge University Press, 2004.
Kim, J. “ Multiple realization and the
metaphysics of reduction,” Philosophy
and Phenomenological Research, v 52,
1992.
Kim, J., Supervenience and Mind. Cam-
bridge University Press, Cambridge,
1993.
Langton, C., "Computation at the Edge
of Chaos: Phase Transitions and Emer-
gent Computation." In Emergent Com-
putation, edited by Stephanie Forest.
The MIT Press, 1991.
Laughlin, R.B., A Different Universe,
Basic Books, 2005.
Laycock, Henry, "Object", The Stanford
Encyclopedia of Philosophy (Winter
2002 Edition), Edward N. Zalta (ed.),
URL as of 9/1/05:
http://plato.stanford.edu/archives/win200
2/entries/object/.
Leibniz, G.W., Monadology, for exam-
ple, Leibniz's Monadology, ed. James Fi-
eser (Internet Release, 1996). URL as of
9/16/2005:
http://stripe.colorado.edu/~morristo/mon
adology.html
Lowe, E. J., “ Things,” The Oxford Com-
panion to Philosophy, (ed T. Honderich),
Oxford University Press, 1995.
Maturana, H. & F. Varela, Autopoiesis
and Cognition: the Realization of the
Living., Boston Studies in the
Philosophy of Science, #42, (Robert S.
Cohen and Marx W. Wartofsky Eds.), D.
Reidel Publishing Co., 1980.
Miller, Barry, "Existence", The Stanford
Encyclopedia of Philosophy (Summer
2002 Edition), Edward N. Zalta (ed.),
URL as of 9/1/05:
http://plato.stanford.edu/archives/sum20
02/entries/existence/.
NASA (National Aeronautics and Space
Administration), “ Hurricanes: The
Greatest Storms on Earth,” Earth Obser-
vatory. URL as of 3/2005
http://earthobservatory.nasa.gov/Library/
Hurricanes/.
Nave, C. R., “ Nuclear Binding Energy” ,
Hyperphysics, Department of Physics
and Astronomy, Georgia State Univer-
sity. URL as of 6/2005:
http://hyperphysics.phy-
astr.gsu.edu/hbase/nucene/nucbin.html.
NOAA, Glossary of Terminology, URL
as of 9/7/2005:
Emergence Explained
57/67
DRAFT
2/12/2006
http://www8.nos.noaa.gov/coris_glossar
y/index.aspx?letter=s.
O'Connor, Timothy, Wong, Hong Yu
"Emergent Properties", The Stanford
Encyclopedia of Philosophy (Summer
2005 Edition), Edward N. Zalta (ed.),
forthcoming URL:
http://plato.stanford.edu/archives/sum20
05/entries/properties-emergent/.
Prigogine, Ilya and Dilip Kondepudi,
Modern Thermodynamics: from Heat
Engines to Dissipative Structures, John
Wiley & Sons, N.Y., 1997.
Ray, T. S. 1991. “ An approach to the
synthesis of life, ” Artificial Life II, Santa
Fe Institute Studies in the Sciences of
Complexity, vol. XI, Eds. C. Langton, C.
Taylor, J. D. Farmer, & S. Rasmussen,
Redwood City, CA: Addison-Wesley,
371--408. URL page for Tierra as of
4/2005:
http://www.his.atr.jp/~ray/tierra/.
Rendell, Paul, “ Turing Universality in
the Game of Life,” in Adamatzky, An-
drew (ed.), Collision-Based Computing,
Springer, 2002. URL as of 4/2005:
http://rendell.server.org.uk/gol/tmdetails.
htm,
http://www.cs.ualberta.ca/~bulitko/F02/
papers/rendell.d3.pdf, and
http://www.cs.ualberta.ca/~bulitko/F02/
papers/tm_words.pdf
Rosen, Gideon, "Abstract Objects", The
Stanford Encyclopedia of Philosophy
(Fall 2001 Edition), Edward N.
Zalta (ed.), URL as of 9/1/05:
http://plato.stanford.edu/archives/fall200
1/entries/abstract-objects/.
Sachdev, S, “ Quantum Phase Transi-
tions,” in The New Physics, (ed G. Fra-
ser), Cambridge University Press, (to
appear 2006). URL as of 9/11/2005:
http://silver.physics.harvard.edu/newphy
sics_sachdev.pdf.
Shalizi, C., Causal Architecture, Com-
plexity and Self-Organization in Time
Series and Cellular Automata, PhD. Dis-
sertation, Physics Department, Univer-
sity of Wisconsin-Madison, 2001. URL
as of 6/2005:
http://cscs.umich.edu/~crshalizi/thesis/si
ngle-spaced-thesis.pdf
Shalizi, C., “ Review of Emergence from
Chaos to Order,” The Bactra Review,
URL as of 6/2005:
http://cscs.umich.edu/~crshalizi/reviews/
holland-on-emergence/
Shalizi, C., “ Emergent Properties,”
Notebooks, URL as of 6/2005:
http://cscs.umich.edu/~crshalizi/noteboo
ks/emergent-properties.html.
Smithsonian Museum, “ Chimpanzee
Tool Use,” URL as of 6/2005:
http://nationalzoo.si.edu/Animals/Think
Tank/ToolUse/ChimpToolUse/default.cf
m.
Summers, J. “ Jason’ s Life Page,” URL
as of 6/2005:
http://entropymine.com/jason/life/.
Trani, M. et. al., “ Patterns and trends of
early successional forest in the eastern
United States,” Wildlife Society Bulle-
tin, 29(2), 413-424, 2001. URL as of
6/2005:
http://www.srs.fs.usda.gov/pubs/rpc/200
2-01/rpc_02january_31.pdf.
University of Delaware, Graduate Col-
lege of Marine Studies, Chemosynthesis,
URL as of Oct 10, 2005:
http://www.ocean.udel.edu/deepsea/level
-2/chemistry/chemo.html
Uvarov, E.B., and A. Isaacs, Dictionary
of Science, September, 1993. URL as of
9/7/2005:
http://oaspub.epa.gov/trs/trs_proc_qry.na
vi-
Emergence Explained
58/67
DRAFT
2/12/2006
gate_term?p_term_id=29376&p_term_c
d=TERMDIS.
Varzi, Achille, "Boundary", The Stan-
ford Encyclopedia of Philosophy (Spring
2004 Edition), Edward N. Zalta (ed.),
URL as of 9/1/2005:
http://plato.stanford.edu/archives/spr200
4/entries/boundary/.
Varzi, A., "Mereology", The Stanford
Encyclopedia of Philosophy (Fall 2004
Edition), Edward N. Zalta (ed.), URL as
of 9/1/2005:
http://plato.stanford.edu/archives/fall200
4/entries/mereology/ .
Wallace, M., “ The Game is Virtual. The
Profit is Real.” The New York Times,
May 29, 2005. URL of abstract as of
6/2005:
http://query.nytimes.com/gst/abstract.ht
ml?res=F20813FD3A5D0C7A8EDDAC
0894DD404482.
Wegner, P. and E. Eberbach, “ New
Models of Computation, ” Computer
Journal, Vol 47, No. 1, 2004.
Wegner, P. and D.. Goldin, “ Computa-
tion beyond Turing Machines ” , Commu-
nications of the ACM, April 2003. URL
as of 2/22/2005:
http://www.cse.uconn.edu/~dqg/papers/c
acm02.rtf.
Weinberg, S., “ Reductionism Redux,”
The New York Review of Books, October
5, 1995. Reprinted in Weinberg, S., Fac-
ing Up, Harvard University Press, 2001.
URL as of 5/2005 as part of a discussion
of reductionism:
http://pespmc1.vub.ac.be/AFOS/Debate.
html
Wiener, P.P., Dictionary of the History
of Ideas, Charles Scribner's Sons, 1973-
74. URL as of 6/2005:
http://etext.lib.virginia.edu/cgi-
local/DHI/dhi.cgi?id=dv4-42.
WordNet 2.0, URL as of 6/2005:
www.cogsci.princeton.edu/cgi-
bin/webwn.
Wolfram, S., A New Kind of Science,
Wolfram Media, 2002. URL as of
2/2005:
http://www.wolframscience.com/nksonli
ne/toc.html.
Woodward, James, "Scientific Explana-
tion", The Stanford Encyclopedia of Phi-
losophy (Summer 2003 Edition), Edward
N. Zalta (ed.). URL as of 9/13/2005:
http://plato.stanford.edu/archives/sum20
03/entries/scientific-explanation/.
Zuse, K., “ Rechnender Raum ” (Vieweg,
Braunschweig, 1969); translated as Cal-
culating Space, MIT Technical Transla-
tion AZT-70-164-GEMIT, MIT (Project
MAC), Cambridge, Mass. 02139, Feb.
1970. URL as of 6/2005:
ftp://ftp.idsia.ch/pub/juergen/zuserechne
nderraum.pdf.
Emergence Explained
59/67
DRAFT
2/12/2006
12 Appendix. Game of Life Patterns
Intuitively, a Game of Life pattern is the
step-by-step time and space progression
on a grid of a discernable collection of
inter-related live cells. We formalize that
notion in three steps.
1. First we define a static construct
called the live cell group. This will
be a group of functionally isolated
but internally interconnected cells.
2. Then we define Game of Life basic
patterns as temporal sequences of
live cell groups. The Game of Life
glider and still-life patterns are ex-
amples
3. Finally we extend the set of patterns
to include combinations of basic pat-
terns. The more sophisticated Game
of Life patterns, such the glider gun,
are examples.
12.1 Live cell groups
The fundamental construct upon which
we will build the notion of a pattern is
what we shall call a live cell group.
A live cell group is a collection of live
and dead cells that have two properties.
1. They are functionally isolated from
other live cells.
2. They are functionally related to each
other.
More formally, we define cells c0 and cn
in a Game of Life grid to be connected if
there are cells c1, c2, …
, cn-1 such that for
all i in 0 .. n-1
1. ci and ci+1 are neighbors, as defined
by Game of Life, and
2. either ci or ci+1 (or both) are alive, as
defined by Game of Life.
Connectedness is clearly an equivalence
relation (reflexive, symmetric, and tran-
sitive), which partitions a Game of Life
board into equivalence classes of cells.
Every dead cell that is not adjacent to a
live cell (does not have a live cell as a
Game of Life neighbor) becomes a sin-
gleton class.
Consider only
those connectedness
equivalence classes that include at least
one live cell. Call such an equivalence
class a live cell group or LCG.
Define the state of an LCG as the spe-
cific configuration of live and dead cells
in it. Thus, each LCG has a state.
No limitation is placed on the size of an
LCG. Therefore, if one does not limit the
size of the Game of Life grid, the num-
ber of LCGs is unbounded.
Intuitively, an LCG is a functionally iso-
lated group of live and dead cells, con-
tained within a boundary of dead cells.
Each cell in an LCG is a neighbor to at
least one live cell within that LCG.
As a consequence of this definition, each
live cell group consists of an “ inside,”
which contains all its live cells (possibly
along with some dead cells), plus a “ sur-
face ” or “ boundary” of dead cells. (The
surface or boundary is also considered
part of the LCG.)
12.2 Basic patterns: temporal se-
quences of live cell groups
Given this definition, we can now build
temporal sequences of LCGs. These will
be the Game of Life basic patterns.
The Game of Life rules define transi-
tions for the cells in a LCG. Since an
LCG is functionally isolated from other
live cells, the new states of the cells in
Emergence Explained
60/67
DRAFT
2/12/2006
an LCG are determined only by other
cells in the same LCG.48
contain multiple elements. If (cid:1) ’ is a
member of Game of Life((cid:1)) write (cid:1) -> (cid:1) ’ .
Suppose that an LCG contains the only
live cells on a Game of Life grid. Con-
sider what the mapping of that LCG by
the Game of Life rules will produce.
There are three possibilities.
1. The live cells may all die.
2. The successor live cells may consist
of a single LCG—
as in a glider or
still life.
3. The successor live cells may parti-
tion into multiple LCGs —
as in the
so-called bhepto pattern, which starts
as a single LCG and eventually stabi-
lizes as 4 still life LCGs and two
glider LCGs.
In other words, the live cells generated
when the Game of Life rules are applied
to an LCG will consist of 0, 1, or multi-
ple successor LCGs.
More formally, if (cid:1) is an LCG, let Game
of Life((cid:1)) be the set of LCGs that are
formed by applying the Game of Life
rules to the cells in (cid:1). For any particular
(cid:1), Game of Life((cid:1)) may be empty; it may
be contain a single element; or it may
48 In particular, no LCG cells have live neighbors
that are outside the LCG. Thus no cells outside
the LCG need be considered when determining
the GoL transitions of the cells in an LCG. A
dead boundary cell may become live at the
next time-step, but it will do so only if three of
its neighbors within the LCG are live. Its
neighbors outside the LCG are guaranteed to
be dead.
If a boundary cell does become live, the next-
state LCG of which it is a member will include
cells that were not part of its predecessor LCG.
For any LCG (cid:1) 0, consider a sequence of
successor LCGs generated in this man-
ner:
(cid:1)0 -> (cid:1)1 -> (cid:1) 2 -> (cid:1)3 -> …
.
Extend such a sequence until one of
three conditions occurs.
1. There are no successor LCGs, i.e.,
Game of Life((cid:1)i) is empty—
all the
live cells in the final LCG die. Call
these terminating sequences.
2. There is a single successor LCG, i.e.,
Game of Life((cid:1)i) = {(cid:1)k}, but that suc-
cessor LCG is in the same state as an
LCG earlier in the sequence, i.e., (cid:1)k =
(cid:1)j, j < k. Call these repeating se-
quences.
3. The set Game of Life((cid:1)i) of successor
LCGs contains more than one LCG,
i.e., the LCG branches into two or
more LCGs. Call these branching se-
quences.
Note that some LCG sequences may
never terminate. They may simply pro-
duce larger and larger LCGs. The so-
called spacefiller pattern, which actually
consists of multiple interacting LCGs,
one of which fills the entire grid with a
single LCG as it expands,49 is an amaz-
ing example of such a pattern. I do not
know if there is an LCG that expands
without limit on its own. If any such ex-
ist, call these infinite sequences.
For any LCG (cid:1)0, if the sequence
49 See
on
pattern
spacefiller
the
http://www.math.com/students/wonders/life/life.ht
ml or http://www.ibiblio.org/lifepatterns.
Emergence Explained
61/67
DRAFT
2/12/2006
(cid:1)0 -> (cid:1)1 -> (cid:1)2 -> (cid:1)3 -> …
.
is finite, terminating in one of the three
ways described above, let seq((cid:1)0) be that
sequence along with a description of
how it terminates. If
(cid:1)0 -> (cid:1)1 -> (cid:1)2 -> (cid:1)3 -> …
.
is infinite, then seq((cid:1)0) is undefined.
Let BP (for Basic Patterns) be the set of
finite non-branching sequences as de-
fined above. That is,
BP = {seq((cid:1)0) (cid:1)0 is an LCG}
Note that it is not necessary to extend
these sequences backwards. For any
LCG (cid:1)0, one could define the pre-image
of (cid:1)0 under the Game of Life rules.
Game of Life-1((cid:1)) is the set of LCGs (cid:1) ’
such that Game of Life((cid:1) ’ ) = (cid:1).
For any chain seq((cid:1)0) in BP, one could
add all the chains constructed by prefix-
ing to seq((cid:1)0) each of the predecessors (cid:1) ’
of (cid:1)0 (cid:1) ’ as long as (cid:1) ’ does not appear in
seq((cid:1) ’ ). But augmenting BP in this way
would add nothing to BP since by defini-
tion seq((cid:1) ’ ) is already defined to be in
BP for each (cid:1) ’ .
We noted above that we do not know if
there are unboundedly long sequences of
LCGs beginning with a particular (cid:1)0.
With
long
to unboundedly
respect
predecessor chains, it is known that such
unbounded predecessor chains (of un-
boundedly large LCGs) exist. The so-
called fuse and wick patterns are LCG
sequences that can be extended arbitrar-
ily far backwards.50 When run forward
50 A simple fuse pattern is a diagonal configuration
of live cells. At each time step, the two end cells
such fuse or wick LCGs converge to a
single LCG. Yet given the original defi-
nition of BP even these LCG sequences
are included in it. Each of these un-
bounded predecessor chains is included
in BP starting at each predecessor LCG.
Clearly BP as defined includes many re-
dundant pattern descriptions. No attempt
is made to minimize BP either for sym-
metries or for overlapping patterns in
which one pattern is a suffix of an-
other—
as in the fuse patterns. In a com-
puter program that generated BP, such
efficiencies would be important.
12.3 BP is recursively enumerable
The set BP of basic Game of Life pat-
terns may be constructed through a for-
mal iterative process. The technique em-
ployed is that used for the construction
of many recursively enumerable sets.
1. Generate the LCGs in sequence.
2. As each new LCG is generated, gen-
erate the next step in each of the se-
quences starting at each of the LCGs
generated so far.
3. Whenever an LCG sequence termi-
nates according to the BP criteria,
add it to BP.
The process sketched above will effec-
tively generate all members of BP. Al-
though theoretically possible, such a
procedure will be so inefficient that it is
useless for any practical purpose.51 The
die; the remaining cells remain alive. A simple
fuse pattern may be augmented by adding more
complex features at one end, thereby building a
pattern that becomes active when the fuse exhausts
itself. Such a pattern can be built with an arbitrar-
ily long fuse.
51 Many much more practical and efficient programs
have been written to search for patterns in the GoL
and
See
automata.
cellular
related
http://www.ics.uci.edu/~eppstein/ca/search.html
for a list of such programs.
Emergence Explained
62/67
DRAFT
2/12/2006
only reason to mention it here is to es-
tablish that BP is recursively enumer-
able. Whether BP is recursive depends
on whether one can in general establish
for any LCG (cid:1)0 whether seq((cid:1)0) will ter-
minate.
12.4 Game of Life patterns: com-
binations of basic patterns
Many of the interesting Game of Life
patterns arise from interactions between
and among basic patterns. For example,
the first pattern that generated an unlim-
ited number of live cells, the glider gun,
is a series of interactions among combi-
nations of multiple basic patterns that
cyclically generate gliders.
To characterize these more complex pat-
terns it is necessary to keep track of how
basic patterns interact. In particular, for
each element in BP, augment its descrip-
tion with information describing
a) its velocity (rate, possibly zero, and
direction) across the grid,
b) if it cycles, how it repeats, i.e., which
states comprise its cycle, and
c) if it branches, what the offspring
elements are and where they appear
relative to final position of the termi-
nating sequence.
Two or more distinct members of BP
that at time step i are moving relative to
each other may interact to produce one
or more members of BP at time step i+1.
The result of such a BP “ collision” will
generally depend on the relative posi-
tions of the interacting basic patterns.
Even though the set BP of basic patterns
is infinite, since each LCG is finite, by
using a technique similar to that used for
generating BP itself, one can (very tedi-
ously) enumerate all the possible BP in-
teractions.
More formally, let (cid:2)f(BP) be the set of
all finite subsets of BP. For each mem-
ber of (cid:2)f(BP) consider all possible (still
only a finite number) relative configura-
tions of its members on the grid so that
there will be some interaction among
them at the next time step. One can then
record all
the possible
interactions
among finite subsets of BP.
These interactions would be equivalent
to the APIs for the basic patterns. We
could call a listing of them BP-API.
Since BP is itself infinite, BP-API would
also be infinite. But BP-API would be
effectively searchable. Given a set of
elements in BP, one could retrieve all the
interactions among those elements. BP-
API would then provide a documented
starting point for using the Game of Life
as a programming language.
As in traditional programming lan-
guages, as more complex interactions are
developed, they too could be docu-
mented and made public for others to
use.
Emergence Explained
63/67
DRAFT
2/12/2006
Figures and Tables
Table 1. Dissipative structures vs. self-perpetuating entities
Self-perpetuating entities
Dissipative structures
Pure epiphenomena, e.g., 2-chamber example. Has functional design, e.g., hurricane.
Artificial boundaries.
Self-defining boundaries
Externally maintained energy gradient.
Imports, stores, and internally distributes
energy.
Emergence Explained
64/67
DRAFT
2/12/2006
Figure 1. Bit 3 off and then on.
Emergence Explained
65/67
DRAFT
2/12/2006
Figure 2. A glider
Emergence Explained
66/67
DRAFT
2/12/2006
Figure 3. Anatomy of a hurricane. [Image from [NASA].]
Emergence Explained
67/67
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.